Amyotrophic Lateral Sclerosis and Frontotemporal Degeneration, 2015; 16: 92–101

Ex post facto assessment of diffusion tensor imaging metrics from different MRI protocols: Preparing for multicentre studies in ALS

JOHANNES ROSSKOPF1, HANS-PETER MÜLLER1, JENS DREYHAUPT2, MARTIN GORGES1, ALBERT C. LUDOLPH1 & JAN KASSUBEK1 1Department

of Neurology, University of Ulm, and 2Institute of Epidemiology and Medical Biometry, University of Ulm, Germany

Abstract Diffusion tensor imaging (DTI) for assessing ALS-associated white matter alterations has still not reached the level of a neuroimaging biomarker. Since large-scale multicentre DTI studies in ALS may be hampered by differences in scanning protocols, an approach for pooling of DTI data acquired with different protocols was investigated. Three hundred and nine datasets from 170 ALS patients and 139 controls were collected ex post facto from a monocentric database reflecting different scanning protocols. A 3D correction algorithm was introduced for a combined analysis of DTI metrics despite different acquisition protocols, with the focus on the CST as the tract correlate of ALS neuropathological stage 1. A homogenous set of data was obtained by application of 3D correction matrices. Results showed that a fractional anisotropy (FA) threshold of 0.41 could be defined to discriminate ALS patients from controls (sensitivity/specificity, 74%/72%). For the remaining test sample, sensitivity/specificity values of 68%/74% were obtained. In conclusion, the objective was to merge data recorded with different DTI protocols with 3D correction matrices for analyses at group level. These post processing tools might facilitate analysis of large study samples in a multicentre setting for DTI analysis at group level to aid in establishing DTI as a non-invasive biomarker for ALS. Key words: Fractional anisotropy, motor neuron disease, magnetic resonance imaging, multicentre study

Introduction Advanced magnetic resonance imaging (MRI) techniques including diffusion tensor imaging (DTI) have greatly improved the understanding of the underlying neuropathology of motor neuron diseases (MNDs) (1,2). Today, DTI is established as a robust non-invasive technique in the analysis of amyotrophic lateral sclerosis (ALS) (3). DTI can quantify the diffusion anisotropy in local microstructures in vivo by fractional anisotropy (FA) (4). Statistical analysis can be performed in an unbiased way by approaches such as whole brain-based spatial statistics (WBSS) or analysis at the group level can be performed by quantitative fibre tracking techniques such as tract-based spatial statistics (TBSS – (5)) or tractwise fractional anisotropy statistics (TFAS – (6)). Consistent DTI based findings in ALS and primary lateral sclerosis (PLS) give rise to a unified picture of white matter alterations in the corticospinal

tract (CST) and in the corpus callosum (CC) (7–11). An FA-based in vivo imaging concept (12) has been applied to the recently introduced neuropathological staging system which has shown that ALS may disseminate in a sequential regional pattern in four disease stages (13). However, MRI has still not reached the level of a neuroimaging biomarker (14) although ALS imaging has been rather successful as a descriptive tool characterizing features of specific ALS phenotypes and genotypes (15). However, the broader use of the technique as a biomarker is limited by the small number of examined subjects in previous studies. Given that it is often a challenge for one centre to acquire a sufficient number of ALS patients due to the strains of the MRI acquisition in severely affected patients (16), the challenge is to increase the number of examined subjects in studies with ALS. In DTI, multicentre studies appear to be the best solution in order to improve statistical power in investigation. However, despite great effort, inter-scanner reliability of different centres is still uncertain

Correspondence: J. Kassubek, Department of Neurology, University of Ulm, Oberer Eselsberg 45, 89081 Ulm, Germany. Fax: 49 731 177 1202. E-mail: [email protected] (Received 5 August 2014 ; accepted 12 October 2014 ) ISSN 2167-8421 print/ISSN 2167-9223 online © 2015 Informa Healthcare DOI: 10.3109/21678421.2014.977297

DTI metrics from different MRI protocols in ALS (17–19). An alternative way to improve the number of patients with ALS is to use retrospective data accumulated from an institute’s database, but scanner protocols might have changed over the years. Specifically, the number of gradient directions (GD) has to be taken into account for a DTI scanning protocol. The objective of this study was to explore the effects of DTI protocol variations in a monocentric study design on FA metrics with the aim to develop a method which allows for pooling of DTI data that had been acquired with different scanning protocols. First, the effects were tested on repeated scans of controls by region of interest (ROI) analyses. Secondly, effects on results of comparisons at group level were investigated and a correction algorithm was developed to reach comparability of FA maps derived from different protocols, which was then applied to test for FA based thresholds in large scale samples.

