Quantifying regional signal with deep learning for sex estimation on lateral cephalometric radiographs: A feasibility study with explainability analysis.

The purpose of this feasibility study was to quantify regional sex-related information in lateral cephalograms using deep learning. In this retrospective study, 600 radiographs from 300 male and 300 female patients were split at the pterygomaxillary fissure into anterior and posterior sub-images, and DenseNet-121 and Vision Transformer B/16 (ViT-B/16) models were trained by five-fold cross-validation. Posterior sub-images yielded higher performance than anterior sub-images for both models. DenseNet-121 achieved accuracies of 94.5% for anterior images and 97.3% for posterior images, whereas ViT-B/16 achieved 85.7% and 96.7%, respectively. Saliency and hit-rate patterns differed between architectures and were interpreted as indicators of model behavior rather than direct anatomical evidence. Within this single-center dataset, which was skewed toward younger individuals, the posterior field retained substantial sex-discriminative information across both architectures. Anatomy-guided input restriction may provide an intervention-based framework for evaluating regional signal sufficiency. However, age-balanced external validation and finer anatomical segmentation are required.