Study on retinal blood vessel segmentation method based on improved U-Net

Aiming at the retinal blood vessel segmentation problem in which microvascular features are difficult to extract, an improved retinal blood vessel segmentation method DA-UNet (deformable attention-UNet) is proposed based on the U-Net model. Firstly, deformable convolution is used instead of traditional convolution in the down-sampling stage to enhance the microvascular feature extraction ability. Secondly, a channel attention mechanism is introduced into the up-sampled and down-sampled feature maps to increase the weight of the channel encoding the vascular location, which improves the segmentation capability. Finally, the dice loss function is utilized to measure the similarity between the segmentation results and the real labeled blood vessels. Comparison experiments are conducted based on the DRIVE dataset, and the accuracy of the experimental results is 95.45%, the specificity is 97.95%, and the sensitivity is 82.83%, which indicates that the method in this paper has certain advantages over the mainstream retinal vessel segmentation methods.