The research, published in the International Journal of Legal Medicine, was led by Suguru Torimitsu of the University of Tokyo and Chiba University, together with collaborators including Eiryo Kawakami and Yohsuke Makino. The team set out to test whether a so-called foundation model—a large, pretrained neural network of the kind that has transformed medical imaging—could simultaneously estimate both sex and age from three-dimensional computed tomographic images of the femur. Their cohort was drawn from postmortem CT scans of 1,489 Japanese individuals, 1,003 males and 486 females, aged between 20 and 90 years, all of whom underwent whole-body PMCT followed by forensic autopsy at the two universities between May 2015 and November 2025.

The technical pipeline behind the study is a showcase of modern deep learning practice. First, the femora were automatically segmented from whole-body scans using the TotalSegmentator library, which isolates anatomical structures without manual tracing. Each segmented femur was then converted into a three-dimensional volume of 112 by 112 by 112 voxels, with voxels outside the bone mask replaced by air-equivalent values to suppress background noise, and intensity normalized using percentile-based scaling. Rather than feeding the model selected two-dimensional slices, the researchers processed the entire volumetric bone, allowing the network to learn from the complete three-dimensional morphology of the femur.

At the heart of the system sits 3DINO, a three-dimensional Vision Transformer foundation model pretrained on large-scale medical imaging data. Vision Transformers work by dividing an image into patches and encoding the relationships between those patches through attention mechanisms, which let the model weigh the importance of different regions dynamically. To adapt this general-purpose backbone to the forensic task, the team used low-rank adaptation, or LoRA, a parameter-efficient fine-tuning technique in which only small, low-rank matrices inserted into the network are updated during training while the original pretrained weights remain frozen. This approach dramatically reduces the number of trainable parameters, lowering the risk of overfitting on a dataset of modest size and shortening training time.

The framework was trained in a multitask learning configuration, meaning a single shared feature-extraction backbone fed two separate output heads: one performing binary classification for sex, and one performing regression for age, with predicted values scaled to the physiological range through a sigmoid function. The total loss was a weighted sum of the classification and regression losses, enabling joint optimization. The intuition is biologically sound: sex-related morphology and age-related structural change are correlated aspects of the same skeleton, so learning them together can produce richer internal representations than learning either in isolation. Training used the AdamW optimizer with a cosine-annealing learning-rate schedule, data augmentation including random cropping and affine transformations, and early stopping to prevent overfitting. The dataset was split into a training-and-validation cohort and an independent test cohort in a four-to-one ratio, with stratified sampling by age.

The results, evaluated on the independent test set, were striking. For sex estimation, the model achieved balanced accuracies of 0.897 for the left femur and 0.929 for the right femur, with area under the receiver operating characteristic curve values of 0.979 and 0.988 respectively—figures the authors describe as favorable compared with previous reports. Earlier CT-based morphometric studies in Japanese populations had produced accuracies of roughly 73 percent from femoral length alone, rising to around 88 to 90 percent when combining measurements such as femoral epicondylar breadth and head diameter. The deep learning approach, which captures complex nonlinear morphological patterns beyond what linear measurements can express, matched or surpassed those benchmarks without any manual feature extraction. Calibration metrics, including Brier scores and expected calibration error, were both below 0.1, indicating that the model’s predicted probabilities were trustworthy rather than merely accurate on average.

Age estimation proved harder, as it always does in forensic anthropology, where adult skeletal indicators are notoriously imprecise. The models achieved a mean absolute error of approximately eight years and a root mean square error of approximately ten years. Notably, performance was better for female individuals than for males, with lower error values and a higher coefficient of determination in the female subgroup. Bland–Altman analysis revealed mean biases close to zero, but also a familiar statistical signature: the model tended to overestimate the ages of younger individuals and underestimate those of older ones, a regression toward the mean that plagues most skeletal aging methods. The authors suggest that age-related morphological variation in the femur may be more consistent in females, while greater interindividual variability in males reduces accuracy; they also note that postmenopausal estrogen deficiency may shape female femoral morphology in ways the model can exploit.

