Classifying Sex from MSCT-Derived 3D Mandibular Models Using an Adapted PointNet++ Deep Learning Approach in a Croatian Population.

Abstract

Accurate sex estimation is critical in forensic anthropology for developing biological profiles, with the mandible serving as a valuable alternative when crania or pelvic bones are unavailable. This study aims to enhance mandibular sex estimation using deep learning on 3D models in a southern Croatian population. A dataset of 254 MSCT-derived 3D mandibular models (127 male, 127 female) was processed to generate 4096-point clouds, analyzed using an adapted PointNet++ architecture. The dataset was split into training (60%), validation (20%), and test (20%) sets. Unsupervised analysis employed an autoencoder with t-SNE visualization, while supervised classification used logistic regression on extracted features, evaluated by accuracy, sensitivity, specificity, PPV, NPV, and MCC. The model achieved 93% cross-validation accuracy and 92% test set accuracy, with saliency maps highlighting key sexually dimorphic regions like the chin, gonial, and condylar areas. A user-friendly Gradio web application was developed for real-time sex classification from STL files, enhancing forensic applicability. This approach outperformed traditional mandibular sex estimation methods and could have potential as a robust, automated tool for forensic practice, broader population studies and integration with diverse 3D data sources.