Certain investigations on Grading of diabetic retinopathy In digital fundus images

Abstract

Diabetic Retinopathy (DR) is the leading cause of blindness. Manual inspection suffers from inter and intraobserver variability. A computer-aided diagnosis system could be very much useful in aiding the pathologist for detecting and diagnosing DR at an early stage. Many proposed approaches addressed different issues and methodology schemes in different datasets. Thus, in this work, we attempted to create a unified framework which could achieve unified performance in detecting the abnormalities from the Fundus images. newlineIn this work, the images from DIARETDB 1, HRFI, and MESSIDOR datasets are considered for the analysis (total of 1319 images). To develop a unified framework all the images are uniformly made to undergo the following steps contrast enhancement, illumination correction, segmentation of blood vessels, pixel feature extraction, classification, and grading of DR conditions. Results show that the CLAHE method shows improved ability to enhance the contrast between the structures in the images without losing any structural information. It is been validated using image quality metrics. The gradient-based gamma correction method is found to exhibit high structural preservation capability during the elimination of illumination effects on the images thus ensuring the preservation of maximum information of the blood vessels. It has been validated using structural similarity indices. Blood vessels in the Fundus images are extracted using the k-means clustering algorithm. Local binary patterns of the blood vessels are extracted and compared to classify normal and DR subjects. It is observed that the local Gabor gray pattern is found to have increased capability in differentiating normal and DR subjects compared to LBP. A neural network classifier is found to identify the DR subjects with an accuracy of 99% and it shows better performance in grading the abnormality newline newline