Automated ICON scoring from orthodontic photographs: a deep learning-based model that learns from the clinical assessment of three orthodontists

Abstract Background Manual scoring of the Index of Complexity, Outcome and Need (ICON) is laborious and observer-dependent. We developed a deep-learning model that learns from three orthodontists’ ICON assessments to classify treatment need from routine clinical photographs. Methods In this retrospective single-centre study, 565 patients (2018–2025) were scored independently by three orthodontists. Five standard photographs per patient (frontal, two lateral, two occlusal) were encoded by a DINOv3-Large Vision Transformer with low-rank adaptation and fused by a two-layer Transformer encoder. Patients were split into a held-out test set ( n = 85) and a development set ( n = 480) used for five-fold cross-validation with two seeds. The decision threshold was pre-specified; 95% confidence intervals (CI) were computed by 2,000-iteration bootstrap. Results Inter-rater agreement at the ICON ≥ 43 cut-off was fair (Fleiss κ = 0.262). On the held-out set, the model achieved an AUC of 0.901 (95% CI, 0.829–0.958) and average precision 0.965 (0.942–0.985); a permutation test rejected chance ( P < 0.001). At the pre-specified threshold, specificity was 1.000 (Wilson CI, 0.857–1.000) but sensitivity was only 0.435 (0.319–0.559). The model also ranked continuous ICON scores (Pearson r = 0.75), indicating that graded severity was captured, although its probability outputs were compressed and were not calibrated to a fixed cut-off. Conclusions A self-supervised Vision Transformer reliably ranks patients by ICON treatment need from routine photographs. Low sensitivity at any fixed cut-off limits screening use; the model is best framed as a referral-prioritisation aid pending multicentre validation.