Macro- and micro-structural insights into primary dystonia: a UK Biobank study

Identification of the UK Biobank dystonia cohort used a clinically validated algorithm derived from a national clinically linked databank [4, 38]. This made use of ICD-10 codes derived from NHS hospital records (UK Biobank data fields 41,202, 41,204 and 41,207) and Primary care records (Read codes v2, data portal field read_2), with inclusion and exclusion diagnoses outlined in Supplementary Table 1. All individuals identified as having been diagnosed with dystonia and who had undergone T₁-weighted and diffusion-weighted brain MRI were included (UK Biobank application number 69610). A 1:4 age- and sex-matched control cohort was derived from the overall cohort following application of the same exclusion criteria as was used for identification of the dystonia cohort [38]. Scans marked by the UKBB as ‘unusable’ (based on pathology or T1 artifact) were visually inspected and only excluded if pathology was identifiable. Participants were excluded from the diffusion-based analyses if they had missing or incomplete diffusion-weighted imaging data or were marked by the UKBB as ‘incompatible’ due to use of a different acquisition protocol.

Imaging was undertaken in one of four centres (Table 1) using a 3 T Siemens Skyra scanner with a standard Siemens 32-channel radiofrequency receive head coil, using a standardised acquisition protocol across sites. A sagittal 3D T₁-weighted MPRAGE was acquired, with field of view 208 × 256 × 256 and voxel size 1mmx1mmx1mm. The dMRI acquisition used a spin echo planar imaging sequence with × 3 multi-slice acceleration, fat saturation, field of view 104 × 104 × 72, and voxel dimensions 2mm × 2mm × 2mm. Fifty distinct diffusion directions were acquired each for b = 1000 s/mm² and b = 2000s/mm², with 5 b = 0 s/mm² images. A 6/8 partial Fourier readout was used [1, 24].

dMRI data are processed within the UK Biobank using standardised pipelines. However, significant variability of approach exists in diffusion imaging analysis, and while there is no current evidence based consensus on optimal approaches, attempts have been made to optimize methodology to reduce the impact of imaging artefacts on dMRI data [2]. For this study, a local, optimized preprocessing pipeline was applied to the raw diffusion data (with facial feature removal prior to UK Biobank data release), described in Fig. 1A. This involved brain extraction, denoising [35], outlier detection (e.g., due to signal dropout from motion) [31], signal drift correction, and correction for Gibbs ringing [18], eddy currents, susceptibility distortion and subject motion [3], and gradient non linearities [5]. Each of the analytical steps were manually reviewed for a randomly selected 10% subset to assess data quality. Parameter estimation used a diffusion kurtosis representation and a constrained weighted least squares fitting approach with weight reduction for outlier measurement [30, 37], demonstrated to enable more accurate estimation of diffusion tensor imaging (DTI) parameters [36] than more standard DTI fitting approaches. Parameter maps were created for FA, mean diffusivity (MD), mean kurtosis (MK), axial kurtosis (AK) and radial kurtosis (RK). These parameter maps were additionally used in the grey matter analysis detailed below.

The raw T₁-weighted images were used for grey matter segmentation using Freesurfer [13], with six region of interest (ROI) masks created from these segmentations (prefrontal, sensorimotor, non-frontal, striatal, thalamic and cerebellar cortical masks, Fig. 1A & B, Supplementary Table 2). Within each ROI, total cortical thickness (summed subregion values) and total volume were calculated, with the latter also calculated for subcortical regions. Grey matter diffusion measures were calculated with the first b = 0 image registered to the brain extracted T₁-weighted image for each participant (using FSL EPI reg [17]), and the inverse of this transform applied to the GM ROIs with nearest neighbour interpolation into diffusion space. Voxels with partial volume with other tissues due to the differing voxel dimensions were excluded and median ROI MD, MK, AK and RK were calculated (from the diffusion kurtosis parameter maps).

