Compressed SVD-based L + S model to reconstruct undersampled dynamic MRI data using parallel architecture.

Area of Science:

• Medical Imaging
• Computational Imaging
• Cardiovascular Imaging

Background:

• Magnetic Resonance Imaging (MRI) offers high resolution and functional insights, particularly for cardiac applications.
• Long MRI scan times pose challenges for patient cooperation and image quality due to motion artifacts.
• Current reconstruction methods for undersampled MRI are computationally intensive, hindering real-time clinical use.

Purpose of the Study:

• To reduce Magnetic Resonance Imaging (MRI) scan time by addressing undersampling artifacts.
• To develop computationally efficient image reconstruction algorithms for dynamic MRI (dMRI).
• To enable real-time clinical applications, such as cardiac MRI, through reduced reconstruction times.

Main Methods:

• Employed a low-rank plus sparse (L+S) matrix decomposition model for reconstructing undersampled dMRI data.
• Integrated Compressed Singular Value Decomposition (cSVD) into the L+S model to decrease reconstruction duration.
• Developed a customized GPU-based parallel architecture to exploit inherent parallelism in the cSVD-enhanced L+S model.

Main Results:

• The proposed GPU-based parallel architecture achieved significant speed-up factors for cardiac perfusion MRI reconstruction.
• Speed-up factors reached up to 19.15x (with memory latency) and 70.55x (without memory latency) compared to conventional CPU reconstruction.
• Reconstruction using the novel method maintained image quality without compromise, even with high acceleration factors (2, 6, 8).

Conclusions:

• The developed parallel reconstruction method substantially reduces MRI reconstruction time.
• This approach is highly suitable for real-time clinical applications, including cardiac MRI.
• The findings pave the way for faster and more efficient diagnostic imaging.