Recommended Implementation of Quantitative Susceptibility Mapping for Clinical Research in The Brain: A Consensus of the ISMRM Electro-Magnetic Tissue Properties Study Group

Abstract

This article provides recommendations for implementing quantitative susceptibility mapping (QSM) for clinical brain research. It is a consensus of the International Society of Magnetic Resonance in Medicine Electro-Magnetic Tissue Properties Study Group. While QSM technical development continues to advance rapidly, the current QSM methods have been demonstrated to be repeatable and reproducible for generating quantitative tissue magnetic susceptibility maps in the brain. However, the many QSM approaches available have generated a need in the neuroimaging community for guidelines on implementation. This article outlines considerations and implementation recommendations for QSM data acquisition, processing, analysis, and publication. We recommend that data be acquired using a monopolar 3D multi-echo GRE sequence and that phase images be saved and exported in DICOM format and unwrapped using an exact unwrapping approach. Multi-echo images should be combined before background field removal, and a brain mask created using a brain extraction tool with the incorporation of phase-quality-based masking. Background fields within the brain mask should be removed using a technique based on SHARP or PDF, and the optimization approach to dipole inversion should be employed with a sparsity-based regularization. Susceptibility values should be measured relative to a specified reference, including the common reference region of the whole brain as a region of interest in the analysis. The minimum acquisition and processing details required when reporting QSM results are also provided. These recommendations should facilitate clinical QSM research and promote harmonized data acquisition, analysis, and reporting.

1. Introduction

Brain quantitative susceptibility mapping (QSM) is increasingly being used to identify calcifications and study the concentration of iron and myelin in tissue, oxygen consumption, as well as changes in these quantities associated with normal brain development, aging, and neurological diseases. These diseases include hemorrhagic stroke, multiple sclerosis (MS), Alzheimer’s disease, Parkinson’s disease, amyotrophic lateral sclerosis, and tumors ^1–15. Particularly, QSM 1) enables differentiation between tissues that are paramagnetic (relative to water), including hemorrhages, and more diamagnetic than water, including calcifications ^16–19, and their quantification with additional R₂ or R₂* mapping ^1,20–29 and 2) improves anatomical visualization ^30–32 and segmentation ^33,34. The latter has proven valuable for presurgical mapping of deep gray nuclei targets in deep brain stimulation ^30,35–39. QSM is also utilized to study imbalances in neurotransmitter levels or metabolism in psychiatric disorders such as psychosis ^40,41 and depression ^42,43, and to investigate neural mechanisms through which alcohol affects the brain ^44–46. An important feature of QSM, when compared to, e.g., R₁ or R₂ relaxometry, is that its results may be more independent of field strength and scanner manufacturer when a consistent reconstruction QSM pipeline is used ^47,48.

Nevertheless, there is currently no consensus or white paper on how to perform brain QSM in the clinical setting, despite extensive reviews on QSM techniques and applications ^{3,6,7,49–62} and strong QSM community interest in identifying the best algorithms ^63–66. QSM is not yet a standard product on most MRI systems, although some scanner vendors are beginning to offer works-in-progress or even product software packages for QSM. However, there are growing demands for a robust “end-to-end” QSM recipe that covers acquisition, processing, and presentation in scientific publications, especially in the neuroimaging community. For example, this need was highlighted at the group discussion of the February 2022 workshop of the North American Imaging in MS Cooperative, as QSM is critical for depicting the paramagnetic rim of chronic active MS lesions ^67–75. In response to this need, QSM investigators in the Electro-Magnetic Tissue Properties Study Group (EMTP SG) of the International Society of Magnetic Resonance in Medicine (ISMRM) established a QSM Consensus Committee to define a consensus approach and determine a series of recommendations.

This paper presents the consensus approach and specific recommendations for implementing QSM in clinical research. The consensus is intended as a guide for researchers with access to technical expertise in MRI acquisition and image processing. The main purpose is to facilitate the use of QSM in patient studies and clinical trials. Guidelines for clinical practice are beyond the current scope.

The QSM Consensus Committee is made up of QSM experts and a community of experienced QSM users and has prioritized its recommendations on 1) clinical applicability with robustness and simplicity, instead of state-of-the-art acquisitions and processing algorithms, 2) description clarity, emphasizing 3 tesla (T), the field strength most widely used in clinical brain research, with only general guidance for other field strengths (1.5T and 7T) where possible, and 3) implementation specificity, focusing on acquisition with multi-echo gradient-recalled echo (GRE), which is very sensitive to the field generated by tissue magnetization gained in MRI in proportion to magnetic susceptibility. Specific issues with no committee consensus (e.g., due to insufficient evidence) are indicated as such in the paper, and these recommendations may provide a starting point for application-specific improvements.

The remainder of this paper is organized into nine sections, covering data acquisition, image processing, image analysis, and presentation of QSM studies in scientific publications (Fig. 1). Data acquisition is split into two sections: Section 2) pulse sequences and protocols, and Section 3) coil combination, saving, and exporting. Image processing is split into four sections corresponding to four processing steps: Section 4) phase unwrapping and echo combination, Section 5) creation of masks, Section 6) background field removal, and Section 7) dipole inversion. Image analysis, Section 8, focuses on the analysis of susceptibility maps. The last section on presentation, Section 9, covers description in scientific publications. Section 10 provides a summary. Each section presents a general overview of the subject matter expounding the consensus approach, and the consensus statement specifying recommendations. Readers interested only in the consensus recommendations may skip the overview and proceed directly to the recommendation.

Figure 1:

The key elements of this consensus paper are described in Sections 2–9. The first two of these sections cover acquisition: 2) pulse sequences and protocol and 3) coil combination, saving, and exporting. The next four sections cover image processing: 4) phase unwrapping & echo combination, 5) creation of masks, 6) background field removal, and 7) dipole inversion. The last two sections cover analysis and presentation in scientific publications: 8) analysis of susceptibility maps, and 9) presentation and publication. The image output from each section is further detailed in the corresponding section.

This paper is accompanied by harmonized pulse sequence protocols for several major platforms (current software versions in 2022) and sample code to perform the recommended processing steps (Supporting Information: the Sepia toolbox ⁷⁶, https://github.com/kschan0214/sepia). The contributions of the members of the QSM Consensus Committee are listed in Supporting Information I, Section S1.1, along with an overview of the history and approach of the initiative (Section S1.3). For many sections, expounded rationale and additional considerations are provided in Supporting Information IV. The version of this manuscript which was endorsed by the EMTP SG, in which additional considerations are integrated into the main text, is archived separately ⁷⁷. The Acknowledgement section lists all individuals who contributed significantly to the consensus recommendations in verbal or written form.

2. Pulse Sequences and Protocol

This section describes recommendations to robustly acquire MRI data for QSM.

2.1. Overview

Most major MRI manufacturers, if not all of them, provide a radiofrequency (RF)-spoiled 3D multi-echo GRE pulse sequence from which one can obtain phase images for QSM in addition to the standard T₂*-weighted magnitude images. The GRE sequence is probably the most elementary sequence in the MRI sequence tree, consisting in its simplest form of an RF excitation pulse followed by acquisition of a gradient recalled echo, which for the multi-echo variant is repeated at various echo times (TEs) before the next excitation pulse is applied. This sequence is often used for calibration steps, such as obtaining a field map for $B_0$ shimming. We recommend the following principles for designing a QSM acquisition protocol based on a RF- and gradient-spoiled 3D multi-echo GRE sequence:

• Aim to set the longest TE (the TE of the last echo) equal to at least the T₂* value of the tissue of interest. The first echo time (TE1) should be as short as possible and the spacing between echoes (ΔTE) should be uniform.

