Multi-GPU Acceleration of Branchless Distance Driven Projection and Backprojection for Clinical Helical CT.

Area of Science:

• Medical Imaging
• Computational Science
• Computer Engineering

Background:

• Model-based image reconstruction (MBIR) offers high-quality imaging with reduced X-ray dose.
• Distance-driven (DD) projection and backprojection are key MBIR components but computationally intensive.
• Current MBIR methods face challenges in clinical adoption due to high computational costs.

Purpose of the Study:

• To develop and implement parallelized distance-driven (DD) projectors for multi-GPU systems.
• To accelerate iterative model-based image reconstruction (MBIR) for clinical applications.
• To reduce the computational burden of DD projection and backprojection.

Main Methods:

• Implemented novel data and voxel partitioning schemes for concurrent execution on multiple GPUs.
• Optimized overlap calculations by projecting detector boundaries directly onto image voxel boundaries.
• Utilized a pre-accumulation technique for image intensities in 2D slabs to enhance parallel processing.
• Employed a parallel multi-GPU alternating minimization (AM) algorithm with penalized likelihood update.

Main Results:

• Achieved significant speedups in combined projection and backprojection using CUDA programming.
• Demonstrated an average speedup of 24x with a single GPU and 74x with three GPUs.
• Successfully reconstructed images from Siemens Sensation 16 patient scan data.

Conclusions:

• The proposed parallelization strategies effectively reduce the computational cost of DD projectors.
• This work paves the way for faster and more efficient MBIR, enabling lower patient X-ray doses.
• The multi-GPU implementation significantly enhances the clinical feasibility of advanced image reconstruction techniques.