Segmentation of structural components of the temporomandibular joint and detection of anteriorly positioned discs in magnetic resonance images using deep learning

Temporomandibular joint disorders (TMJDs) are commonly seen in the population, and using the correct imaging method in diagnosing these diseases is critical for clinical success. Magnetic Resonance Imaging (MRI) is widely recognized as the gold standard for diagnosing soft tissue disorders of the temporomandibular joint (TMJ). However, interpreting TMJ MRI images is challenging even for expert clinicians. Artificial intelligence (AI) systems can be used to standardize MRI interpretation. The primary aim of this study is to automatically segment the structural components of the TMJ region using deep learning models on T2-weighted sagittal MRI images. This study involved segmenting the mandibular condyle, articular eminence, glenoid fossa, and articular disc in both normal and anteriorly displaced positions. The performance of YOLOv5 and YOLOv8 models in this segmentation process was compared to determine which model provides higher accuracy and precision for these structures. The hypothesis tested is that AI based approaches can contribute to diagnostic processes in clinical applications. A dataset comprising 901 T2-weighted sagittal MRI images of the TMJ region was used in the study. Segmentation of the anatomical structures -mandibular condyle, articular eminence, glenoid fossa, and articular disc- was performed using the AI supported Craniocatch (Eskisehir, Turkiye) application with YOLOv5 and YOLOv8 deep learning models. Model performance was evaluated using metrics such as recall, precision, and F1 score. In mandibular condyle segmentation, YOLOv5 had a sensitivity of 1, while YOLOv8 had a sensitivity of 0.98. The precision of YOLOv5 was 0.98, and YOLOv8's precision was 0.97. In terms of the F1 score, YOLOv5 scored 0.99, and YOLOv8 scored 0.98. For the segmentation of the articular eminence, both YOLOv5 and YOLOv8 had a sensitivity of 0.98. In terms of precision, YOLOv5 scored 0.96, while YOLOv8 scored 0.98. The F1 score was 0.97 for YOLOv5 and 0.98 for YOLOv8. In glenoid fossa segmentation, YOLOv5 had a sensitivity of 0.82, while YOLOv8 had a sensitivity of 0.96. The precision for YOLOv5 was 0.90, while it was 0.98 for YOLOv8. The F1 score was 0.86 for YOLOv5 and 0.97 for YOLOv8. In the segmentation of the articular disc in the normal position, YOLOv5 achieved a sensitivity of 1, precision of 1, and F1 score of 1, while YOLOv8 had a sensitivity of 0.98, precision of 1, and F1 score of 0.99. For the anteriorly displaced articular disc, YOLOv5 had a sensitivity of 1, precision of 0.96, and F1 score of 0.98, while YOLOv8 had a sensitivity of 0.96, precision of 0.86, and F1 score of 0.90. This study demonstrates that AI-based YOLOv5 and YOLOv8 models achieve high success in the automatic segmentation of TMJ structures. Specifically, YOLOv5 exhibited superior performance in segmenting the mandibular condyle and anteriorly displaced disc, while YOLOv8 provided better results for the segmentation of the glenoid fossa and articular eminence. The findings suggest that these models can be integrated into clinical and radiological diagnostic processes, potentially accelerating diagnosis and offering significant support to clinicians. Furthermore, the development of more advanced AI models for the TMJ region could lead to more effective early diagnosis and treatment outcomes.