A Hybrid Deep Learning Framework for Pattern Classification of Diabetic Retinopathy Stages Using Retinal Fundus Image Analysis

Diabetic retinopathy (DR) is a progressive microvascular disease of diabetes, and accurately classifying its disease stages from retinal fundus images is difficult because of subtle DR lesions, considerable similarity between disease stages, varying illumination, image quality degradation, and complex retinal morphology. Accurate, timely diagnosis is crucial for prompt clinical evaluation and to minimise the risk of irreversible vision loss. The presents a Hybrid Deep Learning Framework for Pattern Classification of Diabetic Retinopathy Stages, combining complementary deep feature-learning and classification mechanisms to learn discriminative retinal features and stage-specific pathological patterns. The framework preprocesses retinal fundus images and represents their features, followed by hybrid deep feature extraction and multiclass classification to classify various severity stages in DR. Its main advantage is the fusion of complementary learning capabilities, which allows for the development of a more detailed characterization of microaneurysms, haemorrhages, exudates, and other abnormalities of the retina associated with the stage of the disease, and increases the reliability of classification in various imaging conditions. Experiments are performed using the Diabetic Retinopathy Detection dataset, and this dataset is used as a challenging benchmark for automatic retinal disease-stage detection. The proposed framework exhibits good discriminative performance with an accuracy of 95.84%, precision of 94.72%, recall of 94.36% and F1-score of 94.54% in multi-stage DR classification. The automated method thus developed can be used for more efficient screening, can avoid manual evaluation of images, can contribute to early detection of disease, and can offer computational support for larger-scale applications in ophthalmology decision support systems.