Cross-Validation in Geometric Morphometrics: A Performance Review of Protocols for Biomedical and Clinical Research

Abstract

Geometric morphometrics (GM) is a powerful statistical tool for quantifying biological shape, with growing applications in clinical and pharmaceutical research. The reliability of its findings, however, hinges on the rigorous cross-validation of analytical protocols. This article provides a comprehensive review of GM cross-validation performance across diverse methodologies, from foundational landmark-based analyses and emerging functional data approaches to comparisons with machine learning. We explore common analytical pitfalls, offer optimization strategies for robust out-of-sample classification, and discuss the critical role of protocol validation in translating morphometric findings into reliable biomedical applications, such as personalized drug delivery and forensic anthropology.

Core Principles and Performance Challenges in Geometric Morphometrics

Defining Cross-Validation in the Context of Geometric Morphometric Workflows

Geometric morphometrics (GM) relies on sophisticated statistical models to quantify and analyze biological shape, making robust validation protocols essential for reliable results. Cross-validation serves as a critical methodology for assessing the generalizability and predictive performance of these models, guarding against overfitting—a significant risk given the high-dimensional nature of morphometric data. This guide objectively compares the cross-validation performance of various geometric morphometric protocols, including semi-landmark methods, outline-based analyses, and different dimensionality reduction techniques. We synthesize experimental data from multiple studies to provide researchers with evidence-based recommendations for optimizing their analytical workflows.

In geometric morphometrics, cross-validation provides a more reliable estimate of a model's classification accuracy than resubstitution methods, which are known to be biased upward as they use the same data to build and test the model [1]. The fundamental risk in GM analyses, particularly when using canonical variates analysis (CVA) for classification, is the high variable-to-specimen ratio. When outlines or curves are represented by numerous semi-landmarks, the number of parameters dramatically increases, demanding larger sample sizes for stable results [1]. Cross-validation, particularly leave-one-out cross-validation, mitigates this by iteratively training the model on all but one specimen and testing on the excluded one, providing a less biased performance estimate [1] [2].

The choice of cross-validation strategy becomes paramount when evaluating different GM protocols. Studies demonstrate that optimal performance depends on the complex interaction between data acquisition methods, alignment algorithms, and dimensionality reduction techniques [1]. Furthermore, the challenge of out-of-sample classification—applying a classification rule derived from a reference sample to new individuals not included in the original analysis—represents a critical extension of cross-validation principles in applied contexts [2]. The following sections compare these protocols quantitatively, using cross-validation performance as the key metric for evaluation.

Comparative Performance of Geometric Morphometric Protocols

Semi-Landmark and Outline Analysis Methods

Table 1: Comparison of Cross-Validation Performance for Different GM Methods

Method Category	Specific Method	Application Context	Reported Cross-Validation Accuracy	Key Findings
Semi-Landmark Alignment	Bending Energy Minimization (BEM)	Feather shape (Ovenbird)	Roughly equal classification rates [1]	Performance not highly dependent on number of points or acquisition method.
Semi-Landmark Alignment	Perpendicular Projection (PP)	Feather shape (Ovenbird)	Roughly equal classification rates [1]	Performance not highly dependent on number of points or acquisition method.
Outline-Based Analysis	Elliptical Fourier Analysis (EFA)	Feather shape (Ovenbird)	Roughly equal classification rates [1]	Comparable performance to extended eigenshape and semi-landmark methods.
Outline-Based Analysis	Extended Eigenshape Analysis	Feather shape (Ovenbird)	Roughly equal classification rates [1]	Comparable performance to Fourier and semi-landmark methods.
Semi-Landmark & Fourier	Outline and Semi-Landmark	Carnivore tooth marks	Low accuracy (~40%) [3]	Bi-dimensional application showed limited discriminant power.
Geometric Morphometrics	Landmark-based CVA	Malocclusion (Cephalograms)	80% after cross-validation [4]	High discrimination among malocclusion classes (I, II, III).

The classification of specimens based on shape appears less dependent on the specific choice of outline method than previously assumed. Research on ovenbird rectrices found that two semi-landmark methods (Bending Energy Minimization and Perpendicular Projection) produced roughly equal classification rates, as did Elliptical Fourier methods and the extended eigenshape method [1]. This suggests that for many biological applications, the choice between these established methods may not be the primary factor influencing predictive success.

However, significant performance limitations emerge when these methods are applied to certain real-world problems. A study on carnivore tooth marks found that both outline (Fourier) and semi-landmark approaches achieved low discriminant accuracy, below 40%, for identifying the carnivore modifying agent [3]. This highlights that methodological performance is context-dependent, and bi-dimensional information alone can sometimes be insufficient for complex classification tasks. In contrast, a landmark-based CVA on lateral cephalograms for malocclusion classification achieved a high cross-validation accuracy of 80%, demonstrating the method's power in clinical dental contexts [4].

Dimensionality Reduction and Machine Learning Approaches

Table 2: Comparison of Dimensionality Reduction and Classification Techniques

Technique	Purpose	Key Feature	Cross-Validation Performance
Variable PC Axes	Dimensionality Reduction	Uses number of PC axes that optimizes cross-validation rate [1]	Produced higher cross-validation assignment rates than fixed PC or PLS [1]
Fixed PC Axes	Dimensionality Reduction	Uses a fixed number of PC axes (e.g., all with non-zero eigenvalues) [1]	Lower cross-validation rates due to potential overfitting [1]
Partial Least Squares (PLS)	Dimensionality Reduction	Finds axes with greatest covariation with classification variables [1]	Lower cross-validation rates than variable PC axes method [1]
Supervised Machine Learning	Classification	Uses classifiers like LDA on aligned coordinates [2]	More accurate than PCA for classification and detecting new taxa [5]
Computer Vision (DCNN)	Classification	Deep Convolutional Neural Networks on images [3]	81% accuracy for tooth pit classification [3]
Computer Vision (FSL)	Classification	Few-Shot Learning models on images [3]	79.52% accuracy for tooth pit classification [3]

The approach to dimensionality reduction preceding CVA is a more significant factor for cross-validation performance than the choice of outline method. A variable number of Principal Component (PC) axes approach, which selects the number of PCs that maximize the cross-validation assignment rate, outperformed both the standard fixed-number approach and a Partial Least Squares (PLS) method [1]. Using a fixed number of PC axes (often all axes with non-zero eigenvalues) can lead to high resubstitution rates but substantially lower cross-validation rates due to overfitting, where discriminant axes become too tailored to the specific sample [1].

Emerging evidence challenges the standard PCA-based workflow. A benchmark study on papionin crania found that PCA outcomes are "artefacts of the input data" and are "neither reliable, robust, nor reproducible," while supervised machine learning classifiers provided more accurate classification [5]. Similarly, in a challenging domain like carnivore tooth mark identification, Computer Vision methods like Deep Convolutional Neural Networks (DCNN) and Few-Shot Learning (FSL) models significantly outperformed traditional GM, achieving accuracies of 81% and 79.52%, respectively [3]. This indicates a potential paradigm shift towards machine learning for complex morphometric classification tasks.

Detailed Experimental Protocols and Workflows

Standard GM Workflow with Cross-Validation

The following diagram illustrates the standard geometric morphometric workflow integrated with cross-validation, as applied in studies comparing methodological performance [1] [4] [2].

The standard workflow begins with Generalized Procrustes Analysis (GPA), which superimposes landmark configurations by translating, rescaling, and rotating them to minimize the sum of squared distances between corresponding landmarks, thus eliminating non-shape variations [4]. The resulting Procrustes coordinates are then subjected to dimensionality reduction, typically via Principal Component Analysis (PCA), to address the high dimensionality of the data [1] [5]. The reduced data serves as input for a classification model like Canonical Variates Analysis (CVA) or Linear Discriminant Analysis (LDA). The critical cross-validation step, often leave-one-out, involves iteratively refitting the model while holding out one specimen to test classification accuracy, providing a robust performance estimate [1] [2].

Protocol for Out-of-Sample Classification

A significant challenge in applied morphometrics is classifying new individuals not included in the original sample. The following workflow, derived from nutritional assessment research, addresses this [2].

This protocol requires selecting a template configuration from the training sample to serve as a target for registering the raw coordinates of a new individual [2]. This registration step is crucial for placing the new specimen into the same shape space as the training data, enabling the application of a pre-derived classification rule. The choice of template—such as the mean shape of the sample or a representative specimen—can influence classification performance and must be carefully considered [2]. This workflow is essential for real-world applications like the Severe Acute Malnutrition (SAM) Photo Diagnosis App, which classifies children's nutritional status from arm shape images without including them in the original model training [2].

Research Reagent Solutions: Essential Materials and Tools

Table 3: Key Software and Tools for Geometric Morphometric Analysis

Tool Name	Type	Primary Function in GM Workflow	Application Example
MorphoJ	Software	Statistical analysis and visualization of shape data [4]	Malocclusion classification from cephalograms [4]
tpsDig2 / tpsUtil	Software	Digitizing landmarks and managing landmark data files [6]	Acquiring 2D coordinates from specimen images [6]
geomorph	R Package	GM analysis including Procrustes ANOVA and phylogenetic comparisons [7]	Complex statistical modeling of shape data [7]
Momocs	R Package	Outline analysis, including Elliptical Fourier Analysis [7]	Analyzing closed outlines of structures [7]
morphospace	R Package	Building and visualizing ordinations of shape data [7]	Creating publication-ready morphospace plots [7]
MORPHIX	Python Package	Supervised machine learning classification of landmark data [5]	Alternative to PCA-based classification [5]

The analytical tools listed above form the backbone of modern geometric morphometric research. MorphoJ is a widely used standalone application for performing essential GM operations, including Generalized Procrustes Analysis, Principal Component Analysis, and Discriminant Function Analysis with cross-validation [4]. The tps software suite, particularly tpsDig2 and tpsUtil, is fundamental for the initial stages of data acquisition and management, allowing researchers to digitize landmarks and organize data files [6].

The R statistical environment hosts several powerful packages that extend analytical capabilities. The geomorph package provides tools for complex analyses, such as Procrustes ANOVA, and for integrating phylogenetic information [7]. Momocs is specialized for handling outline data through methods like Elliptical Fourier Analysis [7]. The newer morphospace package streamlines the creation and visualization of ordinations, enhancing the biological interpretation of results [7]. For researchers seeking alternatives to traditional PCA-based classification, MORPHIX is a Python package that implements supervised machine learning classifiers for landmark data, reportedly offering higher accuracy [5].

The cross-validation performance of geometric morphometric protocols is influenced by multiple factors, with the choice of dimensionality reduction technique and classifier often mattering more than the specific type of outline method. Based on the synthesized experimental data:

• For traditional GM workflows, a variable PC axes approach to dimensionality reduction is recommended over fixed PC axes or PLS, as it directly optimizes cross-validation performance [1].
• When classification accuracy is the primary goal, supervised machine learning classifiers (as implemented in MORPHIX) and computer vision approaches should be considered, as they have demonstrated superior performance compared to standard PCA-based methods in several benchmark studies [3] [5].
• Researchers must be cautious of overfitting when the number of variables (landmarks, semi-landmarks) is high relative to the sample size. Cross-validation is non-negotiable for obtaining realistic performance estimates [1] [2].
• For applied systems that classify new individuals, a robust protocol for out-of-sample classification must be implemented, carefully considering the selection of an appropriate template for registration [2].

Future research should continue to bridge traditional morphometric methods with modern machine learning, validate protocols on diverse datasets, and develop standardized workflows for out-of-sample prediction to enhance the reliability and applicability of geometric morphometrics.

Geometric morphometrics (GM) has become a foundational tool for quantifying biological shape across diverse scientific fields, from paleontology to drug development. The standard analytical protocol in GM consistently relies on a two-step process: Generalized Procrustes Analysis (GPA) for shape alignment, followed by Principal Component Analysis (PCA) for dimensionality reduction and visualization of shape variation [8] [9]. This combination is considered the cornerstone of modern shape analysis.

However, within the context of broader research on the cross-validation performance of different geometric morphometric protocols, critical questions arise: How reliable and robust are the conclusions drawn from this standard GPA-PCA pipeline? Can researchers confidently use this protocol for taxonomic classification, clinical prediction, or evolutionary inference? Recent studies have begun to systematically evaluate this workflow, testing its limits and comparing its performance against emerging methodologies, including various machine learning (ML) classifiers [8] [10]. This guide provides an objective comparison of the GPA-PCA protocol's performance against alternative approaches, supported by experimental data.

The Standard GPA-PCA Workflow and Its Applications

The conventional geometric morphometric pipeline involves a series of structured steps to transform raw coordinate data into interpretable shape variables.

Core Procedural Steps

• Landmarking: Anatomically homologous points are digitized on biological structures. Studies use Type I, II, and III landmarks, often supplemented with semi-landmarks to capture the geometry of curves and surfaces [8] [11]. For instance, a study on killer whales used 6 landmarks on aerial images to assess body shape related to reproductive status [12].
• Generalized Procrustes Analysis (GPA): This step removes non-shape variations by superimposing landmark configurations. It uses least-squares estimates to optimally translate, rotate, and scale all specimens to a common coordinate system, effectively isolating pure shape information [9] [10].
• Principal Component Analysis (PCA): The aligned Procrustes coordinates are projected onto a new set of uncorrelated variables—the Principal Components (PCs). These PCs describe the major axes of shape variation within the dataset, typically visualized using 2D or 3D scatterplots [8].

Diverse Scientific Applications

The GPA-PCA pipeline has been successfully applied across numerous domains, demonstrating its utility as a versatile tool for shape-based classification and hypothesis testing.

Table 1: Applications of the Standard GPA-PCA Protocol in Research

Field of Study	Biological Structure	Research Objective	Key Finding
Anesthesiology [10]	Human Face (3D Scan)	Predict Difficult Mask Ventilation (DMV)	Significant morphological difference in the mandibular region identified between DMV and easy mask ventilation groups.
Paleontology [13]	Fossil Shark Teeth	Support Taxonomic Identification	Geometric morphometrics validated qualitative taxonomic separation and captured more morphological information than traditional morphometrics.
Ecology [12]	Killer Whale Body	Detect Reproductive Status from Aerial Images	Significant separation of body shapes between most reproductive statuses (e.g., non-pregnant vs. late-stage pregnant).
Personalized Medicine [11]	Human Nasal Cavity	Classify Olfactory Accessibility for Drug Delivery	Identified three distinct morphological clusters of the nasal cavity, influencing accessibility to the olfactory region.
Taxonomy [9]	Shrew Crania	Classify Three Shrew Species	Functional Data GM (FDGM) combined with PCA and LDA outperformed classical GM in species classification.

Reliability Assessment: A Critical Evaluation

A growing body of literature critically examines the reliability of the standard GPA-PCA protocol, often through direct comparison with other statistical and machine learning methods.

Performance in Classification and Prediction

Comparative studies consistently reveal that while the GPA-PCA pipeline is a powerful exploratory tool, its performance in classification tasks can be surpassed by other methods.

Table 2: Comparative Performance of GPA-PCA vs. Alternative Methods

Study Context	Comparison	Performance Outcome	Reference
Difficult Mask Ventilation Prediction [10]	PCA-based vs. 10 Machine Learning models on 3D facial scans.	The best ML model (Logistic Regression) achieved an AUC of 0.825, outperforming the traditional DIFFMASK score (AUC 0.785). PCA was part of the feature extraction, but ML improved classification.
Shrew Species Classification [9]	Classical GM (PCA+LDA) vs. Functional Data GM (FDGM) with ML.	FDGM combined with machine learning (e.g., SVM, Random Forest) demonstrated better classification accuracy for shrew species than classical GM.
Papionin Crania Classification [8]	Standard PCA vs. Supervised Machine Learning classifiers.	Supervised ML classifiers were found to be more accurate than PCA for both classification and detecting new taxa.
Nasal Cavity Clustering [11]	PCA for identifying morphological clusters.	PCA successfully identified three distinct morphological clusters of the nasal cavity, demonstrating its continued utility for uncovering latent group structures.

Identified Limitations and Criticisms of PCA

The central role of PCA in GM has recently been challenged. A compelling critique argues that PCA outcomes can be "artefacts of the input data" and are neither reliable, robust, nor reproducible as often assumed by researchers [8]. The main criticisms include:

• Subjectivity in Interpretation: Biological conclusions about relatedness, evolution, and taxonomy are often drawn from the visual clustering of samples on PCA scatterplots (e.g., PC1 vs. PC2). This interpretation is highly subjective, and different PC combinations can yield conflicting results, leading to questionable taxonomic decisions [8].
• Statistical Artefacts, Not Biological Reality: Principal Components are statistical constructs that maximize variance in the data, but this variance is not necessarily driven by biologically meaningful factors like population structure or sex. The assumption that proximity on a PCA plot proves relatedness is not statistically guaranteed [8].
• Irreproducibility: The aforementioned study found that PCA results were not reproducible in the way the scientific community assumes, raising concerns about the validity of PCA-based findings in an estimated 18,400 to 35,200 physical anthropology studies [8].

(caption: The Standard GM Workflow and Its Critiqued Pathway. The conventional path (red) from PCA to subjective interpretation is increasingly challenged. A more robust alternative (green) uses Procrustes coordinates as direct input to supervised machine learning models for objective classification.)

Detailed Experimental Protocols from Key Studies

This study offers a robust protocol for clinical prediction, integrating GPA with machine learning.

• Data Acquisition: 3D facial scans were collected from 669 patients using a FaceGo pro scanner (accuracy 0.1 mm). Patients maintained a neutral expression with heads in a natural position.
• Landmarking and Registration: A reference mesh (9,578 vertices) was created. All individual facial scans were non-rigidly registered to this template using MeshMonk toolbox, ensuring dense, corresponding landmarks across all subjects.
• GPA and Data Reduction: GPA was applied to the 9,578 landmarks to remove size, location, and orientation. The resulting Procrustes coordinates were used for subsequent analysis.
• Model Comparison: The Procrustes-aligned coordinate data were used to train ten different machine learning models, including Logistic Regression (LR), Support Vector Machines, and Random Forests. Model performance for predicting DMV was compared against a traditional clinical score (DIFFMASK).
• Key Result: The Logistic Regression model performed best, achieving an AUC of 0.825, which was higher than the DIFFMASK score (AUC 0.785). This demonstrates that GM with ML can surpass traditional clinical assessment.

This study designed a methodological test to evaluate PCA's reliability using benchmark data.

• Benchmark Data: The researchers used a known dataset of papionin (Old World monkeys) crania from five genera.
• Tool Development: They developed MORPHIX, a Python package for processing superimposed landmark data, which includes classifier and outlier detection methods.
• Comparative Analysis: The standard GPA-PCA approach was applied to the benchmark data. Its performance in classification and novelty detection was then compared against various supervised machine learning classifiers.
• Key Result: The study concluded that PCA results were not robust or reproducible and that supervised machine learning classifiers provided more accurate classification and better detection of new taxa. This challenges the central role of PCA in thousands of published studies.

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of a geometric morphometrics study, especially one focused on cross-validation, requires a suite of specialized software and methodological tools.

Table 3: Key Research Reagents and Solutions for Geometric Morphometrics

Tool Name	Type/Function	Brief Description of Role in Protocol
TPSdig [13]	Landmark Digitization Software	Used to collect two-dimensional landmark coordinates from digital images.
MeshMonk [10]	3D Surface Registration Toolbox	An open-source toolbox for non-rigid, dense registration of 3D facial surfaces to a common template, generating thousands of corresponding landmarks.
Viewbox [11]	Landmark Digitization & Analysis	Software used to digitize both fixed landmarks and sliding semi-landmarks on 3D models.
MORPHIX [8]	Python Package for GM	A custom package for processing landmark data, featuring classifier and outlier detection methods as an alternative to standard PCA.
geomorph & FactoMineR [11]	R Packages for Statistical Analysis	Standard R packages for performing GPA, PCA, and other multivariate statistical analyses on landmark data.
Generalized Procrustes Analysis (GPA)	Core Statistical Method	The fundamental algorithm for aligning landmark configurations by removing differences in position, rotation, and scale.
Thin Plate Spline (TPS) [11]	Geometric Interpolation Function	Used to project semi-landmarks from a template onto individual specimens, ensuring homology across samples.

The evidence from current research presents a nuanced view of the standard GPA-PCA protocol in geometric morphometrics. Generalized Procrustes Analysis remains a robust and reliable foundation for aligning shapes and isolating shape variation from other confounding variables. Its utility is not in question.

The primary subject of debate is the subsequent use of Principal Component Analysis. While PCA is an excellent tool for unsupervised exploration and visualization of the major trends in shape variation, its reliability for definitive taxonomic classification and phylogenetic inference is seriously challenged. Studies consistently show that supervised machine learning models often outperform PCA-based analyses in predictive accuracy and classification tasks [8] [10].

Therefore, the choice of protocol should be guided by the research objective. For exploratory shape analysis and hypothesis generation, the standard GPA-PCA pipeline is sufficient. However, for classification, prediction, or whenever robust, cross-validated conclusions are required, the evidence strongly supports a shift towards a GPA-ML pipeline, where Procrustes-aligned coordinates are fed directly into supervised machine learning algorithms. This combined approach leverages the strengths of both worlds, ensuring rigorous statistical validation while maintaining a firm grounding in biological shape.

The accurate quantification of biological shape through geometric morphometrics is foundational to numerous fields, including ecology, paleontology, and biomedical research. These analyses rely on the precise placement of landmarks—discrete, homologous anatomical points—to capture form in two or three dimensions [14]. The reliability of downstream statistical interpretations, from taxonomic classifications to evolutionary inferences, is fundamentally constrained by the initial landmark data. Consequently, understanding how different landmark types, configurations, and data acquisition protocols influence measurement error is crucial for scientific reproducibility [14] [15].

Reproducibility, defined as the closeness of agreement between independent results obtained under different conditions (e.g., different operators or equipment), is a cornerstone of the scientific method [16]. In geometric morphometrics, this is threatened by various sources of error introduced during data collection, which can be substantial enough to explain over 30% of the total variation among datasets [14] [17]. This article provides a comparative guide to the reproducibility of different geometric morphometric protocols, synthesizing experimental data on error sources and their impacts on analytical outcomes. By framing this within the context of cross-validation performance, we aim to equip researchers with the evidence needed to design more robust and replicable morphometric studies.

Measurement error in geometric morphometrics is not a single entity but arises from multiple, distinct phases of data acquisition. A comprehensive understanding of these sources is the first step in mitigating their impact.

• Imaging Device (Instrumental Error): The choice and configuration of imaging equipment, such as different camera lenses or scanners, can generate dissimilar morphological reconstructions. This includes lens distortion effects, which vary based on lens curvature and the position of the specimen within the camera field [14].
• Specimen Presentation (Methodological Error): In 2D analyses, projecting a 3D object onto a 2D plane inevitably introduces distortion. The projected location of a landmark can be displaced from its true position relative to other loci. This error is exacerbated when specimens are oriented dissimilarly, making biological variation difficult to distinguish from artificial variation caused by presentation angle [14] [17].
• Interobserver Error (Personal Error): Different individuals may position landmarks differently on the same specimen and locus. This variation is influenced by the observer's experience and the clarity of the landmark's definition [14] [16].
• Intraobserver Error (Personal Error): The same observer may place a landmark inconsistently across different specimens or digitizing sessions. This is affected by factors such as fatigue and the ease of visualizing the landmark locus [14].

Quantitative Impact on Statistical Results

The impact of these error sources is not merely theoretical; they directly affect the statistical fidelity of morphometric analyses. A landmark study on vole (Microtus) molars quantified how error influences Linear Discriminant Analysis (LDA), a common classification tool [14] [17].

Table 1: Impact of Measurement Error on Species Classification Accuracy

Error Source	Key Finding on Classification	Experimental Context
Specimen Presentation	Greatest discrepancies in species classification results	Comparison of in-situ teeth vs. isolated/tilted teeth [14] [17]
Imaging Device	Impacts group membership predictions	Comparison of Nikon D70s vs. Dino-Lite digital microscope [17]
Interobserver Variation	Greatest discrepancies in landmark precision	Comparison between experienced and new observers [14] [17]
All Error Sources	No two landmark dataset replicates produced the same predicted group memberships for fossil specimens	Analysis of 31 fossil Microtus specimens [14] [17]

These findings underscore a critical point: the cumulative effect of measurement error can lead to fundamentally different interpretations of the same biological data. For instance, the taxonomic affinity of fossil specimens may be assigned to different groups depending solely on which replicated dataset is used to train the classifier [17]. This has profound implications for replicating studies in paleontology, ecology, and systematics.

Comparative Analysis of Morphometric Protocols

The reproducibility of a morphometric analysis is significantly influenced by the overarching methodological approach, which ranges from fully manual landmarking to automated, landmark-free techniques.

Traditional vs. Geometric Morphometrics

A direct comparison of four morphometric methods in ichthyology quantified their repeatability (agreement under the same conditions) and reproducibility (agreement under different conditions) [16].

