Diffusion Tensor Imaging Adds Diagnostic Accuracy in Magnetic Resonance Neurography Michael O. Breckwoldt, MD, PhD,* Christian Stock, PhD, MSc,† Annie Xia,* Andreas Heckel, MD, MSc,* Martin Bendszus, MD,* Mirko Pham, MD,* Sabine Heiland, PhD,‡ and Philipp Bäumer, MD, MSc* Objective: The aim of this study was to determine whether quantitative diffusion tensor imaging (DTI) adds diagnostic accuracy in magnetic resonance neurography. Materials and Methods: This prospective study was approved by the institutional review board. We enrolled 16 patients with peripheral polyneuropathy of various etiologies involving the upper arm and 30 healthy controls. Magnetic resonance neurography was performed at 3 T using transverse T2-weighted (T2-w) turbo spin echo and spin echo planar imaging diffusion-weighted sequences. T2-weighted normalized signal (nT2), fractional anisotropy (FA), apparent diffusion coefficient (ADC), radial diffusivity (RD), and axial diffusivity (AD) of the median, ulnar, and radial nerves were quantified after manual segmentation. Diagnostic performance of each separate parameter and combinations of parameters was assessed using the area under the receiver operating characteristic curve (AUC). Bootstrap validation was used to adjust for potential overfitting. Results: Average nT2, ADC, RD, and AD values of the median, ulnar, and radial nerve were significantly increased in neuropathy patients compared with that in healthy controls (nT2, 1.49 ± 0.05 vs 1.05 ± 0.05; ADC, 1.4  10−3 ± 2.8  10−5 mm2/s vs 1.1  10−3 ± 1.3  10−5 mm2/s; RD, 9.5  10−4 ± 2.9  10−5 mm2/s vs 7.2  10−4 ± 1.3  10−5 mm2/s; AD, 2.3  10−3 ± 3.7  10−5 mm2/s vs 2.0  10−3 ± 2.2  10−5 mm2/s; P < 0.001 for all comparisons). Fractional anisotropy values were significantly decreased in patients (0.51 ± 0.01 vs 0.59 ± 0.01; P < 0.001). T2-weighted normalized signal and DTI parameters had comparable diagnostic accuracy (adjusted AUC: T2-w, 0.92; FA, 0.88; ADC, 0.89; AD, 0.84; RD, 0.86). Combining DTI parameters significantly improved the diagnostic accuracy over single-parameter analysis. In addition, the combination of nT2 with DTI parameters yielded excellent adjusted AUCs up to 0.97 (nT2 + FA). Conclusions: Diffusion tensor imaging has high diagnostic accuracy in peripheral neuropathy. Combining DTI with T2 can outperform T2-w imaging alone and provides added value in magnetic resonance neurography. Key Words: magnetic resonance neurography, diffusion tensor imaging, upper arm neuropathy (Invest Radiol 2015;50: 498–504)


eripheral neuropathies encompass demyelinating and axonopathic diseases of the peripheral nervous system with a variety of underlying causes.1 Magnetic resonance neurography (MRN) is an increasingly used diagnostic tool to visualize peripheral nerve pathology including

inflammatory, metabolic, degenerative, compressive, traumatic, or neoplastic conditions.2–6 Several studies have established distinct MRN criteria in various diseases at specific anatomical sites.1,7–10 Spatial identification of nerve lesions for some neuropathies has previously not been possible using conventional clinical and electrophysiological diagnostic tools.3 In imaging studies, T2-weighted (T2-w) imaging is currently the criterion standard for lesion detection and allows the assessment of spatial distribution and severity with good sensitivity and specificity.2,3,7 Recent developments of MRN include more sophisticated “functional” imaging methods that allow the assessment, for example, of the blood-nerve barrier and nerve perfusion imaging in vivo.11 Diffusion tensor imaging (DTI) offers another innovative approach to probe the cellular pathophysiology of disease and microstructural alterations on fiber level.12–17 Diffusion tensor imaging allows the quantification of fractional anisotropy (FA), apparent diffusion coefficient (ADC), and the molecular diffusion parameters axial diffusivity (AD), and radial diffusivity (RD). Although FA and ADC are established parameters in central nervous system imaging, their role and in particular their diagnostic value in peripheral nerve imaging is less clear. So far, DTI has been mainly applied in compression neuropathies with moderate diagnostic accuracy.16,18–20 However, the diagnostic accuracy of DTI in noncompressive neuropathies has seldom been evaluated.21–24 For MRN, a central question in clinical routine is whether DTI provides any added diagnostic value or whether it can be used as a quantitative marker independent from the established T2 signal increase. A recent study demonstrated that DTI is sufficiently sensitive and specific to detect first signs of subclinical demyelination in healthy subjects.13 However, an added value of DTI in clinically manifest neuropathy has not been clearly shown. We hypothesized that DTI might provide such added value in the characterization of nonfocal neuropathies by improving diagnostic accuracy. Therefore, we systematically assessed the diagnostic accuracy of each DTI parameter (FA, ADC, AD, and RD) in peripheral neuropathy of the upper extremity, compared with their performance with the current T2-w criterion standard, and tested whether their combination could further increase diagnostic accuracy.

