Advances in MRI based Electrical Properties Tomography: a Comparison between Physics-supported Learning Approaches

Magnetic resonance imaging (MRI) is widely used in several medical applications, which include the non-invasive and in-vivo investigation of the electrical properties of biological tissues. Such kind of inverse problem can be addressed by means of iterative methods, which are time and memory consuming and solution may converge to local minima. To accelerate the reconstructions and bypass the problem of local minima, we propose and compare two different learning methods to face the inverse problem underlying the MRI based electrical properties tomography, one based on supervised descent method and the other one on a cascade of multi-layers convolutional neural networks. Both methods are trained and tested using 2D simulated data of a human head model and show a good reconstruction capability. Better generalization ability can be achieved by using the CNN-based iterative approach.