DRAGONFLY SWARM INTELLIGENCE ENHANCED SEMI SUPERVISED GENERATIVE ADVERSARIAL NETWORK FOR RETINAL FEATURE OPTIMIZATION

Abstract

Automated diabetic retinopathy analysis demands models capable of high reliability under sparse supervision, where conventional supervised architectures often fail to generalise across limited annotated data. The present study introduces Dragonfly Swarm Optimised Semi Supervised Generative Adversarial Network (DSO-SSGAN), a biologically inspired framework designed for robust multi-stage retinal image classification. The approach integrates dragonfly swarm intelligence into the latent generative process, regulating search trajectories, enhancing feature diversity, and stabilising convergence within adversarial learning. A semi-supervised architecture combines confidence-based filtering, domain alignment, and progressive curriculum scheduling to align labelled and unlabelled representations. The model is trained and validated on the Diabetic Retinopathy Dataset, achieving an accuracy of 77.512%, surpassing conventional semi-supervised baselines with consistent gradient stability and reduced overfitting. Interpretability is strengthened through latent traversal analysis that visualises class transitions and reinforces clinical traceability. The unified framework demonstrates computational resilience and adaptive visual encoding under limited supervision, positioning swarm-driven latent optimisation as an efficient strategy for scalable diabetic retinopathy classification and future medical image intelligence systems.