MATERIALS AND METHODS Study Design and Participants

Received for publication January 14, 2015; and accepted for publication, after revision, February 18, 2015. From the *Department of Neuroradiology, Heidelberg University Hospital, †Institute of Medical Biometry and Informatics, and ‡Section of Experimental Radiology, Heidelberg University Hospital, University of Heidelberg, Heidelberg, Germany. Conflicts of interest and sources of funding: M.O.B., A.H., and P.B. were supported by a postdoctoral stipend (physician-scientist fellowship) by the Medical Faculty of the University of Heidelberg, Germany. M.P. is supported by a memorial stipend from the Else-Kröner-Fresenius foundation and has received a project grant from the EFSD/JDRF/Novo Nordisk Foundation. The authors report no conflicts of interest. Supplemental digital contents are available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal's Web site (www.investigativeradiology.com) Reprints: Michael O. Breckwoldt, MD, PhD, Department of Neuroradiology, Heidelberg University Hospital, Heidelberg University, Im Neuenheimer Feld 400, D-69120, Heidelberg, Germany. E-mail: [email protected] Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. ISSN: 0020-9996/15/5008–0498



This prospectively planned, cross-sectional study was approved by the local ethics committee and was conducted between August 2012 and June 2013. Written informed consent was obtained from all study participants. We included a total of 46 participants in this study: 16 patients with peripheral mononeuropathy or polyneuropathy (8 women; mean age, 51.1 ± 4.7 years; age range, 43–58 years and 8 men; mean age, 50.2 ± 6.2 years; age range, 42–63; see also Table 1 for patients' details) and 30 healthy volunteers (14 women; mean age, 43.5 ± 18.1 years; age range, 20–64 years and 16 men; mean age, 38.2 ± 18.7 years; age range, 24–70 years). Patients were enrolled consecutively with inclusion criteria of a diagnosis of peripheral neuropathy of the upper extremity confirmed by objective clinical and electrophysiological findings. Exclusion criteria for both patients and controls were a history of compression neuropathy and previous trauma to the upper extremity. For control subjects, interviews excluded upper extremity pain, paresthesia, or history of neurologic disease. Investigative Radiology • Volume 50, Number 8, August 2015

Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved.

Investigative Radiology • Volume 50, Number 8, August 2015

Diffusion Tensor Imaging Adds Accuracy in MRN

TABLE 1. Patient Details



Age, y

Clinical Diagnosis

Most Affected Nerve

1 2 3 4 5 6 7 8 9 10 11


43 47 47 52 52 49 51 50 51 46 54

Anterior interosseous syndrome Inflammatory polyneuropathy Plexus neuritis Paraneoplastic polyneuropathy Diabetic polyneuropathy Plexus neuritis Postsurgical polyneuropathy Inflammatory polyneuropathy Parsonage-Turner syndrome Parsonage-Turner syndrome Inflammatory polyneuropathy

Median Median, ulnar Median, ulnar, radial Median, radial Median, ulnar Ulnar Median, ulnar Median, ulnar, radial Median Median, ulnar,radial Radial, ulnar

12 13


50 58

Median, ulnar, radial Median, ulnar, radial

14 15


56 63

Anterior interosseous syndrome Hereditary neuropathy (Charcot-Marie-Tooth type 1) Inflammatory polyneuropathy Inflammatory polyneuropathy




