Quantitative magnetization transfer imaging in relapsing-remitting multiple sclerosis: a systematic review and meta-analysis

Abstract Myelin-sensitive MRI such as magnetization transfer imaging has been widely used in multiple sclerosis. The influence of methodology and differences in disease subtype on imaging findings is, however, not well established. Here, we systematically review magnetization transfer brain imaging findings in relapsing-remitting multiple sclerosis. We examine how methodological differences, disease effects and their interaction influence magnetization transfer imaging measures. Articles published before 06/01/2021 were retrieved from online databases (PubMed, EMBASE and Web of Science) with search terms including ‘magnetization transfer’ and ‘brain’ for systematic review, according to a pre-defined protocol. Only studies that used human in vivo quantitative magnetization transfer imaging in adults with relapsing-remitting multiple sclerosis (with or without healthy controls) were included. Additional data from relapsing-remitting multiple sclerosis subjects acquired in other studies comprising mixed disease subtypes were included in meta-analyses. Data including sample size, MRI acquisition protocol parameters, treatments and clinical findings were extracted and qualitatively synthesized. Where possible, effect sizes were calculated for meta-analyses to determine magnetization transfer (i) differences between patients and healthy controls; (ii) longitudinal change and (iii) relationships with clinical disability in relapsing-remitting multiple sclerosis. Eighty-six studies met inclusion criteria. MRI acquisition parameters varied widely, and were also underreported. The majority of studies examined the magnetization transfer ratio in white matter, but magnetization transfer metrics, brain regions examined and results were heterogeneous. The analysis demonstrated a risk of bias due to selective reporting and small sample sizes. The pooled random-effects meta-analysis across all brain compartments revealed magnetization transfer ratio was 1.17 per cent units (95% CI −1.42 to −0.91) lower in relapsing-remitting multiple sclerosis than healthy controls (z-value: −8.99, P < 0.001, 46 studies). Linear mixed-model analysis did not show a significant longitudinal change in magnetization transfer ratio across all brain regions [β = 0.12 (−0.56 to 0.80), t-value = 0.35, P = 0.724, 14 studies] or normal-appearing white matter alone [β = 0.037 (−0.14 to 0.22), t-value = 0.41, P = 0.68, eight studies]. There was a significant negative association between the magnetization transfer ratio and clinical disability, as assessed by the Expanded Disability Status Scale [r = −0.32 (95% CI −0.46 to −0.17); z-value = −4.33, P < 0.001, 13 studies]. Evidence suggests that magnetization transfer imaging metrics are sensitive to pathological brain changes in relapsing-remitting multiple sclerosis, although effect sizes were small in comparison to inter-study variability. Recommendations include: better harmonized magnetization transfer acquisition protocols with detailed methodological reporting standards; larger, well-phenotyped cohorts, including healthy controls; and, further exploration of techniques such as magnetization transfer saturation or inhomogeneous magnetization transfer ratio.


Introduction
Multiple sclerosis: a heterogeneous disease Multiple sclerosis (MS) is an immune-mediated disease involving widespread focal injury (lesions) to myelin-the fatty sheath which insulates neuronal axons-and nerve fibres within the CNS, accompanied by neuroinflammation. 1 This results in irreversible neurodegeneration.
Demyelination and neuronal damage manifest as heterogeneous clinical disability such as weakness, visual disturbances and cognitive impairment. Acute clinical episodes, or relapses, define the relapsing-remitting MS (RRMS) subtype and are often accompanied by new lesions on MRI. Although diverse in pathological appearance, lesions are indicative of inflammation and demyelination. In RRMS, relapses are interspersed with periods of stability or remission, although the clinical course varies and the choice of effective disease-modifying therapies (DMTs) is currently limited.
Reliable, non-invasive in vivo biomarkers are necessary to predict and track disease progression in individuals, and objectively assess the effectiveness of both current and emerging treatments. 2 The relationship between clinical disability and conventional MRI measures of disease burden such as lesion load visible on T 2 -weighted (T2-w) imaging 3 and atrophy 4 is, however, weak. This reflects a need for validated quantitative MRI metrics which are more sensitive and specific to disease-related pathological microstructural change in RRMS.

Magnetization transfer imaging
Magnetization transfer imaging (MTI) is sensitive to subtle pathological changes in tissue microstructure which cannot typically be quantified with conventional MRI. 5,6 MT signal is indirectly derived from protons 'bound' to macromolecules. 7 Considering a simple two-pool model for hydrogen nuclei in the brain, 8 the so-called 'free' pool of water protons shows relatively unrestricted diffusion and contributes to the bulk source of conventional MRI signal. Hydrogen nuclei in the 'bound' pool, however, are closely coupled to macromolecules (including lipids such as myelin) and have hindered rotational and translational motion, resulting in T2 decays too rapid ( 10 µs) for the signal to be detectable at typical echo times (TEs).
MTI exploits the continuous exchange of magnetization between pools to obtain signal indirectly from this 'bound' pool. Since the frequency spectrum of the 'bound' pool is much broader than the 'free' water peak, an applied offresonance radiofrequency pulse may selectively saturate 'bound' protons. Magnetization exchange between the two pools reduces longitudinal magnetization of the 'free' pool and hence it's signal intensity. Among other factors, the magnitude of this effect depends on the size of the 'bound' pool, which hence provides a surrogate marker of myelin integrity. MTI has therefore been used to study white matter (WM) diseases, including MS. 6,9 Quantifying magnetization transfer Magnetization transfer ratio (MTR), calculated as the percentage change in signal with and without a saturation pulse (Video 1), has been widely applied in clinical studies due to relatively brief acquisition and ease of calculation. MTR is, however, susceptible to field inhomogeneities and T1 relaxation effects, and varies widely depending upon specific acquisition parameters [e.g. repetition time (TR), excitation flip angle, sequence type, saturation pulse offset, power, shape and duration]. 10 Biological interpretation of MTR, as well as inter-site and inter-study comparisons, are therefore challenging, and present a barrier to clinical translation.
Magnetization transfer saturation (MTsat) inherently corrects for B1 inhomogeneities and T1 relaxation, 11 by approximating the signal amplitude and T1 relaxation at low flip angles with an additional T 1 -weighted (T1-w) image. 11,12 MTsat hence addresses some limitations of MTR, within clinically feasible acquisition times and specific absorption rate limits, and the resulting parametric maps have visibly better tissue contrast than MTR (Video 1). 11 Inhomogeneous MTR (ihMTR) exploits observed asymmetry of the broadened spectral line of the bound pool, thought to be driven by dipolar coupling effects, 13 and compares single frequency saturation at positive and negative frequency offsets with simultaneous saturation at two frequencies (+). 14,15 While not yet fully understood, ihMTR 15 is thought to be particularly sensitive to highly restricted protons in lipid chains and therefore more specific to the phospholipid bilayer of myelin than other MTI methods.
Fully quantitative MTI [quantitative magnetization transfer (qMT)] approaches using multi-compartmental models describe MT effects most rigorously by systematically varying the saturation offset and power. Important derived parameters include the fractional pool size ratio (F or PSR), the relative macromolecular content (MMC) and the macromolecular proton fraction ( f ) which provide indicators of myelin content. Calculation of either F or f requires estimation of the longitudinal relaxation rate, R 1 , for each pool. 16 The MT exchange rate from the bound to the free pool (k f ) may also help to gauge myelin status. qMT is timeconsuming to acquire, requires complex analysis and tends not to provide whole-brain coverage. qMT application has therefore mostly been limited to small-scale methodological studies.

