Parallel MRI Reconstruction Memory Efficiency via Asynchronous GPU Streaming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Parallel MRI reconstruction techniques face memory efficiency challenges due to large data sets, limiting their processing on parallel computing platforms despite the potential for faster reconstruction times.
Innovation Solution
The method involves asynchronously streaming coil sensitivity data to the graphical processing device, applying a primal-dual algorithm like Chambolle-Pock for regularization, and using a pre-computed Lipschitz value to optimize memory usage, allowing for overlapping data transfer and processing, and employing a redundant Haar wavelet transform for sparsifying transforms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Parallel MRI reconstruction is performed with high resolution 3D data sets, then image quality and signal-to-noise ratio are improved, but memory requirements exceed the capacity of parallel computing hardware
Solution Approach 1:
The patent segments the large coil sensitivity data into smaller chunks that can be streamed asynchronously to the GPU. Instead of loading all coil data at once, the system divides the data into manageable portions and processes them in sequential batches, allowing high-resolution reconstruction without exceeding GPU memory capacity.
Solution Approach 2:
The patent performs preliminary computations on the host CPU before transferring data to the GPU. Coil sensitivity maps and other preprocessing operations are completed in advance, so that only essential data needs to be transferred to the parallel computing device, reducing the memory burden on the GPU while maintaining reconstruction quality.
2Productivity
If all coil data sets are transferred to the graphical processing device simultaneously, then processing speed is improved, but memory consumption exceeds device capacity
Solution Approach 1:
The patent implements asynchronous streaming that allows continuous processing without idle waiting periods. While the GPU processes one coil data set, the host simultaneously prepares the next data set for transfer. This overlapping of operations maintains high processing throughput while keeping memory usage at manageable levels through sequential batch processing.
Solution Approach 2:
The system dynamically manages data transfer and processing operations, adapting the flow of data between host and device based on current memory availability and processing status. The asynchronous streaming mechanism allows flexible pacing of data transfers, enabling the system to maintain optimal performance while respecting memory constraints.
Data Source
AI summary
A computer-implemented method for reconstructing magnetic resonance images using a parallel computing platform comprising a host unit and a graphical processing device includes receiving a plurality of coil data sets from a magnetic resonance imaging system, each respective coil data set comprising scanner data and a coil sensitivity map associated with a distinct coil included in the magnetic resonance imaging system. An iterative compressed-sensing reconstruction process is applied to reconstruct an image based on the plurality of coil data sets. Each iteration of the iterative compressed-sensing reconstruction process comprises: individually transferring the plurality of coil data sets from the host unit to the graphical processing device using a plurality of asynchronous data streams, overlapping with transfer of the plurality of coil data sets, optimizing a plurality of data fidelity values on the graphical processing device, each respective data fidelity value corresponding to a distinct coil data set, and computing an estimated image on the graphical processing device based on the plurality of data fidelity values.