Free‐Water Imaging in Friedreich Ataxia Using Multi‐Compartment Models

The neurological phenotype of Friedreich ataxia (FRDA) is characterized by neurodegeneration and neuroinflammation in the cerebellum and brainstem. Novel neuroimaging approaches quantifying brain free‐water using diffusion magnetic resonance imaging (dMRI) are potentially more sensitive to these processes than standard imaging markers.

3][4][5][6][7][8] Secondary degeneration has been observed in motor and somatosensory regions of the cerebral and cerebellar cortices. 2,3,7,9he symptoms of FRDA arise as a result of reduced production of frataxin, a mitochondrial membrane protein crucial to the production, storage, and transport of iron-sulfur cluster containing proteins. 10Reduced frataxin triggers the pathological processes underlying FRDA, including persistent oxidative stress and neuroinflammation, which lead to cellular dysfunction and cell death that impacts neurons, oligodendrocytes, and myelin sheaths. 4,11Thus, therapeutic discovery is focused on increasing the availability of frataxin or addressing downstream issues such as oxidative stress or neuroinflammation. 12Alongside such discovery is the ongoing effort to understand the neurobiology of disease expression and progression, both of which are critical to developing effective therapies. 13 novel, and potentially highly sensitive in vivo means of quantifying neuronal and axonal degeneration and neuroinflammation involves measurement of "brain free-water."Increased free-water can result from the atrophy of neurons and neurites (dendrites and axons) which leads to reduced density of the cellular matrix and greater freedom of extracellular fluid movement.5][16][17][18] For example, free-water indices have been shown to be strong discriminators of different forms of parkinsonism, 19 disorders of inflammatory axonal degeneration, 20 and cognitive status in studies of aging and dementia. 14,21,22espite potential as a marker for neurodegeneration and disease progression, free-water modeling has not yet been applied to examine the microstructural and extracellular features of FRDA.
Free-water quantification relies on multi-compartment dMRI analysis techniques, which aim to provide biophysically accurate models of brain tissue.The two most prevalent approaches currently available are neurite orientation dispersion and density imaging (NODDI) 12 and bi-tensor diffusion tensor imaging (btDTI). 23NODDI modeling classifies the fluid diffusion in each voxel into three compartments. 12,24First, a restricted, intracellular compartment describes diffusion that is directionally constrained (ie, within axons or dendrites), giving rise to the "intracellular volume fraction" (FICVF).Second, a pericellular compartment refers to partially hindered diffusion immediately around neurites, providing the "orientation dispersion index" (ODI).Third, a free-water compartment captures the magnitude of unconstrained (isotropic) extracellular fluid diffusion, quantified as the isometric volume fraction (FISO). 12lternatively, the btDTI model relies on a twocompartment model, separating the unconstrained freewater component from the remainder of the diffusion signal. 23Although originally developed to reduce partial-volume effects in the standard DTI pipeline by removing the free-water signal, 23 the value of this signal as a marker of neuropathology has since been realized. 24One advantage of btDTI relative to NODDI is that it can be applied to single-shell dMRI data commonly collected in research contexts and can also be easily translated to clinical use, whereas NODDI requires a more complex multi-shell dMRI acquisition sequence. 12,19,24The NODDI model, however, provides a more biophysically specific representation of cellular and extracellular structure, which potentially enhances its biological interpretability and clinical utility.
The current study applied NODDI and btDTI modeling to brain dMRI images acquired from 14 individuals with FRDA and 14 healthy controls to test the hypothesis that free-water is increased within key regions of the cerebellum and brainstem in FRDA.We also aimed to assess the relative sensitivity of free-water versus conventional microstructural measures (e.g., fractional anisotropy) for quantifying brain abnormalities related to FRDA.A priori analyses were undertaken within 10 cerebellar and brainstem regions-of-interest (ROIs), alongside exploratory whole-brain assessments.To assess the relative discriminatory ability (FRDA vs. controls) of each metric, effect sizes were reported for all NODDI and btDTI outcomes, not just those representing freewater.We additionally explored the relationship between diffusion metrics and neurological disease severity, as assessed by the modified Friedreich Ataxia Rating Scale (mFARS) 13 and Disease Burden score. 25