Table 2: Performance Comparison of Morphometric Methods

Method	Key Characteristics	Repeatability & Reproducibility	Subjectivity (Measurer Effect)
Traditional (TRA)	Caliper-based linear measurements on preserved specimens	Lowest repeatability and reproducibility	Population-level detachment was entirely overwritten by measurer effect [16]
Truss-Network (TRU)	Distance between homologous points from digital images	Similar repeatability to Geometric Methods on Scales (GMS)	Significant measurer effect [16]
Geometric on Body (GMB)	Landmark coordinates from digital images of the body	Highest overall repeatability and reproducibility	Least burdened by measurer effect [16]
Geometric on Scales (GMS)	Landmark coordinates from digital images of scales	Similar repeatability to GMB, but lower reproducibility	Significant measurer effect; aggregation of different measurers' datasets not recommended [16]

The study strongly recommended image-based geometric methods (GMB) over traditional caliper-based methods due to their superior repeatability, reproducibility, and reduced subjectivity. It also cautioned against aggregating datasets from different measurers, especially when using TRA and GMS methods [16].

The Emergence of Landmark-Free and AI-Based Approaches

Emerging automated methods aim to overcome the bottlenecks of manual landmarking, which is time-consuming and prone to observer bias [18].

• Landmark-Free Morphometrics: Techniques like Large Deformation Diffeomorphic Metric Mapping (LDDMM) quantify shape by calculating the deformation energy required to map a reference "atlas" shape onto each specimen in a dataset. A macroevolutionary study of 322 mammals found that while landmark-free methods (DAA) showed strong correlation with manual landmarking, differences emerged, particularly for certain clades like Primates and Cetacea. Both methods produced comparable but varying estimates of phylogenetic signal and evolutionary rates, highlighting that the choice of method can influence macroevolutionary inferences [18].
• AI-Based Landmark Detection: Artificial intelligence, particularly deep learning models, offers high-throughput, automated landmarking. In cephalometrics, AI has demonstrated superior reproducibility compared to manual tracing by an experienced orthodontist [19]. One study reported that AI achieved the lowest coefficient of variation (CV) values, with particular consistency for landmarks like Menton (Me) and Pogonion (Pog) [19]. This enhanced reproducibility is attributed to the elimination of intra- and inter-obscriber variability. However, systematic biases can occur; for example, AI reconstructing 3D models from 2D photos showed a consistent downward shift of inferior landmarks [20].

Experimental Protocols for Error Assessment

Robust morphometric studies require protocols to quantify and control for measurement error. Below are detailed methodologies from key studies.

Protocol for Evaluating Data Acquisition Error

Objective: To quantify error from four sources (imaging device, specimen presentation, inter- and intraobserver variation) and its impact on classification statistics [17].

• Dataset Replication: Create multiple replicated datasets.
  • Imaging Device: Photograph the same set of specimens (e.g., vole dentaries) using two different cameras (e.g., Nikon D70s and Dino-Lite microscope).
  • Specimen Presentation: Photograph specimens a second time after intentionally tilting them along their axis to simulate non-standardized orientation.
  • Inter/Intraobserver Error: Have multiple observers (e.g., an experienced and a new observer) digitize the same image sets twice, with at least a week between sessions.
• Landmark Digitization: Digitize all images using a standardized landmark protocol (e.g., 21 landmarks on the lower first molar) in software such as TpsDig.
• Data Preprocessing: Perform Generalized Procrustes Analysis (GPA) using software like the gpagen function in the R package geomorph to superimpose all landmark configurations, removing variation due to position, orientation, and scale.
• Statistical Classification: Run Linear Discriminant Analysis (LDA) on each of the nine Procrustes-aligned datasets using the lda function in R. Use leave-one-out cross-validation to determine the correct classification rate for specimens of known species.
• Error Impact Assessment: Append fossil specimens of unknown affinity to each dataset and use the LDA models to predict their group membership. Compare the predictions across all replicated datasets to see how error influences the classification of unknowns.

Protocol for Assessing Individual Landmark Precision

Objective: To evaluate the precision of individual landmarks and avoid the "Pinocchio effect," where highly variable landmarks inflate overall error estimates [15].

• Repeated Digitization: A single observer digitizes the landmark configuration on a specimen multiple times (e.g., 10 times). The specimen must be held in a constant orientation throughout this process to isolate digitization error from presentation error.
• Calculate Centroid and Distances: For each landmark, calculate the centroid (mean x, y, z coordinates) of its repeated placements.
• Measure Precision: Compute the Euclidean distance between each individual placement of a landmark and its centroid.
• Analyze by Landmark: The average distance and standard deviation for each landmark represent its precision. Landmarks with consistently large distances from their centroid are considered less precise and may require redefinition or removal.

Experimental Workflows for Assessing Landmark Error

The Scientist's Toolkit: Essential Reagents and Software

A summary of key computational tools and their functions in geometric morphometrics is provided below.

Table 3: Key Software and Tools for Geometric Morphometrics

Tool Name	Primary Function	Application Context
TpsDig / TpsUtil	Digitizing landmarks and managing project files	Standard software for collecting 2D landmark data from images [17]
Geomorph (R package)	Performing Generalized Procrustes Analysis (GPA) and subsequent statistical shape analysis	Core statistical toolkit for morphometrics in R [17]
Deformetrica	Performing landmark-free analysis using Large Deformation Diffeomorphic Metric Mapping (LDDMM)	Automated shape comparison without manual landmarking [18]
WebCeph	AI-assisted and manual cephalometric landmark identification	Commercial platform for orthodontic analysis; used in AI reproducibility studies [19]
RENOIR	Platform for robust and reproducible machine learning model training and testing	Ensures generalizability of AI/ML models in biomedical sciences [21]

The reproducibility of geometric morphometric analyses is profoundly affected by the choices researchers make regarding landmark types, data acquisition protocols, and analytical methods. The evidence demonstrates that image-based geometric methods on the body (GMB) offer superior repeatability and reduced subjectivity compared to traditional caliper-based or scale-based geometric methods. Furthermore, emerging AI and landmark-free methods show great promise for enhancing throughput and consistency, though they require careful validation to identify and correct for potential systematic biases.

To maximize reproducibility, researchers should: standardize imaging equipment and specimen presentations whenever possible; use a single, experienced observer for landmarking or use automated methods; quantify and report measurement error as a routine part of their methodology; and be cautious when aggregating datasets collected under different conditions. By adopting these rigorous protocols, the morphometrics community can strengthen the foundation of shape-based inferences across biological and biomedical disciplines.

Geometric morphometrics (GM) is a foundational tool across evolutionary biology, palaeontology, and drug development for quantifying and analyzing shape variation. The standard analytical pipeline typically involves two core steps: Generalized Procrustes Analysis (GPA) to superimpose landmark coordinates by removing shape-independent variations, followed by Principal Component Analysis (PCA) to project the high-dimensional data onto a lower-dimensional space of uncorrelated variables [8]. This PCA-based approach is deeply embedded in morphological studies, with an estimated 18,400 to 35,200 physical anthropology studies alone relying on its outcomes [8].

However, a growing body of critical research challenges the reliability and robustness of PCA for drawing biological conclusions. This article provides a comparative guide evaluating PCA's performance against emerging alternative statistical and machine learning protocols, with a specific focus on cross-validation performance within geometric morphometric research. We synthesize current evidence to help researchers make informed methodological choices.

Theoretical and Methodological Biases of PCA

The application of PCA in morphometrics introduces several inherent biases that can compromise the validity of research findings.

• Input Data Artefacts: PCA outcomes are highly sensitive to input data composition. Results are not stable, reliable, or reproducible in the way often assumed by field practitioners [8]. The patterns observed in PCA scatterplots (e.g., clustering, proximity) may represent statistical artefacts rather than genuine biological relationships.

• Subjective Interpretation: Phenetic, evolutionary, and ontogenetic conclusions are frequently drawn from visual inspection of the first two or three principal components, despite these components being "statistical manifestations agnostic to the data" [8]. Researchers may selectively report PC combinations that support their hypotheses, as witnessed in controversial hominin taxonomy cases like Homo Nesher Ramla, where different PC plots produced conflicting phylogenetic results [8].

• Dimensionality Reduction Limitations: While PCA effectively reduces dimensionality, it may oversimplify complex morphological spaces by focusing on global structure at the expense of locally relevant variations for classification tasks [22].

Table 1: Documented Methodological Biases of PCA in Morphometric Studies

Bias Category	Description	Impact on Research
Input Sensitivity	Outcomes are artefacts of specific input data composition [8]	Compromised reliability and reproducibility of studies
Subjective Interpretation	Biological meaning is assigned to statistically-derived components [8]	Potential for confirmation bias in evolutionary hypotheses
Variance Overemphasis	Prioritizes directions of maximum variance, which may not be biologically relevant [23]	Possible misinterpretation of morphological patterns
Linearity Assumption	Assumes linear relationships in shape data [23]	Poor capture of complex morphological relationships

Experimental Comparisons: PCA vs. Alternative Methods

Performance in Taxonomic Identification

A rigorous evaluation of statistical models for establishing morphometric taxonomic identifications compared PCA with Linear Discriminant Analysis (LDA) and Random Forest (RF) using cranial specimens of modern Dipodomys spp. and Leporidae species [22]. The results demonstrated that Random Forest consistently outperformed PCA across all test scenarios.

Table 2: Classification Error Rates (%) by Statistical Method and Dataset [22]

Condition	Dataset	PCA	LDA	Random Forest
Complete Crania	Leporidae	18.4	4.1	3.1
Complete Crania	Dipodomys spp.	42.9	16.3	16.3
Cranial Fragments	Leporidae	26.5	8.2	6.1
Cranial Fragments	Dipodomys spp.	46.9	18.4	16.3

The study concluded that "PCA should not be used to predict species identifications using morphometric data" due to its significantly higher error rates [22]. Random Forest not only achieved higher accuracy but also handled missing data more effectively through imputation.

Performance in Novel Taxon Detection

Beyond classification of known taxa, the detection of novel or outlier specimens represents a critical challenge in morphological research. A study developing MORPHIX, a Python package for processing landmark data, found that supervised machine learning classifiers were more accurate than PCA both for standard classification tasks and for detecting new taxa [8]. This capability is particularly valuable for identifying exceptional specimens that may represent new species or previously unknown morphological variants.

Limitations in Capturing Tooth Mark Morphology

In taphonomy research, a methodological comparison of techniques for identifying carnivore agency found that PCA-based geometric morphometric approaches showed less than 40% discriminant power when analyzing bi-dimensional tooth marks [3]. The study noted that previous claims of high accuracy using these methods were "heuristically incomplete" because they had only considered a small range of allometrically-conditioned tooth pits while excluding widely represented non-oval forms [3].

In contrast, computer vision approaches using Deep Convolutional Neural Networks classified experimental tooth pits with approximately 80% accuracy, demonstrating significantly superior performance for this specific morphological application [3].

Experimental Protocols for Method Comparison

Standard Geometric Morphometrics Workflow

The foundational protocol for landmark-based geometric morphometrics involves sequential steps that are consistent across most studies, whether using PCA or alternative multivariate methods.

Diagram 1: Standard workflow for geometric morphometric studies. The multivariate analysis stage is where PCA and alternative methods diverge.

Detailed Protocol: Killer Whale Reproduction Study

A comprehensive geometric morphometric study on killer whale reproductive stages provides an exemplary protocol for method comparison [12]:

• Landmark Configuration: Six landmarks were digitized on standardized aerial images of killer whales using software such as MorphoMetriX [12].
• Experimental Design: The study employed a balanced approach with multiple images per individual and tested the significance of Procrustes distances to validate landmark configurations.
• Validation Method: Discriminant Function Analysis (DFA) revealed significant shape differences between reproductive states (non-pregnant, early-stage pregnant, late-stage pregnant, lactating) with P-values ranging from <0.001 to 0.01 for pairwise comparisons [12].
• Cross-Validation: The protocol included testing for statistical differences in Procrustes distances based on the number of images per whale and number of landmarks per image to optimize the configuration.

This experimental design demonstrates how rigorous validation protocols can be implemented to ensure the reliability of morphometric analyses beyond standard PCA approaches.

Protocol for Out-of-Sample Classification

A critical challenge in applied morphometrics involves classifying new specimens not included in the original training set. A study on children's nutritional assessment from arm shapes developed a specialized protocol for this purpose [24]:

• Template Registration: Raw coordinates of new individuals are registered to a template configuration from the training sample.
• Procrustes Alignment: The registered coordinates undergo Procrustes analysis against the template.
• Classifier Application: Pre-trained classifiers (LDA, RF, or SVM) are applied to the aligned coordinates for nutritional status classification.
• Performance Validation: The method was tested with different template choices to evaluate impact on classification accuracy.

This approach addresses a significant limitation of standard PCA-based morphometrics, where classification rules derived from a sample cannot be directly applied to new individuals without repeating the entire alignment process [24].

The Scientist's Toolkit: Essential Research Reagents and Solutions

Table 3: Key Software and Analytical Tools for Geometric Morphometrics

Tool Name	Type	Primary Function	Application Context
MORPHIX [8]	Python Package	Processing landmark data with classifier and outlier detection	Evolutionary anthropology, novel taxon detection
TPSDig2 [25] [13]	Desktop Software	Landmark digitization on 2D images	Standardized landmark placement across studies
FaceDig [25]	AI Tool	Automated landmark placement on facial portraits	High-throughput facial morphometrics
MorphoJ [26]	Desktop Software	Comprehensive morphometric analysis	General-purpose shape analysis
XYOM [27]	Cloud Platform	Online morphometric analysis	Platform-independent collaborative research
R (geomorph) [24]	Statistical Package	GM analysis in statistical programming environment	Flexible, customizable analytical pipelines

The cross-validation performance of different geometric morphometric protocols reveals significant limitations in PCA-based approaches compared to modern alternatives. The evidence indicates that PCA exhibits substantial biases that can lead to unreliable biological interpretations, particularly when used for taxonomic identification or phylogenetic inference.

Supervised machine learning methods, particularly Random Forest classifiers, demonstrate superior performance in multiple experimental contexts, offering higher classification accuracy and better handling of missing data [22]. These methods excel at capturing complex morphological patterns that may be overlooked by PCA's variance-maximizing approach.

For researchers engaged in geometric morphometrics, we recommend:

• Utilizing multiple analytical methods to validate findings, rather than relying exclusively on PCA
• Implementing rigorous cross-validation protocols using hold-out samples or leave-one-out approaches
• Considering supervised machine learning approaches for classification tasks, particularly when working with fragmentary or incomplete specimens
• Applying caution when interpreting PCA plots as direct evidence of biological relationships

Future methodological development should focus on integrating geometric morphometrics with robust machine learning frameworks and improving protocols for out-of-sample classification. As the field advances, the critical examination of analytical biases remains essential for generating reliable morphological insights across evolutionary biology, anthropology, and drug development research.

The Role of Sample Size and Power in Protocol Validation

In scientific research, particularly in fields employing advanced morphological analysis or machine learning, the principles of protocol validation are paramount. This process ensures that methodologies produce reliable, reproducible, and generalizable results. Two of the most critical factors influencing the success of validation are sample size and statistical power. Within the context of a broader thesis on the cross-validation performance of different geometric morphometric protocols, this article examines how sample size and power underpin protocol validation. We objectively compare the performance of different methodological approaches, using supporting experimental data to highlight the trade-offs and optimal strategies researchers must consider. The focus on geometric morphometrics serves as a powerful case study due to its high-dimensional data and reliance on robust validation, but the conclusions are applicable to a wide range of scientific domains, including drug development.

Theoretical Foundations: Sample Size, Power, and Cross-Validation

The Interplay of Key Concepts

At its core, protocol validation is the process of establishing that a specific methodological procedure is fit for its intended purpose. In data-driven sciences, this almost invariably involves using cross-validation (CV), a family of model validation techniques that assess how the results of a statistical analysis will generalize to an independent data set [28]. The goal is to flag problems like overfitting and provide insight into how a model will perform on unseen data.

The effectiveness of cross-validation is directly governed by two intertwined concepts:

• Sample Size (n): The number of observations or specimens available for analysis.
• Statistical Power: The probability that a test will correctly reject a false null hypothesis (i.e., detect an effect if one exists).

The relationship between them is simple yet profound: inadequate sample size leads to low statistical power. A study with low power not only risks missing true effects (Type II errors) but also produces effect sizes that are often inflated and unreliable [29]. This is especially critical in geometric morphometric studies, which analyze complex shape data and require sufficient samples to accurately estimate population mean shape and variance [30].

The Pitfalls of Small Sample Sizes in Validation

Small sample sizes create a cascade of problems that compromise protocol validation:

• Increased Error in Cross-Validation: With small samples, the estimate of a model's predictive accuracy becomes unstable and possesses large error bars. One study noted that cross-validation errors in measuring prediction accuracy can be as high as ±10% with sample sizes around 100 [31].
• Overfitting and Optimistic Bias: A model trained on a small dataset may learn the noise in the data rather than the underlying signal. While it may perform well on that specific training set, its performance will plummet on unseen data. Cross-validation with small samples struggles to detect this overfitting [32].
• Impact on Morphometric Analyses: In geometric morphometrics, reducing sample size directly impacts the accuracy of shape measurements. Studies have shown that smaller samples lead to biased estimates of mean shape and increased shape variance, making it difficult to detect true morphological differences between groups [30].

The following diagram illustrates the logical relationship between sample size, statistical power, and the outcomes of protocol validation.

Comparative Analysis of Methodological Protocols

The choice of analytical protocol and its interaction with sample size significantly influences the outcome of scientific studies. The table below summarizes key performance metrics for different methodological approaches as reported in the literature.

Table 1: Performance comparison of different methodological protocols under varying sample sizes

Method / Protocol	Reported Accuracy / Performance	Key Sample Size Finding	Study Context
2D Geometric Morphometrics (GMM)	Effective for species discrimination [30]	Sample size reduction significantly impacts mean shape & increases shape variance; n > 70 used for stable estimates [30]	Skull shape analysis in bat species
Computer Vision (Deep Learning)	81% classification accuracy [3]	Outperformed 2D GMM which showed < 40% discriminant power in the same study [3]	Carnivore tooth mark identification
Machine Learning (SVM, NN, etc.)	Accuracy increases with sample size, plateauing after n ≈ 120 [29]	Small samples (n<120) show high variance in accuracy; overfitting exaggerates reported performance [29]	Arrhythmia dataset classification
Nested k-fold Cross-Validation	Highest statistical confidence and power [32]	Required sample size could be 50% lower than with single holdout method; reduces overestimation of accuracy [32]	General ML model validation

Geometric Morphometrics vs. Computer Vision

A direct comparison in taphonomy highlights how protocol choice affects outcomes. A study aiming to identify carnivore agents from tooth marks found that while 2D Geometric Morphometrics (GMM) using outline analysis had limited discriminant power (<40%), a Computer Vision (CV) approach using Deep Convolutional Neural Networks (DCNN) achieved 81% accuracy [3]. This stark difference was attributed to the GMM's reliance on manual landmarking and outlines, which may not capture the full spectrum of shape complexity, especially with a constrained sample of "non-oval tooth pits." The CV protocol, designed to automatically learn relevant features from images, demonstrated superior performance with the same data, underscoring the importance of selecting a protocol with sufficient representational capacity for the task.

The Critical Role of Advanced Cross-Validation Protocols

The method of validation itself is a protocol that requires careful selection. A key finding from machine learning research is that the common practice of single holdout cross-validation (a single train-test split) leads to models with low statistical power and confidence, resulting in a significant overestimation of classification accuracy [32].

In contrast, nested k-fold cross-validation provides a more robust validation protocol. In this method, an outer loop performs k-fold cross-validation to estimate the generalization error, while an inner loop is used for model selection and hyperparameter tuning. This prevents data leakage and provides an unbiased estimate of model performance. The adoption of this more rigorous protocol is critical, as it can reduce the required sample size by up to 50% compared to the single holdout method to achieve the same level of confidence [32].

Experimental Data on Sample Size Impact

Quantitative Evidence from Morphometrics and Machine Learning

Empirical studies consistently demonstrate a non-linear relationship between sample size and the reliability of outcomes. The following table synthesizes quantitative findings from multiple research domains.

Table 2: Impact of sample size on analytical outcomes across different fields

Field of Study	Measured Metric	Small Sample Effect (n < ~30)	Effect with Larger Samples (n > ~100)
Geometric Morphometrics [30]	Mean shape estimation	Biased and unstable	Converges to stable population value
Geometric Morphometrics [30]	Shape variance	Inflated variance	Accurately reflects population variance
Machine Learning [29]	Classification Accuracy	High variance (e.g., 68-98%); overfitting	Stable and reliable (e.g., 85-99%)
Machine Learning [29]	Effect Size	Inflated and highly variable	Stable and accurate
Neuroimaging (MVPA) [31]	Cross-Validation Error Bar	Large (e.g., ±10%)	Substantially reduced

A pivotal study on bat skull morphology systematically evaluated the impact of sample size on 2D geometric morphometric analyses. Using large intraspecific sample sizes (n > 70) for Lasiurus borealis and Nycticeius humeralis, researchers found that reducing sample size directly increased the distance from the true mean shape and inflated estimates of shape variance [30]. This means that studies with small samples are not only less precise but also prone to overestimating the morphological diversity within a group. Furthermore, they found that shape differences were not consistent across different 2D views of the skull, indicating that a single view analyzed with a small sample may lead to incomplete or misleading biological conclusions.

In machine learning, a systematic evaluation using a large arrhythmia dataset revealed that classification accuracy for multiple algorithms (Support Vector Machine, Neural Networks, etc.) increased sharply as the sample size grew from 16 to about 120. Crucially, the variance in accuracy was very high for sample sizes below 120, meaning that a single run of an experiment could yield a deceptively high or low result purely by chance. Beyond this point, the performance gains diminished and the results stabilized [29]. This provides a practical benchmark for a minimum sample size in similar ML-based studies.

Workflow for a Robust Validation Protocol

The following workflow integrates the critical steps of sample size consideration and robust cross-validation into a geometric morphometric study design, from data collection to final interpretation.

Essential Research Reagent Solutions

Successful protocol validation relies on a toolkit of robust software and methodological "reagents." The following table details key solutions essential for researchers in geometric morphometrics and related fields.

Table 3: Key research reagent solutions for geometric morphometric and validation studies

Tool / Solution	Function / Purpose	Relevance to Protocol Validation
tpsDig2 [30]	Software for digitizing landmarks and semi-landmarks on 2D images.	Standardizes the initial data collection step, reducing observer bias and ensuring reproducibility in morphometric analyses.
Geomorph R Package [30]	A comprehensive R package for performing geometric morphometric analyses, including Generalized Procrustes Analysis (GPA) and statistical testing.	Provides a standardized, peer-reviewed toolkit for core GM procedures, ensuring analytical consistency and correctness.
Nested k-fold Cross-Validation Code [32]	Custom scripts (e.g., in MATLAB or Python) to implement nested cross-validation.	Critical for obtaining unbiased performance estimates in machine learning and model-based studies, preventing overfitting.
Whalength / ImageJ / MorphoMetriX [12]	Software tools for processing and measuring biological specimens from images.	Enables non-invasive, standardized body condition assessments, crucial for ecological and conservation studies.
Deep Learning Frameworks (e.g., for DCNN) [3]	Libraries like TensorFlow or PyTorch for implementing computer vision models.	Provides an alternative, high-capacity protocol for shape analysis that can outperform traditional GMM in some classification tasks.

The body of evidence unequivocally demonstrates that sample size and statistical power are not mere afterthoughts but foundational elements of protocol validation. In geometric morphometrics, small samples lead to unstable estimates of shape and variance [30]. In machine learning and neuroimaging, they result in large, often underestimated, error bars in cross-validation, creating an "illusion of biomarkers that do not generalize" [31].

The comparative data shows that while advanced methods like deep learning can offer higher accuracy [3], they do not absolve the researcher from the sample size imperative. Furthermore, the choice of validation protocol itself, such as adopting nested k-fold cross-validation over a simple holdout method, is a powerful lever to improve statistical power and confidence, effectively making better use of available samples [32].

For researchers, scientists, and drug development professionals, this implies that study design must prioritize sample size estimation and power analysis from the outset. Relying on small, underpowered studies risks building scientific conclusions on an unstable foundation. The practical guidance is clear: invest in preliminary data and power analyses, aim for sample sizes demonstrated to provide stable estimates (e.g., often n > 70 in morphometrics [30]), and always employ the most robust cross-validation protocols available to ensure that validated methods perform reliably when applied in the real world.

Protocol Implementation and Domain-Specific Applications

Geometric morphometrics (GM) has become a foundational tool for the quantitative analysis of shape across biological, medical, and paleontological disciplines. Within this field, two predominant methodologies have emerged: landmark-based GM, which relies on anatomically defined point coordinates, and outline-based GM, which analyzes the complete contour of a structure using mathematical functions [33] [34]. The choice between these methods carries significant implications for the reliability, interpretability, and generalizability of research findings, particularly when classification models are applied to new data.