Rationale
Previous reviews provide an overview of qMT, MTI 17 and its specific application in MS. 9,18,19 More recently, Weiskopf et al. 20 have provided a technical review of the concepts, validation and modelling of quantitative MRI, including qMT. The biophysical models used to describe MT effects in tissue, experimental evidence in brain development, ageing and pathology have also been reviewed. 6 Lazari and Lipp 21 and van der Weijden et al. 22 systematically reviewed myelinsensitive MRI validation, reproducibility and correlation with histology in humans and animal populations. Campbell et al. 23 and Mohammadi and Callaghan 24 have addressed incorporation of MTI-derived g-ratio measures to determine relative myelin-to-axon thickness.
The emergence of methods such as MTsat and ihMTR, which provide more specific measures of tissue microstructure than MTR but can be acquired relatively rapidly across the whole brain, present an opportunity to reassess the use of clinical MTI. 11,15,25 An evaluation of the body of evidence for MTI as a marker of disease from diverse studies would allow a better understanding of the effects of technique and other sources of bias across apparently contradictory results in the literature. Moreover, differences in clinical course, 26 current therapeutic approaches [27][28][29] and CSF biomarker profiles reflecting dominant pathophysiology 30 Video 1 Examples of a MTR and a MTsat parametric map from a person with relapsing-remitting multiple sclerosis. Colour scales are compressed to 0-85% for MTR and 0-6.5% for MTsat for high contrast between white matter (in red) and grey matter/demyelinated lesions (in blue). justify specific examination of the different MS subtypes. We believe therefore that a systematic review of myelin-sensitive MTI in RRMS with meta-analyses is warranted.

Purpose
The aim of the present study is thus to systematically review (i) MTI techniques used to assess pathological change in RRMS and (ii) sources of inter-study variability and bias. We then aim to apply meta-analyses to provide consensus on (iii) key cross-sectional and longitudinal pathological findings and (iv) the relationship between MTI and clinical disability in RRMS.

Materials and methods
Approval from an ethics committee was not required for the present review.

Registration and protocol
This review was not registered. The protocol was set a priori as described but not registered externally.

Search strategy and eligibility criteria
This review adhered to PRISMA guidelines. 31,32 The search terms were 'magnetisation transfer' or 'magnetization transfer' and 'brain' (with MeSH terms). The online databases searched were PubMed, Embase and Web of Science.
Search and eligibility criteria were in accordance with a protocol that had been defined a priori. For inclusion, studies had to be primary human research and had to include people with RRMS. Because the focus of the review was on MTI findings and their correlates in RRMS, studies that included people with other MS subtypes (e.g. primary progressive) or post-mortem imaging data, were excluded from the main analysis. Articles in any language were accepted, with a publishing cut-off date of 06/01/ 2021. Exclusion criteria were: inclusion of subjects with non-MS pathology (e.g. brain tumours, traumatic brain injury) where RRMS was not the main focus; paediatric (i.e. ,18 years of age) or paediatric-onset MS; solely inclusion of healthy participants (i.e. without MS patients); the full text was not retrievable; only phantom, in vitro, preclinical in vivo or ex vivo data; study published before 1980; an imaging technique other than MTI used; non-brain imaging only; non-quantitative methodology; theoretical or simulation-only papers; a clinical trial protocol, Phase I or Phase II clinical trial; conference proceedings; a review or opinion article; and, any study clearly irrelevant to the current review. Duplicated datasets were not excluded, as these could not be identified reliably from the study publications.

Search procedure
Search results were imported into EndNote. Duplicate publications were automatically removed using the in-built deduplicator tool, and the remaining duplicates were removed manually. Abstracts were checked by the author (E.N.Y.) and removed when exclusion criteria were met. Full texts were manually retrieved by the author (E.N.Y.) with online searches for article DOIs, PMID or title. If this failed, the abstract was excluded. Full-text articles were screened manually by the author (E.N.Y.) for exclusion criteria and rejected where necessary. The remaining selection was categorized according to the MS subtype. Articles without RRMS cohorts or comprising mixed subtypes were excluded from the main review. MTI data for RRMS patients in excluded studies comprising mixed MS subtypes were, however, included in meta-analyses, where it was possible to identify and analyse these separately.

Data extraction
Data were extracted in detail including demographics, acquisition parameters, MT measure and brain region, statistical methodology, summarized clinical findings and study limitations. Where possible, correlation coefficients, MT mean and standard deviation were extracted to calculate effect sizes for meta-analyses.

