An Effective CUDA Parallelization of Projection in Iterative Tomography Reconstruction

Area of Science:

• Medical Imaging
• Computer Science
• Computational Science

Background:

• Computed Tomography (CT) reconstruction involves computationally intensive projection and back-projection steps.
• Accelerating CT reconstruction is crucial for clinical applications.
• Conventional parallelization in projection faces challenges due to racing conditions and thread unsynchronization.

Purpose of the Study:

• To propose a novel Fixed Sampling Number Projection (FSNP) strategy for efficient Graphical Processing Unit (GPU)-based CT reconstruction.
• To enhance synchronization in ray-driven projection and accelerate interpolation processes.

Main Methods:

• Implementation of a Fixed Sampling Number Projection (FSNP) strategy for ray-driven projection on GPUs.
• Utilization of texture fetching to accelerate interpolation in both projection and back-projection.
• Validation using simulated and real cone-beam CT data.

Main Results:

• The proposed FSNP method with texture fetching achieves 10-16 times speedup compared to conventional global memory approaches.
• Improved synchronization in projection operations was observed.
• Faster interpolation significantly contributes to overall acceleration.

Conclusions:

• The FSNP strategy effectively addresses synchronization issues in GPU-based CT projection.
• The combined FSNP and texture fetching approach offers a significant acceleration for CT reconstruction algorithms.
• This method enables more efficient iterative CT reconstruction processes.