Clinical Quantitative Susceptibility Mapping (QSM) – Biometal
Imaging and its Emerging Roles in Patient Care
Yi Wang, PhD
1,2,*
, Pascal Spincemaille, PhD
1
, Zhe Liu, BS
1,2
, Alexey Dimov, MS
1,2
, Kofi
Deh, BS
1
, Jianqi Li, PhD
3
, Yan Zhang, MMed
4
, Yihao Yao, MD
1,4
, Kelly M. Gillen, PhD
1
, Alan
H. Wilman, PhD
5
, Ajay Gupta, MD
1
, Apostolos John Tsiouris, MD
1
, Ilhami Kovanlikaya,
MD
1
, Gloria Chia-Yi Chiang, MD
1
, Jonathan W. Weinsaft, MD
6
, Lawrence Tanenbaum, MD
7
,
Weiwei Chen, MD, PhD
4
, Wenzhen Zhu, MD
4
, Shixin Chang, MD
8
, Min Lou, MD, PhD
9
, Brian
H. Kopell, MD
10
, Michael G. Kaplitt, MD, PhD
11
, David Devos, MD, PhD
12,13,14,15
, Toshinori
Hirai, MD, PhD
16
, Xuemei Huang, MD, PhD
17,18,19,20
, Yukunori Korogi, MD, PhD
21
,
Alexander Shtilbans, MD, PhD
22,23
, Geon-Ho Jahng, PhD
24
, Daniel Pelletier, MD
25
, Susan A.
Gauthier, DO, MPH
26
, David Pitt, MD
27
, Ashley I. Bush, MD, PhD
28
, Gary M. Brittenham,
MD
29
, and Martin R. Prince, MD, PhD
1
1
Department of Radiology, Weill Cornell Medical College, New York, NY, USA
2
Department of Biomedical Engineering, Ithaca, NY, USA
3
Department of Physics, East China Normal University, Shanghai, China
4
Department of Radiology, Tongji Hospital, Tongji Medical College, Huazhong University of
Science & Technology, Wuhan, China
5
Department of Biomedical Engineering, University of Alberta, Edmonton, AB, Canada
6
Division of Cardiology, Department of Medicine, Weill Cornell Medical College, New York, NY,
USA
7
RadNet, Inc. Los Angeles, CA, USA
8
Department of Radiology, Yueyang Hospital of Integrated Traditional Chinese & Western
Medicine, Shanghai, China
9
Department of Neurology, the Second Affiliated Hospital of Zhejiang University, School of
Medicine, Hangzhou, China
10
Department of Neurosurgery, Mount Sinai Hospital, New York, NY, USA
11
Department of Neurological Surgery, Weill Cornell Medical College, New York, NY, USA
12
Department of Medical Pharmacology, University of Lille, Lille, France
13
Department of Neurology and Movement Disorders, University of Lille, Lille, France
14
Department of Toxicology, Public Health and Environment, University of Lille, Lille, France
15
INSERM U1171, University of Lille, Lille, France
*
To whom correspondence should be addressed: Yi Wang, PhD, 515 E 71th Street, New York, NY, 10021, USA. Tel: 646-962-2631;
yiwang@med.cornell.edu.
HHS Public Access
Author manuscript
J Magn Reson Imaging
. Author manuscript; available in PMC 2018 October 01.
Published in final edited form as:
J Magn Reson Imaging
. 2017 October ; 46(4): 951–971. doi:10.1002/jmri.25693.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
16
Department of Radiology, Faculty of Medicine, University of Miyazaki, Miyazaki, Japan
17
Department of Neurology, Penn State University-Milton S. Hershey Medical Center, Hershey,
PA, USA
18
Department of Pharmacology, Penn State University-Milton S. Hershey Medical Center,
Hershey, PA, USA
19
Department of Neurosurgery, Penn State University-Milton S. Hershey Medical Center, Hershey,
PA, USA
20
Department of Radiology, Penn State University-Milton S. Hershey Medical Center, Hershey,
PA, USA
21
Department of Radiology, School of Medicine, University of Occupational and Environmental
Health, Kitakyushu, Japan
22
Department of Neurology, Hospital for Special Surgery, New York, NY, USA
23
Parkinson's Disease and Movement Disorder Institute, Weill Cornell Medical College, New York,
NY, USA
24
Department of Radiology, Kyung Hee University Hospital at Gangdong, College of Medicine,
Kyung Hee University, Seoul, South Korea
25
Department of Neurology, Department of Neurology, Keck School of Medicine of the University
of Southern California, Los Angeles, CA, USA
26
Department of Neurology and Neuroscience, Weill Cornell Medical College, New York, NY, USA
27
Department of Neurology, School of Medicine, Yale University, New Haven, CT, USA
28
Oxidation Biology Unit, The Florey Institute of Neuroscience and Mental Health, The University
of Melbourne, Parkville, Victoria 3010, AUS
29
Department of Pediatrics, Columbia University, Children's Hospital of New York, New York, NY,
USA
Abstract
Quantitative susceptibility mapping (QSM) has enabled MRI of tissue magnetic susceptibility to
advance from simple qualitative detection of hypointense blooming artifacts to precise quantitative
measurement of spatial biodistributions. QSM technology may be regarded to be sufficiently
developed and validated to warrant wide dissemination for clinical applications of imaging
isotropic susceptibility, which is dominated by metals in tissue, including iron and calcium. These
biometals are highly regulated as vital participants in normal cellular biochemistry, and their
dysregulations are manifested in a variety of pathologic processes. Therefore, QSM can be used to