Participants
Twenty-eight participants aged 18 years and older were recruited: 14 individuals with FRDA (mean age = 28.3AE 8.7 years, six women) and 14 controls (mean age = 27.3AE 6.0 years, five men).One additional participant from each group was removed during quality assessment of the raw images (see the Quality Assessment Section).All individuals with FRDA had molecular confirmation of biallelic GAA repeat expansions in the FXN gene.Exclusion criteria included comorbid neurological illness, history of traumatic brain injury, and history of psychiatric illness other than depression or anxiety.Neurological disease severity in individuals with FRDA was quantified using the mFARS. 13

MRI Data Acquisition
MRI scans were obtained for all participants using a 3T Siemens Biograph (Siemens Erlangen, Germany) with a 32-channel head coil at Monash Biomedical Imaging, Monash University.Multi-shell diffusionweighted images were acquired as follows: TR = 4800 ms; TE = 104 ms; voxel size = 2.5 Â 2.5 Â 2.5 mm 3 with 30 directions at b = 2500 s/mm 2 , 30 directions at b = 1000 s/mm 2 , 6 directions at b = 500 s/mm 2 , 3 directions at b = 200 s/mm 2 , and 5 b = 0 s/mm 2 ; 58 axial slices; slice acceleration factor = 2. Volumes were acquired with right-to-left phase encoding; an additional 5 b = 0 s/mm 2 were acquired with left-to-right phase encoding to allow for subsequent distortion correction.

MRI Preprocessing
Preprocessing of the anatomical T1w and diffusion images was performed using QSIPrep 0.16.0RC3, which is based on Nipype 1.8.1 26 (RRID: SCR_002502).Each stage of the QSIPrep workflow is described in detail in Cieslak and colleagues 27 and in the Supplementary Material.Briefly, the T1w images were corrected for intensity non-uniformity, skullstripped, and spatially normalized to MNI space (ICBM 152 Nonlinear Asymmetrical template version 2009c 28 ).For the diffusion data, images were denoised and corrected for spatial distortions and head motion, followed by eddy current correction.The dMRI timeseries were co-registered to the T1w images.

Quality Assessment
Several quality assessment metrics were calculated as part of the QSIPrep pipeline.These are described in detail in Yeh et al 29 and include framewise displacement (head motion) and slicewise cross-correlations (corrupted data slices).One control and one FRDA participant were removed due to extreme values (>2 SD above the mean) in the number of bad slices (one participant), mean neighbor correlation (one participant), and maximum framewise displacement (both participants).Additionally, dMRI volumes from three participants that were poorly registered to their T1w volumes at the completion of the diffusion preprocessing stage underwent an additional co-registration in Statistical Parametric Mapping software, version 12 (SPM12; http://www.fil.ion.ucl.ac.uk/spm/) using the default parameters.

Derivation of Diffusion Metrics
Conventional FA (FA u ) maps were computed from the preprocessed dMRI volumes using "dtifit" from the Diffusion Imaging in Python imaging library (DIPY) (version 1.5.0,Python 3.10.6).Free-water and free- water-corrected FA maps from the btDTI model were derived using custom scripts implemented in Matlab (R2019b, Mathworks, Natick, MA, USA), as previously described, 23,30 using b-values under 2000 s/mm 2 (ie, 0, 200, 500, 1000) to minimize the influence of non-Gaussian diffusion effects. 31The NODDI metrics (FICVF, FISO, and ODI) were derived from the preprocessed dMRI volumes using the NODDI toolbox for Matlab (version 1.0.5, Matlab R2019b). 12All metric maps were normalized to MNI space by applying the nonlinear transformations estimated during T1w image preprocessing.Diffusion maps were also normalized to SUIT space 32 to allow for an additional voxelwise assessment of group differences within cerebellar and brainstem regions (see Supplementary Methods).

