Parallelized formulation of the maximum likelihood-expectation maximization algorithm for fine-grain message-passing...

Area of Science:

• Computer Science
• Medical Imaging
• Parallel Processing

Background:

• Massively parallel processing systems enable new computational models.
• Maximum Likelihood-Expectation Maximization (ML-EM) algorithms are crucial for image reconstruction.
• High-communications, low-memory (HCLM) architectures present unique computational challenges.

Purpose of the Study:

• To develop and evaluate an ML-EM algorithm optimized for HCLM systems.
• To assess the performance benefits of parallel processing for ML-EM formulations.
• To determine the feasibility of achieving clinically relevant ML-EM reconstruction times.

Main Methods:

• Developed a novel ML-EM algorithm tailored for HCLM execution models.
• Utilized discrete-event simulation on the Pica multiprocessor system for performance evaluation.
• Compared the algorithm's error and speed against sequential ray-driven ML-EM formulations.

Main Results:

• The HCLM ML-EM algorithm achieved normalized least-square errors comparable to or better than sequential methods.
• Demonstrated an effective speedup exceeding two orders of magnitude compared to sequential ML-EM on high-end workstations.
• Simulations were performed on the Georgia Institute of Technology's Pica multiprocessor system.

Conclusions:

• The developed HCLM ML-EM algorithm shows significant performance advantages on parallel systems.
• This approach has the potential to deliver ML-EM reconstructions within practical clinical time-frames.
• Architectural and technological advances in parallel processing open new avenues for medical imaging algorithms.