• Use the minimum repetition time (TR) and set the flip angle to the Ernst angle (${\theta }_{\mathrm{E}\mathrm{r}\mathrm{n}\mathrm{s}\mathrm{t}}={\mathrm{c}\mathrm{o}\mathrm{s}}^{-1}\left(e^{-\mathrm{T}\mathrm{R}/{\mathrm{T}}_1}\right))$ for tissue in the target region.

• While QSM can be achieved with one echo, two is the minimum number of echoes needed to separate the intrinsic transmit RF phase from the magnetic field-induced phase (see Section 3 below). Because phase SNR is maximal when TE = T₂*, the use of multiple echoes ensures that high SNR field estimates are obtained for tissues with a range of T₂* values. Acquisition of echoes with TEs extending well beyond the longest tissue T₂* values is generally avoided as their phase and magnitude signals are often compromised by macroscopic field effects from nearby strong susceptibility sources (e.g., large veins).

• Use the minimum readout bandwidth which generates acceptable distortions. At 3T, 220 Hz/pixel is often sufficient (two-pixel fat-water shift). Such acquisitions negate the need to use fat suppression for brain applications.

• Use isotropic voxels of at most 1 mm to reduce partial volume-related estimation errors ⁷⁸.

• Use 3D acquisition instead of 2D acquisition to avoid potential slice-to-slice phase discontinuities in 2D phase maps ⁷⁹.

• Use a monopolar gradient readout (fly-back) to avoid geometric mismatch and eddy current-related phase problems between even and odd echoes in bipolar acquisitions ⁸⁰.

Consider using flow compensation when targeting vessels, but note that flow compensation is often only available and effective for the first echo, while flow artifacts increase in later echoes ⁸¹. More detailed rationale and additional considerations are provided in Supporting Information IV.

2.2. Consensus Recommendations

The recommendations in this section are based on protocols currently available on commercial scanners which do not require a special research agreement with the scanner manufacturer:

• 2.a.Use 3D multi-echo RF- and gradient-spoiled GRE with monopolar readout. Three or more echoes should be acquired and the TE range should include the T₂* times of the target tissues.
• 2.b.Use the minimum available TR (given selected TEs).
• 2.c.Set the excitation flip angle to the Ernst angle for target tissues (e.g., white matter).
• 2.d.Use whole brain coverage.
• 2.e.Use isotropic resolution with a voxel edge length of at most 1 mm non-interpolated at 3T.
• 2.f.Use accelerated imaging methods (e.g., parallel imaging or compressed sensing) available as product sequences.
• 2.g.Use receive coil arrays with a large number of elements (preferably at least 12) covering the whole brain.

Acquisitions for QSM reconstructions can be performed at all clinical field strengths using the protocol recommendations provided here, but higher field strengths provide several benefits. Both intrinsic SNR and susceptibility contrast increase with the static magnetic field, leading to higher CNR, and acquisition can be more time efficient at higher fields due to the shorter T₂*. The recommended protocols implicitly integrate a compensation for lower SNR at lower $B_0$ through reduced spatial resolution, while keeping the total acquisition times similar. While useful data can be obtained at all field strengths, finer anatomical details in certain applications, such as lesions in multiple sclerosis ⁸², might be difficult to visualize at 1.5T.

Supporting Information I, Section S1.4 contains sample protocols for 1.5T, 3T, and 7T.

3. Coil Combination, Saving, and Exporting

This section recommends effective and practical solutions for generating phase images that can be used for QSM and provides guidance on how to save and convert the format of the phase images in preparation for QSM analysis.

3.1. Overview

Modern MRI systems use phased array coils made up typically of 12, 32 or 64 elements, which provide higher SNR than a single birdcage coil and facilitate parallel imaging ^83–85. However, the images from individual coil elements are sensitive to only a part of the FOV and need to be combined to generate a single image with high SNR throughout the brain, a process that requires consideration of the differences between the coil signals.

Neglecting phase wraps (see Section 4), the phase measured with a particular RF coil element c in a GRE sequence, ${\phi }^c,$ can be approximated by: ${\phi }^c={\phi }_{\mathrm{B}}\left(\mathrm{T}\mathrm{E}\right){+\phi }_0^c$. The first part, ${\phi }_{\mathrm{B}}\left(\mathrm{T}\mathrm{E}\right)$, is the phase shift caused by the deviation $\mathrm{\Delta }{B}_0$ of the magnetic field from the uniform main magnetic field, $B_0$. Neglecting second-order non-linear effects⁸⁶, ${\phi }_{\mathrm{B}}\left(\mathrm{T}\mathrm{E}\right)$ evolves linearly over time and is the term relevant for QSM. The second part, ${\phi }_0^c$, is the phase at TE=0, known as the phase offset (or initial phase) of coil element $c$. This contribution comprises effects that are common to all the RF receive coils, such as the phase of the transmit RF field $B_1^{+}$, the effects of tissue conductivity ⁸⁷, gradient delays and eddy current effects, and contributions that are unique to each receive coil, such as the coil sensitivity. Coil-dependent phase offsets must be removed prior to a complex summation of the coil signals, as shown in Figure 2, to avoid destructive interference. Destructive interference leads to reduced SNR and unphysical phase wraps in regions of the image, which cause artifacts in QSM ⁸⁸. Complete destructive interference is often associated with phase wraps referred to as “open-ended fringe lines,” “phase singularities” or “cusp artifacts” (see Figure 3, left). These ill-behaved wraps cannot be fully removed by unwrapping (see Section 4).

Figure 2:

Steps in coil combination, saving, and exporting (illustrated for eight example coils from a 64-channel array and one echo). Each of the coils generates a phase image (left), which is modified by the coil sensitivity and other terms which make up the initial phase. The initial phase is removed (center left) using methods detailed in the text and referenced publications after which phase images are combined in the manufacturer’s reconstruction and saved for export in DICOM format (center right, with multiple echoes illustrated). QSM analysis software may require the DICOM data to be converted, offline, to NIfTI format (right). Images shown were acquired at 3T (Siemens Healthineers, Erlangen, Germany; Prisma Fit, VE11C) with a head/neck 64 channel coil and the recommended multi-echo GRE sequence (TE₁=5.25ms; echo spacing=5.83ms; 5 echoes) using monopolar readout. The imaging data and a description of the acquisition and analysis protocols can be found in Supporting Information II.

Figure 3:

Some scanner manufacturers’ options for processing and saving phase images (like “Sum of Squares”) do not remove coil sensitivities. This becomes apparent in the combined phase images having open-ended fringe lines (left). Wraps in phase images generated using the recommended methods are quite symmetric across the brain mid-line (right), and (like contours on a topographic map) either begin and end at the edge of the brain tissues, or form closed loops within the brain.

All major 3T MR system vendors have effective solutions for generating phase images from RF array coils on their current software platforms. Most of these remove individual coil sensitivities, which are estimated by referencing to a coil with a relatively homogenous sensitivity over the object (usually the body coil). The reference data is acquired in a separate, fast, automated measurement, and the coil sensitivity correction is carried out on complex data either in k-space ⁸³ or image space ⁸⁹ before extraction of the phase. We recommend using the available ‘on-console’ vendor solutions listed in Section 3.2 and provide alternatives for older scanners and high field (≥ 7T).