Regions of Interest
Binary masks of 10 disease-relevant ROIs were defined in MNI space using available atlases: midbrain, medulla, and pons of the brainstem; 33 the anterior (lobules IÀIV), superior posterior (lobules VI, Crus I/II, VIIB), and inferior posterior (lobules VIIIA/B, IX) cerebellar lobules and dentate nuclei 32,34,35 and the superior (SCP), middle (MCP), and inferior (ICP) cerebellar peduncles. 36Each binary mask was eroded by one voxel to reduce partial volume effects.All masks included both hemispheres.
For the ROI analysis, the mean intensity of voxels covered by each mask for each metric was calculated, excluding voxels with value zero.To ensure no contamination by CSF, voxels outside of the gray and white matter tissue masks were set to zero.
To examine whole-brain effects, voxel-wise analyses for each metric were conducted between individuals with FRDA and controls (with sex and age as covariates of non-interest) using FSL "randomize" with 5000 permutations, and threshold-free cluster-forming enhancement. 38Significance was taken at FWEcorrected P < 0.05.This was replicated in SUIT maps using the same cluster-forming thresholds.
To investigate associations with clinical status, Spearman's rank correlation analyses were conducted for each ROI between all diffusion metrics and both mFARS score and Disease Burden score (disease duration Â GAA1 repeat length 25 ).

ROI Analysis
Free-water changes in people with FRDA, relative to controls, were evident using both the btDTI and the NODDI models.Increased FW was observed across all ROIs (P < 0.005; Fig. 1; Table S2).These effects were most pronounced in the midbrain, dentate, SCP, ICP, and anterior and superior-posterior lobules (P < 0.001).In contrast, significant elevations of FISO were present only in the anterior and superior-posterior cerebellar lobules and SCP.
Regarding metrics of tissue microstructure, FA u was significantly reduced in people with FRDA in all regions of the brainstem, the anterior cerebellar lobe, SCP, and ICP relative to controls.However, FA u differences in the medulla and anterior cerebellar lobe were no longer significant when the free-water signal was removed during btDTI modeling (ie, FA FW ).All other regions that showed significant group differences in FA u remained significant in FA FW , albeit with reduced effect sizes.Of the NODDI metrics, FICVF was decreased in individuals with FRDA within the medulla, midbrain, dentate, SCP, and ICP, which overlapped substantially with the FA u and FA FW findings.ODI was significantly increased in the pons and SCP in individuals with FRDA and, although not reaching corrected-level significance, showed a large effect-size trend toward a decrease in the dentate nuclei (P = 0.01, η 2 = 0.28).
Across all metrics, FW yielded the largest mean effect (mean η p 2 = 0.43; Table S2).The largest and most consistent effects were evident in the SCP, with very large effect sizes detected across all diffusion metrics (η p 2 ≥ 0.46).Descriptive statistics (mean, SD) for each region and metric are provided in Table S1.

Voxel-Wise Whole-Brain Analysis
Voxel-wise whole-brain assessments are shown in Figure 2. Measures of isotropic diffusion (FW and FISO) were significantly elevated in individuals with FRDA relative to controls across the bilateral cerebellum, SCP, and brainstem.Individuals with FRDA also showed significant reductions relative to controls in measures of restricted diffusion (FA u , FA FW , and FICVF) primarily in areas of the SCP, medulla, and midbrain.In addition to this, areas of the thalamus and motor cortex showed significantly lower FICVF in the FRDA group.ODI showed a small number of significantly elevated voxels in the brainstem, anterior cerebellar lobe, and SCP in individuals with FRDA.No other contrasts were significant.Voxel-wise analyses of the cerebellum and brainstem in SUIT space showed comparable results (Supplementary Results and Fig. S1).

