Structural magnetic resonance imaging in dystonia: A systematic review of methodological approaches and findings

Abstract Background and purpose Structural magnetic resonance techniques have been widely applied in neurological disorders to better understand tissue changes, probing characteristics such as volume, iron deposition and diffusion. Dystonia is a hyperkinetic movement disorder, resulting in abnormal postures and pain. Its pathophysiology is poorly understood, with normal routine clinical imaging in idiopathic forms. More advanced tools provide an opportunity to identify smaller scale structural changes which may underpin pathophysiology. This review aims to provide an overview of methodological approaches undertaken in structural brain imaging of dystonia cohorts, and to identify commonly identified pathways, networks or regions that are implicated in pathogenesis. Methods Structural magnetic resonance imaging studies of idiopathic and genetic forms of dystonia were systematically reviewed. Adhering to strict inclusion and exclusion criteria, PubMed and Embase databases were searched up to January 2022, with studies reviewed for methodological quality and key findings. Results Seventy‐seven studies were included, involving 1945 participants. The majority of studies employed diffusion tensor imaging (DTI) (n = 45) or volumetric analyses (n = 37), with frequently implicated areas of abnormality in the brainstem, cerebellum, basal ganglia and sensorimotor cortex and their interconnecting white matter pathways. Genotypic and motor phenotypic variation emerged, for example fewer cerebello‐thalamic tractography streamlines in genetic forms than idiopathic and higher grey matter volumes in task‐specific than non‐task‐specific dystonias. Discussion Work to date suggests microstructural brain changes in those diagnosed with dystonia, although the underlying nature of these changes remains undetermined. Employment of techniques such as multiple diffusion weightings or multi‐exponential relaxometry has the potential to enhance understanding of these differences.


INTRODUC TI ON
Dystonia is a movement disorder involving repetitive or sustained muscle contractions leading to abnormal posturing, with an estimated prevalence of 120/100,000 population [1]. Clinical presentation is heterogeneous, involving single or multiple muscle groups (focal, segmental or generalized), with genetic or idiopathic aetiology, and of childhood or adult onset [2]. Animal models have implicated the cerebellum and basal ganglia in pathogenesis, demonstrating cerebellar Purkinje cell abnormalities, including ectopic dendritic spines and aberrant firing patterns, as well as disrupted striatal gamma aminobutyric acid (GABA) and dopamine neurotransmission [3]. Human postmortem studies support this, with patchy cerebellar cell loss and torpedo bodies in cervical dystonia [4].
In vivo studies show network-based disruption to cerebral motor control pathways in dystonia [5], with differences observed in the primary sensorimotor cortex, putamen, thalamus and cerebellum in functional magnetic resonance imaging (fMRI) and electrophysiological studies, the latter also implicating disruption to normal inhibitory processes [6]. Inhibitory/excitatory imbalances are also suggested by MR spectroscopy and radionucleotide imaging, with changes to GABA neurotransmission observed in cerebellar and sensorimotor cortices [5,7] (Figure 1). Despite this, standard clinical structural MR sequences have not demonstrated gross abnormalities, suggesting changes may be at the microstructural level.
Multiple MRI approaches are used to derive information regarding brain structure in medical research. Diffusion MRI (dMRI) can probe tissue microstructure based on the degree of freedom of molecular movement, deriving properties including mean diffusivity (MD, the overall freedom of diffusion), fractional anisotropy (FA, the degree of orientational preference), axial diffusivity (AxD, the apparent diffusion coefficient along the dominant diffusion axis) and radial diffusivity (RadD, the degree of diffusion in the plane perpendicular to the primary diffusion direction) (Figure 2.1a). Other approaches aim to identify subtle localized volume or size differences, with atlasbased automatic or manual approaches to delineating brain regions and comparing between groups (Figure 2.1b). Thirdly, relaxometry methods aim to provide a quantitative measure of molecular relaxation following excitation with a radiofrequency pulse, inferring information relating to local tissue properties which can influence the speed of this relaxation. Common relaxometry approaches include T2* relaxometry, heavily influenced by susceptibility effects which can be particularly substantially induced by iron, and T2 relaxometry which more closely relates to water content (Figure 2.1c).
Finally, magnetization transfer imaging investigates the effects of macromolecules on unbound free water, the signal for which decays too rapidly to measure directly, allowing signal sensitivity to tissue structures such as membranes and myelin (Figure 2.1d).
This review evaluates structural MRI studies used in the investigation of genetic and idiopathic forms of dystonia to date, with particular emphasis on methodological considerations including study design, imaging acquisition, pre-processing and analysis methods.
Our aim is to synthesize the breadth of work undertaken to date, to critically appraise the imaging methodological approaches used and to highlight consistent anatomical findings which may provide pathophysiological insights of dystonia.

