Improving TMJ Diagnosis: A Deep Learning Approach for Detecting Mandibular Condyle Bone Changes

<b>Objectives</b>: This paper evaluates the potential of using deep learning approaches for the detection of degenerative bone changes in the mandibular condyle. The aim of this study is to enable the detection and diagnosis of mandibular condyle degenerations, which are difficult to observe and diagnose on panoramic radiographs, using deep learning methods. <b>Methods</b>: A total of 3875 condylar images were obtained from panoramic radiographs. Condylar bone changes were represented by flattening, osteophyte, and erosion, and images in which two or more of these changes were observed were labeled as "other". Due to the limited number of images containing osteophytes and erosion, two approaches were used. In the first approach, images containing osteophytes and erosion were combined into the "other" group, resulting in three groups: normal, flattening, and deformation ("deformation" encompasses the "other" group, together with osteophyte and erosion). In the second approach, images containing osteophytes and erosion were completely excluded, resulting in three groups: normal, flattening, and other. The study utilizes a range of advanced deep learning algorithms, including Dense Networks, Residual Networks, VGG Networks, and Google Networks, which are pre-trained with transfer learning techniques. Model performance was evaluated using datasets with different distributions, specifically 70:30 and 80:20 training-test splits. <b>Results</b>: The GoogleNet architecture achieved the highest accuracy. Specifically, with the 80:20 split of the normal-flattening-deformation dataset and the Adamax optimizer, an accuracy of 95.23% was achieved. The results demonstrate that CNN-based methods are highly successful in determining mandibular condyle bone changes. <b>Conclusions</b>: This study demonstrates the potential of deep learning, particularly CNNs, for the accurate and efficient detection of TMJ-related condylar bone changes from panoramic radiographs. This approach could assist clinicians in identifying patients requiring further intervention. Future research may involve using cross-sectional imaging methods and training the right and left condyles together to potentially increase the success rate. This approach has the potential to improve the early detection of TMJ-related condylar bone changes, enabling timely referrals and potentially preventing disease progression.