A Hybrid Retinal Fundus Image Processing and Neural Network Framework for Diabetic Retinopathy Screening

Abstract

Diabetic retinopathy (DR) is a retinal disorder that can be screened from fundus photographs using computer-aided image analysis. This paper presents a texture-based framework that combines image preprocessing, Gray-Level Co-occurrence Matrix (GLCM) feature extraction, Principal Component Analysis (PCA), and neural-network classification for automated retinal image screening. The APTOS 2019 Blindness Detection dataset is used as the experimental data source. The preprocessing stage improves retinal structure visibility through normalization and morphological operations. GLCM is employed to represent spatial texture information, while PCA is used to reduce redundancy in the extracted feature space. The reduced representation is supplied to a feedforward neural-network classification stage for automatic discrimination of retinal disease classes. The framework combines interpretable handcrafted texture descriptors with learned nonlinear classification rather than relying on a single representation. Three experimental executions produced accuracies of 90.00%, 90.63%, and 96.00%, with corresponding specificities of 93.44%, 93.44%, and 95.00% and sensitivities of 91.90%, 95.12%, and 95.50%. The highest observed accuracy was 96.00%. These values are reported as experimental observations, and the best result should be interpreted in the context of the dataset partition and evaluation protocol used.