ME THODS
In line with PRISMA guidelines, studies using structural brain MRI in genetic and idiopathic forms of dystonia were systematically reviewed. Embase and PubMed databases were searched for articles up to January 2022. The full search strategy is detailed in Appendix S1. Inclusion criteria were case-controlled studies using structural MRI in the investigation of dystonia. Exclusion criteria included single case reports, no control cohort, secondary or psychogenic/functional dystonia, no use of structural MRI, deep brain stimulation studies using imaging only as a surgical planning tool, studies with minimal methodological detail, methodological testing, conference proceedings, review articles, where no full paper was available, not written in the English language. Abstracts were screened for inclusion and exclusion criteria by two investigators (C.M. and K.J.P.) working independently with further screening of the full text of the identified articles. Data were extracted from each of the studies to include type of dystonia, number of patients, patient demographics and phenotyping, imaging modality, imaging acquisition features, imaging pre-processing steps, imaging analysis methodology and study findings. Overall risk of bias was assessed by two investigators (C.M. and K.J.P.) based on the risk of bias in non-randomized studies (RoBANS) bias assessment tool [8] with additional consideration of specific imaging methodology features.

Structural MRI modalities
Seventy-seven studies were identified, four of which were animal models of dystonia. Of the 73 human studies, 45 used dMRI, 39 size/ volume-based methodology, three relaxometry-based approaches and one magnetization transfer imaging; 14 studies combined multiple approaches. In interpreting the findings of these studies multiple factors have the potential to impact image and data quality, and ultimately the study findings, including the image acquisition approach, data preprocessing and analysis methodology (Appendices S2 and S3).
The b values employed were in typical ranges of 700-1000 s/mm 2 , with five higher at 1100-1500 s/mm 2 [38,39,[49][50][51]. showing the hydrogen ions aligned in the magnetic field of an MR scanner, with the application of a radiofrequency pulse causing them to come out of alignment with the magnetic field (blue arrows) and their spins coming out of alignment with each other (black arrows). T1 is the time taken for longitudinal relaxation (i.e., the blue arrow to return to alignment with the main magnetic field) and T2 is the time taken for transverse relaxation (i.e., the black arrows to return to being out of phase with each other), with T2* being additionally influenced by local magnetic field differences. Proton density is a measure of how densely packed the protons are. (1d) Magnetization transfer imaging, showing bound and unbound protons in a magnetic field, with a radiofrequency pulse aimed mainly at bound protons applied, and then the transfer of this magnetization to the unbound protons which produces a measurable signal; this is MT weighting. MTR is the difference between an acquisition with and without this off-resonance pulse. (n = 1) [16] or not stated (n = 4) [14,[29][30][31].
With relaxometry, the number of echo times (TEs) or flip angles used to characterize the relaxation profile can influence estimation, with more data points and optimized spacing generally improving accuracy. This is particularly important for T2* acquisitions, as the gradient echo acquisition method is more greatly influenced by inhomogeneities in the magnetic field, resulting in a higher risk of noise influencing the measurements. For the T2* acquisitions the studies used six [69], eight [47] and 12 [48] different TEs; one study also assessed T2 relaxation time with four different TEs, and T1 and proton density (density of protons) with two flip angles [47]. The single magnetization transfer study [15] was based on the magnetization transfer ratio (MTR), involving acquisitions with and without an offresonance radiofrequency pulse targeting the macromolecule pool. This is a semi-quantitative method, more biologically meaningful than magnetization transfer-weighted imaging but very sequence dependent, meaning reproducibility and comparison between studies is limited, and influenced by T1 effects, inhomogeneities in the radiofrequency field and several other acquisition factors.
Quantitative MT approaches aim to ameliorate these issues and provide a more biologically meaningful measure, acquiring multiple images and fitting a quantitative model to the data.

