On developing B-spline registration algorithms for multi-core processors.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Computer Vision

Background:

• Spline-based deformable registration is crucial in medical imaging for its flexibility and robustness.
• Existing methods face challenges due to high computational demands, limiting their practical application.
• Accelerating these registration techniques is essential for clinical workflows and research.

Purpose of the Study:

• To develop and validate accelerated B-spline-based deformable registration methods.
• To reduce the significant computing time associated with traditional spline-based registration.
• To enable efficient implementation on modern multi-core processors, including graphics processing units (GPUs).

Main Methods:

• Introduction of a novel grid-alignment scheme and optimized data structures to decrease algorithmic complexity.
• Development of data-parallel designs for B-spline registration within a stream-processing model.
• Implementation of analytic gradient computations in a data-parallel manner for efficient processing on GPUs and multi-core CPUs.

Main Results:

• The GPU implementation achieved a 15x speedup compared to the single-threaded CPU version for large images.
• The multi-core CPU implementation demonstrated an 8x speedup over the single-threaded CPU version.
• Both CPU and GPU versions yielded near-identical registration quality, measured by RMS differences in vector fields.

Conclusions:

• The proposed grid-alignment scheme and parallel processing designs effectively accelerate B-spline registration.
• The developed GPU and multi-core CPU algorithms offer significant performance improvements without compromising registration accuracy.
• This work facilitates the wider adoption of advanced deformable registration techniques in medical imaging research and practice.