(GD) and one scan with gradient 0 (b0). Five acquisitions were online averaged by the scanner software; repetition time (TR) was 3000 ms, echo time (TE) was 90 ms, and the b-value was 800 s/mm2. • The second scanning protocol (protocol B) contained 52 volumes (64 slices, 128 ⫻ 128 pixels, slice thickness 2.8 mm, in-plane pixel size 2.0 ⫻ 2.0 mm2). Four acquisitions with 13 GD resulted in 48 GD and 4 b0; the four acquisitions had slightly different GD, so averaging was not feasible; TR was 7600 ms, TE was 85 ms, and the b-value was 1000 s/mm2. • The third scanning protocol (protocol C) contained 62 volumes (64 slices, 128 ⫻ 128 pixels, slice thickness 2.5 mm, in-plane pixel size 2.5 ⫻ 2.5 mm2). Two discrete acquisitions with 31 GD resulted in 60 GD and 2 b0; TR was 8700 ms, TE was 102 ms, and the b-value was 1000 s/mm2.

Subject population Data selection. Three hundred and nine datasets were extracted from the database of the Department of Neurology, University of Ulm, Germany. DTI data were considered for inclusion if the data were acquired in patients with definite or probable ALS according to the revised El Escorial diagnostic criteria (20) or in patients with the clinical subtype PLS pursuant to Pringle et al. (21) or in healthy controls. MRI acquisition protocols. All scanning protocols were performed on the same 1.5 Tesla Magnetom Symphony (Siemens Medical, Erlangen, Germany). For all acquisitions, the identical standard receive-only 12-channel circular polarized headcoil was used. Scans with three different acquisition parameter protocols were used: • The first DTI protocol (protocol A) consisted of 13 volumes (45 slices, 128 x 128 pixels, slice thickness 2.5 mm, in-plane pixel size 1.5 ⫻ 1.5 mm2), representing 12 gradient directions

93

Subjects and methods Subjects All subjects gave written informed consent for the MRI protocol according to institutional guidelines that had been approved by the Ethics Committee of the University of Ulm. Descriptive statistics on the subject groups are presented in Table IA. The ALS and PLS patient sample was composed of 170 DTI data sets. Disease duration (mean ⫾ standard deviation) was 2.3 ⫾ 2.6 years (ALS (n ⫽ 140): 1.7 ⫾ 1.3 years; PLS (n ⫽ 30): 4.8 ⫾ 4.7 years), and age of onset of the motor disorder was 60.7 ⫾ 11.5 years. Thirty-four datasets were generated with protocol A (n ⫽ 13 for ALS; n ⫽ 21 for PLS) and 136 datasets were recorded with protocol B (n ⫽ 127 for ALS; n ⫽ 9 for PLS). The control database was composed of 139 healthy adult volunteers’ DTI data, consisting of 75 datasets acquired with protocol A, 13 datasets acquired with protocol B, and 51 datasets

Table I. Characteristics of subjects’ distribution. (A) whole data sample, (B) correction matrix samples, (C) threshold sample, test sample, and application sample. GD – gradient directions; SD – standard deviation. Patients’ DTI-data (N ⫽ 170) (A) subjects GD N Gender (m/f) Age (mean ⫾ SD)

ALS (N ⫽ 140) 13 13 10/3 59 ⫾ 8

N Gender (m/f) Age (mean ⫾ SD)

PLS (N ⫽ 30) 13 21 91/12 59 ⫾ 9

52 9 4/5 63 ⫾ 12

Controls’ data (N ⫽ 40), protocol A (13 GD) 17/23 62 ⫾ 9

(B) Correction matrix sample Gender (m/f) Age (mean ⫾ SD) (C) Samples

52 127 82/45 64 ⫾ 12

Controls’ DTI-data (N ⫽ 139)

Threshold sample ALS/PLS controls 100 59/41 63 ⫾ 11

93 43/50 59 ⫾ 14

13 75 31/44 62 ⫾ 13

52 13 11/2 51 ⫾ 16

62 51 20/31 64 ⫾ 9

Controls’ data (N ⫽ 40), protocols B/C (52/62 GD) 17/23 63 ⫾ 5 Test sample ALS/PLS Controls 50 33/17 62 ⫾ 11

46 19/27 56 ⫾ 14

Application sample ALS/PLS 20 13/7 61 ⫾ 15

94

J. Rosskopf et al.

acquired with protocol C. Of these, nine healthy controls (five male/four female, age at last scan 67.3 ⫾ 7.6 years) were scanned in more than one separate scanning session, yielding a total of 18 DTI scans and one additional scan for subject 6 who was scanned twice with protocol A. In terms of various GDs, three different scanning protocols were applied leading to combinations of acquisitions with protocols A and C and with protocols B and C. Three controls underwent protocols B and C with an averaged time-interval of 20 months; six controls underwent protocols A and C with an averaged time-interval of 42 months; for one subject two scans with protocol A (time-interval 37 months) existed. Data analysis The DTI analysis software Tensor Imaging and Fibre Tracking (TIFT) (22) was used for post processing and statistical analysis. The applied techniques have been previously described in detail (23). ROI analysis In order to assess the effects of the DTI protocol on FA calculation, pairs of healthy controls’ datasets

