Development of a 3D Convolutional Neural Network for the Triage of High-Priority Oral and Maxillofacial CBCT Scans

This project aims to establish a triage system for oral and maxillofacial cone-beam computed tomography (CBCT) scans by developing a neural network to identify high-priority cases. Two hundred scans and reports were reviewed to form two cohorts: Group A included 100 patients with significant oral and maxillofacial findings in the oral and maxillofacial region, and Group B included 100 patients without major findings. Scans with only minor abnormalities were assigned to Group B. CBCT images from the groups were merged and split into training (70%) and validation (30%) sets. Two 3D convolutional neural network (3D-CNN) models were developed using Python and Keras: Model 1, inspired by Zunair et al., and Model 2, based on a modified VGG-16 architecture. Data augmentation, the Adam optimizer, and early stopping were applied during training. Each model was trained five times with randomized dataset shuffling, and performance was evaluated using ROC-AUC, accuracy, precision, recall, and F1-score. A Welch's t-test was used to compare model performance. Both models had excellent performance with Model 2 achieving a slightly higher mean ROC-AUC of 0.918 on validation with no statistically significant inter-model performance (p = 0.078). Model 2 featured advanced architecture with four convolutional blocks and dense layers, optimizing accuracy (0.960) and precision (0.962) in training. In this study, we successfully developed an accurate 3D-CNN based on CBCT images to distinguish between high-priority scans and routine reports. To the best of our knowledge, this is the first study to address this specific task.