SHARQnet - Sophisticated Harmonic Artifact Reduction in Quantitative Susceptibility Mapping using a Deep Convolutional Neural Network

Abstract

Quantitative susceptibility mapping (QSM) reveals pathological changes in widespread diseases such as Parkinson's disease, Multiple Sclerosis, or hepatic iron overload. QSM requires multiple processing steps after the acquisition of magnetic resonance imaging (MRI) phase measurements such as unwrapping, background field removal and the solution of an ill-posed field-to-source-inversion. Current techniques utilize iterative optimization procedures to solve the inversion and background field correction, which are computationally expensive and lead to suboptimal or over-regularized solutions requiring a careful choice of parameters that make a clinical application of QSM challenging. We have previously demonstrated that a deep convolutional neural network can invert the magnetic dipole kernel with a very efficient feed forward multiplication not requiring iterative optimization or the choice of regularization parameters. In this work, we extended this approach to remove background fields in QSM. The prototype method, called SHARQnet, was trained on simulated background fields and tested on 3T and 7T brain datasets. We show that SHARQnet outperforms current background field removal procedures and generalizes to a wide range of input data without requiring any parameter adjustments. In summary, we demonstrate that the solution of ill-posed problems in QSM can be achieved by learning the underlying physics causing the artifacts and removing them in an efficient and reliable manner and thereby will help to bring QSM towards clinical applications.

Overview of quantitative susceptibility mapping using deep learning: Current status, challenges and opportunities

“…Deep learning-based techniques can offer an increase in robustness and speed-up in computation. For example, SHARQnet 69 was designed for background field removal using a 3D convolutional neural network and was trained on synthetic background fields overlaid on top of a brain simulation. When the performance was compared with SHARP, RESHARP and V-SHARP, SHARQnet delivered accurate background field corrections on simulations and on in vivo data.…”

Improving phase‐based conductivity reconstruction by means of deep learning–based denoising of phase data for 3T MRI

“…Recently, several studies 44,47,48 for DL have suggested using simulated data as an alternative to in vivo data for training, which allows us to generate ground-truth information without the need of multiple scans. In addition, a similar non-DL approach (ie, machine learning-based algorithms) using simulated training data termed dictionary-based conductivity mapping 63 has previously been suggested.…”

“…Recently, deep learning (DL) approaches have been introduced for various image reconstruction problems in MR. [37][38][39][40][41][42][43][44][45][46][47] Furthermore, it has recently been suggested to use DL for conductivity reconstructions as an alternative to the computation of the spatial derivatives. [48][49][50] These approaches describe the foundation of alternative algorithms, which allow solving the conductivity reconstruction problem.…”

Accelerating quantitative susceptibility and R2* mapping using incoherent undersampling and deep neural network reconstruction

“…Although many previous deep neural networks [55][56][57][58][59][60][61][62] We also compared DCRNet with one previous deep learning-based MRI phase reconstruction method (Phase-Unet). It was found that the phase wraps in the reconstructed images were disrupted by Phase-Unet, making it problematic for the phase unwrapping process of the QSM pipeline.…”