Statistical analysis
Descriptive statistics were calculated for demographic data, DMTs and steroid usage, and clinical disability measures. Key study findings and limitations were collated according to the MT technique used and the brain region.
When data were available from a sufficient number of studies, random-effects meta-analyses, with brain region as a nested factor, were performed to determine: 1. differences in MT metrics between patients with RRMS and healthy controls (HCs) (significance level, α = 0.05, metafor package in RStudio v1.3.1093). 2. putative relationships between clinical disability and MT metrics, in studies with reported correlation coefficients.
Where the number of studies, k, was .2 for a given brain region, follow-up sub-analyses were carried out to determine regional effect sizes, corrected for multiple comparisons [α = 0.05/(1 + n of sub-analyses)]. The Sidik-Jonkman method was used to assess between-study heterogeneity. Means were standardized (Hedges' g, R meta package) for compartmental qMT metrics and T1 was converted to R1 to ensure consistent directionality.
To assess longitudinal evolution of MT metrics in RRMS, longitudinal data (.1 time-point) were submitted to a mixed-model linear regression with mean MT as the dependent variable, time-point and brain region as fixed effects, and study as a random effect with within-study subgrouping as a nested factor (e.g. active lesions versus reactivated lesions, placebo versus treatment groups; α = 0.05; lmer, RStudio).
Marginal means for each brain region were estimated (ggeffects R package). Follow-up sub-analyses were performed when k ≥ 3 for a given brain region, with time-point as a fixed effect and study as a random effect, with subgrouping as a nested factor [α = 0.05/(1 + n of sub-analyses)]. Formal sensitivity analysis was not considered applicable to these data.

Qualitative assessment
Longitudinal change in MT, the relationship between MT and treatment, its association with disability and the dependence on the MT metric used were qualitatively assessed.

Risk of bias
Risk of bias was determined qualitatively with Joanna Briggs Institute (JBI) Critical Appraisal Checklists, 33,34 stratified by study type (case-control, randomized controlled trial, crosssectional, cohort, case report, case series, or closest match of listed study designs). An overall appraisal was given to each study based on checklist criteria. Funnel plots were used to quantify publication bias across studies included in meta-analyses. The observational nature of the data being examined limited formal evaluation of overall certainty of evidence.

Data availability
Extracted data may be provided upon reasonable request to the corresponding author.

Systematic online literature search results
Initial online database searches yielded 6758 results. Following the removal of duplicates, 3274 studies remained, which was reduced to 780 after abstract screening (Fig. 1). Full articles could not be retrieved for 42 studies and these were excluded. Of the remaining 738 articles, 368 studies met exclusion criteria (Fig. 1), leaving 370 articles for categorization by MS subtype.
As RRMS is the focus of this review, 96 studies that did not include patients with the relapsing-remitting MS subtype were excluded. The remaining selection (k = 274) was refined to 86 studies that only recruited participants with RRMS (and HCs, when included), and which form the foundations of this review. MTI data for RRMS patients from a further 38 studies, which had been excluded from the main review due to comprising mixed MS cohorts (as per the predefined study protocol) were additionally included in meta-analyses. An overview of excluded MS studies with mixed MS subtypes may be found in Supplementary Tables 1 and 2.
In adherence to our protocol, we did not include Phase I or II clinical trials. We nevertheless retrospectively examined these studies for potential inclusion in meta-analyses; however, these studies either did not include analysable MT data, or incorporated duplicate data from cohorts that had already been included in the existing analysis.

Sample characteristics
An overview of sample size, sex ratio, age and study centre location is provided in Supplementary Table 3

Sex
The median female-to-male ratio for analysed MT data was two for RRMS patients (k = 61) and 1.43 for HCs (k = 51, Supplementary Table 3).

Age
The mean age of people with RRMS was 37.15 years (5.63 SD, k = 77). Where mean age was only reported for recruited patients, this was still included; median age was not included. The mean age of HCs was 35.70 years (4.90 SD, k = 47) (Supplementary Table 3).

Clinical disability
The majority of studies (k = 73/86) used EDSS as a measure of disability with median baseline score of 1.5 (k = 64, Supplementary

DMTs and steroid usage
Intra-study and inter-study heterogeneity were apparent in treatment with DMTs and steroids (Table 1 and  Supplementary Table 5 75 Patients in four further studies were either untreated or received homogeneous DMTs which were IfN-α, 76 IfN-β 38,39 and glatiramer acetate. 77 Patients in five studies were treatment-naïve (and not receiving steroid treatment for a minimum of 14 days before imaging), 37,45,46,78,79 and only the placebo arm of a clinical trial was included in one study. 80 Eleven studies allowed steroid treatment for relapses or did not specify usage, but were otherwise treatment-naïve. 40,43,44,48,50,57,65,[81][82][83][84] Many studies did not report DMT or steroid usage (k = 28 and k = 56, Supplementary Table 5 and Table 1, respectively) or did not specify DMTs (k = 5). 59,[85][86][87][88] However, studies that reported steroid usage typically had a washout period of at least 10 days before MR imaging took place.

MTI acquisition protocol parameters
MTI protocols varied across studies (see Supplementary Results); there was heterogeneity in MR system field strength ( Fig. 2A), acquisition sequence design, image contrast, image resolution and MT pulse design, including MT pulse offset frequency (Fig. 2B). Sequence parameter details were often, however, unreported.

Metrics used
The most frequently used quantitative MT metric was MTR (k = 75, Fig. 2C 64,77,87,96,105,106,112,116 including under saturation (k sat , k = 2), 86,108 the equilibrium magnetization of the 'bound' pool and the non-ideal inversion of the 'free' pool signal (M0f and Sf, respectively, k = 2), 105,106 36,94,118 and F (k = 2). 64 MT values across the brain Studies varied as to the brain tissues in which MT was evaluated ( Fig. 2D

MTR in RRMS and HCs
Meta-analysis. Studies that compared MTR crosssectionally between RRMS patients and HCs (k = 46 with available data, n = 1130 RRMS patients/886 HC) were submitted to a random-effects meta-analysis, with brain region as a nested factor. Irrespective of brain region, MTR for RRMS patients was on average 1.17 per cent units [95% confidence interval (CI) −1.42 pu to −0.91 pu] lower than controls (z-value: −8.99, P , 0.001, Fig. 3). Between-study heterogeneity was high (total I 2 = 59.7%).
Since the number of studies examining MTR for most individual brain regions was low (k , 3), follow-up subgroup random-effects meta-analyses were only performed for the thalamus (k = 6) and putamen (k = 3). There was no significant difference in baseline thalamic MTR between RRMS patients and HCs [mean difference −3.97 pu (95% CI −10.07 to 2.12), z-value = −1.28, P = 0.202, n = 132 RRMS/113 HC, Supplementary Fig. 1] and high betweenstudy variance (I 2 = 99.2%). One additional study also found no difference in thalamic MTR between patients and controls (no effect size reported). 51 Similarly, for the putamen, there was no difference between patients and controls [mean difference −5.77 pu (−17.10 to 5.56), z-value = −1.0, P = 0.318, n = 77 RRMS/61 HC] and heterogeneity was high (I 2 = 99.6%). High between-study heterogeneity may be explained by differences in MT sequences used. 85

