CUDA accelerated uniform re-sampling for non-Cartesian MR reconstruction.

Area of Science:

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

Background:

• Non-Cartesian MRI data, such as that acquired with PROPELLER (Periodically Rotated Parallel Lines with Enhanced Reconstruction), requires specialized gridding techniques for accurate image reconstruction.
• Traditional gridding methods can be computationally intensive, limiting the speed of MRI data processing.

Purpose of the Study:

• To introduce a novel grid-driven gridding (GDG) method for uniform re-sampling of non-Cartesian MRI data.
• To develop a CUDA-accelerated version of the GDG method to enhance computational performance.
• To evaluate the performance and speed improvements of the accelerated GDG method.

Main Methods:

• A grid-driven gridding (GDG) approach was developed, computing a trajectory window for each Cartesian grid and using weighted averages of raw data for intensity.
• The GDG method was optimized for parallel processing by leveraging the single instruction multiple data (SIMD) architecture.
• A CUDA-based acceleration strategy was implemented, exploring four thread-block configurations to balance GPU resources and maximize performance.

Main Results:

• The proposed GDG method successfully reconstructed PROPELLER MRI data at two different resolutions.
• While the initial GDG method was more time-consuming than traditional density compensation (DDG), the CUDA-accelerated GDG demonstrated substantial speed improvements.
• Experimental results showed the CUDA-accelerated GDG to be nearly 10 times faster than traditional DDG.

Conclusions:

• The CUDA-accelerated grid-driven gridding method provides a significant performance enhancement for reconstructing non-Cartesian MRI data from PROPELLER acquisitions.
• This accelerated approach offers a more efficient alternative to traditional gridding techniques, potentially reducing MRI scan and reconstruction times.