Internal Validation of a YOLOv8 Model for Classifying Temporomandibular Joint Disc Displacement With Versus Without Reduction on Open‐Mouth Proton‐Density‐Weighted MRI

BACKGROUND: The temporomandibular joint (TMJ) disc plays an important role in maintaining normal joint function. However, disc displacement is a common condition that can lead to pain, restricted jaw movement and chronic TMJ dysfunction. OBJECTIVE: To develop and evaluate a deep learning model capable of autonomously localizing the TMJ region from magnetic resonance (MR) images during mouth opening using the You Only Look Once (YOLO) v8 framework, and to discriminate between TMJ disc displacement with reduction (DDwR) and disc displacement without reduction (DDwoR). METHODS: This study included 237 patients (48 males, 189 females; age range, 19-83 [mean, 44.73 ± 18.3] years) who underwent MRI examination at our hospital for suspected TMJ disorders and were diagnosed with disc displacement based on imaging findings. After exclusion, 439 proton density-weighted MR images during mouth opening were preprocessed and divided on a patient basis into training (280 images), validation (71 images) and test (88 images) datasets. The TMJ region was manually annotated using LabelImg, including the condyles, articular disc and articular tubercle. The YOLOv8 deep learning model was trained for 200 epochs to localize the TMJ region and classify DDwR or DDwoR. RESULTS: The YOLOv8 model achieved promising results for the detection and classification of TMJ disc displacements. The performance metrics were a precision, recall and F1 score of 0.983, 0.988 and 0.987, respectively, indicating a high accuracy in detecting and identifying disc displacement patterns. CONCLUSION: The YOLOv8 deep learning model demonstrated promising performance in automatically localizing and classifying TMJ disc displacement from MR images.