Longitudinal MTR change and therapeutic response
Fourteen studies (n = 563 RRMS) assessed longitudinal change in mean MTR in one or more brain regions, with a maximum of 3 years follow-up. A linear mixed-model revealed that time did not have a significant effect on MTR when all brain regions were considered [β = 0.12 (−0.56 to 0.80), t-value = 0.35, P = 0.724, Supplementary Table 6 and Fig. 2].
Nevertheless, individual studies reported small (e.g. ,1% absolute change over 2 years 47 ) but significant longitudinal decline in whole-brain MTR. 59,76 A slower (non-significant) MTR decline (e.g. 0.02% every 2 months over 14 months 80 ) and inter-subject variation were also reported. 69,76 Additionally, longitudinal stagnation or increase in MTR with treatment compared with longitudinal decreases in MTR in placebo arms was evident in large, placebocontrolled cohorts over 2 years, 91,100 suggesting MTR as a putative therapeutic endpoint. However, one study reported no deterioration in whole-brain MTR with glatiramer acetate treatment but lacked validation against a placebo arm. 75 Longitudinal change in NAWM MTR. Sixteen studies examined the longitudinal evolution of NAWM MTR. 38 Eight studies (n = 100 RRMS) reported appropriate data for a linear mixed-model to assess longitudinal change; NAWM did not change significantly over time [β = 0.037 (−0.14 to 0.22), t-value = 0.41, P = 0.68, Supplementary Table 8]. 45,46,58,66,74,84,96,120 In studies that reported a significant change over time, and in line with a previous report, 98 absolute change in NAWM MTR was small (,1.5% up to 36 months) with reported estimates of an annual decline of 0.1% in early RRMS, possibly preceding clinical onset by years. 38 However, others found no change in NAWM MTR over 2 years in an early MS cohort with minimal disability, after controlling for age and gender. 53 Alternatives to the arithmetic mean such as histogram peak location may, nevertheless, reveal changes over 12-32 months. 78 Longitudinal change in grey matter MTR. A linear mixed-model of all brain regions suggests no effect of time on NAGM MTR but there were insufficient data for follow-up analyses (see 'Longitudinal MTR change and therapeutic response' section). In the literature, however, MTR in grey matter decreases gradually ( 0.18 pu annually, compared with 0.01 pu in controls), 38 although perhaps faster than NAWM MTR in RRMS. 38 However, over 2 years, such a gradual decline is not statistically significant. 53 The longitudinal rate of grey matter change is unaffected by antiphospholipid antibody (APLA) status, 74 or treatment with IfN-β 38 or laquinomod, 100 although the latter may slow decline initially.
Longitudinal change in sub-regional MTR. There was no evidence of longitudinal change in MTR when all brain regions were considered (see 'Longitudinal MTR change and therapeutic response' section). Since there were few studies examining each brain sub-region ( Supplementary Fig. 2), no further meta-analyses of longitudinal change in MTR within brain sub-regions were constructed. However, no significant longitudinal change in MTR has been found in the thalamus, putamen, pallidum or caudate over 2 years. 53 Separately, despite a significant change in thalamic MTR (−0.13 pu/ year) over 2 years, this was not significantly different from the rate of change in control thalamic MTR, 39 and did not differ between those patients who were or were not treated with IfN-β.   Supplementary Figs 2 and 3). MTR of active CELs varies from month-to-month before and after enhancement, 45,46,71,83,93,96 while MTR of GM lesions, 53 'slowly expanding' lesions, 49 T1-w hypointense 75 and T2-w hyperintense 75,80 lesions may remain relatively stable over several years, irrespective of relapses. 80 Increases in lesion MTR may also occur, 84 such as within non-expanding lesions, although this may be accompanied by changes in T1 49 and/or lesion load 61 . MTR increases may be seen with treatment (e.g. fingolimod 66 over 2 years) although not always (e.g. laquinomod 100 ). Steroids can increase CEL MTR 46,71 although certain DMTs, including delayed-release dimethyl fumarate 91 or IfN β-1b 71,73 do not appear to alter CEL MTR. Furthermore, CELs do not tend to recover to NAWM MTR values, 46,72,98 and their longitudinal evolution may be predicted by the change in MTR of the first-month post-enhancement. 46 MTR in reactivated CELs also may deviate from NAWM MTR to a greater extent than new CELs. 96 MTR fluctuations in lesions have been partially ascribed to low reproducibility, changes in interstitial water due to acute inflammation, or perhaps remyelination. 68 Yet, when mixed lesion types are considered, a longitudinal global MTR decrease is typical. 53,54

Clinical correlates of MTR
Thirteen studies reported correlation coefficients between MTR and EDSS permitting a meta-analysis (with the brain region as a nested factor) to be performed. There was a significant negative association between EDSS and MTR across all brain regions; r = −0.32 [95% CI −0.46 to −0.17] (z-value = −4.33, P , 0.001, k = 13, n = 438, Fig. 5) and between-study heterogeneity was low (total I 2 = 0%). Across individual studies, sub-regional results were mixed but in general, suggest that there is no association between EDSS and MTR. 85,88 Whole-brain MTR and clinical correlates. In terms of wholebrain MTR clinical correlates, there is some evidence that NABT MTR correlates with EDSS 65 (Fig. 5) but not retinal nerve fibre layer (RNFL) thickness or low letter contrast acuity. 82 NABT MTR may predict longitudinal memory decline and, in combination with brain parenchymal fraction and 2-year change in ventricular fraction, information processing speed over 7 years. 59 No such association was found between NABT MTR and verbal fluency. 59 However, this study was limited by the lack of comparative longitudinal control data. Furthermore, longitudinal evolution of NABT MTR does not appear to depend on APLA status of patients. 74 NAWM MTR and clinical correlates. Many studies examined the relationship between clinical disability and NAWM MTR (Supplementary Table 4), yet only three studies reported effect sizes. A subgroup meta-analysis for NAWM showed a negative association between EDSS and NAWM MTR [P , 0.05, r = −0.42 (95% CI −0.79 to −0.04), n = 122 RRMS, Fig. 5] with low between-study variance (I 2 = 0%). However, the small number of studies (k = 4) limits the generalisability of this finding, particularly given underreporting of non-significant effect sizes. Indeed, all studies (k = 10/86) which examined the association between NAWM MTR and EDSS found no association, 37,38,58,60,75,78,85,115,119 although one study reported a significant correlation between baseline NAWM MTR and change in EDSS over 18 months (but not baseline EDSS). 48 Evidence of relationships between NAWM MTR and other clinical measures was mixed. For example, NAWM MTR was associated with MSFC z-score at 24-month follow-up but not baseline 58 while, separately, there was no relationship between MSFC z-scores and NAWM MTR 60 or 2-year change in NAWM MTR. 38 Associations may also be region-and model-dependent; for example, temporal lobe MTR was one of several significant predictors of MSFC and SDMT (an attention test) scores, independently, in regression models. 51 In terms of other biomarker correlates, WM MTR was weakly associated with serum neurofilament-a marker of neuronal injury-in RRMS (although not in control subjects), adding to evidence validating MT imaging as a biomarker of myelin integrity. 55 NAWM MTR does not however appear to be related to RNFL thickness or low contrast letter acuity. 82 Grey matter MTR and disability. Eight studies examined the relationship between grey matter MTR and EDSS (Supplementary Table 4) with some demonstrating significant associations 37,109 and others finding no such relationship. 38,57,60,85,89 One study found an association between baseline grey matter MTR and change in EDSS, but not baseline EDSS. 48 A follow-up subgroup random-effects meta-analysis showed no significant association betweenstudy baseline (cortical or cerebral) grey matter MTR and EDSS [P = 0.675, r = −0.10 (95% CI −0.57 to 0.37), n = 82 RRMS, Fig. 5] and low between-study heterogeneity (I 2 = 0%), but the number of studies was small (k = 3).

