to evaluate the performance of an nnU-Net v2–based architecture for the voxel-level detection, lesion-wise detection and 3D segmentation of periapical lesions on CBCT images. In this retrospective study, 150 CBCT scans acquired for various clinical indications were initially evaluated. 10 scans met the exclusion criteria, leaving 140 CBCT scans that were annotated for the presence of periapical lesions. Periapical lesions were manually segmented on a web-based platform by a trained researcher and verified by an experienced endodontist to obtain the ground-truth segmentations. An nnU-Net v2 model, configured in the 3D full-resolution setting, was trained on 80% of the data, validated on 10%, and tested on the remaining 10%. Segmentation performance was assessed at the voxel-wise using Dice coefficient (DC), Intersection over Union (IoU, Jaccard), 95% Hausdorff distance (HD95), sensitivity (recall), precision metrics. Mean, standard deviation (SD), median and interquartile range (IQR) across the 14 test scans were reported, and 95% confidence intervals (CIs) were estimated using a parametric t-distribution. A lesion-wise detection performance was assessed at the root level using sensitivity, precision and F1-score; results were stratified by lesion volume. On the test set, segmentation performance at voxel-level was DC 0.74 (SD 0.14; 95% CI 0.67–0.81), IoU 0.60 (SD 0.15; 95% CI 0.53–0.68), HD95 12.02 mm (SD 18.85 mm; 95% CI 2.15–21.90 mm), sensitivity 0.73 (SD 0.18; 95% CI 0.63–0.82), precision 0.80 (SD 0.11; 95% CI 0.75–0.86). and AUC was 0.86. Lesion-wise detection performance was sensitivity 0.75; precision 0.82, and the lesion-wise F1-score was 0.78. The nnU-Net v2–based model achieved moderate-to-good three-dimensional segmentation performance and good detection perfomance for periapical lesions on CBCT under routine clinical imaging conditions, with high discriminative ability at the voxel level. In its current form, the model appears suitable as an adjunctive tool to support expert interpretation rather than as a stand-alone decision-making system. Not applicable.
Öksüzoğlu et al. (Mon,) studied this question.