GPU-based streaming architectures for fast cone-beam CT image reconstruction and demons deformable registration

Area of Science:

• Medical Imaging
• Computational Science
• Computer Engineering

Background:

• Cone-beam CT (CBCT) reconstruction and 3D deformable image registration are computationally intensive processes.
• Accelerating these processes is crucial for clinical applications and research.

Purpose of the Study:

• To significantly accelerate CBCT reconstruction and 3D deformable image registration.
• To implement and evaluate data-parallel streaming versions of the Feldkamp, Davis and Kress (FDK) and demons algorithms on a commodity graphics processing unit (GPU).

Main Methods:

• Developed data-parallel streaming implementations of the FDK reconstruction and demons deformable registration algorithms.
• Utilized the Brook programming environment for implementation on an NVidia 8800 GPU.
• Compared performance and accuracy against an optimized CPU-based reference implementation.

Main Results:

• Achieved substantial speedups: up to 80x for FDK and 70x for demons compared to a 2.8 GHz Intel processor.
• Demonstrated excellent accuracy for GPU-based implementations, with RMS differences < 0.1 Hounsfield unit for reconstruction and < 0.1 mm for registration.

Conclusions:

• The stream-processing model on GPUs offers significant acceleration for CBCT reconstruction and 3D deformable image registration.
• GPU-based implementations provide a viable and accurate solution for accelerating these critical medical imaging tasks.