Spherical Directional Feature Extraction with Artificial Neural Networkfor Diabetic Retinopathy Classiftcation

Abstract

Since last two decades one of the fast advancing and most sensitive research area is observed to be detection of diabetic retinopathy (DR). In fundus images detection of retinal lesions depends on grading of diabetic retinopathy and computer-aided screening which led to development of automatic telemedicine system. The detection accuracy is still a matter of concern even after existence of huge contribution in area of detection. The existing algorithm for classification of DR images is not able to encode the directional information in 2D and 3D plane. The proposed approach encodes in four different directions (0°, 45°, 90° and 135°) from the reference pixel to its surrounding pixel in 3D plane. The proposed model includes preprocessing, feature extraction using spherical directional local ternary pattern (SDLTP) and classification using traditional distance measure and learning based distance measure using artificial neural network (ANN). SDLTP is used for extracting the directional feature in 3D plane and to reduce the feature vector length, a principle component analysis (PCA) technique is adopted. Further, two techniques are used for classification purpose (distance measure and ANN). The proposed method classification accuracy is measured in terms of precision. From experimental analysis, the proposed method give significant improvement in classification accuracy in both unsupervised and supervised domain because feature extraction in implemented considering the directional information.