Sex estimation from lateral cephalograms via a hybrid multimodel convolutional neural network.

Sex estimation represents a fundamental step of human identification in forensic anthropology, archaeology, and forensic medicine. Lateral cephalograms capture craniofacial morphology that is useful for sex estimation. This study developed a hybrid convolutional neural network (CNN) that combines supervised DenseNet169 and unsupervised EfficientNetB3 with a random forest classifier for automate sex estimation from lateral cephalograms. The dataset comprised 150 cephalograms divided into training (69.33%), validation (20%), and testing (10.67%) subsets. DenseNet169 was trained on annotated images to detect five craniofacial landmarks: nasion (N), sella (S), glabella (G), basion (Ba), and menton (M), and compute linear and triangulation angles-based measurements for sex estimation. EfficientNetB3 integrated with a random forest classifier was trained on unannotated images. The final predictions were determined by majority voting among linear and triangulation angles measurements from DenseNet169 and image-based classification from EfficientNetB3. DenseNet169 achieved 100% accuracy with an Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) of 0.935 for linear measurements and 99.69% accuracy (AUC 0.891) for triangulation angles measurements, while EfficientNetB3 achieved 80.63% accuracy (AUC 0.826). The hybrid multimodel CNN demonstrated enhanced robustness, achieving 99.69% accuracy (AUC 0.978) and 97.83% accuracy (AUC 0.947) on external data. These findings highlight the potential of the proposed hybrid framework for automated sex estimation using lateral cephalograms.