Meta-analysis of the normal diffusion tensor imaging values of the median nerve and how they change in carpal tunnel syndrome

Carpal tunnel syndrome (CTS) leads to distortion of axonal architecture, demyelination and fibrosis within the median nerve. Diffusion tensor imaging (DTI) characterises tissue microstructure and generates reproducible proxy measures of nerve ‘health’ which are sensitive to myelination, axon diameter, fiber density and organisation. This meta-analysis summarises the normal DTI values of the median nerve, and how they change in CTS. This systematic review included studies reporting DTI of the median nerve at the level of the wrist in adults. The primary outcome was to determine the normal fractional anisotropy (FA) and mean diffusivity (MD) of the median nerve. Secondarily, we show how the FA and MD differ between asymptomatic adults and patients with CTS, and how these differences are independent of the acquisition methods. We included 32 studies of 2643 wrists, belonging to 1575 asymptomatic adults and 1068 patients with CTS. The normal FA was 0.58 (95% CI 0.56, 0.59) and the normal MD was 1.138 × 10–3 mm2/s (95% CI 1.101, 1.174). Patients with CTS had a significantly lower FA than controls (mean difference 0.12 [95% CI 0.09, 0.16]). Similarly, the median nerve of patients with CTS had a significantly higher mean diffusivity (mean difference 0.16 × 10–3 mm2/s [95% CI 0.05, 0.27]). The differences were independent of experimental factors. We provide summary estimates of the normal FA and MD of the median nerve in asymptomatic adults. Furthermore, we show that diffusion throughout the length of the median nerve becomes more isotropic in patients with CTS.

Search strategy. The NICE Healthcare Databases (hdas.nice.org.uk) was searched according to Appendix 1 (Supplementary Materials) on 9th October 2020. The medRxiv and bioRxiv preprint archives were searched with the same strategy using the R package medrxivr 14 . Later, CitationChaser 15 was used for forward and backward citation chasing based on the final list of included studies (eFigure 1).

Study selection.
Three review authors (DR, JR and FR) independently screened titles and abstracts for relevance, in accordance with the eligibility criteria. The full texts of potentially eligible articles were obtained and again independently assessed by the same authors. Disagreements were resolved by discussion with RGW. The reasons for excluding studies are outlined in Appendix 2 (Supplementary Materials).
Data extraction. Three review authors (DR, JR and FR) independently double extracted all data. Thereafter, all datapoints were independently checked for accuracy by RGW. DTI parameters were extracted from the following anatomical levels of the median nerve: the distal radio-ulnar joint (DRUJ), the pisiform and the hook of the hamate, as these are three commonly used imaging landmarks which equate to the inlet, mid-point and outlet of the carpal tunnel. The nerve/hand was the unit of analysis 16 . Many studies reported both the number of individuals and wrists scanned (as some studies involved bilateral scanning) but if not otherwise stated we assumed imaging was performed unilaterally. If data were missing, unclear or present in an unfavourable format then the authors were contacted by email with a request for more information. Four authors provided additional information upon request [17][18][19][20] . When no reply was received, estimates were derived from graphs or imputed where possible 21 . In healthy nerves, the axons are enveloped by myelin sheaths and arranged relatively tightly, which restricts the diffusion of water to the long axis of the nerve. Chronic compression leads to distortion of the axonal architecture, demyelination and as such, degradation of physiological barriers to the diffusion of water diffusion. Consequently, more diffusion occurs perpendicular to the long axis of the nerve as water is more free to diffuse around the fibres, reducing the factional anisotropy (FA) and increasing the magnitude of diffusion (mean diffusivity, MD).
Statistical analysis. The raw data are available via the Open Science Framework (https:// osf. io/ vqwkp/). The single study performed at 7 tesla 19 was excluded from all meta-analyses given its clinical disparity. Using the meta suite of Stata v16 (StataCorp, Texas), the mean FA and MD from asymptomatic adults were pooled to estimate the normative values. We performed direct comparisons meta-analysis of the mean differences in FA and MD between asymptomatic adults and patients with CTS. Meta-analyses were subgrouped by anatomical location. Restricted maximum likelihood was used to estimate the between-study variance (tau 2 ), with the Knapp and Hartung modification. Heterogeneity was quantified by I 2 24 .Using the metafor 25 package, mixed-effects meta-regression was used to explore potential reasons for the observed heterogeneity in the direct comparisons meta-analysis of FA; the continuous covariates were age, in-plane resolution (mm 2 ), slice thickness (mm), echo time (TE in ms), b-value (mm 2 /s) and number of diffusion-sensitising gradient directions (N D ). TE and b-value were modelled as an interaction. Confidence intervals (CI) were generated to the 95% level. To investigate the possibility of small-study effects for FA between asymptomatic adults and patients with CTS, a funnel plot was constructed with the pseudo CIs contoured by tau 2 25 . EviAtlas was used to generate a map of the location of the 1st author's institution 26 .
The risk of bias for the included studies is summarised in eFigure 4. The majority of studies were at low risk of methodological bias.  Fig. 3). This difference was again most profound at the mid-point of the carpal tunnel, at the level of the pisiform (mean difference 0.16 × 10 -3 mm 2 /s [95% CI 0.05, 0.27]).
Studies reporting the use of parallel imaging techniques (e.g. GRAPPA, SENSE or ASSET) yielded 5% higher estimates of FA (β 0.05 [95% CI 0.02, 0.08]; I 2 98%, eFigure 17) when compared to studies which did not report this information. Parallel imaging methods were not associated with differences in the MD. There was insufficient data to explore different partial Fourier settings. There was no association between the number of signal averages and FA or MD (eFigures 18 and 19).
Ultimately, mixed-effects multivariable meta-regression showed that having CTS was the strongest independent moderator of the observed heterogeneity in FA (Table 1). Age explained some of the residual between-study variance. The experimental factors we modelled did not explain the residual heterogeneity.

