Deep Learning Photo Processing for Periodontitis Screening
Leran Tao, Yikai Li, Xinyu Wu, Yuting Gu, Yu Xie, Xiao Yu, Hsueh‐Chou Lai, Maurizio S. Tonetti
Abstract
Late detection of periodontitis has significant health implications. Screening via oral images may serve as an accessible nonclinical method. This study tested the hypothesis that diagnostic information in oral images can aid a deep learning algorithm in detecting periodontitis cases. This cross-sectional diagnostic accuracy study involved consecutive subjects seeking care at Shanghai Ninth People's Hospital, China, and their oral digital twins. The index test was a global activation pooling-based multi-instance deep learning model (DLM) based on pretrained ResNet50, developed and tested in 2 independent samples to identify stage II to IV periodontitis. The model did not use annotated landmarks on images but labeled cases based on a reference consisting of a periodontal clinical examination. The external testing dataset included oral images of subjects diagnosed based on panoramic radiographs. The performance was assessed by the area under the receiver-operating curve (AUROC), sensitivity, and specificity. A total of 387 subjects participated in the internal development and testing. The external testing dataset consisted of 183 subjects. DLM processing of a single frontal view oral image accurately identified stage II to IV periodontitis in the internal (AUROC = 0.93, 95% confidence interval [CI] 0.85-0.98) and external dataset (AUROC = 0.93, 95% CI 0.88-0.96). High consistency was observed between the regions of interest identified in the class activation heat maps and a periodontist (internal test: 99.66%; external test: 99.45%). DLM showed better sensitivity and specificity than clinicians with different skill levels. The multimodal combination of images and other nonclinical parameters led to only marginal improvements in accuracy. DLM processing of oral images shows potential for periodontal health screening. Artificial intelligence focuses on the important image areas but seems to capture features that are not apparent to clinicians. More development and validation are needed to introduce this approach as a screening tool to multiple populations worldwide.