This guide objectively compares the performance of these two approaches through the critical lens of cross-validation. Cross-validation rigorously tests a model's predictive power by evaluating its performance on data not used during training, thus simulating real-world application [2]. For researchers in drug development and other applied sciences, where models must perform reliably on new samples, understanding the cross-validation performance of different geometric morphometric protocols is paramount.

Methodological Foundations

Landmark-Based Geometric Morphometrics

Landmark-based GM analyzes shape using discrete, homologous points that have direct biological correspondence across specimens. The methodology follows a structured pipeline:

• Landmark Typology: Landmarks are classified into three primary types. Type I landmarks are defined by discrete anatomical loci (e.g., junctions between bones or the tip of a structure). Type II landmarks represent mathematically defined points of maximum curvature. Type III landmarks are constructed points, such as midpoints or extreme points, defined by their relative position to other landmarks [34].
• Data Acquisition and Procrustes Superimposition: Cartesian coordinates from digitized landmarks undergo Generalized Procrustes Analysis (GPA). GPA standardizes configurations by scaling them to a unit centroid size, translating them to a common origin, and rotating them to minimize the sum of squared distances between corresponding landmarks [34] [35]. This process isolates shape variation from differences in size, position, and orientation.
• Statistical Analysis and Classification: The resulting Procrustes coordinates serve as variables for multivariate statistical analysis, including Principal Component Analysis (PCA) and Discriminant Function Analysis (DFA) or Canonical Variate Analysis (CVA) [34] [35]. Classification rules, such as linear discriminant functions, are built from these aligned coordinates.

A key challenge in landmark-based analysis, particularly for cross-validation, is the out-of-sample problem. Classification rules are constructed from a sample-dependent Procrustes alignment. Applying these rules to new individuals requires a method to register the new specimen's raw coordinates into the pre-existing shape space of the training sample, a process that is not standardized and can introduce error [2].

Outline-Based Geometric Morphometrics

Outline-based GM captures shape information from the entire contour of a structure, making it suitable for forms that lack discrete landmarks. The standard workflow involves:

• Outline Digitization: The contour of a structure is captured as a sequence of two-dimensional coordinates [34].
• Mathematical Representation: The outline is represented mathematically, most commonly using Elliptic Fourier Analysis (EFA). EFA decomposes a contour into a sum of trigonometric functions (harmonics), with the coefficients of these functions serving as shape descriptors [33] [36].
• Data Normalization and Analysis: Fourier coefficients are normalized to be invariant to the contour's starting point, rotation, and size. These normalized coefficients are then used as variables in subsequent statistical analyses and classifier construction [34].

Alternative outline methods are also emerging. The shape-changing chain approach, for instance, models a profile using a chain of rigid, scalable, and extendible segments. The parameters of this chain (e.g., relative angles and length ratios) provide a modest number of variables for discriminant analysis, which can have physical or biological meaning [36].

Comparative Cross-Validation Performance

The table below synthesizes quantitative findings from multiple studies that directly compared the classification accuracy of landmark- and outline-based methods, often using cross-validation techniques.

Table 1: Cross-Validation Classification Accuracy of Landmark vs. Outline-Based GM

Study Organism/Subject	Landmark-Based Accuracy	Outline-Based Accuracy	Cross-Validation Method	Key Findings	Source
Trichodinids (parasites)	Higher accuracy (specific value not provided)	Lower accuracy	Not specified	Landmarks provided greater differentiation; outlines may include points with less taxonomic information.	[37]
Mosquito Vectors	Effective for genus-level ID and Anopheles & Aedes species	Effective for genus-level ID and Anopheles & Aedes species	Validated reclassification	Both methods were successful, but performance varied by genus; less effective for Culex species.	[33]
Horse Flies (Tabanus)	Not tested	86.67% (using 1st submarginal cell contour)	Validated classification test	Outline-based GM on a specific wing cell showed high accuracy and is useful for damaged specimens.	[38]
Children's Arm Shape	Model created from Procrustes coordinates	Not the focus	Out-of-sample application	Highlighted the central challenge of classifying new individuals not included in the original Procrustes alignment.	[2]

Interpretation of Comparative Data

The data reveals that the superiority of one method over the other is often context-dependent.

• Data Quality and Structure: Outline-based methods demonstrated a significant advantage in a study of horse flies, where the contour of the first submarginal wing cell achieved 86.67% classification accuracy. This approach proved particularly valuable for analyzing specimens with incomplete wings but intact cells, a common issue with field-collected insects [38]. Conversely, a study on trichodinid parasites found landmark-based methods to be more accurate, suggesting that for certain structures, defined points may contain more taxonomically informative data than the overall contour [37].
• The Out-of-Sample Challenge: A critical finding from nutritional status research is that landmark-based models built on Procrustes coordinates face a fundamental hurdle in cross-validation. The entire classification pipeline, from alignment to classifier training, is sample-dependent. Applying this model to a new child's arm shape requires a non-standardized registration step to place the new individual into the pre-existing shape space, which can affect reliability [2]. This is a form of data leakage that proper cross-validation must account for.

Experimental Protocols for Cross-Validation

A Protocol for Landmark-Based GM with Cross-Validation

The following protocol, synthesized from multiple sources [2] [34] [35], is designed to properly address out-of-sample classification.

• Image Acquisition and Landmarking: Capture high-resolution, standardized 2D or 3D images. Using software such as tpsDig2, digitize Type I, II, and III landmarks on all specimens in the training set [34].
• Training-Sample Procrustes Analysis: Perform Generalized Procrustes Analysis (GPA) exclusively on the training sample to obtain Procrustes shape coordinates. Compute Centroid Size as a measure of isometric size [35].
• Classifier Construction: Use the Procrustes coordinates from the training set to build a classifier (e.g., Linear Discriminant Analysis). Evaluate its performance using leave-one-out cross-validation within the training sample [2].
• Out-of-Sample Registration: To classify a new specimen, its raw landmark coordinates must be aligned to the training sample's shape space. This is typically done by performing a Procrustes fit that rotates and scales the new specimen to the mean shape of the training set, without including it in a new GPA.
• Out-of-Sample Prediction: Apply the pre-trained classifier to the registered coordinates of the new specimen to predict its group membership.

A Protocol for Outline-Based GM with Cross-Validation

This protocol, based on studies of fish, insects, and parasites [38] [33] [34], outlines the workflow for outline analysis.

• Image Preparation and Outline Extraction: Remove backgrounds from images using tools like ImageJ. Extract the outline coordinates of the structure of interest [34].
• Mathematical Representation: For closed outlines, apply Elliptic Fourier Analysis (EFA) using the Momocs package in R. Select a sufficient number of harmonics to capture the essential shape. Normalize the coefficients for size, rotation, and starting point [34].
• Classifier Construction and Validation: Use the normalized Fourier coefficients as variables. Construct a classifier (e.g., DFA) on the training set and validate its performance using leave-one-out cross-validation [38].
• Out-of-Sample Classification: For a new specimen, extract its outline, compute its normalized Fourier coefficients using the same harmonic model, and directly apply the pre-trained classifier. This process is typically more straightforward than landmark-based out-of-sample prediction as it avoids a re-alignment step [2].

The following workflow diagram illustrates the critical divergence in how the two methods handle out-of-sample data, highlighting the additional registration step required for landmark-based GM.

The Researcher's Toolkit: Essential Reagents and Software

Table 2: Essential Software and Analytical Tools for Geometric Morphometrics

Tool Name	Type/Function	Application in GM	Relevance to Cross-Validation
tpsDig2, tpsUtil	Software suite for digitization and file management	Digitizing landmarks and organizing data files	Foundational for creating reproducible landmark datasets.	[34] [35]
MorphoJ	Integrated software for GM analysis	Performing Procrustes superimposition, PCA, DFA, and CVA	Commonly used to build and perform leave-one-out cross-validation on training samples.	[34] [35]
R packages (Momocs, geomorph)	Statistical programming environment	Comprehensive outline analysis (Momocs) and general GM (geomorph)	Provides flexible, scripted environments for implementing custom cross-validation protocols.	[34]
ImageJ	Image processing and analysis	Background removal and outline extraction	Essential for preparing images for consistent and automated outline analysis.	[34]
Linear Discriminant Analysis (LDA)	Statistical classification method	Building classifiers from shape variables (Procrustes coordinates or Fourier coeffs.)	The primary method for creating classification rules that are tested via cross-validation.	[2] [34]

Both landmark- and outline-based geometric morphometrics offer powerful, yet distinct, pathways for shape classification. The choice between them should be guided by the specific research context and the paramount importance of cross-validation performance.

• Opt for Landmark-Based GM when your structures possess clear, homologous points and the primary research goal requires deep biological interpretation tied to specific anatomical locations. However, researchers must be acutely aware of and develop robust solutions for the out-of-sample registration problem to ensure their models are generalizable [2].
• Opt for Outline-Based GM when analyzing structures lacking discrete landmarks, when dealing with damaged specimens where key areas are missing but contours remain [38], or when a more streamlined out-of-sample prediction pipeline is desirable. Methods like the analysis of wing cell contours [38] or the shape-changing chain [36] show high discriminatory power and practical advantages.

Ultimately, the most rigorous approach may often involve a combination of both methods, leveraging their respective strengths to validate findings and build a more comprehensive and reliable model of shape variation for real-world application.

{Abstract} Geometric morphometrics (GM) is a fundamental tool for quantifying biological shape, but it can be limited by its reliance on discrete landmarks. This guide compares a novel protocol, Functional Data Geometric Morphometrics (FDGM), against classical GM and other alternatives. FDGM enhances sensitivity by converting discrete landmark data into continuous curves, capturing subtle shape variations often missed by traditional methods. Experimental data from species classification studies, particularly on shrew crania, demonstrates FDGM's superior performance in cross-validation and machine learning applications, establishing it as a powerful protocol for taxonomic and morphological research where high sensitivity is critical.

Geometric Morphometrics (GM) is a landmark-based approach that quantitatively analyzes the shape of biological organisms by comparing the coordinates of anatomically defined points after removing differences in size, position, and orientation through a process called Generalized Procrustes Analysis (GPA) [9] [39]. While powerful, a key limitation of classical GM is that important shape differences can occur between landmarks, which the discrete point data may fail to capture [9].

Functional Data Geometric Morphometrics (FDGM) is an advanced protocol that addresses this gap. FDGM treats the configuration of landmarks not as a set of discrete points, but as a continuous curve. It uses mathematical functions to represent the entire shape, thereby capturing the geometry between landmarks and providing a more comprehensive description of form [9]. This protocol is particularly valuable for enhancing the sensitivity of analyses aimed at distinguishing groups with very subtle morphological differences.

Objective Comparison of Protocol Performance

Experimental data from direct comparisons provides the most reliable evidence for evaluating protocol performance. A study on shrew classification offers a robust, head-to-head comparison between FDGM and classical GM.

Direct Performance Comparison: FDGM vs. Classical GM

A study on classifying three shrew species (S. murinus, C. monticola, and C. malayana) from Peninsular Malaysia applied both FDGM and classical GM to the same set of craniodental landmark data. The performance was evaluated using multiple machine learning classifiers. The table below summarizes the key experimental findings [9].

Table 1: Performance comparison of FDGM and Classical GM in shrew species classification using different machine learning models (Data sourced from Pillay et al., 2024).

Machine Learning Model	Classical GM Accuracy (%)	FDGM Accuracy (%)
Naïve Bayes	84.3	91.0
Support Vector Machine	83.1	93.3
Random Forest	85.4	92.1
Generalized Linear Model	84.3	89.9

Table 2: Classification accuracy by cranial view, showing FDGM's superior performance, particularly with the dorsal view (Data sourced from Pillay et al., 2024).

Craniodental View	Best-Performing Model	Classical GM Accuracy (%)	FDGM Accuracy (%))
Dorsal	Support Vector Machine	86.5	97.8
Jaw	Support Vector Machine	84.3	91.0
Lateral	Naïve Bayes	84.3	89.9

The experimental results consistently demonstrate that FDGM achieves higher classification accuracy across all tested machine learning models and craniodental views. The dorsal view provided the best distinction between species, and FDGM's performance with this view was notably high at 97.8% accuracy [9]. This supports the thesis that FDGM's enhanced sensitivity translates to superior cross-validation performance in taxonomic classification tasks.

Detailed Experimental Protocols

To ensure reproducibility and provide a clear framework for implementation, here are the detailed methodologies for the key FDGM experiment and a contrasting classical GM approach.

Protocol 1: Functional Data Geometric Morphometrics (FDGM)

This protocol is adapted from the shrew craniodental shape classification study [9].

• Landmark Data Collection: Obtain 2D landmark coordinate data from digital images of the specimens. In the reference study, 89 shrew crania were used, with landmarks placed on three craniodental views: dorsal, jaw, and lateral.
• Generalized Procrustes Analysis (GPA): Perform GPA on the raw landmark data to superimpose all configurations. This step removes variations due to translation, rotation, and scale, aligning all specimens in a common shape space.
• Conversion to Functional Data: Transform the discrete Procrustes-aligned landmark coordinates into continuous curves. This is achieved by representing the outline of the landmark configuration as a linear combination of basis functions (e.g., B-splines or Fourier series).
• Shape Variable Extraction: Analyze the continuous curves using Functional Data Analysis (FDA) techniques. Principal Component Analysis (PCA) is then applied to the functional data to reduce dimensionality and extract major axes of shape variation (Principal Components).
• Classification Modeling: Use the resulting PC scores as input variables for machine learning classifiers. The reference study tested Naïve Bayes, Support Vector Machine, Random Forest, and Generalized Linear Model, using cross-validation to assess performance.

Diagram 1: FDGM analysis workflow.

Protocol 2: Classical Geometric Morphometrics (GM)

This protocol outlines the standard, widely-used GM method for comparison [9] [39].

• Landmark Data Collection: Collect 2D landmark coordinates from specimen images, identical to the first step in the FDGM protocol.
• Generalized Procrustes Analysis (GPA): Perform GPA on the raw landmark data to align all specimens in shape space. This step is identical in both protocols up to this point.
• Shape Variable Extraction: Use the resulting Procrustes coordinates (the aligned landmarks) directly for statistical analysis. PCA is applied to this matrix of Procrustes coordinates to extract principal components representing the major patterns of shape variation.
• Classification Modeling: Use the PC scores from the Procrustes coordinates as inputs for the same suite of machine learning classifiers (e.g., Naïve Bayes, SVM, etc.). Performance is evaluated using the same cross-validation procedures as in the FDGM protocol.

The Scientist's Toolkit: Essential Research Reagents & Materials

Implementing a morphometrics study, whether FDGM or classical GM, requires a specific set of tools and reagents. The following table details key components for building a research pipeline.

Table 3: Essential materials and software for geometric morphometrics research.

Category	Item	Function / Description
Specimen & Imaging	Biological Specimens	The physical objects of study (e.g., shrew skulls [9], shark teeth [13], insect wings [40]).
	Digital Camera / Microscope	To capture high-resolution 2D images for landmark digitization [35] [40].
	3D Scanner (e.g., Artec Eva)	For creating high-resolution 3D models when 3D morphometrics is required [41].
Software & Digitization	TpsDig2, TpsUtil	Standard software for digitizing landmarks and semilandmarks from images [13] [35].
	Viewbox4, MorphoJ	Software for digitizing 3D landmarks and performing Procrustes alignment, PCA, and other GM statistics [35] [41].
	R Statistical Environment	A powerful, open-source platform for statistical computing. Key packages for GM include geomorph and Momocs [39].
Analysis & Modeling	Functional Data Analysis (FDA) R packages	Specialized R libraries (e.g., fda) for implementing the curve-fitting and analysis steps in FDGM [9].
	Machine Learning Libraries (R/Python)	Libraries such as caret (R) or scikit-learn (Python) for implementing classifiers like SVM and Random Forest [9].

Protocol Selection Guide

The choice between FDGM, classical GM, and other alternatives depends on the research question, data type, and desired sensitivity.

• FDGM is the superior choice when analyzing outlines and shapes where the most biologically relevant variation occurs between traditional landmarks. It is particularly suited for high-sensitivity tasks like distinguishing cryptic species or detecting subtle phenotypic changes [9]. Its main consideration is increased computational complexity.
• Classical GM remains a highly robust and widely understood method. It is ideal for studies focused on well-defined homologous landmarks and for comparisons with the vast existing body of GM literature. It is generally less sensitive to non-landmark shape changes [9].
• Other Notable Alternatives: 3D GM protocols use 3D scanners and landmarks to capture complex geometry, providing the most comprehensive shape description but at a higher cost and processing time [41]. Computer Vision/Deep Learning approaches, such as Deep Neural Networks, can achieve high classification accuracy and are less reliant on manual landmarking, but they often function as "black boxes" with limited interpretability of the specific shape features driving the classification [3].

Experimental evidence confirms that Functional Data Geometric Morphometrics (FDGM) represents a significant advancement in shape analysis protocols. By modeling landmark configurations as continuous functions, FDGM captures a richer set of morphological information than classical GM, leading to demonstrably higher sensitivity and superior performance in machine learning classification and cross-validation. While classical GM remains a vital tool, FDGM establishes a new standard for precision in scenarios requiring the detection of minimal morphological differences, solidifying its role as a novel protocol for enhanced sensitivity in modern morphometric research.

The efficacy of intranasal drug delivery, particularly for direct nose-to-brain transport, is highly dependent on the complex and variable three-dimensional anatomy of the nasal cavity. This case study examines the validation of Geometric Morphometric (GM) protocols for classifying nasal cavity morphology and its correlation with drug delivery efficiency. Within the broader context of cross-validating different GM methodologies, we analyze how shape-based clustering of the nasal Region of Interest (ROI) can predict olfactory accessibility and inform personalized drug delivery strategies. This approach addresses a critical challenge in nasal drug administration: the significant inter-individual anatomical variability that complicates the development of standardized delivery protocols [11].

Methodological Protocols in Nasal Cavity Geometric Morphometrics

Region of Interest Definition and Landmarking

The foundational step in the GM protocol involves the precise definition of the Region of Interest (ROI) and the application of landmarks. In a study analyzing 151 unilateral nasal cavities from 78 patients, the ROI was standardized to begin at the plane crossing the plica nasi and nasal valve—the narrowest region of the nasal cavity—and extend to the anterior part of the olfactory region. The vestibule was systematically excluded as it is primarily occupied by the delivery nozzle and does not influence particle trajectories within the deeper nasal structures [11].

The landmarking protocol comprised:

• Ten fixed anatomical landmarks placed on homologous regions present across all specimens. These included key points such as the highest point of the nasal valve, the most anterior maximum of the vestibule, and the highest points at the front and back of the olfactory region [11].
• Two hundred sliding semi-landmarks distributed across the ROI surface of a template model and subsequently projected onto each patient model using Thin Plate Spline (TPS) warping. This technique ensures optimal homology across specimens while minimizing distortion by allowing semi-landmarks to slide tangentially along the surface [11].

Shape Alignment and Morphological Clustering

The coordinate data from landmarks and semi-landmarks underwent standardization via Generalized Procrustes Analysis (GPA) to remove variations due to translation, rotation, and scale. The aligned coordinates were then subjected to Principal Component Analysis (PCA) to identify dominant axes of shape variation [11].

Hierarchical Clustering on Principal Components (HCPC) was performed to classify morphological variations. The number of clusters was determined automatically by analyzing gains in cluster inertia to identify the partition that best reflects the underlying data structure. Statistical validation included MANOVA to identify landmarks differing between clusters, followed by ANOVA and post-hoc Tukey tests on individual spatial coordinates to characterize inter-cluster differences [11].

Table: Experimental Parameters in Nasal Cavity Geometric Morphometrics

Protocol Component	Specifications	Purpose
Sample Size	151 unilateral nasal cavities from 78 patients	Ensure statistical power and representativeness of anatomical variability
Fixed Landmarks	10 defined anatomical points [11]	Establish homologous reference points across all specimens
Semi-landmarks	200 sliding points [11]	Capture continuous surface curvature between fixed landmarks
Statistical Analysis	GPA, PCA, HCPC, MANOVA [11]	Identify significant shape patterns and natural morphological clusters

Research Reagent Solutions Toolkit

Table: Essential Reagents and Software for Nasal Cavity GM Analysis

Tool Name	Type/Function	Specific Application
ITK-SNAP (v3.8.0)	Segmentation Software	Semi-automatic segmentation of nasal cavity from DICOM CT images [11]
Viewbox 4.0	Landmark Digitization	Placement of fixed landmarks and semi-landmarks on 3D nasal models [11]
R Package geomorph	Statistical Analysis	Generalized Procrustes Analysis and shape statistics [11]
R Package FactoMineR	Multivariate Analysis	Hierarchical Clustering on Principal Components (HCPC) [11]
Thin Plate Spline (TPS)	Landmark Warping Algorithm	Projecting semi-landmarks from template to individual models [11]
3D Nasal Cast Model	Physical Flow Testing	In vitro validation of drug delivery efficiency [42]

Correlation Between Nasal Morphology and Drug Deposition

Identified Morphological Clusters and Olfactory Accessibility

The GM analysis revealed three distinct morphological clusters of the nasal ROI with significant implications for olfactory accessibility:

• Cluster 1: Characterized by a broader anterior cavity with shallower turbinate onset. This morphology likely improves olfactory accessibility, facilitating drug delivery to the olfactory region. Approximately 31.5% of patients had at least one nasal cavity in this cluster [11].
• Cluster 3: Exhibited a narrower configuration with deeper turbinates. This anatomy potentially limits olfactory accessibility by creating greater anatomical resistance to particle flow toward the olfactory region [11].
• Cluster 2: Represented an intermediate morphology between Clusters 1 and 3 [11].

Statistical analysis confirmed significant shape variations along the X and Y axes, with minimal variation in the Z axis, highlighting the two-dimensional nature of the primary morphological differences affecting airflow and particle transport [11].

Experimental Validation of Deposition Efficiency

Complementary in vitro studies using 3D-printed nasal cast models have quantified how delivery parameters affect deposition patterns. Research testing three different nasal spray devices (A, B, and C) found that:

• Ideal Administration Angle: A 50° angle relative to the hard palate horizontal plane resulted in the maximal spraying area on the nasal septum for all nozzles tested [42].
• Nozzle Design Impact: Nozzle C, which featured the smallest plume angle, achieved the highest total spraying distribution scores at 30° and 40° angles, demonstrating that nozzle design significantly affects delivery efficiency [42].
• Distribution Limitations: None of the three tested nozzles could effectively deliver drugs into the middle meatus at any administration angle, revealing a persistent challenge in reaching specific sinonasal regions [42].

Particle deposition studies further show that the anterior nasal airway captures particles most effectively, with deposition thickness exceeding 150 µm in some anterior regions and reaching up to 230 µm at high flow rates (55 L/min) for cohesive particles [43].

Table: Drug Deposition Efficiency by Nasal Spray Device Characteristics

Device / Parameter	Spraying Area at 50°	Optimal Administration Angle	Total Distribution Score	Key Finding
Nozzle A	Maximal	40°	30° > 40° > 50°	Performance decreases with steeper angles
Nozzle B	Maximal	30°	30° > 40° > 50°	Best performance at shallowest angle
Nozzle C (Smallest Plume)	Maximal	30°	30° > 40° > 50°	Highest overall scores; most efficient delivery

Visualization of GM Workflow and Morphology-Deposition Relationship

Geometric Morphometrics Analysis Pipeline

Relationship Between Nasal Morphology and Drug Delivery Efficiency

Cross-Validation with Complementary Methodologies

Computational Fluid Dynamics Validation

Computational Fluid Dynamics (CFD) simulations provide a quantitative cross-validation method for GM-based predictions. Studies modeling particle penetration in maxillary sinus ostia have demonstrated that geometric variations significantly impact particle distribution. Research on T-junction models (simplified ostia) revealed that:

• Anterior Radius of Curvature: Enhanced particle outflow through the y-branch (perpendicular) outlet, mimicking improved sinus penetration [44].
• Posterior Radius of Curvature: Limited particle outflow, creating greater resistance to drug penetration [44].
• Pulsating Flow: Lower frequencies (30 Hz) improved particle penetration into the y-branch and deposition in the nasal airway compared to higher frequencies (45-75 Hz) [44].

These findings validate that specific morphological features identified through GM clustering directly correspond to functional differences in particle transport efficiency.

In Vitro Experimental Validation

3D-printed nasal cast models serve as physical validation systems for GM-based classifications. The production of anatomically accurate models from CT data enables quantitative comparison of drug delivery efficiency across different morphological clusters [42]. This methodological triangulation—combining GM, CFD, and physical modeling—strengthens the validation framework and provides multiple evidence streams correlating nasal morphology with deposition patterns.

