A deep-learning system for dental plaque quantification from disclosed intraoral photographs, validation against clinical plaque scoring.

OBJECTIVES: To determine whether buccal plaque coverage measured automatically from disclosed intraoral photographs agrees with the Turesky-modified Quigley-Hein plaque index (T-QHPI) scored by calibrated dentists. METHODS: Twenty adults were photographed in one frontal and two lateral views after 48 hours without oral hygiene, and a calibrated dentist, blind to the model, scored the T-QHPI on the same teeth. Two U-Net (ResNet-34) models, one per view, were trained under participant-level five-fold cross-validation, and coverage was the stained percentage of each outlined buccal crown. Coverage was compared with the clinical grade by Spearman's ρ and, after conversion to T-QHPI grades under index-definition and calibrated thresholds, by weighted κ; participant means by Pearson's r, the intraclass correlation coefficient ICC(2,1) and Bland-Altman analysis. RESULTS: 438 teeth were compared with the clinical grade. Segmentation Dice was 0.88 in both views. Coverage ranked teeth with the clinical index at ρ = 0.856 (95% CI 0.821 to 0.877). After calibration, weighted κ was 0.812 (0.766 to 0.844) against 0.791 (0.710 to 0.853) between two examiners, and agreement within one grade was 96.8%. Participant means correlated at r = 0.930, with ICC(2,1) 0.916. CONCLUSIONS: Automated buccal plaque coverage from disclosed photographs agreed with the T-QHPI scored by a calibrated examiner, and after threshold calibration the agreement did not differ from that between two examiners. CLINICAL SIGNIFICANCE: Automated plaque grading from disclosed photographs could serve as an auxiliary tool, a second opinion that supports, rather than replaces, the clinician's plaque scoring, pending external validation.