Pre-processing
Pre-processing is undertaken after image acquisition and before data analysis, aiming to correct for sources of imaging artifact that can alter the measured signal ( Figure 3.1d-1f). This can reduce data quality, causing signal dropout, distortions and potential for erroneous conclusions. Motion is a key consideration in a movement disorder cohort, where there is potential for consistent between-group differences. Subject motion can degrade results, potentially causing misalignment between images or slices or signal dropout. Taking steps to correct for this and to identify and remove outlying results reduces the potential for data bias.
Sources of artifact particularly relevant to dMRI relate to the echo planar imaging method, including B0 field inhomogeneities (where local regions in the imaging field are subject to different local magnetic environments), distorting signal and eddy currents (localized currents created by switching diffusion gradients rapidly).
Amongst the dMRI studies, the majority undertook eddy current and motion correction with this either directly stated or inferred based on the analytical software employed. Whilst no studies detailed the specifics of the motion correction undertaken, most are likely to have corrected for between-volume-motion only, with many pre-processing pipelines not accounting for within-volume motion.
Four studies additionally corrected for susceptibility distortions [45,46,50,51]. Other sources of artifact particularly relevant to diffusion imaging, and not accounted for in any of the studies, include noise distribution bias (including thermal noise) related to the 'noise floor' (a measurable signal that remains in the absence of any true signal), gradient deviations due to nonlinearity of the diffusion gradient causing geometrical distortions, and Nyquist ghosts-an echo-planar-imaging-specific artifact source where slight timing inconsistencies lead to the appearance of a 'ghost' image halfway across the main image.
The volumetric studies predominantly did not outline the preprocessing steps to account for artifacts, stating that standard approaches were used. This would be presumed to include motion correction which is particularly key to enable accurate image registration and alignment for volumetric comparisons. Two studies undertook denoising [25,71], and three inhomogeneity correction [25,71,72]. T1 relaxometry approaches are particularly influenced by non-uniformity in the flip angle, and T2* relaxometry by magnetic field inhomogeneities; correction for these was only reported in one paper [47]. MTR imaging is particularly susceptible to radiofrequency field non-uniformities, but only correction for motion was reported.
A number of additional potential artifacts can degrade data quality, and were not corrected for by any of the studies reviewed. These include signal drift, when the measured signal gradually changes over time (e.g. due to system heating), and Gibbs ringing, involving signal oscillations in regions with sharp boundaries in measured signal.
Analysis methods A regional or whole brain focus can be taken during data analysis (Figure 3.2 gives example analysis approaches). The former involves targeted analysis of predefined locations/tracts, allowing a hypothesis-driven approach, whilst a whole brain approach F I G U R E 3 Structural imaging approaches. 3.1 Examples of factors influencing measured parameters. (1a-1c) Examples of the effect of differing voxel size and anisotropy on partial volume effect in white matter, with (1a) showing small isotropic voxels, (1b) larger voxels and (1c) anisotropic voxels, with more inclusion of tracts with different orientations in the larger and anisotropic voxels, which would influence the fractional anisotropy (FA) measured. (1d) Example of the effect of motion on a T2-weighted image, (1e) a gradient deviation map showing variation in the magnetic field and (1f) an example of the signal removed from the Gibbs ringing artifact in diffusion MRI. 3.2 DTI as an example of the range of potential analysis approaches: (2a) a map of MD (mean diffusivity) values, (2b) a map of FA values, (2c) T directional orientation colour-coded FA map, (2d) tract-based spatial statistics, (2e) graph theoretical analysis, (2f) a region of interest delineated in the cerebellum and (2g) an example of a tractography reconstruction using regions of interest to define way-points along the tract [Colour figure can be viewed at wileyonlinelibrary.com] studies combined these approaches [9-11, 14, 15, 26, 30, 33, 36, 37, 40, 46, 47, 49, 52, 53, 56, 58, 61, 71, 80, 81], with five identifying the more specific regions post hoc using data from the whole brain analysis [9-11, 33, 53], an approach which has the potential to bias results.

