Emerging Methods: 3D Scanning and Geometric Morphometrics

Photogrammetry reconstructs three-dimensional geometry from a set of two-dimensional photographs. The method exploits the principle that if the same surface point appears in two or more photographs taken from different positions, and if the camera positions are known or can be estimated, the 3D position of the point can be computed by triangulation. Modern photogrammetry software (Structure from Motion, or SfM, is the computational framework) automates this process: the analyst photographs an object from multiple angles with substantial overlap between adjacent images, imports the image set into the software, and the software extracts feature points, matches them across images, estimates camera positions, and builds a dense point cloud, then a textured mesh, of the object's surface.

Agisoft Metashape Professional (Agisoft LLC, Saint Petersburg, Russia) is the most widely cited photogrammetry software in published forensic anthropology papers. Its workflow is: import images, align photos (SfM feature matching), build dense cloud, build mesh, build texture. A full cranium shot with 60 to 80 images on a rotating turntable, each image with 50 to 60 per cent overlap with adjacent images, processes in 30 to 90 minutes on a mid-range workstation and produces a mesh with accuracy of approximately 0.2 to 0.5 mm depending on camera resolution and lighting quality. The software is available at approximately USD 3,500 for a permanent licence.

RealityCapture (Capturing Reality, Slovakia, acquired by Epic Games 2023) is a faster alternative, processing the same image set in a fraction of the time using GPU acceleration, and is the preferred software in high-throughput DVI contexts where dozens or hundreds of bone elements need to be digitised rapidly. COLMAP (open-source, maintained by ETH Zurich) provides a zero-cost alternative with somewhat longer processing times, used by under-resourced laboratories in South Asia, sub-Saharan Africa, and Latin America.

The accessibility of photogrammetry using smartphone cameras is significant for equity in the discipline. A practitioner at the AIIMS forensic medicine department in New Delhi, the CFSL in Kolkata, or a state FSL without access to a dedicated 3D scanner can produce a photogrammetric model of a skull using an iPhone 14 or Samsung Galaxy S23 camera, a DIY turntable, and COLMAP or a low-cost cloud photogrammetry service. Published validation studies (De Oliveira, 2020; Colman, 2021) have confirmed that smartphone photogrammetry achieves accuracy of 0.3 to 0.6 mm on skeletal targets, adequate for most biological-profile and trauma documentation purposes. This matters because the largest unmet casework need for non-contact documentation is in laboratories in South Asia, sub-Saharan Africa, and Latin America, where scanner hardware is either unavailable or cost-prohibitive.

Computed tomography applied to skeletal material has been used in forensic anthropology since the 1980s, initially on clinical CT scanners with voxel sizes of 0.5 to 1.0 mm. Micro-computed tomography (micro-CT), operating at voxel sizes of 10 to 100 micrometres, opens the internal architecture of cortical and trabecular bone to quantitative analysis at a resolution that the clinical scanner cannot achieve.

For age estimation, the most significant application is the measurement of trabecular microstructure in weight-bearing bones. Bone density peaks in the early third decade and then declines through a remodeling imbalance (formation slightly less than resorption) that progressively reduces trabecular bone volume (BV/TV), increases trabecular separation (Tb.Sp), and decreases trabecular thickness (Tb.Th) and number (Tb.N). These parameters, measured from micro-CT scans of the femoral head, the iliac crest, or the first rib, correlate with biological age in published reference samples. A 2017 paper by Stull, L. et al. in the Journal of Forensic Sciences demonstrated that trabecular BV/TV and Tb.N in the iliac crest provided age estimates with standard errors of approximately 12 years, comparable to the surface morphological methods on the pubic symphysis, but without the examiner subjectivity.

For trauma analysis, micro-CT resolves internal features that are invisible to the surface examiner. A fracture that appears to have clean margins on the cortical surface may show internal fragmentation and a microcrack network consistent with a different mechanism when viewed in the micro-CT cross-section. Cut marks that are ambiguous on the surface, whether they penetrate into the cortex or are confined to the periosteal layer, are definitively resolved in cross-section. Gunshot entrance and exit beveling, which is assessed on the external cortical surface in standard examination, can be confirmed or modified by the internal CT cross-section showing the extent of beveling into the diploe and inner table.

