Abstract: Aiming at the problem of weak robustness of encoder-decoder networks in feature extraction from complex medical images, such as complex textures, blurred boundaries, low contrast with surrounding tissues, and background noise interference, which led to low segmentation accuracy of lesion region, we proposed a medical image segmentation network based on multi-scale semantic representation. Firstly, a multi-scale context-aware module was used to enhance the representation ability of different scale contexts. Secondly, by calculating the feature difference between adjacent layers, the difference in semantic features between different layers was highlighted and the redundancy of feature information was reduced. Finally, a hybrid attention module was used to enhance the boundary information of the lesion region and the semantic perception ability of complex features by the network. Experimental results show that the network has high segmentation accuracy and strong robustness in  complex medical image segmentation.

Key words: medical image segmentation, semantic representation, multi-scale context, feature difference, hybrid attention