Three-Dimensional Segmentation of Pulp Stones in Tooth-Level Volumes on Cone-Beam Computed Tomography Using nnU-Net v2: A Preliminary Study

Abstract

Objectives: This study aimed to develop and evaluate an nnU-Net version 2 (nnU-Net v2)-based model for three-dimensional segmentation of pulp stones on cone-beam computed tomography (CBCT) images, with particular emphasis on segmentation performance within preselected tooth-level volumes of interest (VOIs) known to contain pulp stones. Methods: This retrospective methodological study included 90 CBCT examinations in which pulp-stone presence was confirmed according to predefined eligibility criteria. Reference annotations were established through consensus among three experienced observers. Segmentation was initially evaluated on complete CBCT volumes. Subsequently, 238 pulp stone-positive tooth-level VOIs were generated and divided at the VOI level into training and held-out test datasets comprising 215 and 23 VOIs, respectively. Because the allocation was performed at the VOI level, different VOIs from the same patient could be represented in both datasets. Model performance was assessed using voxel-level overlap, classification, surface-distance, and receiver operating characteristic metrics. Results: The full-volume model demonstrated poor segmentation performance, with a Dice similarity coefficient (DSC) of 0.183, Jaccard index of 0.119, precision of 0.254, and recall of 0.205. Within the held-out VOI-level test set of preselected pulp stone-positive tooth VOIs, the model achieved a DSC of 0.637, Jaccard index of 0.485, precision of 0.695, and recall of 0.635. The average surface distance and average symmetric surface distance were 0.256 mm and 0.286 mm, respectively. Voxel-level accuracy was 0.998, and the area under the Receiver Operating Characteristic (ROC) curve was 0.967. However, these classification metrics represented pulp stone-versus-background voxel classification within positive tooth VOIs and did not represent tooth-level diagnostic performance. Conclusions: The nnU-Net v2 model demonstrated moderate segmentation performance within preselected pulp stone-positive tooth-level VOIs, whereas the poor full-volume results showed that reliable automatic localization of pulp stones in complete CBCT examinations remains unresolved. The findings provide preliminary evidence for computer-assisted pulp-stone delineation within preselected tooth volumes but do not establish an end-to-end detection approach or clinical readiness. Further optimization and external validation using both pulp stone-positive and pulp stone-negative teeth are required.