Mao Wei-Yu, Fang Yuan-Yuan, Wang Zhong-Zhen, Liu Mu-Qing, Sun Yu, Wu Hong-Xin, Lei Jie, Fu Kai-Yuan
Department of Oral & Maxillofacial Radiology, Peking University School & Hospital of Stomatology, Beijing 100081, PR China; National Center for Stomatology & National Clinical Research Center for Oral Diseases, Beijing 100081, PR China; National Engineering Research Center of Oral Biomaterials and Digital Medical Devices, Beijing 100081, PR China; Beijing Key Laboratory of Digital Stomatology, Beijing 100081, PR China.
LargeV Instrument Corp. Ltd., Beijing 100084, PR China.
J Dent. 2025 Mar;154:105592. doi: 10.1016/j.jdent.2025.105592. Epub 2025 Jan 25.
In this study, artificial intelligence (AI) techniques were used to achieve automated diagnosis and classification of temporomandibular joint (TMJ) degenerative joint disease (DJD) on cone beam computed tomography (CBCT) images.
An AI model utilizing the YOLOv10 algorithm was trained, validated and tested on 7357 annotated and corrected oblique sagittal TMJ images (3010 images of normal condyles and 4347 images of condyles with DJD) from 1018 patients who visited Peking University School and Hospital of Stomatology for temporomandibular disorders and underwent TMJ CBCT examinations. This model could identify DJD as well as the radiographic signs of DJD, namely, erosion, osteophytes, sclerosis and subchondral cysts. The diagnosis and classification performances of the model were evaluated on the test set. The accuracy of the model for evaluating images with one to four DJD signs was also evaluated.
The accuracy, precision, sensitivity, specificity, F1 score and mean average precision (mAP) at an intersection over union (IoU) threshold of 0.5 of the model for DJD detection all exceeded 0.95. The accuracies for identifying erosion, osteophytes, sclerosis and subchondral cysts were 0.91, 0.96, 0.91 and 0.96, respectively. The precisions, specificities and F1 scores for the DJD signs were all >0.90. The sensitivity ranged from 0.88 to 0.95, and the mAP (IoU=0.5) ranged from 0.87 to 0.97. The accuracies of the model for detecting one to four DJD signs in one image were 94 %, 84 %, 66 % and 63 %, respectively.
A deep learning model based on the YOLOv10 algorithm can not only detect the presence of TMJ DJD on CBCT images but also differentiate the typical radiographic signs of DJD, including erosion, osteophytes, sclerosis and subchondral cysts, with acceptable accuracy.
TMJ DJD is a very common disease that causes joint pain and mandibular dysfunction and affects patients' quality of life; therefore, early diagnosis and intervention are particularly important. However, identifying radiographic signs of early-stage TMJ DJD is difficult. AI can quickly review CBCT images and assist in the accurate and rapid diagnosis and classification of TMJ DJD.
在本研究中，利用人工智能（AI）技术在锥形束计算机断层扫描（CBCT）图像上实现颞下颌关节（TMJ）退行性关节病（DJD）的自动诊断和分类。
使用YOLOv10算法的AI模型在来自1018名因颞下颌关节疾病就诊于北京大学口腔医学院及口腔医院并接受TMJ CBCT检查的患者的7357张标注并校正的斜矢状位TMJ图像（3010张正常髁突图像和4347张患有DJD的髁突图像）上进行训练、验证和测试。该模型能够识别DJD以及DJD的影像学征象，即侵蚀、骨赘、硬化和软骨下囊肿。在测试集上评估该模型的诊断和分类性能。还评估了该模型对具有一至四个DJD征象的图像的评估准确性。
该模型用于DJD检测在交并比（IoU）阈值为0.5时的准确率、精确率、灵敏度、特异性、F1分数和平均精度均值（mAP）均超过0.95。识别侵蚀、骨赘、硬化和软骨下囊肿的准确率分别为0.91、0.96、0.91和0.96。DJD征象的精确率、特异性和F1分数均>0.90。灵敏度范围为0.88至0.95，mAP（IoU = 0.5）范围为0.87至0.97。该模型对一张图像中检测一至四个DJD征象的准确率分别为94%、84%、66%和63%。
基于YOLOv10算法的深度学习模型不仅可以在CBCT图像上检测TMJ DJD的存在，还能以可接受的准确率区分DJD的典型影像学征象，包括侵蚀、骨赘、硬化和软骨下囊肿。
TMJ DJD是一种非常常见的疾病，会导致关节疼痛和下颌功能障碍，影响患者生活质量；因此，早期诊断和干预尤为重要。然而，识别早期TMJ DJD的影像学征象具有难度。AI可以快速查看CBCT图像，并协助对TMJ DJD进行准确、快速的诊断和分类。
