Benchmarking Vanilla GAN, DCGAN, and WGAN Architectures for MRI Reconstruction: A Quantitative Analysis

Technical Summary: Benchmarking Vanilla GAN, DCGAN, and WGAN for MRI Reconstruction

Problem Statement
Magnetic Resonance Imaging (MRI) is a vital diagnostic tool for visualizing soft tissues in the knee, brain, and heart. However, conventional MRI acquisition is time-consuming, often leading to motion artifacts, increased costs, and patient discomfort. Traditional reconstruction methods, such as Compressed Sensing (CS-MRI) and iterative techniques, frequently struggle to maintain optimal clarity, contrast, and detail preservation. While Generative Adversarial Networks (GANs) have emerged as a promising solution for reconstructing high-quality images from undersampled data, there is a lack of direct, side-by-side comparisons of standard GAN architectures (Vanilla GAN, DCGAN, and WGAN) evaluated across multiple anatomical regions under a unified preprocessing pipeline.

Methodology
The study employs a comparative experimental framework to evaluate three distinct GAN architectures:

1. Vanilla GAN: Utilizes a standard adversarial setup with fully connected layers and binary cross-entropy loss.
2. Deep Convolutional GAN (DCGAN): Replaces fully connected layers with convolutional and transposed convolutional layers to better capture spatial dependencies and anatomical details. It incorporates batch normalization and LeakyReLU activations to stabilize training.
3. Wasserstein GAN (WGAN): Replaces the standard discriminator with a "critic" and utilizes the Wasserstein distance (Earth Mover's distance) as a loss function. This approach enforces Lipschitz continuity via weight clipping to ensure stable convergence and mitigate mode collapse.

Datasets and Preprocessing
The models were trained and tested on three distinct, anonymized MRI datasets to assess generalizability:

• Knee: 1,000 images from the NYU Langone Health fastMRI database (k-space data).
• Heart: 805 images from the Sunnybrook Cardiac Data (SCD) dataset.
• Brain: 90 images from the Jordan University Hospital (JUH) dataset (T2-weighted sequences).

All datasets underwent preprocessing, including normalization, restructuring, and conversion from DICOM to PNG format. The data was split into 70% for training and 30% for testing. Hyperparameters were tuned specifically for each model, with Vanilla GAN trained for 2,000 epochs (due to slower convergence) and DCGAN/WGAN for 1,000 epochs.

Key Contributions

• Cross-Organ Benchmarking: This work provides the first cross-organ benchmark of baseline GANs (Vanilla, DCGAN, WGAN) across knee, heart, and brain MRI datasets using a consistent preprocessing pipeline.
• Quantitative Analysis: The study offers a rigorous statistical evaluation of these architectures using Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Inception Score (IS).
• Statistical Validation: One-way ANOVA and post-hoc Tukey's HSD tests were conducted to validate the significance of performance differences between the models.

Results
The quantitative analysis reveals distinct performance hierarchies among the architectures:

• Structural Similarity (SSIM): WGAN achieved the highest score (0.99), followed closely by DCGAN (0.97), while Vanilla GAN scored significantly lower (0.84).
• Peak Signal-to-Noise Ratio (PSNR): WGAN demonstrated superior noise reduction with a PSNR of 49.3, compared to DCGAN (43.5) and Vanilla GAN (26).
• Inception Score (IS): Both DCGAN and WGAN achieved high scores (9.0), indicating high image diversity and sharpness, whereas Vanilla GAN lagged at 2.9.
• Statistical Significance: ANOVA results confirmed that the differences in SSIM and PSNR between the models are statistically significant (p < 0.00001). Post-hoc tests confirmed that both DCGAN and WGAN significantly outperform Vanilla GAN, with WGAN showing a statistically significant, albeit smaller, edge over DCGAN in SSIM and PSNR.

Visual comparisons across all three anatomical regions (Figures 5–7) corroborate these metrics, showing that DCGAN and WGAN produce reconstructions with superior anatomical fidelity and fewer artifacts compared to Vanilla GAN.

Significance and Claims
The paper claims that DCGAN and WGAN-based frameworks are highly promising for MR image reconstruction due to their ability to generate high-quality images with superior accuracy. The study emphasizes that while Vanilla GAN serves as a foundational baseline, it suffers from training instability and lower image quality. In contrast, the architectural enhancements in DCGAN (convolutional layers) and the loss function improvements in WGAN (Wasserstein distance) effectively resolve training instabilities and mode collapse.

The authors position this work as a reproducible benchmark for future research, specifically for the development of hybrid GANs and clinical MRI applications. By establishing a solid baseline across multiple body regions, the study aims to guide future modifications of GAN models to address real-world limitations in medical imaging. The paper modestly acknowledges that while Inception Score was used as a perceptual metric, it is not specifically designed for grayscale medical images, suggesting future work could benefit from domain-specific alternatives like FID or clinical scoring.