(Figure 1A) were statistically analysed in terms of FA values in ROIs placed in the same anatomical location. Intra-subject variability was calculated from the scans of a control that was repeatedly scanned with protocol A; intra-operator variability was assessed by calculating the variation in FA values when varying the position of the respective ROI by two millimetres in all spatial directions. The selected ROI positions were in anatomical regions that are prone to be affected in ALS stage 1, i.e. CST (e.g. (12)), and in anatomical regions that have been reported to show differences at group level, i.e. corpus callosum (CC) (e.g. (24)) and frontal lobes (1). ROIs are imaged in Supplementary Figure 1 – which is only available in the online version of the journal. Please find this material with the following direct link to the article: http://informahealthcare.com/doi/ abs/10.3109/21678421.2014.977297. Post processing Spatial normalization to the Montreal Neurological Institute (MNI) (25) stereotaxic standard space was performed by creating a study-specific (b0) template and FA template in an iterative manner (26). FA maps were smoothed with an 8-mm full width at

Figure 1. Fractional anisotropy correction and sample tests. Distribution of FA maps: (A) for ROI analyses, (B) to set up correction matrices, threshold sample and test sample and fi nally the application sample.

DTI metrics from different MRI protocols in ALS half-maximum Gaussian filter in order to achieve a good balance between sensitivity and specificity (24). Differences between scanning protocols at the group level In order to calculate differences between scanning protocols at the group level, FA maps of controls that underwent different scanning protocols were group averaged for each of the scanning protocols and an averaged difference matrix was calculated. Step 1. Recalibration of FA maps by calculation of a 3D correction matrix. For the description of the following analysis methods, it is of note that differences in FA maps were observed mainly between protocol A (13 GD) and protocols B and C (52 GD and 62 GD, respectively); no differences were found between protocols B and C (see Results section). Therefore, FA maps of protocols B and C were merged for the following analysis steps. In order to calculate a 3D correction matrix to recalibrate differences between FA maps of protocol A and protocols B/C, an age- and gendermatched controls’ data sample of 40 FA maps derived from protocol A (13 GD) and 40 FA maps derived from protocols B/C (52/62 GD) was assembled (Table IB). Differences of FA maps derived from different DTI protocols were assumed to be approximated by a polynomial approach: y ⫽ a0 ⫹ a1 x ⫹ a2 x2 ⫹ … ⫹ an xn

(1)

where yi and xi were the voxel-FA values of the different samples at MNI position i; an were the respective coefficients. The polynomial approach offers the possibility to define the differences between FA maps of the different samples (from two scan protocols A and B) as: 1) linear shift – a0 ≠ 0; a1 ⫽ 1; an ⫽ 0 (n ⬎ 1) (2) 2) linear regression –a0, a1 ≠ 0; an ⫽ 0 (n ⬎ 1) (3) 3) higher order differences. In this way, two 3D correction matrices were defined in the following: 1) a linear shift 3D correction matrix: each voxel position represents the difference of averaged FA values of 40 controls (protocol A) and averaged FA values of 40 controls (protocols B/C) (linear shift, Equation 2), 2) a linear regression 3D correction matrix: at each voxel position FA values of 40 controls (protocol A) and FA values of 40 controls (protocols B/C) were separately sorted (ascending order) to 40 pairs, and a regression line was calculated for values between fifth and 95th percentile (36 data sets); two 3D correction matrices (for a0 and a1 each) were calculated (linear regression, Equation 3).

95

With calculation of the 3D correction matrices, all FA maps acquired with protocol A could be recalibrated to FA maps acquired with protocol B (Figure 1B) by application of Equation 1. Step 2. Comparisons at the group level Whole brain-based spatial statistics (WBSS). WBSS was performed with voxelwise comparison by Student’s t-test. FA values below 0.2 were not considered for statistical analysis as cortical grey matter shows FA values up to 0.2 (27). Results were corrected for multiple comparisons using the false discovery rate (FDR) algorithm (28) at p ⬍ 0.05; a clustering procedure for further reduction of type I and type II errors was applied with a threshold cluster size of 512 voxels, corresponding to a sphere with radius of approximately 2 acquisition voxels (24). Fibre tracking and tractwise fractional anisotropy statistics (TFAS). For fibre tracking of the CST, an averaged DTI data set was calculated from control data sets by arithmetic averaging of the MNI transformed data while preserving directional information of individual data sets (for details, see (6)), i.e. Eigenvectors and Eigenvalues were calculated for each voxel position that represented the average of all controls’ data sets. These averaged control DTI data sets were then used to identify the CST with a seed-to-target approach for which seed and target region had a radius of 10 mm each defining a tract of interest (TOI), i.e. the CST. For the fibre tracking technique, a modified deterministic streamline tracking approach was used that takes the directional information of neighboured FTs into account (29). Parameters for FT were an FA threshold of 0.2 (27) and an Eigenvector scalar product threshold of 0.9. Seed and target coordinates were MNI ⫾ 22/–8/9 and MNI ⫾ 30/–20/56, respectively (Supplementary Figure 2B – which is only available in the online version of the journal. Please find this material with the following direct link to the article: http:// informahealthcare.com/doi/abs/10.3109/21678421. 2014.977297). In a consecutive step, the technique of tractwise fractional anisotropy statistics (TFAS) (6) was applied to quantify the tractography results by use of the TOI to select FA values underlying the FTs. FA threshold definition With the objective of generating and testing an FA based threshold, the entire set of data (170 ALS and PLS patients’ data and 139 controls’ data) was randomly split into the ‘threshold sample’ that incorporated 93 controls’ data and 100 patients’ data and into the ‘test sample’, consisting of the remaining 64 controls’ data and 50 patients’ data (Table IC). The distribution of data into threshold sample and test

