Sex estimation from hand-wrist radiographic images: an artificial intelligence approach.

Sex estimation is a fundamental component of forensic identification. This study aimed to evaluate the performance of three state-of-the-art deep learning architectures-YOLOv5, YOLOv8, and YOLOv11-for automated sex classification using hand-wrist radiographs of individuals aged 8-18 years. A total of 1,800 hand-wrist radiographs were retrospectively collected and categorized into male and female classes based on verified hospital records. Images were preprocessed, resized to 224 × 224 pixels, and divided into training, validation, and test subsets. Each YOLO architecture was trained using identical hyperparameters and evaluated using accuracy, precision, recall, F1-score, confusion matrices, and area under the ROC curve (AUC). Performance varied substantially among models. YOLOv5 showed limited discriminative ability, achieving an F1-score of 0.69 and an AUC of 0.61, indicating insufficient feature extraction for sex-related skeletal cues. YOLOv8 demonstrated near-perfect performance (F1-score: 0.99, AUC: 0.99) with stable training behavior and balanced classification across both sexes. YOLOv11 also achieved excellent results, with F1-scores of 0.98 for both classes and an AUC of 0.98, reflecting strong sensitivity to subtle morphological dimorphism. Both YOLOv8 and YOLOv11 outperformed YOLOv5 across all metrics, demonstrating superior feature extraction capacity and robust convergence. The findings indicate that next-generation YOLO architectures (YOLOv8 and YOLOv11) enable highly accurate and fully automated sex estimation from hand-wrist radiographs without the need for manual annotation or region-of-interest selection. The proposed pipeline provides a fast and reproducible approach that may support sex estimation in selected forensic cases when preferred anatomical regions are unavailable.