Distributed MLEM: an iterative tomographic image reconstruction algorithm for distributed memory architectures.

Area of Science:

• Medical Imaging
• High-Performance Computing
• Computational Science

Background:

• Positron Emission Tomography (PET) image reconstruction speed is crucial for clinical applications.
• Current multi-graphics processing unit (GPU) approaches face scalability limitations due to inter-GPU data transfer.
• Existing methods do not efficiently leverage large-scale GPU clusters for PET image reconstruction.

Purpose of the Study:

• To develop a novel reformulation of the Maximum Likelihood Expectation Maximization (MLEM) algorithm for enhanced scalability in PET image reconstruction.
• To reduce inter-node communication overhead in multi-GPU cluster environments.
• To achieve near-linear performance scaling with an increasing number of GPUs.

Main Methods:

• Reformulation of the MLEM algorithm to minimize inter-GPU data transfer requirements.
• Implementation of the reformulated algorithm on a multi-GPU cluster.
• Mathematical validation ensuring convergence to the same solution as the standard MLEM algorithm.

Main Results:

• The proposed algorithm demonstrates improved scalability on multi-GPU clusters compared to traditional parallel MLEM.
• Significant reduction in inter-node communication frequency was achieved.
• Experimental results validate the effectiveness and efficiency of the novel approach for PET image reconstruction.

Conclusions:

• The reformulated MLEM algorithm offers a viable solution for accelerating PET image reconstruction on large-scale GPU clusters.
• This approach overcomes the communication bottleneck, enabling more efficient parallel processing.
• The findings pave the way for faster and more accessible advanced medical imaging.