https://orcid.org/0009-0000-3873-2301
Figures
Abstract
Background
Studies using quantitative susceptibility mapping (QSM) to investigate thalamic iron levels in people with multiple sclerosis (pwMS) have yielded inconsistent results. It has been speculated that cohort differences are responsible for these inconsistencies, leading to the phenomenological “early-rise late-decline” hypothesis, which posits that cohort age differences explain conflicting thalamic susceptibility findings. In a recent replication study, the authors failed to reproduce elevated thalamic susceptibility in pwMS previously reported by one of the only two QSM-based studies, despite matching cohort characteristics and processing, weakening the support for the phenomenological hypothesis.
Objective
To investigate if the outcome of the recent replication study is robust with respect to different QSM algorithms and analysis methodologies.
Methods
Using the same MRI dataset as the previous replication study, we assessed thalamic susceptibility across 83 pwMS and 44 healthy controls. To comprehensively evaluate methodological variability, we tested combinations of three background field removal (BFR) algorithms using various regularization parameters, four dipole inversion algorithms, three reference regions, and two segmentation methods. Each unique combination of a BFR algorithm (with its specific parameter) and a dipole inversion algorithm constituted a distinct pipeline, yielding a total of 19,558 susceptibility maps across 154 different pipelines.
Results
Thalamic susceptibility was lower in pwMS compared to controls independent of the chosen methodology, with differences in effect sizes primarily driven by the background field removal algorithms and their regularization parameters, reference region, and segmentation method. The impact of dipole inversion algorithms was minimal.
Citation: Salman F, Bergsland N, Dwyer MG, Reeves JA, Ramesh A, Jakimovski D, et al. (2025) Impact of processing and analysis methodology on thalamic susceptibility assessment in multiple sclerosis. PLoS One 20(11): e0332478. https://doi.org/10.1371/journal.pone.0332478
Editor: Md Nasir Uddin, University of Rochester, UNITED STATES OF AMERICA
Received: May 6, 2025; Accepted: August 31, 2025; Published: November 14, 2025
Copyright: © 2025 Salman et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Computer code used for the processing of QSM data in this manuscript has been made available at https://zenodo.org/records/15127577. As for the imaging data, at the time of our IRB approval, we did not explicitly request permission to be able to share data and due to the nature of the high resolution MRI data in our dataset, it is potentially identifying (i.e., reconstruction of facial features from the MRI data can lead to re-identification of individuals). Therefore, the imaging data supporting the findings are available upon reasonable request from the corresponding author or the UB IRB office (ub-irb@buffalo.edu). As per the IRB restrictions, the data are not publicly available due to the above mentioned reason.
Funding: Research reported in this publication was performed in part at the University at Buffalo’s Center for Computational Research (NIH award S10OD024973 and NSF award 1724891), and was supported by the National Institute of Neurological Disorders and Stroke of the National Institutes of Health under Award Number R01NS114227. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. There was no additional external funding received for this study.
Competing interests: R.Z. has received personal compensation from Bristol Myers Squibb, EMD Serono, Sanofi, Mapi Pharma, Sana Biotechnologies and Filterlex for speaking and consultant fees. He received financial support for research activities from Bristol Myers Squibb, EMD Serono, Mapi Pharma and Protembis and Filterlex. M.GD. received personal compensation from Bristol Myers Squibb, Novartis, EMD Serono and Keystone Heart, and financial support for research activities from Bristol Myers Squibb, Novartis, Mapi Pharma, Keystone Heart, Protembis, and V-WAVE Medical. The funder provided support in the form of salaries for authors (M.G.D and R.Z), but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
1. Introduction
Iron is essential for normal brain function [1], supporting oxygen metabolism [2], myelination [3], and neurotransmitter synthesis [4], but its dysregulation contributes to multiple sclerosis (MS) pathogenesis [5–8]. In deep gray matter (DGM) structures, abnormal iron accumulation or depletion may drive neurotoxicity, oxidative stress, and neurodegeneration [4]. Quantitative susceptibility mapping (QSM), a technique exquisitely sensitive to paramagnetic and diamagnetic tissue components [9], has consistently demonstrated elevated iron levels throughout DGM structures in people with MS (pwMS) compared to healthy controls – with the notable exception of the thalamus [10,11].
The thalamus serves as a central hub for cognitive functions [12,13], and its atrophy has been recognized as an early and reliable marker of disease progression in MS [14–17]. As such, assessing thalamic iron levels may provide valuable insights into underlying pathophysiological mechanisms and help improve prognostic assessments. However, findings on thalamic iron levels in MS remain controversial. Most QSM- and R2*-based studies indicated decreased [18–26] or unchanged iron levels in pwMS [27–34]. In contrast, based on systematic reviews of QSM and R2* studies [10,11], only two studies—Rudko et al. [35] and Cobzas et al. [36]—have reported increased thalamic iron levels, highlighting the rarity of this finding. To reconcile these discrepancies, a phenomenological “early-rise late-decline” hypothesis was proposed, suggesting that thalamic iron levels increase early in the disease course in young individuals before decreasing in later stages in older individuals [24]. This hypothesis was supported by reports of increased thalamic susceptibility and R2* in younger pwMS (average ages below 40 years) [35,36] compared to controls, and decreased susceptibility and R2* in older pwMS (above 40 years) [18–26].
Salman et al. [37] recently replicated the 2014 study by Rudko et al., one of the two QSM studies reporting increased thalamic susceptibility in pwMS. Despite closely matching Rudko et al.‘s methodology, the study found decreased (not increased) thalamic susceptibility in pwMS, challenging the ‘early-rise’ hypothesis. Considering the notion that QSM outcomes may depend on methodological differences [38–49], the present study followed up on the replication study [37] with a comprehensive investigation of the robustness of the finding of reduced thalamic susceptibility. This study aimed to both confirm the robustness of the replication study’s findings and elucidate factors contributing to variability in QSM-based iron measurements. Using the same MRI data as the replication study (“younger cohort” therein), we systematically evaluated the impact of choosing different background field removal (BFR) algorithms [50–52], BFR algorithmic regularization parameters, dipole inversion algorithms [53–55], reference regions [35,38], and segmentation techniques [43,56] on thalamic susceptibility measurements.
2. Study design
To address the impractical computational cost of testing all possible combinations of QSM algorithms and analytical methods, we implemented a hierarchical approach. First, we reconstructed magnetic susceptibility maps using the same QSM algorithms as the replication study [37] (BFR = sophisticated harmonic artifact reduction for phase [SHARP] [50] and dipole inversion = morphology enabled dipole inversion [MEDI] [55]) and examined whether alternative analytical approaches—different segmentation methods and referencing regions—influenced group-level susceptibility findings. Subsequently, we investigated the impact of different BFR algorithms, their regularization parameters, and dipole inversion algorithms. To isolate algorithm-based effects, we retained the replication study’s segmentation method (FMRIB’s Integrated Registration and Segmentation Tool [FSL-FIRST] [56]) and referencing region (frontal deep white matter [FDWM]), ensuring any observed discrepancies would stem solely from the BFR and dipole inversion algorithm choices rather than analytical variations. All analyses performed in this study focused exclusively on the thalamus.