96

J. Rosskopf et al.

Figure 2. Differences between scanning protocols. (A) Averaged intra-individual difference. (B) Shift correction matrix calculated from differences at the group level between two control groups, protocols A and B/C, respectively.

sample was chosen at a rate of 2:1 in order to obtain a comparatively high weighting to the threshold sample. The scans of the last acquired 18 ALS and two PLS patients were defined as the ‘application sample’. The randomized distribution was chosen because an age- and gender-matched threshold sample would imply a strained test sample with the constraint that distinctive testing would not be possible any more. The attempt of defining a threshold for discriminating ALS patients from controls was performed in two respects: 1)

2)

FA values in four ROIs in the CST in MNI normalized FA maps at MNI coordinates ⫾ 22/–17/34 and ⫾ 22/–17/1 (following (10,30–32)) (Supplementary Figure 2A – which is only available in the online version of the journal. Please find this material with the following direct link to the article: http:// informahealthcare.com/doi/abs/10.3109/ 21678421.2014.977297) were calculated to define sensitivity and specificity for variation of the FA threshold. The technique of TFAS in the CST (Supplementary Figure 2B which is only available in the online version of the journal. Please find this material with the following direct link to the article: http://informa healthcare.com/doi/abs/10.3109/21678421. 2014.977297) was used to calculate sensitivity and specificity for various FA thresholds.

The optimum FA thresholds for group separation were obtained by use of Receiver Operating Characteristics (ROC) curves and the Youden Index which represents the sum of sensitivity and specificity minus 1 and shows optimum group separation at its maximum for variation of the thresholds (33). In the final step, the optimum FA thresholds that were obtained from the threshold samples were applied on the test sample and the application sample to assess how reliable the calculated threshold is to distinguish between controls’ data and ALS patients’ data.

Results Effects of DTI scanning protocol: intra-subject variability Averaged FA values were calculated for each ROI separately, and differences between subjects/scans were arithmetically averaged. The intra-subject variability of two scanning sessions with the protocol A was 0.009 (2.2%). For pairs of controls’ datasets, the averaged intra-subject variability between protocols B and C was 0.008 (2.4%). Intra-operator variability (calculated from variations of ROI position) was 0.003 (0.9%). In this way, deviations below 3.0%, corresponding to intra-subject and intraoperator variability, were defined as normal and did not have to be corrected; deviations greater than 3.0% were above intra-subject and intra-operator variability and thus could result from different scan protocols and had to be corrected (Figure 1A). The intra-operator reproducibility and intra-subject reproducibility of this study was in line with previous studies (34). Protocols B (52 GD) and C (62 GD) showed FA maps with a variability of less than 3.0%, and data of these protocols could be merged. FA maps from protocol A scans showed an increment of FA values of 0.018 (5.0%) on average compared to scans from protocols B and C. This increment was higher than deviations from intra-subject and intra-operator variability and thus has to be corrected prior to group averaging of FA maps. Effects of DTI scanning protocol: differences of controls’ FA maps at the group level Differences of the FA maps of the nine controls with repeated scans, i.e. averaged FA values of controls from protocol A minus FA values of controls from protocols B/C, showed higher FA values throughout the entire brain (Figure 2A). Differences of the FA maps between unpaired controls’ data (40 controls’ data recorded with protocol A vs. 40 controls’ data recorded with protocol B/C) showed a similar pattern (Figure 2B). Thus, it was assumed that these

