HetSI: Heterogeneous System Identification for the Analysis of Magnetic Resonance Elastography Data

Magnetic resonance elastography (MRE) is a non-invasive method for estimating mechanical parameters of tissues. To infer the parameters one performs an inversion; finding system parameters from the output displacement field. Inverse solving methods typically rely on forward models for the system dynamics. An accurate model is required to properly infer the system parameters, typically based on mechanical models. SINDy is a method for inferring differential equations from sampled data, providing insight into an appropriate forward model. Our work implements an extension of the SINDy framework, called HetSI, incorporating heterogeneous parametric fields, continuous representations and a stateful library, increasing the usefulness for MRE. The method was tested on samples from four public multifrequency MRE datasets regarding system identification, indicating an ability to identify a model, as well as regions of improper forward modeling or insufficient information. Inversion capabilities were compared to two baseline methods using one simulated sample. Estimates were comparable to baselines, and seem to exceed them in regions of rapid change. HetSI forms a new approach in MRE, providing insight into forward models as well as parametric distributions.

First annual seminar

Opponents: PhD student Henrik Palme, Postdoc Zhongzheng Wang