Weakly-supervised Learning for Single-step Quantitative Susceptibility Mapping

TL;DR

This paper introduces a weakly-supervised deep learning method for direct single-step QSM reconstruction from MRI data, overcoming limitations of previous methods that relied on synthetic or multi-orientation data.

Contribution

The proposed wTFI method enables direct QSM reconstruction without background field removal, using local fields as supervision, and improves accuracy near brain edges.

Findings

High-quality susceptibility maps generated across neuroimaging contexts

Effective recovery of magnetic susceptibility near brain edges

Outperforms existing methods in reliability and quality

Abstract

Quantitative susceptibility mapping (QSM) utilizes MRI phase information to estimate tissue magnetic susceptibility. The generation of QSM requires solving ill-posed background field removal (BFR) and field-to-source inversion problems. Because current QSM techniques struggle to generate reliable QSM in clinical contexts, QSM clinical translation is greatly hindered. Recently, deep learning (DL) approaches for QSM reconstruction have shown impressive performance. Due to inherent non-existent ground-truth, these DL techniques use either calculation of susceptibility through multiple orientation sampling (COSMOS) maps or synthetic data for training, which are constrained by the availability and accuracy of COSMOS maps or domain shift when training data and testing data have different domains. To address these limitations, we propose a weakly-supervised single-step QSM reconstruction method, denoted as wTFI, to directly reconstruct QSM from the total field without BFR. wTFI uses the BFR method RESHARP local fields as supervision to perform a multi-task learning of local tissue fields and QSM, and is capable of recovering magnetic susceptibility estimates near the edges of the brain where are eroded in RESHARP and realize whole brain QSM estimation. Quantitative and qualitative evaluation shows that wTFI can generate high-quality local field and susceptibility maps in a variety of neuroimaging contexts.