Clinical Correlation Analysis
Spearman's rank correlations of ROI mean value versus mFARS score in individuals with FRDA did not reach statistical significance.However, as depicted in Figure 3 and Table S3, medium effect sizes (r > 0.3) were observed with FW in the cerebellar lobules and dentate nucleus, and with FISO in the midbrain, anterior and superior-posterior lobes, and SCP.Individual correlation plots are shown in Figures S2-S7.Negative correlation between Disease Burden score and FA u in the superior posterior lobe reached significance (r = À0.75,P = 0.003) (Fig. 4).This region also showed strong positive correlations in FW and FISO maps, but did not reach significance after multiple comparison correction.Further nonsignificant positive trends with medium effects sizes were observed between the clinical metrics and FW in the cerebellar lobules, dentate nuclei, and SCP and ICP, with FISO in the midbrain, cerebellar lobules, dentate nuclei, SCP and ICP.Numeric outcomes and plots of correlations between diffusion metrics and Disease Burden score are shown in Table S5 and Figures S8-S13, respectively.

Discussion
In this study, we report for the first time significantly elevated free-water in established areas of brain pathology in people with FRDA.Across all ROIs, the bi-tensor DTI model had a greater between-group discriminatory capacity (mean effect size) relative to NODDI's three-compartment model.Free-water effect sizes also surpassed conventional (ie, FA u ) and novel (eg, FICVF) measures of tissue microstructure in cerebellar regions, but not in the brainstem or cerebellar peduncles.Free-water represents a compelling target for further longitudinal analyses and more detailed characterization as a sensitive microstructural measure of FRDA expression and progression.
0][41] In the current study, elevated free-water was observed in key gray and white-matter regions of the brainstem and cerebellum, consistent with patterns of atrophy observed in volumetric studies in FRDA. 3,42hese findings are also in line with increased [ 18 F]-FEMPA binding, a measure of neuroinflammation, in the dentate nuclei, cerebellar cortex, brainstem, and SCP in people with FRDA. 41The free-water measures may therefore reflect the aggregate impact of both neurodegeneration and neuroinflammation in FRDA, rendering them more sensitive to disease expression and progression than other imaging measures.Indeed, in other neurodegenerative diseases, free-water has been shown to be more sensitive to disease onset or differential diagnosis, more strongly correlated with disease symptomology, and more predictive of future outcome than other measures of gray and white-matter volume or integrity. 19,21,43The current work provides evidence of similar utility in FRDA and motivates further characterization in larger and longitudinal cohorts.Although structural (T1) imaging has consistently found widespread macroscopic structural changes in individuals with FRDA, 3,7,44 multi-compartment metrics such as the bi-tensor and NODDI models provide sensitivity to change in tissue microstructure that may precede macroscopic change. 44These metrics additionally provide quantitative measures of white-matter integrity and organization, which enable insights into the biological impact and progression of the disease not possible with structural scans alone.
The two diffusion models we employed provided both concordant and divergent findings across the different ROIs.Generally, we found a high agreement between FW and FISO values across all ROIs in the patient sample (Table S4), and in all regions in the control sample aside from the dentate nuclei (Table S4).We also found an agreement between FW and FISO in between-group differences in the cerebellar lobules (both elevated in FRDA) and SCP.FW alone, however, was significantly elevated in the brainstem, dentate nuclei, ICP, and MCP.Our whole-brain voxel-wise FISO group differences also generally overlapped with FW outcomes, but with fewer significant voxels.Similar discrepancies in the sensitivity of these two metrics have previously been noted in a recent study involving individuals with Parkinson's disease, whereby FW, but not FISO, was also increased in the dentate nucleus and MCP. 19Taken together, these metrics seem to have a general correspondence and thus reflect similar aspects of the DWI signal, as expected, but in our sample, FW appears to be more sensitive to FRDA pathology than FISO.Differences between the btDTI and NODDI outcomes may relate to how the diffusion signal is apportioned to the two or three compartments of the models, respectively.For example, in the dentate nuclei, the btDTI model suggested an extra-cellular pathology (FW was elevated, but there were no effects for FA FW ), whereas NODDI posited cellular effects (FICVF and ODI were reduced in the absence of increased FISO).The discrepant outcomes between diffusion models may reflect the cellular heterogeneity of the dentate nuclei and sensitivity of these models to particular tissue types, combined with the complex neuropathy that accompanies FRDA.Of the regions examined in this study, the dentate nucleus poses a particular challenge to imaging-based measures of pathology due to its complex cellular structure, which consists of a dense, thin, highly convoluted ribbon of gray matter surrounding a white matter core.As such, future studies are necessary to further resolve these discrepancies, but our findings indicate that different biophysical models of the diffusion signal are not entirely interchangeable when studying FRDA, and potentially other neurodegenerative pathologies.
