Forward-Projection Architecture for Fast Iterative Image Reconstruction in X-ray CT

Area of Science:

• Medical Imaging
• Computer Engineering
• Computational Science

Background:

• Iterative reconstruction in X-ray computed tomography (CT) enhances image quality but is computationally intensive.
• The lengthy computation time of 3D forward- and back-projection hinders widespread clinical adoption of iterative CT methods.

Purpose of the Study:

• To develop a highly parallel hardware architecture for accelerating the forward-projection step in CT image reconstruction.
• To reduce the computational bottleneck of iterative CT, making it more suitable for routine clinical use.

Main Methods:

• Analyzed CT geometry to identify parallelism for a high-performance hardware architecture.
• Implemented a water-filling buffer and out-of-order sectored processing to optimize efficiency and minimize memory access.
• Performed floating-point to fixed-point conversion based on simulations to reduce implementation costs.

Main Results:

• Prototyped a 5-stage, 55-way parallel forward-projector on an FPGA achieving 925.8 million voxel projections/s.
• Demonstrated a 4.6x throughput increase compared to an optimized multi-threaded CPU implementation.
• Achieved comparable image quality with reduced implementation costs through fixed-point conversion.

Conclusions:

• The proposed algorithm and hardware architecture offer significant acceleration for CT forward-projection.
• Similar architectures are applicable to back-projection and can be adapted for GPUs and DSPs.
• This advancement paves the way for faster, routine clinical use of iterative CT image reconstruction.