Automated Classification of Orthodontic Treatment Need Based on PAR Estimated from Intraoral Photographs: A Multi-View Deep Learning Framework Benchmarked Against Human Inter- and Intra-Rater Reliability

Background: The Peer Assessment Rating (PAR) index quantifies malocclusion severity and is widely used to determine orthodontic treatment outcome. Manual scoring of estimated PAR components on intraoral photographs is time-consuming and subject to substantial inter-examiner variability, and automation is further complicated by the need to integrate five disparate clinical views. This study developed and evaluated a deep-learning framework for automated classification of orthodontic treatment need based on estimated PAR from routine intraoral photographs. Methods: A total of 2965 photographs from 593 patient visits at a single centre were scored by a board-certified orthodontist using the weighted PAR index and dichotomised at PAR > 10. The model was trained and evaluated as a binary treatment-need classifier on this dichotomised label, not as a regressor of the continuous PAR score or its individual components. A self-supervised foundation model isolated the dental region; the resulting views were embedded and aggregated by an attention-based multiple instance learning layer to yield a visit-level prediction. A held-out set of 80 visits was rescored independently by a second orthodontist and again by the first after 6 weeks. Agreement was quantified with Cohen’s κ; 95% confidence intervals (CIs) and paired differences (Δκ) were estimated by bootstrap (2000 resamples). The final model was compared against input-, early-, and late-fusion baselines. Results: The model achieved an AUROC of 0.82. Agreement with the training annotator (an optimistic reference) was κ = 0.55 (95% CI 0.36–0.72) and with the independent rater (the conservative reference) κ = 0.41 (0.22–0.60), compared with κ = 0.52 between the two orthodontists and κ = 0.70 within the same rater across sessions. Agreement with the independent rater did not differ significantly from human inter-rater agreement. Conclusions: Automated classification of treatment need based on photograph-estimated PAR reached moderate agreement with orthodontists. The model reproduces an expert’s photograph-based estimate and is not a clinically validated assessment of treatment need.