Parsonage-Turner syndrome

Median Median, ulnar Median

Muscle Atrophy Yes (pronator teres/quadratus, flexor pollicis longus) Yes (pronator flexor pollicis longus) None Yes (extensor digitorum, extensor carpi radialis) No No Yes (pronator teres) No Yes (supraspinatus/infraspinatus, brachialis) Yes (pronator teres, biceps, triceps) Yes (biceps, flexor carpi ulnaris, digitalis profundus, extensor digitorum, extensor carpi radialis) No No Yes (musculus supinator) Yes (triceps, flexor digitorum profundus, extensor carpi ulnaris, extensor pollicis longus, anconeus) Yes (biceps, triceps)

F indicates female; M, male.

Magnetic Resonance Imaging

MRI Postprocessing

Magnetic resonance imaging (MRI) examinations were performed on 3 Tesla systems (Magnetom TIM Trio and Verio; Siemens Healthcare, Erlangen, Germany). T2-weighted and diffusion-weighted sequences covered an identical region of the upper arm. Subjects were examined in the prone position with the arm extended centered in a knee 15-channel transmit/receive phased array radiofrequency coil. To minimize artificial signal increase in the T2-w images related to the so-called magic angle effect,25 the longitudinal axis of the arm was aligned at an angle of less than or equal to 10 degrees relative to the B0 field direction. Parameters for the transverse T2-w turbo spin echo sequence were as follows: repetition time, 7020 milliseconds; echo time, 52 milliseconds; spectral fat saturation; number of slices, 45; slice thickness, 3.0 mm; interslice gap, 1.2 mm; field of view, 150  150 mm2; acquisition matrix, 512  358; in-plane resolution, 0.25  0.25 mm2; number of excitations, 3; and acquisition time, 6:23 minutes. For DTI, we customized a spin echo planar imaging sequence from the Advanced Diffusion work-in-progress package developed by Siemens Healthcare (ASP 511 E). Parameters were optimized for peripheral nerve imaging as published previously13; these are as follows: repetition time, 3800 milliseconds; echo time, 94 milliseconds; b values, 0 and 1200 s/mm2 (encoded in monopolar 19 directions); noise threshold, 20; readout bandwidth, 1395 Hz/pixel; number of slices, 15–18; slice thickness, 4 mm; interslice gap, 1.2 mm; field of view, 150  150 mm2; acquisition matrix, 128  128; in-plane resolution, 1.14  1.14 mm2; and number of excitations, 2. Generalized autocalibrating partially parallel acquisition of 2 (38 reference lines) and a phase partial fourier of 7/8 were used (acquisition time, 2:53 minutes). Imaging slabs were centered at identical positions for both sequences. Scans were repeated if motion artifacts occurred. This was the case in 3 (~6%) of the 45 subjects. One control subject was excluded from the analysis because of persisting motion artifacts. The scanned area included the upper arm ~10 to 15 cm proximal to the elbow (scan field ~10 cm, for an illustration, see Fig. 1).

Diffusion tensor imaging postprocessing was performed using FSL diffusion toolbox of the FMRIB free software library (FSL), version 5.0.6 (FSL, Oxford, United Kingdom, http://fsl.fmrib.ox.ac.uk/fsl). The diffusion tensor model was fitted on the diffusion weighted imaging images yielding maps of the first, second, and third eigenvalues. From these images, parameter maps of FA, mean diffusivity (subsequently called apparent diffusion coefficient, ADC as this term is more common in clinical routine), RD, and AD were computed. Specifically, FA corresponds to a normalized standard deviation of the 3 eigenvalues (see standard formula in Basser and Pierpaoli26), AD is identical to the first eigenvalue, RD equals the average of the second and third eigenvalue, and ADC equals the average of all 3. Manual segmentation masks of upper arm nerves was performed in 15 to 18 slices per scan on b = 0 images by 1 reader (M.O.B.), blinded to the history of the patient, using FSL's FSLView. T2weighted images were considered for anatomical reference in cases of doubt. Care was taken to avoid partial volume artifacts and to exclude surrounding fat, vessels, or muscles from the segmentation masks, which were used to extract DTI parameter values (FA, AD, ADC, and RD) using customized FSL scripts. Likewise, nerves were segmented at identical positions in T2-w images by a blinded reader (A.X.) to determine nerve cross-sectional area (CSA) and T2-w signal intensities. An additional ROI was placed in the biceps muscle for T2-w signal normalization. If signs of biceps denervation were present (apparent in 3 of 16 patients), the ROI was placed in the triceps muscle or in the least affected part of the muscle. T2-weighted images were analyzed with OsiriX DICOM viewing software, and the nerve signal normalized to muscle was calculated as follows:

