Automated recognition of lesions in retinal images using artificial neural networks

This study presents an automatic recognition system for different disease lesions (hard exudate (HE), hemorrhage (HMR), microaneurysm (MA), soft exudate (SE) and non-lesion (normal, vessel, optic disc, macula) (N) patterns in retinal images. Proposed method consists of thresholding, morphological operations, filtering, image enhancement, optic disc and macula localization, segmentations, optic disc and vessel elimination, region growing, classification and recognition for four different disease lesions and non-lesion patterns. Artificial Neural Networks (ANNs), Support Vector Machines (SVM) and Radial Basis Function (RBF) were used as classifier to recognize. The features are extracted from the images and fed to input of the ANN. Results were compared with expert ophthalmologists´ hand-drawn ground-truth. Experimental results obtained were presented and recognition performances of the system for Multi-Layer Perceptron (MLP), Radial Basis Function (RBF) and Support Vector Machines (SVM) classifiers were compared and discussed.