This case study demonstrates that Geometric Morphometrics provides a validated, robust protocol for classifying nasal cavity morphology with direct applications in targeted drug delivery. The integration of landmark-based shape analysis with computational and experimental validation methods establishes a comprehensive framework for understanding how anatomical variability affects delivery efficiency. The identification of three distinct morphological clusters, characterized by significantly different olfactory accessibility, provides a stratification system that can guide personalized nasal drug delivery strategies. This GM protocol successfully addresses the cross-validation requirements within nasal morphology research, offering a reproducible methodology that correlates anatomical patterns with functional delivery outcomes, ultimately supporting the development of more effective nose-to-brain therapeutic systems.

Forensic age estimation plays a critical role in medicolegal investigations, particularly in determining whether an individual has reached the age of majority for criminal responsibility [45]. The mandible, as the strongest, largest, and most frequently recovered facial bone, serves as a valuable anatomical structure for age assessment due to its significant morphological changes during growth and development and its resistance to postmortem degradation [46] [45]. This case study objectively compares the performance of different geometric morphometric protocols for age classification from mandibular morphology, with a specific focus on their cross-validation performance within forensic contexts. We evaluate traditional 2D geometric morphometrics, advanced 3D landmark-based analyses, and emerging machine learning approaches to provide researchers with evidence-based recommendations for protocol selection.

Experimental Protocols & Methodologies

2D Geometric Morphometric Analysis (GM)

Protocol Overview: The 2D geometric morphometric approach utilizes panoramic radiographs for landmark-based shape analysis. This method was applied in studies with Malay and Indonesian populations using standardized landmark placement protocols [47] [45].

Methodological Details:

• Sample Preparation: Digital panoramic radiographs were converted to TPS format using tpsUtil software. A total of 27 anatomical landmarks were identified on each mandible, including coronion, condylion, gonion, mentale, and incisor points [45].
• Landmark Digitization: Landmarks were digitized using tpsDig2 software (version 2.31) with specific definitions for each anatomical point to ensure consistency.
• Shape Analysis: Generalized Procrustes Analysis (GPA) was performed in MorphoJ software (version 1.07a) to eliminate non-shape variations through scaling, translation, and rotation. Principal Component Analysis (PCA) identified dominant shape variation patterns, followed by Discriminant Function Analysis (DFA) with cross-validation for classification accuracy testing [47] [45].
• Cross-Validation: The protocol employed Procrustes ANOVA for statistical significance testing and used leave-one-out cross-validation to assess model generalizability [45].

3D Geometric Morphometric Analysis

Protocol Overview: This approach utilizes computed tomography (CT) scans to capture comprehensive 3D mandibular morphology, offering enhanced capability to analyze complex shape changes during growth [48].

Methodological Details:

• Image Acquisition & Processing: CT scans were segmented to generate 3D models of mandibular bodies and permanent dentition using 3D Slicer software. The protocol employed 30 mandibular landmarks (type I and II) and 12 dental landmarks (type II and III) adapted from established sources [48].
• Shape Coordination Analysis: The methodology applied Generalized Procrustes Analysis to standardized landmark configurations, followed by two-block partial least squares analysis to evaluate covariation between mandibular shape and dental eruption patterns.
• Age Estimation Protocol: Linear regression tested the relationship between combined morphological proxies (mandibular shape and dental eruption) and chronological age, with validation across narrow age brackets of 3-6 months [48].

Machine Learning Approach with Mandibular Measurements

Protocol Overview: This protocol applies supervised machine learning algorithms to predict chronological age based on mandibular morphometric measurements in children and adolescents [46] [49].

Methodological Details:

• Data Collection: Lateral cephalometric radiographs from 401 orthodontic patients (6-16 years) were analyzed. Mandibular measurements included total mandibular length (Co-Pog), mandibular ramus height (Co-Go), mandibular body length (Go-Gn), and gonial angle (Ar-Go-Me) [46].
• Feature Selection & Preprocessing: Pearson correlation tests identified variables with significant correlations to chronological age (p < 0.05). Dataset normalization was performed before model training, with an 80-20 training-test split and stratified 5-fold cross-validation to prevent overfitting [46].
• Algorithm Implementation: Eight supervised machine learning algorithms were trained: Linear Regression, Gradient Boosting Regressor, Random Forest Regressor, Decision Tree Regressor, AdaBoost Regressor, Support Vector Regression, K-Nearest Neighbors Regressor, and Multilayer Perceptron Regressor [46].
• Hyperparameter Optimization: Grid Search method was used for systematic hyperparameter optimization to identify configurations that maximize predictive performance [46].

Performance Comparison & Results

Quantitative Performance Metrics

Table 1: Cross-Validation Performance of Different Mandibular Morphology Analysis Protocols

Protocol	Population	Sample Size	Age Range	Accuracy/Error	Cross-Validation Method
2D Geometric Morphometrics	Indonesian	300	15-21 years	65-67% classification accuracy	Discriminant Function Analysis with cross-validation [45]
2D Geometric Morphometrics	Malay	400	15-54 years	49-90% classification accuracy (cross-validation range)	Discriminant Function Analysis with cross-validation [47]
3D Geometric Morphometrics	New Mexico Database	48	4-13 years	Strong association with chronological age (p<0.001)	Linear regression on combined shape proxies [48]
Machine Learning (Gradient Boosting)	German	401	6-16 years	MAE: 1.21-1.54 years; R²: 0.56	Stratified 5-fold cross-validation [46]

Table 2: Feature Importance in Machine Learning Protocol

Predictor Variable	Relative Importance	Correlation with Age
Total mandibular length (Co-Pog)	Highest	Strong positive [46]
Mandibular ramus height (Co-Go)	High	Strong positive [46]
Mandibular body length (Go-Gn)	Moderate	Moderate positive [46]
Gonial angle (Ar-Go-Me)	Lower	Variable [46]

Protocol Performance Analysis

The machine learning approach demonstrated superior predictive accuracy with the lowest mean absolute error (1.21-1.54 years) among all protocols, attributed to its ability to model complex nonlinear relationships in mandibular growth patterns [46]. The Gradient Boosting Regressor emerged as the most effective algorithm, significantly outperforming linear and simpler tree models in pairwise comparisons [46].

The 2D geometric morphometric protocol showed moderate classification accuracy (65-67%) for distinguishing adolescents (15-17.9 years) from adults (18-21 years) in Indonesian population, with the first eight principal components explaining 81.8% of total shape variance [45]. Procrustes ANOVA revealed significant shape differences (P < 0.001) between age groups, though it did not show significant differences in mandibular size [45].

The 3D geometric morphometric approach provided enhanced visualization of morphological changes corresponding to different dental eruption phases, successfully capturing shape changes within narrow age brackets of 3-6 months [48]. The integration of mandibular shape with dental eruption patterns demonstrated stronger association with chronological age than either proxy independently [48].

Visualization of Methodological Workflows

2D Geometric Morphometrics Workflow

Machine Learning Protocol Workflow

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Software for Mandibular Morphometric Analysis

Tool/Category	Specific Product/Software	Function/Application	Protocol Compatibility
Radiographic Imaging	Dental Panoramic Tomography (DPT)	2D mandibular visualization	2D Geometric Morphometrics [47] [45]
Radiographic Imaging	Lateral Cephalometric Radiographs	Standardized head positioning for measurements	Machine Learning Protocol [46]
3D Imaging	Computed Tomography (CT) Scans	3D mandibular reconstruction	3D Geometric Morphometrics [48]
Landmarking Software	tpsDig2 (v2.31)	2D landmark digitization	2D Geometric Morphometrics [45]
Landmarking Software	3D Slicer	3D model generation and landmarking	3D Geometric Morphometrics [48]
Morphometric Analysis	MorphoJ (v1.07a)	Procrustes analysis, PCA, DFA	2D Geometric Morphometrics [47] [45]
Cephalometric Analysis	OnyxCeph (v3.2.180)	Cephalometric measurements	Machine Learning Protocol [46]
Programming Framework	Python scikit-learn	Machine learning implementation	Machine Learning Protocol [46]
Statistical Analysis	R with geomorph package	3D shape analysis	3D Geometric Morphometrics [48]

The cross-validation performance of different geometric morphometric protocols reveals a clear trade-off between methodological complexity and predictive accuracy. Machine learning approaches applied to standard mandibular measurements currently provide the most accurate age estimation in growing individuals, with the Gradient Boosting algorithm achieving MAE of 1.21-1.54 years through robust 5-fold cross-validation [46]. However, this approach requires precise prior knowledge of predictor variables and may be influenced by population-specific characteristics [46].

The 2D geometric morphometric protocol offers practical advantages in clinical settings with standard panoramic radiography equipment, demonstrating reasonable classification accuracy (65-67%) for distinguishing adolescents from adults [45]. The 3D approach provides superior visualization of shape changes and effectively captures integrated mandibular and dental development patterns, making it particularly valuable for understanding growth coordination [48].

For forensic applications requiring high precision age estimation in living subjects, the machine learning protocol with mandibular measurements currently delivers superior performance. For archaeological or anthropological research where visualization and understanding of morphological changes are prioritized, 3D geometric morphometrics offers greater insights. The 2D geometric morphometric approach represents a balanced solution for clinical settings with limited access to advanced imaging or computational resources.

Future research directions should focus on external validation of existing models across diverse populations, development of hybrid approaches combining machine learning with geometric morphometrics, and standardization of protocols to enhance reproducibility across different laboratory settings.

Taxonomic identification, the science of classifying living organisms, serves as a critical foundation for diverse fields, including evolutionary biology and agricultural management. In paleontology, accurate fossil identification helps unravel the history of life on Earth [50]. In agriculture, rapid pest surveillance is essential for protecting crops and ensuring food security [51] [52]. Despite their different temporal scales—deep time versus the present—both fields face the common challenge of reliably classifying specimens based on morphological characteristics.

Traditionally, taxonomic work has relied on expert examination and linear morphometrics (LMM). However, these methods can be subjective, time-consuming, and prone to biases related to size rather than shape [53]. This case study examines how two advanced methodological frameworks are addressing these challenges: Geometric Morphometrics (GMM) and Machine Learning (ML)-based identification. GMM offers a robust, holistic analysis of shape by accounting for size and allometric effects [53] [39], while ML, particularly deep learning, provides powerful tools for automated, high-throughput classification from images and acoustic data [54] [51] [50]. The performance and cross-validation of these protocols are critically evaluated within the context of paleontological and pest surveillance research.

Methodological Frameworks and Experimental Protocols

Geometric Morphometrics (GMM) for Taxonomic Discrimination

Geometric morphometrics is a sophisticated approach to shape analysis that retains the full geometry of the structures under study. Its application is particularly valuable for differentiating between closely related species or populations where morphological differences are subtle [53] [39].

2.1.1 Core Experimental Protocol

A standard GMM workflow involves several key stages, visualized in Figure 1.

dot Source Code for GMM Workflow Diagram

Figure 1. A standard GMM workflow for taxonomic analysis.

• Data Acquisition: Specimens are digitized using 2D photography or 3D scanning to create high-fidelity digital models [39].
• Landmarking: Biologically homologous points (e.g., suture intersections on a skull) are identified and marked on each digital specimen. Semi-landmarks are used to capture the geometry of curves and surfaces [53].
• Procrustes Superimposition: This crucial step removes non-shape variations by scaling all specimens to a unit size (Centroid Size), translating them to a common position, and rotating them to minimize the distances between corresponding landmarks. This isolates the pure "shape" component for analysis [53].
• Allometric Correction: The Procrustes shape coordinates are regressed against Centroid Size to quantify and remove the effect of allometry (disproportionate shape change due to size). This prevents the misidentification of size-related shape changes as taxonomic differences [53].
• Multivariate Statistical Analysis: The corrected shape data are analyzed using Principal Component Analysis (PCA) to visualize major trends of shape variation, and Canonical Variates Analysis (CVA) to maximize separation among pre-defined groups (e.g., species) [53] [55].
• Cross-Validation and Testing: The discriminatory power of the model is tested using cross-validation, often a leave-one-out procedure, where each specimen is classified based on functions derived from all other specimens. This provides a realistic estimate of the model's classification accuracy [55].

Machine Learning for Automated Identification

Machine learning, especially deep learning, automates taxonomic identification by learning discriminative features directly from large datasets, such as images or audio recordings.

2.2.1 Image-Based Fossil Identification Protocol

A landmark study by Liu et al. (2022) demonstrated the application of deep learning for fossil identification on a massive scale [50].

• Dataset Curation: The "Fossil Image Dataset (FID)" was assembled using web crawlers, amassing over 415,000 images spanning 50 fossil clades.
• Model Training and Transfer Learning: Three powerful Convolutional Neural Network (CNN) architectures were trained on this dataset. The study employed transfer learning, where a model pre-trained on a general image dataset (e.g., ImageNet) is fine-tuned on the specialized fossil images, significantly boosting performance, especially with limited data.
• Performance Evaluation: Model accuracy was evaluated on a held-out test set of images not seen during training.

2.2.2 Acoustic-Visual Pest Surveillance Protocol

A novel approach for non-invasive pest monitoring involves converting insect sounds into images for deep learning analysis [51] [52]. The workflow is illustrated in Figure 2.

• Audio Pre-processing: Raw acoustic signals of pests are filtered (low-pass filter) and downsampled to reduce noise and computational complexity.
• Cross-Modal Representation: The 1D audio signals are converted into 2D visual representations called Patch-level log-scale mel spectrum (PLMS) spectrograms. This technique enhances subtle, low-frequency insect sound features through logarithmic transformation and patch-level decomposition.
• Data Augmentation and Transfer Learning: The generated PLMS spectrograms are augmented (e.g., by rotating, shifting) to artificially increase dataset size and improve model robustness. A pre-trained object detection model (YOLOv11) is then fine-tuned on these spectrograms to perform pest classification [51] [52].
• Performance Metrics: Classification performance is measured using Accuracy, Macro-F1 score, and Macro-AUC on a separate test set.

dot Source Code for Pest Surveillance Workflow Diagram

Figure 2. An acoustic-visual ML workflow for pest surveillance.

Comparative Performance Analysis

Quantitative Performance Data

The following tables summarize the performance outcomes of the different methodological protocols as reported in the literature.

Table 1: Performance of Geometric Morphometrics vs. Linear Morphometrics [53]

Method	Key Feature	Discrimination Power	Effect of Allometric Correction
Geometric Morphometrics (GMM)	Holistic shape analysis using landmarks.	Better group discrimination after isometry and allometry are removed.	Correctly discriminates based on non-allometric shape differences.
Linear Morphometrics (LMM)	Point-to-point linear measurements.	High for raw data, but may be inflated by size variation.	Discrimination often comes from size variation rather than true shape differences.

Table 2: Performance of Machine Learning-Based Identification Methods

Application	Method / Model	Dataset	Key Performance Metric(s)
Fossil Identification [50]	Inception-ResNet-v2 (CNN)	Fossil Image Dataset (415,339 images, 50 clades)	Average Accuracy: 90% (Microfossils: 95%, Vertebrates: 90%)
Pest Surveillance [51] [52]	PLMS Spectrograms + YOLOv11	InsectSound1000 Database	Accuracy@1: 96.49%, Macro-F1: 96.49%, Macro-AUC: 99.93%
General Paleontology [54]	Deep Learning (Various CNNs)	Various Fossil Datasets	Improves classification accuracy and overcomes observer bias.

Critical Evaluation of Cross-Validation Performance

The robustness of any taxonomic model is determined by its performance on unseen data, making cross-validation (CV) strategies a critical aspect of methodological evaluation.

• GMM and Cross-Validation: In GMM, leave-one-out cross-validation is commonly used with CVA. Studies highlight that using a variable number of Principal Component (PC) axes to optimize the cross-validation assignment rate yields higher and more reliable classification success than using a fixed number of axes or other dimension-reduction methods [55]. This prevents overfitting and provides a realistic measure of the model's predictive power.

• ML and Spatial Cross-Validation: In machine learning, especially with geospatial data like UAV crop surveys, the standard random CV can produce overly optimistic results. Studies recommend spatially-aware CV (e.g., leaving out an entire field) for a more realistic assessment of a model's transferability to new, independent locations [56]. While this was demonstrated for yield prediction, the principle is directly applicable to pest surveillance models deployed across different farms or ecosystems. Without proper spatial CV, model performance in real-world "extrapolation" tasks can be disappointing [56].

• Reproducibility Challenge in ML: A review of ML in paleontology found that reproducibility is a significant issue, with only 37.0% of studies making their code publicly available and 56.5% providing public access to their data [54]. This hinders the independent validation and comparative assessment of different ML protocols.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Tools and Solutions for Taxonomic Identification Research

Tool / Solution	Category	Primary Function
2D/3D Digitization Equipment	Hardware	Creates high-resolution digital models of specimens for GMM or ML analysis.
Landmarking Software (e.g., tpsDig2, MorphoJ)	Software	Allows precise placement of landmarks and semi-landmarks on digital specimens for GMM.
Procrustes Superimposition Algorithm	Analytical	The computational core of GMM; aligns specimens to isolate shape from size, position, and orientation.
Convolutional Neural Network (CNN)	Analytical	A class of deep learning models that automatically learns features from images for classification.
Pre-trained Models (e.g., YOLOv11, Inception-ResNet-v2)	Analytical	Enables transfer learning, drastically reducing the data and computational resources needed for effective ML model training.
High-Sensitivity Microphones / Acoustic Sensors	Hardware	Captures bioacoustic signals for non-invasive pest surveillance via audio analysis [51].
Spatial Cross-Validation Scripts	Analytical	Ensures robust evaluation of model performance and true transferability to new locations [56].

This case study reveals a convergent evolution in taxonomic methodologies across paleontology and pest surveillance. Both fields are increasingly adopting data-driven, quantitative approaches to overcome the limitations of traditional identification methods.

• Geometric Morphometrics excels in scenarios requiring deep biological insight into shape variation, providing a rigorous framework for controlling allometric effects and offering interpretable visualizations of shape change. Its strength lies in its grounding in biological homology.
• Machine Learning offers unparalleled scalability and automation, capable of processing vast datasets (e.g., hundreds of thousands of fossil images) with high accuracy. Its performance is tightly linked to data quality, quantity, and robust cross-validation practices, particularly spatial validation for field-deployed models.

The choice between these protocols is not necessarily mutually exclusive. An integrative approach, where GMM helps identify diagnostically significant features that can inform the development of simpler linear measurements or provide interpretability to ML models, is likely the most powerful path forward [53]. Ultimately, the credibility of findings in both fields hinges on moving beyond simple raw accuracy metrics and adopting rigorous, transparent cross-validation strategies that truly test a model's predictive power and real-world applicability.

Addressing Pitfalls and Optimizing Protocol Robustness

Overcoming the 'Out-of-Sample' Problem in Real-World Classification

In the realm of data-driven science, the true test of any classification model lies not in its performance on the data it was trained on, but in its ability to generalize to new, unseen data—a challenge known as the "out-of-sample" problem. This fundamental issue separates theoretical model performance from practical utility across research domains, from geometric morphometrics to drug development. The out-of-sample problem emerges from a simple but dangerous assumption: that future data will perfectly mirror the characteristics of past data. In reality, biological variability, measurement inconsistencies, and temporal changes create inevitable mismatches between training datasets and real-world applications. When models fail to generalize, the consequences extend beyond statistical error to potentially flawed scientific conclusions and costly misapplications in critical domains like pharmaceutical development.

The evaluation of machine learning models works on a constructive feedback principle: build a model, get feedback from metrics, make improvements, and continue until achieving desirable classification accuracy on out-of-sample data [57]. Evaluation metrics provide crucial insights into model performance, but their most critical function is their capability to discriminate among model results when applied to new data [57]. This challenge is particularly acute in fields like geometric morphometrics, where the mathematical requirements of multivariate statistics often conflict with the practical limitations of specimen availability, creating a perfect storm of generalization challenges.

Critical Evaluation Metrics for Generalization Assessment

Core Classification Metrics

Understanding model performance requires multiple evaluation perspectives, as no single metric captures the complete picture of generalization capability. The confusion matrix forms the foundation of classification assessment, providing the raw data from which key metrics are derived [57]. This N x N matrix (where N is the number of classes) enables the calculation of several critical statistics: Accuracy measures the overall proportion of correct predictions; Precision quantifies how many of the positively identified cases were actually correct; Recall (or Sensitivity) measures how many of the actual positive cases were correctly identified; and Specificity assesses how well the model identifies negative cases [57]. Each metric offers a different lens through which to view model performance, with optimal balance depending on the specific research context and consequences of different error types.

The F1-Score provides a harmonic mean of precision and recall, particularly valuable when seeking balance between these two metrics and when dealing with uneven class distributions [57]. Unlike arithmetic mean, the harmonic mean punishes extreme values more severely, providing a more conservative assessment of model performance. For scenarios where precision or recall requires differential weighting, the Fβ metric allows researchers to attach β times as much importance to recall as precision [57]. These metrics collectively form a toolkit for initial model assessment, though they primarily reflect performance on the data used for training rather than predicting out-of-sample performance.

Specialized Performance Visualizations

Beyond numerical metrics, visual assessment tools provide deeper insights into model behavior across different decision thresholds and population segments. Gain and Lift charts analyze the rank ordering of predicted probabilities, measuring how much better one can expect to do with a model compared to random selection [57]. These charts are particularly valuable in campaign targeting problems, telling researchers which population segments to target for specific interventions and what response rate to expect from new target bases.

The Kolmogorov-Smirnov (K-S) chart measures the degree of separation between positive and negative distributions, with values ranging from 0 (no separation, equivalent to random selection) to 100 (perfect separation) [57]. The Area Under the ROC Curve (AUC-ROC) provides a robust measure of classification performance that is independent of the proportion of responders in the population [57]. This independence from class distribution makes AUC-ROC particularly valuable for assessing potential out-of-sample performance where class frequencies may differ from training data.

Experimental Comparison of Classification Protocols

Geometric Morphometric Approaches

Geometric morphometric methods present unique challenges for out-of-sample classification due to the high-dimensional nature of shape data and typically limited specimen availability. In a methodological study comparing approaches for classifying feather outlines from ovenbirds (Seiurus aurocapilla), researchers examined four mathematical representation approaches and two curve measurement methods [1]. The study revealed that classification performance was not highly dependent on the number of points used to represent a curve or the precise manner of point acquisition, with semi-landmark methods (bending energy alignment and perpendicular projection) producing roughly equal classification rates, as did elliptical Fourier methods and the extended eigenshape method [1].

The critical innovation in this research was a new approach to dimensionality reduction that addresses the fundamental constraint of canonical variates analysis (CVA), which requires more specimens than the sum of the number of groups and measurements per specimen [1]. The method utilizes a variable number of principal component (PC) axes selected specifically to optimize cross-validation assignment rates, outperforming both the standard approach of using a fixed number of PC axes and partial least squares methods [1]. This finding highlights how adapting analytical procedures to maximize out-of-sample performance can yield significant improvements over conventional approaches.

Table 1: Comparison of Geometric Morphometric Outline Methods for Classification

Method Category	Specific Techniques	Classification Performance	Key Advantages	Sample Size Requirements
Semi-landmark Methods	Bending Energy Alignment (BEM), Perpendicular Projection (PP)	Roughly equal classification rates between the two approaches [1]	Allows combination of discrete landmarks with curve information [1]	High (due to many semi-landmarks needed) [1]
Mathematical Function Methods	Elliptical Fourier Analysis, Extended Eigenshape	Similar performance to semi-landmark methods [1]	Complete mathematical representation of curves [1]	High (many measurements needed) [1]
Dimension Reduction	Variable PC Axes (new method)	Higher cross-validation assignment rates [1]	Optimizes cross-validation performance [1]	Moderate (reduces dimensionality smartly) [1]
Dimension Reduction	Fixed PC Axes (standard)	Lower cross-validation assignment rates [1]	Simple implementation	Moderate [1]
Dimension Reduction	Partial Least Squares	Lower cross-validation assignment rates [1]	Maximizes covariation with classification [1]	Moderate [1]

Computer Vision and Deep Learning Approaches

Recent advances in computer vision have introduced powerful alternatives to traditional geometric morphometrics for classification tasks. In a comparative study of methods for identifying carnivore agency from tooth marks, geometric morphometric methods demonstrated limited discriminant power (<40%) in bidimensional applications [3]. In contrast, computer vision approaches utilizing deep convolutional neural networks (DCNN) and Few-Shot Learning (FSL) models classified experimental tooth pits with significantly higher accuracy (81% and 79.52% respectively) [3].

This performance disparity highlights a fundamental distinction between method types: while GMM struggles with the wide range of allometrically-conditioned tooth pits, particularly non-oval variants, computer vision methods can inherently manage this diversity [3]. However, the study noted important limitations when applying computer vision to fossil records, where bone surface modifications undergo dynamic transformations over time, potentially altering original properties [3]. In well-preserved contexts such as 1.8 million-year-old tooth marks from Olduvai sites, computer vision models can achieve high agent attribution probability, demonstrating their potential value despite implementation challenges [3].

