High Dice, Broken Canals: Progressive Model Assisted Annotation and Novel Post-Processing for Continuous Mandibular Canal Segmentation in CBCT

Accurate localization of the inferior alveolar nerve (IAN) within the mandibular canal is essential for safe dental implant planning. Most existing canal segmentation methods tend to produce discontinuous segmentation masks, often leaving gaps along the predicted nerve pathway. We used a progressive, model-assisted labelling process to create our training dataset. The trained model achieved a Dice of 0.984 on an independent test dataset of 60 CBCT volumes. However, the model predicted discontinuous masks in 14 out of 60 samples, indicating that Dice alone does not reflect structural quality. Expanding the training set raised overall Dice but did not help the most challenging cases, which motivated a dedicated post-processing refinement method. The proposed novel refinement method restored continuity by tracing minimum-cost paths through the network's own probability field, for bridging the gaps. As a result, the percentage of continuous canal sides rose from 85.0\% to 97.5\% in the Test dataset, with a mean bridge centerline difference of 0.14 mm and no regression on correct segmentations. At the same time, its conservative safety rule refrained from connecting genuinely disrupted canals and flagged them for review.