3. Methods
3.1. Subjects and MRI data acquisition
We used the exact same data as in the replication study by Salman et al. (Section 2.2 therein) [37]. In brief, the replication study had assembled a cohort that matched that in Rudko et al.’s study [35], referred to as younger cohort therein, using a large institutional database of scans collected in previous IRB-approved studies, with written informed consent obtained from all participants.. The cohort was designed to align with the group-level clinical and demographic characteristics reported in the original study by Rudko et al. All MRI data were collected between 2008-10-10 and 2019-02-07.
The resulting pwMS cohort had an average age (± standard deviation) of 37.4 ± 4.5 years and an Expanded Disability Status Scale (EDSS) of 1.7 (range: 0–6.0). The age- and sex-matched healthy control group (N = 44) had an average age of 36.9 (±4.5) years. Details are summarized in Table 1.
All scans were performed on the same 3T MRI scanner (Signa Excite HD 12.0; General Electric, Milwaukee, WI, USA) using an eight-channel head-and-neck coil, with no hardware or software upgrades over the acquisition period. Raw k-space data for QSM were acquired using an unaccelerated 3D single-echo spoiled gradient-recalled echo (SPGR) sequence with first-order flow compensation in the readout and slice directions. Acquisition parameters included a matrix size of 512 × 192 × 64, nominal resolution of 0.5 × 1 × 2 mm³ (FOV = 256 × 192 × 128 mm³), flip angle = 12°, TE/TR = 22/40 ms, bandwidth = 13.89 kHz, and total scan time of 8 minutes and 46 seconds. Additionally, a 3D T1-weighted (T1w) sequence was acquired using previously described parameters [21].
3.2. Analytical methods
The investigation of analytical methods used susceptibility maps reconstructed using SHARP [50] with a 7 mm kernel radius as the BFR and MEDI [55] as the dipole inversion with N = 7 regularization parameters (SHARP: λS [log-scaled range: 0.0057–0.0191] and MEDI: λD [250–4500]) that passed a visual quality assessment in the replication study [37] (acceptable pipelines therein), excluding pipelines (combinations of SHARP+MEDI parameters) with under- or over-regularized reconstructed maps. Specifically, a two-tiered visual quality assessment was performed: first, BF-corrected field maps were reviewed to identify and exclude BFR parameters yielding under- or over-regularized background field (BF)-corrected field maps; second, the final susceptibility maps were assessed to exclude inversion parameters that produced similarly suboptimal results.
3.2.1. Thalamic segmentation.
We performed thalamic segmentation using FSL FIRST, following the methodology described in the replication study [37]. In addition, we used an advanced bi-parametric (QSM-T1w) multi-atlas segmentation technique optimized for QSM, Bi-parametric Joined Label Fusion (B-JLF; https://gitlab.com/R01NS114227/antsjointlabelfusion_biparametric). B-JLF utilizes nine QSM atlases for extreme ranges in age, volume, and DGM susceptibility to minimize segmentation biases [43]. We applied Advanced Normalization Tools (ANTs) joint label fusion technique using all atlases (N = 9) and propagated thalamic labels to the native subject spaces using subject-specific bi-modal warp field computations. The subject-space labels were then applied to all reconstructed susceptibility maps.
3.2.2. Referencing region.
We first established a baseline by following the replication study’s referencing methodology – referencing bi-lateral susceptibility values from both segmentation methods (Section 3.2.1) to the FDWM region. Subsequently, we repeated the analyses using whole brain (WB) and cerebrospinal fluid (CSF) as alternative reference regions, selected for their frequent use in the literature [39,44], endorsement in recent QSM consensus guidelines [38], and practical advantages—WB offers largest voxel coverage, while CSF being away from the sinus cavity. The same FDWM mask was used as in the replication study (see Referencing in Section 2.4 therein) [37], while the WB and CSF masks were generated as described elsewhere [44]. All referencing masks are displayed in Fig 1a and c (bottom row: FDWM; top row: WB/CSF).
a) T1-weighted (T1w) and quantitative susceptibility mapping (QSM) contrasts of the study-specific template generated using the bi-parametric approach. b) Representative subject’s background field (BF)-corrected field map from one of the SHARP parameters (λS = 0.0128) and c) susceptibility map from one of the SHARP+MEDI pipelines (λS/λD = 0.0128/1000), in native space (41-year-old female, RMS patient [EDSS = 1.5]), with approximately matched slice views to those in (a). Top row: (a) and (c) display the manually delineated cerebrospinal fluid (CSF) region using the T1w contrast of the template (left column in a) and the subject-specific CSF label (right-most in c) obtained from the dual-contrast atlas in (a), respectively. Additionally, the whole brain (WB) mask (middle column in c) was obtained using the QSM pipeline, following the 2024 QSM Consensus Recommendations. Bottom row: (a) and (c) show the manually delineated frontal deep white matter (FDWM) region using the QSM contrast of the template (right column in a) and the subject-specific FDWM label (middle and right-most columns in c) segmented using the dual-contrast atlas in (a), respectively. Bi-lateral thalamic masks from FMRIB’s Integrated Registration and Segmentation Tool (FIRST) and Bi-parametric Joint Label Fusion (B-JLF) (violet) are displayed in the middle and right-most columns of (c), respectively, where mis-segmentation is evident in FIRST compared to B-JLF. The susceptibility template and map contrast range is in parts per million (ppm), while the field map contrast range is in radians (rad).
3.3. Voxel-wise analysis
Voxel-based statistics provide a detailed spatial assessment of susceptibility changes, reducing reliance on segmentation accuracy. To confirm that observed group differences were not biased by mis-segmentation, we conducted voxel-wise statistical analysis.
Each subject’s susceptibility maps from all pipelines were referenced to the FDWM (following the replication study) and spatially registered to the study-specific template. The normalized images were then analyzed using a nonparametric permutation-based analysis, as implemented in the FSL randomise program [57], with 10,000 permutations per analysis and age included as a covariate. The choice of 10,000 permutations was made to ensure the margin-of-error remained below 10% of the nominal alpha [57]. Threshold-free cluster enhancement was employed [58], and the analysis was restricted to the thalamus using a standard-space mask. This approach allowed us to determine group differences at P < 0.05, controlling for family-wise error rate. MRICron (v1.0.20201102) was used for the visualizations.
3.4. Additional QSM algorithms
We processed the data using two additional BFR algorithms alongside the SHARP BFR algorithm, as well as three additional dipole inversion algorithms (excluding MEDI).
We chose the BFR and inversion algorithms based on a recent study in which they demonstrated the highest reproducibility and sensitivity toward susceptibility changes over a period of 10 years across 6 BFR and 21 dipole inversion algorithms [44].
