Multi-task learning for forensic age and sex estimation using large foundation models on pelvic radiographs: a multi-center study.

Age and sex estimation are essential tasks in forensic science, law enforcement and disaster victim identification. Previous methods hardly implement the two tasks simultaneously due to limitations of models and the complexity of the multitask. We propose a novel framework based on large foundation model (LFM) to conduct multi-task estimation workflow to address the problem. In this study, large-scale pelvic radiographs (6712 samples) aged from 6.00 to 30.99 years from three independent medical centers were collected and evaluated. The method achieved mean absolute error (MAE) of 0.569 (95% CI: 0.540-0.598) years and root mean squared error (RMSE) of 0.749 in age estimation, with an MAE of 0.564 (95%CI: 0.522-0.606) for males and 0.575 (95%CI: 0.535-0.615) for females. For sex estimation, an accuracy of 0.905 and F1 score of 0.904 were achieved, with accuracy ranging from 0.790 to 1.000 across age groups. The proposed framework demonstrated competitive performance compared with the evaluated deep learning and manual methods on the studied datasets. The Grad-CAM heatmaps show that the model's regions of interest (ROIs) mainly concentrate on bony structures including the pubic symphysis, iliac crest, greater trochanter of the femur, and pelvic shape, providing preliminary interpretability evidence for the model. The proposed LFM-based framework demonstrates promising performance for multi-task age and sex estimation and may serve as a supportive tool for forensic assessment, subject to further validation in broader populations and real-world forensic settings.