Machine learning identification of diabetic retinopathy from fundus images

Diabetic retinopathy may potentially lead to blindness without early detection and treatment. In this research, an approach to automate the identification of the presence of diabetic retinopathy from color fundus images of the retina has been proposed. Classification of an input fundus image into one of the three classes, healthy/normal, Non-Proliferative Diabetic Retinopathy (NPDR) and Proliferative Diabetic Retinopathy (PDR) has been achieved. Blood vessel segmentation from the input image is achieved by Gaussian filtering. An adaptive, input - driven approach is considered for the mask generation and thresholding is accomplished using local entropy. The processed image obtained is characterized by second order textural feature, contrast, in four different orientations- 0°, 45°, 90° and 135° and structural features namely, fractal dimension and lacunarity. The research incorporates a three layered artificial neural network (ANN) and support vector machines (SVM) to classify the retinal images. The efficiency of the proposed approach has been evaluated on a set of 106 images from the DRIVE and DIARETB1 databases. The experimental results indicate that this method can produce a 97.2% and 98.1% classification accuracy using ANN and SVM respectively invariant of rotation, translation and scaling in input retinal images as opposed to a fixed mask based on the matched filter method.