assess important tissue functions and disease. To facilitate QSM clinical translation, this review
aims to organize pertinent information for implementing a robust automated QSM technique in
routine MRI practice and to summarize available knowledge on diseases for which QSM can be
used to improve patient care. In brief, QSM can be generated with postprocessing whenever
gradient echo MRI is performed. QSM can be useful for diseases that involve neurodegeneration,
inflammation, hemorrhage, abnormal oxygen consumption, substantial alterations in highly
Wang et al. Page 2
J Magn Reson Imaging
. Author manuscript; available in PMC 2018 October 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
paramagnetic cellular iron, bone mineralization, or pathologic calcification; and for all disorders in
which MRI diagnosis or surveillance requires contrast agent injection. Clinicians may consider
integrating QSM into their routine imaging practices by including gradient echo sequences in all
relevant MRI protocols.
Introduction
Quantitative susceptibility mapping (QSM) solves the deconvolution or inverse problem
from magnetic field to susceptibility source to map a local tissue magnetic property (1,2).
This local property is fundamentally different from the nonlocal property of traditional
gradient echo (GRE) MRI, including susceptibility weighted imaging (SWI), the closely
related GRE magnitude T2*-weighted imaging (T2*w), and GRE phase imaging (Phase),
although both QSM and traditional GRE MRI are regarded as being sensitive to
susceptibility (3-5). Without deconvolution, traditional GRE MRI generally suffers from
blooming artifacts, which 1) may generate contrasts at neighboring locations without
susceptibility sources, in addition to at locations with susceptibility sources; 2) strongly
depend on imaging parameters, including field strength, voxel size and echo time; and 3)
deceptively vary with object orientations, where tissue interfaces with susceptibility
differences perpendicular to the main field B0 have much greater contrasts than interfaces
parallel to B0 (6). With deconvolution, QSM eliminates the problem of blooming artifacts
and provides quantitative distribution of susceptibility sources in tissue. Without
deconvolution, traditional GRE MRI can only detect the presence of susceptibility interfaces
perpendicular to B0, and cannot localize or quantify any susceptibility source. With
deconvolution, QSM can precisely localize and quantify these sources.
The long-standing desire to determine susceptibility sources in tissue arose in the early days
of MRI (7). Despite this, the quest to quantify susceptibility as an inverse problem may not
have begun in earnest until 2001 (8). Early efforts did not lead to successful susceptibility
mapping (9-12), because they failed to identify additional information needed to solve the
ill-posed field-to-source inverse problem. A major technological breakthrough came in 2008
when the Bayesian inference with a morphological prior was introduced to form the
foundation for QSM (1,13-15). Bayesian inference is a statistical method to optimally
estimate susceptibility from both field data that is noisy and incomplete and tissue structure
information that also has its uncertainty. Since 2008, research efforts to develop the details
of the Bayesian QSM approach have mushroomed, including robust field extraction from
MRI signal and effective morphological regularization (6,16-36). The tremendous QSM
development efforts in the past 8 years, as evidenced by an exponential growth in the
number of QSM papers, have propelled QSM technology from basic research to adaptation
and investigation for clinical applications.
QSM accurately maps strong isotropic susceptibility sources in human tissue –
predominantly biometals that are highly paramagnetic (mainly iron in ferritin or
deoxygenated heme) or present in high concentrations (mainly calcium in mineralization or
calcification). QSM of biometals has been valuable in studying disease processes. QSM is
shown to be reproducible across scanner makers, models, field strengths, and sites (37-40).
Wang et al. Page 3
J Magn Reson Imaging
. Author manuscript; available in PMC 2018 October 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
QSM can be automated, making it ready for wide dissemination to evaluate its diagnostic
and therapeutic value in clinical practice. This will enable clinical investigations both
longitudinally and across-centers, ushering in a new era of clinical QSM applications.
QSM can be used to study susceptibility sources other than biometals, particularly white
matter (WM) fibers with anisotropic susceptibilities (17). However, anisotropic
susceptibility imaging may require much more technical development to overcome the
requirement of multiple orientations before it can be applied in clinical studies (29,41).
