Comparative Analysis of Loss Functions in BigGAN for Medical Image Synthesis: A Case Study on Brain MRI Scans
Aakriti Pandey, Jatin Trehan, Pintu Kumar Ram, Sofia Singh · 2025
Acquiring high-quality medical images such as MRI scans is essential for diagnostics and treatment planning but remains challenging due to high costs, health risks, and data scarcity. To address this, we propose the use of Generative Ad-versarial Networks (GANs) to synthesize MRI images, alleviating the limitations of medical data availability. This paper presents a fine-tuned BigGAN model, optimized to generate MRI scans representing different stages of Alzheimer's disease. We explore the impact of various loss functions, including Adversarial, Hinge, and Mean Squared Error (MSE) Loss, on model performance. While the outputs showcase the potential of GAN s to generate high-quality medical images, challenges such as mode collapse and overfitting remain. The synthesized images offer significant promise for enhancing AI model training and medical research, particularly in scenarios where acquiring real medical images is difficult.