For personal identification, micro-CT enables comparison between case material and antemortem clinical CT scans. A unique trabecular pattern in the femoral head, or a specific cortical density distribution in the tibia, can be matched to an antemortem scan even when the external surface morphology has been modified by taphonomy. The Swedish National Forensic Centre and the Federal Institute of Legal Medicine in Lausanne (Switzerland) use micro-CT routinely in personal identification casework, and the method has been applied in the 9/11 WTC fragment identification work at the OCME New York.

Clinical CT scanners (available in any hospital with a radiology department) provide a useful intermediate step where micro-CT is not available. At 0.5 to 1.0 mm voxel resolution, clinical CT can confirm gross trauma features (cranial beveling, fracture pattern, metallic foreign bodies), document pathological lesions relevant to the biological profile, and produce a dataset that can be shared with other institutions or jurisdictions without physical transport of the bone. Indian AIIMS hospitals, Australian state forensic pathology services, and UK coroner's court supported pathology services all have clinical CT capability that is in routine use for forensic anthropology-adjacent casework.

Geometric morphometrics (GMM) is a set of methods for the statistical analysis of biological shape, in which shape is represented by the coordinates of homologous landmarks on a structure, independent of the size, position, and orientation of the specimen. The method has been applied to cranial and postcranial skeletal morphology in forensic anthropology since the early 2000s and has substantially changed how population-affinity and sex-estimation studies are conducted.

The workflow begins with landmark placement. A set of landmarks is defined, each representing a biologically homologous point on the skeletal surface (for example, the most anterior point of the nasion, or the most inferior point of the post-orbital constriction). The analyst identifies each landmark on the 3D surface model (from laser scan, photogrammetry, or CT) and records its x, y, and z coordinates. For cranial analysis, standard landmark sets have been published by Slice (2001), Bookstein and colleagues, and the AAAS Geometric Morphometrics Open Course; forensic applications typically use between 40 and 90 cranial landmarks.

Procrustes superimposition removes the non-shape variation, size (by scaling all specimens to the same centroid size), position (by translating all centroids to the origin), and orientation (by rotating specimens to minimise the sum of squared distances between corresponding landmarks). After superimposition, all specimens sit in the Procrustes shape space, a multidimensional space where the distance between two points represents the shape difference between two specimens.

Principal component analysis (PCA) in this shape space identifies the principal axes of variation: the first principal component (PC1) accounts for the largest proportion of shape variation in the sample, PC2 the next largest, and so on. For crania, PC1 typically accounts for 15 to 30 per cent of total shape variation and often relates to overall cranial breadth and vault height. Plotting specimens in the PC space (typically PC1 vs PC2 or a three-dimensional PC1/PC2/PC3 plot) visualises population-level clustering without the investigator deciding in advance which measurements are relevant.

Thin-plate spline (TPS) warping visualises the shape difference between a reference specimen and any other specimen in the shape space by warping a reference mesh to match the target landmark configuration. This is used in forensic anthropology to show, for instance, the average shape difference between male and female crania in a given population, or between European and East Asian cranial morphology, in a form that is more intuitive than a table of discriminant-function coefficients.

The R package geomorph, published by Adams and Otarola-Castillo in 2013 and maintained actively through 2024, is the primary computational tool for GMM analysis in forensic anthropology research. It implements Procrustes superimposition, PCA, multivariate regression of shape on size and other covariates, phylogenetic methods, and a set of visualisation functions. Published applications include sex estimation from cranial shape (Kimmerle, 2008), population-affinity discrimination (Harvati and Weaver, 2006; Guyomarc'h and Bruzek, 2011), age estimation from trabecular shape (using CT landmark data rather than surface landmarks), and trauma pattern classification.

