(PR095) Detection of External Cervical Resorption Using a CNN Model on Intra-oral Radiographs: A Pilot Study

The purpose of this pilot study was to identify a convolutional neural network (CNN) capable of detecting all stages of external cervical resorption (ECR) using intraoral radiographs (IOs) and to assess the accuracy, sensitivity, and specificity of the model. This observational cross-sectional study utilized de-identified intraoral radiographs from Boston University’s database and private dental clinics, including 255 cases of permanent teeth with various stages of ECR and 197 healthy control cases. Two calibrated endodontists detected ECR lesions on IOs, using cone-beam computed tomography (CBCT) as the ground truth for validation. The labeled data was used to train a pretrained ResNet50 model, with the weights of layers 2, 3, and 4 adjusted for training to achieve optimal accuracy. Final hyperparameters included a batch size of 16, a learning rate decay step of 8, a learning rate decay factor of 0.2, a dropout rate of 0.4, and the Adam optimizer. K-fold cross-validation was employed to enhance the model’s generalization performance. The CNN model achieved an average accuracy of 60%, with peak performance of 68% after code refinement and retraining. Accuracy was calculated by dividing the sum of true positives and true negatives by the total cases. Model performance was evaluated using sensitivity, specificity, a confusion matrix, and area under the curve (AUC) metrics. This pilot study demonstrated the potential of the CNN model as a diagnostic tool for automated detection of ECR. However, further refinement is needed to enhance its accuracy, sensitivity, and specificity, improving its clinical utility for early diagnosis and treatment planning of ECR.