Machine Learning Approach for Automated Recognition of Non-Proliferative Diabetic Retinopathy

Diabetic Retinopathy (DR) is one of the most rampant ophthalmic disease found in diabetic patients. The damage of blood venule in the light-sensitive cell at the back of the eye causes DR. Vision impairment and vision loss are the results due to DR. Vision degradation can be prevented if diagnosed early and treated promptly. Normally clinical diagnosis is done by the visual examination of the fundus manually by an ophthalmologist. This technique is time-consuming and costly. Today emerging technologies in health care aims to reduce the cost of treatment and early diagnosis. For solving this problem, a computerized detection method is proposed for the early prediction of DR, i.e., non-proliferative diabetic retinopathy (NPDR). Many studies are being conducted to help in the early detection of DR. Any computer-assisted detection technique include segmentation, extraction of features, and classification. These approaches, on the other hand, are unable to capture the deep complex features and can only classify DR’s prediction with a poor degree of accuracy. In this work, forty-nine features of each candidate object are extracted and classify them as normal verse abnormal using machine learning algorithms. MESSIDOR dataset having 1200 fundus images are used for the classification of DR. The experimental findings reveal that this technique outperforms KNN (K- Nearest Neighbour), SVM (Support Vector Machine), and NB (Naive Bayes) in recognising the existence of DR. In this method SVM performs high accuracy of 98% in prediction of DR.