[Paper Review] DeepcomplexMRI: Exploiting deep residual network for fast parallel MR imaging with complex convolution

Shanshan Wang, Huitao Cheng+6 more|arXiv (Cornell University)|Jun 11, 2019

Advanced MRI Techniques and Applications52 references17 citations

TL;DR

DeepcomplexMRI proposes a deep residual convolutional neural network with complex-valued convolutions to accelerate parallel MRI reconstruction by leveraging multi-channel k-space data and enforcing data consistency across network layers. It achieves superior image quality by modeling the correlation between real and imaginary components of MR images, outperforming state-of-the-art methods on in vivo datasets.

ABSTRACT

This paper proposes a multi-channel image reconstruction method, named DeepcomplexMRI, to accelerate parallel MR imaging with residual complex convolutional neural network. Different from most existing works which rely on the utilization of the coil sensitivities or prior information of predefined transforms, DeepcomplexMRI takes advantage of the availability of a large number of existing multi-channel groudtruth images and uses them as labeled data to train the deep residual convolutional neural network offline. In particular, a complex convolutional network is proposed to take into account the correlation between the real and imaginary parts of MR images. In addition, the k space data consistency is further enforced repeatedly in between layers of the network. The evaluations on in vivo datasets show that the proposed method has the capability to recover the desired multi-channel images. Its comparison with state-of-the-art method also demonstrates that the proposed method can reconstruct the desired MR images more accurately.

Motivation & Objective

• To accelerate parallel MRI acquisition by reducing scan time while preserving image quality.
• To exploit the correlation between real and imaginary parts of MR images through complex-valued neural networks.
• To improve reconstruction accuracy by enforcing k-space data consistency within a deep residual network architecture.
• To train a deep learning model using large-scale multi-channel MR image pairs without relying on coil sensitivity maps or predefined transforms.
• To achieve state-of-the-art performance in fast MRI reconstruction using an end-to-end trainable framework.

Proposed method

• A deep residual network is employed with complex-valued convolutions to model the joint relationship between real and imaginary components of MR images.
• The network is trained end-to-end using a large number of multi-channel k-space data pairs as supervised labels.
• Data consistency is enforced iteratively between network layers by projecting reconstructed images back into the k-space domain.
• Complex-valued filters are used to maintain phase and magnitude information throughout the network, improving representation fidelity.
• The architecture alternates between complex convolutional layers and data consistency layers to refine reconstructions progressively.
• The method does not require prior knowledge of coil sensitivities or transformation domains such as wavelets or Fourier.

Experimental results

Research questions

• RQ1 Can a deep residual network with complex-valued convolutions effectively reconstruct multi-channel MR images from undersampled k-space data?
• RQ2 How does enforcing k-space data consistency within the network layers affect reconstruction accuracy and image quality?
• RQ3 To what extent can a data-driven approach outperform traditional methods that rely on coil sensitivity maps or transform-domain priors?
• RQ4 Does modeling the correlation between real and imaginary parts of MR images lead to improved reconstruction over real-valued networks?
• RQ5 How does the proposed method compare to state-of-the-art deep learning methods in terms of quantitative metrics and visual quality?

Key findings

• The proposed DeepcomplexMRI method achieves higher reconstruction accuracy than state-of-the-art methods on in vivo MRI datasets.
• The use of complex-valued convolutions significantly improves image quality by preserving phase and magnitude relationships across the network.
• Iterative enforcement of k-space data consistency within the residual network leads to better convergence and reduced artifacts.
• The method outperforms real-valued networks in both quantitative metrics and visual assessment, demonstrating the importance of complex-valued learning.
• The model generalizes well across different anatomical regions and acceleration factors without requiring retraining.
• The approach achieves superior performance without relying on coil sensitivity maps or predefined transform priors, relying solely on large-scale labeled data.