Deep learning-based sex estimation from multi-planar cranial CT images in a Thai population

Developments in deep learning and medical imaging have created new opportunities for automated image analysis in both medical and forensic applications. In forensic investigation, the cranium is widely recognized as a valuable skeletal element for sex estimation. Therefore, this study aimed to develop and evaluate a deep learning-based framework for sex estimation using multi-planar cranial CT images in a Thai population. A total of 250 cranial CT datasets (125 males and 125 females) obtained from the Osteology Research and Training Center (ORTC), Faculty of Medicine, Chiang Mai University, were analyzed. Sagittal, coronal, and horizontal CT images were reconstructed from each dataset. A ResNet-18 architecture was used for image classification. For each imaging plane, datasets were randomly divided into 80% training and 20% validation subsets. An independent blind test set consisting of 26 additional CT datasets (13 males and 13 females) was used for external validation and model performance. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to evaluate model interpretability. Among the three orthogonal planes, sagittal cranial CT images achieved the highest classification performance, with an accuracy of 96% in the validation set and 92.31% in the independent blind test set. Moreover, Grad-CAM analysis demonstrated that the convolutional neural network focused on cranial regions corresponding to established sexually dimorphic traits used in forensic anthropology. These findings demonstrate the potential of deep learning-based cranial CT analysis for automated forensic sex estimation and may serve as a reproducible decision-support tool for forensic anthropological applications.