Phase and magnitude images should be exported in DICOM format. However, most QSM tools require data to be in NIfTI format, and a number of tools are available to perform the necessary conversion. We recommend DCM2NIIX ⁹⁰ (https://github.com/rordenlab/dcm2niix), which is a well-maintained open-source software with compiled versions for all major operating systems. Generally, phase changes related to positive susceptibility sources (those which are paramagnetic relative to water) are defined as being positive, but the possibility exists that vendors may use the opposite sign convention,⁹¹ in which case the sign of all phase values should be reversed.

3.2. Consensus Recommendations

We recommend the following manufacturers’ methods for combining phase images from array coils and saving phase data, listing the software versions for which these solutions have been tested and whether a research agreement is required. Detailed step-by-step descriptions and solutions for higher field scanners and older systems are provided in Supporting Information III.

• 3.a
  The recommended solutions for saving phase images are, for

  • Canon: SPEEDER, a version of SENSE which is available from MPower version 2.3 onwards and allows phase images to be reconstructed through a vendor-provided service password.
  • GE: ASSET, a SENSE-similar solution which reconstructs magnitude, phase, real, and imaginary images without a research key on platforms MR30 onwards.
  • Philips: SENSE, which provides well-combined phase images ⁸⁴ without the need for a research key from software version 5 onwards.
  • Siemens: “Adaptive-combined with prescan normalize” ⁹², which is available from software version VE11 onwards in the product GRE sequence.
  • United Imaging: an inter-coil referencing and weighted correction approach which is available from software version v9 onwards without the need for a research key.
• 3.bExporting data: Data should be exported in (classic) DICOM format.
• 3.cFormat conversion: if the analysis pipeline requires NIfTI data, DICOM data should be converted to NIfTI using DCM2NIIX.

4. Phase Unwrapping and Echo Combination

This section describes the methods used to resolve phase aliasing and calculate a field map from multi-echo GRE data.

4.1. Overview

MRI phase measurements are constrained to an interval of $2\pi$ radians and are, therefore, subject to phase wraps or phase aliasing artifacts, i.e., the measured phase $\phi =\left({\phi }_{\mathrm{B}}\left(\mathrm{T}\mathrm{E}\right)+{\phi }_0\right)$ modulo $2\pi$. Such phase wraps introduce a phase difference of an integer multiple of $2\pi$ between the measured phase, $\phi ,$ and the true phase ${\phi }_{\mathrm{B}}\left(\mathrm{T}\mathrm{E}\right)+{\phi }_0$. Phase wraps are usually visible as discontinuous phase jumps in the phase images (Figs. 3 and 4). To a first-order approximation, ${\phi }_{\mathrm{B}}\left(\mathrm{T}\mathrm{E}\right)=\gamma ·\mathrm{T}\mathrm{E}·\mathrm{\Delta }{B}_0$, where $\mathrm{\gamma }$ is the proton gyromagnetic ratio in rad/T/s. To obtain an accurate estimate of the field shift, $\mathrm{\Delta }{B}_0,$ for QSM, both phase wraps and the phase offset ${\phi }_0$ need to be removed from the measured phase, $\phi$ ⁵⁶.

Figure 4.

A) Example wrapped phase images at different echo times after proper coil combination (same images as shown on the right-hand side of Figure 2). More phase wraps are observed at later echoes (bottom). B) Example unwrapped phase images (using ROMEO template phase unwrapping with MCPC3D-S phase offset correction). C) Frequency map of the total field estimation, i.e., $\frac{\gamma \Delta B_0}{2\pi }$ in the unit of Hz, after echo combination using weighted echo averaging.

Over the years, different phase unwrapping methods have been adapted, refined and applied to MR phase imaging; including time-domain unwrapping methods (with multi-echo acquisition) such as CAMPUS ⁹³ and UMPIRE ⁹⁴, and spatial-domain unwrapping methods such as region-based PRELUDE ⁹⁵ and SEGUE ⁹⁶, path-based PHUN ⁹⁷, SPUN ⁹⁸, BEST-PATH ⁹⁹ and ROMEO ¹⁰⁰, and Laplacian unwrapping ¹⁰¹. The Laplacian unwrapping method is robust and gives wrap-free phase results even under low SNR conditions, but can result in high-frequency errors that propagate into susceptibility maps¹⁰². It is also noted that Laplacian unwrapping only gives an approximation of the underlying unwrapped phase, especially when using the commonly used Fourier-based implementation, while region-based and path-based unwrapping give quantitatively more accurate estimates of the unwrapped phase ^56,100. Region-based and path-based methods are termed “exact unwrapping methods” below ¹⁰⁰. When comparing exact unwrapping methods to Laplacian unwrapping, unwrapping errors (e.g., in veins and hemorrhages) are observed to be smaller in the former, improving QSM quantification accuracy, e.g., for oxygenation estimation ^103,104.

Multi-echo phase images can be combined to achieve a more accurate estimate of the underlying field shift, $\mathrm{\Delta }{B}_0,$ than can be obtained from single-echo phase images ¹⁰⁵. This is because combining multi-echo phase images can remove the phase offset contribution and give higher SNR in the estimated tissue field and susceptibility maps (Fig. 4). The optimum approach may depend on the application, but two echo combination methods have been widely used for QSM – nonlinear complex data fitting ¹⁰⁶ and weighted echo averaging ¹⁰⁷.

The nonlinear complex data fitting approach takes into account the Gaussian noise in the complex images ¹⁰⁶ and estimates the field shift, $\mathrm{\Delta }{B}_0,$ and phase offset, ${\phi }_0,$ together as parameters from fitting the complex MR signal over echo time, with the requirement of having acquired three or more echoes ¹⁰⁶. This approach usually needs spatial phase unwrapping to be performed after the fitting, e.g., on a phase estimate $\gamma ·\delta TE·\Delta B_0$ proportional to the fitted field shift, with $\delta TE$ being the echo spacing, wrapped again between $-\pi$ and $+\pi$. Nonlinear complex data fitting is more robust than linear phase fitting for guarding against phase noise or phase errors due to partial volume effect and intra-voxel phase aliasing at long TEs and around large susceptibility sources, e.g., veins and hemorrhages.

If echoes are acquired over a useful range of TE values (depending on T₂* values of the tissues of interest, see Section 2), the weighted echo averaging approach gives higher SNR for estimating the field shift, $\Delta B_0,$ than the complex data fitting approach^107–109. Unlike nonlinear complex data fitting, this approach needs the phase data at each TE to be spatially unwrapped first. It also requires explicit removal of the phase offset through subtraction of the estimated phase offset (by extrapolating the linear phase evolution to zero echo time) ¹¹⁰ from the phase measured at each TE. Unwrapping errors at longer TEs in voxels with large field shifts and more pronounced noise can be reduced by the “template” unwrapping approach used in ROMEO, which performs path-based spatial unwrapping on an early echo and unwraps other echoes on the basis of an expected linear phase evolution ¹⁰⁰. This, combined with weighted echo averaging, reduces the effect of such errors in the estimated field map.

4.2. Consensus Recommendations

• 4.a.Use an exact phase unwrapping method.
• 4.b.Perform echo combination before background field removal.
• 4.c.The optimal pipeline for phase unwrapping and echo combination depends on the acquisition and application. In most situations, to achieve better SNR, we recommend using weighted echo averaging after template phase unwrapping and explicit phase offset removal. When phase errors persist due to pathology with strong susceptibility sources, we recommend nonlinear complex data fitting followed by spatial phase unwrapping.

