A Segmentation Method of 3D Liver Image Based on Multi-scale Feature Fusion and Coordinate Attention Mechanism

Meng Zhang, Xiaolong Zhang, He Deng, Hongwei Ren

Abstract: Due to the high similarity of organs in 3D liver image and the use of simple connection by U-Net to fuse different semantic features, the segmentation accuracy of network needs to be improved. To solve these problems, this paper proposes a 3D liver semantic segmentation method based on multi-scale feature fusion and coordinate attention mechanism. Firstly, in the encoder section of U-Net, the multi-scale feature fusion module was used to capture multi-scale features; Then, coordinate attention mechanism was used to fuse low-level features and high-level features to locate regions of interest; Finally, the segmentation effect of edge details was improved through a deep supervision mechanism. The experimental results show that: on the LiTS dataset, the dice similarity coefficient (DSC) of this method reaches 96.5%. Compared with the U3-Net + DC method, the DSC increases by 0.1%, and the relative volume difference (RVD) decreases by 1.09%; On the CHAOS dataset, the DSC of this method reaches 96.8%, and compared with CANet, the DSC increases by 0.2%; On the MRI dataset of a hospital, the DSC of this method reaches 97.2%.