Talk: Ass. Prof. Dr. Andreas Mang

On June 20, 2018 3 pm ct, Ass. Prof. Dr. Andreas Mang from the University of Houston Texas, Department of Mathematics will give a talk on "CLAIRE: A Parallel Solver for Constrained Diffeomorphic Image Registration".

Location: MIC-Arena, Maria-Goeppert-Str. 3, 23562 Lübeck

Title: "CLAIRE: A Parallel Solver for Constrained Diffeomorphic Image Registration"

I will present algorithms for diffeomorphic image registration that efficiently run on multi-core systems and have been shown to scale on high-performance computing platforms with thousands of cores.

We use a PDE-constrained formulation. In its simplest form, these constraints are a hyperbolic transport equation for the image intensities. The control variable is the velocity field. I will present different formulations and discuss our strategy for their solution.

Our code is implemented in C++ and uses the message passing interface library for parallelism. Our solver is based on a globalized, matrix-free, preconditioned Newton--Krylov method to solve the optimization problem. We use a spectral discretization in space and solve the transport equations that appear in our optimality systems using an unconditionally stable, semi-Lagrangian scheme. We use a matrix-free, two-level multigrid preconditioner for the iterative inversion of the reduced space Hessian.

We will study the rate of convergence, time-to-solution, registration quality, and the scalability of our solver for real and synthetic test problems.

This is joint work with George Biros.



09.05.2018 13:04