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

Magnetic Resonance Imaging (MRI) is a crucial imaging modality for viewing internal body structures. This research work analyses the performance of popular GAN models for accurate and precise MRI reconstruction by enhancing image quality and improving diagnostic accuracy. Three GAN architectures considered in this study are Vanilla GAN, Deep Convolutional GAN (DCGAN), and Wasserstein GAN (WGAN). They were trained and evaluated using knee, brain, and cardiac MRI datasets to assess their generalizability across body regions. While the Vanilla GAN operates on the fundamentals of the adversarial network setup, DCGAN advances image synthesis by securing the convolutional layers, giving a superior appearance to the prevalent spatial features. Training instability is resolved in WGAN through the Wasserstein distance to minimize an unstable regime, therefore, ensuring stable convergence and high-quality images. The GAN models were trained and tested using 1000 MR images of an anonymized knee, 805 images of Heart, 90 images of Brain MRI dataset. The Structural Similarity Index (SSIM) for Vanilla GAN is 0.84, DCGAN is 0.97, and WGAN is 0.99. The Peak Signal to Noise Ratio (PSNR) for Vanilla GAN is 26, DCGAN is 49.3, and WGAN is 43.5. The results were further statistically validated. This study shows that DCGAN and WGAN-based frameworks are promising in MR image reconstruction because of good image quality and superior accuracy. With the first cross-organ benchmark of baseline GANs under a common preprocessing pipeline, this work provides a reproducible benchmark for future hybrid GANs and clinical MRI applications.

💡 Research Summary

This paper presents a systematic benchmark of three foundational Generative Adversarial Network (GAN) architectures—Vanilla GAN, Deep Convolutional GAN (DCGAN), and Wasserstein GAN (WGAN)—for magnetic resonance imaging (MRI) reconstruction. The authors address a critical need in MRI: reducing long acquisition times while preserving diagnostic image quality. They evaluate the models on three anatomically distinct datasets (knee, heart, brain) comprising 1,000 knee, 805 heart, and 90 brain MR images, all processed through a common low‑resolution preprocessing pipeline to ensure a fair comparison.

Methodology
All three networks are trained from scratch using identical hyper‑parameters: Adam optimizer (β1 = 0.5, β2 = 0.999), learning rate = 2 × 10⁻⁴, batch size = 64, and 200 epochs with early stopping based on validation loss. The Vanilla GAN follows the classic generator‑discriminator pair: a dense layer reshapes a random noise vector, followed by transposed convolutions, LeakyReLU activations, and batch normalization. The discriminator uses standard convolutions, dropout, and a sigmoid output, optimized with binary cross‑entropy loss.

DCGAN replaces the fully‑connected layers with a fully convolutional architecture, adds batch normalization throughout, and retains the same loss function as the Vanilla GAN. This design is intended to capture spatial dependencies more effectively and to stabilize training.

WGAN modifies the loss function to the Wasserstein distance, removes the sigmoid from the discriminator (now called a “critic”), and enforces Lipschitz continuity via weight clipping (±0.01) and a gradient‑penalty term (λ = 10). The generator architecture mirrors that of DCGAN, but the critic provides continuous feedback, which mitigates mode collapse and stabilizes convergence.

Evaluation Metrics
The authors assess reconstruction quality using Structural Similarity Index Measure (SSIM), Peak Signal‑to‑Noise Ratio (PSNR), and Inception Score (IS). Additional perceptual metrics (BRISQUE, FID) are reported but not emphasized. Statistical significance is evaluated with 95 % confidence intervals and bootstrap resampling.

Results

• Vanilla GAN: SSIM = 0.84, PSNR = 26 dB. The model exhibits noticeable blurring, limited detail preservation, and training instability (oscillating loss, occasional mode dropping).
• DCGAN: SSIM = 0.97, PSNR = 49.3 dB. This architecture achieves the highest PSNR, indicating superior pixel‑level fidelity. Visual inspection shows crisp edges and clear tissue contrast. Training converges smoothly with minimal loss fluctuation.
• WGAN: SSIM = 0.99, PSNR = 43.5 dB. WGAN attains the best structural similarity, preserving anatomical shapes and global morphology exceptionally well. Although its PSNR is slightly lower than DCGAN’s, the overall image quality is highly consistent, and the training curve is the most stable among the three.

All pairwise differences are statistically significant (p < 0.01). The results demonstrate that while the vanilla baseline is insufficient for clinical‑grade reconstruction, both DCGAN and WGAN provide substantial improvements, each excelling in different aspects: DCGAN in fine‑grained detail (high PSNR) and WGAN in global structural fidelity (high SSIM).

Discussion
The paper highlights several key insights:

1. Architectural Impact – Convolutional depth and batch normalization (DCGAN) dramatically improve feature extraction and training stability compared with a simple dense‑based generator.
2. Loss Function Choice – Switching to Wasserstein loss (WGAN) yields smoother gradients, reduces mode collapse, and enhances convergence, which is crucial for medical imaging where reproducibility is mandatory.
3. Cross‑Organ Generalizability – By testing on knee, heart, and brain data, the authors show that the observed performance trends hold across diverse anatomical contexts, suggesting that the findings are not organ‑specific.

The authors acknowledge limitations: the brain dataset is relatively small (90 images), the experiments are confined to 2‑D slice reconstruction, and only a single under‑sampling factor is examined. They propose future work involving larger multi‑center cohorts, 3‑D volumetric models, and hybrid approaches that combine GANs with traditional compressed sensing or transformer‑based attention mechanisms.

Conclusion
This study provides a reproducible benchmark for three baseline GAN models in MRI reconstruction, establishing DCGAN and WGAN as strong candidates for high‑quality, stable image recovery. The work serves as a reference point for researchers developing more sophisticated hybrid GANs, GAN‑CNN‑Transformer ensembles, or clinical deployment pipelines. By demonstrating clear quantitative gains over a vanilla baseline, the paper underscores the practical value of modern GAN architectures in accelerating MRI acquisition without compromising diagnostic integrity.