MSA-UNet: Retinal vessel segmentation method based on multiple spatial attention mechanism

The high-precision segmentation of the macro and macro structure of the retina and its fine vascular structure is very important for the identification of some early developing eye diseases, such as diabetes and hypertension. In this article, we improve the original U-Net network and propose the MSA-UNet network, which introduces multiple spatial attention mechanisms. This network does not require a large amount of training data. We enhance and amplify the basic dataset to enable it to work efficiently in situations with limited data. MSA-UNet introduces multiple spatial attention mechanism modules in each of the U-Net decoder up-sampling phase inserts SA modules that infer the attention map along the spatial dimension and multiply the attention map with the input feature map for adaptive feature refinement, enabling each up-sampling phase to suppress unnecessary information. In addition, the network uses structured discard convolution fast instead of U-Net raw convolution blocks, which can effectively prevent network overfitting. We evaluate the MSA-UNet based on retinal vascular extraction (DRIVE) datasets. The results show that the proposed MSA-UNet achieves a more advanced performance on this datasets.