[Paper Review] Denoising Simulated Low-Field MRI (70mT) using Denoising Autoencoders (DAE) and Cycle-Consistent Generative Adversarial Networks (Cycle-GAN)

Fernando Vega, Abdoljalil Addeh+1 more|arXiv (Cornell University)|Jul 12, 2023

Advanced MRI Techniques and Applications4 citations

TL;DR

This study proposes a denoising Cycle-GAN framework to reconstruct high-field, high-SNR MRI images from simulated low-field (70mT) MRI scans without requiring paired training data. By leveraging unpaired 3T and low-field MRI images, the model outperforms traditional denoising autoencoders (DAE), achieving a 14.62% higher mean PSNR and superior image quality in contrast and structural fidelity, demonstrating the feasibility of generative deep learning for low-field MRI enhancement.

ABSTRACT

In this work, a denoising Cycle-GAN (Cycle Consistent Generative Adversarial Network) is implemented to yield high-field, high resolution, high signal-to-noise ratio (SNR) Magnetic Resonance Imaging (MRI) images from simulated low-field, low resolution, low SNR MRI images. Resampling and additive Rician noise were used to simulate low-field MRI. Images were utilized to train a Denoising Autoencoder (DAE) and a Cycle-GAN, with paired and unpaired cases. Both networks were evaluated using SSIM and PSNR image quality metrics. This work demonstrates the use of a generative deep learning model that can outperform classical DAEs to improve low-field MRI images and does not require image pairs.

Motivation & Objective

• To address the challenge of low signal-to-noise ratio (SNR) and low resolution in low-field MRI (70mT) without relying on paired image data.
• To improve image quality in low-field MRI by leveraging deep generative models that do not require image registration or alignment.
• To evaluate whether Cycle-GAN can outperform classical denoising autoencoders (DAE) in reconstructing high-field quality images from simulated low-field data.
• To develop a pipeline that enables high-quality MRI reconstruction from low-field scans using unpaired data, reducing dependency on expensive high-field scanners.

Proposed method

• Simulated low-field MRI data were generated by resampling 3T T1-weighted MRI images (1mm³ isotropic) to 1.5mm³ resolution and adding Rician noise to emulate a 70mT field strength.
• A 3D Cycle-GAN architecture was implemented using the MONAI deep learning framework, consisting of two generators and two discriminators trained end-to-end over 500 epochs.
• The model uses 13 layers with 9 residual blocks as a bottleneck, diverging from U-Net architectures common in DAEs, and operates on unpaired high-field and low-field image distributions.
• The generators learn to translate low-field noisy images into high-field-like denoised images and vice versa, while discriminators distinguish real from generated images.
• The training process enforces cycle consistency, ensuring that translating an image and back preserves its original content.
• Performance was evaluated using unseen test images with SSIM and PSNR metrics, comparing the Cycle-GAN against a standard DAE trained on paired data.

Experimental results

Research questions

• RQ1 Can a Cycle-GAN-based approach effectively denoise simulated low-field MRI (70mT) without requiring paired training images?
• RQ2 How does the image quality of Cycle-GAN-generated images compare to that of a traditional DAE in terms of PSNR and SSIM?
• RQ3 Does the Cycle-GAN preserve anatomical structure and contrast better than a DAE in low-SNR, low-resolution MRI reconstructions?
• RQ4 To what extent can unpaired generative models outperform paired autoencoder-based methods in low-field MRI denoising?

Key findings

• The Cycle-GAN model achieved a mean PSNR that was 14.62% higher than the DAE on unseen test images, indicating superior noise reduction performance.
• Although the DAE scored 1.15% higher in SSIM, PSNR is a more sensitive metric for noise comparison, and the Cycle-GAN's higher PSNR indicates better denoising capability.
• Visual inspection confirmed that Cycle-GAN generated images exhibited better contrast, sharper edges, and more accurate anatomical shape preservation than DAE outputs.
• The Cycle-GAN successfully reconstructed high-field quality images from low-field simulations without requiring image pairing or registration, demonstrating robustness to distribution shifts.
• The model's performance on unpaired data suggests it can generalize well to real-world low-field MRI applications where paired data are often unavailable.
• The results indicate that Cycle-GANs can serve as a viable alternative to DAEs for low-field MRI reconstruction, particularly when paired data are impractical to obtain.