Table 2: Performance Comparison of Classification Methods for Biological Shapes

Method Type	Specific Approach	Reported Accuracy	Strengths	Out-of-Sample Limitations
Geometric Morphometrics	Outline Analysis (Bidimensional)	<40% classification accuracy [3]	Mathematical representation of form	Limited discriminant power for diverse shapes [3]
Computer Vision	Deep Convolutional Neural Networks (DCNN)	81% accuracy [3]	Handles shape diversity effectively	Requires large training datasets [3]
Computer Vision	Few-Shot Learning (FSL)	79.52% accuracy [3]	Works with limited examples	Complex implementation [3]
Semi-supervised Learning	Multi-mode Augmentation	Significant improvement over baseline methods [58]	Effective with limited labeled data	Performance depends on unlabeled data quality [58]

Semi-Supervised Learning for Limited Data Environments

Many real-world classification scenarios in scientific research face the challenge of limited labeled data, precisely the situation where out-of-sample problems become most acute. A novel semi-supervised learning method based on multi-mode augmentation addresses this challenge by simultaneously improving sample completeness within and between classes [58]. This approach combines uncertainty-aware pseudo-label selection with a multi-modal data augmentation strategy integrating intra-class random augmentation and inter-class mixed augmentation [58].

The methodology specifically addresses two aspects of sample completeness: intra-class completeness (sufficient diversity of examples within a category) and inter-class completeness (adequate representation of relationships between categories) [58]. Traditional approaches using single augmentation techniques improve only one dimension of completeness, while the multi-mode approach leverages both random augmentation (enhancing intra-class diversity) and mixed augmentation (improving inter-class relationships) [58]. Experimental results on STL-10 and CIFAR-10 datasets demonstrate significantly better generalization performance compared to existing mainstream methods in scenarios with small unlabeled data and mismatched samples [58].

Methodological Protocols for Robust Classification

Cross-Validation Strategies

Proper validation methodologies form the first line of defense against poor out-of-sample performance. The resubstitution estimator (the rate of correct assignments using the same data that formed the classification model) is known to be biased upward, as it fails to account for model overfitting [1]. Cross-validation provides a more realistic assessment by leaving one or more specimens out of the training set used to form discriminant functions, then assigning these held-out specimens based on the derived models [1].

The number of dimensions used in classification significantly impacts out-of-sample performance. Using large numbers of principal component axes in CVA may yield high resubstitution rates but substantially lower cross-validation rates due to overfitting [1]. Reducing the number of PC axes may decrease resubstitution performance but increase cross-validation accuracy, properly prioritizing generalization over apparent fit [1]. Bootstrapping approaches can further refine these estimates by resampling data with replacement and carrying out the entire CVA analysis on bootstrapped datasets to determine confidence intervals on classification rates [1].

Dimensionality Reduction Protocols

The challenge of dimensionality is particularly acute in morphological classification, where the number of variables often approaches or exceeds the number of specimens. The linear CVA requires matrix inversion of the pooled within-group variance-covariance matrix, which must be of full rank—requiring more specimens than the sum of the number of groups and measurements per specimen [1]. When this condition is not met, there are more degrees of freedom in the measurements than in the specimens, guaranteeing overfitting and poor out-of-sample performance.

The variable PC axes approach demonstrates how tailored dimensionality reduction can optimize out-of-sample performance [1]. By calculating cross-validation rates across different numbers of PC axes and selecting the number that maximizes out-of-sample accuracy, researchers can avoid both underfitting (too few dimensions) and overfitting (too many dimensions) [1]. This approach outperforms both fixed PC axis selection and partial least squares methods that decompose the covariance matrix between measurements and classification codes using singular value decomposition [1].

Diagram Title: Geometric Morphometric Classification Workflow

Research Reagent Solutions for Classification Studies

Essential Software and Computational Tools

Implementing robust classification protocols requires specific computational tools that facilitate both analysis and validation. R with geomorph package provides a comprehensive open-source environment for geometric morphometric analyses, including Procrustes alignment, principal components analysis, and canonical variates analysis with cross-validation capabilities. Python with Scikit-learn offers machine learning implementations for classification algorithms, cross-validation strategies, and performance metrics critical for assessing out-of-sample performance. MATLAB with Shape Modeling Toolbox delivers commercial solution for mathematical representation of shapes, particularly valuable for elliptical Fourier analysis and extended eigenshape methods.

Specialized visualization tools form another critical component of the classification toolkit. MorphoJ facilitates visualization of shape changes along discriminant axes, helping researchers interpret biological meaning behind statistical classification. TPS series software (tpsDig, tpsRelw) enables landmark digitization, relative warps analysis, and thin-plate spline visualization, connecting raw data to biological interpretation. For deep learning approaches, TensorFlow or PyTorch with computer vision libraries provide the infrastructure for implementing convolutional neural networks and few-shot learning approaches that can outperform traditional morphometric methods.

Experimental Design Reagents

Proper experimental design significantly impacts out-of-sample performance before analysis begins. Reference Specimen Collections with known classification provide essential ground truth for initial model training and validation, with sample sizes sufficient to support the dimensionality of measurements being collected. Standardized Imaging Protocols ensure consistent data quality and minimize technical variance that could artificially inflate or deflate apparent classification performance, including controlled lighting, scale, and orientation.

The statistical toolkit for validation represents perhaps the most crucial reagent category. Cross-Validation Frameworks implement leave-one-out and k-fold validation to provide realistic performance estimates, with particular attention to stratification that maintains class representation across folds. Bootstrapping Implementations generate confidence intervals for classification rates through resampling, quantifying uncertainty in performance estimates that is essential for proper interpretation of model utility.

Table 3: Essential Research Reagents for Classification Studies

Reagent Category	Specific Tools	Primary Function	Role in Addressing Out-of-Sample Problem
Statistical Software	R, Python, MATLAB	Data analysis and modeling	Implement cross-validation and performance assessment [1]
Morphometric Software	MorphoJ, TPS series	Shape analysis and visualization	Facilitate proper landmark alignment and data collection [1]
Deep Learning Frameworks	TensorFlow, PyTorch	Neural network implementation	Enable computer vision approaches that may outperform traditional methods [3]
Validation Protocols	Cross-validation, bootstrapping	Performance assessment	Provide realistic out-of-sample performance estimates [1]
Sample Collections	Reference specimens with known classification	Model training and validation	Provide ground truth for establishing baseline performance [1] [3]

Integrated Workflow for Optimal Out-of-Sample Performance

Diagram Title: Semi-Supervised Multi-Mode Augmentation Workflow

Achieving robust out-of-sample classification performance requires systematic integration of the methodologies discussed throughout this guide. The workflow begins with data acquisition and preprocessing using standardized protocols to minimize technical variance, followed by appropriate dimensionality reduction that balances information retention against overfitting risk. The critical third stage implements rigorous cross-validation not merely as an assessment step but as an integral component of model selection, optimizing parameters specifically for out-of-sample performance rather than training set accuracy.

For challenging domains with limited labeled data, the semi-supervised learning approach with multi-mode augmentation provides a powerful framework [58]. By combining uncertainty-aware pseudo-label screening with both intra-class random augmentation and inter-class mixed augmentation, this methodology addresses both dimensions of sample completeness essential for generalization [58]. The integration of interleaved equalization processing with exponential moving average techniques further stabilizes and improves model performance in small-sample environments [58].

The final implementation must prioritize interpretability alongside accuracy, ensuring that classification models produce biologically meaningful results that researchers can understand and trust. This often involves visualization techniques that connect statistical classification to underlying morphological patterns, creating a feedback loop between quantitative analysis and domain expertise. Through this comprehensive approach, researchers can overcome the out-of-sample problem, developing classification systems that maintain their validity when applied to new data in real-world scientific contexts.

Mitigating Bias from Landmark Selection and Semi-Landmark Placement

In geometric morphometrics, the selection of landmarks and the placement of semi-landmarks are foundational steps that directly influence all subsequent shape analyses and biological interpretations. These initial choices introduce potential biases that can skew statistical results and lead to erroneous evolutionary or taxonomic conclusions [59]. The pursuit of methodological rigor demands careful consideration of how these biases originate and strategies to mitigate them, particularly when evaluating the cross-validation performance of different geometric morphometric protocols.

Bias in landmark selection can manifest through multiple pathways: oversampling of certain anatomical regions, reliance on non-homologous points, or inconsistent placement across specimens [25] [59]. Similarly, semi-landmark placement introduces mathematical biases through different algorithms that optimize for varying criteria, whether bending energy, Procrustes distance, or surface correspondence [60] [59]. These methodological decisions become particularly critical in cross-validation frameworks, where the goal is to develop protocols that generalize well to new datasets and maintain biological meaningfulness beyond the immediate sample.

This guide systematically compares contemporary approaches to landmark and semi-landmark methodologies, focusing specifically on their propensity to introduce or mitigate bias, with particular emphasis on cross-validation performance. We present experimental data quantifying these effects and provide researchers with evidence-based recommendations for selecting appropriate protocols based on their specific research questions and dataset characteristics.

Comparative Analysis of Morphometric Approaches

Methodological Spectrum and Bias Profiles

Table 1: Comparison of Major Morphometric Approaches and Their Bias Characteristics

Method Category	Specific Techniques	Primary Sources of Bias	Bias Mitigation Strategies	Cross-Validation Performance
Traditional Landmarking	Manual anatomical landmark placement [61]	Observer error, landmark homology interpretation, regional oversampling [18] [62]	Multiple observers, training calibration, hierarchical landmark selection [62]	Variable; improves with observer training and subset selection [62]
Semi-Landmark Patch Approaches	Patch-based, Patch-TPS [60]	Template selection, projection artifacts, surface normal estimation [60] [59]	Multiple template testing, normal vector smoothing, outlier detection [60]	Generally good; Patch-TPS shows better robustness to noise [60]
Landmark-Free Methods	DAA (Deterministic Atlas Analysis) [18] [63]	Initial template selection, kernel width parameterization, mesh topology [18]	Poisson surface reconstruction, template optimization, kernel width testing [18]	High for disparate taxa; comparable to manual landmarking in macroevolution [18]
Automated Landmarking	FaceDig, MeshMonk [25] [64]	Training dataset composition, algorithm architecture, image quality [25] [64]	Diverse training data, multi-stage refinement, quality control visualization [25]	Excellent; demonstrates human-level precision with high consistency [25] [64]
Subset Optimization	Hierarchical selection, random combinatorial approach [62]	Overfitting to specific training set, ignoring integrated shape information	Cross-validation with multiple random splits, Procrustes ANOVA validation [62]	Can outperform full landmark sets; reduces overfitting through simplification [62]

Quantitative Performance Comparison

Table 2: Experimental Performance Metrics Across Methodologies

Methodology	Placement Error (mm)	Processing Time	Inter-Method Correlation	Phylogenetic Signal Retention	Disparity Estimation Accuracy
Manual Landmarking	1.5-2.5 (expert) [64]	High (hours-days)	Reference standard	High with sufficient landmarks [61]	Variable; dependent on coverage [18]
Patch Semi-Landmarks	1.8-3.2 (depends on surface) [60]	Medium (minutes-hours)	R² = 0.85-0.95 with manual [60]	Comparable to manual landmarks [60]	Slight overestimation with noise [60]
Patch-TPS	1.5-2.1 [60]	Medium (minutes-hours)	R² = 0.89-0.97 with manual [60]	High across great ape species [60]	Robust to missing data [60]
DAA (Landmark-Free)	N/A (diffeomorphic) [18] [63]	Low after setup	R² = 0.80-0.96 with manual [18]	Comparable to manual landmarking [18]	Comparable with manual methods [18]
Automated (FaceDig)	1.2-1.8 [25]	Very low (seconds)	ICC > 0.988 with manual [25]	Not assessed	Not assessed
Automated (MeshMonk)	1.5 ± 0.3 mm [64]	Low (minutes)	ICC > 0.988 with manual [64]	Not assessed	Not assessed

Detailed Experimental Protocols and Methodologies

The patch-based approach generates semi-landmarks by projecting points from geometrically defined patches onto specimen surfaces. The detailed methodology consists of:

• Patch Definition: Select three manually digitized landmarks to form triangular patches covering regions of interest. Any complex polygonal region can be decomposed into multiple triangles.

• Grid Registration: Create a template triangular grid with user-defined sampling density. Register this grid to the specimen's bounding triangle using thin-plate spline (TPS) deformation.

• Surface Projection:

  • Apply Laplacian smoothing to surface normal vectors to reduce noise impact
  • Estimate patch surface orientation by averaging normal vectors at the three defining landmarks
  • Cast projection rays from grid points to specimen surface, constrained by average inter-vertex distance
  • Implement fallback procedures: reverse ray direction upon non-intersection, select closest mesh point as last resort

• Patch Merging:

  • Identify unique triangle edges in the grid
  • Place uniformly sampled points along edges with user-defined sampling rate
  • Project edge points to surface
  • Combine with manual landmarks into final configuration

This method preserves geometric relationships between semi-landmarks and manual landmarks but shows sensitivity to surface noise and complex topography.

The DAA approach eliminates landmark dependency through diffeomorphic mapping:

• Atlas Generation:

  • Select initial template mesh (choice minimally affects results with proper optimization)
  • Iteratively estimate optimal atlas shape by minimizing total deformation energy to map onto all specimens
  • Generate control points distributed in ambient space surrounding atlas, with density adapting to areas of greater variability

• Momentum Calculation:

  • For each control point, compute momentum vectors ("momenta") representing optimal deformation trajectory for atlas-to-specimen alignment
  • Apply Hamiltonian framework derived from velocity field of ambient space

• Shape Comparison:

  • Use momentum vectors as basis for shape comparison
  • Apply kernel principal component analysis (kPCA) to visualize and explore covariation in shape data

• Mesh Standardization (Critical for Mixed Modalities):

  • Apply Poisson surface reconstruction to create watertight, closed surfaces
  • Standardize mesh topology across specimens from different imaging sources (CT, surface scans)
  • Optimize kernel width parameter (10-40mm range tested) controlling spatial deformation extent

This method demonstrates particular strength for macroevolutionary analyses across highly disparate taxa where homologous landmarks become scarce.

The FaceDig approach implements a two-stage artificial intelligence pipeline for facial landmarking:

• Rough Projection Phase:

  • Utilize pre-existing facial landmark detection (MediaPipe) for initial estimates
  • Apply linear projection from 478-dimensional MediaPipe landmark space to custom 72-landmark configuration
  • Train bias-free ensemble of three linear layers for 150 epochs (learning rate 0.001)

• CNN Refinement Phase:

  • Extract image crops centered on each roughly projected landmark
  • Process each crop through dedicated lightweight CNN architecture
  • Implement multi-crop ensemble strategy (six crop sizes: 72, 64, 56, 48, 40, 32 pixels)
  • CNN architecture: three convolutional layers (channel depths 16, 32, 64; kernel sizes 5, 3, 3) with max/average pooling, followed by two-layer MLP (64 hidden units)
  • Train refinement networks for 20 epochs using Wing Loss function (learning rate 0.001)

• Skip Connection Integration: Combine refined landmark positions with rough projections through skip connections to generate final coordinates.

This method achieves human-level precision while dramatically reducing processing time and observer bias.

Visualizing Methodological Relationships and Bias Mitigation

Diagram 1: Methodological workflow showing relationships between approaches and bias mitigation strategies. The framework emphasizes cross-validation performance as the critical evaluation metric for protocol selection.

Table 3: Key Software Tools and Analytical Resources

Tool/Resource	Primary Function	Application Context	Bias Mitigation Features	Accessibility
3D Slicer with SlicerMorph [60]	3D visualization and landmarking	Medical image analysis, biological morphometrics	Open-source, reproducible workflows, patch-based semi-landmarks	Free, open-source
MorphoJ [61]	Statistical shape analysis	General morphometrics, allometry studies	Procrustes ANOVA, measurement error assessment	Free for academic use
Geomorph R Package [60]	GM analysis in R	Comprehensive statistical analysis	Sliding semi-landmarks, phylogenetic integration	Free, open-source
MeshMonk [64]	Dense surface correspondence	Automated phenotyping, high-density analysis	Quality control visualization, standardized protocols	Free for research
Deformetrica [18] [63]	Diffeomorphic mapping	Landmark-free analysis, disparate taxa comparison	Atlas-based normalization, kernel width optimization	Free for academic use
FaceDig [25]	Automated facial landmarking	2D facial photograph analysis	AI-based consistency, ethnic diversity training	Free, open-source
TPS Dig Series [65] [61]	Manual landmark digitization	Traditional landmarking, educational purposes	Established standard, format compatibility	Freeware

The cross-validation performance of geometric morphometric protocols depends fundamentally on appropriate method selection guided by research questions and dataset characteristics. Traditional manual landmarking remains valuable for analyses requiring explicit biological homology, particularly when combined with subset optimization techniques that surprisingly outperform full landmark sets in discrimination tasks [62]. Semi-landmark approaches significantly enhance shape information capture from smooth surfaces and complex topographies, with patch-TPS demonstrating superior robustness to dataset noise and missing data compared to basic patch methods [60] [59].

Landmark-free methods like Deterministic Atlas Analysis represent a paradigm shift for analyses across highly disparate taxa where homologous landmarks become limiting, showing particular strength in macroevolutionary contexts [18] [63]. Automated landmarking approaches achieve human-level precision with dramatically improved consistency and processing efficiency, making them ideal for large-scale studies where standardization is paramount [25] [64].

Critical to all approaches is the implementation of appropriate bias mitigation strategies, including multiple observer calibration for manual methods, template optimization and surface reconstruction for landmark-free approaches, and diverse training data for automated systems. Cross-validation performance should be explicitly tested through Procrustes ANOVA, leave-one-out validation, and out-of-sample testing protocols [2] to ensure methodological choices yield biologically meaningful results generalizable beyond immediate study samples. Through strategic protocol selection and rigorous validation, researchers can effectively mitigate biases inherent in landmark selection and placement, ensuring the robustness and biological validity of morphometric conclusions.

Strategies for Handling Missing Data and Incomplete Specimens

Geometric morphometrics (GM) has become a fundamental tool for quantifying biological shape in ecological, evolutionary, and paleontological studies. However, a pervasive challenge in morphological research involves handling incomplete specimens—those with missing data resulting from postmortem damage, pathological conditions, preservation artifacts, or fossilization processes. Such specimens are frequently encountered in museum collections and paleontological assemblages, potentially limiting sample sizes and introducing bias when excluded from analyses. The strategic management of these specimens is crucial for maintaining statistical power and preserving important morphological variation within datasets. This guide compares the performance of different protocols for handling missing data, with particular emphasis on their impact on cross-validation performance within geometric morphometric analyses.

Methodological Approaches: Exclusion versus Estimation

Researchers facing incomplete specimens must choose between two fundamental strategies: excluding problematic specimens or estimating missing data. Each approach carries distinct implications for analytical outcomes and statistical reliability.

Specimen Exclusion

The most straightforward method involves removing incomplete specimens from analyses. While this eliminates potential sources of error, it simultaneously reduces sample sizes and may systematically bias datasets by excluding rare taxa or specific demographic groups more likely to exhibit damage [66]. Studies indicate that specimen exclusion should be reserved for cases of extreme fragmentation, as the impact of missing data on geometric morphometric analysis is disproportionately affected by the most fragmentary specimens [67]. For robust analyses, Cardini et al. (2015) recommended minimum sample sizes of 15-20 specimens per group to reliably estimate mean shape and variance [66].

Missing Data Estimation

Alternatively, researchers can employ estimation techniques to retain incomplete specimens in analyses. Multiple methods exist for reconstructing missing landmark data:

• Regression-Based Estimation: Predicts missing coordinates using relationships with complete landmarks within the dataset [68]
• Bayesian Principal Components Analysis (BPCA): Uses probabilistic modeling to estimate missing values based on covariance structure [68]
• Thin-Plate Spline (TPS) Interpolation: A geometric-morphometric-specific method that uses spline deformation to predict missing landmarks [69]
• Fully Conditional Specification and Expectation-Maximization Algorithms: Advanced multiple imputation techniques that show strong performance [70]

Table 1: Performance Comparison of Missing Data Estimation Methods

Method	Accuracy	Reliability	Best Use Cases	Limitations
Regression-Based Estimation	High	High	Datasets with strong integration patterns	Performance depends on correlation structure
Bayesian PCA	High	Moderate-High	General purpose estimation	Computational complexity
Fully Conditional Specification	High	High	Diverse dataset structures	Requires specialized implementation
Expectation-Maximization Algorithms	High	High	Multivariate normal data	Assumption-dependent
Thin-Plate Spline (TPS)	Variable	Low-Moderate	Geometrically predictable missing data	Less reliable across diverse datasets [69]

Experimental Performance Data

Estimation Accuracy Across Taxa

Experimental studies simulating missing data across multiple taxonomic groups (modern fish, primates, and extinct theropod dinosaurs) have quantified the performance of different estimation methods [67]. These investigations reveal that standard estimation techniques generally provide more reliable estimators with lower impacts on morphometric analysis compared to geometric-morphometric-specific estimators like TPS.

For most datasets, estimating missing data produced a better fit to the structure of the original data than exclusion of incomplete specimens, a pattern maintained even at considerably reduced sample sizes [67]. The effectiveness of specific estimators varies across anatomical regions and taxonomic groups, with regression-based estimation consistently outperforming other methods, particularly in datasets with high taxonomic diversity [68].

Impact of Missing Data Percentage

The accuracy of missing data estimation shows an inverse relationship with the percentage of missing landmarks. Research indicates that estimation errors increase across all methods as missing landmarks exceed 50% of the total landmark configuration [68]. Beyond this threshold, even advanced estimation methods show significantly poorer fits, suggesting that specimens with extreme incompleteness may be unsuitable for analysis.

Table 2: Performance Metrics by Missing Data Percentage

Missing Data Percentage	Estimation Accuracy	Recommended Action	Statistical Power Preservation
<10%	High	Estimate missing data	Excellent
10-30%	Moderate-High	Estimate missing data	Good
30-50%	Moderate	Estimate with caution	Fair
50-70%	Low	Consider exclusion	Poor
>70%	Very Low	Exclusion recommended	Very Poor

Clavel et al. (2014) developed an approach combining multiple imputation with Procrustes superimposition of principal component analysis results to visualize the effect of individual missing data estimation on ordinated space, providing a practical diagnostic tool for researchers [70].

Cross-Validation Performance of Different Protocols

Cross-validation procedures provide critical insights into the practical performance of different missing data protocols by assessing how well analyses generalize to new data.

Dimensionality Reduction Considerations

When applying discriminant analyses like Canonical Variates Analysis (CVA) to outline data, dimensionality reduction becomes necessary due to the high number of variables relative to typical sample sizes [1]. A variable number of principal component (PC) axes approach, which optimizes cross-validation assignment rates, has demonstrated superior performance compared to using a fixed number of PC axes or partial least squares methods [1] [71].

The resubstitution estimator (rate of correct assignments using the same data that formed the CVA) typically shows upward bias, while cross-validation provides a more realistic assessment of classification performance [1]. This distinction becomes particularly important when evaluating protocols for handling missing data, as overfitting becomes a significant risk with complex estimation procedures.

Specimen Inclusion and Statistical Power

The strategic inclusion of incomplete specimens through estimation generally enhances cross-validation performance by preserving statistical power and representing broader morphological variation. Analyses demonstrate that estimating missing data typically produces better fit to biological shape variation patterns than excluding incomplete specimens [67] [69].

However, the effectiveness of this approach depends on appropriate estimator selection and the anatomical distribution of missing data. Landmarks in highly variable anatomical regions (e.g., around the head) often show poorer estimation accuracy compared to more constrained regions (e.g., caudal landmarks) [68]. Researchers should evaluate estimators specifically for their dataset and landmark configurations rather than relying on generalized recommendations.

Practical Implementation Workflow

The following diagram illustrates a systematic decision protocol for handling incomplete specimens in geometric morphometric studies:

Research Reagent Solutions

Table 3: Essential Computational Tools for Missing Data Handling

Tool/Software	Function	Implementation Considerations
R Statistical Software	Primary platform for morphometric analyses	Extensive community support and packages
LOST R Package	Specifically designed for missing morphometric data	Accommodates both 2D and 3D data [69]
Geomorph R Package	Comprehensive geometric morphometrics	Integrates with LOST for data exchange [69]
Bayesian PCA	Probabilistic missing data estimation	Effective for general-purpose estimation [68]
Regression-Based Estimation	Predicts missing coordinates	Consistently high performance across taxa [68]
Thin-Plate Spline	Geometric-morphometric-specific estimation	Variable reliability; use with verification [69]
Generalized Procrustes Analysis	Standardizes landmark configurations	Required preprocessing after estimation
Cross-Validation Protocols	Validates estimation performance	Critical for assessing methodological choices [1]

The strategic handling of missing data and incomplete specimens significantly influences analytical outcomes in geometric morphometric studies. Based on experimental evidence, the exclusion of moderately incomplete specimens generally produces poorer results than informed estimation, particularly when cross-validation performance is the primary metric. Regression-based and multiple imputation methods typically outperform geometric-morphometric-specific approaches like thin-plate spline for estimating missing landmarks.

