Opening a new window on MR-based Electrical Properties Tomography with deep learning

TL;DR

This paper introduces a deep learning approach for Electrical Properties Tomography that produces high-quality, noise-robust tissue property maps from MRI data, enabling new diagnostic possibilities.

Contribution

It presents a novel neural network-based method for reconstructing tissue electrical properties from MRI, overcoming noise issues and enabling permittivity imaging at 3 Tesla.

Findings

Deep learning yields accurate EP maps with high precision.

DL-EPT enables permittivity imaging at 3 Tesla, not possible with traditional methods.

The approach is highly noise-robust, allowing faster MRI acquisitions.

Abstract

Electrical properties (EPs) of tissues, conductivity and permittivity, are modulated by the ionic and water content, which change in presence of pathologies. Information on tissues EPs can be used e.g. as an endogenous biomarker in oncology. MR-Electrical Properties Tomography (MR-EPT) aims to reconstruct tissue EPs by solving an electromagnetic inverse problem relating MR measurements of the transmit radiofrequency RF field to the EPs. However, MR-EPT reconstructions highly suffer from noise in the RF field maps, which limits the clinical applicability. Instead of employing electromagnetic models posing strict requirements on the measured quantities, we propose a data driven approach where the inverse transformation is learned by means of a neural network. Supervised training of a conditional generative adversarial neural network was performed using simulated realistic RF field maps and realistic human head dielectric models. Deep learning EPT (DL-EPT) reconstructions are presented for in-silica MR data and MR measurements at 3 Tesla on phantoms and human brains. DL-EPT shows high quality EP maps, demonstrating good accuracy and greatly improved precision compared to conventional MR-EPT. Moreover, DL-EPT allows permittivity reconstructions at 3 Tesla, which is not possible with state-of-art MR-EPT techniques. The supervised learning-based approach leverages the strength of tailored electromagnetic simulations, allowing inclusion of a priori information (e.g. coil setup) and circumvention of inaccessible MR electromagnetic quantities. Since DL-EPT is highly noise-robust, the requirements for MRI data acquisitions can be relaxed, allowing faster acquisitions and higher resolutions. We believe that DL-EPT greatly improves the quality and applicability of EPT opening a new window for an endogenous biomarker in MRI diagnostics that reflects differences in ionic tissue content.