3.4.1. Background field removal.
We applied SHARP using only the regularization parameters (λS) discussed earlier (Section 3.2) and variable radius SHARP (VSHARP [50,51]) with a variable radius of 1:7 mm—within the optimal in-vivo range [59] and matching the SHARP’s fixed kernel to ensure comparability—across the full range of regularization parameters (including non-acceptable; λV) as used for SHARP in the replication study (see Background field removal in Section 2.3 therein) [37], to further facilitate direct comparison. We also applied Regularization Enabled SHARP (RESHARP [52]) with a 3 mm kernel radius, using uniformly distributed 14 regularization parameter (λR) values in log-scale, ranging from 0.0000001–0.1, and added 4th-order 3D polynomial fitting to suppress non-harmonic transceive phase contributions.
The RESHARP parameters were chosen to capture the full spectrum of BF-corrected field map appearances—including under-regularized (residual background fields), well-regularized (visually usable), and over-regularized (excessively smoothed) outputs—as was done for SHARP in the replication study and for VSHARP in this study. Refer to Table 2 for algorithm-specific regularization parameters utilized in this study.
We decided to adopt the same strategy as the replication study, as also described in Section 3.2 herein, limiting our analysis of the additional BFR algorithms to only the visually acceptable maps across all BFR-specific λ parameters (λV and λR). The λ parameters were deemed acceptable after reconstructing all BF-corrected field maps and having two raters (F.Sa. and F.Sc.) visually inspect those from a randomly selected subset of subjects. During the visual inspection, the raters were blinded to the pipeline-specific group differences to ensure unbiased outcomes.
3.4.2. Dipole inversion.
We applied Least Squares QR (LSQR [53]), Homogeneity Enabled Incremental Dipole Inversion (HEIDI [53]), and Approximate Message Passing with Parameter Estimation (AMPPE [54]). For LSQR and HEIDI, default parameters were used. For AMPPE, we used high-order (‘db1’) instead of the default low-order (‘db2’) wavelet bases its developer (personal communication with developer Shuai Huang from Emory University, Atlanta). We did not investigate the impact of different regularization parameters as observations with MEDI indicated that variations in regularization parameters (λD) had a minimal impact on effect sizes [44](Fig 3 in the Replication Study) [37].
Lower values (−1; blue – falsifies the “early rise” hypothesis) represent higher absolute negative effect size and white for effect size of 0. Each vertical panel corresponds to a reference region listed at the top, with the left-most panel using frontal deep white matter (FDWM) referencing as in the replication study, middle using whole brain (WB), and right-most using cerebrospinal fluid (CSF) referencing. Within each panel, each row corresponds to a regularization parameter value used for MEDI inversion algorithm (λD, listed on the left-hand side), while each column represents a regularization parameter value used for SHARP BFR algorithm (λS, listed at the bottom). Significant group differences were marked with asterisks (*).
Subject-specific masks utilized for the above algorithms were generated and applied in compliance with the 2024 QSM Consensus Recommendations [38], as described in the replication study.
3.5. QSM statistical analysis
Thalamic susceptibility values for each subject were calculated using FSLstats [60] with bi-lateral binary masks in native QSM image space, obtained from each segmentation method (only FIRST for additional algorithms in Section 3.4). Subsequently, susceptibility values from each segmentation method were referenced to each of the reference regions (only FDWM for Section 3.4).
Following the replication study’s protocols, we then 1) averaged bi-lateral referenced thalamic susceptibility values, 2) applied age correction, and 3) conducted independent t-tests for group comparisons. Subsequently, we calculated Cohen’s d effect sizes across pipelines to quantify the impact of parameter choices on study outcomes and visualized them using heat maps (Python Seaborn v0.12.2) [61].
Finally, median absolute effect sizes were calculated for all methods (referencing region, segmentation method, BFR algorithms, and inversion algorithms) using only the significant effect sizes (P ≤ 0.05). Following this, the percentage difference of median absolute effect sizes was calculated for all new methods used in this study, relative to the originally used methods in the replication study (BFR = SHARP, inversion = MEDI, segmentation = FIRST, and reference region = FDWM).
3.6. Scientific rigor
All processing and analyses were fully automated to ensure consistency and reproducibility, as described in the replication study. Specifically, the QSM processing was fully automated and executed in a standardized computational environment using containerized computing on a high-performance cluster [62]. Additionally, subject-specific QSM-based montages were generated for inspection, comprising both BF-corrected field maps (without regional overlays) and susceptibility maps (with and without thalamus, FDWM, WB, and CSF overlays). These montages were carefully reviewed by an experienced QSM user (F.Sa., 6.5 years of neuroimaging experience) to confirm the absence of artifacts (e.g., Gibbs ringing, motion, and/or transceive phase inhomogeneities). Scans were excluded if DGM regions were substantially affected by such artifacts. These montages also served to verify accurate placement of segmentation masks on native susceptibility maps.
A total of 19,558 susceptibility maps were reconstructed in this study: 6,223 maps for Section 3.2 (127 subjects × [7 λS × 7 λD]) and 13,335 maps for Section 3.4 (127 subjects x [7 λS + 14 λV + 14 λR] × 3 dipole inversion algorithms). These susceptibility maps underwent 50,673 total analyses: 37,338 in Section 3.2 (6,223 maps × 2 segmentation methods × 3 reference regions) and 13,335 in Section 3.4 (13,335 maps × 1 segmentation method × 1 reference region).
4. Results
4.1. Quality control
Figs 1b and c displays a representative native-space BF-corrected field map and susceptibility map reconstructed using the SHARP (λS = 0.0128) + MEDI pipeline (λS/λD = 0.0128/1000), respectively, along with all corresponding segmented regional masks overlayed on the susceptibility map (middle and right-most columns in Fig 1c) – including the thalamic region of interest (ROI) derived from both FIRST and B-JLF segmentation methods. Quality assessment revealed no artifacts in either the BF-corrected field maps (Fig 1b) or susceptibility maps (left-most column in Fig 1c); therefore, no subjects were excluded from the study.
The middle and right-most columns in Fig 1c demonstrate subject-specific mask placement for all ROIs (thalamus, FDWM, WB, and CSF). However, clear thalamic mis-segmentation was observed from FIRST (bottom middle panel in Fig 1c), extending into neighboring white matter (WM) region, specifically into the internal capsule, as observed previously [40,63]. Conversely, B-JLF did not extend outside the thalamic boundary (bottom right-most panel in Fig 1c). Importantly, none of the subjects’ thalamic masks from either FIRST or B-JLF, nor any of the reference region masks, showed incorrect placement (i.e., not centered on the intended regions or shifted).
4.2. Findings across referencing regions and segmentation methods
Figs 2 and 3 summarize the thalamic effect sizes from each pipeline across different reference regions and the two segmentation methods, respectively. To summarize the impact of reference regions and segmentation methods on effect sizes, Fig 4 illustrates the median percentage difference in absolute effect sizes for (a) WB and CSF referencing relative to FDWM with FIRST segmentation, and (b) B-JLF segmentation with each reference region compared to FIRST segmentation and the corresponding reference region.
