Automated retinal disease classification from OCT images using particle swarm-optimized deep learning

Abstract Accurate retinal disease classification from optical coherence tomography (OCT) images depends critically on CNN hyperparameter settings; hand-picked configurations yield inconsistent results, and grid search over an eight-parameter space requires 384 model evaluations. This paper proposes an Adaptive Multi-Strategy Particle Swarm Optimization (AMS-PSO) framework that tunes CNN hyperparameters by assigning each particle to one of three update strategies based on its fitness improvement rate and population diversity. On a large-scale OCT dataset (84,495 images, four pathology classes), AMS-PSO achieves 95.24% test accuracy using only 88 model evaluations, outperforming Bayesian optimization (94.18%), grid search (92.87%), and standard PSO (93.87%). Ablation and benchmark experiments confirm each component’s contribution and the generality of the optimization gains.