© 2015 Wolters Kluwer Health, Inc. All rights reserved.

nT2 ¼ T2 − w signalnerve = T2 − w signalmuscle:

Statistical Analysis For each nerve, the average over the segmented 15 to 18 slices was calculated and used for statistical analysis. We compared mean www.investigativeradiology.com

Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved.


Investigative Radiology • Volume 50, Number 8, August 2015

Breckwoldt et al

FIGURE 1. Scheme on the left shows the scanned anatomical region of the upper arm. Representative cross-sectional T2-w images, FA maps, and colored FA (col-FA) maps of a control subject (A) and a patient (B) with peripheral neuropathy. Insets show magnifications of the median, ulnar, and radial nerve (dotted outline). Note the fascicular swelling and signal increase of the nerves (T2-w insets). FA images show the decreased anisotropy in the affected nerves (FA insets). In this patient, mainly the median and ulnar nerves are affected. Figure 1 can be viewed online in color at www.investigativeradiology.com.

values for all parameters of interest (nerve CSA, T2-w normalized signal [nT2], FA, ADC, AD, and RD) across the 2 groups using standard descriptive statistics and unpaired Student t test. Mean values ± SEM are shown unless stated otherwise. For the assessment of diagnostic accuracy and further statistical analysis, mean values of nerves were pooled even though not all nerves were affected to the same degree; that is, 1 of the 3 nerves might have shown normal values in a given patient. Pearson correlation coefficients were calculated to assess the relationships between parameters. The area under the receiver operating characteristic (ROC) curve (AUC) was calculated from binary logistic regression models with disease status as the outcome and the parameters as explanatory variables. In addition to single-parameter models, multiparameter models were fitted to assess the joined diagnostic capabilities when multiple parameters are taken into account. Confidence intervals (CIs) for the AUC were computed using the method of Delong et al.27 Model comparison was performed by likelihood ratio tests in case of nested models and by the Akaike information criterion (AIC) in case of nonnested models. The AIC is a measure of the relative quality of statistical models that penalizes the number of parameters in the model, thus more complex models are not necessarily considered as “better” because they provide better fit to the data (lower values of the AIC indicate better models). Internal validation of the classification models was performed using the 0.632+ bootstrap algorithm (with 500 bootstrap samples) to adjust the apparent AUC for overfitting and to increase generalizability of the results.28 Bootstrap validation is considered to have several statistical advantages over cross validation and split-sample techniques for internal validation.29 Confidence intervals for bootstrap-corrected AUCs are not reported because, to the author's knowledge, no established method exists for their calculation. A sensitivity analysis was performed to assess the impact of age differences in the 2 groups specifically in models involving FA, which is known to be negatively correlated with age. We determined an ageadjusted value of FA that we labeled FA*. This value was calculated as the mean FA plus the observed FA minus the predicted FA from an 500


age-adjusted linear regression model. We performed single-parameter and multiparameter models for FA* as described previously. All statistical tests were performed 2-sided. Because the study was still exploratory in design, and multiple tests were performed, the reported P values can generally be interpreted only descriptively. For the main analyses, a conservative Bonferroni correction was applied (5 tests of model coefficients in single-parameter models and 7 likelihood ratio tests in multiparameter models). To assure an overall type 1 error rate of 5%, the obtained P values were compared with the nominal significance level of 0.05/12 = 0.0041. The analyses were performed using the commercial software package PRISM (GraphPad) and the R language and environment for statistical computing (version 3.1.1, July 10, 2014).30 The R packages pROC (version 1.7.3) and Daim (version 1.1.0) were used to visualize ROC curves and to internally validate AUC estimates, respectively.31,32