DTI metrics from different MRI protocols in ALS differences resulted from the different acquisition protocols (A and B/C) and had to be corrected. Effects of DTI scanning protocol: correction for differences at the group level In order to correct for the differences in scanning protocols A and B/C at group level, 3D correction matrices were calculated according to Equation 2 (linear shift approach) and Equation 3 (linear regression approach). Whole brain-based spatial statistics for ALS patients vs. controls Using FA maps not corrected with 3D correction matrices, comparison at the group level for the threshold data sample, i.e. 100 ALS patients vs. 93 controls, demonstrated a significant FA decrease along the CST with frontal cluster extensions (Figure 3A), i.e. areas where intra-individual differences according to the scanning protocol had already been detected (compare Figure 2A). After 3D linear shift correction by use of the 3D shift matrix (Figure 2B) of all data recorded with protocol A, the comparison at the group level showed less pronounced frontal cluster extensions, whereas the CST alterations remained unchanged (Figure 3B). Next order 3D correction (linear regression) demonstrated no frontal cluster extensions (Figure 3C). Therefore, 3D FA map correction with linear regression was performed for all DTI data sets from protocol A. Sensitivity and specificity ROC curves and Youden Index were calculated for group differences in the threshold sample (100 ALS

97

patients and 93 controls) (Figure 4) both for ROI based CST analyses and for FT of the CST (Supplementary Figure 2 to be found at online http:// informahealthcare.com/doi/abs/10.3109/21678421 .2014.977297). This resulted in an FA threshold of 0.409 for the average of the four ROIs and an FA threshold of 0.341 for FT. Sensitivity of 74% and specificity of 72% were obtained for the ROIs, and sensitivity of 68% and specificity of 66% were obtained for the FTs. The application to the test sample FA maps yielded a sensitivity of 68% and a specificity of 74% for ROI analysis and a sensitivity of 68% and a specificity of 61% for TFAS (Table II). Application of the ROI analysis to the application sample showed a sensitivity of 70%; TFAS showed a sensitivity of 75%. The FA values of patients’ data correlated significantly with the ALSFRS-R (n ⫽ 127), with Spearman correlation coefficients of 0.22 (p ⫽ 0.005) for the ROI analysis and 0.24 (p ⫽ 0.003) for TFAS. No significant correlation was found with disease duration. Analysis of the whole data sample Final WBSS based analysis on the whole data sample of 170 ALS patients and 139 controls demonstrated a large significant cluster along the CST, corresponding to ALS stage 1 (12,13), together with the CC in the so-called ‘horseshoe configuration’ (Figure 5). Further clusters in the frontal lobes and in the brainstem were partially interconnected to this large area. All clusters are listed in Supplementary Table I – which is only available in the online version of the journal. Please find this material with the following direct link to the article: http://informahealthcare. com/doi/abs/10.3109/21678421.2014.977297). Averaged FA values in the four ROIs allowed for a group separation with a sensitivity of 72% (64%– 80%) and a specificity of 73% (64%–82%) between ALS patients and controls at an FA threshold of 0.409. Discussion

Figure 3. Whole brain-based spatial statistics (WBSS), comparison of FA maps from 100 ALS patients vs. 93 controls, p ⬍ 0.05, FDR-corrected. (A) No correction was applied. (B) FA maps of DTI protocol A (13 GD) were corrected with the shift correction matrix (Figure 2B). (C) FA maps of DTI protocol A were corrected with the linear regression correction matrices.

MRI is a leading non-invasive, accessible tool to probe the motor and extramotor damage in ALS, and it has tangible potential as a source for diagnosis of motor and extramotor lesion in ALS and thus potential as a monitoring biomarker (14,34). The purpose of ALS imaging is two-fold: first is to develop an imaging based technology that enhances individualized diagnostic accuracy beyond best clinical practice and secondly is to further progress our understanding of disease pathology and pathophysiology for which group analysis is appropriate (15). The most consistent results in ALS have come from studies using FA; this measure is a marker for disruption of the normal architecture of white matter tracts (3). The

98

J. Rosskopf et al.

Figure 4. Receiver operator characteristics (ROC) curves (left column) and Youden-Index (right column) for ROI analysis and for fibre tracking (FT) analysis (tractwise fractional anisotropy statistics – TFAS).

majority of recent systematic reviews on the topic are technique based (11). A meta-analysis performed by Foerster et al. (35,36) reported a pooled sensitivity of 65% and a pooled specificity of 67% and postulated that the discriminatory capability of DTI to make a diagnosis of ALS is only modest. Highest values for sensitivity (92%) and specificity (88%) for study group classification were reported in a discriminant analysis combining radial diffusivity, fractional anisotropy and voxel based morphometry (37). This multiparametric MRI study was applied to a sample of only 24 ALS patients but shows the potential of MRI based classification in ALS.

