AnImproved Method for Retinal Vascular Segmentation in U-Net

The fine-grained characteristics of blood vessels are difficult to obtain, and the details of the blood vessels are obscured when the current mainstream methods of retinal vascular segmentation are employed. This paper proposes an improved U-Net model algorithm to address these problems. The convolution layer of quadratic-cycle residual difference was used to replace the original convolutional layer in the upper and lower sampling of U-Net to improve the utilization rate of the features. A multichannel attention model was introduced in the decoding part to improve the segmentation effect of small blood vessels with low contrast. Results show that the accuracies of the algorithm in DRIVE (Digital Retinal Images for Vessel Extraction) and STARE (Structured Analysis of the Retina) databases are 96.89% and 97.96%, the sensitivities are 80.28% and 82.27%, and the AUC performances are 98.41% and 98.65%, respectively. All these parameters are higher than those of existing advanced algorithms. The proposed algorithm can effectively improve the segmentation accuracy of fine blood vessels in fundus images.