Abstract
INTRODUCTION: The availability of large-scale medical imaging datasets is often constrained by privacy regulations, high acquisition costs, and ethical concerns. Synthetic medical image generation using generative adversarial networks (GANs) offers a promising solution to overcome these limitations. This study investigates the effectiveness of a Parameter-Optimized Generative Adversarial Network (POP-GAN) and compares its performance with state-of-the-art architectures, including StyleGAN2, multi-stream GAN (mustGAN), and Conditional GAN (cGAN), for realistic MRI image synthesis. METHODS: The proposed framework integrates progressive growing strategies with optimized hyperparameters, including a batch size of 256, learning rate of 1 × 10(-4), dropout rate of 0.3, and a buffer size of 6,000. All models were trained to generate MRI images at a resolution of 128 × 128. Performance was evaluated using quantitative metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), and Fréchet Inception Distance (FID), along with expert-based clinical realism scoring. RESULTS: POP-GAN demonstrated a 27% reduction in MSE compared with the baseline model (from 6.58 × 10(-3) to 4.81 × 10(-3)), achieved higher PSNR, and reduced FID from 32.91 to 24.36. cGAN achieved the lowest MAE (3.50 × 10(-3)), indicating superior reconstruction accuracy. mustGAN produced the strongest resolution fidelity, while StyleGAN2 delivered the highest perceptual realism. POP-GAN also attained a clinical realism score of 4.13 out of 5. DISCUSSION: The results demonstrate that parameter optimization and progressive training substantially enhance synthetic MRI quality. POP-GAN provides a balanced trade-off between reconstruction accuracy, perceptual realism, and clinical relevance, supporting its potential for privacy-preserving dataset augmentation and robust medical imaging research.