Automatic Detection of Hard Exudates from Fundus Images Using Adaptive Neuro-Fuzzy Inference System

Abstract

Diabetic retinopathy is one of the most significant factors contributing to blindness, and so diagnosis and timely treatment is particularly important to prevent visual loss. Exudates are a primary sign of diabetic retinopathy and early detection could limit the disease’s severity. Hard exudates are detected using a combination of colour histograms in the HSV (hue, saturation, value) colour space, and the split and merge technique with the predicate function involving contrast measures to calculate edge pixels. Intensity, textural features using second-order statistics are extracted from fundus images and given as input to an adaptive neuro-fuzzy inference system (ANFIS) to detect the exudates pixels and non-exudates pixels. The proposed method is evaluated on 415 real-time images comprising 200 normal and 215 non-proliferative eyes, and these are compared with the performance of human graders. Performance of the classifier is measured by the receiver operating characteristic curve. ANFIS classifier provides classification with 99.1% accuracy and less convergence time when compared with the performance of back-propagation network. The method can also be used as a screening tool in the analysis and diagnosis of retinal images.