Early Stage Detection of Diabetic Retinopathy Using an Optimal Feature Set

Diabetic Retinopathy (DR) is the most common source of blindness in the current population worldwide. The development of an automated system will assists to ophthalmologists. DR is a worsening disease, hence early detection is important for diagnosis and proper treatment to prevent blindness. Microaneurysms (MAs) are the first signs of DR; hence their accurate detection is necessary for early stage detection of DR. This paper proposes a three stage system to detect all MAs in the retinal fundus image. First stage extracts all possible candidates using morphological operations and Gabor filter. Feature vector using statistical, gray scale and wavelet features for each candidate is formed in second stage. In the last stage, classification of these candidates as MAs and non MAs is performed using a multilayered feed forward neural network (FFNN) classifier and support vector machine (SVM) classifier. The main objective of the proposed work is to propose a list of important and optimal features for MA detection using the most common features used in the literature. The experiments have been performed on the database DIARETDB1 to evaluate the proposed system. The evaluation parameters accuracy, sensitivity and specificity are obtained as 92%, 79%, 90% and 95%, 76%, 92% respectively for FFNN and SVM classifiers.