Fast, accurate and shift-varying line projections for iterative reconstruction using the GPU.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Computational Science

Background:

• List-mode processing is efficient for sparse projections in emission tomography image reconstruction.
• Iterative algorithms, like expectation-maximization, are computationally intensive due to extensive line-projection operations.
• Current methods face challenges with the high computational demands of processing individual recorded events.

Purpose of the Study:

• To investigate the use of graphics processing units (GPUs) for accelerating list-mode image reconstruction in positron emission tomography (PET).
• To implement a fully-3D ordered-subsets expectation-maximization algorithm optimized for GPU acceleration.
• To incorporate spatially-varying resolution kernels into the reconstruction to model physical processes accurately.

Main Methods:

• Developed a GPU-accelerated reconstruction approach for fully-3D list-mode ordered-subsets expectation-maximization in PET.
• Designed algorithms to efficiently handle line-projection operations on the GPU, accommodating arbitrary line endpoints.
• Integrated spatially-varying and shift-varying resolution kernels to model complex physical processes.

Main Results:

• The GPU implementation achieved over 50 times speedup compared to a CPU implementation.
• Image quality and accuracy were virtually identical between the GPU and CPU methods.
• The approach demonstrated suitability for sparse projection data scenarios, including high-resolution, dynamic, and time-of-flight PET.

Conclusions:

• GPU acceleration offers a significant computational advantage for list-mode PET image reconstruction.
• The developed method effectively models physical processes using resolution kernels while maintaining high performance.
• This GPU-based strategy is particularly beneficial for advanced PET applications requiring efficient processing of sparse data.