YOLOv9c for Reliable Caries Detection on Intraoral Periapical Radiographs

Abstract

 Automated detection of dental caries on intraoral periapical radiographs must couple high sensitivity to tiny, low-contrast radiolucencies with low latency for chairside use. While one-stage detectors have shown promise, YOLOv9 has not been systematically evaluated on periapical radiographs. We benchmark four YOLOv9 variants (e/m/s/c) under a unified, single-class protocol and identify YOLOv9c—which combines a GELAN backbone, training-time PGI, and decoupled multi-scale heads—as the most accurate and computationally efficient choice for this modality. The dataset comprises de-identified periapical radiographs collected at the Sibar Institute of Dental Sciences (Guntur, India), curated under institutional ethics approval with image-wise 80/10/10 splits; images were quality-screened and contrast-normalized prior to training. On a held-out test split, YOLOv9c attains Precision = 0.9830, Recall = 0.9974, F1 = 0.9902, mAP@0.5 = 0.9935, and mAP@0.5-0.95 = 0.8528, outperforming YOLOv9s/m/e and a two-stage Faster R-CNN (ResNet-18) baseline by large margins on both thresholded and strict-IoU criteria. Wilson 95\% confidence intervals show high centers with tight bounds for precision and recall, and pairwise significance tests corroborate statistically reliable gains—especially in recall and tight-overlap localization—indicating that improvements are not sampling artefacts. Architecturally, GELAN’s efficient, detail-preserving aggregation and PGI’s gradient conditioning enhance shallow-stride features critical for subtle lesions, without adding inference cost. These results fill a key gap by providing the first modality-specific evaluation of YOLOv9 on periapicals and establish YOLOv9c as a strong, real-time baseline for clinical deployment and future research on periapical caries detection.