5. Creation of masks

This section provides recommendations on creating masks for background field removal (Section 6), dipole inversion (Section 7), and visualization (Section 9).

5.1. Overview

Masking is often overlooked when describing a QSM pipeline, but is a crucial step ¹¹¹, particularly for background field removal (see Section 6). Masking refers to selecting a region of interest (ROI) within the whole field of view and applying a process or function only within this ROI. In QSM, field maps, $\Delta B_0$, are masked primarily because most background field removal algorithms require a mask. By design, background field removal eliminates all sources of magnetic susceptibility outside of the mask, leaving only the local magnetization within the mask. In general, masks should cover the largest ROI possible to prevent exclusion of brain tissue with a sufficient signal-to-noise ratio to have reliable phase/field values.

Masks may be created by using heuristic thresholding operations on available subject images, including magnitude images, T₂* maps, quality maps, or SNR maps. In addition, segmentation algorithms may be based on pre-learned shapes or on the optimization of functionals ^112–118. In particular, the Brain Extraction Tool (BET) ¹¹⁹ from the FMRIB Software Library (FSL) is a widely-used method for brain masking (skull stripping), although it may fail when pathologies or injuries are present ^120,121. BET is a magnitude-image based algorithm that effectively removes non-brain tissues, air, and bone from magnitude images of the head. Using the last-echo magnitude image for BET masking is a robust way to remove regions with signal dropout. ¹²² However, this is undesirable if such regions are of interest, e.g., the cortex right under the skull, which may be compromised or removed with this approach, and thus this approach is not usually recommended. A more balanced approach is to use magnitude images combined across TEs (e.g., using sum of squares or weighted averaging), although using the first-echo magnitude is a good alternative for just the deskulling action. Alternatives to BET include standard template-based brain-extraction ¹²³. Deep learning-based segmentation, a rapidly developing field ^124–127, may also be considered.

QSM is also vulnerable to errors and artifacts arising from unreliable phase data that may not be directly reflected in the corresponding magnitude data. These may be caused by coil combination errors, flow in vessels, and other factors. For this reason, it has been proposed to use phase-based quality maps in addition to magnitude-based de-skulling (i.e, BET masking using the first echo) to achieve a more reliable brain mask ¹²². A straightforward method to obtain a phase quality map is to threshold the inverse of the noise map provided by the complex nonlinear multi-echo fitting algorithm (described in Section 4) at its mean value ^106,122. This thresholding maintains an adequate number of voxels, as it is applied to the entire field of view (FOV), and the distribution of values exhibits bimodal characteristics. This approach effectively distinguishes between reliable and unreliable voxels, serving as a suitable initial approximation. However, modifying the threshold factor (e.g., by increasing its value by 20%) may improve results. To fine-tune the threshold value, it may be needed to inspect both the background field removal and dipole inversion results, as these will more easily reveal noisy voxels and artifacts arising from unmasked unreliable regions. Some exact phase unwrapping algorithms also provide phase-based quality maps ^100,128, which can again be thresholded to identify voxels within the brain with unreliable phase values, and to provide a better estimation of the brain boundary ^111,122,129.

Masking imperfections may result in small exclusion areas (“holes”) in the region of interest that need special attention. It is important to fill all holes, as the background field removal algorithms will suppress any susceptibility sources within these holes. This may lead to the inadvertent removal of pathologies and other clinically-relevant information. These holes can be reintroduced at the dipole inversion step to prevent streaking artifacts. Please refer to the Additional Considerations section for more details on this subject.

Finally, although some recent deep learning-based single-step QSM approaches have shown that explicit masking can be avoided ^130–134, these methods require further study and validation to be considered for clinical applications and are therefore not included in the recommendations in this manuscript.

5.2. Consensus Recommendations

The recommendations below are summarized in Figure 5, and differences between masks (Masks 1–4 in Fig. 5) are highlighted in Figure 6.

Figure 5:

Block diagram of the masking stages. 1) Create an initial mask using FSL BET (Mask 1). 2) Threshold a phase-based quality map to create a mask of reliable phase values (Mask 2). 3) Multiply Mask 1 with Mask 2 and fill in holes for background field removal. 4) Erode by one or two voxels according to the output of the background field removal algorithm (and, optionally, reintroduce holes) for dipole inversion. Use Mask 4 without holes filled in for display and reporting susceptibility values. The magnitude and phase images shown are the same as those in Figure 4.

Figure 6:

Differences between masks in Figure 5. By using the reliable phase (Mask 2), the initial BET mask (Mask 1) can be further improved by removing unreliable phase data near the boundary (red). Mask 3 is used for background field removal. After removal, Mask 3 may need to be further eroded depending on the output of the background field removal algorithms (eroded region shown in blue). This is used for visualization of the results and reporting. Unreliable phase data inside the brain can also be masked out for dipole inversion (holes in green, with the final Mask 4 used for dipole inversion in white).

• 5.a.Create an initial brain mask (Mask 1) by applying a whole-brain segmentation tool (such as BET) to either the combined (sum of squares) or the first echo magnitude image. The goal of this initial mask is to remove air, skull and other tissues, while preserving cortical areas. Further refinement is performed in the following steps.
• 5.b.Create a mask of reliable phase values (Mask 2) by thresholding the phase quality map generated by the multi-echo combination method in Section 4. Multiply Mask 1 by Mask 2.
• 5.c.After multiplication, holes should be filled to obtain the input mask for background field removal algorithms (Mask 3).
• 5.d.If the background field removal method produces a smaller mask, as the reliable ROI, use this for dipole inversion (Mask 4). To avoid streaking artifacts from unreliable phase data within the brain, holes from Mask 2 can be reintroduced (i.e., multiply Mask 4 by Mask 2). For increased accuracy of susceptibility values inside pathological regions of low signal, e.g., hemorrhages and calcifications, mixing data from reconstructions with and without the holes can be performed.
• 5.e.The calculated susceptibility map should be multiplied by the mask used for background field removal (without holes; Mask 3) before display, reporting of susceptibility values, or further analysis. While small holes can be ignored, since their values are extrapolated from neighboring voxels, large holes should be masked out or flagged (e.g., using NaN).

6. Background Field Removal

This section provides recommendations for the background field removal step.

6.1. Overview

In QSM, the background field is defined as the field generated by susceptibility sources outside the brain ⁵⁷, while the field generated by brain sources is the tissue field (often called the local field). The total field map is the sum of these two fields. The susceptibility difference between brain tissue and one such source, air, is approximately 9 ppm ¹³⁵, which is almost two orders of magnitude larger than the naturally occurring brain tissue susceptibility differences. Therefore, background fields can be significantly larger than the tissue field. However, pathologies such as hemorrhages can create a tissue field that is similar in magnitude to the background fields. Removal of the background field from the total field map, $\mathrm{\Delta }{B}_0$, allows the inversion to focus on the effects of spatial susceptibility variations in the brain (Figure 7). When background fields are not completely removed from $\mathrm{\Delta }{B}_0$, most dipole inversion methods will result in shadowing artifacts and/or experience a slow convergence rate.

Figure 7.