Figure 3 Random-effects meta-analysis of the difference in mean MTR in between relapsing-remitting MS patients and control
subjects in NAWM and all brain tissue types. Study baseline data were used. One study (Catalaa 78 ) was included twice as separate protocols and cohorts were used. A random-effects model with brain region as a nested factor showed that mean MTR was 1.17 per cent units [z-value = −8.99, P , 0.001, 46 studies (including grey matter and whole brain studies in Fig. 4), 1130 RRMS/886 HC] lower for people with RRMS than HCs across all brain tissue types. A random-effects model for NAWM alone showed that mean MTR was 1.25 per cent units (z-value = −7.55, P , 0.001, 31 studies/n = 32; 651 RRMS/491 HC) lower for people with RRMS than HCs. NAWM, normal-appearing white matter; RE, random-effects; RRMS, relapsing-remitting multiple sclerosis. *Averaged over sub-regions.
Four studies examined the relationship between grey matter MTR and the MSFC. 37,38,57,60 MSFC z-score did not correlate with cerebral NAGM, 37 cortical NAGM 60 or voxels of NAGM for which the MTR differed from controls. 57 Furthermore, neither change in MSFC nor its cognitive component correlated with change in MTR in NAGM over 2 years. 38 Figure 4 Random-effects meta-analysis of the difference in mean MTR between relapsing-remitting MS patients and control subjects in grey matter and whole brain. Random-effects models of study baseline data showed that mean MTR was lower for people with RRMS than HCs in whole brain (mean difference  Fig. 3 for estimate across all brain tissue types, including NAWM. GM, grey matter; NAWM, normal-appearing white matter; RE, random-effects; RRMS, relapsing-remitting multiple sclerosis; WB, whole brain. *Averaged over sub-regions. Regarding other clinical variables, NAGM MTR was significantly correlated with age 85 as well as RNFL thickness of eyes affected by optic neuritis. 82 Female subjects may also have higher NAGM MTR 37 although this was not a consistent finding. 85 In addition, NAGM MTR correlates with T1 and myelin water fraction. 97 On the other hand, grey matter MTR did not correlate with low contrast letter acuity, 82 RNFL of eyes unaffected by optic neuritis, 82 serum neurofilament levels, 55 immune cell brain-derived neurotrophic factor (BDNF) secretion, 102 APLA status, 74 fatigue 44 or disease duration. 37,57,85 Change in NAGM MTR was not associated with relapse rate, baseline T2 lesion volume or change in T2 lesion volume over 2 years 38 nor APLA status over 3 years. 74 MTR in other sub-regions and disability. MTR within other sub-regions such as the internal capsule, 43,88 cerebral corticospinal tract, 62 caudate, pallidum, putamen, accumbens, hippocampus and amygdala 85 and corpus callosum 88 was not associated with EDSS. There was a negative association between thalamic MTR and EDSS averaged over 2 years, 39 although 2-year change in thalamic MTR was not associated with EDSS at follow-up, 39 possibly reflecting a lack of change in thalamic MTR over 2 years. 53 Regarding other clinical correlates, no relationship was found between thalamic MTR or rate of change of MTR over 2 years and MSFC. 39 Nevertheless, the walk component of the MSFC was negatively associated with thalamic MTR. 39 In the cerebral corticospinal tract, MTR was associated with walk velocity and Two Minute Walk Test but not Pyramidal Functional Systems Score, gender or symptom duration, but perhaps slightly dependent on age. 62 MTR of the corpus callosum was positively associated with PASAT (the cognitive component of the MSFC) score, although possibly mediated by lesion load. 63 Cognitively impaired RRMS patients may also have marginally reduced MTR in the corpus callosum compared with unimpaired patients. 63 There may be an influence of age on MTR in the basal ganglia, thalamus and hippocampus. 85 Finally, MTR in an area of the cerebellum thought to be involved in movement trajectories was associated with performance on the MSFC arm component. 56 Clinical and other imaging correlates of lesion MTR. In lesions, any relationship between clinical disability and MTR is at most weak. 85,119,35,51,58,85,101,115 Only two studies reported a correlation coefficient (Fig. 5) for an association with EDSS and hence a meta-analysis was not performed for lesion MTR alone.
This relationship may depend on lesion type, characteristics 52 and location. 85 For example, cortical, but not WM, lesion MTR was related to EDSS, after adjusting for demographic factors. 85 Furthermore, when lesions were grouped according to their inflammatory and neurodegenerative characteristics, lesions with low MTR were found to predict attention deficits (SDMT) and general disability (MSFC), when combined with age and depression score. 52 The timescale of the study, disease duration 85 and treatment of confounding variables may affect the strength of association. A longitudinal relationship between MTR in lesions and clinical disability developed with longer disease duration in one study when not present at baseline. 58 Lesion MTR, when combined with T2-w lesion and NAWM measures, was also related to longitudinal change in deambulation (MSFC T25FW). 53 However, baseline T2-w lesion MTR was not a significant predictor of change in memory, verbal fluency or information processing speed over 7 years. 59 More generally, the association between MTR and clinical disability may depend on which clinical measure(s) are used. For example, lesion MTR was not significantly different between cognitively impaired and unimpaired patients, when assessed by an extensive battery of neuropsychological tests. 65 Similarly, MTR within (mixed-type) lesions did not correlate with motor tasks (finger tapping rate or 9HPT), 50 and was not a significant predictor in regression models to predict general clinical disability (MSFC), attention (SDMT) or fatigue (Fatigue Scale for Motor and Cognitive functions). 51 Some studies indicate associations between MTR as a measure of myelin integrity and other imaging markers of disease in MS. Weak evidence suggests that the uptake of radiotracer 18 F-PBR111, which binds to the 18-kD translocator protein, is greater in around 60% of T2-w fluid-attenuated inversion recovery (FLAIR) hyperintense regions compared with non-lesional regions with high MTR. 35 Higher uptake of 18 F-PBR111 is suggestive of a pathological increase in macrophages and microglia. Single-subject MR spectroscopy has shown elevated choline and lactate/lipids suggestive of demyelination and injury to cell membranes, alongside decreases in N-acetyl compounds, creatine and myoinositol indicating axonal loss and increased glial cell infiltration, and decreased MTR compared with NAWM in a tumefactive CEL. 72 MTR in lesions is strongly associated with other imaging metrics such as MMC, 93 and k f 87,93,112 and, to a lesser extent, quantitative T1 93,97,112 and myelin water fraction. 97 Lesion MTR is negatively correlated with relative activation on functional MRI in motor areas suggestive of functional adaptations to loss of myelin integrity, although perhaps confounded by lesion volume. 50 MTR correlates weakly with diffusion-weighted imaging metrics including fractional anisotropy 110 in large T2-w lesions and mean diffusivity 115 in chronic lesions, but not significantly with susceptibility-weighted phase imaging values, despite a negative trend. 115 Additionally, T2-w and T1-w 'black hole' lesion volume, as well as 2-year change in T2-w lesion volume may predict lesion MTR 13 years later, although uncorrected for baseline lesion MTR. 61 Nevertheless, as a general trend across the RRMS literature, MTR within lesions does not tend to correlate with other disease biomarkers. T2-w lesion MTR is not significantly associated with age, 85,115 time since diagnosis, 101 visual contrast acuity or RNFL thickness, 82 immune cell BDNF secretion, 102 or APLA status (+). 74 MTR in CELs was not associated with anti-CD3 plus anti-CD28 stimulated BDNF secretion, despite a negative trend. 102 MTR in T1-w 'black holes' is not associated with RNFL thickness or visual contrast acuity. 82 There is some evidence that APLA+ patients show greater reduction in MTR in T1 'black holes' compared with APLA-patients over 3 years, but this may be driven by lesion volume changes. 74 Evidence for associations between lesion MTR and disease duration or gender is mixed, and may depend upon acquisition parameters and lesion type. 85,115 Magnetization transfer saturation Three studies used MTsat (Fig. 2C), 11,111,114 beginning with Helms et al. 11 who showed that, on a whole-brain histogram, the WM MTsat mode appeared visually reduced in a RRMS patient compared with controls. Furthermore, compared with NAWM, MTsat in a CEL and non-enhancing lesions was visually lower on a parametric map. 11 Saccenti et al. 114 confirmed that MTsat was significantly lower in WM 'plaques' and periplaques than NAWM. Yet, MTsat did not correlate with EDSS or disease duration in plaque, periplaque or NAWM ROIs. 114 MTsat may additionally correlate with radial diffusivity, T1w/T2w ratio and synthetic MR-derived myelin volume fraction, although this was stronger in plaques than NAWM. 114 Finally, Kamagata et al. 111 used MTsat as a surrogate for myelin volume fraction to calculate the tract-averaged MR g-ratio within WM in a small RRMS cohort. 111 The g-ratio was increased (indicating myelin degradation and/or axonal loss) compared with HCs, in motor somatosensory, visual and limbic regions. Subnetwork g-ratio strongly negatively correlated with WM lesion volume, but not with disease duration or EDSS, although the latter was correlated with g-ratio connectome nodal strength mainly in motor, visual and limbic regions.

