A novel AI-powered radiographic analysis surpasses specialists in stage II–IV periodontitis detection: a multicenter diagnostic study
Yuan Li, Zhiming Cui, Lanzhuju Mei, Yu Xie, Lorenzo Marini, George Pelekos, Wen Gu, Xiaoyu Yu, Xinyu Wu, Xindi Wei, Leran Tao, Ke Deng, Andrea Pilloni, Dinggang Shen, Maurizio S. Tonetti
Abstract
Missed periodontitis diagnoses are common, and AI dental radiography systems based on clinical standards can enhance reliable detection. We introduce and evaluate HC-Net+, a deep-learning model that mimics clinical pathways while integrating localized tooth lesion analyses with broad contextual understanding. This is the first AI model developed from orthopantomograms (OPGs) linked to clinical diagnoses, pre-trained and fine-tuned with 10,881 OPGs, and tested with dual benchmarking against 382 clinically labeled OPGs covering 10,198 teeth and 760 radiographically labeled OPGs from four diverse international centers. It outperformed periodontal specialists' diagnostic accuracy (AUROC: 94.2% vs. 85.6%, p < 0.01). The system significantly improved early periodontitis detection across training and experience levels, enabling junior dentists to match specialist performance with AI support. Performance remained consistent in the multicenter evaluation, achieving >92.4% accuracy across all locations. HC-Net+'s ability to surpass specialist accuracy while making diagnostic expertise more accessible positions it as a transformative tool for precision dentistry. The diagnostic trials were registered at ClinicalTrial.gov (NCT05513599) on 08/23/2022 and (NCT06306677) on 03/12/2024.