TA B L E 1 Animal models
Amongst the dMRI studies that took a regional approach, either  [14,25,52,58,72,75]. Differences in these measures are often attributed to hypertrophy or atrophy of associated brain regions, with inference of underlying pathological processes. However, image alignment and partial volume effects have particular propensity to influence these measures and therefore should be rigorously considered during analysis. For these studies, the segmentation process plays a vital role in the measured results.
Manual segmentation has the potential for more anatomically accurate delineation of structures where anatomical boundaries may be unclear using automated methods; however, this introduces potential for inter-rater variability, avoided with an automated approach.
All papers used automated segmentation approaches, with one study undertaking additional manual thalamic segmentation.
Of the relaxometry studies, all three used a regional approach, with one additionally undertaking whole brain analysis [47], and the single MTR study employed a whole brain voxel-by-voxel comparison of the MTR with additional ROIs [15].

Animal model imaging
These all involved DYT1 models, including knock-out (KO) (n = 2) and knock-in (KI) (n = 2) designs ( Table 1). Two studies [82,83]  KO models identified lower striatal free water and an associated higher MD [83], whilst higher FA was identified in the caudate, putamen, sensorimotor cortex and brainstem [85]. By contrast, elevated free water in the cerebellum and striatum [82] and lower FA in the superior cerebellar peduncle, sensorimotor cortex, caudate and putamen were identified in the KI forms [84].

TA B L E 2 Genetic dystonia
The two studies involving DYT12 cohorts revealed widespread elevated MD on TBSS analysis, one with corresponding lower FA [36] and the other lower FA in the fornix, anterior thalamic radiation, corticospinal tract and superior corona radiata [32]. Use of more targeted probabilistic tractography between the paralimbic/ sensorimotor cortex and caudate/putamen found increased fibre counts [36]. A single diffusion study of SGCE-mutation-positive patients (DYT11) found higher FA and lower MD in the subthalamic brainstem and subgyral sensorimotor cortex respectively, as well as an increased subthalamic WM volume [67]; a further study of grey matter (GM) volume identified no differences, although  higher disease severity was associated with greater putaminal volume [74]. For those with DYT27 mutations, a single TBSS study found lower FA in the cerebellar peduncles, pons, midbrain, cerebral peduncles, thalamus, internal capsule, and frontal and parietal WM, with more targeted predefined tractography between thalamus/putamen and cortex revealing lower FA between the dentate nucleus and thalamus [49].

Task-specific focal dystonias
Diffusion studies involving individuals with writer's cramp have found conflicting results. One identified no FA differences between patients and controls [38], whereas another noted lower FA in the tracts between the middle frontal gyrus and putamen [51].  areas tracking to either the primary sensorimotor cortex or brainstem [11], and higher volumes in the posterior putamen and globus pallidus [81].
Of the studies investigating spasmodic dysphonia, lower FA and higher MD were identified in the right internal capsule in the patient group using TBSS, with higher MD in the corona radiata, internal capsule, thalamus, cerebral peduncle and cerebellum [14,56]. Language regions, namely the inferior frontal gyrus, were also implicated, with corpus callosal differences associated with the presence or absence of a tremor. Others have found a lower FA and higher MD and RadD in the corpus callosum and WM tracts, with higher AxD in the more anterior WM regions [14]. Another study used more targeted probabilistic tractography focused on regions involved in speech between the insula and cortex, noting no significant differences between patients and controls [35]. Volumetric studies have shown some overlap in findings with increased cortical surface area and GM thickness in the inferior frontal gyrus and primary sensory and motor cortices [14,25] and, on occasion, implication of the superior and middle temporal gyri, superior frontal gyrus, putamen and pallidum [54].
Studies focused on musician's dystonia included embouchure dystonia, noting lower AxD between the primary somatosensory cortex and putamen and higher AxD between the supplementary motor area (SMA) and the superior parietal cortex [50]. VBM volumetric studies have identified greater sensorimotor cortex and putaminal GM volume compared to both unaffected musicians and non-musicians [71].

