Parameter-optimized generative adversarial network framework for synthetic MRI generation: fine-tuning critical...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical imaging dataset availability is limited by privacy, cost, and ethics.
• Generative Adversarial Networks (GANs) offer a solution for synthetic medical image generation.
• Existing GAN architectures face challenges in achieving high-fidelity and realistic MRI synthesis.

Purpose of the Study:

• To investigate the effectiveness of a Parameter-Optimized Generative Adversarial Network (POP-GAN) for realistic MRI image synthesis.
• To compare POP-GAN's performance against state-of-the-art GAN architectures: StyleGAN2, mustGAN, and cGAN.
• To evaluate the impact of parameter optimization and progressive growing strategies on synthetic MRI quality.

Main Methods:

• Developed POP-GAN integrating progressive growing with optimized hyperparameters (batch size 256, learning rate 1e-4, dropout 0.3, buffer size 6,000).
• Trained all models to generate MRI images at 128x128 resolution.
• Evaluated performance using quantitative metrics (MSE, MAE, PSNR, FID) and expert clinical realism scoring.

Main Results:

• POP-GAN achieved a 27% reduction in MSE and reduced FID from 32.91 to 24.36 compared to baseline.
• cGAN showed the lowest MAE (3.50e-3), indicating superior reconstruction accuracy.
• StyleGAN2 provided the highest perceptual realism, while mustGAN excelled in resolution fidelity. POP-GAN scored 4.13/5 in clinical realism.

Conclusions:

• Parameter optimization and progressive training significantly improve synthetic MRI quality.
• POP-GAN offers a balanced approach to reconstruction accuracy, perceptual realism, and clinical relevance.
• POP-GAN shows potential for privacy-preserving dataset augmentation and advancing medical imaging research.