Uncertainty-aware ordinal deep learning for automated periapical index assessment on dental radiographs

Abstract

Abstract Automated radiographic assessment of periapical health may support scalable review, but the five-grade Periapical Index (PAI) is ordinal and should not be treated as nominal classification. We present PAI-SegScore as a single-cohort framework developed on DenPAR radiographs with an endodontic expert-verified reference standard. The frozen cohort comprised 550 training, 98 validation, and 200 locked-test radiographs, containing 2,451, 434, and 907 annotated apices, respectively. The pipeline separates class-agnostic apex detection from apical-crop grading. Six matched grading conditions (C1–C6) were evaluated across five seeds; C6 added grade-agnostic SAM2 pseudo-mask features generated from image pixels and the apex box only, without PAI -label input. On oracle crops, C5 achieved accuracy $$0.6456\pm 0.0137$$ , macro-F1 $$0.4988\pm 0.0263$$ , QWK $$0.5603\pm 0.0438$$ , and MAE $$0.4551\pm 0.0221$$ ; C6 achieved $$0.6520\pm 0.0268$$ , $$0.5073\pm 0.0328$$ , $$0.5781\pm 0.0518$$ , and $$0.4509\pm 0.0490$$ , respectively. The common Stage 1 detector matched 792 of 907 reference apices at IoU 0.30 (recall, 87.32%); complete-pipeline grading used these same predicted boxes for C5 and C6. The ablation produced small, metric-specific differences rather than general ordinal superiority. PAI-SegScore remains a radiographic decision-support research framework requiring clinical correlation and prospective validation; C6 pseudo-masks are auxiliary feature support, not clinical lesion or periodontal-ligament segmentations.