DTI scans of ALS patients were acquired for almost one decade in many centres around the world with different acquisition protocols. This study is an ex post facto approach to use the high potential for merging and pooling of DTI with different technical specifications data to large scale subject studies in ALS, as proposed in the NeuroImaging Society in ALS initiative (3). We focused on FA as a prominent DTI based parameter; however, other DTI metrics could also be of potential interest and strategies could be developed to transfer the FA based analysis methodology to further DTI metrics. Multi-platform, multi-protocol studies of different MRI metrics are frequently topics of research:

Table II. Results of sensitivity and specificity by use of an FA-based threshold for threshold sample, test sample, and application sample; numbers in brackets indicate the confidence interval (95%). ROI ⫽ region of interest; FT ⫽ fiber tracking. Shift corrected data Threshold Threshold sample Sensitivity (N ⫽ 100) specificity (N ⫽ 93) Test sample Sensitivity (N ⫽ 50) specificity (N ⫽ 46) Application sample Sensitivity (N ⫽ 20)

Linear regression

Corrected data

ROIs 0.408

FT 0.341

ROIs 0.409

FT 0.341

71% [61%-80%] 77% [68%-86%]

66% [56%-75%] 69% [60%-79%]

74% [64%-82%] 72% [62%-81%]

68% [58%-77%] 66% [55%-75%]

68% [53%-80%] 80% [66%-91%]

64% [49%-77%] 63% [48%-77%]

68% [53%-80%] 74% [59%-86%]

68% [53%-81%] 61% [46%-75%]

70% [45%-89%]

75% [50%-92%]

DTI metrics from different MRI protocols in ALS

Figure 5. Whole brain-based spatial statistics (WBSS) of FA maps from 170 ALS patients vs. 139 controls at p ⬍ 0.05, FDRcorrected. Significant clusters are displayed on an averaged b0 background.

studies of multiple sclerosis (e.g. (38)) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (e.g. (39)) have been integrating cross-platform data. Recent DTI studies in other diseases than ALS also reported replicability, reliability, and stability of DTI based FA measurements in multicentre environments (17–19,40,41), but the challenge has not been yet solved how to compensate for differences detected for FA maps. The present study introduces novelties in two respects. First, a methodological framework was provided that allows for pooling of FA maps that had been recorded with different DTI protocols at the same scanner. In this study, first and second order corrections were suggested, and the effect on results of WBSS could be shown. Nevertheless, higher order 3D correction matrices could be set up in different scenarios with the gain to pool DTI based metrics. Secondly, with the sizeable sample of DTI data sets investigated, FA thresholds were defined both in ROIs as well as in the CST FT bundle in which FA values could be used for biomarking ALS in vivo with a good sensitivity and specificity. Limitations In an ex post facto approach, assembling of data samples is driven by the boundary condition of the data base. Prospective studies could plan with statistical arguments to obtain statistically more specific group samples. Therefore, the retrospective design caused two limitations concerning group assembling, i.e. the limited number of DTI data acquired in more than one scanning sessions, and the lack of controls’ paired data with the combination of data derived from acquisition with protocols A and B. By introducing 3D correction matrices, differences in FA maps between protocols (A and B/C) could be corrected by a first

99

order approximation; herein, the amount of data used for creating the 3D correction matrix, i.e. 80 FA maps, appeared to provide a stable stock of data in order to size up the distance between the data acquired with different GD and thus to define a first-order 3D correction matrix. A limitation of the study is that an optimal demonstration of the sensitivity of the corrective algorithm could not be performed: it was not possible to repeat scans with healthy controls under identical conditions, as in the past, because several software upgrades had been performed during the years of data acquisition. Thus, repeated scans of healthy controls with a comparatively long follow-up time were used as the reference. However, three controls had scans of protocols B and C with an average timeinterval of 20 months. Since only differences ⬍ 0.008 in FA values between protocols B and C could be detected, which were in the same range as differences in the intra-subject variability of protocol A (i.e. ⬍ 0.009), FA maps obtained from protocols B and C were considered to be equivalent. Effects of aging have to be considered as below the detection sensitivity. As reported from multicentre studies (e.g. (42)), FA could differ although a standardized DTI protocol is performed on identically structured scanners of various sites. As is customary in multicentre studies, in order to analyse data, the technique of manual setting of ROIs was applied to regions that are prone to be affected in ALS such as CST, CC, and frontal lobes. In this way, the intra-operator reproducibility and intra-subject reproducibility of the study (although surveyed in a comparatively small number of subjects) was in line with previous studies (43). This study was performed as a monocentre feasibility study where scanner-specific factors did not have to be considered. From the methodological point of view, the technique of 3D correction matrices could be transferred to a multicentric study design. This study constitutes proof of concept for an ex post facto merging of FA maps in ALS. Therefore, the multicentre extrapolation has to be tested in a separate study design. In order to obtain a uniform database in terms of differences caused by various acquisition parameters, data were recalibrated by 3D correction matrices based on a polynomial approach. This approximation has already shown significant effects in WBSS. However, we are aware that our concept could be extended for an improved 3D correction of DTI metrics derived from different protocols and also from multicentre settings. Thus, this solution represents a framework, and an approach for higher order corrections should be a topic of future research. Sensitivity and specificity From more than 300 DTI datasets of ALS patients and controls, an FA based threshold was generated

