CLAIRE: Constrained Large Deformation Diffeomorphic Image Registration on Parallel Computing Architectures.

Area of Science:

• Computational imaging
• Medical image analysis
• Scientific computing

Background:

• Image registration is crucial for comparing medical images.
• Large deformation registration presents significant computational challenges.
• Existing frameworks may lack scalability or flexibility.

Purpose of the Study:

• Introduce CLAIRE, a computational framework for Constrained Large deformation diffeomorphic Image Registration.
• Highlight CLAIRE's support for parallel CPU and GPU architectures.
• Detail the numerical methods and optimization strategies employed by CLAIRE.

Main Methods:

• Utilizes MPI for distributed-memory parallelism across multi-node CPU and GPU systems.
• Employs semi-Lagrangian interpolation and high-order finite differences/FFTs for computational kernels.
• Integrates Newton-Krylov solvers with PETSc and TAO for numerical optimization.

Main Results:

• Demonstrates scalability on thousands of cores and multiple GPU devices.
• Achieves optimized computational throughput via efficient linear algebra operations and preconditioners.
• Supports various regularization schemes and similarity measures for flexible registration.

Conclusions:

• CLAIRE provides a powerful, scalable, and flexible platform for large deformation image registration.
• Its optimized computational kernels and numerical solvers enable efficient and accurate registration.
• The framework is publicly available, facilitating further research and application.