[Paper Review] Over-and-Under Complete Convolutional RNN for MRI Reconstruction

Pengfei Guo, Jeya Maria Jose Valanarasu+4 more|arXiv (Cornell University)|Jun 16, 2021

Advanced MRI Techniques and Applications30 references4 citations

TL;DR

This paper proposes Over-and-Under Complete Convolutional RNN (OUCR), a novel deep learning architecture for accelerated MRI reconstruction that combines an overcomplete and an undercomplete Convolutional Recurrent Neural Network (CRNN) to enhance local feature learning and global structure preservation. The method achieves state-of-the-art image quality with fewer parameters than existing deep learning and compressed sensing methods.

ABSTRACT

Reconstructing magnetic resonance (MR) images from undersampled data is a challenging problem due to various artifacts introduced by the under-sampling operation. Recent deep learning-based methods for MR image reconstruction usually leverage a generic auto-encoder architecture which captures low-level features at the initial layers and high-level features at the deeper layers. Such networks focus much on global features which may not be optimal to reconstruct the fully-sampled image. In this paper, we propose an Over-and-Under Complete Convolutional Recurrent Neural Network (OUCR), which consists of an overcomplete and an undercomplete Convolutional Recurrent Neural Network(CRNN). The overcomplete branch gives special attention in learning local structures by restraining the receptive field of the network. Combining it with the undercomplete branch leads to a network which focuses more on low-level features without losing out on the global structures. Extensive experiments on two datasets demonstrate that the proposed method achieves significant improvements over the compressed sensing and popular deep learning-based methods with less number of trainable parameters.

Motivation & Objective

• To address the challenge of reconstructing high-quality MR images from highly undersampled k-space data, which often results in aliasing artifacts and loss of fine details.
• To improve upon existing deep learning-based MRI reconstruction methods that prioritize global features at the expense of local, low-level structures.
• To leverage overcomplete representations for enhanced local feature extraction while preserving global context through an undercomplete branch.
• To develop a parameter-efficient architecture that outperforms both compressed sensing and leading deep learning methods in reconstruction quality.

Proposed method

• The OUCR framework consists of two parallel branches: an overcomplete CRNN (OC-CRNN) that restricts receptive field to emphasize local features, and an undercomplete CRNN (UC-CRNN) that captures global structures.
• The overcomplete branch uses upsampling layers instead of max-pooling to maintain fine-grained spatial information and enhance local feature learning.
• The undercomplete branch employs downsampling to progressively expand the receptive field and learn high-level representations.
• A residual module (RM) is applied to both branches to improve gradient flow and feature refinement.
• The outputs of the two branches are fused and jointly optimized through end-to-end training to reconstruct fully-sampled MR images from undersampled k-space data.
• The model is trained using a combination of L2 loss and perceptual losses to preserve image details and structural fidelity.

Experimental results

Research questions

• RQ1 Can an overcomplete CRNN architecture improve the recovery of fine details in undersampled MRI reconstruction compared to standard auto-encoder style networks?
• RQ2 Does combining overcomplete and undercomplete CRNN branches lead to better reconstruction performance than using either branch alone?
• RQ3 Can the proposed OUCR architecture achieve superior image quality with fewer trainable parameters than state-of-the-art deep learning and compressed sensing methods?
• RQ4 How does the inclusion of a residual module affect the performance and convergence of the OUCR framework?

Key findings

• On the HPKS dataset at acceleration factor 8, OUCR achieved a PSNR of 32.14 dB and SSIM of 0.9044, significantly outperforming CS (24.96 dB, 0.7125) and UNet (29.47 dB, 0.8249).
• On the fastMRI dataset at AF=8, OUCR achieved PSNR of 30.59 dB and SSIM of 0.6634, surpassing KIKI-Net (29.27 dB, 0.6355) and D5C5 (29.65 dB, 0.6457).
• The ablation study confirmed that combining UC-CRNN and OC-CRNN with the residual module yielded the highest PSNR (39.33 dB) and SSIM (0.9747) on HPKS at AF=4.
• The OUCR model achieved state-of-the-art performance with only 1.192 million parameters, making it more parameter-efficient than UNet (8.634M), Kiu-net (7.951M), and D5C5 (2.237M).
• K-space analysis showed that OUCR outperformed all baselines in reconstructing both low- and high-frequency components, indicating improved preservation of image contrast and fine details.
• Statistical significance tests confirmed that the improvements of OUCR over other methods were highly significant (p < 10^-5).