Machine learning, in the context of forensic anthropology biological profile estimation, means applying statistical learning algorithms to skeletal measurement data (linear measurements, GMM landmark coordinates, CT density values) to produce estimates of sex, age, ancestry, and stature that are validated on labelled reference samples and that come with documented accuracy metrics.

The Detroit-Bone Sex/Maturity Logic (DBSML) is a classification system developed at Wayne State University using logistic regression on cranial and postcranial linear measurements from the Forensic Anthropology Data Bank (FADB). DBSML combines metric data from multiple skeletal elements into a single sex-estimation probability, providing a posterior probability of male or female rather than a binary classification, which allows the uncertainty to be explicitly incorporated into the report. Published accuracy on holdout samples from the FADB is approximately 90 to 95 per cent depending on the elements available, comparable to the best traditional discriminant-function approaches but with a more explicit uncertainty quantification.

LASSO (Least Absolute Shrinkage and Selection Operator) regression applied to cranial landmark GMM coordinates selects the subset of landmark-coordinate features most predictive of a biological profile parameter (sex, ancestry, stature) while suppressing irrelevant or redundant features. This addresses a specific problem in high-dimensional GMM datasets: with 80 landmarks each with x, y, z coordinates, there are 240 predictor variables, far more than any traditional multivariate regression should handle without regularisation. LASSO shrinks the coefficients of non-predictive variables toward zero, effectively performing feature selection and regression simultaneously. Published forensic studies (Biwasaka 2018; Coquerelle 2019) report sex-estimation accuracy of 90 to 97 per cent from LASSO models on cranial landmark data, with the advantage that the selected features have an anatomical interpretation that a morphologist can examine and defend.

Deep-learning age estimation from CT scans is the most technically ambitious application. A 2020 paper by Stull, Tise, Ali, and Horton in Forensic Science International demonstrated that a convolutional neural network (CNN) trained on a set of annotated CT scans of the hand and wrist (growth-plate fusion, bone density, cortical thickness) produced age estimates with a mean absolute error of approximately 1.8 years on sub-adult specimens, superior to all traditional radiographic methods on the same sample. For adult age estimation, a 2023 paper in Forensic Science International (building on earlier radiology deep-learning work from the German AGFA-NDT group) reported mean absolute errors of 7 to 10 years on CT-derived age estimation in the 30 to 80 year range, comparable to the best traditional methods but with much greater throughput: a trained model produces an estimate in seconds rather than the 15 to 30 minutes required for traditional morphological scoring.

A 2024 paper in Nature Methods (Rizk and colleagues) applied a graph neural network to geometric morphometric data from crania to estimate population affinity in a sample spanning 14 global populations. The reported cross-validation accuracy was 94.7 per cent on a held-out test set. The paper attracted both interest and immediate critique: the population-affinity categories used in the training data reflected existing skeletal collection groupings, which carry the same demographic and provenance limitations as the collections themselves, and the 94.7 per cent accuracy figure was achieved on a relatively small and geographically unbalanced test set. The corresponding authors acknowledged these limitations in the paper's discussion; the downstream use of such a model in casework would require population-specific validation far beyond the published study.

The admissibility of machine-learning-derived forensic anthropology opinions raises questions that are distinct from the accuracy of the underlying model. Even if a CNN age-estimator achieves a mean absolute error of 8 years in a controlled validation study, a court must evaluate several things that the accuracy figure does not capture.

First, the validation sample question. The Daubert standard requires that the method's error rate be known, but error rates in machine learning are sample-specific. A model trained and cross-validated on the FADB (predominantly European-American and African-American) may have a different error rate on South Asian or East Asian specimens from those reported in the training-sample validation. The published error rate is not the operational error rate if the case specimen is drawn from a different population distribution than the training data. This is the same problem that affects traditional discriminant-function methods, but it is more acute for neural networks because the model's internal representation is less interpretable and the analyst cannot easily assess which features it is using to make the estimate.

