EARLY VS. DEEP DENTAL CARIES DETECTION ON PANORAMIC RADIOGRAPHS: A CRITICAL ANALYSIS OF DEEP LEARNING PERFORMANCE

Automated caries detection on panoramic radiographs is usually reported with a single aggregate score, although early and deep lesions present very different detection problems: deep lesions form large, high-contrast radiolucencies, whereas early lesions cause subtle density changes close to the resolution limit of panoramic imaging. This analytical report re-examines a two-stage deep learning pipeline published earlier by the author [18]: a YOLOv8m detector localizes individual teeth, and a UNet++ model with a pretrained ResNet50 encoder segments early and deep caries on the resulting crops. The pipeline was trained and evaluated on the DC1000 dataset (593 panoramic radiographs, approximately 17,399 tooth crops). No new experiments were performed; the published results are analyzed separately for each lesion class. Deep caries reached an IoU of 0.635 (Dice 0.777, precision 82.6%, recall 73.3%), compared with 0.279 for early caries (Dice 0.436, precision 38.2%, recall 50.9%), a 2.3-fold gap. The error profiles also differ: early caries were over-segmented (predicted area about 1.33 times the reference), whereas deep caries were slightly under-segmented (about 0.89 times). The benefit of the two-stage decomposition was strongly class-dependent: relative to the best single-stage result for each class, it raised deep-caries IoU by 64% but early-caries IoU by only 17%, and a recall-oriented loss function improved deep-caries IoU by 85% while lowering early-caries IoU by 28%. The aggregate mean IoU of 0.457 and tooth-level F1-score of 74.5% therefore average two very different regimes and say little about early lesions, which are the most valuable target for screening. Early-caries detection on panoramic radiographs remains an open problem, and future studies should report severity-stratified performance.