Retinal Blood Vessel Segmentation Based on Modified CNN and Analyze the Perceptional Quality of Segmented Images

Abstract

Diabetic retinopathy is a major issue faced all over the world peoples that causes permanent blindness. With the onset of symptoms of diabetic retinopathy and the illness advances to an extreme level, it is difficult to recognize diabetic retinopathy at an earlier level. This paper presents the automatic detection of blood vessel segmentation based on U-net architecture. First, the retina blood vessels were segmented using a U-Net Architecture with the encoder/decoder module of multiple convolutional neural networks. For segmentation, binary conversion techniques are used. For the classification, deep learning models were proposed, namely ResNet50, Inception V3, VGG-16, and modified CNN. The final results are measured on a standard benchmark DRIVE dataset that contains 2865 retinal blood vessel images. For image classification, the proposed modified CNN performed better for DRIVE datasets with an accuracy score of 98%. Precision of 98%, Recall is 94.5% and F1-score is 95%. This paper evaluates the perceptional quality of segmented retinal images using SSIM. In this study pixel intensity was measured using RMSE, and PSNR to assess the quality of the retinal vessel segmented image.