Spectral Normalisation and Efficient Perceptual Conditional Generative Adversarial Networks for High‐Quality Medical Image Generation

Conditional generative adversarial networks (cGANs) extend conventional GANs by incorporating class information into the image generation process, enabling controlled and class‐consistent synthesis. In the medical domain, cGANs are increasingly important for addressing the scarcity of high‐quality annotated datasets, which limits the development of robust diagnostic models. However, existing cGAN‐based approaches often suffer from unstable training, mode collapse, and insufficient preservation of fine‐grained diagnostic features. In this study, we propose Spectral Normalisation & Efficient Perceptual‐cGAN (SNEP‐cGAN), which is designed to generate high‐fidelity medical images. The model integrates spectral normalisation for training stability, residual up‐sampling and down‐sampling blocks with squeeze‐and‐excitation attention for enhanced feature representation, and a perceptual loss computed using a pretrained EfficientNetB0 network to improve visual realism. A gradient penalty is further employed to enforce Lipschitz continuity. The proposed model is evaluated on the NIH Malaria Cell Dataset and a breast cancer imaging dataset, achieving a remarkably low Fréchet Inception Distance (FID) of 4.28 and strong class‐wise diversity, notably outperforming baseline models. These results demonstrate that SNEP‐cGAN provides a robust and effective solution for medical image synthesis, offering strong potential for dataset augmentation and improved computer‐aided diagnosis in healthcare applications.