Abstract
The worldwide loss in human vision is primarily associated with Diabetic Retinopathy (DR). It occurs due to accelerated levels of blood sugar thereby causing perforation, bulging and leakage of retinal blood vessels (BVs). DR commences with the emergence of small blood spots on the retinal surface known as Microaneurysms (MAs) that are subsequently transformed into heavy blood deposits called Hemorrhages (HGs). This paper proposes an optimized and computationally inexpensive digital image processing (DIP) technique for detection and classification of 'Retinal Red Lesions' (RRLs) i.e. MAs and HGs using green channel of the digital fundus images. The basic essence of the proposed technique revolves around regional spatial transformations detection performed through region based spatial filtering, matching features and neural networks classification. The proposed technique comprises of five main stages i.e. Pre-processing, Regional Spatial Transformations, Optimization, Features extraction and Classification. Speed Up Robust Features (SURF) algorithm has been used for features selection & extraction while Feed-forward Back-propagation Artificial Neural Network (FFBP ANN) has been used for classification. The proposed technique has been successfully applied on commercially available digital fundus image data-set and has yielded 98.4% 'Sensitivity' (SE), 94% 'Specificity' (SP) and 98% `Accuracy' (AC). The SE, SP and AC have also been compared with other RRLs detection methods and has shown highly promising and encouraging results.