RESULTS Quantification of Nerve Parameters Patients showed increased CSA of median, ulnar, and radial nerve (mean area, 9.3 ± 0.5 mm2 vs 7.2 ± 0.3 mm2 in healthy controls; P < 0.001) with visually appreciable increased T2-w signal and decreased FA values at upper arm level (Fig. 1, A and B). Different degrees of pathological involvement could be discerned for each nerve. In addition, some patients (~62%) showed acute signs of muscle denervation as indicated by an increased T2 signal of target muscles (Supplementary Fig. 1, A–D, Supplemental Digital Content 1, http://links.lww.com/RLI/A200). Quantitative analysis showed increased average nT2 in all 3 nerves (nT2: 1.49 ± 0.05 vs 1.05 ± 0.05; P < 0.001; Table 2), whereas the mean FA values were decreased in the corresponding anatomical region (0.51 ± 0.01 vs 0.59 ± 0.01; P ≤ 0.05 for all nerves; Fig. 2, A and B). The mean values for the other diffusion parameters ADC, AD, and RD were significantly increased in patients (ADC: 1.4  10−3 mm2/s vs 1.1  10−3 mm2/s; RD: 9.5  10−4 mm2/s vs 7.2  10−4 mm2/s; AD: © 2015 Wolters Kluwer Health, Inc. All rights reserved.

Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved.

Investigative Radiology • Volume 50, Number 8, August 2015

Diffusion Tensor Imaging Adds Accuracy in MRN

TABLE 2. Quantitative Assessment of Different Magnetic Resonance Sequences in Patients and Controls Nerve Median Ulnar Radial Average

nT2 1.6 ± 0.08 1.1 ± 0.03 1.6 ± 0.09 1.1 ± 0.02 1.2 ± 0.07 0.9 ± 0.03 1.5 ± 0.05 1.0 ± 0.02

ADC, mm2/s

FA 0.54 ± 0.01 0.60 ± 0.02 0.48 ± 0.01 0.59 ± 0.01 0.50 ± 0.01 0.59 ± 0.02 0.51 ± 0.01 0.59 ± 0.01


AD, mm2/s −5

1.1  10 ± 5.3  10 1.0  10−3 ± 2.1  10−5 1.4  10−3 ± 4.3  10−5 1.1  10−3 ± 2.1  10−5 1.3  10−3 ± 4.5  10−5 1.0  10−3 ± 1.9  10−5 1.4  10−3 ± 2.8  10−5 1.1  10−3 ± 1.3  10−5


RD, mm2/s −5


2.3  10 ± 6.3  10 2.0  10−3 ± 3.2  10−5 2.3  10−3 ± 7.2  10−5 2.0  10−3 ± 3.4  10−5 2.2  10−3 ± 5.4  10−5 1.8  10−3 ± 3.2  10−5 2.3  10−3 ± 3.7  10−5 2.0  10−3 ± 2.2  10−5

CSA, mm2 −5

9.0  10 ± 5.7  10 7.3  10−4 ± 2.4  10−5 9.7  10−4 ± 4.3  10−5 7.4  10−4 ± 2.1  10−5 9.8  10−4 ± 5.1  10−5 6.7  10−4 ± 2.3  10−5 9.5  10−4 ± 2.9  10−5 7.2  10−4 ± 1.3  10−5

11.5 ± 0.9 8.5 ± 0.3 8.8 ± 0.9 5.9 ± 0.2 7.5 ± 0.6 4.6 ± 0.2 9.3 ± 0.5 6.3 ± 0.2

Upper row, patients; lower row, control subjects. Values are expressed as mean ± SEM; P < 0.01 for all comparisons. nT2 indicates T2-weighted normalized signal; FA, fractional anisotropy; ADC, apparent diffusion coefficient; AD, axial diffusivity; RD, radial diffusivity; CSA, cross-sectional area.

2.3  10−3 mm2/s vs 2.0  10−3 mm2/s; P ≤ 0.01 for all nerves; Fig. 2, C–E, Table 2). T2-weighted normalized signal and FA values correlated and demonstrated a clear separation between patient and control population (R = −0.55, CI −0.72 to −0.30; P < 0.001; Fig. 2F). This correlation was weaker for nT2 and FA than for nT2 and ADC, AD, or RD, respectively, suggesting that of all diffusion parameters, FA might have most additional explanatory and discriminative power compared with T2-w (Table 3).

