Iterative Reconstruction of Micro Computed Tomography Scans Using Multiple Heterogeneous GPUs.

Area of Science:

• Medical Imaging
• Computer Science
• High-Performance Computing

Background:

• Iterative reconstruction of ultrahigh-resolution micro computed tomography (CT) scans requires significant computational resources.
• Graphics processing units (GPUs) offer massive parallelism suitable for these demanding calculations.
• Current methods using ordered subsets expectation maximization (OSEM) can be limited by single-GPU memory constraints and data transfer latency.

Purpose of the Study:

• To propose a novel finite state automaton (FSA) method for efficient iterative reconstruction on heterogeneous multi-GPU platforms.
• To address the memory-bound limitations of single GPUs in ultrahigh-resolution CT image reconstruction.
• To enhance the speed and efficiency of CT image reconstruction using parallel processing.

Main Methods:

• Developed a finite state automaton (FSA) method for flow control in parallel threading across heterogeneous GPUs.
• Implemented parallelized matrix calculations derived from a ray tracing system of ordered subsets.
• Utilized an event-triggered FSA approach for managing tasks on multiple GPUs simultaneously.
• Compared reconstruction performance across single-GPU, heterogeneous multi-GPU with FIFO queues, and heterogeneous multi-GPU with job queues.

Main Results:

• The proposed heterogeneous multi-GPU approach with job queues achieved reconstruction speeds up to five times faster than a single-GPU environment.
• This method was nine times faster than heterogeneous multi-GPUs using FIFO queues for device scheduling control.
• The FSA method effectively minimized task launch latency and reduced data transfer between system and GPU memory.
• The memory-bound issue of single GPUs was successfully overcome.

Conclusions:

• The event-triggered finite state automaton (FSA) method enables efficient, high-throughput iterative reconstruction for ultrahigh-resolution CT scans on multiple heterogeneous GPUs.
• This parallel processing strategy significantly accelerates reconstruction times and overcomes memory limitations inherent in single-GPU systems.
• The proposed method demonstrates successful simultaneous execution of reconstruction routines across multiple GPUs.