Perhaps the most scientifically intriguing findings came from the attention maps, which reveal where in the bone the network was looking when it made its decisions. For sex estimation, attention concentrated overwhelmingly in the distal femur—the lower end near the knee—regardless of age group, a result consistent with decades of osteometric research showing that the epicondylar breadth of the distal femur is among the most sexually dimorphic measurements available. For age estimation, however, attention shifted to the femoral shaft, the long diaphyseal midsection, which in the right femur accounted for the highest attention in nearly all cases across both age groups. This dissociation suggests the model is genuinely learning distinct biological signals rather than relying on a single anatomical shortcut, and it points forensic researchers toward diaphyseal features as an underexplored source of age-related information.

One unexpected observation was a laterality effect: the right femur model consistently outperformed the left, and attention maps from the left femur showed greater variability in the contributing regions, particularly among older individuals. The authors caution that the reasons remain unclear and that further work is needed to determine whether this reflects biology, imaging conditions, or methodological artifacts. They are equally candid about other limitations. The study included only intact femora from Japanese individuals, so it remains unknown whether the models work on fragmented or decomposed remains—precisely the conditions most common in real forensic casework. Skeletal sexual dimorphism also varies across populations, meaning standards derived from Japanese data may not transfer directly elsewhere, and factors such as nutrition, lifestyle, occupation, and cultural habits were not examined.

Even with those caveats, the implications are considerable. Automated segmentation and foundation-model feature extraction remove the reliance on manual measurements, improving reproducibility and objectivity while enabling rapid analysis of large datasets—capabilities that are especially valuable in time-sensitive scenarios such as disaster victim identification, where hundreds of remains may need assessment under acute pressure. The multitask design means a single scan of a single bone yields both components of the biological profile at once. The authors argue that extending AI-based approaches to understudied skeletal elements could advance the field further, and they call for larger, more heterogeneous datasets, including multi-population cohorts and incomplete skeletal material, to validate and generalize the approach. If those efforts succeed, the humble thigh bone—already the bone most likely to survive environmental degradation—may become the backbone of a new generation of automated forensic identification, in which an algorithm reads the story written in bone faster, and perhaps more consistently, than any human expert.

Subject of Research: Deep learning–based estimation of age and sex from femoral postmortem CT images in a Japanese population

Article Title: Age and sex estimation from Japanese femoral computed tomographic images using foundation models

Article References: Torimitsu, S., Ikari, H., Tsuneya, S., Uemura, Y., Okada, M., Kojima, M., Al Mansoori, R., Iwase, H., Kawakami, E., & Makino, Y. (2026). Age and sex estimation from Japanese femoral computed tomographic images using foundation models. International Journal of Legal Medicine. https://doi.org/10.1007/s00414-026-03998-5

Image Credits: AI Generated

DOI: 10.1007/s00414-026-03998-5

Keywords: forensic anthropology, deep learning, foundation models, postmortem CT, femur, sex estimation, age estimation, Vision Transformer, LoRA, multitask learning, biological profile, victim identification

The research, led by Alperen Tekin and colleagues at Istanbul Medeniyet University, addresses a persistent challenge in forensic anthropology. When investigators recover skeletal remains, estimating sex is one of the first and most important steps, because it narrows the pool of missing persons and shapes every subsequent analysis of age, stature, and ancestry. The most reliable methods rely on the pelvis and skull, but these bones are frequently missing, fragmented, or damaged in disaster scenes, fires, and archaeological contexts. Vertebrae, which are dense and often survive, have therefore attracted growing attention as alternative sex indicators, and the second cervical vertebra, known as the axis or C2, has emerged as a particularly promising candidate.

What makes the new study distinctive is its source of imaging data. Cone-beam computed tomography, or CBCT, is a low-dose scanning technology used routinely in dentistry and maxillofacial radiology, and the axis vertebra falls within the field of view of many head-and-neck CBCT examinations. This means that vast archives of existing clinical scans, acquired for entirely unrelated dental reasons, could in principle be mined for forensic information without exposing anyone to additional radiation. The researchers exploited this opportunity by retrospectively analyzing 140 adult CBCT examinations, 68 from female patients and 72 from male patients, all acquired on a single scanner under homogeneous imaging conditions.

