Coarse-to-Fine Tranformer for articular disc of the temporomandibular joint Segmentation

The most important subtypes of joint abnormalities in patients with temporomandibular disorders are different forms of disc displacement and deformation. An effective segmentation model for jaw joint detection to support the diagnosis of TMJ disease on magnetic resonance imaging is very crucial. Data for this study were obtained from 204 MRI images of patients with articular discs and the corresponding MRI segmentation labels of the temporomandibular joints. These images were used to evaluate four deep learning-based semantic segmentation methods. Using a multi-scale structured C2Ftrans segmentation model transformed from coarse to fine, it describes medical image segmentation as a coarse to fine process. It is able to perform accurate target boundary segmentation with lower computational complexity. Tested on this dataset, comparing U-Net, Unet ++ and Attention-U net models for data segmentation results show the C2Ftrans model performs best with the highest dice of 73.5% and the lowest computational complexity.