GPU-Accelerated SPECT Image Reconstruction via Compressed PSF
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high computational cost of existing SPECT image reconstruction algorithms, such as OS-EM, limits their clinical applicability due to the time required for producing high-resolution images, despite efforts to improve efficiency through simulators and convergence strategies.
Innovation Solution
The method involves generating a compressed point-spread function matrix and accumulated attenuation factor, which are processed using a graphics processing unit (GPU) for accelerated image projection operations, including re-projection and back-projection, to reduce the computational burden and enhance reconstruction speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative maximum likelihood expectation maximization (ML-EM) reconstruction algorithm with ordered-subsets (OS) strategy is used, then image reconstruction quality is improved, but computational cost and reconstruction time increase significantly
Solution Approach 1:
The patent replaces the traditional CPU-based iterative computation system with a GPU-based parallel processing system. The GPU's architecture with thousands of cores enables simultaneous execution of multiple projection and back-projection operations, substituting the sequential mechanical computation process with a massively parallel computational approach. This substitution maintains the OS-EM algorithm's reconstruction quality while achieving speedups of 10-100x in reconstruction time.
Solution Approach 2:
The patent divides the image reconstruction process into multiple independent segments that can be processed in parallel on the GPU. The volume is segmented into multiple regions, and the projection operations for each region are computed simultaneously by different GPU cores. This segmentation allows the computationally intensive OS-EM algorithm to be executed efficiently without compromising the iterative refinement quality.
2Manufacturing precision
If high-resolution image reconstruction is performed, then image detail and diagnostic value are improved, but computational complexity and processing time increase
Solution Approach 1:
The patent uses GPU parallel processing to replace traditional CPU computation for handling the exponentially increasing computational complexity associated with high-resolution reconstructions. The GPU's ability to perform thousands of simultaneous floating-point operations enables the calculation of projection data for millions of voxels at high resolution without the time penalty that would normally accompany increased detail.
3Device complexity
If conventional CPU-based processing is used for image reconstruction, then system complexity is kept low, but processing speed and clinical applicability are limited
Solution Approach 1:
The patent introduces a GPU as an intermediary processing unit between the data acquisition system and the final image display. This intermediary device handles the computationally intensive projection and back-projection operations, allowing the main CPU system to remain relatively simple while offloading the heavy computational burden to the specialized GPU hardware. The GPU acts as a mediator that bridges the gap between data acquisition and rapid image reconstruction.
Data Source
AI summary
A method for reconstructing an image from emission data includes generating a compressed point-spread function matrix, generating an accumulated attenuation factor; and performing at least one image projection operation on an image matrix of the emission data using the compressed point-spread function matrix and the accumulated attenuation factor. The image projection operation can include rotating an image matrix and an exponential attenuation map to align with a selected viewing angle. An accumulated attenuation image is then generated from the rotated image matrix and rotated exponential attenuation map and a projection image is generated for each voxel by multiplying the accumulated attenuation image and point spread function matrix for each voxel. The rotating and multiplying operations can be performed on a graphics processing unit, which may be found in a commercially available video processing card, which are specifically designed to efficiently perform such operations.