Coevolutionary algorithms in generative adversarial networks for medical image analysis with limited labels

Abstract We propose NU-SSL GAN, a Neuroevolutionary semi-supervised generative adversarial network designed for effective learning under extreme label scarcity. We evaluate NU-SSL GAN on two challenging medical image classification tasks: the HAM10000 skin lesion dataset (pigmented lesion images) and the BreCaHAD breast cancer histopathology dataset (microscopic biopsy images). Our results show that NU-SSL GAN achieves the highest classification accuracy among the GAN-based SSL methods tested, reaching 88.4% on HAM10000 and 98.5% on BreCaHAD under a 7% labeled regime. in both classification accuracy and image generation quality, and attains the highest classification accuracy on both datasets while producing realistic and diverse synthetic images with only a tiny fraction of labeled examples. Notably, NU-SSL GAN markedly narrows the gap with fully supervised performance despite using far fewer labels and training epochs. We attribute NU-SSL GAN’s success to its dynamic evolution of generator and discriminator architectures, which foster robust learning dynamics, avoid mode collapse, and ensure stable convergence under severe label constraints.