Inhomogeneous MTR
Two studies employed ihMTR as a measure of myelin status in RRMS. 88,119 ihMTR was reduced in lesions and NAWM compared with control WM, and reduced in lesions compared with NAWM. 119 Within sub-regions, single-slice ihMTR was lower for patients in the thalamus, frontal, temporal and occipital lobes compared with controls, but not as EDSS score. A multi-level random-effects model with brain region as a nested factor within each study showed a significant negative association (r = −0.32, z-value = −4.33, P , 0.001, 13 studies, 438 RRMS) between MTR and EDSS across all brain regions. Studies which did not report a correlation coefficient were not included. Random-effects sub-analyses showed a significant correlation between EDSS and NAWM MTR (r = −0.42, z-value = −2.17, P = 0.030, four studies, 122 RRMS), and not grey matter (r = −0.10, z-value = −0.42, P = 0.675, three studies, 82 RRMS). Sub-analyses were not performed when the number of studies, k , 3. *MTR values were averaged over sub-regions of NAWM. GM, grey matter; NABT, normal-appearing brain tissue; NAWM, normal-appearing white matter; WML, white matter lesions; RE, random effects; CI, confidence interval. different in the corpus callosum, internal capsule or putamen. 88 ihMTR varied across WM tracts, but was highest in the internal and external capsule and lowest in the genu of the corpus callosum. 88,119 ihMTR in WM lesions, but not NAWM, was negatively associated with EDSS. 119 However, when sub-regions were considered, EDSS was significantly associated with ihMTR (but not MTR) in frontal and temporal NAWM, the corpus callosum, internal capsule and the thalami. 88 Quantitative magnetization transfer qMT metrics examined varied across studies (see 'Quantitative measures of magnetization transfer: metrics used' section). Sled and Pike 116 first modelled the compartmental MT signal in RRMS in two lesions on a single-slice proton density-weighted image for a RRMS patient. Compared with frontal WM, lesions had reduced k f , F, R1 free and T2 bound and increased T2 free . Parameter estimates were higher for the newer lesion compared with the older lesion for k f , F and R1 free , but lower for T2 free and T2 bound . Indeed, other studies also show lower k f and k sat lesions than NAWM and HC WM, while T1 free and T1 sat present the inverse pattern. 86,87,112 Up to 4 months before the appearance of new or reactivating CELs, k f may even decrease while T1 free increases. 96 However, changes are subtle, and month-by-month change may be less predictable for reactivating CELs.
Increasing are also reduced in lesions compared with NAWM and control WM, with reduced F and R 1free in T2 hyperintense lesions visible on selective inversion recovery-derived parametric maps. 104,105 Finally, MMC is reduced in CELs but may recover post-enhancement. 93 The relationship between pathology and qMT-derived metrics is evidently complex, but may still differentiate between lesions with similar MTR, particularly when lesions are T1-w isointense. 112 Differences between NAWM and control WM qMT are, however, subtle. Some studies report differences for qihMT, 119 119 Nine studies were submitted to a random-effects meta-analysis to compare qMT in NAWM and WM. 36,86,87,94,112,116,118 There was a significant difference between patients and controls across all qMT metrics [standardized mean difference −0.60 (95% CI −0.95 to −0.25), z-value: −3.51, P , 0.005, n = 87 RRMS/98 HCs, Fig. 6]. Additional follow-up models for metrics where k ≥ 3, however, showed no significant difference for R1 free , R2 bound , f and k f (α = 0.0125, Fig. 6) despite a trend for k f . Other brain regions were not assessed due to limited data.
In cortical grey matter, k f , F, R1 free and T2 bound appear lower and T2 free higher than in lesions and frontal WM. 116 RRMS patients have lower k f than controls in cortical grey matter but F does not differ, except for patients with high disability. 64 No differences between patients and controls were found in cerebral or cerebellar grey matter for f, T1 free or T2 bound . 36 In deep grey matter, f was lower for patients than controls. 118 However, differences in methodology can results in over-or underestimation of f in certain ROIs (e.g. thalami). 118 Few studies have examined the relationship between qMT and clinical disability in RRMS. Cortical grey matter k f may be negatively associated with EDSS and Choice Reaction Time, but not SDMT or PASAT. 64 Associations between EDSS and both qMT and qihMT in lesions, but not NAWM have also been reported. 119 Combining qMT parameters, and including covariates such as lesion load and age may improve models 94 but collinearity (e.g. between f and T2 bound or k f and T1 free ) may be problematic if used in the same model. 36,112 Risk of bias Seven studies (8.1%) were given an 'excellent' rating based on JBI Critical Appraisal Checklist criteria (Supplementary Table 10). The majority of studies rated 'good' or 'ok' (k = 33, 38.4% each) and 13 studies (15.1%) were given a 'poor' rating. The latter result, however, was partly driven by methodological 'proof of principle' studies for which there was no specific checklist.
Overall, the main sources of bias, where relevant, were inadequate examination of confounding factors, poor standardization and reliability of MTI outcomes, inappropriate statistical analyses, particularly concerning no correction for multiple comparisons, poor matching of cases and controls, and a lack of detail regarding setting/site description. Funnel plots also suggest that case-control studies with high precision are lacking, particularly for analyses of grey matter ( Supplementary Fig. 4). Similarly, there appears to be a bias towards small, less powerful studies which examined the relationship between clinical disability and MTI in WM ( Supplementary Fig. 5). In contrast, studies that used compartmental models had relatively high precision, particularly R1 and MTsat ( Supplementary Fig. 6).

