[Paper Review] Deep Quantitative Susceptibility Mapping for Background Field Removal and Total Field Inversion

Juan Liu, Kevin M. Koch|arXiv (Cornell University)|May 28, 2019

Advanced MRI Techniques and ApplicationsMedicine38 references3 citations

TL;DR

This paper proposes QSMAllNet, a deep learning framework that jointly performs background field removal and total field inversion for quantitative susceptibility mapping (QSM) from single-contrast, clinical SWI data. Trained on physics-based synthetic data derived from in-vivo QSM scans, the model achieves superior accuracy and artifact reduction compared to state-of-the-art methods, enabling robust, high-resolution QSM reconstruction without additional scans beyond standard clinical protocols.

ABSTRACT

Quantitative susceptibility mapping (QSM) utilizes MRI signal phase to estimate local tissue susceptibility, which has been shown useful to provide novel image contrast and as biomarkers of abnormal tissue. QSM requires addressing a challenging post-processing problem: filtering of image phase estimates and inversion of the phase to susceptibility relationship. A wide variety of quantification errors, robustness limitations, and artifacts constraints QSM clinical translation. To overcome these limitations, a robust deep-learning-based QSM reconstruction approach is proposed to perform background field removal and susceptibility inversion simultaneously from input MRI phase images. Synthetic training data based on in-vivo data sources and physics simulations were used for training. The network was quantitatively tested using gold-standard in-silico labeled dataset against established background field removal and QSM inversion approaches. In addition, the algorithm was applied to a QSM challenge data and clinical susceptibility-weighted imaging (SWI) data. When quantitatively compared against gold-standard in-silico labels, the proposed algorithm outperformed the existing comparable background field removal approaches and QSM reconstruction algorithms. The QSM challenge data and clinical SWI data demonstrated that the proposed approach was able to robustly generate high quality local field and QSM with improved accuracy.

Motivation & Objective

• To address the limitations of existing QSM reconstruction methods, including parameter tuning, long computation times, and residual artifacts in regions with large susceptibility variations.
• To overcome the lack of ground-truth training data for QSM by generating synthetic training data using in-vivo QSM datasets and physical susceptibility-to-field models.
• To develop a single, end-to-end deep learning model that performs both background field removal and susceptibility inversion simultaneously, reducing error propagation from sequential processing.
• To enable high-quality, artifact-free QSM reconstruction directly from routinely acquired clinical SWI data, without requiring additional scans or multi-orientation acquisitions.
• To improve diagnostic utility of QSM in clinical settings by enhancing image quality and quantification accuracy, especially near pathological regions like hemorrhages and calcifications.

Proposed method

• The model is trained on synthetic datasets generated from 200 in-vivo QSM datasets using a physics-based forward model that maps tissue susceptibility to local magnetic fields.
• A U-Net-like convolutional neural network architecture is employed, with dilated convolutions to expand the receptive field for capturing non-local susceptibility effects.
• The network is trained to predict both local field and susceptibility maps end-to-end from raw phase images, sharing a common latent space for joint optimization.
• The training objective minimizes L2 loss between predicted and ground-truth local field and susceptibility maps in the synthetic data.
• The model leverages a single network for both background field removal and susceptibility inversion, avoiding the need for separate networks or sequential processing.
• The architecture is designed to be memory-efficient and computationally lightweight, enabling fast inference on clinical data.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on synthetic data achieve superior performance in background field removal and QSM inversion compared to established two-step and single-step methods?
• RQ2 To what extent does the proposed end-to-end network reduce artifacts such as streaking and shading compared to conventional QSM reconstruction techniques?
• RQ3 Can the model generalize to clinical SWI data with non-isotropic resolution and complex pathology, such as intracranial hemorrhages and calcifications?
• RQ4 Does the joint optimization of background field removal and susceptibility inversion eliminate error propagation from separate processing steps?
• RQ5 Can the model produce high-quality QSM maps from standard clinical SWI sequences without requiring additional acquisitions or multi-orientation data?

Key findings

• On the gold-standard in-silico labeled dataset, QSMAllNet outperformed existing background field removal and QSM reconstruction methods in quantitative metrics, including lower mean absolute error and higher correlation with ground truth.
• In clinical SWI data, QSMAllNet produced QSM maps with improved image sharpness, clearer tissue structures, and no visible streaking or shading artifacts near hemorrhagic regions, unlike SS-TV-QSM, PDF+MEDI, RESHARP+TKD, and LN-QSM.
• The method successfully removed residual background fields and eliminated shading artifacts in regions with large susceptibility variations, such as near microbleeds and calcifications, as confirmed by visual and quantitative assessment.
• Compared to two-step methods like RESHARP+TKD and PDF+MEDI, QSMAllNet achieved better accuracy and reduced error propagation by performing both steps jointly in a single network.
• The model demonstrated robust performance on non-isotropic clinical SWI data, producing diagnostically useful QSM maps without requiring specialized acquisition protocols.
• The use of dilated convolutions enabled effective non-local susceptibility estimation, contributing to improved artifact suppression and structural preservation.