Brain tumor segmentation based on CBAM-TransUNet

Abstract

Brain tumor is one of the most serious brain diseases, and accurate brain tumor segmentation is crucial in clinical planning treatment and evaluating treatment outcomes in brain tumor patients. In this paper, we propose a 3D visual transducer model (CBAM-TransUNet) that incorporates an attention mechanism for 3D multimodal brain tumor edge detection and segmentation, to improve the accuracy of brain tumor segmentation. In our proposed model based on the framework of the U-Net model (Ronneberger O et al., 2015), Swin Transformer module (LIU Z et al., 2021) is introduced in the process of the encoder and decoder of the model, and the convolution block attention module (WOOS et al., 2018) is applied at the bottleneck layer. Comprehensive experiments are implemented on the BraTS 2021 dataset and it shows that the proposed model obtains competitive results: the Dice coefficients of whole tumor, core tumor and enhanced tumor segmentation are 93.08%, 91.49% and 87.76%, respectively, and the other 95% Hausdorff distances are 2.93mm, 4.20mm, 4.91mm. The proposed CBAM-TransUNet model can effectively improve the accuracy of brain tumor segmentation.