Optimisation of SOUP-GAN and CSR-GAN for High Resolution MR Images Reconstruction

A research team including Muneeba Rashid and Hina Shakir has published a paper detailing significant optimizations to two Generative Adversarial Network (GAN) architectures—SOUP-GAN and CSR-GAN—specifically for reconstructing high-resolution Magnetic Resonance Imaging (MRI) scans. The work addresses a critical bottleneck in medical diagnostics: MRI outputs are often degraded by patient motion artifacts and equipment limitations, which can obscure crucial details. The researchers' core announcement is that through meaningful architectural modifications and advanced training techniques, they have enhanced these models' ability to generate cleaner, more detailed images from suboptimal scan data, with CSR-GAN showing particular strength in reconstructing high-frequency details.

The technical enhancements involved deepening both the generator and discriminator networks within each GAN by adding convolutional layers and optimizing filter sizes. To combat common GAN training issues like mode collapse and improve gradient flow, the team implemented the LeakyReLU activation function and applied spectral normalization. Rigorous hyperparameter tuning, including a reduced learning rate and optimal batch size, further stabilized training. The results are quantified by standard image quality metrics: CSR-GAN achieved a Peak Signal-to-Noise Ratio (PSNR) of 34.6 and a Structural Similarity Index (SSIM) of 0.89, excelling in detail reconstruction. SOUP-GAN, with a PSNR of 34.4 and SSIM of 0.83, performed best in delivering structurally sound images with minimal noise. The implications are direct for radiology, offering a software-based method to enhance scan clarity for more accurate disease detection, potentially reducing the need for repeat scans and improving diagnostic confidence.