Three-dimensional automatic segmentation of root canals with focus on the second mesiobuccal canal using nnU-Netv2 on CBCT images: deep learning approach

Artificial intelligence (AI) has the potential to reduce interpretation errors and save time during the evaluation of cone beam computed tomography (CBCT) images. This study aimed to assess the performance of AI in identifying and segmenting the second mesiobuccal canal (MB2), with concurrent segmentation of the main root canals, in the maxillary first molar prior to endodontic treatment. In this study, 202 CBCT images that met the inclusion criteria were obtained from an anonymized database provided by Craniocatch (Eskişehir, Türkiye), with no associated personal data. The nnU-Netv2 model implemented with the PyTorch library was used for the detection and three-dimensional (3D) automatic segmentation of root canals. Owing to the narrow structure of the MB2 canal, labels were preprocessed via binary dilation with SciPy (v1.10.1), and training was conducted in two stages by applying different dilation levels. The performance of the artificial intelligence model was evaluated via the confusion matrix and further assessed with additional metrics, including the Dice score (DC), Jaccard index (JI), 95% Hausdorff distance (HD), and area under the curve (AUC). In this study, the nnU-Netv2 model achieved a sensitivity of 0.538, a precision of 0.719, a DC of 0.616, a JI of 0.445, a 95% HD of 0.874, and an AUC of 0.8 for 3D automatic segmentation of MB2. This study is the first to apply the nnU-Netv2 model for 3D automatic segmentation of the MB2 canal in untreated teeth and highlights its potential utility in endodontic imaging. Further refinements in these systems may enable rapid and reliable 3D automatic segmentation of MB2 and enhance endodontic treatment quality and patient outcomes.