Conclusion

This study provides the first systematic comparison of YOLOv8-hybrid and nnU-Net v2 systems for maxillary sinus-based biological profiling in forensic anthropology. YOLOv8 is suitable for clinical screening with high consistency (mAP@50 = 98.19%, CV = 0.77%). nnU-Net v2 should be preferred for forensic reporting with superior classification performance (AUC = 0.927, DeLong p = 0.024, Cohen’s d = 0.48). Both systems provided robust age estimation (MAE ≈ 7.2 years). Bilateral area difference is the most determinant feature (sex: 0.42, age: 0.51). Inter-observer reliability was excellent (ICC = 0.94–0.97), and external validation confirmed generalizability.

With robust methodology (patient-level splitting, 5-fold CV, transfer learning, Bayesian optimization, bootstrap 95% CI, SHAP, dual-annotator verification) and > 120 radiomic/morphometric features, this study proposes a comprehensive methodological framework for automated forensic biological profiling. Multi-center prospective validation in diverse ethnic populations is warranted.