100

J. Rosskopf et al.

and tested. The well-known bilateral white matter alterations of patients with ALS in the CST were used as ROIs (43–46). In this study, we focused on ALS stage 1 (13) with the CST as the tract correlate assessable by DTI (12) since the CST is (compared to tract correlates of the further ALS stages) the initial tract correlate to be affected during the disease course. Thus, the CST has the highest probability to show differences at group level in a large study sample which will be per se a mixture of ALS patients in various disease stages. This was also supported by comparisons of the full study sample at group level where the CST, as the tract correlate of ALS stage 1, was found to be the brain structure with the most prominent affectations. The tract of interest based analysis of the tract correlates of further ALS stages (12) was, however, not the topic of investigation in the present study. According to Prokscha et al. (47), DTI based data already delivered a sensitivity of 100% and a specificity of 92% in discriminating only 12 controls’ data and 13 patients’ data. However, the small number of examined subjects makes this study expandable. The recent meta-analysis performed by Foerster et al. (36,37) concluded a pooled sensitivity of 65% and a pooled specificity of 67%. Findings of the present study, however, fall between both extremes, which is a promising result given the inhomogeneity of the included DTI data sets. WBSS of all ALS patients’ FA maps vs. controls’ FA maps revealed major significant results along the CST, in the CC as well as in the frontal lobes (30). Furthermore, measured FA values correlated significantly with the patients’ clinical characteristics quantified by the ALSFRS-R score. Although longitudinal studies in ALS have addressed this topic and come to inconsistent conclusions (44,48–50), the present study indicates a significant correlation between FA values and stage of disease. Conclusion The aim of future DTI based studies on patients with ALS should be to establish DTI as a neuroimaging surrogate marker for ALS. This ex post facto analysis was able to set up a technique for pooling of data recorded with different MRI protocols in terms of calculation of 3D correction matrices for detection of differences in ALS stage 1 related structures, i.e. the CST. This correction matrix technique can be extended for pooling multicentre DTI data from different MRI protocols due to local scannerspecific conditions. Acknowledgement This study was supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG Grant Number LU 336/15-1) and the German

Network for Motor Neuron Diseases (BMBF 01GM1103A). Declaration of interest: The authors report no confl icts of interest. The authors alone are responsible for the content and writing of the paper. References 1. Agosta F, Chiò A, Cosottini M, de Stefano N, Falini A, Mascalchi M, et al. The present and the future of neuroimaging in amyotrophic lateral sclerosis. AJNR Am J Neuroradiol. 2010;31:1769–77. 2. Kassubek J, Ludolph AC, Müller HP. Neuroimaging of motor neuron diseases. Ther Adv Neurol Disord. 2012;5: 119–27. 3. Turner MR, Grosskreutz J, Kassubek J, Abrahams S, Agosta F, Benatar M, et al. First Neuroimaging Symposium in ALS (NISALS). Towards a neuroimaging biomarker for amyotrophic lateral sclerosis. Lancet Neurol. 2011; 10:400–3. 4. Le Bihan D, Mangin JF, Poupon C, Clark CA, Pappata S, Molko N, et al. Diffusion tensor imaging: concepts and applications. J Magn Reson Imaging. 2001;13:534–6. 5. Smith SM, Jenkinson M, Johansen-Berg H, Rueckert D, Nichols TE, Mackay CE, et al. Tract based spatial statistics: voxelwise analysis of multi-subject diffusion data. Neuroimage. 2006;31:1487–505. 6. Müller HP, Unrath A, Sperfeld AD, Ludolph AC, Riecker A, Kassubek J. Diffusion tensor imaging and tractwise fractional anisotropy statistics: quantitative analysis in white matter pathology. Biomed Eng Online. 2007;6:42. 7. Ulug˘ AM, Grünewald T, Lin MT, Kamal AK, Filippi CG, Zimmerman RD, et al. Diffusion tensor imaging in the diagnosis of primary lateral sclerosis. J Magn Reson Imaging. 2004;19:34–9. 8. Senda J, Ito M, Watanabe H, Atsuta N, Kawai Y, Katsuno M, et al. Correlation between pyramidal tract degeneration and widespread white matter involvement in amyotrophic lateral sclerosis: a study with tractography and diffusion tensor imaging. Amyotroph Lateral Scler. 2009;10:288–94. 9. Iwata NK, Kwan JY, Danielian LE, Butman JA, Tovar-Moll F, Bayat E, Floeter MK. White matter alterations differ in primary lateral sclerosis and amyotrophic lateral sclerosis. Brain. 2011;134:2642–55. 10. Müller HP, Unrath A, Huppertz HJ, Ludolph AC, Kassubek J. Neuroanatomical patterns of cerebral white matter involvement in different motor neuron diseases as studied by diffusion tensor imaging analysis. Amyotroph Lateral Scler. 2012;13:254–64. 11. Turner MR, Agosta F, Bede P, Govind V, Lulé D, Verstraete E. Neuroimaging in amyotrophic lateral sclerosis. Biomark Med. 2012;6:319–37. 12. Kassubek J, Müller HP, del Tredici K, Brettschneider J, Pinkhardt EH, Lulé D, et al. Diffusion tensor imaging analysis of sequential spreading of disease in amyotrophic lateral sclerosis confirms patterns of TDP-43 pathology. Brain. 2014;137:1733–40. 13. Brettschneider J, del Tredici K, Toledo JB, Robinson JL, Irwin DJ, Grossman M, et al. Stages of pTDP-43 pathology in amyotrophic lateral sclerosis. Ann Neurol. 2013;74:20–38. 14. Turner MR. MRI as a frontrunner in the search for amyotrophic lateral sclerosis biomarkers? Biomarkers in Medicine. 2011;5:79–81. 15. Bede P, Hardiman O. Lessons of ALS imaging: pitfalls and future directions. A critical review. Neuroimage Clin. 2014;4:436–43. 16. Wijesekera LC, Leigh PN. Amyotrophic lateral sclerosis. Orphanet J Rare Dis. 2009;4:3.