Within the white matter pathways, we evaluated both whole tract averages (tractography) and along-tract profiling (tractometry), investigating evidence of localised changes in tract morphology, as has been shown in genetically homogenous dystonia cohorts [9]. Tractseg [40] was used to segment the motor tracts: middle cerebellar peduncle, bilateral inferior cerebellar peduncles, superior cerebellar peduncles, frontopontine tracts, corticospinal tracts, anterior thalamic radiations, superior thalamic radiations, thalamoprefrontal tracts, thalamopremotor tracts, thalamoprecentral tracts, thalamopostcentral tracts, striatopremotor, striatoprecentral tracts and striatopostcentral tracts (Fig. 1C). Optic radiations were included as a non-motor comparison tract [23]. Multi-shell multi-tissue constrained spherical deconvolution was applied to the preprocessed diffusion images to extract the white matter fODF, with subsequent peak-extraction limited to a maximum of three peaks/voxel. A fully-connected convolutional neural network was used to create a tract probability image for each orientation and tract, with start and end regions segmented and fibre orientation maps calculated within the segmented regions [42]. Probabilistic tractography [41] with 5000 streamlines was performed within each segmented region, with median parameter diffusion tensor values (FA and MD) and diffusion kurtosis values (MK, AK and RK) extracted from the streamlines. In addition, along-tract profiling (tractometry) was performed [11, 39]: the tracts were portioned into 20 segments along their length with the centroid of all streamlines within each segment identified, the end two segments excluded and median value across streamlines calculated for each of the 18 remaining segments [11]. The thalamoprefrontal, thalamoprecentral, striatoprecentral, thalamopostcentral and striatopostcentral tracts were excluded from tractometry analysis due to their geometry not enabling consistent segmentations to be produced.

Tracts where significant differences were observed between the dystonia and control cohorts were further examined using white matter compartmental parameters (OD—orientation dispersion index, ND—neurite density index, and FWF—free water signal fraction), estimated using post-hoc Neurite Orientation, Dispersion and Density Imaging (NODDI) [44], hypothesized to better disentangle biological features. This was performed for tracts that demonstrated significant differences in any of the original analyses. As the tractometry analyses demonstrated visually corresponding tract profiles for left and right tracts, the two were combined for this post hoc analysis.

Initially, 78 individuals diagnosed with dystonia and 320 control participants with imaging data were identified. Of these, two dystonia and 9 control cases were excluded due to visible pathology (cases n = 2 (extensive white matter changes n = 1, hydrocephalus n = 1), controls n = 3 (extensive white matter changes n = 2, congenital malformation n = 1)), incompatible acquisition protocols (n = 4) or absent/incomplete diffusion data (n = 2), Therefore, a total of 76 individuals with dystonia and 311 healthy controls were included for onward analysis. Median age at scanning for both patient and control groups was 64 (SD 8.87 and 8.91 respectively), with a male to female ratio of 1:1.44 for both groups. Dystonia cases were subdivided based on available clinical diagnoses, with the two largest groups being cervical dystonia (n = 40) and dystonic tremor (n = 17) (Table 1).

No significant differences in regional volume or cortical thickness were observed when comparing the overall dystonia, cervical or tremor cohorts to unaffected controls. Diffusion kurtosis measures showed significant differences only in the striatum for the tremor cohort, with a higher RK values (p = 5.84 × 10^–4) compared to controls (Fig. 2).

No significant diffusion tensor (FA or MD) or diffusion kurtosis (MK, AK, RK) differences were observed between the overall cohort, cervical or dystonic tremor cohorts, compared to the unaffected control group (Supplementary Fig. 1).

For tracts in which significant differences were observed, along-tract profiles are included in Figs. 3, 4, 5 (Supplementary Figs. 2, 3, 4 display all tract analyses).

Figure 3 demonstrates higher FA in the mid right superior cerebellar peduncle (p = 2.69 × 10^–4), while NODDI identified significant changes to ODI measures across the same tract (lower mid-tract (p = 7.19 × 10^–4) and higher distal tract (p = 3.3 × 10^–7), as well as significantly lower NDI (p = 6.8 × 10^–4) and FWF (p = 0.002) values in the dystonia cohort. ODI differences were also demonstrated in the anterior thalamic radiations (higher proximal-mid (p = 6.53 × 10^–5) and distal (p = 1.96 × 10^–4) values) and striatopremotor (lower mid-tract values (p = 0.001) tracts in the dystonia cohort.

No significant differences were observed for DTI and DKI measures (Fig. 4). NODDI identified differences in ODI in the superior cerebellar peduncles (higher proximal (p = 0.002) and lower mid (p = 4.17 × 10^–4) and distal (p = 1.2 × 10^–4) values), and anterior thalamic radiations (lower distal values (p = 7.37 × 10^–4)) in the cervical dystonia cohort compared to controls Fig. 5.

