DescriptionFor designing efficient system solvers for linear systems from 3D image registration, one may analyze structures of general operators or matrices used by the algorithms. In our case, although these systems are often sparse and structured, they are very large and ill-conditioned. Thus, their solvers are time consuming and their complexity in operation counts is polynomial. As a consequence, fast and superfast direct methods reveal numerical instabilities and thus lead to breakdowns or inaccurate solutions.
The most used alternative are iterative system solvers that enable a reduction of the number of expensive operations such as matrix-vector products and thus a speed up of the registration process. Although iterative solvers provide only an approximation of the solution, they are well suited for very large systems when cheap and well suited preconditioners are available. Preconditioners may be stationary (Jacobi, Gauss-Seidel or SOR) and non-stationary (polynonmial or low-rank tensors).
In this study, we are especially interested in benchmarking iterative system solvers on a large set of 3D medical images. These system solvers are based on the Conjugate Gradient method but different from the preconditioning techniques .
The ongoing results confirm that there is no single system solver that is the best for all the problems. However, the results indicate which solver has the highest probability of being the best within a factor $f \in [1,\infty[$ and considering a limited computional budget in terms of time and storage requirements.
|Période||1 févr. 2018 → 2 févr. 2018|
|Conservé à||University of Liege, 32nd ORBEL conference, Belgique|