Process of background field removal estimates the background field component of the total field (first row; same images as shown on the right-hand side of Figure 4; unit is Hz) relative to a chosen region of interest (brain mask, third row) and subtracts it from the total field, resulting in the tissue field (fourth row; unit is Hz). The tissue field encodes the spatially varying susceptibility within the brain but is much smaller than the background field. This is illustrated by showing cross-sections (indicated by the dotted lines in the field images) in the total field (second row) and the tissue field (last row). The background field was calculated using the V-SHARP method.

Because the background field is spatially smooth, high-pass filtering has been a popular method to suppress it. However, high-pass filtering also removes the low spatial frequency components of the tissue field, leading to reduced accuracy in QSM. Methods that are based on the fact that background fields are harmonic functions have replaced heuristic filtering methods ⁵⁷. Harmonic functions satisfy the Laplace equation within the brain and are completely determined by their values at the brain boundary ^57,136.

The SHARP (Sophisticated Harmonic Artifact Reduction for Phase data) method, ¹³⁷ and variants thereof use the spherical mean value property of harmonic functions. This property states that the average of a harmonic function over a sphere with arbitrary radius centered at any location but still contained within the brain is equal to the value of the harmonic function at that location. In practice, the need for a radius that is several voxels wide to overcome discretization effects leads to an erosion of the brain mask in which the tissue field can be computed. The most common variant of this method is V-SHARP ¹³⁸, which involves multiple partial applications of SHARP with different radii to mitigate this erosion. E-SHARP ¹³⁹ and other variants of SHARP ^140,141 overcome the remaining erosion of one voxel required for V-SHARP. Other variants like HARPERELLA ¹⁴² combine SHARP with phase unwrapping. In general, SHARP-based methods perform less well at the boundary of the region of interest ⁵⁷ and perform implicit low pass filtering due to the regularized deconvolution inherent in SHARP¹⁴³.

The PDF (Projection onto Dipole Fields) method ^144,145 finds a susceptibility distribution outside the brain that mimics the field inside that region. The field generated by those outside sources are approximately orthogonal to those generated by local sources, allowing background field removal to be formulated as a weighted linear least squares problem. Because the orthogonality breaks down at the boundary of the brain, like SHARP, this method performs less well at the boundary ^57,144.

The LBV (Laplacian Boundary Value) method ¹³⁶ assumes that the field at the boundary of the brain is entirely background field and finds a harmonic function that satisfies this boundary condition ¹³⁶. While LBV can perform better than PDF and SHARP in some situations⁵⁷, its performance has been observed to be highly dependent on the mask and on the quality of the field estimates at the mask boundary ¹⁴⁶.

6.2. Consensus Recommendations

• 6.a.Use V-SHARP to achieve good results in many situations, as it is less sensitive to imperfections in brain masking. This comes at a cost of a one-voxel erosion of the brain mask used for dipole inversion (Mask 4 in Fig. 5) at the brain surface and reduced accuracy at the edge of the brain.
• 6.b.When whole brain mapping (including the cortex or veins near the brain’s surface) is desired, use PDF. This method will be slightly more accurate throughout the brain. PDF requires a good brain mask.
• 6.c.Depending on the application, tissue field quality, i.e., the phase SNR especially near the boundary, must be balanced against mask erosion.

7. Dipole Inversion

This section provides recommendations for the field-to-susceptibility inversion step (Fig. 8): the inversion input is the tissue field, $\Delta B_t$ (with background fields removed; see previous section), and output is tissue magnetic susceptibility $\chi \left(\mathit{r}\right)$ (relative to a reference value ^147,148, see Section 8).

Figure 8.

The process of dipole field inversion starts from the tissue field (first row, same images as shown in the bottom row of Figure 7; unit is Hz) and estimates the susceptibility map (second row; unit is ppm).

7.1. Overview

The tissue field at a specific location is a summation of weighted contributions from all surrounding magnetic susceptibility sources, therefore a nonlocal quantity. Mathematically, ^149–152 this summation is a spatial convolution ($*$) of the susceptibility with the unit dipole kernel $d\left(\mathit{r}\right)=\frac{1}{4\pi }\frac{3{cos}^2\theta -1}{{\mathit{r}}^3}$:

$$\Delta B_t\left(\mathit{r}\right)=d\left(\mathit{r}\right)\mathrm{*}\chi \left(\mathit{r}\right).$$ [1]

Spatial deconvolution with the dipole kernel is needed to determine the local tissue susceptibility from the tissue field $.$ However, the dipole kernel value is zero at the cone surface of the magic angle (54.7⁰) relative to the main field and very small close to it, making this deconvolution a poorly conditioned inverse problem ^152–156. The measured tissue field contains deviations from perfect dipole patterns, particularly in regions with low signal-to-noise ratio (SNR). While the regions with large deviations are usually eliminated through masking (see Section 5 above), remaining deviations cause deconvolution errors in the calculated susceptibility, manifesting as streaking and shadowing artifacts ^55,157.

Additional information about the susceptibility map can be incorporated through regularization, which helps to reduce streaking and shadowing artifacts in the solution ^{18,61,109,138,153,158–162}. An optimization approach for incorporating this additional information is Bayesian inference, which can be formulated as the following minimization problem:

$$\chi \left(\mathit{r}\right)={\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{m}\mathrm{i}\mathrm{n}}_{\chi \left(\mathit{r}\right)}{\Vert w\left(\mathit{r}\right)\left(\Delta B_t\left(\mathit{r}\right)-d\left(\mathit{r}\right)\mathrm{*}\chi \left(\mathit{r}\right)\right)\Vert }_2^2+\lambda R\left(\chi \left(\mathit{r}\right)\right).$$ [2]

Here the first term is the data fidelity term, which approximates the noise in the field as Gaussian with weighting $w\left(\mathit{r}\right)$; the second term is regularization with strength $\lambda$ ^109,159. The minimization problem is iteratively solved with the number of iterations determined by the desired convergence. The regularization strength depends on anatomy, susceptibility contrast, and SNR, and should be optimized to balance artifact suppression and image sharpness in each imaging protocol and application, using, e.g., the L-curve method ^163,164, the measurement of image feature sharpness, ¹⁶⁵ or radiological image scoring.

Various regularization strategies have been developed for QSM ^{18,55,63,102,109,138,153,156,158,159,161–163,166–173}, as implemented in QSM software packages including FAst Nonlinear Susceptibility Inversion (FANSI) ¹⁶¹ (https://gitlab.com/cmilovic/FANSI-toolbox), Morphology Enabled Dipole Inversion (MEDI) ^163,174 (http://pre.weill.cornell.edu/mri/pages/qsm.html), and STI Suite ¹⁷⁵ (https://people.eecs.berkeley.edu/~chunlei.liu/software.html). Sparsity regularization is used in the openly accessible source codes for FANSI and MEDI; total variation (TV) as a specific sparsity regularization performed favorably in the two QSM reconstruction challenges ^63,66. Zero-referencing to the cerebrospinal fluid (CSF) may be included as an extra regularization, which provides the CSF-uniformity and associated reduction of streaking and shadowing artifacts, ^147,176 see further discussion in the next section.

QSM can be improved using 1) nonlinear forward signal modeling, particularly in tissue of large susceptibility values, such as hemorrhages or calcifications ^63,66, and 2) deep learning-based reconstruction ^{54,132,177,178}, which may require careful management of generalization issues related to limited training data ^63,66.

