Deep learning for sex estimation from cranial CT scout images

Sex estimation is a key component of biological profiling in forensic science, and routinely acquired computed tomography (CT) scout images may contain cranial information for automated assessment. This retrospective study evaluated deep learning-based sex estimation from cranial CT scout images in 764 patients (336 females, 428 males). Separate EfficientNet-B0 models were trained for anteroposterior (AP) and lateral scout projections after standardized preprocessing. Performance was assessed in an independent internal test set of 115 patients, and the lateral model was evaluated externally in 138 patients from another institution. On internal testing, the AP model achieved an accuracy of 0.939 and ROC-AUC of 0.985, whereas the lateral model achieved an accuracy of 0.931, precision of 0.983, recall of 0.892, and ROC-AUC of 0.992, with no significant difference between projections (McNemar test, p > 0.05). In external validation, the lateral model achieved an accuracy of 0.805, precision of 0.743, recall of 0.927, and ROC-AUC of 0.902. Grad-CAM visualizations indicated frontal and parieto-occipital regions contributing to predictions, but these findings were interpreted cautiously. These results suggest that cranial CT scout images may provide complementary information for deep learning-based sex estimation, although broader multicenter validation is required before forensic implementation.