4.2.1. Initial thalamic findings in pwMS - replication study [37].
Thalamic susceptibility was consistently lower in patients than in controls, with effect sizes ranging from −0.35 to −0.91 (min and max variation: 117.07% and 160.00% variation with λD = 4500 and 250, respectively, across λS values). These findings are summarized in the left-most panel of Fig 2 (FDWM referencing).
4.2.2. Effect of changing the referencing region.
Changing the reference regions from FDWM to WB and CSF did not affect the overall findings of lower thalamic susceptibility in patients, as shown in the middle and right-most panels of Figs 2 and 3, respectively.
The absolute effect sizes were 23.53% higher (median difference) with WB referencing and 11.76% lower with CSF referencing, compared to FDWM referencing, as shown in Fig 4a. Notably, with all reference regions, the highest absolute effect sizes were observed at the higher end of the acceptable range for λS, with λS = 0.0156 yielding the largest effect sizes, ranging from −0.88 to −0.91 with FDWM, −0.86 with WB, and −0.79 to −0.80 with CSF referencing, across λD values (Fig 2).
4.2.3. Effect of the segmentation method.
As shown in Fig 4b, compared to the FIRST segmentation outcomes, B-JLF segmentation resulted in 23.81% lower effect sizes (median difference) across all reference regions, with the largest difference observed with WB referencing (−45.16%), followed by minor differences with CSF (−13.30%) and FDWM referencing (−9.30%).
While the choice of segmentation method substantially affected the magnitude of effect sizes, it did not alter the qualitative outcomes, i.e., the declining trend remained unaffected. Additionally, similar comparability in effect sizes was observed across reference regions with B-JLF segmentation (Fig 3), as with FIRST.
4.3. Voxel-wise analysis
Fig 5 displays FDWM-referenced voxel-wise analysis from all λS values with λD = 1000 (default for MEDI), focusing on the thalamic region using the B-JLF segmented mask due to its superior segmentation performance compared to FIRST segmentation, as discussed in Section 4.1.
Lower values (−1; blue – falsifies the “early rise” hypothesis) represent higher absolute negative effect size and white for effect size of 0. Each vertical panel corresponds to a reference region listed at the top (frontal deep white matter [FDWM]; whole brain [WB]; cerebrospinal fluid [CSF]). Within each panel, each row corresponds to a regularization parameter value used for MEDI inversion algorithm (λD, listed on the left-hand side), while each column represents a regularization parameter value used for SHARP BFR algorithm (λS, listed at the bottom). Significant findings were marked with asterisks (*).
Median percentage differences are indicated outside the bars, with negative values indicating a decrease in absolute effect sizes and positive values indicating an increase.
Results were corrected for age and are shown at P < 0.05, corrected for family-wise error rate. Areas of significantly lower susceptibility in patients compared with HCs are shown in blue. Darker shades are indicative of smaller P values. The contrast range is in part-per-million (ppm).
Voxel-wise analysis revealed significant clusters of reduced thalamic susceptibility in pwMS compared to controls, with the most prominent differences localized to the lateral thalamic nucleus. All voxel-wise visuals corroborated our B-JLF-specific FDWM-referenced group differences (Fig 3 – left-most panel).
4.4. Effect of BFR and inversion algorithm
4.4.1. Visual inspection of field and susceptibility maps.
Fig 6 illustrates the absence of artifacts in the BF-corrected field maps generated using VSHARP and RESHARP across various λV and λR values, respectively. However, clear signs of under- and over-regularization were observed within the field maps from VSHARP, particularly with λV values of 0.0038 and lower (assuming this pattern persists at lower values), and 0.0287–0.0430. In contrast, RESHARP’s tolerance level led to the results being stagnant around λR = 2.0e-5 and lower, without overly under-regularized solutions.
Green boxes indicate acceptable λ values for both BFR algorithms, respectively. Contrast range is in radians (rad). See Fig 7 for susceptibility maps from AMPPE, LSQR, and HEIDI inversion algorithms across all three BFR algorithms (select regularization parameter).
Visually acceptable appearances of field maps (green box) were observed with λV values between 0.0057 and 0.0191 and λR values of 1.0e-3 and lower, hence, the subsequent analysis was limited to the acceptable parameter ranges.
Fig 7 displays the susceptibility maps reconstructed using inversion algorithms with specific BFR algorithmic λ values. Susceptibility maps were visually comparable across all algorithmic combinations (BFR+inversion), with no artifacts observed in any reconstruction.
Contrast range is in part-per-million (ppm).
4.4.2. No effect on qualitative outcome.
Fig 8 portrays FDWM-referenced FIRST-segmented thalamic effect sizes from all BFR algorithms and three inversion algorithms (AMPPE, HEIDI, and LSQR). The trend of lower susceptibility in pwMS compared to the control group, what was observed previously in the replication study (using SHARP+MEDI) [37], remained unchanged regardless of the choice of the BFR algorithm, its parameter, or the inversion algorithm, with only the magnitude of effect sizes being affected.
Each horizontal panel portrays the results of one specific BFR algorithm that is denoted at the left-hand side of the panel (Sophisticated Harmonic Artifact Reduction for Phase [SHARP]; Variable radius SHARP [VSHARP] and Regularized Enabled SHARP [RESHARP], respectively). Within each panel, each row corresponds to a specific inversion algorithm (listed on the left-hand side; Approximate Message Passing with Parameter Estimation [AMPPE], Least Squares QR [LSQR], Homogeneity Enabled Incremental Dipole Inversion [HEIDI], respectively), while each column represents an acceptable BFR-specific regularization parameter listed at the bottom of the horizontal panel (RESHARP = λR; same for SHARP [λS] and VSHARP [λV]– see top row in Fig 6). Lower values (−1; blue – falsifies our “early rise” hypothesis) represent higher absolute negative effect size and white for effect size of 0.
4.4.3. Variable λ ranges for SHARP-based BFR algorithms with higher effect sizes.
Each inversion algorithm exhibited a similar trend between SHARP (top panel in Fig 8) and VSHARP (middle panel), with λS and λV = 0.156 yielding the highest effect sizes, consistent with the SHARP+MEDI pipelines (Figs 2 and 3; see Section 4.2). As for RESHARP (bottom), the highest effect size was observed at λR = 0.001 (AMPPE = −0.75, LSQR = −0.69, HEIDI = −0.73; P < 0.05).
In summary, effect sizes varied by 162.86% with SHARP (range: −0.35 to −0.92 across all inversion algorithms), 212.90% with VSHARP (−0.31 to −0.97), and 150.00% with RESHARP (−0.30 to −0.75), with the lowest effect size observed with the RESHARP+HEIDI/LSQR pipeline (−0.30; P < 0.05) and the highest with the VSHARP+HEIDI pipeline (−0.97; P < 0.05). Among all inversion algorithms across BFR methods, HEIDI was the most robust inversion method, yielding the highest number of significant effect sizes (N = 23).