In addition to measures of free-water, btDTI and NODDI models provide more biophysically principled measures of cellular integrity relative to conventional DTI metrics, such as fractional anisotropy (FA).Here, we identified significant decreases in the traditional FA metric, FA u , across all areas of the brainstem, cerebellar lobules, and cerebellar peduncles in people with FRDA, largely replicating previous reports. 7,45However, when correcting for the confounding influence of local increases in free-water, effect sizes in these regions were diminished.As has been widely reported elsewhere, FA after correcting for free-water is more specifically representative of white matter changes. 14,23As such, the traditional FA measure may conflate multiple processes (white matter integrity loss and free-water increases), potentially making it a more discriminative disease biomarker (higher between-group effect sizes), but obscuring biological interpretation relative to free-water corrected values.Similar to FA u , FICVF was elevated in individuals with FRDA with large effects in the brainstem and cerebellar peduncles.FICVF has been shown to be sensitive to diffuse axonal loss and secondary demyelination in other neurodegenerative diseases (such as MS). 24,46,47ur results showed that microstructural metrics from all diffusion models were sensitive to changes in whitematter regions typically involved in early stages of FRDA, most notably the SCP. 3 Alterations in graymatter areas that are more typical of later stages of the disease were less evident.Interestingly, the pattern of changes in the pons was unique, with a decrease in ODI alongside nonsignificant FICVF, suggesting axonal disorganization rather than volume loss in this area.This is consistent with the relative volumetric and microstructural sparing of the pons (and MCP) in comparison with the substantial changes evident in the medulla (and ICP) and midbrain (and SCP) in FRDA, 3,7 alongside the known breakdown of the corticospinal tract, which passes through the pons. 3verall, we found that microstructural metrics from both bi-tensor and NODDI models mapped onto similar brain regions in our sample and were particularly sensitive to signal loss in white-matter regions of the cerebellum and brainstem.Further effects were observed in supratentorial regions in the whole-brain voxel-wise analysis, whereby regions of the cerebellothalamo-cortical tract, corpus callosum, and thalamic radiations were shown to have greater intracellular integrity (FICVF) in controls relative to patients (Fig. 2), consistent with past literature. 7,44,48Interestingly, none of the extracellular metrics revealed substantial alterations in supratentorial regions, suggesting that the presence of free-water may be dependent on factors such as degree of white matter atrophy and disease staging.
Our study found a significant negative correlation between Disease Burden score and FA u in the superiorposterior cerebellar lobe, which additionally showed strong, yet nonsignificant positive correlations with FW and FISO.These correlations provide some indication of progressive microstructural declines that occur over the course of FRDA.Our results suggest that the posterior cerebellar lobe suffered from the most pronounced loss of general microstructural integrity (FA u ) coupled with elevated free-water (FW and FISO).This is in contrast to recent longitudinal analyses in FRDA cohorts, which typically show greatest atrophy in white-matter tracts and the dentate nucleus, 44 although both FW and FISO also showed medium effects in the dentate nuclei, SCP and ICP in the current analysis.Although we failed to find significant correlations between the free-water metrics and clinical ataxia severity (mFARS), medium effect sizes were evident for FW in the anterior and posterior cerebellar lobules and dentate nuclei, and for FISO in the anterior and superior-posterior lobes.Our general lack of significant findings may be attributable to poor statistical power resulting from our modest sample size, but the observed trends represent important preliminary evidence for a possible relationship between free-water in key regions of the cerebellum and disease status in FRDA that warrant future follow-up.
Several additional limitations and potential areas for further work must be considered.As the design was cross-sectional, no firm inferences could be made on the progressive nature of these indices in the context of FRDA.Although we speculate that elevations of freewater and microstructural change may interact with disease duration and tissue type, we must be cautious when extending imaging outcomes to underlying biological mechanisms, and corroborate these preliminary findings with longitudinal analyses.Future studies should additionally consider how markers of neuroinflammation (eg, [18F]-FEMPA binding and quantification of blood plasma inflammatory cytokines) and regional volume correlate with free-water volume in key brain sites.Our sample was also small and restricted to adults with an average disease duration of more than 13 years.As such, replication of this study in independent cohorts is necessary to ensure the generalizability of our findings.Understanding the expression and evolution of these measures earlier in the disease course will be essential to understand disease progression and allow for broader generalizability.Larger, longitudinal studies will be necessary to resolve these issues in future work.
Taken together, the current study presents compelling evidence for the sensitivity of free-water to neuropathology in FRDA, and its overall advantage over traditional microscopic measures.Multi-compartment diffusion models should be employed as standard approaches in future FRDA research, both for their potentially increased sensitivity to disease expression and progression, and for their more targeted biological interpretability.These metrics have the potential to be used to identify therapeutic targets in individuals with FRDA and supplement existing measures of disease severity.

