WSC-Trans: A CNN-Transformer Structure-Based 3D Multi-Structural Automatic Segmentation Model for Temporal Bone CT

Cochlear implantation is the most effective treatment for severe deafness, which requires accurate localization of temporal bone anatomy. Preoperative CT image segmentation is an essential technique to determine the location of relevant tissues in the temporal bone. However, manual segmentation is usually time-consuming and suffers from low accuracy due to the complex and small structures of these tissues in temporal bone CT. To address this issue, we proposed a CNN-Transformer structure-based 3D multi-structured model for the automatic segmentation of fine and complex tissues such as the cochlea, facial nerve, ossicles, vestibule and semicircular canal in the temporal bone CT. Our model adopts a new Transformer deformation structure, which effectively utilizes the spatial attention mechanism to capture feature dependencies, and uses the channel attention mechanism to fuse different channel semantic representations to improve segmentation accuracy. Extensive experiments on a private temporal bone CT dataset show that our model achieves higher DSS and JSS scores, and lower HD95 and ASSD scores for all targets compared with other existing segmentation methods, demonstrating its superior performance.