Efficient rendering of digitally reconstructed radiographs on heterogeneous computing architectures using central...

Area of Science:

• Medical Physics
• Medical Imaging
• Computational Science

Background:

• Digitally reconstructed radiographs (DRRs) are crucial for patient alignment in image-guided radiation therapy, typically using direct volume rendering (DVR) with O(N3) complexity.
• High-quality DRR generation faces performance limitations with traditional DVR methods, hindering real-time clinical applications.
• Existing k-space methods for DRR generation have vendor or software limitations.

Purpose of the Study:

• To develop and present a high-performance, vendor-neutral DRR generation pipeline utilizing k-space and the central slice theorem.
• To enable efficient DRR generation across heterogeneous computing architectures via OpenCL.
• To overcome the performance bottlenecks associated with traditional DVR methods for DRR creation.

Main Methods:

• Implementation of a k-space-based DRR generation pipeline using OpenCL for broad hardware compatibility.
• Leveraging the central slice theorem for an improved computational complexity of O(N2logN).
• Testing the pipeline on various computing architectures, including CPUs and GPUs.

Main Results:

• The k-space DRR pipeline achieves significantly faster rendering times compared to DVR methods.
• DRR generation for a 5123 CT volume was completed in 6 ms (CPU), 2.7 ms (mid-range GPU), and 0.68 ms (high-end GPU).
• The OpenCL implementation demonstrates high performance and flexibility across different hardware.

Conclusions:

• The presented k-space DRR pipeline offers a substantial performance improvement for generating digitally reconstructed radiographs.
• This approach enhances the feasibility of rapid DRR generation for clinical applications, particularly in image-guided radiation therapy.
• The OpenCL-based, heterogeneous implementation provides a versatile and efficient solution for modern radiation therapy workflows.