Discussion
This study demonstrates that throughout the length of the median nerve at the wrist, patients with CTS have more isotropic diffusion than asymptomatic adults. The largest differences for both fractional anisotropy and mean diffusivity were observed at the mid-point of the carpal tunnel, at the level of pisiform where CTS patients had lower FA and higher MD. Of clinical importance, we demonstrate that these real-world differences were independent of age and experimental (acquisition) conditions. Therefore, we believe that aberrations in both FA and diffusivity could be used to identify patients with median nerve neuropathy at the wrist.
There are inherent problems with clinicians diagnosing CTS given that the constellation of symptoms and clinical signs of the syndrome, and the available tests are largely unreliable. For example, nocturnal paraesthesias and many classical tests such as Phalen and Tinel, the scratch-collapse and sensory threshold testing have poor diagnostic value 54,55 . Despite the widespread use of electrodiagnostic studies in patients with suspected carpal tunnel syndrome, it remains an invasive test which evokes pain and anxiety, and controversy still exists regarding its accuracy and the normal values 56,57 . For these reasons, surgeons still perform decompression surgery in the presence of normal provocative and electrodiagnostic tests 58 . More recently, measurement of the cross sectional area of the median nerve using ultrasound is gaining popularity given that it has good inter-rater and intra-rater reliability 59 , and validity 60,61 . Despite these benefits, ultrasound has not been adopted into routine clinical practice because several aspects remain unclear, such as (i) how the cross-sectional area is affected by other factors such as age, sex, diabetes, sonographer technique [pressure applied, measurement angle, etc.] and the hardware, (ii) whether the epineurium should be included in the measurement, and (iii) how these measurements relate to severity, subjective and objective outcomes. Moreover, sonographically derived cross-sectional area still only provides morphological information (size and shape) which is inherently insensitive to nerve function and microstructure. Therefore, considerable effort has been directed towards the development of DTI because it characterises tissue microstructure and generates reproducible 4-8 proxy measures of nerve 'health' which are sensitive to myelination, axon diameter, fibre density and organisation [9][10][11] (Fig. 1). DTI metrics outperform standard morphological imaging in patients with peripheral neuropathy 35 and are independent of age and acquisition parameters, something which cannot be said for electrodiagnostics 62 or sonography. In the UK, the cost of a non-contrast MRI of the extremity is less than an electrodiagnostic exam (£389 versus £444) but more than sonography (£220) 63      www.nature.com/scientificreports/ (anatomical) scans, contrast-free angiography, sequences which characterise muscle (fat fractions, elastography, iron deposition, etc.) and the topography of the sensorimotor cortex (e.g., using functional MRI) to determine whether there is central capacity to regenerate following peripheral nerve surgery. Ultimately, we have not examined the diagnostic test accuracy of each modality head-to-head and this must be performed before comments about relatively accuracy and cost effectiveness can be made. Overall, we suggest that DTI might provide additional valuable information for the diagnosis, grading and management of (at least unclear or complex) patients with CTS. However, incorporating DTI in the real-world management of CTS would be difficult and require significant training for clinicians, changes to infrastructure and clinical pathways. None-the-less, we show that DTI yields unique information about the 'health' of the median nerve which could of significant clinical value. Initially, this technology could be used in patients with an unclear diagnosis or bilateral symptomatology, and those who don't improve after treatment. We observed high statistical heterogeneity which has many potential explanations. The majority of the (statistical) heterogeneity was explained by the presence of CTS and it is plausible that the remainder is explained by the 'severity' of disease, which we were unable to capture. For example, we speculate that patients with more severe CTS (e.g., symptoms for years, resulting in profound demyelination, axonal loss and fibrosis) are likely to have lower FA and higher MD than patients with recent-onset mild CTS. Age also explained some of the observed heterogeneity and this is unsurprising, given that FA is known to fall in aging peripheral nerves 64 , just as it does in the white matter tracts of the brain 65,66 . This is because aging axons lose integrity, undergo demyelination and there is a simultaneous increase in extra-cellular fluid. Importantly, we showed that DTI metrics were sensitive to CTS after adjusting for age. Finally, in highly controlled and extreme conditions, user-specified factors 67 such as the SNR 68 , b-value 69,70 , N D 71,72 , distortion correction pipelines 73,74 , tensor fitting methods 68 and partial volume effects 75 have been shown affect the DTI parameter estimates, which may explain some of the remaining heterogeneity. Although we could not explore the effects of all these factors, in general we observed that experimental conditions had little or no significant effect on the measured FA or MD. Therefore, despite the statistical heterogeneity, DTI appears to be reliably sensitive to the microstructural changes of the median nerve which occur in CTS.
There were no significant associations between FA and MD, and several core elements of the pulse sequence. Therefore, we suggest that clinicians and researchers wishing to acquire DTI could optimise their sequence as follows. As tensors are robust to varying b-values (in the hindered range) we suggest a b-value of 300-800 mm 2 /s; smaller b-values enable a shorter TE, which improves SNR and mitigates the effects of T2 shine-through at the expense of less diffusion-weighting. Reducing the TE might also enable users to take advantage of other vendor-specific options to improve data quality and reduce distortions. Given that the median nerve has no crossing fibres to model, it is not tortuous (within or between voxels) and increasing the N D has little effect on simple tensor fitting 76 , we see no reason for the N D to exceed approximately 15. The normal median nerve has a cross-sectional area of 9 mm 2 (3.4 mm diameter) 77 and this increases with CTS 78 , so we recommend an in-plane resolution of approximately 1 mm 2 to ensure that at least 1 voxel is not affected by partial voluming 75 . As that the median nerve is orthogonal to the imaging plane (if data are acquired axially), the slice thickness could be comfortably increased until there is adequate SNR because it appears to have little effect on the resultant metrics. Until work is published to show the exact relaxation properties (T1, T2, T2*) of the median nerve we suggest that TR is set to approximately 4500 ms to reduce scan time. Ideally, users specify an even number of signal averages (full datasets), divided equally between opposing phase-encoding directions (i.e. 1 signal acquisition blip-up and 1 blip-down or 2 averages blip-up and 2 blip-down, etc.) as this would allow offline concatenation and exploitation of the various corrections available in the FMRIB Software Library (FSL) 79 . We advocate capturing such data via ssEPI because it is more time efficient than current implementations of rsEPI and distortions associated with ssEPI can be ameliorated using various softwares. It should be noted that thicker slices, longer repetition times and more signal averages are associated with lower estimates of MD, if this is of importance to users.