Streaking artifacts may arise from strong susceptibility sources near borders and within the brain, where the Gaussian noise model breaks down and field estimation becomes erroneous ¹⁰⁶. These artifacts may be suppressed by masking out or reducing the weight of less trustworthy voxels ¹⁰⁶, improving the brain mask ^111,179,180, preconditioning ^181,182, and using in-painting techniques such as MERIT ¹⁰⁶ and L₁ data fidelity ¹⁸³.

Shadowing artifacts may arise from residual background fields. This shadowing may be reduced by improving background field removal such as harmonic incompatibility removal ^102,184 and by suppressing slowly-varying spatial frequency components through regularization ¹⁸⁵ or preconditioning ^55,181.

Some algorithms do not incorporate spatial constraints for suppressing streaking and shadowing artifacts but explicitly modify the dipole kernel instead ^55,157, for example, thresholded k-space division ^{145,156,169,186–188}. Implicit regularizations based on the number of iterations may work, but these methods have limited denoising capabilities and may be less robust than the sparsity regularization optimization approach ^61,168,189.

The above-mentioned QSM reconstruction challenges ^63,66 include publicly available data that can be used to ensure that custom pipelines achieve results similar to the methods recommended in this consensus.

7.2. Consensus Recommendations

• 7.a.Use an optimization approach for dipole inversion with a sparsity-type regularization that is commonly used in compressed sensing ⁵⁵. Specific sparsity types include L1-norm, total variation, and generalized total variation, which likely provide similar outcomes. Further algorithm developments and evaluations are needed to provide a more specific consensus on the sparsity type.
• 7.b.Use the default sparsity type, regularization strength and number of iterations in a QSM software package, such as the processing pipelines recommended here (Supporting Information II), including FANSI, STI Suite, and MEDI, where these default parameters have been optimized for common brain protocols. If the acquisition protocol recommended here (Supporting Information I, Section S1.4) is substantially altered, researchers should perform an L-curve optimization or other method on at least one typical case with the specific study protocol to fine tune the regularization strength and number of iterations and then fix these parameters for the same protocol.

8. Analysis of Susceptibility Maps

This section provides recommendations for quality control, referencing of susceptibility maps, and the quantification of susceptibility values for clinical research, e.g., when performing group studies. Possible tools for susceptibility quantification are provided in Supporting Information I, Section S1.5. Figure 9 summarizes the recommendations of this section.

Figure 9:

Schematic for susceptibility map analysis in case of a study interested in susceptibility values of the putamen (blue ROI on the susceptibility map, ${\chi }_{ref}$) and globus pallidus (red ROI). Data with streaking artifacts that affect the ROIs need to be excluded (or recalculated when applicable to all study data). ROI generation benefits from the inclusion of susceptibility contrast, e.g., by calculation of hybrid images (blue) or use of T₁-weighted and susceptibility data (green). Susceptibility maps need to be referenced, then regional average susceptibility values (${\chi }_{ROI}$) can be computed from referenced susceptibility maps (${\chi }_{ref}$). The shown susceptibility map without artifacts is the same as the one in Figure 8.

8.1. Overview

8.1.1. Quality control

Only a few tools exist that allow fully automatic QSM calculation and evaluation directly from scanner DICOM data, with all the steps outlined in Sections 2 to 8^111,190. Most QSM applications still require multiple processing steps, which can result in error amplification/propagation or inconsistencies between steps, rendering QSM workflows prone to i) reconstruction artifacts (Table 1) and ii) errors in measurement of region-specific susceptibility values. Particularly, the use of one or several masks to exclude unreliable phase data and for background field removal can result in missing areas in computed susceptibility maps. When voxels in those regions are not properly excluded in subsequent analyses, regional mean values may be biased by these erroneously included zero-valued voxels. This can be resolved by incorporating the eroded background-correction mask (Mask 4 without holes in Fig. 5) in the ROI masks. Deviation from radiological orientation (right-left flip) in the final susceptibility maps can be another potential issue arising from the combination of toolboxes. These flips can be revealed easily by comparing brain features between QSM and the original GRE magnitude images on the scanner console.

Table 1.

Reconstruction artifacts, their possible sources, and strategies to identify, mitigate these artifacts and criteria for excluding data.

Artifact	Streaking and shadowing artifacts	Incorrect susceptibility values	(Regional) strong noise
Typical sources	- large susceptibility differences (air-tissue, calcification, hemorrhage etc.) - poor brain mask - use of late TEs and dynamic field fluctuations (especially at higher field strength) in combination with unsuitable mask - inversion algorithm unsuitable for the data (e.g. presence of strong susceptibility sources) - poor choice of reconstruction parameters (for phase unwrapping, background field removal or inversion algorithm) - incorrect coil combination	- Mismatch between acquired and assumed TE values during the reconstruction, e.g., due to unanticipated acquisition protocol changes	- incorrect coil combination - suboptimal image acquisition (coverage, 2D)
Identification	- manual/visual quality control - automatic detection of outlier regions on phase images or QSM 190 (beware of outliers due to pathology) - automated histogram analysis - use of image quality measures such as the structural similarity index - manual/visual quality control	- outlier detection (based on mean/median ROI values)	(see streaking and shadowing artifacts)
Mitigation	- adjust masking, reconstruction algorithm and parameters (e.g., exclude late TEs from echo combination) - use appropriate coil combination	- verification of imaging parameters from DICOM header (manually or automatic) - pull imaging parameters from data instead of hard coding it in the pipeline	- use appropriate coil combination - use of recommended acquisition protocol
Data exclusion	- exclude if ROI affected and recalculation of the entire study cohort with the adjusted pipeline not possible (to avoid bias)	- exclude if recalculation of the entire study cohort with the corrected pipeline not possible (to avoid bias)	- exclude if recalculation of the entire study cohort with the adjusted pipeline not possible (to avoid bias)

8.1.2. Referencing and choice of reference region

QSM can only assess relative susceptibility differences between tissues, as phase data reflect field distortions caused by spatial susceptibility differences^58,148. To obtain susceptibility values that are comparable between repeated measurements, subjects, and scanners, consistent referencing of susceptibility maps is required. In QSM, internal reference regions are used, and external reference regions generally are not. The ideal choice of a reference region for brain QSM is still under debate¹⁴⁸. Different regions used in the literature come with certain advantages and disadvantages and will lead to different susceptibility values in the resulting susceptibility map.

In the case of widespread pathology such as in multiple sclerosis or Alzheimer’s disease, there might not be an ideal choice of reference region. Large reference regions are generally advantageous over small regions, which are more affected by potential local lesions, reconstruction inhomogeneities and other artifacts (less averaging), which are then propagated to all other regions in the map by the referencing process. This issue reduces statistical power and therefore 3D segmentation of reference regions is advisable to include a greater number of voxels. Consequently, whole-brain referencing (using the largest possible mask) is considered stable and reproducible.

The dependence of white matter apparent susceptibility on the fiber orientation with respect to the main magnetic field^191–195 can be a source of additional variability when using a reference that includes white matter regions. In the case of widespread pathology, where no ideal reference region exists, two reference regions could be used to evaluate whether the choice of reference region affects results. If both choices reveal the same significant findings, these can be assumed with greater confidence to originate from the presence of pathology, instead of being an artifact from susceptibility referencing¹⁹⁶.

For local pathology, the use of contralateral or surrounding tissues as a reference is an effective strategy to avoid introducing artificial susceptibility differences.

Table 2 lists advantages and disadvantages of common reference regions. More details on referencing can be found in dedicated literature^147,148.