Assessment of Diagnostic Accuracy Single-Parameter Models T2-weighted normalized signal by itself had high diagnostic accuracy as quantified by ROC analysis (AUC, 0.95; 95% CI, 0.86–1; Table 4, Fig. 3A). Individual DTI parameters showed a comparable but slightly lower diagnostic accuracy (AUC for FA, 0.92 [95% CI, 0.84–1]; ADC, 0.91 [95% CI, 0.80–1]; RD, 0.89 [95% CI, 0.78–1]; and AD, 0.87 [95% CI, 0.75–1]). In single-parameter models, all parameters were statistically significantly associated with disease status at a nominal significance level of 0.0041. In addition to the highest AUC, nT2 was also associated with the lowest AIC value, indicating its highest predictive value.

Multiparameter Models We further assessed whether the combination of different imaging parameters can yield additional diagnostic accuracy. In models consisting of 2 DTI parameters, the diagnostic accuracy could be improved in comparison to models with only 1 DTI parameter. Results for 2-parameter combinations containing FA were most favorable (eg, FA + ADC, AUC of 0.97; Fig. 3B, Table 4). After Bonferroni correction, the added value of ADC and AD was still statistically significant; however, this was not the case for RD. The combination of FA + ADC as well as FA + AD yielded also lower AIC values than nT2 alone suggesting a superior diagnostic performance of the combined diffusion parameters. Because of the limited sample size, models with more than 2 parameters could not be fitted. Adding individual DTI parameters to nT2 significantly increased the diagnostic accuracy compared with a model with nT2 alone. For all parameters apart from AD, these results remained statistically significant even after Bonferroni correction. Combinations of nT2 + FA, nT2 + ADC, nT2 + AD, and nT2 + RD were tested (Fig. 3C, Table 4). The combination nT2 + FAyielded a perfect separation of cases and controls in the present study population with an AUC of 1. In addition, the AIC values of the models nT2 + FA, nT2 + ADC, and nT2 + AD indicated substantial improvements over a model with nT2 alone. Although this particular model seems promising, given its zero deviance for nT2 + FA, the results need to be interpreted with some caution.

FIGURE 2. Box plots of nT2 (A), FA (B), ADC (C), AD (D), and RD (E) for patients (black boxes) and controls (gray boxes). Correlation analysis (Pearson) of mean FA and nT2 for the study cohort (F). Dotted ovals indicate the 2 populations of patients versus controls that are well separated. *P < 0.05, **P < 0.01, ***P < 0.001.

© 2015 Wolters Kluwer Health, Inc. All rights reserved.


Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved.


Investigative Radiology • Volume 50, Number 8, August 2015

Breckwoldt et al

TABLE 3. Pearson Correlation Coefficients of Diffusion Parameters With nT2






−0.55 −0.72 to −0.30

0.70 0.5 to 0.82

0.61 0.38 to 0.77

0.66 0.45 to 0.80

Upper row, Pearson R; lower row, 95% confidence interval. Values are expressed as mean ± SEM; P < 0.001 for all correlations. nT2 indicates T2-weighted normalized signal; FA, fractional anisotropy; ADC, apparent diffusion coefficient; AD, axial diffusivity; RD, radial diffusivity.

Effect of Age As in previous reports,16,20,33 we found that FA decreases moderately with age (Pearson correlation coefficient, −0.38; P = 0.01). T2-weighted normalized signal, ADC, and AD did not correlate with age (P > 0.05). To determine to which extent the difference in age between our patient (mean age, 51 years) and control group (mean age, 43 years) impacts our results, we repeated the analyses that involved FA using age-adjusted FA values (FA*). The sensitivity analysis yielded an AUC estimate of 0.92 in a single-parameter ROC analysis. When we retested T2 + FA*, we found also an AUC of 1.0.

Adjustment of AUCs for Overfitting To increase the generalizability of our data, we adjusted the AUC estimates for potential overfitting. The bootstrap adjustment of the estimated AUCs resulted in a lower AUC estimate of 0.92 for nT2. The adjusted AUCs for single DTI parameters ranged from 0.84 (for AD) to 0.89 (for ADC) (Table 4). For combinations of DTI parameters, the corrected AUC was either only slightly lower or of the same order (in case of FA + AD with an AUC of 0.93) than for the model with nT2 alone. Adjusted AUCs for combinations of nT2 and individual diffusion parameters consistently yielded higher estimates compared with the single-parameter model with nT2. However, the magnitude of the improvement tended to be smaller than for nonadjusted models. A considerably better accuracy was observed for the combination of nT2 + FA with a corrected AUC of 0.97 over T2 alone (AUC, 0.92; P < 0.001; Fig. 3D).

