Imbalanced medical datasets significantly impair the performance of diagnostic models by biasing predictions towards majority classes. This study introduces DGAN-WP-TL, a multi-domain generative adversarial network that integrates Wasserstein loss with gradient penalty, VGG19-based perceptual regularisation, and transfer learning to generate high-quality synthetic medical images across three modalities: BUSI ultrasound, CBIS-DDSM mammography, and Alzheimer MRI. Quantitative evaluation shows that DGAN-WP-TL achieves substantial improvements over StyleGAN2 on BUSI and Alzheimer MRI datasets, reducing KID from 0.3797 to 0.1448 and FID from 330.94 to 179.42 on BUSI, and lowering KID from 0.7097 to 0.3567 and FID from 458.74 to 289.79 on Alzheimer MRI. LPIPS (real – fake) and (fake – fake) scores indicate both enhanced perceptual realism and diversity. On CBIS-DDSM, DGAN-WP-TL surpasses StyleGAN2 in LPIPS and MS-SSIM diversity metrics, while StyleGAN2 retains slightly better KID and FID. An ablation study confirms the contribution of each architectural component, with perceptual regularisation and transfer learning yielding the most pronounced gains in realism and diversity. Downstream classification experiments using the augmented datasets demonstrate measurable performance improvement. DGAN-WP-TL offers a robust, multi-domain solution for synthetic medical data generation in low-data regimes. Future research will focus on incorporating domain-specific priors, attention-based synthesis, and clinical validation.
Ryspayeva et al. (Fri,) studied this question.