Discussion
Our search demonstrated a broad literature of MS-specific MTI studies, a considerable number of which were excluded due to the lack of distinctions between MS subtypes or grouped subtypes in analyses and results. Eighty-six studies used MTI to investigate cerebral RRMS pathology, the vast majority (87%) of which used MTR. We also incorporated in meta-analyses additional RRMS data from a further 38 studies which included mixed MS subtypes.

Common findings
Lesion MT was found to be lower than in NAWM. MT was also generally reduced in non-lesional brain for patients compared with HCs, indicative of subtle loss in microstructural integrity. Conversely, smaller sub-regions (e.g. thalamus, putamen) did not show such differences. The absolute sensitivity of MT metrics to pathological changes in the brain of people with MS is modest; the difference in MTR between patients with RRMS and HCs is estimated to be small ( 0.5-2%) compared with inter-study variability. Meta-analyses did not support a significant annual longitudinal decline in MT in RRMS despite qualitative evidence to the contrary and a trend in NABT. In lesions, MT is inclined to fluctuate over time.
Although associations between MT measures and clinical disability in RRMS were apparent, relationships were weak, and confounded by factors such as age. This association may be limited by the lack of longitudinal data over sufficient time periods for divergence in disability to become apparent.
Studies examining longitudinal change and clinical correlates were limited to MTR; we did not identify any such studies using other techniques, such as MTsat, ihMTR or qMT.

Sample characteristics
Overall, patient sample sizes across the RRMS MTI literature were small, with a median of ,20 subjects, and many studies were statistically underpowered. Research with a technical or proof-of-concept focus tended to include a single subject or handful of participants (e.g. 11,42,105,106,116,118 ). Conversely, international clinical trials recruited much larger cohorts (e.g. 91,92 ), but at the expense of standardized, welldocumented MTI protocols.
Comparisons between MS and (typically) age-matched HC subjects featured in a number of studies, albeit often with smaller control than patient groups. Such well-matched control data are important to account for confounding variables such as age, 85 and may additionally provide reference measures to help improve comparability of MT metrics across studies and centres.
Treatment effects are a further potential confound of MT microstructure measures, and inter-and intra-study heterogeneity was apparent in DMT and steroid usage which is an additional source of variability. Although some studies control for treatment effects, greater consistency is required in studies whose primary focus is imaging biomarker validation.