Non-task-specific focal forms of dystonia
Amongst cervical dystonia cohorts lower WM FA has been identified compared to controls in regions including the superior cerebellar peduncles, thalamus, middle frontal gyrus, corpus callosum, and prefrontal and visual cortices [34,37,39,40]. By contrast, higher WM FA has been observed in the substantia nigra [37], putamen [12], pons, thalamus, supplementary motor cortex, middle temporal gyrus and cingulate gyrus [15]. Probabilistic tractography studies have corroborated abnormalities relating to thalamic projections, including lower fibre counts between the thalamus, middle frontal gyrus and brainstem [34]. MD value differences have been conflicting, with higher values reported in the basal ganglia and cerebello-thalamo pathways [39,40], whilst others have reported lower values in the caudate, pallidum and putamen [12]. Volumetric studies have yielded similarly contrasting results with some identifying larger [24], and others smaller, volumes in the caudate, putamen, globus pallidus and primary motor cortex [15], primary sensory cortex, premotor cortex, SMA, medial temporal gyrus and prefrontal cortex [20]. Specific focus on the cerebellum found smaller GM volumes in the anterior and VI lobules and smaller cerebellar peduncles [21]. A single longitudinal study identified a reduction in left primary sensorimotor cortex volumes over 5 years [20]. Finally, no significant differences were observed using a T1, T2, T2* relaxometry approach and proton density maps [47], whereas others identified higher R2* values in the globus pallidus [48], potentially implicating increased brain iron deposition.
Studies of blepharospasm have not identified FA differences [30,57] but did note reduced average tract volumes and streamline counts between the brainstem and motor cortex [30], and increases in local diffusion homogeneity in multiple regions, correlating with disease severity [57]. Volumetric studies have again produced mixed findings including higher GM volume in the putamen, cingulate and middle frontal gyrus, lower orbitofrontal and occipital cortical volumes, variable primary sensorimotor cortex volumes, and lower cortical thickness in frontal and temporal regions [18,22,28,30].
Studies examining idiopathic paroxysmal kinesigenic dyskinesia (PKD) have found higher FA in the thalami and right anterior thalamic projections [52], with no cortical thickness differences compared to controls. Using VBM-based morphological network matrices, global differences were noted, including shorter path length and higher local efficiency in the clinically affected cohort [77].

