Transformer Based Position Information Enhancement for Medical Image Segmentation

The U-Net structure is widely used in the field of image segmentation, and the medical image segmentation task is more sensitive to the accuracy of location information, which has become a research difficulty in the field. In order to improve the accuracy of medical image segmentation, a new medical image segmentation algorithm LIE-Trans is proposed: the algorithm utilizes CNN and Swin-Transformer network in the encoder part to extract features from the image; fuses the multi-scale location information using LIE Module to enhance the location information; decodes the feature map using Swin-Transformer module. Transformer module is used to decode the feature map, and the position information is enhanced during the decoding process; the decoded feature map is fused with multi-scale features for segmentation. Experiments were performed on Synapse abdominal multi-organ segmentation dataset with higher accuracy and better results compared to other algorithms.