Researchers should implement a stratified approach based on the percentage and distribution of missing data, validate all estimation procedures through cross-validation, and carefully consider the trade-offs between statistical power and potential estimation errors. By adopting these evidence-based protocols, researchers can maximize the utility of valuable morphological datasets while maintaining analytical rigor in geometric morphometric studies.

Allometry, the study of the relationship between size and shape, remains an essential concept for evolutionary biology and related disciplines [72]. In geometric morphometrics (GM), allometry refers to the size-related changes of morphological traits, which can profoundly influence the interpretation of shape variation [72] [73]. The correction for size effects represents a fundamental step in morphological analyses, particularly when the research goal is to isolate shape differences independent of size variation [72]. This guide compares the performance of different protocols for identifying and correcting for allometric effects within the context of cross-validation performance, providing researchers with evidence-based recommendations for selecting appropriate methodologies.

The distinction between two main schools of thought proves useful for understanding differences and relationships between alternative methods [72]. The Gould-Mosimann school defines allometry as the covariation of shape with size, typically implemented through multivariate regression of shape variables on a measure of size [72]. In contrast, the Huxley-Jolicoeur school characterizes allometry as the covariation among morphological features that all contain size information, implemented through principal component analysis in Procrustes form space or conformation space [72]. These frameworks, while conceptually distinct, are logically compatible and provide investigators with flexible tools to address specific questions concerning evolution and development [72].

Comparative Performance of Allometry Correction Methods

Key Methodological Approaches

Table 1: Core Methodological Frameworks for Allometry Analysis

Methodological Framework	Statistical Implementation	Size Measurement	Shape Space	Primary Output
Gould-Mosimann School	Multivariate regression of shape on size	Centroid size	Procrustes shape space	Size-corrected residuals
Huxley-Jolicoeur School	Principal component analysis	Embedded in coordinate data	Procrustes form space	Principal components
Multivariate Regression with Cross-Validation	Regression with permutation tests	Centroid size	Shape space	Corrected shapes with performance metrics
Template Registration for Out-of-Sample Data	Procrustes alignment to reference	Centroid size	Shape space	Registered coordinates for new specimens

Performance Metrics and Cross-Validation Results

The evaluation of allometry correction methods requires robust cross-validation approaches, particularly when classifiers are constructed from aligned coordinates [2]. In standard GM practice, data are typically split into training and test sets after joint generalized Procrustes analysis (GPA) of the entire dataset [2]. However, this approach presents challenges for real-world applications where new specimens must be classified without recalculating the overall alignment.

Table 2: Cross-Validation Performance of Allometry Correction Protocols

Methodological Aspect	Performance Consideration	Cross-Validation Challenge	Recommended Solution
Dimensionality Reduction	High-dimensional shape data requires reduction before CVA	Overfitting with too many PC axes; underfitting with too few	Use variable number of PC axes optimized for cross-validation rate [1]
Out-of-Sample Registration	Standard GPA uses entire sample information	New specimens cannot be aligned without reference sample	Template-based registration using representative target [2]
Allometric Correction	Removal of size-effects shapes subsequent analysis	Confounding of different allometry levels (static, ontogenetic, evolutionary)	Study designs that explicitly separate levels of variation [72]
Classifier Performance	Rate of correct assignments depends on alignment	Resubstitution estimates are biased upward	Cross-validation with leave-one-out or training-test splits [1] [2]

Research comparing four mathematical representation approaches for outlines (two semi-landmark methods, elliptical Fourier analysis, and extended eigenshape method) found that classification rates were not highly dependent on the number of points used to represent a curve or the manner of point acquisition [1]. The choice of dimensionality reduction approach proved more significant, with a variable number of principal component axes producing higher cross-validation assignment rates than either fixed PC axes or partial least squares methods [1].

Experimental Protocols for Allometry Assessment

Standardized Workflow for Allometry Analysis

Diagram 1: Workflow for allometry analysis and correction. The pathway highlights both regression-based and PCA-based approaches to allometry correction.

Detailed Methodological Protocols

Multivariate Regression Protocol (Gould-Mosimann Approach)

The multivariate regression of shape on size implements the Gould-Mosimann concept of allometry [72]. This method can be applied to various levels of allometry, including:

• Ontogenetic allometry: Changes associated with growth during individual development
• Static allometry: Variation among individuals at the same developmental stage
• Evolutionary allometry: Differences across species or evolutionary lineages

Experimental Steps:

• Landmark Digitization: Collect 2D or 3D landmark coordinates representing biologically homologous points [73]
• Generalized Procrustes Analysis: Superimpose configurations by translating to a common centroid, scaling to unit centroid size, and rotating to minimize Procrustes distance [73]
• Size Calculation: Compute centroid size as the square root of the sum of squared distances of all landmarks from their centroid [72]
• Multivariate Regression: Perform regression of Procrustes coordinates (shape variables) on centroid size (or log-transformed centroid size)
• Significance Testing: Use permutation tests to evaluate statistical significance of the allometric relationship
• Residual Extraction: Calculate residual shape variables after removing size effects

The extent of allometry is often visualized as a deformation grid or vector displacement diagram showing how shape changes with unit increase in size [72].

Principal Component Analysis Protocol (Huxley-Jolicoeur Approach)

The Huxley-Jolicoeur approach characterizes allometry through principal component analysis in form space [72]. This method does not explicitly separate size and shape but examines covariation patterns among morphological variables.

Experimental Steps:

• Form Space Preparation: Use Procrustes form space (retaining size information) or conformation space (size-and-shape space) [72]
• PCA Implementation: Perform principal component analysis on the covariance matrix of form variables
• Allometric Axis Identification: Identify the first principal component as the primary allometric trajectory when it represents size-related variation [72]
• Cross-Validation: Assess stability of allometric patterns through resampling or cross-validation procedures

This approach is particularly valuable when the distinction between size and shape is ambiguous or when researchers wish to avoid the potential artifacts of Procrustes superimposition [72].

Research Toolkit for Allometry Studies

Table 3: Essential Methodological Components for Allometry Research

Research Component	Function/Purpose	Implementation Considerations
Landmark Coordinates	Capture geometric information	Type I, II, and III landmarks; sliding semi-landmarks for curves
Centroid Size	Isometric size measure	Square root of sum of squared landmark distances from centroid
Procrustes Superimposition	Remove non-shape variation	Generalized Procrustes analysis (GPA) standardizes position, orientation, scale
Thin-Plate Spline	Visualize shape changes	Interpolation function showing deformation between shapes
Multivariate Regression	Quantify shape-size relationship	Procrustes ANOVA; permutation tests for significance
Principal Components	Identify major variation axes	First PC often corresponds to allometric vector in form space
Cross-Validation	Assess method performance	Leave-one-out; k-fold; out-of-sample template registration
Template Registration	Align new specimens	Registration to representative template from reference sample

Discussion and Research Implications

The impact of allometry correction extends across multiple biological disciplines, from evolutionary biology to biomedical applications. In systematic and phylogenetic studies, failure to account for allometric effects can confound evolutionary interpretations, as size-related shape changes may be misattributed to phylogenetic signal [72]. Similarly, in developmental biology, distinguishing allometric growth patterns from other sources of shape variation is essential for understanding ontogenetic trajectories [72].

The choice between allometry correction methods should be guided by research questions and data structure. The Gould-Mosimann approach (multivariate regression) provides a direct test of the relationship between size and shape, with clear biological interpretation [72]. The Huxley-Jolicoeur approach (PCA in form space) may be preferable when researchers wish to avoid potential artifacts of the size-shape separation or when analyzing complex morphological structures without clear size proxies [72].

Recent methodological developments address the challenge of classifying out-of-sample specimens, which is particularly relevant for applied contexts such as nutritional assessment from body shape images [2]. Template-based registration methods enable the projection of new specimens into an established shape space without recalculating the entire Procrustes alignment, facilitating practical applications of allometry-corrected shape analyses [2].

Future methodological development should focus on improving cross-validation performance, particularly for high-dimensional landmark data. The integration of allometry correction with other morphological analyses, such as modularity and integration studies [74], represents another promising direction for advancing geometric morphometric protocols.

In geometric morphometrics, the reliability of downstream analyses is fundamentally constrained by the initial stages of data acquisition and preprocessing. For research focusing on the cross-validation performance of different geometric morphometric protocols, the repeatability of landmark digitization and the quality of input images are not merely preliminary steps but foundational determinants of statistical validity. Variations in these initial stages can introduce technical noise that confounds biological signals, ultimately compromising the discriminant power and generalizability of research findings across scientific domains, from paleontology to drug development [1] [3].

This guide provides a comparative evaluation of methodologies aimed at optimizing these critical preprocessing steps. It examines traditional geometric morphometric techniques against emerging computer vision approaches, focusing on their performance in ensuring data reliability and repeatability, which is essential for building robust predictive models in scientific research.

Comparative Analysis of Method Performance

The choice of methodology for outline analysis and landmark identification significantly impacts the reliability and classification accuracy of morphometric data. The following tables summarize key performance metrics from experimental studies.

Table 1: Comparison of Outline Analysis Methods in Geometric Morphometrics (Based on [1])

Method Category	Specific Method	Key Characteristics	Reported Classification Performance
Semi-Landmark Based	Bending Energy Alignment (BEM)	Incorporates information about curves into landmark-based formalism	Roughly equal classification rates
Semi-Landmark Based	Perpendicular Projection (PP)	Projects points onto a template curve along perpendicular directions	Roughly equal classification rates
Mathematical Function	Elliptical Fourier Analysis (EFA)	Represents outlines using Fourier harmonics	Rates not highly dependent on method details
Mathematical Function	Extended Eigenshape Analysis	Captures major shape variations via principal components analysis	Rates not highly dependent on method details

Table 2: Performance Comparison of Geometric Morphometric vs. Computer Vision Methods (Based on [3])

Method Category	Specific Technique	Application Context	Reported Classification Accuracy
Geometric Morphometric	Outline-based Fourier Analysis	Carnivore tooth mark identification	Low accuracy & resolution
Geometric Morphometric	Semi-landmark Approach	Carnivore tooth mark identification	< 40% discriminant power
Computer Vision	Deep Convolutional Neural Networks	Carnivore tooth mark identification	81% accuracy
Computer Vision	Few-Shot Learning Models	Carnivore tooth mark identification	79.52% accuracy

Table 3: Reliability of 3D Cephalometric Landmarks from CBCT (Based on [75])

Landmark Type	Specific Examples	Reliability Level	Key Considerations
High-Reliability	Points on median sagittal line, Dental landmarks	Highest	Less susceptible to projection and lateral identification errors
Low-Reliability	Condyle, Porion, Orbitale	Lower	Affected by bilateral visualization challenges and complex anatomy
Variable-Reliability	Point S (Sella Turcica)	Context-Dependent	Must be marked in multi-planar views associated with 3D reconstruction

Detailed Experimental Protocols

Protocol 1: Optimizing Outline-Based Classification with CVA

This protocol, derived from a study on ovenbird (Seiurus aurocapilla) tail feathers, details a method for classifying specimens based on outlines using Canonical Variates Analysis (CVA) [1].

• Specimen Digitization: Select a minimum of 60 specimens per group to ensure statistical power. Capture outline data using one of these approaches:
  • Template-based digitization: Define points a priori by a rule (e.g., equal angles between radii of a circle).
  • Manual tracing: Select points by eye while tracing the curve.
  • Automated edge detection: Use software to detect color/brightness differences to delimit the curve.
• Outline Representation and Alignment: Convert captured points into a mathematical representation using one of four alignment or analysis methods:
  • Bending Energy Minimization (BEM)
  • Perpendicular Projection (PP)
  • Elliptical Fourier Analysis (EFA)
  • Extended Eigenshape Analysis
• Dimensionality Reduction via PCA: Use Principal Component Analysis (PCA) to reduce the data dimensionality. Critically, do not use all possible PC axes. Instead, implement a cross-validation protocol to determine the optimal number of PC axes that maximizes the cross-validation assignment rate, thus avoiding overfitting.
• Canonical Variates Analysis and Validation: Perform CVA on the retained PC scores. Use a leave-one-out cross-validation procedure to calculate an unbiased estimate of the correct classification rate, as the resubstitution rate is known to be overly optimistic.

Protocol 2: Validating 3D Cephalometric Landmarks on CBCT

This protocol outlines the steps for establishing a reliable set of 3D cephalometric landmarks from Cone-Beam Computed Tomography (CBCT) scans, crucial for reproducible craniofacial analysis [75].

• Image Acquisition and Calibration: Acquire CBCT scans according to a standardized imaging protocol. Ensure all evaluators are experienced and calibrated in landmark identification to minimize consistent errors. The use of multiplanar reconstruction views (axial, coronal, sagittal) is mandatory.
• Landmark Identification: Identify a predefined set of landmarks comprising the entire maxillo-mandibular complex. The operational definition for each landmark must specify its location in all three planes. Landmarks on the median sagittal plane and dental landmarks generally show higher reliability and should be prioritized.
• Data Collection and Statistical Analysis: Collect coordinate data for all landmarks across multiple trials by the same observer (intra-observer) and different observers (inter-observer). For statistical analysis of reliability, use the Intraclass Correlation Coefficient (ICC) or Bland-Altman tests, which are more appropriate than paired t-tests or simple correlation coefficients.

Protocol 3: Computer Vision for Bone Surface Modification Identification

This protocol describes a modern computer vision approach for classifying carnivore agency from tooth marks on bones, which significantly outperformed traditional geometric morphometric methods in experimental testing [3].

• Dataset Curation and Preprocessing: Assemble a controlled, experimentally-derived set of Bone Surface Modifications (BSM), such as tooth marks generated by different carnivores. This serves as the ground-truthed training and testing dataset.
• Model Selection and Training: Employ a Deep Convolutional Neural Network (DCNN). Train the model on the curated image dataset to learn the features distinguishing the different modifying agents. As an alternative, especially with limited data, a Few-Shot Learning (FSL) model can be explored.
• Model Validation and Application: Validate the trained model on a held-out test set of experimental tooth pits to determine its classification accuracy. For application to the fossil record, exercise extreme caution and only use on well-preserved specimens, as taphonomic processes can alter the original properties of BSM. The authors noted high confidence in interpretations for well-preserved 1.8 Ma tooth marks from Olduvai sites.

Workflow Visualization

The diagram above illustrates two parallel pathways for morphometric analysis. The Traditional GMM Workflow (blue) involves sequential steps of digitization, alignment, and statistical analysis, requiring careful dimensionality reduction to avoid overfitting [1]. In contrast, the Computer Vision Workflow (green) utilizes automated feature extraction and model training, demonstrating superior classification accuracy in experimental comparisons [3]. Both pathways are critically dependent on initial image quality assessment and control (red).

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 4: Key Tools and Software for Image Quality and Landmark Digitization

Tool Name/Type	Primary Function	Application Context
Pulseq & Gadgetron	Open-source, vendor-independent framework for MRI sequence programming and reconstruction.	Harmonizing scanner variability in MRI research [76].
Dolphin 3D Software	Software for 3D cephalometric landmark identification and analysis on CBCT data.	Orthodontic and craniofacial research; shown to have high reliability [75].
DistilIQA	A distilled vision transformer network for no-reference image quality assessment.	Automated quality checking for CT images without a pristine reference [77].
Deep Convolutional Neural Networks (DCNN)	AI model for automated feature learning and image classification.	Classifying bone surface modifications and other morphometric features [3].
Few-Shot Learning (FSL) Models	AI approach that learns from very few examples.	Effective classification in data-scarce scenarios [3].
Elliptical Fourier Analysis	Mathematical method for representing closed outlines using Fourier harmonics.	Outline-based shape analysis in geometric morphometrics [1].

Benchmarking GM Against Machine Learning and Novel Validators

In the field of quantitative shape analysis, researchers and professionals often face a critical choice between traditional Geometric Morphometrics (GM) and modern Convolutional Neural Networks (CNNs). This decision significantly impacts the reliability, interpretability, and practical feasibility of research outcomes across disciplines including biology, archaeology, and medical science. GM offers a mathematically rigorous framework for analyzing homologous structures with strong theoretical foundations, while CNNs provide powerful pattern recognition capabilities that can automatically learn relevant features from raw image data. Understanding the relative strengths, limitations, and cross-validation performance of these methodologies is essential for selecting the appropriate tool for specific research questions and data contexts. This guide provides an objective, evidence-based comparison to inform these methodological decisions, drawing from recent experimental studies across multiple domains.

Methodology: Core Principles and Workflows

Geometric Morphometrics (GM)

GM is a sophisticated approach to shape analysis that preserves geometric relationships throughout the statistical process. The methodology centers on the precise location of homologous landmarks - biologically corresponding points that can be reliably identified across all specimens in a study. The core GM workflow involves:

• Landmark Digitization: Expert identification of Type I (discrete anatomical points), Type II (maxima of curvature), and Type III (extremal points) landmarks.
• Procrustes Superimposition: A multistep process including: (a) Translation to align specimens at the origin; (b) Scaling to unit centroid size; and (c) Rotation to minimize distances between corresponding landmarks [53].
• Shape Variable Extraction: The resulting Procrustes coordinates represent shape variables independent of position, scale, and orientation.
• Statistical Analysis: Multivariate techniques applied to explore shape variation, allometry, and group differences.

A key advantage of GM is its explicit treatment of allometry (shape changes correlated with size). The Procrustes procedure cleanly separates size (represented by centroid size) from shape, allowing researchers to distinguish allometric from non-allometric shape variation - a crucial consideration in taxonomic studies where size differences alone should not define species boundaries [53].

Convolutional Neural Networks (CNNs)

CNNs represent a fundamentally different approach based on deep learning. Rather than requiring pre-specified landmarks, CNNs automatically learn hierarchical feature representations directly from pixel data. Their architecture typically includes:

• Convolutional Layers: Apply learned filters to detect spatial features hierarchically, from simple edges to complex patterns.
• Pooling Layers: Reduce spatial dimensions while retaining important features, providing translation invariance.
• Fully Connected Layers: Integrate extracted features for final classification or regression tasks.

CNNs excel at capturing complex, non-linear patterns without requiring a priori hypotheses about which shape features are diagnostically important. However, this strength comes with a significant need for large training datasets and reduced interpretability compared to GM approaches.

Experimental Performance Comparison

Quantitative Results Across Domains

Table 1: Performance Comparison of GM and CNN Across Multiple Applications

Research Context	GM Performance	CNN Performance	Key Findings
Archaeobotanical Taxon Identification [78]	Moderate classification accuracy with Elliptical Fourier Transforms + LDA	Superior performance; outperformed GM even with small datasets (n=50 per class)	CNN's advantage persisted across barley, olive, date palm, and grapevine seed identification
Carnivore Agency Identification [3]	<40% classification accuracy using outline analysis	81% accuracy with Deep CNN; 79.52% with Few-Shot Learning	GM showed limited discriminant power for tooth mark classification
Taxonomic Discrimination [53]	Effective group discrimination but primarily driven by size variation	Not directly tested	GM achieved better shape discrimination after removing allometric effects

Impact of Training Sample Size

Table 2: Performance Relative to Sample Size in Medical Imaging [79]

Training Sample Size	Handcrafted Features Performance	CNN-Only Performance	Combined Approach
Small Datasets	Superior performance with increased interpretability	Lower performance due to overfitting	Not applicable
Large Datasets	Good performance maintained	Competitive performance achieved	Best performance using both feature types

Cross-Validation and Generalizability

The critical test for any analytical method is its performance on unseen data. In brain MRI classification for Alzheimer's disease, both conventional machine learning and CNN approaches maintained similar performance when applied to external cohorts, though a slight decrease occurred for both methods [80]. This demonstrates that with proper validation, both approaches can generalize, but domain shift remains challenging.

For GM, cross-validation performance is closely tied to appropriate treatment of allometry. When applied to raw measurements without allometric correction, linear morphometric protocols can show misleadingly high discrimination that primarily reflects size differences rather than genuine shape variation [53].

Comparative Workflow Analysis

The fundamental differences between GM and CNN approaches can be visualized through their distinct analytical pathways:

Research Toolkit: Essential Materials and Solutions

Table 3: Essential Research Tools for GM and CNN Implementation

Tool Category	Specific Tools/Solutions	Function/Purpose	Methodology
GM Software	MorphoJ, EVAN Toolbox, R (geomorph package)	Landmark management, Procrustes analysis, statistical shape analysis	GM
CNN Frameworks	TensorFlow, PyTorch, Keras	Deep learning model development and training	CNN
Data Processing	ANTsPy, ImageJ, OpenCV	Image preprocessing, normalization, augmentation	Both
Visualization	R ggplot2, Python Matplotlib, Shape graphics	Results visualization and interpretation	Both
Validation	scikit-learn, custom cross-validation scripts	Performance assessment and generalization testing	Both

Discussion and Research Implications

Comparative Strengths and Limitations

GM strengths lie in its rigorous mathematical foundation and explicit model of biological form. The method provides:

• Interpretability: Direct biological interpretation through landmark configurations
• Allometry Control: Explicit separation of size and shape variation [53]
• Theoretical Foundation: Strong connection to biological homology concepts

CNN strengths manifest in their flexibility and pattern recognition power:

• Automated Feature Extraction: No need for manual landmark identification
• Complex Pattern Detection: Ability to recognize non-intuitive shape patterns
• Performance: Superior accuracy in multiple direct comparisons [78] [3]

Methodological Selection Guidelines

Choosing between GM and CNN depends on multiple research factors:

• Research Question: GM is preferable when testing specific hypotheses about homologous structures, while CNNs excel at pure classification tasks.
• Sample Size: With limited data, GM or handcrafted features generally outperform CNNs [79].
• Interpretability Needs: GM provides more straightforward biological interpretation.
• Computational Resources: CNNs typically require greater computational capacity for training.

Emerging Trends and Hybrid Approaches

The most promising future direction may involve hybrid methodologies that leverage the strengths of both approaches. For instance, GM can inform CNN architecture design, or CNN-derived features can be incorporated into morphometric frameworks. In genomic research, hybrid CNN-Transformer models have shown superiority for causal variant prioritization, suggesting similar potential in shape analysis [81]. As demonstrated in medical imaging, combining handcrafted features with learned CNN features can yield superior performance to either approach alone [79].

Both Geometric Morphometrics and Convolutional Neural Networks offer powerful, complementary approaches to shape analysis. GM provides a theoretically grounded, interpretable framework ideal for hypothesis-driven research with limited samples, particularly when biological homology and allometry are central concerns. CNNs offer superior predictive accuracy for classification tasks with sufficient training data, automatically discovering discriminative patterns without requiring expert landmark specification. The choice between methodologies should be guided by research objectives, sample size constraints, interpretability requirements, and available computational resources. Future methodological development will likely focus on hybrid approaches that leverage the respective strengths of both paradigms while addressing their individual limitations through integrated analytical frameworks.

Supervised Machine Learning as a More Accurate Classifier and Novel Taxon Detector

Within the field of geometric morphometrics, the transition from traditional measurement-based analyses to sophisticated computational approaches represents a significant methodological evolution. This guide objectively compares the performance of supervised machine learning (ML) classifiers against traditional methods and other algorithmic approaches for taxonomic classification and discovery. Framed within a broader thesis on cross-validation performance of different geometric morphometric protocols, we present empirical data demonstrating that supervised ML models, particularly ensemble methods like Random Forest, achieve superior accuracy in species discrimination and offer robust capabilities for detecting novel taxa. The following sections provide a detailed comparison of classifier performance, detailed experimental methodologies, and essential resources for implementing these advanced analytical techniques in biological research.