FIG. 1 . 2 )
FIG. 1. Region of interest (ROI) between-group (Friedreich ataxia vs. controls) effect size maps (η p 2 ) controlling for sex and age.Mean effect size across all ROIs is shown in bottom left corner of each panel.Regions showing significance P < 0.005 are denoted by a blue asterisk.See Table S2 for numeric values.Metrics: FW, bi-tensor free-water volume; FA FW , bi-tensor free-water corrected FA; FICVF, NODDI intracellular volume fraction; FISO, NODDI extracellular/isotropic volume fraction; ODI, NODDI orientation and dispersion index; FA u , free-water un-corrected FA.Sample size: 14 participants in each group.Brain regions, as depicted on the right: SCP, superior cerebellar peduncle; MCP, middle cerebellar peduncle; ICP, inferior cerebellar peduncle; Ant.Lobe, anterior lobe; SP Lobe, superior posterior lobe; IP Lobe, inferior posterior lobe.[Color figure can be viewed at wileyonlinelibrary.com]

FIG. 2 .
FIG. 2. Whole-brain voxel-level group comparisons (Friedreich ataxia [FRDA] vs. controls) for each metric.Only voxels that survived voxel-level p FWE < 0.05 are depicted.Larger values in FRDA versus controls are shown in red, smaller in green.Slice numbers in MNI space are shown in white text.Key: FW, bi-tensor free-water volume; FA FW , bi-tensor free-water corrected FA; FICVF, NODDI intracellular volume fraction; FISO, NODDI extracellular/isotropic volume fraction; ODI, NODDI orientation and dispersion index; FA u , free-water un-corrected FA.Sample size: 14 participants in each group.[Color figure can be viewed at wileyonlinelibrary.com]

TABLE 1
Descriptive statistics for each clinical scale related to individuals with FRDA (n = 14) 25te: mFARS, Modified Friedreich Ataxia Rating Scale; GAA1, size of the FXN GAA repeat length on the smaller allele; GAA2, size of the FXN GAA repeat length on the larger allele; age of onset, age at which participant/carer first noted symptoms; disease duration, current ageÀage of onset; disease burden, disease duration Â GAA125; SEM, standard error of the mean.