Multiple regions of significant difference were observed across the superior cerebellar peduncle and anterior thalamic radiation, including FA (right, higher in tremor (p = 5.22 × 10^–4)), MD (left, higher in tremor (p = 3.43 × 10^–4)), MK (left, lower in tremor (p = 4.59 × 10^–4)) and AK (left, lower in tremor (p = 1.92 × 10^–4)) in the anterior thalamic radiations, with higher FA (p = 6.61 × 10^–5) found in the right superior cerebellar peduncle. These were coupled with corresponding changes in NODDI measures, with lower ODI (p = 4.31 × 10^–4) and higher NDI (p = 8.44 × 10^–5) in the superior cerebellar peduncles, and lower ODI (p = 8.3 × 10^–5) and lower NDI (p = 5.96 × 10^–4) in the anterior thalamic radiations. Differences were additionally observed in the right striatopremotor tract (right tract, higher FA (p = 4.71 × 10^–4))) and optic radiations (left tract, higher FA (p = 1.78 × 10^–7), and higher distal FWF (p = 1.6 × 10^–7)) in the dystonic tremor group compared to controls.

Morphometric analyses demonstrated no between group differences for the overall or subgroup analyses. This is partially at odds with existing literature, particularly in genetic and task specific forms of dystonia [16, 27]. However, non-task-specific dystonias, such as cervical dystonia and blepharospasm, have demonstrated substantial variability in both the presence and direction of any identified volumetric differences. Minimal existing work has investigated GM in dystonic tremor, identifying increases in the inferior frontal cortex [19], motor cortex [10] and striatal [19] volumes/thickness. Not previously applied in a dystonia cohort, grey matter diffusion kurtosis measures here identified tremor-specific striatal differences, with significantly higher RK values. Previous work has shown diffusion kurtosis measures to be correlated with brain microstructural changes, including during normal brain development [25] and in white matter in multiple sclerosis [22]. The differences observed may indicate subtle morphological striatal changes, particular to tremor, such as in cell body density or dendritic branching complexity, however, the MR acquisition methodology used as part of the UK Biobank study is insufficient to apply modelling approaches to ascertain the nature of this signal difference. Approaches which provide additional information on tissue heterogeneity, including compartmentalisation into neurites and soma, may facilitate better understanding and insight into these observed differences.

Tractography identified no significant differences between groups, although multiple tracts had significantly higher FA values in the dystonia cohort compared to controls prior to correction for multiple comparisons. Tractometry, aimed at identifying regional variation along individual tracts, identified localized differences, most notably higher FA in the superior cerebellar peduncles, anterior thalamic radiations and striatopremotor tracts (Fig. 6). These higher FA values are potentially explained by disrupted inhibition resulting in increased pathway activation and subsequent increased muscle activity, or conversely could reflect a mechanism of attempted compensation for muscle overactivation. While studies involving cohorts of individuals diagnosed with genetic dystonias have often identified lower (rather than higher) FA values, those involving idiopathic forms (comprising the majority of this cohort) have reported more mixed results with some identifying lower MD or higher FA in corroboration with our findings, while only a few have found lower FA values [7, 12, 28]. The dissonance may reflect the heterogeneity of different dystonia diagnoses with varying underlying pathological processes, including, for example, disruption to white matter tract integrity in some forms, with others such as adult onset idiopathic dystonias involving differing mechanism such as reinforcement of selected projections. Alternatively, aberrance of earlier life neurodevelopmental processes may be involved, such as the pruning of dendritic connections and increases in the complexity of WM pathway course, which later decompensate and clinically manifest as a focal dystonia. The optic radiations, included as a non-motor comparison, unexpectedly demonstrated distally higher FA in the tremor cohort, with a few studies previously implicating visual regions in dystonia [28]. This may reflect compensatory visual system processes given the role of visual feedback in movement, or evidence of more diffuse microstructural changes within the brain.

