EG-UNet: A Medical Image Segmentation Model with Enhanced Feature Fusion

Accurate medical image segmentation is crucial to improving diagnosis and treatment results. The existing CNN -based Rolling - Un et model performs well in medical image segmentation, but it still has certain limitations in capturing long-distance dependencies and processing local features. To solve these problems, this paper proposes an improved model EG-UNet. Specifically, we propose a channel feature enhancement module (CBRE encoder-decoder), which is responsible for dynamically assigning channels to features and improving the representation ability of channel features in CNN; and a core feature fusion module (GLFF), which captures long-distance dependencies through the interaction of DOR-MLP and PRM modules, while further enhancing the capture ability of local features, and using the residual structure to retain more semantic information, thereby improving the ability to capture boundary details. Experimental results show that on the ISIC20 18 and PH2 data sets, the Fl of this paper's network reached 89.96% and 94.52% respectively, and the ASD reached 0.47% and 0.38% respectively, showing significant performance improvement.