Fast l₁-SPIRiT compressed sensing parallel imaging MRI: scalable parallel implementation and clinically feasible...

Area of Science:

• Medical Imaging
• Computational Imaging
• Magnetic Resonance Imaging (MRI)

Background:

• Iterative reconstruction methods in MRI, such as compressed sensing (CS) and parallel imaging (PI), offer improved image quality but often suffer from high computational costs.
• Long runtimes for these advanced reconstruction techniques present a significant barrier to their widespread clinical adoption.
• Efficient auto-calibration techniques are crucial for optimizing PI reconstruction.

Purpose of the Study:

• To introduce l₁-SPIRiT, a novel algorithm designed for efficient auto calibrating parallel imaging (acPI) and compressed sensing (CS) MRI reconstruction.
• To achieve clinically feasible runtimes for advanced MRI reconstruction without compromising image quality.
• To explore and implement parallelization strategies for l₁-SPIRiT on multi-GPU and multi-core CPU systems.

Main Methods:

• Developed a CS objective function that leverages cross-channel joint sparsity in the wavelet domain.
• Employed iterative soft-thresholding for reconstruction, integrating seamlessly with iterative self-consistent parallel imaging (SPIRiT).
• Implemented and evaluated parallelization strategies for l₁-SPIRiT on multi-GPU and multi-core CPU architectures, focusing on cache usage and parallelization overheads.

Main Results:

• Demonstrated that l₁-SPIRiT enables efficient MRI reconstruction with clinically feasible runtimes.
• Achieved high image quality, validated through clinical experimentation using a 3D Fast Spoiled Gradient Echo (3DFT SPGR) sequence.
• Successfully implemented parallelized versions of l₁-SPIRiT, showing performance dependent on processor architecture, image matrix size, and number of PI channels.

Conclusions:

• l₁-SPIRiT offers a computationally efficient approach to acPI and CS MRI reconstruction, addressing the clinical barrier of long runtimes.
• The algorithm's integration with SPIRiT and its parallelization strategies contribute to faster, high-quality MRI scans.
• The findings highlight the importance of optimizing the trade-off between computational efficiency and parallelization overheads for clinical applicability.