CLAIRE

CLAIRE performs velocity-based large-deformation diffeomorphic image registration in three dimensions for precise, scalable image alignment and analysis.

Key Features:

• Implementation: Implemented in C/C++ with highly optimized parallel computational kernels for multi-node CPU and multi-node multi-GPU architectures.
• Registration model: Velocity-based diffeomorphic image registration for three-dimensional images.
• Distributed parallelism: Uses MPI (Message Passing Interface) for distributed-memory parallelism and scales effectively to thousands of cores and GPU devices.
• Multi-GPU communication: Multi-GPU implementation that utilizes device direct communication to optimize performance across multiple GPUs.
• Semi-Lagrangian integration: Employs semi-Lagrangian time integration for interpolation.
• Differentiation operators: Uses high-order finite difference operators and Fast-Fourier-Transforms (FFTs) for differentiation.
• Optimization solver: Uses a Newton–Krylov solver for numerical optimization.
• Regularization: Incorporates various regularization schemes to address the control problem in diffeomorphic transformations.
• Similarity measures: Supports multiple similarity measures for image matching.
• Preconditioning: Implements several preconditioners for the reduced-space Hessian to accelerate convergence.
• Linear algebra: Utilizes PETSc (Portable, Extensible Toolkit for Scientific Computation) for scalable linear algebra operations and solvers.
• Optimization toolkit: Uses TAO (Toolkit for Advanced Optimization) for numerical optimization tasks.

Scientific Applications:

• Three-dimensional image alignment: Precise diffeomorphic registration of 3D volumetric images.
• Large-scale image registration: Large-scale image registration and analysis on multi-node CPU and multi-GPU systems.
• Diffeomorphic control problems: Solving control problems inherent to diffeomorphic transformations that require regularization and Hessian preconditioning.
• Bioinformatics imaging: Applications in bioinformatics that require accurate, large-scale image registration.

Methodology:

Semi-Lagrangian time integration for interpolation; high-order finite difference operators and Fast-Fourier-Transforms (FFTs) for differentiation; Newton–Krylov solver with preconditioners for the reduced-space Hessian; MPI-based distributed-memory parallelism and multi-GPU device-direct communication; uses PETSc and TAO; implemented in C/C++.