Classification of Fundus Images Based on Severity Utilizing SURF Features from the Enhanced Green and Value Planes

Abstract

Diabetic retinopathy (DR) is a complication of longstanding diabetes that affects vision. In DR, the small blood vessels that supply the nutrients and oxygen to the retina get damaged, which can blur the vision. If not treated in the preliminary stages, DR can even lead to a complete loss of vision. Hence, a concise technique for detecting and grading the severity level of DR is necessary. The current paper focuses on an automated DR detection and severity classification technique on retinal fundus images using the Support Vector Machine (SVM) classifier. Our proposed system comprised pre-processing of the fundus images and merging the enhanced green and value color planes. The Average Grey Value Extraction (AGVE) algorithm was applied to the merged image to extract the important information from the eye. Then, the Speeded-Up Robust Features (SURF) was used to extract the strongest feature points in the fundus image. SVM was trained to classify the DR fundus images into various levels of severity. The experimentation results on the DIARETDB1 dataset obtained an average accuracy of 98.68% and an average F1 score of 0.99. A novel red score was found out to be a good indicator of severity. By combining the enhanced green and value color plane, the features extracted by SURF were more accurate for predicting the severity. Thus, the proposed system will assist the doctors in detecting the severity level of DR efficiently and reliably, enabling them to start medication on time.