ABSTRACT Generative adversarial networks (GANs) have shown remarkable potential in medical image synthesis but face persistent challenges in achieving diagnostic‐grade quality, particularly in preserving anatomical edges and avoiding mode collapse. To address these limitations, we present a stable edge‐aware GAN architecture featuring two key innovations: dynamically learnable bilateral kernels that adaptively enhance structural gradients during training, and a layered generator with interpolated skip connections to maintain spatial coherence. Our methodology leverages adversarial training with Wasserstein regularization on the CT Kidney Dataset (12,446 images), optimizing for both global fidelity and local precision. Comprehensive experiments demonstrate the model's superiority through quantitative metrics—achieving a 43% improvement in Fréchet Inception Distance (FID = 87 vs. DCGAN's 149, p < 0.01), 16% higher edge sharpness (Sobel gradient magnitude 45.2 ± 3.1 vs. 38.9 ± 4.2), and 0.82 ± 0.05 SSIM scores. Clinical validation by board‐certified radiologists confirmed 89% diagnostic plausibility for synthetic images, with particular praise for tumor boundary delineation. The architecture also shows exceptional training stability, reducing loss fluctuations by 34% compared to conventional GANs while efficiently scaling to 128 × 128 resolution. These results establish a new benchmark for privacy‐preserving medical data augmentation, offering immediate value for scenarios with limited annotated datasets. Future directions include extension to 3D volumetric synthesis and integration with diffusion models for multi‐modal applications, potentially revolutionizing how healthcare institutions generate and share synthetic patient data without compromising privacy.
Vavekanand et al. (2026) studied this question.