High performance computing for deformable image registration: towards a new paradigm in adaptive radiotherapy.
Sanjiv S Samant1, Junyi Xia, Pinar Muyan-Ozcelik
1Department of Nuclear and Radiological Engineering, University of Florida, Gainesville, Florida 32611-8300, USA. samant@ufl.edu
Medical Physics
|September 10, 2008
Summary
Graphics processing unit (GPU) acceleration significantly speeds up deformable image registration (DIR), making real-time applications in radiation therapy feasible. This GPU implementation using CUDA is 34-61 times faster than CPU methods.
Area of Science:
- Medical Imaging
- Computational Science
Background:
- 4D imaging is crucial for radiotherapy planning and adaptive radiotherapy (ART).
- Deformable image registration (DIR) is vital for medical imaging applications like motion correction but is limited by long computation times.
- Graphics processing units (GPUs) offer parallel processing capabilities suitable for accelerating DIR.
Purpose of the Study:
- To implement and evaluate the Demons algorithm for DIR on a GPU using CUDA.
- To compare the performance of the GPU implementation against single-threaded and multi-threaded CPU implementations.
- To assess the feasibility of real-time DIR for clinical applications.
Main Methods:
- Implemented the Demons algorithm for DIR on an NVIDIA 8800 GTX GPU using the CUDA programming language.
- Compared GPU performance with C-based single-threaded and multi-threaded CPU implementations on an Intel dual-core processor.
- Evaluated performance using volumetric clinical lung images from 4DCT scans.
Main Results:
- The GPU implementation achieved computation times of 1.8–13.5 seconds for 100 iterations on images ranging from 2.0 x 10^6 to 14.2 x 10^6 pixels.
- GPU registration was 55–61 times faster than single-threaded CPU and 34–39 times faster than multi-threaded CPU.
- Achieved a Time Per Megapixel Per Iteration (TPMI) of 0.00916 spmi on GPU, significantly lower than CPU implementations (0.527 spmi and 0.335 spmi).
Conclusions:
- GPU-based DIR using CUDA offers substantial speed improvements over traditional CPU methods.
- This acceleration enables real-time DIR, opening new clinical applications in medical imaging and radiotherapy.
- Optimized memory handling under CUDA contributes to the high computational efficiency observed.
