(PR064) Artificial Intelligence-Assisted Detection of Second Mesiobuccal Canals in Cone-Beam Computed Tomography Images Using MultiResUNet

Objectives - The detection of second mesiobuccal (MB2) canals in maxillary molars presents a challenge due to their variable anatomy and frequent under-detection in conventional imaging. This study aims to train a MultiResUNet deep learning model for detection of MB2 canals in segmented cone-beam computed tomography (CBCT) images of maxillary first molars. Methods - CBCT images of 15 maxillary first molars with unobturated MB2 canals and 15 maxillary first molars without MB2 canals were included. Segmentation of the MB roots was performed, generating 90 segmentations in total, with 3 segmentations per tooth. For each tooth, two segmentations were randomly selected, and MATLAB's image labeler was used to label MB canals in these 60 segmentations to train the MultiResUNet deep learning model. The remaining 30 segmentations were used as testing samples to evaluate the model's accuracy. Additionally, the average prediction time per segmentation was measured.

Results: The model reports an overall accuracy of 0.92. It also achieved sensitivity of 0.85, specificity of 1, PPV of 1 and NPV of 0.87 in identifying the MB2 canal. Notably, our model achieved this with an average processing time of 0.088 seconds per segmentation, demonstrating its feasibility for real-time use.

Conclusion: The MultiResUNet deep learning model shows promise in detecting MB2 canals in segmented CBCT images, combining high accuracy with real-time performance. This approach could enhance diagnostic precision and streamline workflows in endodontic practice. Future studies with larger datasets and fully automated segmentation are needed to facilitate its clinical applicability.