MRI Image Segmentation Algorithm for Temporomandibular Joint Discs and Condyles Based on the Improved nnU-Net Framework

Accurate segmentation of the articular disc from MRI is crucial for diagnosing and treating temporomandibular joint disorders (TMD), but it remains challenging because of the small structure size, blurred boundaries, and class imbalance. This study aims to improve articular disc segmentation accuracy for computer-aided TMD diagnosis. We propose an enhanced nnU-Net-based method with three modifications: (1) fewer downsampling layers to preserve spatial detail, (2) a DynFocus module with channel-and-spatial attention at the bottleneck to enhance disc-background discrimination, and (3) a joint loss combining Dice focal and boundary losses to improve sensitivity and boundary delineation. On 50 temporomandibular joint MRI cases, the proposed method improved disc segmentation Dice from 0.69 to 0.71, maintained condyle Dice at 0.89, and achieved an average Dice of 0.80. These results indicate that the proposed enhancements improve disc segmentation performance while preserving condyle segmentation accuracy, supporting their usefulness for MRI-based TMD assessment.