Since most other susceptibility sources in human tissue are much weaker than the dominant
biometals, we choose to focus on biometal QSM for timely and promising clinical QSM
developments, while emphasizing the connection between pathogenic biometals and patient
care that is beyond the reach of conventional MRI. We aim to provide readers with basic
information on how to 1) implement a robust and automated QSM in their practice, 2)
understand the roles of biometals in human health and diseases, and 3) use QSM
measurements of biometals in clinical applications.
Robust and Automated QSM
In this technical section on QSM, we aim to provide a conceptual appreciation of the
principles of robust QSM based on the Bayesian approach. For integration into daily clinical
workflows, we describe an automated QSM that can be implemented across a wide range of
major MRI manufacturers, including GE, Philips and Siemens, at both 1.5 and 3 Tesla. The
automation and standardization in implementing QSM for biometal imaging is fortunately
made possible by the results from rich variations in the Bayesian approach (2).
Fundamental principles of robust QSM
The main idea underlying QSM is to extract the susceptibility source from its blooming
artifacts on traditional GRE MRI through rigorous biophysical modeling of the MRI signal
phase. Phase has historically been largely discarded in routine MRI practice, though MRI
data is inherently complex, consisting of half phase and half magnitude. Yet, phase data
provides rich insight into tissue properties that are complementary to magnitude data (42).
Recalling that signal in clinical MRI comes from water (and sometimes fat) protons, phase
reflects the inhomogeneous magnetic field experienced by protons. The field sources consist
of tissue molecular electron clouds and background sources outside tissue. They become
magnetized in the MRI main field B0 according to their magnetic susceptibilities and
contribute to the magnetic field as dipoles according to Maxwell's equation. The tissue field
and background field can be separated according to their source location difference
(background field removal). Therefore, MRI phase can be processed to generate the tissue
field, which can be analyzed according to the dipole field model to determine tissue
magnetic susceptibility (Fig. 1).
The magnetic field at a location is the sum of contributions from all surrounding dipole
sources. Mathematically speaking, the field is a convolution between the susceptibility
spatial distribution and the field of a unit dipole (dipole kernel). Consequently, the
determination of tissue susceptibility requires deconvolution of the tissue field with the
dipole kernel. Deconvolution in image space is division in k-space (the Fourier convolution
Wang et al. Page 4
J Magn Reson Imaging
. Author manuscript; available in PMC 2018 October 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
theorem). The challenge for this dipole kernel division is that the dipole kernel is zero when
an observation point relative to the dipole source is at ±54.7° (magic angles) with respect to
the B0 direction. The observed field contribution at the magic angles should be zero, but
there is always noise in the measured data. The resulting division-by-zero of noise (and other
data deviation from the dipole field type) leads to streaking artifacts along the magic angles
in k-space.
These streaking artifacts are cone-like surfaces distinct from tissue surfaces, manifesting as
prominent lines in image space along the complementary magic angles in the sagittal and
coronal views and rings in the axial view. Early efforts in solving the field-to-susceptibility
inverse problem were not effective in identifying and minimizing streaking artifacts; in fact,
the truncated k-space division method amplified the streaking artifacts by increasing the
deviation from the dipole field type (12,20). The Bayesian approach enables robust
suppression of streaking artifacts by tenaciously searching for a solution of minimal
streaking (1,14,15). Mathematically, minimal streaking is characterized by penalizing
interfaces distinct from tissue interfaces depicted on an anatomic MRI during the search for
a susceptibility distribution that satisfies the measured field data. Both noise in the field data
and uncertainty in the definition of tissue interfaces are considered in a balanced manner
(discrepancy principle) during this tenacious search or numerical optimization, which is
termed “Bayesian machine learning” in signal processing or data science (43). While this
Bayesian reconstruction is robust (convex optimization), its computation is much costlier
than Fourier transform in standard MRI reconstructions. Fortunately, modern numerical
optimization tools have allowed the search to be completed within a few minutes on a
reasonably equipped desktop computer, now enabling robust QSM in a clinical setting.
Automated QSM processing
Until a commercial product is available to automatically generate QSM, we recommend the
following steps to implement automated QSM on the major scanners for clinical
investigations: QSM can be regarded as a postprocessing technique for GRE MRI. The most
important factor for enabling QSM is to save faithfully the complex data (both real and
imaginary parts, or both magnitude and phase parts) acquired by a GRE MRI, particularly
without adulteration of the phase data.
Once QSM protocols are setup on the scanner to produce these images in DICOM format, a
technologist, or ideally an automated image management program on the scanner, can
forward these images to a dedicated DICOM image server that is listening for incoming
GRE images, from which it reconstructs the QSM images and sends them back to the
scanner. The process is automatic and is usually completed within 5 to 10 minutes
depending on the computing performance of the server, the connection bandwidth between
the scanner and the server, and the matrix size of the GRE data. The advantage of using
DICOM is that it is available on all scanner platforms, does not require installation of extra
software on the scanner, and has high quality open source implementations.
Wang et al. Page 5
J Magn Reson Imaging
. Author manuscript; available in PMC 2018 October 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript