Performance-aware programming for intraoperative intensity-based image registration on graphics processing units.

Area of Science:

• Medical imaging
• Computational anatomy
• Image processing

Background:

• Intensity-based image registration is crucial for aligning medical images but is computationally intensive.
• Slow registration times hinder real-time applications like intra-operative navigation.
• Existing methods struggle to overcome the inherent computational demands of registration algorithms.

Purpose of the Study:

• To accelerate intensity-based image registration by optimizing the diffeomorphic log-demons algorithm on GPUs.
• To address the computational bottlenecks in non-rigid image registration.
• To demonstrate performance-aware programming techniques for GPU acceleration.

Main Methods:

• Implemented the diffeomorphic log-demons algorithm using performance-aware programming techniques on a GPU.
• Systematically exploited GPU features, including computation units, memory, and data bandwidth.
• Focused on resolving identified computation bottlenecks within the registration algorithm.

Main Results:

• Achieved a ~95x speed-up for 1.3-M voxel images and a ~258x speed-up for 37-M voxel images compared to CPU.
• Registered a 1.3-M voxel image in 286 ms and a 37-M voxel image in 8.56 s.
• Demonstrated effective utilization of GPU resources to overcome computational limitations.

Conclusions:

• Pinpointed and resolved computation bottlenecks in diffeomorphic log-demons using GPU programming techniques.
• The developed fast computation methods enhance diffeomorphic log-demons and can benefit other intensity-based registration approaches.
• An open-source implementation is available, facilitating further research and application.