An End-to-End Segmentation Network for the Temporomandibular Joints CBCT Image based on 3D U-Net

The temporomandibular joints segmentation from CBCT images plays an important role in medical diagnosis of related diseases, such as temporomandibular disorders. However, due to the weak contrast and heterogeneous shapes, this task is considerably challenging for conventional image processing algorithms. In this paper, we proposed a novel end-to-end deep learning segmentation network based on 3D U-Net for the TMJ segmentation in CBCT volumes. The proposed network takes advantages of the symmetrical architecture to achieve precise and voxel-wise prediction. We demonstrated the well segmentation results of the proposed method on our clinical CBCT image datasets. Without any post-processing, our method attained an average Dice Coefficient 0.9760 which performed much better than the super-pixel segmentation algorithm and the active contour model.