Rather than measuring the bone manually with calipers, as traditional morphometric studies do, the team turned to radiomics, a computational approach that converts medical images into hundreds of quantitative features. Radiomics captures characteristics that the human eye cannot reliably assess, including the statistical distribution of gray values, the coarseness and uniformity of bone texture, and the precise three-dimensional shape of the segmented structure. From each scan, the researchers extracted 107 original-image radiomic features from two different regions of interest: the entire second cervical vertebra, and the odontoid process alone, the tooth-like projection that rises from the vertebral body and allows the head to rotate.

The central question was methodological as much as forensic. If a forensic team only has access to the odontoid process, perhaps because the rest of the vertebra is damaged or outside the scan volume, does segmenting just that smaller structure sacrifice the discriminative information contained in the whole bone? To find out, the researchers built paired prediction models for each region of interest, using regularized machine learning algorithms, including least absolute shrinkage and selection operator, or LASSO, regression and elastic-net regression, both of which are designed to handle datasets where the number of features is large relative to the number of samples.

The results, validated through a rigorous internal testing framework, were strikingly consistent. Across 100 repeated paired validations with a balanced 80/20 train-test split, the mean area under the receiver operating characteristic curve, or AUC, reached 0.929 for the combined whole-C2 models, while the odontoid-only models achieved mean AUCs of 0.924 for texture-only features, 0.918 for combined LASSO models, and 0.905 for combined elastic-net models. An AUC of 0.5 would indicate performance no better than a coin flip, while 1.0 represents perfect discrimination, so values above 0.9 indicate strong internal classification ability. Five-fold nested cross-validation, a more conservative scheme that guards against information leaking from the test data into model training, produced AUCs of 0.918 for the whole-C2 combined LASSO model and 0.914 for the odontoid-only equivalent, confirming that the advantage of the whole vertebra was marginal at best.

The team also took deliberate steps to guard against the inflated performance estimates that have plagued much of the radiomics literature. All preprocessing steps were fitted exclusively on the training set, label-permutation testing was used to verify that the models were genuinely learning signal rather than noise, and the permutation null distributions were centered near 0.50, with a permutation p-value of 0.003 for both regions of interest. The authors were candid about the limits of their most eye-catching numbers: the highest AUCs observed on the primary test split, which ranged from 0.985 to 0.995, were derived from only 28 test subjects and were explicitly interpreted as optimistic estimates rather than realistic expectations of casework performance.

Perhaps the most consequential finding is that the odontoid process alone, a structure smaller than a fingertip, carries nearly as much sex-related information as the entire vertebra. The whole-C2 region retained more shape information, which makes anatomical sense given that the axis vertebra shows well-documented sexual dimorphism in its overall dimensions, but the texture features computed from the odontoid process alone performed on a par with the full-bone models. For forensic practice, this matters because the odontoid process is compact, anatomically well-defined, and easy to segment, and because it may be the only part of the vertebra preserved or captured in certain imaging scenarios. Previous studies had already shown that odontoid volume and morphometry differ between sexes, and the new results suggest that subtle internal bone texture adds a further layer of discriminative signal.

At the same time, the researchers are careful not to oversell the approach, and their caveats deserve as much attention as their headline numbers. The study was retrospective and single-device, meaning all scans came from one scanner and one institution under homogeneous imaging conditions, so the models have not yet faced the variability of different machines, scanning protocols, and populations. No external validation was performed, the reliability subsets used to assess feature reproducibility were too small to support comparative conclusions between the two regions of interest, and the biological meaning of many texture descriptors remains uncertain. High dimensionality and possible instability in feature selection further complicate interpretation, and the authors note that these limitations restrict the applicability of the models to real forensic casework for now.