DISCUSSION Magnetic resonance neurography is increasingly used in the diagnostic workup of disorders of the peripheral nervous system. Currently, the diagnosis is commonly based on fat-saturated T2-w sequences.3 Here, we demonstrate added value of DTI independently of the present imaging standard T2. Combining T2-w imaging with DTI in patients with peripheral polyneuropathy involving the upper extremity resulted in a significant increase in diagnostic accuracy over T2-w imaging alone (combinations nT2 + FA, nT2 + ADC, and nT2 + RD). Interestingly, the combination of the diffusion parameter FA with either ADC, AD, or RD also yielded improved diagnostic accuracy at least equivalent to T2-w imaging (based on the AIC statistics and including adjustment for overfitting) as well as a statistically significant improvement over single DTI parameter analysis. This indicates that each diffusion parameter can add information and independently improve “anatomical” T2-w imaging and therefore is likely to represent different aspects of the underlying pathology. However, the adjusted AUC estimates of the 2-parameter models suggested that the magnitude of the improvement may only be modest for most DTI parameters given the already high level of diagnostic accuracy for T2-w imaging. Nevertheless, the combination of nT2 and FA yielded a substantially higher adjusted AUC of 0.97. Our study population was directly recruited from clinical routine MRN examinations and represents a typical patient population for 502


nonfocal peripheral neuropathy. The study participants included patients with neuropathy of various etiologies including, for example, hereditary, infective, autoimmune, metabolic, and paraneoplastic. The heterogeneity is typical for peripheral neuropathy for which a manifold of etiologies exist.1 This is similar to other neuroinflammatory diseases where disease course and timing are heterogeneous. We chose to investigate the upper arm because in cases of disseminated neuropathy, involvement of the upper extremity is frequently seen. In addition, neuropathy of the upper extremity can have severe debilitating effects that account for the high morbidity of the disease.7,34 Diagnosis remains a challenge and is mainly based on clinical and electrophysiological signs, the latter being particularly difficult to be performed and evaluated at proximal extremity sites. We hypothesized that MRN and specifically DTI could be an additional diagnostic tool. To our knowledge, there has been no previous study on polyneuropathy of the upper extremity using DTI. Two previous studies used DTI at the lower extremity in neuropathy21,22 and found changes in DTI values on group levels. However, none of these studies assessed diagnostic accuracy, especially compared with the current criterion standard T2-w imaging. As in other locations, T2-w signal alterations are thought to mainly reflect tissue edema, axonopathy, demyelination, or widening of the extracellular space2 and are thus nonspecific for the underlying pathophysiology. Regarding the individual DTI parameters, FA showed the best diagnostic accuracy in our study. In addition, it had the weakest correlation with nT2, suggesting that it has least overlap with T2-w imaging and thereby may provide most additional information. Consistent with this notion, combined nT2 and FA performed best in ROC analysis and showed the highest AUC values. The mean values of nT2 and FA revealed changes as expected in patients. The decrease in FA is due to an increase in RD, which is more

TABLE 4. Diagnostic Performance of nT2 and Diffusion Parameters Unadjusted AUC Model

Est (95% CI)

Single-parameter models nT2 0.946 (0.860–1) FA 0.922 (0.841–1) ADC 0.910 (0.800–1) AD 0.866 (0.746–1) RD 0.886 (0.777–1) Multiparameter models FA + ADC 0.970 (0.931–1) FA + AD 0.983 (0.956–1) FA + RD 0.953 (0.897–1) nT2 + FA 1 (1–1) nT2 + ADC 0.987 (0.960–1) nT2 + AD 0.953 (0.881–1) nT2 + RD 0.989 (0.967–1)

Adjusted AUC




0.004 0.003 0.001

Diffusion Tensor Imaging Adds Diagnostic Accuracy in Magnetic Resonance Neurography.

The aim of this study was to determine whether quantitative diffusion tensor imaging (DTI) adds diagnostic accuracy in magnetic resonance neurography...
711KB Sizes 1 Downloads 40 Views