A two-stage deep learning approach for forensic age and sex determination from hand-wrist radiographs: an anatomical pose estimation and multi-task model

Biological profiling in disaster victim identification (DVI) scenarios is challenged by soft tissue decomposition and skeletal fragmentation, limiting traditional anthropological methods. This study developed a two-stage deep learning system for automated skeletal age and sex estimation from hand radiographs, designed to function independently of soft tissue presence. A two-stage architecture integrating YOLOv8x for anatomical pose estimation (Stage 1: 17 keypoints, mAP50-95: 99.3%; 19-bone segmentation, Dice > 92%) and EfficientNet-B0 for biological profiling (Stage 2) was developed. Training utilized 13,865 radiographs from RSNA Pediatric Bone Age Challenge and Roboflow HandBone datasets, with aggressive regularization (dropout 0.65, L2 0.05) to ensure population generalizability. Performance was evaluated on 200 hold-out samples. Age estimation achieved MAE 7.75 months (95% CI: 6.82–8.68), R² 0.942, and ICC 0.993, comparable to expert-level inter-rater agreement. Sex classification demonstrated 94% accuracy (Cohen’s κ: 0.88, “almost perfect” agreement). Age-stratified analysis confirmed consistent performance across 0–19 years (MAE < 10 months all groups). Bland-Altman analysis revealed negligible systematic bias (Cohen’s d: 0.128, p = 0.060). The segmentation module’s isolation of skeletal structures validates the method’s independence from soft tissue artifacts. This forensically-oriented system provides objective, reproducible biological profiles from hand radiographs without relying on soft tissue integrity. The two-stage architecture enhances court admissibility through anatomical interpretability while maintaining expert-level accuracy, offering potential utility in DVI operations involving decomposed or skeletonized remains.