Even with those caveats, the study represents a meaningful step toward a future in which routine dental scans, already acquired for millions of patients, double as a forensic resource. The work aligns with international standardization efforts such as the Image Biomarker Standardization Initiative and follows modern reporting frameworks for prediction models, signaling a maturing field that is increasingly aware of its own methodological pitfalls. If future multi-center studies can confirm that C2 radiomics generalizes across scanners and populations, forensic investigators may one day load an anonymous CBCT scan into software, segment a bone the size of a thumb, and receive a statistically grounded estimate of sex within seconds, turning an everyday dental image into a silent witness that helps put a name to the nameless.

Subject of Research: CBCT radiomics of the second cervical vertebra for forensic sex estimation

Article Title: Whole-axis versus odontoid process CBCT radiomics for forensic sex estimation: a retrospective paired-ROI model-development study

Article References: Tekin, A., Deveci, B., Tekin, G., & Ay, F. (2026). Whole-axis versus odontoid process CBCT radiomics for forensic sex estimation: a retrospective paired-ROI model-development study. International Journal of Legal Medicine. https://doi.org/10.1007/s00414-026-03999-4

Image Credits: AI Generated

DOI: 10.1007/s00414-026-03999-4

Keywords: forensic anthropology, radiomics, cone-beam computed tomography, second cervical vertebra, odontoid process, sex estimation, machine learning, forensic radiology, LASSO, elastic net, bone texture analysis, International Journal of Legal Medicine

The study, led by Yadigar Kastamoni and Kenan Öztürk of the Department of Anatomy at Süleyman Demirel University, together with Ahmet Dursun, Onur Can Şanlı and radiologist Veysel Atilla Ayyıldız, set out to reveal sex-related characteristics through morphometric measurements taken on the mandible, and to predict sex in individuals from different age groups using those variables. The work addresses one of the enduring challenges of forensic anthropology: identifying reliable, population-appropriate skeletal markers of sexual dimorphism that remain robust across the adult lifespan.

The research team analysed head and neck computed tomography images from 152 individuals, 77 males and 75 females, aged between 20 and 98 years. From these scans, the researchers generated three-dimensional reconstructions of each mandible using the RadiAnt DICOM viewer, a software platform that converts the two-dimensional slice data produced by CT scanners into volumetric models that can be measured in three dimensions. This approach reflects a broader shift in forensic science away from measurements taken with calipers on dry skulls and toward virtual anthropology, in which clinical imaging archives become rich sources of skeletal data.

The specific landmarks chosen for measurement were carefully selected regions of the mandible with well-established anatomical identities. The team measured dimensions relating to the foramen mentale, the mental foramen through which the mental nerve and vessels emerge on the front of the jaw; the condylus mandibulae, the rounded condyles that articulate with the skull at the temporomandibular joints; the incisura mandibulae, the notch between the condyle and the coronoid process on the upper margin of the ramus; and the angulus mandibulae, the angle of the jaw at the back and lower corner of the bone. In addition, the researchers measured the bigonial distance, the straight-line breadth of the jaw between the two gonion points at the angles on either side. All measurements, except those involving the mandibular angle, were carried out in RadiAnt DICOM Viewer 2020.1.1, and the resulting data were analysed statistically using SPSS 20.0 for Windows.

The results were strikingly consistent with the known biology of the human mandible. When the parameters were compared between the sexes, the mean values of all measurements, with the single exception of the left mandibular angle, were larger in males than in females. This pattern mirrors the influence of testosterone-driven bone growth during puberty, which produces a generally larger, squarer and more robust jaw in men, while hormonal and muscular factors in women tend to produce a jaw with a more obtuse gonial angle and finer proportions. The only measurement that failed to show a statistically significant sex difference was the gonial angle itself on the left side, a finding that reinforces a long-standing caution in the literature: angular measures of the jaw are notoriously variable and are influenced by age-related remodelling, tooth loss and chewing forces.

The degree of sexual dimorphism varied considerably across the measured parameters. The highest dimorphism rate was observed in the incisura mandibulae, the mandibular notch, which differed between the sexes by 14.39 percent on the right side and 14.29 percent on the left. At the opposite extreme, the angulus mandibulae showed almost no dimorphism, with differences of just 0.12 percent on the right and 0.57 percent on the left. This gradient is informative for forensic practice: it suggests that linear distances spanning the ramus, particularly between the mandibular condyle and the gonion, carry far more sex-diagnostic information than the angle of the jaw, which should be treated with caution when sexing fragmentary remains.

