Retinal Vessel Segmentation Based on Recurrent Convolutional Skip Connection U-Net

At present, the efficiency of manual diagnosis of eye diseases is low, subjective, and easy to cause misdetection and missed diagnosis. The image segmentation of blood vessels is performed on the retinal fundus image so that the machine vision results can be used for the prevention and early diagnosis of hypertension, diabetes, and other diseases, which has a very high reference value. Inspired by Recurrent Convolutional Neural Network (RCNN), a Recurrent Convolution Skip Connection UNet (RCSCU-Net) is proposed, which improves the problem that the coupling between traditional U-Net skip connection and convolution layer is too large. It connects the skip connection with the input layer directly and obtains more information through the recurrent convolutional operation. Besides, the residual convolution module is used to replace the ordinary convolution module to avoid overfitting due to too deep a network. The network ’s accuracy on the DRIVE and STARE datasets is 0.9624 and 0.9673, proving that the network structure has a better segmentation effect than the existing network structures such as U-Net and HED.