4.4.4. Outcomes compared to the originally implemented algorithms.
Fig 9 plots the median percentage difference in absolute effect sizes from algorithms utilized for this sub-analysis (BFR = SHARP, VSHARP, RESHARP – Fig 9a; inversion = AMPPE, HEIDI, and LSQR – Fig 9b) compared to the ones from the replication study (SHARP and MEDI), using the originally used analytical methods (FIRST segmentation and FDWM referencing).
(a) Effect size percentage differences for Variable radius (V) Sophisticated Harmonic Artifact Reduction for Phase (SHARP; VSHARP) and Regularized Enabled SHARP (RESHARP) relative to SHARP, calculated using the median effect sizes across all three additional dipole inversion algorithms (Approximate Message Passing with Parameter Estimation [AMPPE], Homogeneity Enabled Incremental Dipole Inversion [HEIDI], and Least Squares QR [LSQR]) for each background field removal (BFR) algorithm. (b) Effect size percentage differences for AMPPE, HEIDI, and LSQR relative to morphology enabled dipole inversion (MEDI). Since MEDI was only used with SHARP, the median effect size for each additional inversion algorithm was computed across SHARP and compared to SHARP+MEDI. Median percentage differences are indicated inside the bars, with negative values indicating a decrease in absolute effect sizes and positive values indicating an increase.
Focusing on the BFR algorithms, VSHARP portrayed 1.07% lower effect sizes, while RESHARP portrayed 7.73% lower effect sizes compared to SHARP across all inversion algorithms within this sub-analysis. On the other hand, for inversion algorithms with the SHARP BFR algorithm, AMPPE exhibited 6.45% higher effect sizes compared to MEDI, while HEIDI and LSQR showed no differences relative to MEDI (0%).
Notably, the findings of the current sub-analysis align with those of the replication study, where the BFR algorithm’s regularization parameters had a substantially greater impact on effect sizes (160.00% variation) than the choice of inversion algorithm (or its parameter; ≤ 15%).
5. Discussion
5.1. Study outcome
This study builds on Salman et al.’s recent replication work, which tested the “early-rise late-decline” hypothesis and failed to reproduce findings of increased thalamic susceptibility in a relatively young cohort, refuting the “early-rise” aspect. The results of the present study demonstrated remarkable robustness of reduced thalamic susceptibility findings in pwMS compared to healthy subjects across diverse QSM processing and analytical approaches. Consequently, differences in processing methods, their parameters, or analysis routines are unlikely to explain the heterogeneity of thalamic susceptibility findings in the literature.
These findings strengthen the potential of thalamic susceptibility as an imaging biomarker for MS-related neurodegeneration. Given the thalamus’ early involvement in MS and its association with cognitive decline and disability progression [16,64], reliable detection of altered susceptibility could support earlier and more accurate disease staging, monitoring, and treatment response evaluation.
5.2. Consistent trends but variable effect sizes across analytical methods
We observed that the choice of reference region (FDWM to WB=+23.53% and CSF = −11.76% median difference; Fig 4a) and segmentation method (FIRST to B-JLF = −23.81%; Fig 4b) substantially influenced the effect size magnitude.
FIRST’s performance is critically dependent on anatomical contrast [49,65], which is particularly problematic in the thalamus where contrast is suboptimal on T1-weighted images. Recent studies indicated poor QSM-based DGM segmentation from FSL FIRST, which were shown to lead to biased susceptibility findings in pwMS, especially in the thalamus, where atrophy begins from the earliest stages of MS [40,63].
The higher absolute effect sizes observed with FIRST may stem from systematic mis-segmentation, where labels extended into neighboring white matter (WM) regions such as the internal capsule (Fig 1b, right-side section’s bottom middle panel), consistent with previous observations [40]. Notably, both this mis-segmented region (e.g., internal capsule) and the FDWM are known to be affected by MS pathology [38,39]. When used in demyelinating disorders, these compromised regions may artificially elevate thalamic susceptibility measurements in MS patients, consequently amplifying observed group differences and effect sizes [40,63]. This could explain why effect sizes were not only higher with FIRST segmentation but also with FDWM referencing compared to WB and CSF referencing.
5.3. Consistent qualitative study outcomes across QSM processing
Contrary to the notion that QSM-based qualitative study outcomes (such as significantly decreased thalamic susceptibility in pwMS) are dependent on parameter settings and algorithmic choices [41–43,48], thalamic outcomes (Figs 2, 3 and 8) were highly consistent across regularization parameters of the different BFR algorithms and dipole inversion algorithm in the present study. Moreover, thalamic outcomes remained robust even across dipole inversion algorithms.
5.4. BFR algorithm is a major source of effect size variation
Our findings indicate that BFR regularization parameters contributed to the highest variation in effect sizes, with a minimum variation of 102.56% across λS values (λD = 4500 FIRST/B-JLF + CSF; Figs 2 and 3 [significant findings]). This surpassed the variation observed due to the dipole inversion parameters (MEDI), which reached a maximum of 20.00% across λD values (λS = 0.0085 with FIRST+WB; Fig 2), segmentation methods, which showed a maximum variation of 56.96% when comparing FIRST to B-JLF at λS/λD = 0.191/250 pipeline (WB in Figs 2 and 3), and reference regions, where the largest variation was 29.55% between FDWM and WB at λS/λD = 0.128/1000 pipeline (B-JLF; Fig 3).
Moreover, in line with the replication study [37] and another recent study [44], effect sizes—likely influenced by differences in variance [45]—were primarily affected by the BFR algorithms (SHARP to VSHARP: −1.07%; RESHARP: −7.73%; Figs 8 and 9a) rather than the dipole inversion algorithms (MEDI to AMPPE: + 6.45%; HEIDI and LSQR: 0% median difference; Fig 9b). In summary, the BFR algorithms and their regularization parameters play the most critical role in group difference detection. Consequently, greater emphasis should be placed on optimizing BFR algorithms and their parameters over dipole inversion algorithms [46,47].
5.5. Toward QSM pipeline standardization
While our findings emphasize the substantial impact of QSM pipeline choices on susceptibility values, full standardization remains challenging. The main concern is not variation in effect size or noise, but rather discrepancies in absolute susceptibility values across pipelines, which undermine reproducibility and cross-study comparability. Based on our results and prior studies [38,44,66], we recommend using SHARP-based BFR methods—especially RESHARP—paired with a stable dipole inversion method such as HEIDI. Parameter selection should prioritize visually acceptable map quality, especially in settings where full numerical optimization of effect sizes is impractical due to time or computational constraints. Although we do not recommend a single standardized pipeline, clear reporting of all methodological components and transparent rationale for their selection are essential steps toward harmonization. Long-term efforts should focus on minimizing cross-study differences in absolute values (excluding acquisition-related noise), potentially through consensus protocols or phantom-based calibration.
