We present CLAIRE, a parallel solver for constrained large deformation diffeomorphic image registration in three dimensions. Our contributions are the following: (i) We present an improved implementation of our memory-distributed, globalized, preconditioned Newton--Krylov solver. (ii) We present effective techniques to precondition the reduced space Hessian. (iii) We study numerical accuracy, rate of convergence, time-to-solution, inversion quality, and scalability of our solver. We use a PDE-constrained formulation for diffeomorphic image registration. The PDE constraints are the transport equations for the image intensities. The control variable is the velocity field. The discretization of the optimality system leads to high-dimensional, ill-conditioned, multiphysics systems that are challenging to solve in an efficient way. Our code is implemented in C/C++ and uses the message passing interface (MPI) library for parallelism. We study the performance of our solver for multi-subject registration problems in neuroimaging. We will see that our solver is competitive in terms of time-to-solution and registration quality. We will see that we can solve problems on clinical images (50 million unknowns) in less than two minutes on a workstation with 40 cores. If we use 512 MPI tasks we can reduce the runtime to under 2 seconds, paving the way to tackle real-time applications.

This presentation is part of Minisymposium “MS28 - Diffeomorphic Image Registration: Numerics, Applications, and Theory (2 parts)”
organized by: Andreas Mang (Department of Mathematics, University of Houston) , George Biros (Institute for Computational Engineering and Sciences, University of Texas at Austin) .

Authors:

Andreas Mang (Department of Mathematics, University of Houston)

Amir Gholami (UC Berkeley)

George Biros (Institute for Computational Engineering and Sciences, University of Texas at Austin)

Keywords:

image registration, inverse problems, nonlinear optimization