A Feature-Fusion Approach Combining Domain-Specific Descriptors and RNNs for Automated Diabetic Retinopathy Classification

Abstract

The increasing  prevalence of diabetes has significantly raised the incidence of diabetic retinopathy (DR) is essential for preventing irreversible vision impairment and blindness.This study presents a robust binary classification framework for automated DR detection using the Augmented Messidor retinal fundus image dataset. The proposed methodology integrates a heterogeneous set of handcrafted features, including matched filter based vascular descriptors, Gray Level Co-occurrence Matrix (GLCM) texture features, discrete wavelet transform (DWT) based multiresolution features, Histogram of Oriented Gradients (HOG), Local Binary Pattern (LBP) descriptors, patch-level contrast features, and blind image quality assessment metrics. These complementary features are fused into a unified feature representation that captures both structural and textural characteristics of retinal images. The fused feature vector is subsequently classified using a customized Recurrent Neural Network (RNN), which effectively models sequential relationships within the high dimensional feature space to enhance classification performance. Experimental evaluation demonstrates that the proposed framework achieves an overall classification accuracy of 96.11%, with a sensitivity of 95.95%, specificity of 94.20%, precision of 94.78%, and an F1 score of 95.35%, outperforming or matching existing state of the art approaches. Furthermore, binary classification across individual DR severity grades yields accuracies of 94.44% (Grade 0), 97.08% (Grade 1), 95.83% (Grade 2), and 97.78% (Grade 3), demonstrating the robustness of the proposed framework under varying disease conditions. The findings confirm that the integration of carefully engineered handcrafted features with a customized deep learning architecture significantly enhances diabetic retinopathy detection and provides a reliable foundation for future multi class DR grading systems and real world clinical decision support applications.