Eg-Trans3Dunet: A Single-Staged Transformer-Based Model For Accurate Vertebrae Segmentation From Spinal Ct Images

Multi-label vertebrae segmentation has been a fundamental research for spinal image analysis and surgical intervention. However, most existing methods have some drawbacks including poor boundary segmentation results and identification of the 25th vertebra. In this paper, we propose a single-staged network based on vision Transformer and UNet. Besides, we exert supervision on vertebrae Edges and introduce additional Global information to this network, called EG-Trans3DUNet. Specifically, Transformer encoder can better extract global semantic information compared with pure UNet structure. Edge detection module helps to refine segmentation boundaries and keep segmentation consistency in each vertebra. Global information from the whole case can improve model performance on the identification of all vertebrae. We demonstrate that our proposed model outperforms other single-stage methods on the segmentation accuracy of VerSe'20 datasets.