U-Net-based blood vessel segmentation using an improved data augmentation with green channel images

Diabetic Retinopathy (DR) detection is challenging due to retinal fine blood vessel structure. In this case, segmentation of the vascular tree in early diabetic retinopathy detection details initial DR stage neovascularization abnormalities. However, identifying thick and thin vessel parts from the retina is also challenging in blood vessels using segmentation methods. Thus, we propose U-Net-based retinal blood vessel segmentation in this paper using an improved data augmentation method with green channel images. The green channel extraction has been done in preprocessing, which detects and extracts all pixels of the vasculature to overcome the issue of low-contrast images and thin vessel identification. We propose an improved data augmentation technique for green channel images. The proposed preprocessing and data augmentation techniques, such as flip, distortion, and elastic transform, help the U-Net to learn all the thick and thin vessel parts. The proposed U-Net model has trained on an augmented dataset consisting of green channel images that provide better accuracy for blood vessel segmentation. The proposed model gives a segmentation training accuracy of 96% and validation accuracy of 97% on the DRIVE dataset with 140 training and 20 test images. The proposed model can be applied with real-time data to identify diabetic retinopathy at an earlier phase.