Deep learning for automated instance segmentation of residual roots in panoramic radiographs using mask R-CNN: A retrospective diagnostic accuracy study

Wenhui Li, Yaqin Chang, Rui Li, Lu Mm, Jiao Zhang

Background: This study aimed to evaluate the efficacy of Mask regions with convolutional neural network (R-CNN) for the automated detection and segmentation of residual dental roots in panoramic radiographs and to compare its diagnostic performance against the semantic segmentation benchmark, U-shaped network (U-Net). Methods: A retrospective dataset comprising 224 patients with 505 annotated residual roots was utilized. Image preprocessing involved adaptive contrast enhancement using contrast limited adaptive histogram equalization in the Commission Internationale de l’Éclairage lab color space to improve root-to-bone definition. A Mask R-CNN model utilizing a ResNet-50 backbone and Feature Pyramid Network was trained using K-fold cross-validation. Performance was compared to U-Net based on sensitivity, specificity, accuracy, dice similarity coefficient, and receiver operating characteristic analysis. Result: The Mask R-CNN model significantly outperformed U-Net across all evaluated metrics. It achieved an accuracy of 98.67% and a dice similarity coefficient of 91.34%. Most notably, the model demonstrated a sensitivity of 91.16%, presenting a marked improvement over U-Net (78.54%), while maintaining a specificity of 99.12%. The area under the curve was calculated at 0.9599, indicating superior discriminative capability. Conclusion: Mask R-CNN provides a robust solution for identifying residual roots, effectively addressing challenges related to low contrast and anatomical noise. By combining high sensitivity with high specificity, the system significantly reduces false negatives without causing alert fatigue, thereby serving as a reliable automated assistant for enhancing surgical safety and planning efficiency.