Retinal Vessel Segmentation with Differentiated U-Net Network

In this study, an improved method based on U-Net architecture was applied for retinal vessel segmentation and the results were compared with other methods. In the preprocessing phase of the applied method, color fundus images were converted to LAB space and CLAHE (Contrast Limited Adaptive Histogram Equalization) was applied to the L channel of the image, then the channels were converted back to RGB space and the Gaussian and median filtering processes were used to reduce the noise. In the developed U-Net architecture, feature maps that were obtained by up-sampling (un-pooling) and maximum pooling operations were concentrated on the jump connections of the architecture. The accuracy, sensitivity, specificity, dice and jaccard percentage values were 97.87, 84.11, 9939, 88.70, 79.69, respectively that were obtained from the method. The results show that the method performs an efficient segmentation according to the literature we know.