Enhancing forensic sex identification through AI-based analysis of the foramen magnum

Sex estimation from skeletal remains is an essential task in forensic anthropology. Traditional morphological analysis, while effective, can be time-consuming and subject to inter-observer variability. This study evaluates artificial intelligence (AI)-based methods, specifically object detection and instance segmentation, for sex estimation using the foramen magnum (FM). A total of 600 adult dry skull images (300 males, 300 females) were labeled and augmented to create a dataset of 2280 images, which was split into training (92 %), validation (5 %), and test (3 %) sets. The models were trained using Roboflow and assessed based on sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV), with additional validation performed on 30 independent skulls. The object detection model demonstrated strong performance, achieving high precision (95.0 %) and recall (100.0 %) in training, with precision values of 93.0 % and 89.0 % in validation and test sets, respectively, while maintaining 100.0 % recall across datasets. In the independent test set, the model achieved 75.0 %specificity. The instance segmentation method yielded lower performance, with specificity of 68.75 %. The overall accuracy of the object detection method was 65.68 % (95 % CI: 46.19 % - 81.64 %), outperforming the instance segmentation method, which achieved an accuracy of 62.69 % (95 % CI: 43.22 % - 79.55 %). Although AI-based methods, particularly object detection, show potential for forensic sex estimation from foramen magnum, the results indicate that their accuracy remains lower than traditional morphometric approaches. Future research should focus on integrating additional cranial features and expanding the training dataset to enhance model reliability and generalizability.