A Fused Deep Learning Architecture for the Detection of the Relationship between the Mandibular Third Molar and the Mandibular Canal

The study aimed to generate a fused deep learning algorithm that detects and classifies the relationship between the mandibular third molar and mandibular canal on orthopantomographs. Radiographs (n = 1880) were randomly selected from the hospital archive. Two dentomaxillofacial radiologists annotated the data via MATLAB and classified them into four groups according to the overlap of the root of the mandibular third molar and mandibular canal. Each radiograph was segmented using a U-Net-like architecture. The segmented images were classified by AlexNet. Accuracy, the weighted intersection over union score, the dice coefficient, specificity, sensitivity, and area under curve metrics were used to quantify the performance of the models. Also, three dental practitioners were asked to classify the same test data, their success rate was assessed using the Intraclass Correlation Coefficient. The segmentation network achieved a global accuracy of 0.99 and a weighted intersection over union score of 0.98, average dice score overall images was 0.91. The classification network achieved an accuracy of 0.80, per class sensitivity of 0.74, 0.83, 0.86, 0.67, per class specificity of 0.92, 0.95, 0.88, 0.96 and AUC score of 0.85. The most successful dental practitioner achieved a success rate of 0.79. The fused segmentation and classification networks produced encouraging results. The final model achieved almost the same classification performance as dental practitioners. Better diagnostic accuracy of the combined artificial intelligence tools may help to improve the prediction of the risk factors, especially for recognizing such anatomical variations.