A framework of hybrid deep learning and nature-inspired algorithms for accurate multi-class fundus disease classification.

Fundus imaging is an essential technique for detecting anatomical changes indicative of various ophthalmological diseases. These alterations-including changes in the macula, optic disc, fovea, and blood vessels-can signal conditions such as diabetic retinopathy, glaucoma, age-related macular degeneration, cataracts, and myopia. In this paper, we propose a novel hybrid framework for classifying multi-class fundus images by combining deep learning feature extraction with nature-inspired algorithms. Pre-trained ResNet-18 and GoogLeNet architectures are employed to extract features, which are subsequently refined using nature-inspired algorithms including Particle Swarm Optimization, Grey Wolf Optimizer, Differential Evolution, Firefly Algorithm, and Genetic Algorithm. Finally, machine learning classifiers, such as K-Nearest Neighbor, Gaussian Naive Bayes, Support Vector Machine, and Logistic Regression, are applied to the optimized feature sets. Our extensive experiments on the ODIR and RFMiD datasets demonstrate that the proposed framework-specifically the combination of the Firefly Algorithm with K-Nearest Neighbor (FAKNN)-achieves state-of-the-art performance, reaching 100% accuracy, recall, precision, Kappa, F1 score, and Area Under the Curve. This approach not only improves diagnostic accuracy but also demonstrates significant potential for clinical application, facilitating early intervention and better patient outcomes.