5.6. Limitations
Although we did not explicitly investigate thalamic sub-nuclei, our additional voxel-wise analysis revealed susceptibility changes primarily in non-pulvinar regions (Fig 5), partially complementary to previous reports by Rudko et al. and Cobzas et al., which demonstrated R2*-based voxel-wise differences in the pulvinar, but not in lateral regions. These observations contribute to a nuanced understanding of the heterogeneity of intra-thalamic iron changes in pwMS [24,28,67]. It is possible that differences in nucleus-specific atrophy rate between study cohorts may have contributed to differences in QSM findings between cohorts [41,42,68–71]. Furthermore, future research should assess whether different QSM reconstruction and analytical methods yield consistent results in other brain regions.
While the use of a single-echo acquisition—as opposed to the multi-echo approach used by Rudko et al.—may result in susceptibility underestimation due to increased sensitivity to noise and artifacts [72,73], this bias is likely systematic and consistent across groups, as supported by Rua et al.’s supplementary materials (mmc1.xlsx) [73]. As such, it would primarily affect the absolute susceptibility values at the individual level without altering group differences or their direction. Nonetheless, we acknowledge this as a potential limitation and encourage further investigation.
We cannot exclude the possibility that our additional QSM reconstruction-based outcomes (Section 4.4) may have been influenced by not employing different regularization parameters for additional dipole inversion algorithms (Section 3.4.2). Although we observed minimal variation in effect sizes across MEDI parameters (maximum variation: 20%), each algorithm has unique nuances that could impact susceptibility estimates. Therefore, further investigation is warranted to determine whether different regularization choices for other dipole inversion algorithms might yield varying outcomes.
6. Conclusion
Despite employing alternative QSM reconstruction and analytical methods, we consistently observed reduced thalamic susceptibility in young pwMS. Our results highlight the critical role of BFR algorithms and their parameters in obtaining high effect sizes, underscoring the need for their careful optimization in QSM studies. Furthermore, our study suggests that QSM is a highly robust and reproducible technique for clinical research, supporting the comparability of published studies conducted with different reconstruction and analysis methodologies. This reinforces QSM’s potential as a reliable quantitative technique for investigating tissue susceptibility changes in neurological disorders.
Acknowledgments
We are grateful to Dr. Hongfu Sun (University of Newcastle, Australia) for suggesting the inclusion of 3D polynomial fitting for the field maps from RESHARP, and Shuai Huang (Emory University, Atlanta; personal communication) for support with the AMPPE implementation and optimization.
References
- 1. McCann S, Perapoch Amadó M, Moore SE. The Role of Iron in Brain Development: A Systematic Review. Nutrients. 2020;12(7):2001. pmid:32635675
- 2. Roemhild K, von Maltzahn F, Weiskirchen R, Knüchel R, von Stillfried S, Lammers T. Iron metabolism: pathophysiology and pharmacology. Trends Pharmacol Sci. 2021;42(8):640–56. pmid:34090703
- 3. Connor JR, Menzies SL. Relationship of iron to oligondendrocytes and myelination. Glia. 1996;17(2):83–93.
- 4. Ward RJ, Zucca FA, Duyn JH, Crichton RR, Zecca L. The role of iron in brain ageing and neurodegenerative disorders. The Lancet Neurology. 2014;13(10):1045–60.
- 5. Williams R, Buchheit CL, Berman NEJ, LeVine SM. Pathogenic implications of iron accumulation in multiple sclerosis. J Neurochem. 2012;120(1):7–25. pmid:22004421
- 6. Tang C, Yang J, Zhu C, Ding Y, Yang S, Xu B, et al. Iron metabolism disorder and multiple sclerosis: a comprehensive analysis. Front Immunol. 2024;15:1376838. pmid:38590521
- 7. Nnah IC, Wessling-Resnick M. Brain Iron Homeostasis: A Focus on Microglial Iron. Pharmaceuticals (Basel). 2018;11(4):129. pmid:30477086
- 8. Hare D, Ayton S, Bush A, Lei P. A delicate balance: Iron metabolism and diseases of the brain. Front Aging Neurosci. 2013;5:34. pmid:23874300
- 9. Liu C, Wei H, Gong N-J, Cronin M, Dibb R, Decker K. Quantitative Susceptibility Mapping: Contrast Mechanisms and Clinical Applications. Tomography. 2015;1(1):3–17. pmid:26844301
- 10. Voon CC, Wiltgen T, Wiestler B, Schlaeger S, Mühlau M. Quantitative susceptibility mapping in multiple sclerosis: A systematic review and meta-analysis. Neuroimage Clin. 2024;42:103598. pmid:38582068
- 11. De Lury AD, Bisulca JA, Lee JS, Altaf MD, Coyle PK, Duong TQ. Magnetic resonance imaging detection of deep gray matter iron deposition in multiple sclerosis: A systematic review. J Neurol Sci. 2023;453:120816. pmid:37827008
- 12. Hwang K, Bertolero MA, Liu WB, D’Esposito M. The Human Thalamus Is an Integrative Hub for Functional Brain Networks. J Neurosci. 2017;37(23):5594–607. pmid:28450543
- 13. Fama R, Sullivan EV. Thalamic structures and associated cognitive functions: Relations with age and aging. Neurosci Biobehav Rev. 2015;54:29–37. pmid:25862940
- 14. Hänninen K, Viitala M, Paavilainen T, Karhu JO, Rinne J, Koikkalainen J, et al. Thalamic Atrophy Predicts 5-Year Disability Progression in Multiple Sclerosis. Front Neurol. 2020;11:606. pmid:32760339
- 15. Azevedo CJ, Cen SY, Khadka S, Liu S, Kornak J, Shi Y, et al. Thalamic atrophy in multiple sclerosis: A magnetic resonance imaging marker of neurodegeneration throughout disease. Ann Neurol. 2018;83(2):223–34. pmid:29328531
- 16. Amin M, Scullin K, Nakamura K, Ontaneda D, Galioto R. Cognitive profiles associated with faster thalamic atrophy in multiple sclerosis. Mult Scler Relat Disord. 2025;103:106662. pmid:40753887
- 17. Zivadinov R, Bergsland N, Dolezal O, Hussein S, Seidl Z, Dwyer MG, et al. Evolution of cortical and thalamus atrophy and disability progression in early relapsing-remitting MS during 5 years. AJNR Am J Neuroradiol. 2013;34(10):1931–9. pmid:23578679
- 18. Bergsland N, Schweser F, Dwyer MG, Weinstock‐Guttman B, Benedict RHB, Zivadinov R. Thalamic white matter in multiple sclerosis: A combined diffusion‐tensor imaging and quantitative susceptibility mapping study. Human Brain Mapping. 2018;39(10):4007–17.
- 19. Burgetova A, Dusek P, Vaneckova M, Horakova D, Langkammer C, Krasensky J, et al. Thalamic Iron Differentiates Primary-Progressive and Relapsing-Remitting Multiple Sclerosis. AJNR Am J Neuroradiol. 2017;38(6):1079–86.
- 20. Hagemeier J, Ramanathan M, Schweser F, Dwyer MG, Lin F, Bergsland N, et al. Iron-related gene variants and brain iron in multiple sclerosis and healthy individuals. NeuroImage: Clinical. 2018;17:530–40.
