Accelerated Computing in Magnetic Resonance Imaging: Real-Time Imaging Using Nonlinear Inverse Reconstruction.

Area of Science:

• Medical Imaging
• Computational Science
• Biophysics

Background:

• Dynamic MRI requires efficient image reconstruction for high temporal resolution.
• Current reconstruction methods can be computationally intensive, limiting frame rates.
• Accelerating reconstruction is crucial for advanced dynamic MRI applications.

Purpose of the Study:

• To develop generalizable GPU-based optimization strategies for MRI image reconstruction.
• To implement and evaluate an accelerated nonlinear inversion (NLINV) algorithm for high-frame-rate dynamic MRI.
• To demonstrate the effectiveness of parallel decomposition techniques for dynamic MRI.

Main Methods:

• Optimized and ported the NLINV algorithm to a multi-GPU server.
• Implemented data domain decomposition along the channel dimension for multi-GPU processing.
• Utilized temporal domain decomposition by relaxing regularization for parallel frame processing.
• Developed an autotuning method to combine decomposition strategies for optimal performance.

Main Results:

• Successfully deployed the NLINV algorithm on a multi-GPU system for online image reconstruction.
• Achieved high frame rates up to 30 frames per second with low latency.
• Demonstrated real-time image reconstruction capabilities for dynamic MRI.

Conclusions:

• Introduced novel parallel decomposition methods applicable to iterative dynamic MRI algorithms.
• Successfully parallelized the NLINV algorithm on multiple GPUs for high-frame-rate reconstruction.
• Enabled online image reconstruction, significantly advancing dynamic MRI capabilities.