IDMUNet : An effective network for liver tumor segmentation
WU Cai-hong, Changming Song, Dongxu Cheng, Siyv Ren, Zenghui Li, Kang Chen, Hao Li · 2023
Most existing approaches for medical image segmentation has been proposed, but they can't get desirable result due to the network. The classical network lose much information during downsampling process. And the medica images segmentation is easy to be contaminated by heavy noise, poor boundaries and so on. To address these issues, we propose an effective network named IDMUNet, for liver tumor segmentation in CT images, This network introduces an information supplement(IS) module to guide the network to efficiently fuse global and local information. After that, the features obtained from each encoding layer of the encoder are fed into a dense attention-based atrous spatial pyramid pooling(DAASPP) module to further extract features on channel and spatial domains, thus it can not only obtain effective features in channel and spatial dimensions, but also suppress irrelevant weight information. What's more, to obtain better edge information, a multi-task learning(MTL) module is introduced to enable the network to achieve better performance. Experiments on the public liver tumor dataset, LiTS2017, demonstrate that the effectiveness of the proposed IDMUNet and the superiority of our method against the SOTA methods.