Post hoc NODDI was undertaken as it is hypothesized to provide more microstructurally meaningful interpretation of signal differences, demonstrating more pronounced localised differences in the tracts implicated in DTI and DKI analysis. The superior cerebellar peduncles and thalamic radiations were the most dominantly implicated, with ODI the measure in which greatest variation between groups was observed, with overall lower values in the dystonia cohort. This corresponds with the FA differences outlined above, and the suggestion of ‘over-use’ of subsets of neuronal pathways in giving rise to dystonia. Further interpretation of the ODI NODDI measure could suggest that white matter changes in idiopathic dystonia are related to disruption to the organisation and dispersion of fibres within tracts, rather than differences in axonal morphology. However,—in the superior cerebellar peduncles for the overall dystonia cohort, and for the superior cerebellar peduncles, anterior thalamic radiations and the optic radiations for the tremor cohort compared to controls—differences in NDI and FWF measures were also observed, potentially indicating a degree of morphological change to the axons themselves in these regions, which may reflect secondary adaptation or an alternative primary pathological process. No previous work has applied NODDI in a dystonia cohort for comparison, although one fixel-based analysis approach has identified a lower Fibre Density value in the striatal region [45].

Differences, compared to the control cohort, were most frequently observed in the tremor cohort, observed across both grey and white matter measures. Several more recent studies have indicated that distinct imaging changes are observed when both tremor and focal dystonia are observed, compared to isolated, focal dystonia alone. Suggested explanations for this include adaptive changes, particular to tremor, rather than of the dystonic syndrome itself. The limited dystonia subgroup sizes and phenotypic detail available within the UK Biobank dataset doesn’t allow for a direct tremor versus no tremor comparison, however, a previous comparison of dystonic and essential tremor identified higher FA in cerebellar projections, regardless of tremor type but to a greater extent within the dystonic tremor cohort, with wider evidence existing of differences between tremulous and non-tremulous dystonia forms [7]. One limitation of the limited phenotypic data available in the UK Biobank is the inability to determine body part affected by tremor, with this broad diagnostic category likely encompassing multiple phenotypic subgroups which may have differing imaging correlates. Additionally, whilst a diagnosis of essential tremor is applied as an exclusion criteria (Supplementary material 1), there is potential that a portion of this group would fall into this diagnostic category, and so the tremor group and resultant findings may represent a broader range of the tremor spectrum than dystonic tremor alone.

As we observed, whole tract average tractography approaches potentially overlook localized pathway changes, with along-tract profiling (tractometry) providing more anatomically specific localised assessment, undertaken here for the first time in dystonia. Despite this advantage, diffusion tensor and kurtosis measures remain non-specific even when profiled along a tract. Biophysical models such as NODDI aim to overcome some limitations of traditional DTI and DKI approaches by modelling tissue microstructural properties such as intra- and extra-axonal compartments. However, estimating microstructural parameters with acquisitions at low-to-moderate b-values (as used in this acquisition) is notoriously difficult due to inherent degeneracy (where multiple physically plausible parameter settings result in the same signal decay) and requires a priori fixing or linking of values, limiting the ability to capture true biophysical changes [20]. Hence, these measures should be interpreted with caution, and more extensive acquisitions are necessary to map more specific microstructural features. A further limitation of this dataset is the lack of in-depth clinical phenotyping data available, with the heterogeneous spectrum of dystonic disorders reduced to broad diagnostic categories providing minimal insight into the specific motor (or non-motor) phenotype of the participants. This limits potential interpretation of the findings seen here and requires further work in well phenotyped and homogenous cohorts in future to delineate syndrome and symptom specific differences. This is a particularly dominant factor within the overall cohort analysis, where potentially differences between the dystonia syndromes may have differing impacts, either cancelling or diluting the resultant imaging differences seen. To assess whether the cervical dystonia cases, comprising the majority of the overall cohort, were having excess impact on the overall analysis, we performed a Cohen’s D effect size calculation between the cervical and non-cervical dystonia cases for two key regions, finding no meaningful between group differences (ODI measure in the mid anterior thalamic radiation: 0.029; FA in the mid-distal right superior cerebellar peduncle: 0.097). Additionally, whether the differences seen are causative or compensatory cannot be established via a cross-sectional imaging study, with longitudinal and animal model studies necessary to attempt to disentangle this point. Cross site image acquisition differences have additional potential to influence findings in a study such as this, however, previous work in UK Biobank data has demonstrated that this variation in image centre has had minimal impact on study findings [2].

Overall, this work demonstrates microstructural motor pathway differences in idiopathic forms of dystonia, with additional changes that appear to be specific to when dystonia and tremor are observed together. Further work is needed to develop and apply more biologically meaningful and specific in vivo human imaging measures, in robustly phenotyped clinical cohorts, that may better allow disentanglement of potential underlying brain morphological and microstructural differences and advance mechanistic understanding in dystonia.