Efficient Probabilistic and Geometric Anatomical Mapping Using Particle Mesh Approximation on GPUs.

Area of Science:

• Medical image analysis
• Computational anatomy
• Biomedical engineering

Background:

• Deformable image registration faces challenges with contrast variations and large anatomical changes.
• Intensity-based methods struggle with nonlinear shape variations and high computational costs.
• Existing techniques are often computationally infeasible for practical, time-sensitive applications.

Purpose of the Study:

• To develop a robust and computationally efficient deformable image registration framework.
• To address limitations of intensity-dependent methods in handling significant anatomical deformations.
• To enable practical application of advanced registration in time-critical scenarios.

Main Methods:

• A novel registration method mapping anatomies via multicompartment models (class posterior images and geometries).
• Implementation using particle mesh approximation on Graphics Processing Units (GPUs) for computational efficiency.
• Quantitative validation on neonatal to 2-year-old infant MRI datasets.

Main Results:

• The proposed method achieves superior registration accuracy, maintaining anatomical structure consistency over time.
• Transformations generated by this method better preserve structures undergoing large deformations compared to intensity-only methods.
• A three-orders-of-magnitude speedup was achieved using GPU acceleration versus a CPU implementation.

Conclusions:

• The developed registration technique offers a robust solution for complex anatomical changes and contrast differences.
• GPU-accelerated particle mesh approximation significantly enhances computational feasibility.
• This method is suitable for time-critical medical imaging applications requiring precise deformable registration.