Automated Detection and Ordinal Severity Grading of Periapical Lesions on Panoramic Radiographs Using a Lesion-Preserving Tiling and Ordinal-Aware YOLO Framework.

Background/

Objectives

Apical periodontitis is among the most prevalent dental pathologies, and its radiographic severity is graded on panoramic radiographs using the ordinal periapical index (PAI 3 < 4 < 5). Existing computer-aided systems typically treat these grades as unrelated classes and down-sample the radiograph to a fixed input, which erases the small periradicular radiolucencies that define early disease. This study evaluated a lesion-preserving, ordinal-aware YOLO framework for the simultaneous detection and severity grading of periapical lesions.

Methods

A total of 3924 expert-annotated panoramic radiographs (PAI 3/4/5) were processed with an adaptive lesion-preserving tiling algorithm that mapped each radiograph onto native-resolution 640 × 640 tiles so that every lesion appeared intact within exactly one tile, yielding 4641 tiles. Six consecutive YOLO generations (YOLOv8, YOLOv9, YOLOv10, YOLO11, YOLO12, and YOLO26) were benchmarked under one identical pipeline on a locked hold-out test set of 985 lesions, of which 628 were PAI 3, 274 PAI 4, and 83 PAI 5. YOLO12m, which attained the highest point estimates and the smallest run-to-run variation, was then progressively augmented with an ordinal-aware detection loss, a high-resolution P2 head, severity class-weighting and test-time augmentation. Localization was measured with COCO mAP and severity grading with severity-weighted average precision (Sev-wAP), quadratic weighted kappa (QWK), a catastrophic off-by-two error rate, and per-grade recall.

Results

Lesion-preserving tiling improved detection for every generation, raising the mAP@0.50 of YOLO12m from 0.494 to 0.615, the highest point estimate among the six generations, although the margin over the next two was within the resolution of this test set. Cumulatively adding the ordinal loss, the P2 head, severity weighting, and test-time augmentation raised mAP@0.50 to 0.664, mAP@0.50:0.95 to 0.333, Sev-wAP to 0.704, and QWK to 0.780. PAI 5 average precision rose from 0.268 to 0.331 and PAI 5 recall from 0.602 to 0.663, 55 of 83 severe lesions, while catastrophic confusion between mild and severe lesions stood at 0.41%, 4 of the 985 reference lesions.

Conclusions

Coupling lesion-preserving tiling with ordinal severity-aware detection improves localization and severity-weighted precision by resolvable margins, shifts every severe-grade measure in a clinically preferable direction, and provides a reproducible reference protocol for comparing detector architectures on this task.