A Fully GPU-Based Ray-Driven Backprojector via a Ray-Culling Scheme with Voxel-Level Parallelization for Cone-Beam CT...

Area of Science:

• Medical Imaging
• Computer Vision
• Scientific Computing

Background:

• Ray-driven backprojection is computationally intensive due to ray-voxel intersection tests.
• Existing methods face challenges in balancing computational load and reconstruction speed.

Purpose of the Study:

• To develop a highly efficient GPU-based ray-driven backprojector.
• To reduce computational burden in tomographic reconstruction using a novel ray-culling scheme.

Main Methods:

• Implemented a fully GPU-based ray-driven backprojector utilizing a ray-culling technique.
• Employed axis-aligned bounding boxes (AABBs) to exclude irrelevant ray-voxel intersection tests.
• Optimized ray-culling using GPU shared memory for enhanced performance.

Main Results:

• Achieved a 26x speedup compared to standard CPU-based backprojectors and 7.5x faster than existing GPU methods.
• Reconstructed a 280x280x176 volume in 77 seconds from 680 projections.
• Demonstrated superior image quality with higher contrast-to-noise ratio and universal quality index compared to the Feldkamp-Davis-Kress algorithm.

Conclusions:

• The proposed GPU-based ray-driven backprojector with ray-culling offers significant speed improvements for tomographic reconstruction.
• The method achieves high-quality reconstructions, outperforming traditional algorithms.
• This approach is well-suited for GPU implementation, enabling efficient voxel-level parallelization.