Abstract: Aiming at the problem of the loss of semantic information in the feature extraction process of U-shaped network, which affected the accuracy of liver image segmentation, we  proposed a liver image segmentation method based on multi-scale feature fusion, introducing embedded spatial attention and improved channel attention. Firstly,  multi-scale convolution modules were used  to extract semantic information of different scales and stitch them together in the encoding stage of the network. Secondly,  spatial attention and improved channel attention were introduced into the whole network model to strengthen the weight of meaningful semantic information from the perspectives of spatial and channel domains. Finally, the liver segmentation results were output through a convolutional layer. The experimental results show that the segmentation method has a better segmentation effect on the liver image data set, improves the liver segmentation accuracy, and effectively improves the difficulty of small-area liver segmentation.

Key words: deep learning, image segmentation, multi-scale convolution, attention mechanism