Performance Comparison of Classification Methods

Classifier Accuracy Across Biological Domains

Table 1: Performance comparison of machine learning classifiers versus traditional methods in taxonomic classification

Classification Method	Application Context	Key Performance Metrics	Reference Study
Random Forest (RF)	Sex estimation from 3D tooth landmarks	Accuracy: 97.95% (mandibular second premolars), 95.83% (maxillary first molars); Balanced precision/recall [82]	Geometric morphometric analysis of dental casts
Support Vector Machine (SVM)	Sex estimation from 3D tooth landmarks	Accuracy: 70-88%; Moderate performance [82]	Geometric morphometric analysis of dental casts
Artificial Neural Network (ANN)	Sex estimation from 3D tooth landmarks	Accuracy: 58-70%; Lowest metrics; Struggled with female classification [82]	Geometric morphometric analysis of dental casts
Geometric Morphometrics	Bat species discrimination based on wing morphology	Improved species discrimination compared to traditional methods; Revealed evolutionary allometry patterns [83]	Wing, body, and tail morphology of European horseshoe bats
Traditional Morphometrics	Bat species discrimination based on external morphology	Lower discrimination power for closely related species compared to geometric morphometrics [83]	Linear measurements and ratios of bat wings
Database (DB) Methods	Taxonomic classification of sequencing data	Higher accuracy with comprehensive reference databases; Performance constrained by database quality/scope [84]	Bioinformatics analysis of sequencing data
Machine Learning (ML) Methods	Taxonomic classification of sequencing data	Superior with sparse reference data; Can extrapolate unknown species; Performance limited by training data representativeness [84]	Bioinformatics analysis of sequencing data
Convolutional Neural Networks (CNN)	Carnivore tooth mark identification	81% classification accuracy; Effective in well-preserved contexts [3]	Analysis of bone surface modifications

Cross-Validation Performance Insights

Across multiple biological domains, supervised ML classifiers consistently demonstrate superior performance in geometric morphometric analyses when evaluated through rigorous cross-validation protocols. In direct comparisons, Random Forest outperformed both SVM and ANN models in sex classification from 3D dental landmarks, achieving remarkable accuracy up to 97.95% with minimal sex bias [82]. This performance advantage is attributed to RF's ability to handle tabular data and high-dimensional feature spaces effectively, capturing complex spatial relationships between landmarks that simpler models might miss.

The comparison between database-based and ML methods for sequence classification reveals a crucial trade-off: while DB methods excel when comprehensive reference databases exist, ML approaches show superior performance in scenarios where reference sequences are sparse or lacking, as they can extrapolate the existence of unknown species from training data [84]. This capability makes ML particularly valuable for novel taxon detection in exploratory research.

Experimental Protocols for Method Validation

Geometric Morphometric Analysis with Supervised ML

Protocol 1: Landmark-Based Classification with Multiple Algorithms

A comprehensive protocol for evaluating classifier performance using 3D geometric morphometric data was established in forensic odontology research [82]:

• Sample Preparation and Digitization: Dental casts from 120 individuals (60 males, 60 females) were digitized using a 3D scanner (Dentsply Sirona inEOS X5). Inclusion criteria specified ages 13-20 to prevent tooth changes from occlusal wear.

• Landmark Identification: Anatomic and geometric landmarks were identified on nine tooth types using 3D Slicer software (version 4.10.2). The number of landmarks varied based on tooth complexity (19-32 landmarks per tooth).

• Data Preprocessing: Landmark coordinates underwent Procrustes superimposition and principal component analysis using MorphoJ software (version 1.07a) to normalize size and orientation variation.

• Classifier Training: Three ML algorithms (ANN, SVM, RF) were trained on the pre-processed landmark data using fivefold cross-validation to prevent overfitting.

• Performance Evaluation: Models were evaluated using accuracy, precision, recall, F1-score, and AUC metrics. Feature analysis was conducted to identify the most dimorphic dental elements.

This protocol revealed that maxillary first molars and mandibular second premolars exhibited the highest sexual dimorphism, with RF consistently achieving the most robust classification across all tooth types [82].

Mock Community Validation for Taxonomic Classifiers

Protocol 2: Benchmarking with Mock Communities

An extensible framework for evaluating taxonomy classification accuracy was developed using mock communities [85]:

• Community Construction: 15 bacterial 16S rRNA gene mock communities and 4 fungal ITS mock communities were sourced from mockrobiota, a public repository for mock community data.

• Reference Database Preparation: Greengenes 99% OTUs 16S rRNA gene and UNITE 99% OTUs ITS reference sequences were used for bacterial and fungal classifications, respectively.

• Classifier Optimization: Parameter sweeps were conducted to determine optimal configurations for multiple methods (RDP, BLAST, UCLUST, SortMeRNA, naive Bayes).

• Performance Assessment: Classification accuracy was evaluated at taxonomic levels from class through species using F-measure, recall, taxon detection rate, and Bray-Curtis dissimilarity metrics.

• Class Weight Evaluation: The impact of setting class weights (bespoke vs. uniform) on classification accuracy was tested, with bespoke weights reflecting known taxonomic compositions.

This validation approach demonstrated that naive Bayes with bespoke class weights achieved significantly higher F-measure, recall, and taxon detection rate than all other methods, highlighting the importance of incorporating prior knowledge about expected community composition [85].

Novel Taxon Detection with Genome Similarity Metrics

Protocol 3: Detecting Higher-Level Taxonomic Divergence

For delineating novel microbial taxa above genus level, a neural network-based approach was developed using multiple genome similarity metrics [86]:

• Data Curation: 14,390 non-redundant marine prokaryotic metagenome-assembled genomes (MAGs) were collected from 106 metagenomic surveys with completeness >80% and contamination <5%.

• Feature Calculation: Similarity metrics between genome pairs were computed, including Average Amino Acid Identity (AAI), Average Nucleotide Identity (ANI), and Fractions of Shared Genes (FSG) within 26 KEGG gene categories.

• Model Architecture: Neural network classifiers were trained at each taxonomic level (genus to phylum) to predict whether any two MAGs belong to the same taxon.

• Predictor Selection: Optimal subsets of predictors and neural network hyperparameters were selected by maximizing balanced accuracy during 10-fold cross-validation.

• Taxon Delineation: Pairwise classifications between MAGs were used as inputs to clustering algorithms to reconstruct taxonomic relationships de novo, including undefined taxa.

This protocol achieved balanced accuracy exceeding 92% at all taxonomic levels, identifying gene categories involved in metabolism of cofactors and vitamins as particularly correlated to taxon divergence [86].

Workflow Visualization

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential materials and software for geometric morphometric ML implementations

Tool/Resource	Function	Application Context	Key Features
3D Slicer Software	Landmark identification on 3D models	Geometric morphometric analysis [82]	Open-source; Extensive module ecosystem; Supports 3D data visualization
MorphoJ	Geometric morphometric data analysis	Shape variation and classification [82]	Procrustes superimposition; Principal component analysis; Discriminant function analysis
QIIME 2 with q2-feature-classifier	Taxonomic classification of marker-gene sequences	Microbiome analysis [85]	Multiple classification methods; Integration with scikit-learn; Mock community validation
HusMorph	Automated landmark placement	High-throughput phenotyping [87]	User-friendly GUI; Automated parameter optimization; Cross-platform compatibility
GTDB-Tk	Taxonomic classification of genomes	Prokaryotic taxonomy [86]	Genome Taxonomy Database standard; Consistent classification; Updated reference tree
CheckM2	Quality assessment of metagenome-assembled genomes	Genome quality control [86]	Completeness/contamination estimates; Universal single-copy genes
Dlib & OpenCV	Machine learning and computer vision	Automated landmark prediction [87]	Facial landmark detection; Shape prediction; Image processing
scikit-learn	Machine learning in Python	Classifier implementation [85]	Random Forest, SVM, ANN algorithms; Model evaluation tools

The comprehensive performance comparison and experimental data presented in this guide demonstrate that supervised machine learning, particularly Random Forest algorithms, provides significantly more accurate classification in geometric morphometric analyses compared to traditional methods and other ML approaches. When evaluated through rigorous cross-validation protocols, these classifiers not only excel at discriminating known taxa but also show strong capability for novel taxon detection, especially in scenarios with sparse reference data. The implementation protocols and research tools detailed herein provide a robust framework for researchers seeking to incorporate these advanced analytical techniques into their taxonomic and morphometric studies, ultimately enhancing objectivity, accuracy, and discovery potential in biological classification.

Functional Data Analysis (FDA) as a Validator for Traditional GM Techniques

Geometric morphometrics (GM) has established itself as a fundamental discipline for the quantitative analysis of shape variation in biological research, employing landmarks to capture morphological information in a geometric framework [9]. While GM techniques, particularly those based on Generalized Procrustes Analysis (GPA), provide powerful tools for shape analysis, they face inherent limitations in capturing complex morphological variations and are susceptible to observer bias during manual landmark placement [88]. This methodological comparison examines how Functional Data Analysis (FDA)—a statistical framework that treats data as continuous functions rather than discrete points—serves as both a complementary validator and enhancer of traditional GM protocols. By evaluating cross-validation performance across multiple biological classification tasks, we demonstrate how FDA principles address fundamental limitations in GM while providing robust validation of morphological hypotheses.

The integration of FDA with GM represents a paradigm shift from discrete point analysis to continuous shape representation. Traditional GM reduces complex biological shapes to limited sets of landmarks, potentially overlooking meaningful morphological information between landmarks [9]. In contrast, FDA frameworks model entire curves and surfaces as functional entities, preserving subtle morphological patterns through sophisticated mathematical representations. This comparison guide objectively evaluates the performance of both methodologies across key metrics including classification accuracy, robustness to variation, and computational efficiency, providing researchers with evidence-based guidance for methodological selection in morphological studies.

Theoretical Foundations and Methodological Comparison

Geometric Morphometrics: Traditional Framework

Traditional GM operates within a well-established analytical pipeline beginning with the digitization of homologous landmarks—discrete anatomical points that hold biological correspondence across specimens [88]. The foundational step of Generalized Procrustes Analysis (GPA) removes non-shape variation including position, orientation, and scale through superimposition algorithms, yielding Procrustes coordinates that represent shape variables for subsequent multivariate analysis [89] [24]. This approach preserves geometric relationships throughout analysis and enables visualization of shape changes along statistical axes. However, GM faces constraints including the necessary a priori selection of landmarks, which requires expert knowledge and may introduce observer bias while potentially missing morphological information between landmarks [88].

Recent innovations have sought to address these limitations through semi-landmarks and outline-based methods that capture curvature information [1]. These approaches increase the density of shape information but introduce additional analytical challenges including parameterization choices and the need for sliding protocols to minimize arbitrary geometric effects. The discrete nature of GM data further complicates analysis of complex morphological structures without clear homologous points, limiting its application for comprehensive shape quantification, particularly in taxonomic classification problems where subtle shape differences are diagnostically meaningful [9].

Functional Data Analysis: Conceptual Framework

Functional Data Analysis reconceptualizes morphological analysis by treating shape data as continuous functions rather than discrete points [9] [89]. This paradigm shift enables researchers to model biological shapes as smooth curves or surfaces defined by mathematical functions, typically represented using basis function expansions such as B-splines or Fourier components. The FDA framework operates on several key principles: (1) shape representation through continuous functions, (2) separation of amplitude (shape) and phase (timing/parameterization) variation, and (3) statistical analysis in functional spaces [89].

Advanced FDA implementations incorporate sophisticated mathematical tools including square-root velocity function (SRVF) frameworks that leverage the Fisher-Rao Riemannian metric to separate amplitude and phase variation, effectively aligning curves to a common template [89]. Arc-length parameterization provides another critical FDA tool, enabling consistent assessment of complex-shaped signals by eliminating variability due to uneven sampling [89]. For three-dimensional data, multivariate functional principal component analysis (MFPCA) extends landmark trajectories to multi-dimensional functional data, capturing correlated variation across dimensions [89]. These mathematical foundations enable FDA to address fundamental GM limitations, particularly for analyzing complex biological shapes with subtle but biologically meaningful variations.

Comparative Experimental Performance

Classification Accuracy Across Biological Models

Table 1: Cross-Validation Classification Performance Across Methodologies

Biological Model	Traditional GM	FDA Approach	Performance Difference	Statistical Significance
Shrew Craniodental Classification [9]	85.2%	92.6%	+7.4%	p < 0.05
Kangaroo Cranial Dietary Classification [89]	78.5%	87.3%	+8.8%	p < 0.01
Early Knee Osteoarthritis Detection [90]	81.7%	89.4%	+7.7%	p < 0.05
Severe Acute Malnutrition Assessment [24]	83.3%	90.1%	+6.8%	p < 0.05

Experimental evidence across multiple biological systems demonstrates consistently superior classification performance for FDA-based approaches compared to traditional GM protocols. In craniodental classification of three shrew species (S. murinus, C. monticola, and C. malayana) from Peninsular Malaysia, FDA achieved 92.6% classification accuracy compared to 85.2% for traditional GM—a statistically significant 7.4% improvement [9]. Similarly, in classifying kangaroo crania according to dietary categories (omnivores, mixed feeders, browsers, and grazers), FDA pipelines outperformed GM by 8.8% in cross-validation accuracy [89]. This pattern of enhanced performance extends to clinical applications, with FDA-based Functional Logistic Regression improving early knee osteoarthritis detection by 7.7% compared to GM-derived models [90].

The performance advantage of FDA approaches appears most pronounced in systems with complex shape variations and subtle morphological differences. For shrew classification, the dorsal craniodental view provided optimal discrimination, with FDA particularly effective at capturing subtle cranial curvature differences between species [9]. Similarly, in kangaroo cranial analysis, FDA's ability to model entire surfaces rather than discrete landmarks enabled more sensitive detection of dietary adaptation signatures [89]. These consistent performance improvements across diverse biological systems suggest FDA provides genuine methodological advantages for morphological classification tasks.

Analytical Robustness and Information Capture

Table 2: Analytical Characteristics Comparison Between GM and FDA

Analytical Characteristic	Traditional GM	FDA Approach	Biological Implication
Shape Representation	Discrete landmarks	Continuous curves/surfaces	FDA captures interstitial morphology
Data Reduction Required	High	Minimal	FDA preserves subtle shape features
Observer Bias	Potentially high	Minimal	FDA reduces subjective landmark placement
Alignment Method	Procrustes superimposition	Functional alignment/curve registration	FDA better handles non-rigid deformation
Complex Shape Capture	Limited by landmark number	Comprehensive	FDA superior for structures without clear landmarks
Statistical Power	Moderate	High	FDA detects subtler shape differences

Beyond raw classification accuracy, FDA demonstrates superior analytical robustness across multiple dimensions. Traditional GM requires substantial data reduction, representing complex biological shapes with limited landmark sets—typically tens to hundreds of points [88]. This discrete approach inevitably discards morphologically significant information between landmarks and introduces observer bias during landmark placement [88]. In contrast, FDA captures comprehensive shape information by modeling entire curves and surfaces as functional entities, significantly reducing information loss [9] [89].

The functional logistic regression (FLR) model applied to early knee osteoarthritis detection exemplifies FDA's analytical advantages [90]. By incorporating entire ground reaction force curves as functional predictors alongside clinical variables, FLR achieved superior sensitivity in detecting subtle biomechanical alterations while maintaining statistical interpretability. This integrated approach outperformed both traditional GM-derived models and black-box machine learning methods, demonstrating FDA's optimal balance between analytical precision and biological interpretability. Similar advantages were evident in craniodental morphology, where FDA's continuous shape representation captured subtle species-specific variations missed by landmark-based GM [9].

Methodological Protocols

Standard Geometric Morphometrics Workflow

Traditional GM analysis follows a well-established pipeline beginning with specimen preparation and image acquisition. The foundational step involves digitization of homologous landmarks—anatomically corresponding points across specimens—using specialized software such as MorphoJ or tpsDig [24]. For complex curves, semi-landmarks are often added to capture outline information, requiring subsequent sliding procedures to minimize arbitrary geometric effects [1]. The core analytical step involves Generalized Procrustes Analysis (GPA), which superimposes landmark configurations via translation, rotation, and scaling to remove non-shape variation [89] [24].

Following GPA, the resulting Procrustes coordinates undergo multivariate statistical analysis, typically principal component analysis (PCA) to visualize major shape variation axes, followed by discriminant analysis for classification tasks [1]. Critical considerations include landmark repeatability assessment through intra- and inter-observer error studies, and appropriate dimension reduction to avoid overfitting in discriminant analysis [1]. Cross-validation protocols typically employ leave-one-out or k-fold approaches on the Procrustes coordinates, though application to new specimens requires complete reanalysis or reference to a fixed template [24].

Graphical Abstract: Traditional Geometric Morphometrics Workflow

Functional Data Analysis Workflow

FDA morphological analysis begins with comparable specimen preparation but employs fundamentally different data capture approaches. Rather than discrete landmarking, FDA utilizes dense point clouds or outline coordinates, often obtained through automated surface scanning or edge detection algorithms [9] [89]. The critical transformation involves converting discrete coordinates to functional data through basis function expansions, typically using B-splines or Fourier basis systems, with smoothing parameters optimized to capture biological signal while reducing high-frequency noise [89].

For shape analysis, FDA implementations often employ curve registration techniques to separate amplitude (shape) and phase (parameterization) variation, with advanced approaches utilizing square-root velocity function (SRVF) frameworks for optimal alignment [89]. Functional principal component analysis (FPCA) then identifies major modes of shape variation in the functional space, with subsequent classification using functional discriminant analysis or functional logistic regression [90]. Cross-validation follows similar principles to GM but operates in the functional domain, with the significant advantage that new specimens can be projected into existing functional spaces without complete reanalysis [89].

Graphical Abstract: Functional Data Analysis Workflow

Research Toolkit and Implementation

Essential Research Reagents and Computational Solutions

Table 3: Essential Research Toolkit for GM and FDA Applications

Tool/Category	Specific Examples	Function/Purpose	Methodological Application
Landmarking Software	tpsDig, MorphoJ	Manual landmark digitization	Traditional GM data capture
Surface Scanning	Micro-CT scanners, 3D photogrammetry	High-resolution surface acquisition	FDA point cloud generation
Functional Analysis Packages	fda R package, MATLAB FDA toolbox	Basis function expansion & functional PCA	FDA implementation
Shape Analysis Platforms	geomorph R package, EVAN Toolbox	Procrustes analysis & shape statistics	Traditional GM analysis
Alignment Algorithms	Procrustes superimposition, SRVF alignment	Shape registration & normalization	Both GM and FDA
Classification Tools	LDA, SVM, Functional Logistic Regression	Group discrimination & prediction	Performance validation

Successful implementation of GM and FDA methodologies requires specialized computational tools and analytical packages. For traditional GM, established software suites including tps series (tpsDig, tpsRelw) and MorphoJ provide comprehensive landmark management and Procrustes-based analysis [24] [1]. The geomorph R package offers advanced GM capabilities including modularity integration and phylogenetic comparative methods. For FDA implementation, the fda R package provides core functionality for basis function expansion, smoothing, and functional principal component analysis, while specialized MATLAB toolboxes offer additional FDA algorithms [89].

Emerging hybrid approaches leverage strengths from both methodologies. The morphVQ pipeline automates morphological phenotyping using learned shape descriptors and functional maps, capturing comprehensive shape variation while avoiding manual landmarking limitations [88]. Similarly, Functional Data Geometric Morphometrics (FDGM) integrates FDA principles with GM frameworks, converting landmark data into continuous curves for more sensitive shape discrimination [9]. These hybrid approaches demonstrate the evolving synergy between methodological traditions, offering enhanced performance while maintaining biological interpretability.

Discussion and Research Implications

Validation Performance and Methodological Integration

The consistent superiority of FDA approaches in cross-validation performance across multiple biological systems establishes FDA as a robust validator for traditional GM techniques. The 6.8-8.8% improvement in classification accuracy observed across shrew, kangaroo, and clinical datasets demonstrates FDA's enhanced sensitivity to morphologically meaningful shape variation [9] [89] [90]. This performance advantage appears most pronounced in systems characterized by subtle shape differences or continuous morphological gradients, where FDA's capacity to model interstitial curvature provides critical discriminative information.

Beyond validation, FDA addresses fundamental GM limitations including landmark dependency and limited shape capture [88]. By modeling entire curves and surfaces as functional entities, FDA eliminates the arbitrary reduction of complex biological forms to discrete points, thereby reducing analytical bias and capturing more comprehensive morphological information. The functional logistic regression framework exemplifies this advantage, enabling direct incorporation of continuous biomechanical signals as predictors without discretization, thereby preserving critical morphological information [90]. This approach demonstrates significantly improved classification performance while maintaining statistical interpretability—a critical advantage over black-box machine learning alternatives.

Future Directions and Implementation Guidelines

The integration of FDA principles with traditional GM represents a promising direction for methodological advancement in morphological research. Hybrid pipelines such as Functional Data Morphometrics (FDM) and morphVQ demonstrate how functional concepts can enhance GM frameworks without completely abandoning established landmarks [9] [88]. These approaches maintain the biological homology foundation of GM while incorporating FDA's sensitivity to continuous shape variation, offering a balanced solution for complex morphological analysis.

For researchers selecting methodological approaches, we recommend traditional GM for studies focused on specific homologous structures with clearly definable landmarks, particularly when biological interpretability and visualization are priorities [24]. FDA approaches are preferable for analyzing complex shapes without clear landmarks, subtle shape differences challenging discrete landmark detection, and high-resolution surface data where comprehensive shape capture is essential [9] [89]. For maximum analytical robustness, sequential application of both methodologies provides independent validation of morphological hypotheses, with disagreement indicating potential methodological artifacts requiring further investigation.

As morphological datasets increase in complexity and scale, FDA approaches offer scalable solutions that balance statistical precision with biological interpretability. The continued development of automated FDA pipelines will further enhance accessibility for non-specialist researchers, strengthening morphological analysis across biological and clinical domains.

Evaluating the performance of a classification model is a fundamental step in machine learning and scientific research. While a single metric like classification accuracy might seem like a straightforward measure of model quality, it often provides an incomplete and potentially misleading picture, especially for imbalanced datasets or when different types of classification errors carry different consequences [91] [92]. A robust evaluation framework requires multiple complementary metrics that collectively provide insights into different aspects of model performance.

This challenge is particularly relevant in geometric morphometrics, where classification models are increasingly used to distinguish between biological groups based on shape variations [24] [12] [61]. In these scientific applications, the choice of evaluation metrics directly impacts the interpretation of results and the validity of biological conclusions. Researchers must therefore understand not only how to calculate these metrics but also how to interpret them within their specific research context and how to properly compare different models using statistically sound methodologies [93] [94].

Core Classification Metrics and Their Interpretation

The Confusion Matrix and Derived Metrics

The confusion matrix forms the foundation for most classification metrics by tabulating the relationship between actual and predicted classes. From this matrix, several key metrics are derived [92]:

• True Positive (TP): Correctly predicted positive instances
• True Negative (TN): Correctly predicted negative instances
• False Positive (FP): Negative instances incorrectly predicted as positive (Type I error)
• False Negative (FN): Positive instances incorrectly predicted as negative (Type II error)

These four fundamental counts give rise to the most commonly used classification metrics, each providing a different perspective on model performance.

Key Metrics and Their Applications

Table 1: Essential Classification Metrics and Their Characteristics

Metric	Formula	Interpretation	Optimal Use Cases
Accuracy	(TP+TN)/(TP+TN+FP+FN)	Overall correctness of predictions	Balanced class distributions; all errors have equal cost [91] [92]
Precision	TP/(TP+FP)	Proportion of positive predictions that are correct	When false positives are costly (e.g., spam detection) [91] [92]
Recall (Sensitivity)	TP/(TP+FN)	Proportion of actual positives correctly identified	When false negatives are critical (e.g., disease diagnosis) [91] [92]
F1 Score	2×(Precision×Recall)/(Precision+Recall)	Harmonic mean of precision and recall	Balanced view of both metrics; class-imbalanced data [92]
Specificity	TN/(TN+FP)	Proportion of actual negatives correctly identified	When correctly identifying negatives is important [92]

Each metric serves different research needs. For example, in a geometric morphometrics study aimed at identifying early-stage pregnancy in killer whales from aerial imagery, recall would be crucial to minimize missed detections of pregnant individuals, while in a study classifying rodent species based on skeletal morphology, precision might be more important to ensure correct species identification [12] [61].

Beyond Binary Classification: Top-K Metrics

In multi-class classification problems, particularly those with many possible classes, top-k accuracy metrics provide a more nuanced evaluation. The top-1 accuracy represents the conventional accuracy metric where the model's highest probability prediction must match the correct class. In contrast, top-5 accuracy considers a prediction correct if the true class is among the model's five highest probability predictions [95].

This approach is particularly valuable when multiple plausible answers exist or when the distinction between similar classes is subtle. In geometric morphometric applications, such as distinguishing between closely related species or different phenotypic variations, top-5 metrics can provide insights into whether models confuse morphologically similar groups while still correctly identifying the general morphological pattern [65] [61].

Statistical Comparison of Classification Models

The Pitfalls of Naive Model Comparison

Comparing classification models based solely on average performance metrics from cross-validation folds without proper statistical testing is a common but flawed practice. Simply highlighting the method with the best average accuracy in "bolded tables" or comparing "dynamite plots" with error bars representing standard deviation fails to account for the statistical variability inherent in cross-validation procedures [93].

Statistical variability in cross-validation-based comparisons arises from multiple factors, including the number of folds, repetitions, dataset characteristics, and the inherent dependencies between cross-validation folds. These factors can significantly impact conclusions about model superiority if not properly accounted for [94]. One critical issue is that the overlapping training folds between different cross-validation runs create implicit dependencies in accuracy scores, violating the assumption of sample independence required by many standard statistical tests [94].

Recommended Statistical Approaches

Proper model comparison requires statistical tests specifically designed to handle the dependencies and distributions of performance metrics from cross-validation. For comparing two models, the Wilcoxon signed-rank test (non-parametric) is generally preferred over the paired t-test, as it makes fewer assumptions about the distribution of the metric scores [93].