Second, the transparency problem. Traditional methods, such as the Suchey-Brooks six-phase scoring system, are transparent: the phase descriptions are published, illustrated, and reproducible. A second analyst can read the same phase description, examine the same symphysis, and either confirm or dispute the primary analyst's conclusion. A neural network's conclusion is a number (a posterior probability or a point estimate) derived from a computation across tens of millions of learned parameters. It cannot be re-examined in the way a morphological score can; what a second analyst can do is run the same model on the same CT image and confirm that the model produces the same output, which is useful for replication but does not allow the second analyst to independently assess whether the model's conclusion is correct.

Third, the chain-of-custody problem for digital data. A 3D scan or CT dataset used as input to a machine-learning model must be maintained with the same chain-of-custody discipline as a physical bone specimen. The digital file must be demonstrably unaltered between capture and analysis; the model version used must be documented (a model updated after the analysis produces different output from the same input); and the analysis pipeline must be reproducible from the case record.

These problems are solvable in principle. Explainability methods (LIME, SHAP, attention visualisation for CNNs) can partially decompose a model's output into contributions from identifiable input features, providing at least a partial answer to the "why" question. Population-specific validation studies can be conducted before a model enters casework. Digital chain of custody can be enforced through cryptographic hashing of input files and logging of model version numbers. Regulatory bodies in the US (OSAC), UK (Forensic Science Regulator), and Australia (ANZPAA NIFS) have begun issuing guidance on the validation requirements for computational forensic methods including machine-learning estimators.

In India, the BSA 2023 framework does not specifically address computational methods, but the Supreme Court of India's general requirement that expert opinions be supported by reasons (Ramesh Chandra Agrawal v. Regency Hospital, 2009) applies to machine-learning outputs as readily as to traditional morphological opinions: the expert must be able to explain, in terms accessible to the court, the basis for the conclusion. Indian forensic practitioners planning to use ML estimators in casework should validate the model's performance on an Indian skeletal reference sample and should prepare a plain-language explanation of the method's principles before the testimony.

The equipment-access gap in non-contact osteology is real and well-documented. A micro-CT scanner costs between USD 200,000 and 600,000. An Artec Spider costs approximately USD 25,000. An Agisoft Metashape licence costs approximately USD 3,500. The FADB, which underlies FORDISC and many discriminant-function studies, is predominantly North American. The OSAC Anthropology Subcommittee meets in Washington DC and its published standards are in English.

The practical consequence is that the laboratories with the least resource and the most caseload are the least able to access the new methods. A state FSL in a Tier 2 Indian city, a government forensic laboratory in sub-Saharan Africa, or a university forensic anthropology programme in Latin America may have neither the budget for scanner hardware, the licensing budget for commercial photogrammetry software, nor the connectivity to participate in OSAC or ENFSI FAWG working group discussions.

Two technology families within the emerging methods stack are partially exempt from this access gap. Photogrammetry with smartphone cameras and COLMAP costs essentially nothing beyond the time to capture and process the images. Published validation studies (Colman 2021, De Oliveira 2020, Schutz and Gruber 2022) confirm that smartphone photogrammetry achieves accuracy of 0.3 to 0.6 mm on skeletal targets, adequate for trauma documentation and GMM landmark placement. The geomorph R package is free and open-source, runs on any laptop, and its landmark-analysis methods are publishable in peer-reviewed journals.

The second partial exception is clinical CT. Most regional hospitals and government medical colleges in India, South Africa, Brazil, and Mexico have clinical CT scanners as part of their radiology infrastructure. At 0.5 to 1.0 mm voxel resolution, clinical CT cannot resolve individual trabeculae as micro-CT does, but it can document gross trauma, produce a 3D surface model from the DICOM data (using free software such as 3D Slicer), and enable GMM landmark placement on the rendered surface. Several Indian forensic anthropology research groups (including groups at AIIMS New Delhi, SGPGI Lucknow, and Kasturba Medical College Manipal) have published sex-estimation and age-estimation studies using clinical CT data and geomorph-based GMM analysis, demonstrating that a substantive part of the emerging-methods toolkit is achievable within existing infrastructure.

