Automated Identification of Accessory Mental Foramen Using Cone-Beam Computed Tomography and Convolutional Neural Networks

To develop and evaluate a deep learning-based system for automatic detection of the accessory mental foramen (AMF) using cone-beam computed tomography (CBCT) images, and to compare the detection accuracy and clinical reliability performance of two convolutional neural network (CNN) architectures for this model. A total of 3000 CBCT scans were retrospectively screened. After expert evaluation, 700 CBCT scans exhibiting AMFs were identified. For comparative analysis, 700 CBCT scans with normal mental-foramen anatomy were selected as the matched control group. A custom lightweight CNN and a ResNet-50 model were trained for binary classification of AMF presence. Model performance was evaluated by determining accuracy, precision, recall, and the F1-score. Gradient-weighted class activation mapping (Grad-CAM) visualisation was employed to assess the anatomical relevance of the models’ attention maps. Statistical analyses were performed to compare the diagnostic performance of the two networks. The ResNet-50 model achieved superior performance (overall accuracy: 85.8% for ResNet-50 vs 71.1% for the custom CNN). With the ResNet-50 model, anomaly recall improved from 0.68 to 0.88, reducing missed detections by 63%. Grad-CAM analysis demonstrated that the models focused primarily on anatomically valid regions around the MF, confirming the interpretability and clinical relevance of the models. Automatic detection of the AMF using CBCT and deep learning represents a reliable, objective, and efficient diagnostic approach that minimises observer bias and enhances clinical decision-making. Deep learning-based detection of AMFs on CBCT can enhance diagnostic accuracy and reduce the risk of surgical complications by providing consistent, observer-independent evaluations.