Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cine cardiac MRI reconstruction using a convolutional recurrent network with refinement

Yuyang Xue, Yuning Du, Gianluca Carloni, Eva Pachetti, Connor Jordan, Sotirios A. Tsaftaris

TL;DR

This work tackles cine cardiac MRI reconstruction under k-space undersampling by exploiting temporal correlations with a convolutional recurrent neural network (CRNN) and augmenting it with a lightweight refinement module. A novel loss strategy, including a high-pass split of the ℓ1 loss and a polar perp-loss, is explored to emphasize high-frequency details and phase information. The proposed CRNN–refinement architecture outperforms a 2D U-Net baseline and yields about 4% SSIM and 3.9% NMSE improvements over the plain CRNN, particularly for short-axis views and higher acceleration factors, while memory constraints shape model choices. The study highlights the potential of combining temporal models with efficient post-processing refinements for practical, fast cardiac MRI reconstruction, and points to avenues for further gains via per-view training and advanced loss tuning.

Abstract

Cine Magnetic Resonance Imaging (MRI) allows for understanding of the heart's function and condition in a non-invasive manner. Undersampling of the $k$-space is employed to reduce the scan duration, thus increasing patient comfort and reducing the risk of motion artefacts, at the cost of reduced image quality. In this challenge paper, we investigate the use of a convolutional recurrent neural network (CRNN) architecture to exploit temporal correlations in supervised cine cardiac MRI reconstruction. This is combined with a single-image super-resolution refinement module to improve single coil reconstruction by 4.4\% in structural similarity and 3.9\% in normalised mean square error compared to a plain CRNN implementation. We deploy a high-pass filter to our $\ell_1$ loss to allow greater emphasis on high-frequency details which are missing in the original data. The proposed model demonstrates considerable enhancements compared to the baseline case and holds promising potential for further improving cardiac MRI reconstruction.

Cine cardiac MRI reconstruction using a convolutional recurrent network with refinement

TL;DR

This work tackles cine cardiac MRI reconstruction under k-space undersampling by exploiting temporal correlations with a convolutional recurrent neural network (CRNN) and augmenting it with a lightweight refinement module. A novel loss strategy, including a high-pass split of the ℓ1 loss and a polar perp-loss, is explored to emphasize high-frequency details and phase information. The proposed CRNN–refinement architecture outperforms a 2D U-Net baseline and yields about 4% SSIM and 3.9% NMSE improvements over the plain CRNN, particularly for short-axis views and higher acceleration factors, while memory constraints shape model choices. The study highlights the potential of combining temporal models with efficient post-processing refinements for practical, fast cardiac MRI reconstruction, and points to avenues for further gains via per-view training and advanced loss tuning.

Abstract

Cine Magnetic Resonance Imaging (MRI) allows for understanding of the heart's function and condition in a non-invasive manner. Undersampling of the -space is employed to reduce the scan duration, thus increasing patient comfort and reducing the risk of motion artefacts, at the cost of reduced image quality. In this challenge paper, we investigate the use of a convolutional recurrent neural network (CRNN) architecture to exploit temporal correlations in supervised cine cardiac MRI reconstruction. This is combined with a single-image super-resolution refinement module to improve single coil reconstruction by 4.4\% in structural similarity and 3.9\% in normalised mean square error compared to a plain CRNN implementation. We deploy a high-pass filter to our loss to allow greater emphasis on high-frequency details which are missing in the original data. The proposed model demonstrates considerable enhancements compared to the baseline case and holds promising potential for further improving cardiac MRI reconstruction.
Paper Structure (19 sections, 2 equations, 4 figures, 3 tables)

This paper contains 19 sections, 2 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Problem formulation and dataset
  3. Data
  4. Data pre-processing
  5. Methodology
  6. Model exploration
  7. Model architecture
  8. Loss function
  9. Implementation details
  10. Results
  11. Model choice and weight sharing
  12. Loss function investigation
  13. Introduction of the refinement module
  14. Validation results
  15. Discussion
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Final model architecture: BCRNN, CRNN, and CNN units with a data consistency (DC) step from Qin2019 for primary reconstruction. "t" and "i" denote time and iterations, respectively. The low-cost refinement module, inspired by Bilecen2023, includes downsampling (DS), CNN, and upsampling (US) units.
  • Figure 2: Log loss during exploratory training of modified CineNet and CRNN (with and without weight-sharing between kernels). Note that the implementation is not identical to the original works.
  • Figure 3: Reconstruction (top) and associated error maps (bottom) for the initial network investigation. (a) 8 $\times$ undersampled LAX input (b) fully sampled ground truth (c,d) CineNet model (6 cascades) (e,f) CRNN model (weight-sharing between cascades) (g,h) CRNN model (no weight-sharing).
  • Figure 4: Reconstruction (top) and associated error maps (middle) for the U-Net baseline and CRNN models. Finer details (bottom) are not resolved by the U-Net, which are partially captured by the plain CRNN model. The refinement module subsequently deblurs the image and provides better resolution at boundaries. (a) 10 $\times$ undersampled SAX input (b, c) fully sampled ground truth (d, e, f) U-Net (g, h, i) 6 cascades with combined refinement (j, k, l) 7 cascades, no refinement.