Blood Vessel Segmentation from Retinal Images using U-NET Architecture

Abstract

Diabetic Retinopathy has emerged as a substantial cause of blindness on a global scale. According to some recent studies automatic detection of blood vessels in the preprocessing step can significantly improve the classification accuracy of diabetic retinopathy. Furthermore, accurately segmenting blood vessels in retinal images allows healthcare professionals to identify subtle changes in the blood vessel structure that may indicate the onset of diabetic retinopathy. In this paper, the U-NET architecture model has been implemented on the DRIVE and the Aptos 2019 datasets. Additionally, the performance of the model has been evaluated on the proposed KUBERI dataset. The images in the KUBERI dataset have been collected from the Bangladesh Eye Hospital in Khulna, comprising a total of 54 images divided into five diabetic retinopathy classes. The proposed dataset is a valuable addition to the research community since there was no previous retinal image dataset from Bangladesh before that. Our results indicate that the U-NET model achieved an impressive accuracy of 98.57% on the DRIVE dataset and 97.73% on the Aptos 2019 dataset. Particularly noteworthy is the exceptional performance of the implemented model on the proposed KUBERI dataset, where it achieved an accuracy of 91.11%, precision of 85.20%, and recall of 79.31%.