Prediction of Diabetic Retinopathy Using Fusion of Hierarchically Reduced Color Texture Features

This paper proposes a novel hierarchical feature dimensionality reduction (FDR) methodology for fusion of features from multiple color spaces. The objective is to automatically recognize diabetic retinopathy (DR) stages; an eye disease that attacks one in three diabetes carriers in USA. Our work examines: (1) the role of FDR in extracting discriminatory features and (2) the discriminatory information carried in different color spaces on fundus images. Effectiveness of the proposed methodology is evaluated on the challenging fine grained diabetic retinopathy image dataset. Features are extracted from color retina images by local binary pattern descriptors at multiple scales, on three channels of each color space, and organized as a feature hierarchy. Proposed hierarchical FDR methodology performs feature reduction by PCA, LDA, random forest, and multi-layer perceptron on the hierarchy of features. Our experiments show (i) the hierarchical FDR methodology with multi-layer perceptron efficiently transforms features to lower dimensional space by achieving best accuracy. (ii) To investigate the discriminatory power of color spaces; FDR methodology and classification is evaluated on RGB, Lab*, HSV, YIQ, and rgb color spaces with best classification accuracy of 74.26 % on RGB, and (iii) the novel combination of the proposed FDR method with fusion of multiple color features achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 5. 6 9 \%}$</tex> accuracy by one-vs-one SVM classifier.