From Diagnosis to Decision-Making: AI-Driven Triage of Oral Mucosal Lesions Detected on Intraoral Photographs.

BACKGROUND

Managing oral mucosal lesions in general dental practice is challenging, as visually similar conditions may require very different treatments. We developed an AI-based triage system connecting intraoral image analysis with guideline-based treatment recommendations, tested it externally, and compared its performance with clinicians' recommendations.

METHODS

YOLOv8 and YOLOv11 were trained on 4079 intraoral photographs to classify 11 oral mucosal lesion types. These were mapped into five treatment tiers defined by urgency and management aim: T1 specialist referral for malignant and high-risk lesions, T2 interdisciplinary referral for specialist-led non-cancer management, T3 risk-adapted surveillance, T4 dental treatment with reassessment, and T5 no treatment. Model performance was compared with two general dentists and four specialists on an internal test set. YOLOv11 was additionally tested on an external dataset.

RESULTS

YOLOv11 outperformed YOLOv8 in macro-averaged precision, recall, and F1 score, achieving 0.72, 0.74, and 0.73 respectively. However, 25% of T1 high-risk cases were under-triaged. YOLOv11 significantly outperformed all clinicians, whose F1 scores ranged from 0.33 to 0.69. External performance decreased markedly (F1 score: 0.58), consistent with domain shift.

CONCLUSION

Guideline-oriented AI triage is feasible but requires multicentre training for improved generalizability, as well as safety-focused optimization, and prospective user evaluation before clinical use.