- 21. Hagemeier J, Zivadinov R, Dwyer MG, et al. Changes of deep gray matter magnetic susceptibility over 2 years in multiple sclerosis and healthy control brain. NeuroImage Clin. 2018;18:1007–16.
- 22. Pontillo G, Cocozza S, Lanzillo R, Russo C, Stasi MD, Paolella C, et al. Determinants of Deep Gray Matter Atrophy in Multiple Sclerosis: A Multimodal MRI Study. AJNR Am J Neuroradiol. 2019;40(1):99–106. pmid:30573464
- 23. Pudlac A, Burgetova A, Dusek P, Nytrova P, Vaneckova M, Horakova D, et al. Deep Gray Matter Iron Content in Neuromyelitis Optica and Multiple Sclerosis. Biomed Res Int. 2020;2020:6492786. pmid:32509866
- 24. Schweser F, Raffaini Duarte Martins AL, Hagemeier J, Lin F, Hanspach J, Weinstock-Guttman B, et al. Mapping of thalamic magnetic susceptibility in multiple sclerosis indicates decreasing iron with disease duration: A proposed mechanistic relationship between inflammation and oligodendrocyte vitality. Neuroimage. 2018;167:438–52. pmid:29097315
- 25. Zivadinov R, Tavazzi E, Bergsland N, Hagemeier J, Lin F, Dwyer MG, et al. Brain Iron at Quantitative MRI Is Associated with Disability in Multiple Sclerosis. Radiology. 2018;289(2):487–96. pmid:30015589
- 26. Schweser F, Hagemeier J, Dwyer MG, Bergsland N, Hametner S, Weinstock-Guttman B, et al. Decreasing brain iron in multiple sclerosis: The difference between concentration and content in iron MRI. Hum Brain Mapp. 2021;42(5):1463–74. pmid:33378095
- 27. Langkammer C, Liu T, Khalil M, Enzinger C, Jehna M, Fuchs S, et al. Quantitative susceptibility mapping in multiple sclerosis. Radiology. 2013;267(2):551–9. pmid:23315661
- 28. Elkady AM, Cobzas D, Sun H, Blevins G, Wilman AH. Discriminative analysis of regional evolution of iron and myelin/calcium in deep gray matter of multiple sclerosis and healthy subjects. J Magn Reson Imaging. 2018;:10.1002/jmri.26004. pmid:29537720
- 29. Al-Radaideh AM, Wharton SJ, Lim S-Y, Tench CR, Morgan PS, Bowtell RW, et al. Increased iron accumulation occurs in the earliest stages of demyelinating disease: an ultra-high field susceptibility mapping study in Clinically Isolated Syndrome. Mult Scler. 2013;19(7):896–903. pmid:23139386
- 30. Burgetova A, Dusek P, Uher T, Vaneckova M, Vejrazka M, Burgetova R, et al. Oxidative Stress Markers in Cerebrospinal Fluid of Newly Diagnosed Multiple Sclerosis Patients and Their Link to Iron Deposition and Atrophy. Diagnostics (Basel). 2022;12(6):1365. pmid:35741175
- 31. Elkady AM, Cobzas D, Sun H, Blevins G, Wilman AH. Progressive iron accumulation across multiple sclerosis phenotypes revealed by sparse classification of deep gray matter. J Magn Reson Imaging. 2017;46(5):1464–73. pmid:28301067
- 32. Elkady AM, Cobzas D, Sun H, Seres P, Blevins G, Wilman AH. Five year iron changes in relapsing-remitting multiple sclerosis deep gray matter compared to healthy controls. Mult Scler Relat Disord. 2019;33:107–15. pmid:31181540
- 33. Fujiwara E, Kmech JA, Cobzas D, Sun H, Seres P, Blevins G, et al. Cognitive Implications of Deep Gray Matter Iron in Multiple Sclerosis. AJNR Am J Neuroradiol. 2017;38(5):942–8.
- 34. Cho J, Nguyen TD, Huang W, Sweeney EM, Luo X, Kovanlikaya I, et al. Brain oxygen extraction fraction mapping in patients with multiple sclerosis. J Cereb Blood Flow Metab. 2021;42(2):338–48.
- 35.
Rudko DA, Solovey I, Gati JS, Kremenchutzky M, Menon RS. Multiple sclerosis: improved identification of disease-relevant changes in gray and white matter by using susceptibility-based MR imaging. 2014;272(3).
- 36. Cobzas D, Sun H, Walsh AJ, Lebel RM, Blevins G, Wilman AH. Subcortical gray matter segmentation and voxel-based analysis using transverse relaxation and quantitative susceptibility mapping with application to multiple sclerosis. J Magn Reson Imaging. 2015;42(6):1601–10. pmid:25980643
- 37. Salman F, Bergsland N, Dwyer MG, Reeves JA, Ramesh A, Jakimovski D, et al. Thalamic iron in multiple sclerosis: Waning support for the early-rise late-decline hypothesis. Neuroimage Clin. 2025;46:103771. pmid:40187193
- 38. Bilgic B, Costagli M, Chan KS. Recommended implementation of quantitative susceptibility mapping for clinical research in the brain: a consensus of the ISMRM Electro-Magnetic Tissue Properties Study Group. ArXiv. 2023.
- 39. Straub S, Schneider TM, Emmerich J, Freitag MT, Ziener CH, Schlemmer H-P, et al. Suitable reference tissues for quantitative susceptibility mapping of the brain. Magn Reson Med. 2017;78(1):204–14. pmid:27529579
- 40. Lyman C, Lee D, Ferrari H, Fuchs TA, Bergsland N, Jakimovski D, et al. MRI‐based thalamic volumetry in multiple sclerosis using FSL‐FIRST: Systematic assessment of common error modes. Journal of Neuroimaging. 2021;32(2):245–52.
- 41. Magon S, Chakravarty MM, Amann M, Weier K, Naegelin Y, Andelova M, et al. Label‐fusion‐segmentation and deformation‐based shape analysis of deep gray matter in multiple sclerosis: The impact of thalamic subnuclei on disability. Human Brain Mapping. 2014;35(8):4193–203.
- 42. Blyau S, Koubiyr I, Saranathan M, Coupé P, Deloire M, Charré-Morin J, et al. Differential vulnerability of thalamic nuclei in multiple sclerosis. Mult Scler. 2022;29(2):295–300.
- 43. Salman F, Thomas K, Adegbemigun A. Bi-parametric joint label fusion a comprehensive segmentation tool for deep gray matter in QSM. EMTP QMR. 2024.
- 44. Salman F, Ramesh A, Jochmann T. Sensitivity of quantitative susceptibility mapping for clinical research in deep gray matter. Hum Brain Mapp. 2025.
- 45. Salman F, Ramesh A, Jochmann T. Impact of regularization parameter choice on real-world sensitivity of QSM. QMR EMTP. 2024.