Mixed idiopathic cohort studies
A number of studies have compared task-and non-task-specific dystonias, identifying lower FA in the middle/inferior frontal gyrus, corpus callosum, putamen and premotor cortex in task-specific forms, and in the middle cingulate gyrus and primary sensorimotor cortex in non-task-specific forms [55,58]. Volumetric GM comparisons have found widespread higher GM volume and cortical thickness in task-specific forms, including sensory and premotor cortex, parietal and temporal regions, basal ganglia and thalamus, whilst cerebellar measurements appear to vary [55,58].
Several studies have collectively assessed cranio-cervical dystonias (cervical, blepharospasm and oromandibular). In combination, one study found no consistent differences compared to unaffected controls [53], whilst analysis of the individual forms identified both higher [70] and lower [13] MD in key motor regions in cervical dystonia, and lower FA and higher MD in the basal ganglia in blepharospasm [66,70]. Volumetric studies have identified higher volumes in the caudate and lower volumes in the putamen in both cervical dystonia and blepharospasm, with a higher thalamic GM volume in cervical dystonia and lower in blepharospasm [19]. Others have identified increased cerebellar GM volumes, reduced cortical thickness and an overall tendency towards smaller GM volumes [72,73,86].
Other comparisons have included (i) a mixed group of cervical and laryngeal dystonia, identifying lower thalamic volumes compared to healthy controls [26], (ii) laryngeal dystonia in musicians and non-musicians demonstrating lower FA in the musicians in the  (Figure 4a).
Some differences appear independent of the manifestation of motor symptoms, with NMC exhibiting changes to motor WM pathways to a lesser degree than in clinically manifesting forms.
Amongst sporadic dystonia cohorts results were more variable but do demonstrate widespread motor pathway abnormalities, most commonly lower FA, involving the WM between the brainstem, cerebellum, basal ganglia/thalamus and sensorimotor cortex, and higher GM volumes in the supplementary/secondary motor regions ( Figure 4b). There are also indications of differences in task-and non-task-specific forms of dystonia, with a tendency towards larger volumes in implicated regions in task-specific forms and connected by WM pathways in which dMRI abnormalities have been identified [87]. Magnetoencephalography in writer's cramp has shown lower post-movement-event-related synchronization of beta activity, potentially indicating impaired deactivation of the motor cortex [6], GABA spectroscopy has shown reduced GABA levels in the sensorimotor cortex and lentiform nuclei [88] and positron emission tomography with flumazenil (which binds to GABA-A receptors) showed reduced binding in the cerebellum and sensorimotor cortex in focal hand dystonia [7], indicating that inhibitory changes may be involved in pathogenesis.
Contribution to bias from selection of participants potentially impacted a subset of studies, with relatively small sample sizes [24,32,34,38,40,43,49,64,65,68]. Very few studies considered the non-motor phenotype of participants which may confound results, and likewise sub-optimal imaging methodology has the potential to introduce performance bias [9-13, 53, 65]. Measurement of exposure was generally adequate, with selection of participants based on diagnosis by a specialist. Blinding of outcome measures was not commented on but has less relevance with studies almost exclusively using automated ROI selection and analysis approaches.
Attrition bias was relevant only to studies collecting data at multiple timepoints [20,42,43], with one study having substantial dropout in the 5-year follow-up (seven out of 19) [20]. There was no overt evidence of selective outcome reporting in the literature.
This review particularly serves to highlight the methodological limitations of many dystonia structural imaging studies to date.
Notably, lower field strength and large anisotropic voxels limit a number of studies, resulting in potential compromise in data quality compared to higher resolution acquisitions. Whilst most studies did employ the key basic pre-processing steps, other steps aimed at reducing artifacts and distortions would probably further improve data quality. In particular, few studies reported outlier detection and rejection, relevant for movement disorder cohorts in identifying and removing from analysis any acquisitions with substantial signal dropout, which could lead to false conclusions regarding systematic between-group difference. There was also generally no documented correction for inhomogeneity in magnetic or radiofrequency fields which can particularly influence T2* relaxometry and MTR imaging respectively. The approach taken to data analysis also varied substantially, with several studies using a less hypothesis-driven whole brain approach; additionally some of the ROI-based studies only did so following a whole brain analysis. Whilst this approach allows for additional information to be gathered regarding the identified regions, it does reduce the validity of these as hypothesis-driven areas of in- terest. An additional consistent limitation is the use of biologically non-specific measures, such as FA, MD, volume, T2*, MTR, which do not necessarily enable inference of the nature of any underlying abnormality, aetiology or pathophysiological process (Figure 2.2).
Of the range of potential imaging approaches available, the majority of the studies focused on WM diffusion or GM volume-based approaches, with only a small number utilizing other approaches such as relaxometry and magnetization transfer imaging. Amongst the applied approaches, there is substantial scope for application of more advanced or optimized methodologies, to enhance the level of biological meaning attributable to measured signals and provide insights into any potential role for impaired neurodevelopmental processes resulting in differences in cellular morphology. For example, newer methods for analysing diffusion data overcome several of the technical limitations, such as tractometry (involving segmentation and subsequent analysis of WM tracts), fixel-based analysis (determines fibre-specific measures within a single voxel) and biophysical models that attribute diffusion signals to particular underlying tissue properties. More biological specificity could be attained using techniques such as multi-shell dMRI (collecting data at multiple diffusion gradient strengths), relaxometry utilizing multi-exponential T2 decay and quantitative magnetization transfer methods.

CON CLUS ION
Magnetic resonance imaging differences are evident between dys-

This work was supported by an ABN/Guarantors of Brain Clinical
Research Training Fellowship (520286) and a Wellcome Trust

CO N FLI C T O F I NTE R E S T
No conflict of interest to declare.

FI N A N CI A L D I SCLOS U R E
No financial disclosures to declare.

DATA AVA I L A B I L I T Y S TAT E M E N T
N/A.

CO N S ENT S TATEM ENT
Not applicable.