Residual based attention-Unet combing DAC and RMP modules for automatic liver tumor segmentation in CT

Abstract

<abstract> <p><italic>Purpose</italic>: Due to the complex distribution of liver tumors in the abdomen, the accuracy of liver tumor segmentation cannot meet the needs of clinical assistance yet. This paper aims to propose a new end-to-end network to improve the segmentation accuracy of liver tumors from CT. <italic>Method</italic>: We proposed a hybrid network, leveraging the residual block, the context encoder (CE), and the Attention-Unet, called ResCEAttUnet. The CE comprises a dense atrous convolution (DAC) module and a residual multi-kernel pooling (RMP) module. The DAC module ensures the network derives high-level semantic information and minimizes detailed information loss. The RMP module improves the ability of the network to extract multi-scale features. Moreover, a hybrid loss function based on cross-entropy and Tversky loss function is employed to distribute the weights of the two-loss parts through training iterations. <italic>Results</italic>: We evaluated the proposed method in LiTS17 and 3DIRCADb databases. It significantly improved the segmentation accuracy compared to state-of-the-art methods. <italic>Conclusions</italic>: Experimental results demonstrate the satisfying effects of the proposed method through both quantitative and qualitative analyses, thus proving a promising tool in liver tumor segmentation.</p> </abstract>