Table 2.

Commonly used reference regions in the literature.

Reference region	advantages	disadvantages
cerebrospinal fluid 147,168,216–221	· automatic pipelines available 147 · no orientation dependence · susceptibility of CSF unlikely to be significantly affected by disease	· ventricles can be small in young subjects, resulting in segmentation inaccuracies · partial volume effect because of possibly small ventricles in young subjects or compression of ventricles by pathology · CSF flow artifacts · choroid plexus can affect CSF susceptibility assessment in lateral ventricles
global white matter regions (not restricted to internal capsule) 31,222,223	· large region	· orientation dependence · might be affected by pathology, e.g. demyelination, gliosis, hemorrhage, atrophy
internal capsule 23,224–226		· orientation dependence · might be affected by pathology, e.g. demyelination, gliosis, hemorrhage, atrophy, focal lesions · Relatively small region
whole brain 196,202,202,227	· no extra mask required, brain mask from previous processing steps can be used · intrinsic for some methods · large region	· might be affected by pathology and age (e.g. myelination, global demyelination, gliosis, iron accumulation, hemorrhage) · due to large WM fraction similar limitations as “white matter” above.

8.1.3. Effect of segmentation on susceptibility quantification (iron, white matter changes, lesions, vessels, oxygenation)

An accurate segmentation of ROIs is essential to uncover subtle changes in regional susceptibility values that might indicate pathology, or to establish normative values. Many available automated neuroimaging segmentation tools are optimized for use with T₁-weighted images or require T₁-weighted input data¹⁹⁷. However, when using these methods for the analysis of susceptibility maps, the segmentation and registration accuracy in many structures of interest (e.g., basal ganglia) can depend on T₁ contrast¹⁹⁸, which is also affected by tissue iron^199,200, and the generally low visibility of some deep gray matter regions on T₁-weighted images²⁰¹. Previously, it has been shown that the use of a QSM or hybrid QSM-T₁-weighted contrasts for template generation improves atlas and voxel-based analyses^33,34. Consequently, methods that rely solely on T₁-weighted contrast may be biased and suffer from inaccuracies^33,34. Therefore, using multi-contrast segmentation can be considered the best approach to avoid template bias²⁰². Furthermore, partial volume effects could be corrected by eroding of ROIs⁸¹, only using high susceptibility voxels (in case of positive susceptibility)²⁰³, or using a “partial volume map” for correction²⁰⁴.

8.2. Consensus Recommendations

• 8.a.When ROIs are affected by artifacts, exclude data by automated detection of outliers or outlier regions, use of image quality measures or visual inspection.
• 8.b.Ensure that analysis methods do not include voxels of the susceptibility map with unreliable values, e.g., those that lie outside of the eroded background field removal mask.
• 8.c.Always reference susceptibility maps to an internal reference region before performing further analyses.
• 8.d.When choosing a reference region, consider the study design, influence and possible bias of pathology, and discuss accordingly. For widespread pathology, cross-checking results using two different reference regions can be considered a safe means to exclude bias.
• 8.e.Segment reference regions in 3D.
• 8.f.Always include commonly-used reference regions in the analysis and report mean and standard deviation in these regions.
• 8.g.Consider incorporating QSM contrast in ROI segmentation or ensure that T₁w-based methods are accurate.

9. Presentation and Publication

The purpose of the recommendations in this section is to facilitate the interpretation and replicability of future findings with QSM, future meta-analyses, and the comparison among studies by recommending how findings should be presented.

9.1. Overview

The general recommendation is to report as much information as possible regarding data acquisition (hardware and scan parameters), reconstruction pipeline, analysis procedure and results. Among the information entities relevant for QSM, the Consensus Organization Committee identified those that should always be reported in any manuscript, and assigned a “traffic light ranking” (green, orange, red) to items that were not considered essential, as described in Section 9.3.1, to help identify the information that should be prioritized when space is limited. The last part of this section reviews important aspects, potential limitations and confounds that should be considered when presenting QSM findings in scientific papers.

9.1.1. Acquisition and reconstruction pipeline

Ideally, the acquisition hardware should be described in a single sentence reporting the scanner field strength, model, vendor, software release version, and type of coil used (including the number of channels). The QSM Consensus Organization Committee considered it as essential to indicate acquisition sequence type and several acquisition parameters including the number of echoes, TEs, TR, flip angle, bandwidth, resolution and scan duration.

It is considered essential to describe the toolbox and reconstruction pipeline, and list the algorithms used. The numerical values of parameters used should be listed, even if they were the default parameters.

Tables 3, 4 and 5 provide an overview of recommendations pertaining to the description of acquisition hardware, sequence type/parameters, and reconstruction/analysis pipelines, respectively. A representative paragraph presenting this information in a scientific paper is provided in Supporting Information IV, Section 9.3.2.

Table 3.

Recommendations for reporting of parameters of the acquisition hardware. The reporting of specific items is considered essential if there was unanimous consensus in reporting them. Items that were not considered essential were assigned a “traffic light ranking” (green, orange, red; detailed in Supporting Information IV, Section 9.3.1): it is recommended that items with a green or orange flag are reported.

Item	Notes and examples	Recommended
Field strength	DICOM tag (0018,0087)	essential
Vendor	DICOM tag (0008, 0070)	essential
Scanner model	DICOM tag (0008, 1090)
Software release	DICOM tag (0018, 1020)
Type of coil(s) used, including information on number of channels	e.g. “... a transmitting body-coil and a 64-channel head-and-neck receiving coil”
Gradient system	e.g. “... a gradient system with maximum amplitude = 50 mT/m and slew rate = 200 mT/m/ms”

Table 4.

Recommendations for reporting of parameters of the acquisition sequence. The reporting of specific items is considered essential if there was unanimous consensus in reporting them. Items that were not considered essential were assigned a “traffic light ranking” (green, orange, red; detailed in Supporting Information IV, Section 9.3.1): it is recommended that items with a green or orange flag are reported.

Item	Notes	Recommended
Acquisition sequence type	2D vs 3D; GRE vs EPI etc.;	essential
Acquisition sequence commercial name	e.g. “SWAN”, “MERGE”, “SWIp”…
k-space sampling trajectory scheme	cartesian vs spiral vs radial etc.	essential, if not cartesian
Acquisition orientation	pure axial vs sagittal vs oblique
Number of echoes, TE1:ΔTE:TEmax	e.g. 7 echoes, TE = 5:5:35 ms	essential
TR		essential
FA
Pixel Bandwidth or Receiver Bandwidth [Hz]	DICOM tag (0018, 0095)
Spatial coverage (FOV) and acquisition matrix size		essential
Voxel Size	NOTE: it can be different from “FOV divided by matrix size” because images are often interpolated at reconstruction	essential
Monopolar vs bipolar echoes	Indicate if the sequence produces monopolar of bipolar echoes
Average ± std center frequency [MHz]	In multi-scanner studies, mean ± std center frequency shall be reported for data from each scanner. For example, Siemens “3T” scanners systematically operate at <2.9T. DICOM tag (0018,0084)	essential for multi-scanner studies; unnecessary for single-scanner studies
Flow compensation	Yes / no; if yes, please indicate the compensated echo(s): all vs only the first one; and direction (full, phase)
Acceleration type and factor	Yes / no. If yes: SENSE (or ASSET) vs GRAPPA (or ARC), compressed SENSE, etc.; indicate phase factor and slice factor (if 3D)	essential
Partial Fourier factor	Use should be avoided. If used, indicate partial Fourier factors in phase and slice direction	essential, if used
Partial echo (GE/Philips) aka Asymmetric echo (Siemens) aka Half echo (Hitachi)	Use should be avoided.	essential, if used
Elliptical k-space shutter	Yes / no.
Phase stabilization	Option available only in particular implementations. If the option is available, indicate Yes / no	(essential if used)
Excitation pulse	Fat-sat vs Water-only
Scan duration		essential

Table 5.