- 46. Langkammer C, Schweser F, Shmueli K, Kames C, Li X, Guo L, et al. Quantitative susceptibility mapping: Report from the 2016 reconstruction challenge. Magn Reson Med. 2018;79(3):1661–73. pmid:28762243
- 47. QSM Challenge 2.0 Organization Committee, Bilgic B, Langkammer C, Marques JP, Meineke J, Milovic C, et al. QSM reconstruction challenge 2.0: Design and report of results. Magn Reson Med. 2021;86(3):1241–55. pmid:33783037
- 48. Feng X, Deistung A, Dwyer MG, Hagemeier J, Polak P, Lebenberg J, et al. An improved FSL-FIRST pipeline for subcortical gray matter segmentation to study abnormal brain anatomy using quantitative susceptibility mapping (QSM). Magn Reson Imaging. 2017;39:110–22. pmid:28188873
- 49. Li X, Chen L, Kutten K, Ceritoglu C, Li Y, Kang N, et al. Multi-atlas tool for automated segmentation of brain gray matter nuclei and quantification of their magnetic susceptibility. Neuroimage. 2019;191:337–49. pmid:30738207
- 50. Schweser F, Deistung A, Lehr BW, Reichenbach JR. Quantitative imaging of intrinsic magnetic tissue properties using MRI signal phase: an approach to in vivo brain iron metabolism? Neuroimage. 2011;54(4):2789–807. pmid:21040794
- 51. Li W, Wu B, Liu C. Quantitative susceptibility mapping of human brain reflects spatial variation in tissue composition. Neuroimage. 2011;55(4):1645–56. pmid:21224002
- 52. Sun H, Wilman AH. Background field removal using spherical mean value filtering and Tikhonov regularization. Magn Reson Med. 2014;71(3):1151–7. pmid:23666788
- 53. Schweser F, Sommer K, Deistung A, Reichenbach JR. Quantitative susceptibility mapping for investigating subtle susceptibility variations in the human brain. Neuroimage. 2012;62(3):2083–100. pmid:22659482
- 54.
Huang S, Lah JJ, Allen JW, Qiu D. Robust Quantitative Susceptibility Mapping via Approximate Message Passing with Parameter Estimation. Published online May 30, 2023. [Accessed February 7, 2024]. http://arxiv.org/abs/2207.14709
- 55. Liu J, Liu T, de Rochefort L, Ledoux J, Khalidov I, Chen W, et al. Morphology enabled dipole inversion for quantitative susceptibility mapping using structural consistency between the magnitude image and the susceptibility map. Neuroimage. 2012;59(3):2560–8. pmid:21925276
- 56. Patenaude B, Smith SM, Kennedy DN, Jenkinson M. A Bayesian model of shape and appearance for subcortical brain segmentation. Neuroimage. 2011;56(3):907–22. pmid:21352927
- 57. Winkler AM, Ridgway GR, Webster MA, Smith SM, Nichols TE. Permutation inference for the general linear model. Neuroimage. 2014;92(100):381–97. pmid:24530839
- 58. Smith SM, Nichols TE. Threshold-free cluster enhancement: addressing problems of smoothing, threshold dependence and localisation in cluster inference. Neuroimage. 2009;44(1):83–98. pmid:18501637
- 59. Özbay PS, Deistung A, Feng X, Nanz D, Reichenbach JR, Schweser F. A comprehensive numerical analysis of background phase correction with V-SHARP. NMR Biomed. 2017;30(4):10.1002/nbm.3550. pmid:27259117
- 60. Jenkinson M, Beckmann CF, Behrens TEJ, Woolrich MW, Smith SM. FSL. Neuroimage. 2012;62(2):782–90. pmid:21979382
- 61. Waskom M. seaborn: statistical data visualization. JOSS. 2021;6(60):3021.
- 62.
Center for Computational Research, University at Buffalo, http://hdl.handle.net/10477/79221
- 63. Burggraaff J, Liu Y, Prieto JC, Simoes J, de Sitter A, Ruggieri S, et al. Manual and automated tissue segmentation confirm the impact of thalamus atrophy on cognition in multiple sclerosis: A multicenter study. Neuroimage Clin. 2021;29:102549. pmid:33401136
- 64. Mahajan KR, Nakamura K, Cohen JA, Trapp BD, Ontaneda D. Intrinsic and Extrinsic Mechanisms of Thalamic Pathology in Multiple Sclerosis. Ann Neurol. 2020;88(1):81–92. pmid:32286701
- 65. Derakhshan M, Caramanos Z, Giacomini PS, Narayanan S, Maranzano J, Francis SJ, et al. Evaluation of automated techniques for the quantification of grey matter atrophy in patients with multiple sclerosis. Neuroimage. 2010;52(4):1261–7. pmid:20483380
- 66. Milovic C, Fuchs PS, Arsenov O. Exhaustive Comparison of QSM Background Field Removal and Masking Using a Realistic Numerical Head Phantom. ISMRM. (4178).
- 67. Cagol A, Ocampo-Pineda M, Lu P-J, Weigel M, Barakovic M, Melie-Garcia L, et al. Advanced Quantitative MRI Unveils Microstructural Thalamic Changes Reflecting Disease Progression in Multiple Sclerosis. Neurol Neuroimmunol Neuroinflamm. 2024;11(6):e200299. pmid:39270143
- 68. Krijnen EA, Salim Karam E, Russo AW, Lee H, Chiang FL, Schoonheim MM, et al. Intrinsic and extrinsic contributors to subregional thalamic volume loss in multiple sclerosis. Ann Clin Transl Neurol. 2024;11(6):1405–19. pmid:38725151
- 69. Koubiyr I, Yamamoto T, Blyau S, Kamroui RA, Mansencal B, Planche V, et al. Vulnerability of Thalamic Nuclei at CSF Interface During the Entire Course of Multiple Sclerosis. Neurol Neuroimmunol Neuroinflamm. 2024;11(3):e200222. pmid:38635941
- 70. Levy S, Sandry J, Beck ES, Brandstadter R, Katz Sand I, Sumowski JF. Pattern of thalamic nuclei atrophy in early relapse-onset multiple sclerosis. Mult Scler Relat Disord. 2022;67:104083. pmid:35940020
- 71. Trufanov A, Bisaga G, Skulyabin D, Temniy A, Poplyak M, Chakchir O, et al. Thalamic nuclei degeneration in multiple sclerosis. J Clin Neurosci. 2021;89:375–80. pmid:34090763
- 72. Gilbert G, Savard G, Bard C, Beaudoin G. Quantitative comparison between a multiecho sequence and a single-echo sequence for susceptibility-weighted phase imaging. Magn Reson Imaging. 2012;30(5):722–30. pmid:22459441
- 73. Rua C, Clarke WT, Driver ID, Mougin O, Morgan AT, Clare S, et al. Multi-centre, multi-vendor reproducibility of 7T QSM and R2* in the human brain: Results from the UK7T study. Neuroimage. 2020;223:117358. pmid:32916289