Ferumoxytol-Enhanced Cardiac Cine MRI Reconstruction Using a Variable-Splitting Spatiotemporal Network
Chang Gao1,2, Zhengyang Ming1,2, Kim-Lien Nguyen1,2,3
1Department of Physics and Biology in Medicine, University of California Los Angeles, Los Angeles, California, USA.
Journal of Magnetic Resonance Imaging : JMRI
|March 4, 2024
Summary
A new deep learning network, VSNet, was developed for cardiac MRI reconstruction. It shows superior image quality and accurate functional measurements for Ferumoxytol-enhanced gradient echo cine imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Radiology
- Cardiac MRI
Background:
- Balanced steady-state free precession (bSSFP) is standard for cardiac cine MRI but suffers from artifacts.
- Ferumoxytol-enhanced (FE) gradient echo (GRE) offers an alternative, necessitating efficient reconstruction methods.
- Leveraging existing bSSFP data can improve FE GRE cine imaging through advanced network development.
Purpose of the Study:
- To develop a variable-splitting spatiotemporal network (VSNet) for cardiac MRI image reconstruction.
- To train VSNet on bSSFP cine images and ensure its applicability to FE GRE cine images.
- To create a computationally efficient network for improved cardiac MRI.
Main Methods:
- Retrospective and prospective study design.
- Network training on 41 patients' bSSFP cine images; testing on 31 patients and 5 healthy subjects' FE GRE cine images.
- Comparison of VSNet against other reconstruction methods (total variation loss, compressed sensing, low rank) at 14x acceleration, using GRAPPA images as reference.
Main Results:
- VSNet significantly outperformed other methods in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), qualitative ranking, and latent scores.
- Quantitative analysis showed VSNet achieved comparable left ventricular (LV) and right ventricular (RV) end-systolic volume (ESV) and ejection fraction (EF) to the reference.
- VSNet demonstrated a statistically significant but small difference in end-diastolic volume (EDV) compared to the reference.
Conclusions:
- VSNet achieved superior image quality and more accurate functional measurements for FE GRE cine images compared to other tested 14x accelerated reconstruction methods.
- The developed VSNet offers a promising solution for artifact reduction and improved diagnostic accuracy in cardiac MRI.
- This study highlights the potential of deep learning for enhancing MRI techniques.
