Efficient acceleration of mutual information computation for nonrigid registration using CUDA

Area of Science:

• Medical Imaging
• Computer Vision
• High-Performance Computing

Background:

• Nonrigid registration aligns medical images with complex anatomical variations.
• Multimodal image registration is crucial for integrating information from different imaging modalities.
• Existing methods often face computational challenges, limiting their clinical applicability.

Purpose of the Study:

• To propose an efficient graphics processing unit (GPU)-accelerated method for nonrigid registration of multimodal images.
• To optimize normalized mutual information (NMI) computation and hierarchical B-spline deformation using on-chip memory.
• To achieve significant speedups in image registration tasks.

Main Methods:

• Implementation of a Compute Unified Device Architecture (CUDA) program for nonrigid registration.
• Efficient parallelization strategies, including hierarchical data organization, data reuse, and multiresolution representation.
• Utilization of on-chip GPU memory for NMI computation and B-spline deformation.

Main Results:

• A 12-fold increase in speed compared to an off-chip memory version.
• Enhanced parallel execution efficiency from 4% to 46%.
• Approximately 14 times faster than a fully optimized four-core CPU-based implementation.

Conclusions:

• The proposed GPU-accelerated method significantly enhances the efficiency of nonrigid multimodal image registration.
• Exploitation of on-chip memory is key to achieving substantial speedups.
• The method enables rapid image alignment within seconds, facilitating clinical applications.