Polat Sema, Öksüzler Fatma Yasemin, Öksüzler Mahmut, Çoban Önder, Tunç Mahmut, Yüksel Hazal Duyan, Özşahin Esin, Göker Pinar
Cukurova University Faculty of Medicine, Department of Anatomy, Adana, Turkey.
Izmir Democracy University Buca Seyfi Demirsoy Training and Research Hospital, Department of Radiology, Izmir, Turkey.
Medicine (Baltimore). 2024 Dec 13;103(50):e40846. doi: 10.1097/MD.0000000000040846.
The study aimed to compare the morphometric and morphologic analyses of the bone structures of temporomandibular joint and masticatory muscles on Cone beam computed tomography (CBCT) in 62 healthy subjects and 33 subjects with temporomandibular dysfunction (TMDS) aged between 18 and 56 years. In addition, a machine learning (ML) pipeline involving the Random Forest classifier was used to automatically detect TMDS. Thirty parameters (including age and gender) associated with the condylar process, articular tubercle, mandibular fossa, ramus mandible, joint space, and masticatory muscles were examined using CBCT. Well-known steps including scaling, feature selection, and feature extension are used to build the ML pipeline. Among 30 parameters, angle between mediolateral axes of both the head of mandible, medial pterygoid muscle thickness (PMT), distance between the most superior point of head of the mandible and the mandibular fossa bone surface opposite, medial joint space, lateral joint space, articular tubercle inclination, mandibular fossa depth head of the mandible's length, and angle between the ramus mandible long axis and the coronal plane values showed significant differences between healthy subjects and TMDS. Additionally, from the above measurements, all parameters (except PMT) were significantly lower in TMDS than in healthy subjects. Moreover, the results show that it is possible to automatically detect temporomandibular dysfunction with an f1-score of 0.967 when arming our ML pipeline with feature selection and extension. The reference values of the condylar process, articular tubercle, mandibular fossa, ramus of mandible, and joint space may play a key role in increasing of the success of the surgical procedure, or the assessment/differentiating of the TMD. ML is capable of detecting TMD in an automatic and highly accurate way. Hence, it is also concluded that ML can be useful for cases requiring making automatic and highly correct predictions.
该研究旨在比较62名年龄在18至56岁之间的健康受试者和33名颞下颌关节紊乱病（TMDS）患者颞下颌关节和咀嚼肌骨骼结构的形态测量和形态学分析。此外，使用包含随机森林分类器的机器学习（ML）管道自动检测TMDS。使用锥形束计算机断层扫描（CBCT）检查了与髁突、关节结节、下颌窝、下颌支、关节间隙和咀嚼肌相关的30个参数（包括年龄和性别）。构建ML管道时使用了包括缩放、特征选择和特征扩展等知名步骤。在30个参数中，双侧下颌头内外侧轴之间的角度、翼内肌厚度（PMT）、下颌头最上点与相对的下颌窝骨表面之间的距离、内侧关节间隙、外侧关节间隙、关节结节倾斜度、下颌窝深度、下颌头长度以及下颌支长轴与冠状面之间的角度值在健康受试者和TMDS患者之间存在显著差异。此外，从上述测量结果来看，TMDS患者的所有参数（除PMT外）均显著低于健康受试者。而且，结果表明，当我们的ML管道进行特征选择和扩展时，能够以0.967的F1分数自动检测颞下颌关节紊乱病。髁突、关节结节、下颌窝、下颌支和关节间隙的参考值可能在提高手术成功率或评估/鉴别颞下颌关节紊乱病方面发挥关键作用。ML能够以自动且高度准确的方式检测颞下颌关节紊乱病。因此，还得出结论，ML对于需要进行自动且高度正确预测的情况可能会很有用。