Recommendations for reporting of parameters of the reconstruction and analysis pipelines. The reporting of specific items is considered essential if there was unanimous consensus in reporting them. Items that were not considered essential were assigned a “traffic light ranking” (green, orange, red; detailed in Supporting Information IV, Section 9.3.1): it is recommended that items with a green or orange flag are reported.

Item	Notes	Recommended
Toolbox used	Specify toolbox name and version (or download date), e.g. FANSI, STISuite, MEDI, etc.	essential
Algorithms used	For each step of the recon pipeline (phase reconstruction, echo combination, masking, phase unwrapping, background field removal, dipole inversion), please specify the algorithm used. Indicate the numerical values of relevant parameters (even if default values were used), e.g. regularization parameters.	essential, at least for non-default algorithms and parameters
Further processing	If further processing was necessary to make images compatible with image review environments (such as PACS) used in the study, any data manipulation (including geometrical transformations, interpolation, header data changes, etc.) should be reported
Referencing	Magnetic susceptibility values should always be reported in either ppm or ppb (parts-per-billion) and the reference region (see the Section 8) should be explicitly stated, even in the case the adopted method did implicit whole brain referencing. When the reference region used in the study is not the whole-brain mask, its [mean ± std] susceptibility value when referenced to the whole-brain mask should be reported, to enable post-hoc re-referencing for meta-analyses. Generally, it should be discussed in the Discussion section how potential pathological changes within the reference region may have biased the study outcome.	essential
Data inclusion/exclusion criteria	Details on data inclusion/exclusion criteria should be reported. For example: which artifacts were taken into consideration, and which level of artifact severity was considered as a threshold for inclusion/exclusion. The description of this aspect, which is study-specific, can be supported by images with representative cases in the Supporting Information.	essential, in studies where datasets were excluded based on image quality, or when datasets with visible artifacts were deemed acceptable for inclusion
Artifacts and SNR	Relevant artifacts and SNR issues should be described.	essential, at least for studies where systematic artifacts and SNR issues in the regions of interest may influence the interpretation of results.

9.1.2. Displaying figures

When displaying quantitative susceptibility maps and the underlying phase images, avoid using rainbow, jet, or similar nonlinear colormaps, as they create false edges in some areas, conceal existing edges in others, and lack intuitive attribution ^205–207. Using a linear grayscale map maintains consistency with most of the published literature. When using color, unless there is a specific reason to use a different approach, opt for a linear, perceptually uniform colormap. Adjust contrast windowing to prevent saturation of relevant brain areas. For healthy brain QSM, a typical window range is [−0.2, +0.2] ppm. Always report the windowing using an intensity bar or providing information in the figure caption. Display susceptibility values only within the background field removal mask (refer to Section 5) to avoid presenting meaningless results.

9.1.3. Interpretation of results

A potential confound that can affect the extraction of quantitative susceptibility values from MRI phase/frequency data arises from the fact that the apparent field measured in a voxel depends on the subvoxel distribution and visibility of water protons (the sensors of the MRI signal) around susceptibility perturbers (such as iron and myelin). This can cause a phase shift due to biased sampling when sensor and perturber distributions spatially correlate in an anisotropic manner ^208,209. An example is water in and around myelinated fibers, whose anisotropic distribution leads to a fiber orientation-dependent shift in apparent frequency which can exceed 10 ppb ^{191,194,195,208–210}. Furthermore, substantial T₁, MT, or T₂* weighting may differentially affect water visibility in different compartments and render apparent frequency shifts dependent (in a non-linear manner) on TR or TE ^86,211,212. Caution is advised when interpreting QSM values within and around fibers perpendicular to B0, because pathological changes in myelin structure in the absence of changes in myelin content in such fiber bundles may lead to QSM changes without actual changes in tissue susceptibility.

Caution is necessary also with interpretation of QSM near the edge of analyzed regions (see Section 5). A notable example involves consideration of areas near the brain’s surface, where phase data is unreliable (due to, e.g., the prevalence of paramagnetic blood in pial veins), or unavailable (due to the lack of signal in skull), or the tissue phase was partially removed in the background field removal step. Because of this, QSM values in cortical grey matter may be incorrect.

Lastly, when strong regularization or prior information is used in QSM dipole inversion, potential smoothing and resolution loss may occur or new features may be added ^163,213,214 (see Section 7.1). Some anatomical detail, visible in phase or magnitude GRE data, may therefore be lost in the QSM.

9.2. Consensus Recommendations

• 9.a.Always report at least the essential information regarding the acquisition hardware (Table 3), acquisition sequence type and parameters (Table 4), reconstruction pipeline and analysis (Table 5).
• 9.b.Representative susceptibility maps and the underlying background field-corrected phase images should be shown in all articles. Linear gray scale is often seen as best, but is not specifically recommended. The intensity windowing should always be reported using a scale bar or with pertinent information in the figure caption.

To facilitate documenting the reconstruction pipeline, we encourage software developers to output the values of all relevant parameters in Tables 4 and 5 (including default parameters) and provide suggested descriptions of their toolboxes, which users can re-utilize in their publications.

10. Summary and Conclusion

This consensus paper has been developed by the QSM Consensus Paper Committee with consideration of suggestions from the whole QSM research community (see Acknowledgements and Supporting Information I, Section S1.3). The paper provides recommendations for all steps essential in setting up a successful QSM study in a clinical research setting. The recommendations, intended for a robust but not necessarily state-of-the-art QSM, are based on the current state of the field (as of 2023) and should be updated as the QSM field progresses.

In summary, we recommend that data be acquired using a monopolar 3D multi-echo GRE sequence, and that phase images be saved and exported in DICOM format and unwrapped using a quantitative approach. Echoes should be combined before background removal, and a brain mask created using a brain extraction tool with the incorporation of phase-quality based masking. Fields from sources outside the brain should be removed using a SHARP-based or PDF technique and the optimization approach to dipole inversion should be employed with a sparsity type regularization. Susceptibility values should be measured relative to a specified reference, including the common reference region of whole brain as a region of interest in the analysis, and QSM results should be reported with – as a minimum – the acquisition and processing details listed in Tabs. 3–5.

The recommended steps for data acquisition, data preparation and post processing are intended to provide a uniform robust reference starting point for a brain-focused QSM study performed with a clinical scanner. Specialty applications, such as the depiction of small structures might require higher spatial resolution than that recommended here ²¹⁵. In this regard, limitations and further considerations are included in each section, but thorough testing of the processing pipeline is recommended before beginning a large patient study.

We hope that the recommendations here will enable many medical research centers to perform comparable QSM studies on scanners from different vendors, and that the standardized acquisition protocols and the processing pipeline provided along with this article will facilitate these studies (see Supporting Information I, Section S1.4. and Supporting Information II). As more clinical QSM studies are performed, analyzed, and presented in scientific publications, and current and future technical innovations become mature, these QSM recommendations will need to be updated.