Imaging acquisition protocols
Systematic comparison of MTI in RRMS demonstrates substantial heterogeneity of MTI acquisition protocols. There was wide variation in magnetic field strength, pulse sequence, image weighting, excitation flip angle, TR and TE. With the rapid evolution of MRI hardware and techniques, such sources of variation are inevitable and well-recognized in the quantitative MRI literature. The nature of MT acquisition, however, makes MT measurements particularly sensitive to these factors. For example, simulations suggest that the difference between grey and WM MTR at 3 T at an offset frequency of 1.5 kHz is around 43% larger than at 1.5 T. 117 Use of proprietary hardware and pulse sequences allows broader access of MTI to research groups with limited MRI pulse programming expertise, but typically fixes, restricts and even conceals important pulse sequence parameters.
MT measurements are especially sensitive to characteristics of the MT pulse. Quantification typically assumes selective saturation of the 'bound' pool with minimal direct saturation of the 'free' water pool. The extent to which this is achieved in vivo and the resulting tissue-type contrast, however, depends on the complex relationship between tissue properties, hardware, sequence parameters and MT pulse design features including the offset frequency, power, pulse duration and shape. 98 In particular, our finding of the wide variance in NAWM MTR in RRMS cohorts is suggestive of sequence parameter dependence. Early experiments with relatively low offsets (e.g. 110,113 ) are likely to have a greater direct saturation effect. Improved harmonization and standardization of MT protocols between centres would help to minimize these sources of variability.
The majority of large-scale MT studies in RRMS to date have used MTR, which is relatively easy to acquire and analyse. Importantly, however, MTR signal is markedly dependent on T1 and B1 effects in addition to magnetization transfer processes, which limits its specificity as a microstructural imaging marker of myelin integrity.
qMT provides the most accurate modelling of MT processes and is helpful for probing microstructure in healthy and pathological tissue; however, prolonged acquisition is needed at multiple pulse powers and offset frequencies with adequate spatial resolution. Whole-brain coverage is therefore not currently feasible for clinical imaging in patients.
Emerging MT methods such as MTsat and ihMTR provide potentially more robust and specific measures of myelin integrity than MTR within clinically feasible acquisition times. 11,121 Histological validation in felines has shown that MTsat is sensitive to demyelination, 122 and, in mice, ihMTR signal is more specific to myelin than MTR. 121 Both techniques, however, require further validation with histology and study in larger patient and HC cohorts.

Tissue types and definitions
The substantial variation observed in MTR values for different tissue types is likely due not only to varying acquisition parameters discussed above, but also how tissue type is defined, and variations in methods by which the regions are segmented from structural imaging. For example, individual studies examine different combinations of WM, NAWM, cortical and deep grey matter structures, atlas-based ROIs, and whole-brain analyses. Moreover, a number of different 'lesion types' are recognized in RRMS, as defined by their signal characteristics; for example, T2-w or FLAIR hyperintensities, T1-w hypointense lesions or 'black holes', and contrast-enhancing lesions. A clear definition of lesion subtypes is therefore important for the interpretation of their MT characteristics.

Sources of bias and limitations
Study quality, including assessment ratings of application of methods to minimize bias, was variable; the large majority of studies classified as 'good' or 'ok', and those rated 'poor' were largely associated with small methodologically focused papers.
Bias was apparent towards small sample sizes, and also towards studies using MTR compared with other techniques. Overall, high precision case-control studies were lacking and bias was apparent towards small, less wellpowered studies correlating clinical disability with MTI measures. Overall, the small number of studies that used compartmental MTI models showed relatively high precision compared with MTR. Inadequate examination of confounding factors, poor standardization and reliability of acquisition methods, flawed statistical analyses, poor matching of cases and controls and lack of detail regarding the research setting were also identified in a significant number of studies.
Across studies, there was a near-universal bias towards European and North American populations, which is likely to reflect the geographical prevalence of MS, the attention given to the disease within healthcare systems, and access to MRI and research protocols. Importantly, analysis of the location of study centres highlights possible bias due to data duplication from multiple or overlapping analyses of cohorts. This is rarely overtly reported, but may influence the calculation of effect sizes.
With regard to the review process, the literature search procedure was carried out by a single reviewer which may have led to bias in study selection, and influence overall certainty of evidence. Meta-analyses were limited by large interstudy protocol heterogeneity and missing data, and also did not take into account patient or control group demographics. The scope of the present review is also limited to results in RRMS patients. Data from progressive MS subtypes were excluded, but may still provide insights on how MT metrics reflect microstructural damage in MS.

Implications for future studies using MT in RRMS
The findings of this review indicate the potential for MT measures of microstructure as useful disease markers in MS, but equally highlight large variability in quantitative findings compared with modest effect sizes.
Major sources of systematic differences and variance in MTR measured across studies are technical variation in acquisition protocols, and confounding magnetic field homogeneity (B1) and magnetization relaxation processes (notably T1); relaxation processes, in particular, may lead to bidirectional longitudinal fluctuations in MTR. These effects, combined with variability in cohort characteristics and experimental design, contribute to weak association with clinical measures of disease.
Harmonizing MTR acquisition protocols across participating centres will go some way to mitigate this variability, although will not address the confounds of B1 and T1 effects. Signal from more quantitative, clinically applicable MT methods such as MTsat and ihMT is less confounded by these technical features and other tissue characteristics, and hence provide more specific biomarkers of myelin status. These methods, however, require further evaluation, with rigorous validation against tissue reference data, and other biomarkers of MS disease activity and neurodegeneration.
Cohorts which are adequately powered to detect predicted effect sizes are likely to require large multicentre studies of highly characterized patients with defined MS disease subtypes. Further optimization, harmonization and cross-site validation of MTI protocols across multiple MRI platforms, will allow assessment of inter-site variance and potential systematic differences in measures across centres.
Adoption of more consistent definitions and methods for segmenting tissues of interest will also facilitate comparability across sites and studies.
We, therefore, expect that moving towards more quantifiable, harmonized MT protocols in large well-defined and annotated cohorts will provide a more reliable indication of the relationships between MT and clinical features in MS, and hence their potential utility in patient stratification and clinical trial platforms.
Moreover, we suggest that in order for MTI to evolve as a useful imaging tool in MS and other diseases, there is a need to establish consensus standards for image acquisition, analysis and reporting from an international group of experts working across centres, as has been successfully achieved with other quantitative MRI methods such as diffusion and perfusion imaging. [123][124][125] Conclusion This systematic review demonstrates a substantial literature on MTR applied to RRMS. The evidence evaluated suggests that MT imaging can detect subtle disease-related differences. There is, however, large measurement variability due to differences in technique; this dominates over small effect sizes which, in turn, limit clinical and biological interpretation. The implementation of more robust emerging quantitative techniques, and consensus regarding optimized, harmonized protocols in large well-characterized patient cohorts will be required to establish the value of MTI as a useful microstructural marker in RRMS, for translation into wider clinical use.