MGAU-ResNet-multi-grained attention U-shaped residual networks for medical image segmentation.

The complexity and variability of medical image data pose a challenge to image segmentation. Deep networks have significantly progressed in image segmentation in recent years. However, the limitations of the fixed perceptual field and the unknown network width of the convolutional kernel in U-Net make it difficult to achieve fine segmentation of medical image lesion target boundaries. While the Transformer segmentation method based on the self-attentive mechanism improves the segmentation accuracy, the extensive network parameters are difficult to apply on small devices. This paper proposes a multigranular attention U-shaped residual network to address the above problems. Firstly, the multi-grain residual module obtains more local features from the image; secondly, the multi-grain attention mechanism enhances global and local information features; and finally, the network fuses feature from different branches. The multi-granularity feature module embedded in the U-Net architecture allows for the generalisation of image cuts. We have evaluated MGAU-ResNet on several medical images' datasets, such as retinal vasculature, skin lesions and cancer cell nuclei datasets. The experimental results show that our model outperforms other U-Net in ACC, MIOU and DICE metrics and achieves better segmentation results on different medical datasets. [ABSTRACT FROM AUTHOR]