Automated Detection of Middle Mesial Canals in Mandibular Molars on CBCT using nnU-Net: A Retrospective Diagnostic Accuracy Study.

BACKGROUND

This study aimed to develop and validate a fully three-dimensional (3D) convolutional neural network for automated detection of the mid-mesial canal (MMC) in mandibular molars on cone-beam computed tomography (CBCT).

METHODS

In this retrospective diagnostic accuracy study, 248 CBCT volumes were used for model development, with an independent test set of 28 cases (1.611 axial slices). Ground-truth MMC status was defined by endodontist-radiologist consensus; multiclass manual masks (dentin, mesiobuccal-, mesiolingual-, and mid-mesial-canal) were created in 3D Slicer. A 3D nnU-Net-based model was trained using canal-focused patch sampling and a combined Dice plus weighted cross-entropy loss. MMC detection was evaluated at slice and case levels using sensitivity, specificity, accuracy, positive predictive value/negative predictive value, and Cohen's κ; paired comparisons used McNemar's test.

RESULTS

Interobserver agreement for slice-level MMC labeling was excellent (κ = 0.892). For case-level MMC detection using a prespecified threshold (≥5 AI-positive slices), performance was sensitivity 94.7%, specificity 100%, accuracy 96.4%, and κ = 0.920. At slice level, agreement remained high versus individual observers (κ ≈ 0.80) and consensus references (κ = 0.812 for OR; κ = 0.792 for AND). No significant difference was observed between AI and the OR-consensus reference at slice level, nor between AI and the case-level ground truth.

CONCLUSIONS

A fully 3D nnU-Net framework achieved clinician-level MMC detection on CBCT and may support standardized, efficient identification of this frequently overlooked anatomical variation.