Indeed, the statistically significant differences between the sexes were found specifically in the distances from the condylus mandibulae to the gonion on both sides of the jaw, alongside the bigonial distance. These findings give forensic anthropologists a small set of hardy landmarks, robust bony points that survive decomposition and trauma relatively well, from which reliable linear measurements can be taken. Because the condyle and gonion are prominent structures on the ramus, they remain measurable even when the delicate alveolar portion of the jaw bearing the teeth has been damaged or lost, a common scenario in archaeological and forensic contexts.

What elevates the study beyond a simple descriptive anatomy exercise is its predictive modelling. The researchers built a logistic regression model from their measured variables, which correctly classified 91.4 percent of individuals in the development sample. A parallel discriminant function analysis achieved 90.1 percent classification accuracy on the original sample, and, crucially, retained 85.5 percent accuracy under leave-one-out cross-validation, a stringent test in which each individual is classified by a model trained on everyone else. The modest drop between original and cross-validated performance indicates that the model generalises reasonably well rather than simply memorising the training data, an important consideration for any method intended for real casework.

The authors conclude that sex estimation can be successfully performed using external measurements of the mandible, and they highlight the bigonial distance and the gonion-to-condyle measurements as particularly valuable parameters for sex differentiation. They suggest the method may prove useful in anatomy, forensic medicine and anthropology, especially in studies requiring sex estimation. This positions the mandible alongside the pelvis and skull as a first-rank bone for sex estimation, with particular value in contexts where those more classically dimorphic elements are missing. The mandible is also among the densest and most durable bones in the human skeleton, frequently surviving fire, fragmentation and long burial, which makes any validated method for reading sex from it disproportionately useful.

The study also carries broader implications for how forensic science will develop in the coming years. Computed tomography is now routinely performed in hospitals worldwide, and archives of clinical scans represent vast, untapped populations of reference data. Virtual osteometry of this kind does not require a physical skeleton at all: it can be applied ante-mortem to living patients in identification contexts, and post-mortem using clinical or forensic CT, avoiding destructive sampling of evidence. Earlier work cited by the team, including validation studies of three-dimensional CT measurements and postmortem CT morphometry of the mandible in Japanese populations, had already established the technical reliability of imaging-based measurement; the new study extends that framework with a large, wide-ranging adult sample spanning nearly eight decades of age, from 20 to 98 years.

As with all skeletal sex-estimation methods, population specificity remains the central caveat. The dimensions of the mandible vary across ancestral and geographic groups, and the classification functions derived here were developed on a Turkish sample; forensic practitioners applying them elsewhere would need to validate or recalibrate the model against local reference data before relying on it in identification casework. The retrospective design, approved by the Human Research Ethics Committee of Süleyman Demirel University with informed consent waived for anonymised data, also means the findings describe a clinical imaging population rather than a random demographic one. Nevertheless, with cross-validated accuracy approaching ninety percent from a handful of simple linear measurements, the message of the study is clear and, for a field where even a single additional clue can resolve an identification, quietly revolutionary: the jawbone remembers who you were, and it is willing to tell.

Subject of Research: Sex estimation from morphometric measurements of the human mandible in three-dimensional computed tomography images

Article Title: Sex estimation from the mandible in 3D computed tomography images

Article References: Kastamoni, Y., Dursun, A., Şanlı, O. C., Ayyıldız, V. A., & Öztürk, K. (2026). Sex estimation from the mandible in 3D computed tomography images. International Journal of Legal Medicine. https://doi.org/10.1007/s00414-026-04001-x

Image Credits: AI Generated

DOI: 10.1007/s00414-026-04001-x

Keywords: sexual dimorphism, mandible, computed tomography, forensic anthropology, 3D imaging, morphometry, sex estimation, bigonial distance, logistic regression, discriminant analysis, human identification, osteometric analysis