Limitations.
The main limitation of this study is the inherent and pervasive problem of CTS diagnosis which may have biased the findings. At present there is no internationally agreed diagnostic criteria for CTS and as such, there is clinical variation which is present in the includes studies. We planned to capture disease severity from the original studies, but this information was not available. As a matter of urgency, the community should work towards a consensus on objective criteria which constitute a diagnosis of CTS. Thereaftere, a reference standard for the diagnosis can be defined and this would enable studies of diagnostic test accuracy to be done, comparing DTI to other available tests such as electrophysiology, ultrasound and more.
It is widely known that diffusion metrics in the brain are strongly dependent on preprocessing pipelines (i.e., software) 79 but still there is no consensus on the minimum or indeed ideal suite of corrections to perform. This issue is compounded in the limb owing to an absence of research on the topic and hardware limitations. The majority of the included studies did not describe any form of distortion correction, how the diffusion data were reconstructed or how metrics were extracted from the median nerve. Before DTI can be used clinically, variations in these pipelines should be tested and a universal pipeline and standards for reporting diffusion data should be agreed by consensus. Some readers will criticise our choice to pool estimates of FA and MD in the presence of high statistical heterogeneity. This was done because forest plots provide an important graphical representation of the variability of measurements in relation to experimental conditions (e.g., b-values and N D ), they summarise a large amount of information in an easy-to-interpret format and moreover, meta-regression facilitates the exploration of heterogeneity.