International collaboration provides a third pathway. The INTERPOL DVI Standing Committee, the ICRC's Forensic Coordinator network, and bilateral research partnerships between Western forensic anthropology programmes and universities in South Asia and Africa have, in several cases, provided scanner access through shared-equipment agreements or scanning-visit protocols, where a researcher from a partner institution travels to a facility with scanner access and digitises a reference sample from a home collection. The resulting digital dataset can then be returned to the home institution for analysis.

A fully integrated non-contact osteology workflow for a casework submission looks as follows. On receipt of the skeletal material, before any physical examination, the forensic anthropologist photographs the specimen set for photogrammetric documentation using a standardised multi-angle protocol. The resulting 3D models are processed (Metashape or RealityCapture for well-resourced labs; COLMAP for others) and stored as part of the case file. These models form the permanent digital record of the specimen's surface condition at the time of receipt, independent of any subsequent physical examination.

For trauma analysis, key elements, particularly the cranium, any vertebrae with suspicious lesions, long-bone fracture surfaces, and any element with suspected cut marks or gunshot defects, are scanned at higher resolution (NextEngine or Artec Spider if available, or a second photogrammetry pass at closer range). The resulting models allow sub-millimetre measurement of kerf profiles, beveling angles, and fracture-surface morphology without touching the original bone after the initial scan.

GMM landmark placement for the biological profile uses the 3D models rather than the bone in hand. This enables inter-laboratory review (the 3D model can be emailed to a second expert), eliminates the physical access constraint (the bone does not need to be available at the time of review), and provides a record that is auditable against the original digital capture. Landmark placement can be performed in Avizo (Thermo Fisher), Checkpoint (Stratovan), or the free MorphoJ software.

The biological profile estimation integrates traditional morphological scoring (Suchey-Brooks, Phenice, Walker) with metric methods (discriminant functions, FORDISC) and, where validated ML estimators are available and explainable, with the ML output as a corroborating estimate. The report documents all three tiers, specifies the reference collection and error rates for each method, and expresses the final estimate in the OSAC/SWGANTH qualified-opinion language with an explicit uncertainty range.

1. Photogrammetric documentation on receipt
   Before physical examination: 60-80 overlapping photographs per key element, turntable or hand-held circuit. Process in Metashape, RealityCapture, or COLMAP. Store processed 3D models with the case file as a permanent surface-condition record.
2. High-resolution scan of trauma elements
   For elements with suspected trauma (cut marks, gunshot defects, fracture surfaces): NextEngine fixed-stage scan or Artec Spider handheld scan at 0.05-0.2 mm resolution. Alternatively, close-range photogrammetry pass. Document kerf profiles and beveling angles from the model.
3. Clinical or micro-CT acquisition (if available)
   Route key elements through a clinical CT scanner for gross trauma confirmation, internal fracture documentation, and DICOM dataset creation. For research-quality age estimation: micro-CT of trabecular structures in the proximal femur, iliac crest, or first rib.
4. GMM landmark placement on 3D models
   Load 3D surface models into Checkpoint, Avizo, or MorphoJ. Place the landmark set (40-90 cranial landmarks, or appropriate postcranial set). Export landmark coordinate files (.nts, .tps, or .csv). All placement on digital model, not on bone in hand.
5. Statistical analysis in geomorph (R) or PAST
   Procrustes superimposition, PCA, and discriminant-function analysis in the geomorph package (R) or PAST (Hammer et al., open-source). Generate sex and ancestry estimates with posterior probabilities. Cross-validate against the relevant reference sample.
6. Integration with traditional morphological scoring
   Conduct traditional Suchey-Brooks, Phenice, Walker, and Iscan-Loth scoring directly on the bone or on the 3D model where validated. Record traditional and metric estimates separately. Where ML estimator is used: document training data, validation accuracy, and method limitations.
7. Report with full method documentation
   Report includes: 3D model capture protocol, instrument model and settings, GMM landmark set reference, statistical software and version, all method reference publications, reference sample demographics and consent status, known error rates, and final estimate in OSAC four-tier qualified-opinion language with explicit uncertainty range.