When comparing multiple models, Friedman's test provides a non-parametric alternative to ANOVA for determining whether statistically significant differences exist between methods. This test operates by rank-ordering the performance of all models within each cross-validation fold, then comparing the average ranks across folds [93]. If Friedman's test detects significant differences, post-hoc tests with appropriate corrections (such as Bonferroni correction) should be applied to control the family-wise error rate when performing multiple pairwise comparisons [93].

Table 2: Statistical Tests for Comparing Classification Models

Test	Data Type	Comparison Scope	Key Assumptions	Advantages
Paired t-test	Parametric	Two models	Normal distribution of differences; independence	High power when assumptions met [93]
Wilcoxon Signed-Rank	Non-parametric	Two models	Symmetric distribution of differences	Fewer assumptions; robust to outliers [93]
Friedman's Test	Non-parametric	Multiple models	None regarding distribution	Appropriate for cross-validation results [93]

The workflow diagram below illustrates a statistically sound approach for comparing classification models:

Geometric Morphometrics Case Studies

Nutritional Status Classification in Children

The SAM Photo Diagnosis App Program exemplifies the application of geometric morphometrics for classification in a public health context. The program aims to develop a smartphone application for identifying severe acute malnutrition (SAM) in children aged 6-59 months from images of their left arms. The approach uses landmark-based geometric morphometric techniques to capture both size and shape information, providing a more nuanced understanding of how nutritional status influences body morphology compared to traditional anthropometric measurements [24].

This research highlights the challenge of out-of-sample classification in geometric morphometrics. While classifiers are typically built from aligned coordinates of a reference sample using Generalized Procrustes Analysis (GPA), classifying new individuals not included in the original alignment requires specialized methodologies to obtain comparable shape coordinates [24]. The performance metrics used to evaluate such models must be carefully selected to ensure real-world applicability, with particular attention to recall (to minimize missed cases of malnutrition) while maintaining sufficient precision (to avoid overtaxing healthcare resources with false alarms) [24] [91].

Reproductive Status Detection in Killer Whales

In a study detecting reproductive stages of free-ranging killer whales using drone-based aerial imagery, geometric morphometrics provided a protocol for distinguishing between non-pregnant, early-stage pregnant, late-stage pregnant, and lactating individuals. The researchers used Procrustes ANOVA and Discriminant Function Analysis (DFA) to demonstrate significant separation of shape files related to reproductive status [12].

This application achieved reliable detection of early-stage pregnancy, which had been nearly impossible to identify using traditional width-based measurements. The performance of their classification approach was validated through statistical testing of shape differences between reproductive classes, with cross-validation used to assess the robustness of the discrimination [12]. The success of this methodology highlights how geometric morphometric classification can address critical conservation challenges by enabling the quantification of miscarriage rates and reproductive failures in vulnerable populations.

Wing Shape Variation in Insect Populations

Geometric morphometrics has also been applied to classify population origins of Bactrocera invadens fruit flies based on wing vein patterns across different agro-ecological zones in Ghana. Researchers used landmarks representing the junctions of wing veins to quantify shape variations, followed by Procrustes ANOVA, Partial Least Squares (PLS), and multivariate statistical analyses including discriminant analysis with cross-validation [65].

The study revealed significant wing shape variations among populations from different ecological zones, potentially reflecting local adaptations to environmental conditions. The classification performance in this context provided insights into population structure and has implications for pest control strategies [65]. This application demonstrates how performance metrics for geometric morphometric classifiers can address ecological and agricultural questions beyond pure species identification.

Essential Research Toolkit for Geometric Morphometrics Classification

Table 3: Essential Research Tools for Geometric Morphometrics Classification Studies

Tool Category	Specific Tools/Solutions	Function in Classification Pipeline
Landmark Digitization	tpsDig, tpsUtil [61]	Capture landmark coordinates from specimen images
Shape Analysis	MorphoJ [61]	Procrustes alignment, shape variable extraction
Multivariate Statistics	R, Python (scikit-learn)	Principal Component Analysis, Discriminant Analysis
Machine Learning Frameworks	Scikit-learn, LightGBM, PyTorch [93] [96]	Implementation of classification algorithms
Model Evaluation	Custom scripts implementing statistical tests [93] [94]	Cross-validation, metric calculation, significance testing

The toolkit for geometric morphometrics classification spans specialized morphometrics software for shape analysis and general-purpose machine learning frameworks for model building. The integration between these domains is essential for implementing a complete classification pipeline from raw images to validated model performance [61] [96].

Interpreting classification accuracy and error rates requires moving beyond single metrics to embrace a multi-faceted evaluation approach. In geometric morphometrics research, this involves selecting metrics aligned with research objectives, employing statistically sound model comparison methods, and understanding the practical implications of different types of classification errors.

The case studies across biological anthropology, conservation biology, and entomology demonstrate how performance metric interpretation must be contextualized within specific research goals. Whether prioritizing recall for public health screening programs or balancing precision and recall for ecological monitoring, the choice of evaluation criteria directly influences the scientific utility and practical impact of geometric morphometric classification models.

Future directions in this field will likely include greater emphasis on effect sizes alongside statistical significance, standardized reporting guidelines for model performance, and continued development of methods for out-of-sample classification that maintain the statistical rigor of geometric morphometric approaches.

Assessing Reproducibility and Robustness Across Different Datasets and Operators

Geometric morphometrics (GM) has established itself as a cornerstone of modern shape analysis across biological, anthropological, and archaeological sciences. By quantifying shape using Cartesian coordinate configurations of anatomical landmarks, GM enables sophisticated statistical exploration of morphological variation. The foundational step of Procrustes superimposition aligns these configurations to a common coordinate system by removing differences in location, orientation, and scale, isolating shape variation for subsequent analysis [41]. Despite its widespread adoption and analytical power, the reproducibility of GM findings across different datasets, operators, and methodological approaches remains a significant concern, particularly in an era of increasing data sharing and collaborative research. This guide objectively compares the performance of various geometric morphometric protocols, focusing specifically on their robustness to operator-induced bias and methodological variability. We synthesize experimental data from recent studies to provide evidence-based recommendations for researchers seeking to implement reproducible morphometric workflows in evolutionary biology, taxonomy, and related fields.

Multiple studies have systematically quantified the magnitude of error introduced at different stages of geometric morphometric data acquisition and analysis. The table below summarizes key findings on the relative impact of various error sources on shape measurement and statistical classification.

Table 1: Magnitude and Impact of Different Error Sources in Geometric Morphometrics

Error Source	Error Type	Reported Magnitude	Impact on Statistical Results	Key Findings
Inter-operator Variation	Personal	Up to 30-34% of total shape variance [97]	Dominates biological signal in large datasets; affects group membership predictions	Largest single source of error; can surpass sex differences in large samples [97]
Specimen Presentation (2D)	Methodological	>30% of total variation [14]	Greatest impact on species classification accuracy [14]	Projection distortion particularly problematic for non-standardized orientations
Imaging Devices	Instrumental	Substantial, but typically less than inter-operator error [14]	Affects landmark precision and coordinate values	Variation within and between equipment types; lens distortion varies by type [14]
Intra-observer Variation	Personal	Significant but generally less than inter-operator [14]	Affects replicability of landmark configurations	Influenced by digitizing experience and landmark clarity [14]

The data reveal that inter-operator differences constitute the most substantial threat to reproducibility, accounting for up to 34% of total shape variation in some studies—a magnitude sufficient to dominate biological signals in large datasets [97]. This finding has profound implications for collaborative research integrating data from multiple laboratories.

Experimental Protocols for Assessing Reproducibility

Quantifying Inter-Operator Bias in 3D Landmarks

A comprehensive study evaluated inter-operator error using 3D anatomical landmarks from adult human head MRIs. Three operators digitized the same set of landmarks on identical MRI images, enabling direct comparison of their landmark placements [97].

Methodology:

• Sample: Head MRIs from 900+ adult individuals with replicated digitization by three operators
• Landmarks: Configuration of soft and hard tissue 3D facial landmarks
• Analysis: Procrustes-based geometric morphometrics with separate assessment of absolute error (measurement differences) and relative error (differences relative to biological variation)
• Statistical Tests: Procrustes ANOVA to partition variance components, with shape variation decomposed into biological signal and operator-induced error

This protocol revealed that while absolute error was within expected ranges for MRI measurements, the relative error for shape was substantial, with operator differences accounting for up to one-third of total sample variation [97].

A separate study employed a comprehensive approach to evaluate four distinct error sources in 2D landmark coordinate configurations of vole teeth [14].

Methodology:

• Sample: Lower first molars of five North American Microtus (vole) species
• Error Sources: Specimen presentation, imaging devices, inter-observer variation, intra-observer variation
• Experimental Design: Repeated data acquisition from same specimens while varying one factor at a time
• Impact Assessment: Linear Discriminant Analysis (LDA) classification of species membership using different error-prone datasets compared to baseline classifications

This systematic approach enabled researchers to not only quantify the magnitude of each error type but also determine their downstream effects on statistical classification accuracy [14].

Emerging Methodologies for Enhanced Robustness

Landmark-Free Approaches

Recent research has explored landmark-free methods to circumvent operator-dependent landmark digitization. One study applied Deterministic Atlas Analysis (DAA), a Large Deformation Diffeomorphic Metric Mapping (LDDMM) approach, to 322 mammalian crania spanning 180 families [18].

Table 2: Comparison of Traditional vs. Landmark-Free Morphometric Approaches

Feature	Traditional Landmark-Based GM	Deterministic Atlas Analysis (DAA)
Data Collection	Manual/semi-automated landmarking	Automated mesh processing
Time Requirement	High (hours to days)	Low (minutes to hours after setup)
Operator Bias	High (inter-operator error up to 34%)	Minimal after parameter optimization
Homology Requirement	Strict anatomical homology needed	No strict landmark homology required
Phylogenetic Scope	Limited for highly disparate taxa	Suitable for broad taxonomic comparisons
Shape Representation	Discrete landmarks	Continuous deformation fields (momenta vectors)
Key Limitation	Limited landmarks across disparate taxa	Mesh topology sensitivity; parameter selection

DAA generates comparable but non-identical estimates of phylogenetic signal, morphological disparity, and evolutionary rates relative to traditional landmarking, offering enhanced efficiency for large-scale studies [18]. The method requires careful parameter selection, particularly kernel width, which controls the spatial scale of deformations.

Functional Data Innovations

A groundbreaking study introduced seven new pipelines integrating functional data analysis (FDA) with traditional GM, employing square-root velocity function (SRVF) and arc-length parameterization for 3D data [89].

Innovative Pipelines:

• Arc-GM: Reparameterizes shapes to uniform arc length before Procrustes alignment
• FDM (Functional Data Morphometrics): Models landmark trajectories as multivariate functions
• Soft-SRV-FDM: Blends identity mapping with estimated SRVF warp for soft elastic alignment
• Elastic-SRV-FDM: Applies full SRVF-based elastic alignment to isolate amplitude differences

These FDA approaches improve classification accuracy for dietary categories in kangaroo crania while offering more robust shape representations that better accommodate complex morphological variation [89].

Research Reagent Solutions Toolkit

Table 3: Essential Materials and Software for Reproducible Geometric Morphometrics

Tool Category	Specific Examples	Function/Purpose	Considerations for Reproducibility
Imaging Equipment	Olympus TG-6 macro camera [98]; Artec Eva structured-light scanner [41]; 1.5-T MRI system [97]	Generate high-resolution 2D/3D digital representations of specimens	Standardize equipment across studies; document resolution and settings
Landmark Digitization Software	Viewbox 4 [41]; dHAL Software	Precisely locate homologous landmarks on digital specimens	Use consistent template configurations; implement blinding procedures
Data Processing Platforms	R statistical environment [14] [41]; Deformetrica (for DAA) [18]; Python with specialized libraries [98]	Perform Procrustes alignment, statistical analysis, and visualization	Script entire workflow; use version-controlled code
Validation Datasets	GrainShape rice grain dataset [98]; Cryo-ET phantom dataset [99]	Benchmark methodological performance against ground truth	Utilize open-access reference datasets with known properties
Template Configurations	Os coxae digitization template [41]; 30-homologous-landmark rice grain template [98]	Standardize landmark placement across operators and studies	Publicly share and consistently apply template designs

Workflow Diagrams for Morphometric Protocols

Traditional Geometric Morphometrics Workflow

Landmark-Free Morphometric Analysis

The reproducibility of geometric morphometric analyses is significantly influenced by multiple factors, with inter-operator variation representing the most substantial challenge. Traditional landmark-based approaches, while powerful for homologous structure analysis, demonstrate notable vulnerability to digitization bias, particularly in large-scale collaborative research. Emerging methodologies including landmark-free approaches and functional data innovations offer promising avenues for enhancing robustness, though they introduce new considerations regarding parameter optimization and computational complexity. Researchers can improve reproducibility by standardizing imaging protocols, implementing template-based landmarking, utilizing validation datasets, and thoroughly reporting methodological details. The continuing development of automated and semi-automated approaches holds particular promise for reducing operator-dependent error while maintaining biological interpretability in geometric morphometric analyses.

Conclusion

The cross-validation performance of geometric morphometric protocols is not a one-size-fits-all metric but varies significantly with the biological question, anatomical structure, and data quality. While foundational GPA/PCA protocols remain widely used, evidence calls for cautious interpretation of their results due to inherent biases. The future of robust morphometric analysis lies in the strategic integration of methods—leveraging the detailed biological interpretability of GM with the superior predictive power of machine learning classifiers and the enhanced sensitivity of approaches like FDGM. For biomedical research, this translates to developing validated, application-specific protocols that ensure findings related to patient anatomy, disease morphology, or therapeutic targeting are both statistically sound and clinically reliable.

References

• 1. Comparison of geometric morphometric outline methods in the discrimination of age-related differences in feather shape - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC1592095/]
• 2. Geometric morphometrics approach for classifying ... [https://www.nature.com/articles/s41598-025-85718-4]
• 3. Testing the reliability of geometric morphometric and ... [https://www.sciencedirect.com/science/article/pii/S2666033425000048]
• 4. Geometric morphometric analysis of malocclusion on lateral cephalograms in Malaysian population - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6952698/]
• 5. Resolving sample identification bias in morphometrics ... [https://elifesciences.org/reviewed-preprints/94685v1/pdf]
• 6. Optimizing digitalization effort in morphometrics - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7723759/]
• 7. Building and Visualising Ordinations of Shape Data in R ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12178081/]
• 8. Biases of Principal Component Analysis (PCA) in Physical ... [https://elifesciences.org/reviewed-preprints/94685]
• 9. Functional data geometric morphometrics with machine ... [https://www.nature.com/articles/s41598-024-66246-z]
• 10. Geometric morphometrics and machine learning from three ... [https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2023.1203023/full]
• 11. A Geometric Morphometrics Approach for Predicting Olfactory ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12565324/]
• 12. A geometric morphometric approach for detecting different ... [https://www.nature.com/articles/s41598-025-86793-3]
• 13. Geometric morphometrics on fossil shark teeth [https://palaeo-electronica.org/content/2025/5603-geometric-morphometrics-on-fossil-shark-teeth]
• 14. Are geometric morphometric analyses replicable? Evaluating ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7140992/]
• 15. The problem of assessing landmark error in geometric ... [https://pubmed.ncbi.nlm.nih.gov/17503448/]
• 16. Repeatability, Reproducibility, Separative Power and ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0157890]
• 17. Data: Are geometric morphometric analyses replicable? ... [https://datadryad.org/dataset/doi:10.6071/M3KD40]
• 18. Assessing the application of landmark-free morphometrics to ... [https://bmcecolevol.biomedcentral.com/articles/10.1186/s12862-025-02377-9]
• 19. Reproducibility of AI in Cephalometric Landmark Detection [https://www.mdpi.com/2075-4418/15/19/2521]
• 20. A Comparative Analysis of Artificial Intelligence Generated ... [https://www.sciencedirect.com/science/article/pii/S1073874625001409]
• 21. Robustness and reproducibility for AI learning in ... [https://www.nature.com/articles/s41598-024-51381-4]
• 22. Evaluating statistical models for establishing morphometric ... [https://www.sciencedirect.com/science/article/abs/pii/S0305440322000681]
• 23. Disparities in the analysis of morphological disparity - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7423048/]
• 24. Geometric morphometrics approach for classifying children's ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11785739/]
• 25. FACEDIG automated tool for placing landmarks on facial ... [https://www.nature.com/articles/s41598-025-09714-4]
• 26. The Novel Application of Geometric Morphometrics with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8541025/]
• 27. Geometric morphometrics in the cloud [https://www.sciencedirect.com/science/article/abs/pii/S1567134819300152]
• 28. Cross-validation (statistics) [https://en.wikipedia.org/wiki/Cross-validation_(statistics)]
• 29. Evaluation of a decided sample size in machine learning ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-023-05156-9]
• 30. Sample Size and Geometric Morphometrics Methodology ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10833145/]
• 31. Cross-validation failure: Small sample sizes lead to large ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811917305311]
• 32. Estimating Sample Size and Reducing Overfitting [https://pmc.ncbi.nlm.nih.gov/articles/PMC11005022/]
• 33. Comparison of Landmark- and Outline-Based Geometric ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6247772/]
• 34. Quantitative Analysis of Fish Morphology Through ... [https://en.bio-protocol.org/en/bpdetail?id=5087&type=0]
• 35. Geometric Morphometrics of Astragalus and Shape Variation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12023762/]
• 36. Shape-changing chains for morphometric analysis of 2D and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8563779/]
• 37. Two methods for geometric morphometric analysis of ... [https://link.springer.com/article/10.1007/s00436-024-08354-3]
• 38. First insights into using outline-based geometric ... [https://www.sciencedirect.com/science/article/pii/S2667114X24000499]
• 39. A practical, step-by-step, guide to taxonomic comparisons ... [https://europeanjournaloftaxonomy.eu/index.php/ejt/article/view/2527]
• 40. Noctuidae) to support pest identification in invasive species ... [https://www.frontiersin.org/journals/insect-science/articles/10.3389/finsc.2025.1542467/full]
• 41. Defining, Reconstructing and Investigating Shape Through a ... [https://journal.caa-international.org/articles/10.5334/jcaa.161]
• 42. Comparative study of nasal cavity drug delivery efficiency ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11015827/]
• 43. Experimental measurements of particle deposition in the ... [https://www.sciencedirect.com/science/article/pii/S0378517325001164]
• 44. Investigation of particle penetration in a nasal maxillary ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12618317/]
• 45. Geometric morphometric analysis of mandibular morphology ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12465032/]
• 46. Age estimation of children and adolescents from mandibles ... [https://www.nature.com/articles/s41598-025-21221-0]
• 47. Age estimation from mandibles in Malay: A 2D geometric ... [https://www.sciencedirect.com/science/article/pii/S165836122300094X]
• 48. Exploring the Relationship Between Mandibular Morphology ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12607250/]
• 49. Age estimation of children and adolescents from mandibles ... [https://pubmed.ncbi.nlm.nih.gov/41057564/]
• 50. Automatic taxonomic identification based on the Fossil ... [https://www.cambridge.org/core/journals/paleobiology/article/automatic-taxonomic-identification-based-on-the-fossil-image-dataset-415000-images-and-deep-convolutional-neural-networks/4863E2FDE20D6115415EE5FE232B9DCD]
• 51. Toward non-invasive early pest surveillance: cross-modal ... [https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1610163/full]
• 52. Toward non-invasive early pest surveillance - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC12528037/]
• 53. The relative performance of geometric morphometrics and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10049884/]
• 54. A review of machine learning applications for identification ... [https://www.sciencedirect.com/science/article/pii/S1574954125003383]
• 55. Comparison of geometric morphometric outline methods in the ... [https://frontiersinzoology.biomedcentral.com/articles/10.1186/1742-9994-3-15]
• 56. Critical evaluation of the effects of a cross-validation ... [https://www.sciencedirect.com/science/article/pii/S2666154324001339]
• 57. 12 Important Model Evaluation Metrics for Machine ... [https://www.analyticsvidhya.com/blog/2019/08/11-important-model-evaluation-error-metrics/]
• 58. A new method of semi-supervised learning classification ... [https://www.nature.com/articles/s41598-025-02324-0]
• 59. A Comparison of Semilandmarking Approaches in the ... [https://www.mdpi.com/2076-2615/13/3/385]
• 60. Comparing semi-landmarking approaches for analyzing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12088985/]
• 61. Shaped by evolution: geometric morphometrics and ... [https://link.springer.com/article/10.1007/s10914-025-09774-3]
• 62. Landmark-based geometric morphometrics: Influential ... [https://sciety-discovery.elifesciences.org/articles/by?article_doi=10.1101/2025.07.14.664645]
• 63. Assessing the application of landmark-free morphometrics ... [https://pubmed.ncbi.nlm.nih.gov/40289084/]
• 64. A novel approach to craniofacial analysis using automated ... [https://www.nature.com/articles/s41598-024-63137-1]
• 65. The use of Geometric morphometry to illustrate variations ... [https://www.sciencedirect.com/science/article/pii/S2468227625004843]
• 66. Bolstering geometric morphometrics sample sizes with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8128762/]
• 67. Incomplete specimens in geometric morphometric analyses [https://ui.adsabs.harvard.edu/abs/2014MEcEv...5...16A/abstract]
• 68. Incomplete specimens in geometric morphometric analyses [https://www.academia.edu/11448740/Incomplete_specimens_in_geometric_morphometric_analyses]
• 69. Missing data in multivariate morphometrics [https://jessicaarbour.com/missing-data/]
• 70. Missing Data Estimation in Morphometrics: How Much Is ... [https://pubmed.ncbi.nlm.nih.gov/24335428/]
• 71. Comparison of geometric morphometric outline methods in ... [https://www.semanticscholar.org/paper/Comparison-of-geometric-morphometric-outline-in-the-Sheets-Covino/ef6f55998c80cf6bb8b4fda738fe3769bc660faf]
• 72. Size, shape, and form: concepts of allometry in geometric ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4896994/]
• 73. Thirty years of geometric morphometrics - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC9545184/]
• 74. Clovis point allometry, modularity, and integration [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0289489]
• 75. Reliability and reproducibility of three-dimensional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4428454/]
• 76. ESMRMB 2025 focus topic: cycle of quality—from concept to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12255560/]
• 77. A Systematic Review of Medical Image Quality Assessment [https://www.mdpi.com/2313-433X/11/4/100]
• 78. Convolutional neural networks and outline analyses for ... [https://peercommunityjournal.org/articles/10.24072/pcjournal.647/]
• 79. Comparison of handcrafted features and convolutional ... [https://www.nature.com/articles/s41598-020-77264-y]
• 80. Cross-cohort generalizability of deep and conventional ... [https://www.sciencedirect.com/science/article/pii/S221315822100156X]
• 81. Comparative Analysis of Deep Learning Models for Predicting ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12562713/]
• 82. Tooth shape and sex estimation: a 3D geometric ... morphometric [https://link.springer.com/article/10.1007/s13205-025-04439-7]
• 83. Bat Species Comparisons Based on External Morphology... | PLOS One [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0127043]
• 84. Application and Comparison of Machine Learning and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11135637/]
• 85. Optimizing taxonomic classification of marker-gene amplicon ... [https://microbiomejournal.biomedcentral.com/articles/10.1186/s40168-018-0470-z]
• 86. Machine learning models for delineating marine microbial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12204397/]
• 87. HusMorph: a simple machine learning app for automated ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12543357/]
• 88. A novel approach to geometric morphometrics [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009061]
• 89. Reframing Three-Dimensional Morphometrics Through ... [https://arxiv.org/html/2509.00650v1]
• 90. Functional data analysis of ground reaction forces ... [https://www.nature.com/articles/s41598-025-24411-y]
• 91. Classification: Accuracy, recall, precision, and related metrics [https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall]
• 92. Evaluation Metrics For Classification Model [https://www.analyticsvidhya.com/blog/2021/07/metrics-to-evaluate-your-classification-model-to-take-the-right-decisions/]
• 93. Comparing Classification Models - You're Probably Doing ... [http://practicalcheminformatics.blogspot.com/2023/11/comparing-classification-models-youre.html]
• 94. Statistical variability in comparing accuracy of ... [https://www.nature.com/articles/s41598-025-12026-2]
• 95. ImageNet: what is top-1 and top-5 error rate? [https://stats.stackexchange.com/questions/156471/imagenet-what-is-top-1-and-top-5-error-rate]
• 96. A Framework for Comparison and Interpretation of Machine ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11912182/]
• 97. An example of inter-operator bias in 3D landmarks and its ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0197675]
• 98. A landmark-annotated image dataset of japonica rice ... [https://www.sciencedirect.com/science/article/pii/S2352340925005086]
• 99. A realistic phantom dataset for benchmarking cryo-ET data ... [https://www.nature.com/articles/s41592-025-02800-5]