Biological Sex Determination in Cadavers Using Deep Learning Algorithms from Computed Tomography Images of Pelvis and Skull

Technical Summary: Biological Sex Determination in Cadavers Using Deep Learning Algorithms from Computed Tomography Images of Pelvis and Skull

Problem Statement
In forensic anthropology, determining biological sex is a critical step in building a biological profile for human identification. Traditional methods rely on morphological and metric analysis of skeletal structures, particularly the skull and pelvis. However, these classical approaches are observer-dependent, time-consuming, and often fail in challenging forensic scenarios involving mass disasters, charred remains, fragmentation, or severe decomposition. Furthermore, post-mortem taphonomic processes and trauma can alter anatomical structures, increasing image complexity and reducing the efficacy of manual analysis. While Artificial Intelligence (AI) and deep learning have shown promise in medical imaging, many existing studies are limited to specific populations or highly controlled acquisition conditions, lacking validation on real-world forensic casework featuring anatomical disfigurements and imaging artifacts.

Methodology
This study evaluated the robustness of deep learning architectures, enhanced by transfer learning, for the automated determination of biological sex using forensic Computed Tomography (CT) scans.

• Dataset: The study utilized a retrospective dataset of 141 autopsied cadavers (104 males, 37 females) from the Aristoclides Teixeira Institute of Forensic Medicine (IMLAT) in Goiânia, Brazil. The dataset included CT scans of the skull and pelvis, acquired via a Siemens SOMATOM go.Now scanner. The sample was intentionally heterogeneous, retaining cases with significant anatomical disfigurements and trauma-related artifacts to reflect real-world forensic conditions, excluding only cases with extensive bone destruction that precluded landmark identification.
• Data Preprocessing: Three-dimensional reconstructions of the skull and pelvis were generated using Volume Rendering Technique (VRT). To optimize computational efficiency, these 3D volumes were converted into standardized 2D profile projections. A forensic specialist manually curated the data to identify a "pivot image" (0° frontal view) for each subject, from which a rotational window of ±45° (11 images per anatomical region) was extracted. The final dataset comprised 3,059 images.
• Model Architectures: Seven distinct architectures were benchmarked: EfficientNetV2, ConvNeXt-Tiny, ViT-B16, VGG16, ResNet50, YOLO11, and YOLO26. All models utilized transfer learning, initialized with weights pre-trained on ImageNet.
• Classification Strategies: Two scenarios were tested:
  1. Binary Classification: Predicting sex only (Male/Female).
  2. Quaternary Classification: Jointly predicting sex and anatomical region (Male Skull, Female Skull, Male Pelvis, Female Pelvis).
• Training and Optimization: To address class imbalance and limited sample size, data augmentation techniques were employed, including a custom offline pipeline and native YOLO augmentations. Hyperparameter optimization was conducted using a Genetic Algorithm (GA) via the Ultralytics auto-tuning tool. To handle class imbalance, a Weighted Cross-Entropy Loss was implemented.
• Evaluation Protocol: The dataset was split at the subject level (75% training, 10% validation, 15% testing) to prevent data leakage. Patient-level diagnosis was achieved via soft voting, aggregating the softmax probabilities of the 11 angular projections per subject. Model selection was further refined using multi-criteria decision-making methods (TOPSIS and VIKOR) to rank configurations based on accuracy, precision, recall, and F1-score. Interpretability was assessed using Grad-CAM heatmaps.

Key Contributions

1. Validation on Real-World Forensic Data: The study validates deep learning architectures on a diverse dataset of actual forensic cases, including cadavers with significant anatomical disfigurements and trauma-related imaging artifacts, rather than idealized or controlled samples.
2. Comparative Analysis of Anatomical Regions: The work provides a comparative analysis of the discriminative power between 2D projections of the pelvis and the skull, evaluating their respective efficacy in sex estimation.
3. Assessment of Generalization: The study assesses the generalization capabilities of transfer learning in the presence of forensic-specific image noise and artifacts, demonstrating the feasibility of the methodology under complex conditions.

Results

• Architecture Performance: Among the baseline models, YOLO26 emerged as the top performer, consistently outperforming other architectures (including ConvNeXt, ViT, and ResNet) with an initial accuracy exceeding 89%.
• Classification Strategy: While binary classification yielded high overall accuracy, the quaternary classification strategy (separating skull and pelvis) demonstrated superior metric equity between sexes. It mitigated the performance disparity caused by the male-heavy dataset, preventing the model from developing a bias toward the majority class.
• Optimal Configuration: Multi-criteria decision analysis (TOPSIS and VIKOR) identified the YOLO26 Nano model with default YOLO data augmentation as the optimal configuration. This configuration achieved the best balance of performance and robustness.
• Performance Metrics: The final optimized model achieved the following patient-level metrics:
  • Accuracy: 95.65%
  • Recall: 92.86%
  • F1-Score: 94.36%
  • Precision: 97.22%
  • Pelvic Performance: The model achieved perfect recall (100%) for both male and female pelvic structures at the patient level.
  • Cranial Performance: Male skulls achieved 100% recall, while female skulls showed lower recall (71.43%) at the patient level, attributed to the concentration of misclassifications in a small subset of borderline female cases.
• Interpretability: Grad-CAM heatmaps confirmed that the models focused on established anthropological landmarks (e.g., subpubic angle, pubic symphysis, glabella, supraorbital margin, and mandibular angle), validating that predictions were grounded in valid anatomical features rather than spurious artifacts.

Significance and Claims
The paper claims that the proposed methodology demonstrates the technical feasibility of using deep learning for objective, high-speed skeletal analysis in forensic contexts. The study concludes that deep learning models can serve as a reliable, reproducible support tool for forensic experts, particularly in cases where traditional morphological assessment is hindered by decomposition or trauma.

The authors emphasize that the system is intended as an assistive instrument, not a substitute for the forensic expert. They note that predictions are probabilistic and require individualized validation. The study acknowledges limitations, specifically that the data was sourced from a single institution and a single CT scanner, necessitating further validation across multiple centers and scanners to assess the generalizability of the approach. The work aims to pave the way for streamlining forensic triage and supporting medical examiners while preserving the expert's decision-making role.