Pang Ya-Ning, Yang Zhen, Zhang Ling-Xiao, Liu Xiao-Qiang, Dong Xin-Shu, Sheng Xun, Tan Jian-Guo, Mao Xin-Yu, Liu Ming-Yue
Institute of Applied Electronics, School of Electronics, Peking University, Beijing, China.
Department of Prosthodontics, Peking University School and Hospital of Stomatology & National Center for Stomatology & National Clinical Research Center for Oral Diseases & National Engineering Research Center of Oral Biomaterials and Digital Medical Devices & Beijing Key Laboratory of Digital Stomatology & NHC Key Laboratory of Digital Stomatology & NMPA Key Laboratory for Dental Materials, Beijing, China.
J Dent Sci. 2025 Jan;20(1):477-486. doi: 10.1016/j.jds.2024.05.013. Epub 2024 May 21.
BACKGROUND/PURPOSE: Artificial intelligence (AI) can assist in medical diagnosis owing to its high accuracy and efficiency. This study aimed to develop a diagnostic system for automatically determining the degree of tooth wear (TW) using intraoral photographs with deep learning.
The study included 388 intraoral photographs. A tooth segmentation model was first established using the Mask R-CNN architecture, which incorporated U-Net and SGE attention mechanisms. Subsequently, 2774 individual tooth images output from the segmentation model were included into the classification task, labeled and randomized into training, validation, and test sets with 6.0:2.0:2.0 ratio. A vision transformer model optimized using a mask mechanism was constructed for TW degree classification. The models were evaluated using the accuracy, precision, recall, and F1-score metrics. The time required for AI analysis was calculated.
The accuracy of the tooth segmentation model was 0.95. The average accuracy, precision, recall, and F1-score in the classification task were 0.93, 0.91, 0.88, and 0.89, respectively. The F1-score differed in different grades (0.97 for grade 0, 0.90 for grade 1, 0.88 for grade 2, and 0.82 for grade 3). No significant difference was observed in the accuracy between different surfaces. The AI system reduced the time required to grade an individual tooth surface to 0.07 s, compared to the 2.67 s required by clinicians.
The developed system provides superior accuracy and efficiency in determining TW degree using intraoral photographs. It might assist clinicians in the decision-making for TW treatment and help patients perform self-assessments and disease follow-ups.
背景/目的：人工智能（AI）因其高精度和高效率可辅助医学诊断。本研究旨在开发一种利用深度学习的口腔内照片自动确定牙齿磨损（TW）程度的诊断系统。
本研究纳入388张口腔内照片。首先使用结合了U-Net和SGE注意力机制的Mask R-CNN架构建立牙齿分割模型。随后，将分割模型输出的2774张单颗牙齿图像纳入分类任务，进行标注并按6.0:2.0:2.0的比例随机分为训练集、验证集和测试集。构建了一种使用掩码机制优化的视觉Transformer模型用于TW程度分类。使用准确率、精确率、召回率和F1分数指标对模型进行评估。计算AI分析所需时间。
牙齿分割模型的准确率为0.95。分类任务中的平均准确率、精确率、召回率和F1分数分别为0.93、0.91、0.88和0.89。不同等级的F1分数有所不同（0级为0.97，1级为0.90，2级为0.88，3级为0.82）。不同牙面之间的准确率无显著差异。与临床医生所需的2.67秒相比，AI系统将单颗牙齿表面分级所需时间缩短至0.07秒。
所开发的系统在利用口腔内照片确定TW程度方面具有卓越的准确性和效率。它可能有助于临床医生在TW治疗决策中提供帮助，并帮助患者进行自我评估和疾病随访。
