Atlantis Institut des Sciences Fictives

Hernández, Mónica ; Olmos Gassó, Salvador

13 June 2008

Abstract: In this article, we propose a numerical implementation of Gauss-Newton's method for optimization in diffeomorphic registration in the Large Deformation Diffeomorphi c Metric Mapping framework. The computations of the G\^ateaux derivatives of the o bjective function are performed in the tangent space of the Riemannian manifold of diffeo morphisms. The resulting algorithm has been compared to gradient descent optimization in br ain MRI anatomical images. The experiments have shown similar accuracy for both tech niques at steady-state while Gauss-Newton has resulted to be more robust with a faster rate of convergence.

Keyword(s): diffeomorphic registration ; gauss-newton ; hilbert spaces ; optimization methods