Abstract

Info

Author(s):: ZHOU Jinzhi1; 2; HU Zhen1; 2; GUO Lili1; 2; GONG Li1; 2; ZHANG Wengrong1; 2; 1. School of Information Engineering, Southwest University of Science and Technology, Mianyang 621000, China 2. Sichuan Provincial Key Laboratory of Robot Technology Used for Special Environment, Mianyang 621000, China

Keywords:: medical image segmentation deep learning generative adversarial network attention mechanism liver tumor

Abstract:: Aiming at the problems of under-segmentation, over-segmentation, boundary ambiguity, and low segmentation accuracy in the existing liver CT image segmentation methods, an automatic segmentation algorithm for liver tumors based on a generative adversarial network (GAN) is proposed. Firstly, the image is pre-segmented to reduce the effect of irrelevant information. Secondly, the GAN-generative network uses Dual Attention Unet (DAUnet), which introduces dual attention mechanisms in skip connections to enhance the features of liver tumors. Finally, the predicted tumor images become more accurate through the generative adversarial training of GAN and the introduction of the mixed loss function in the training process. The experimental results on the LiTS dataset show that the Dice similarity coefficient of the proposed algorithm reaches 76.15%, which is 3.63% higher than that of Unet. DAUnet for generative adversarial training can effectively improve the performance of tumor segmentation in liver images.