GPU-based 3D cone-beam CT image reconstruction for large data volume.

Area of Science:

• Medical Imaging
• Computational Science

Background:

• 3D cone-beam CT image reconstruction speed is a significant bottleneck for clinical use.
• Graphics Processing Units (GPUs) offer potential for accelerating reconstruction, but large datasets pose memory challenges.

Purpose of the Study:

• To develop and evaluate a new GPU-based partitioning scheme for accelerating 3D cone-beam CT image reconstruction of large data volumes.
• To address the limitations of existing methods like Octree partitioning for handling data exceeding GPU memory.

Main Methods:

• A novel partitioning scheme utilizing geometric symmetries in circular cone-beam projection data is proposed.
• Projection data and image volumes are divided into subsets, packed into GPU RGBA channels for chunk-wise reconstruction.
• Reconstruction is performed on subsets, followed by combining individual results.

Main Results:

• The GPU implementation achieves reconstruction speeds 110-120 times faster than traditional CPU implementations for circular cone-beam scans.
• The method successfully reconstructs 3D images from both simulated and real micro-CT data.
• Original image precision is maintained with the proposed GPU acceleration technique.

Conclusions:

• The developed GPU-based method offers a significant speed-up for 3D cone-beam CT image reconstruction, making it more viable for clinical applications.
• The novel partitioning strategy effectively handles large data volumes that exceed GPU memory limitations.
• This approach demonstrates the potential of GPU computing for high-speed, high-precision medical image reconstruction.