DTI metrics from different MRI protocols in ALS 17. Teipel SJ, Reuter S, Stieltjes B, Acosta-Cabronero J, Ernemann U, Fellgiebel A, et al. Multicentre stability of diffusion tensor imaging measures: a European clinical and physical phantom study. Psychiatry Res. 2011;194:363–71. 18. Vollmar C, O’Muircheartaigh J, Barker GJ, Symms MR, Thompson P, Kumari V, et al. Identical, but not the same: intra-site and inter-site reproducibility of fractional anisotropy measures on two 3.0T scanners. Neuroimage. 2010; 51:1384–94. 19. Fox RJ, Sakaie K, Lee JC, Debbins JP, Liu Y, Arnold DL, et al. A validation study of multicentre diffusion tensor imaging: reliability of fractional anisotropy and diffusivity values. AJNR Am J Neuroradiol. 2012;33:695–700. 20. Brooks BR, Miller RG, Swash M, Munsat TL; World Federation of Neurology Research Group on Motor Neuron Diseases. El Escorial revisited: revised criteria for the diagnosis of amyotrophic lateral sclerosis. Amyotroph Lateral Scler Other Motor Neuron Disord. 2000;1:293–9. 21. Pringle CE, Hudson AJ, Munoz DG, Kiernan JA, Brown WF, Ebers GC. Primary lateral sclerosis. Clinical features, neuropathology and diagnostic criteria. Brain. 1992;115: 495–520. 22. Müller HP, Unrath A, Ludolph AC, Kassubek J. Preservation of diffusion tensor properties during spatial normalization by use of tensor imaging and fibre tracking on a normal brain database. Phys Med Biol. 2007;52:99–109. 23. Müller HP, Kassubek J. Diffusion tensor magnetic resonance imaging in the analysis of neurodegenerative diseases. J Vis Exp. 2013; doi: 10.3791/50427. 24. Brett M, Johnsrude IS, Owen AM. The problem of functional localization in the human brain. Nat Rev Neurosci. 2002;3:243–9. 25. Gorges M, Müller H-P, Ludolph AC, Rasche V, Kassubek J. Intrinsic Functional Connectivity Networks in Healthy Elderly Subjects: A Multiparametric Approach with Structural Connectivity Analysis. BioMed Res Int. 2014:947252. 26. Unrath A, Müller HP, Riecker A, Ludolph AC, Sperfeld AD, Kassubek J. Whole brain-based analysis of regional white matter tract alterations in rare motor neuron diseases by diffusion tensor imaging. Hum Brain Mapp. 2010;31: 1727–40. 27. Agosta F, Pagani E, Petrolini M, Caputo D, Perini M, Prelle A, et al. Assessment of white matter tract damage in patients with amyotrophic lateral sclerosis: a diffusion tensor MR imaging tractography study. AJNR Am J Neuroradiol. 2010;31:1457–61. 28. Kunimatsu A, Aoki S, Masutani Y, Abe O, Hayashi N, Mori H, et al. The optimal trackability threshold of fractional anisotropy for diffusion tensor tractography of the corticospinal tract. Magn Reson Med Sci. 2004;3:11–7. 29. Genovese CR, Lazar NA, Nichols T. Thresholding of statistical maps in functional neuroimaging using the false discovery rate. Neuroimage. 2002;15:870–8. 30. Canu E, Agosta F, Riva N, Sala S, Prelle A, Caputo D, et al. The topography of brain microstructural damage in amyotrophic lateral sclerosis assessed using diffusion tensor MR imaging. AJNR Am J Neuroradiol. 2011;32: 1307–14. 31. Li J, Pan P, Song W, Huang R, Chen K, Shang H. A metaanalysis of diffusion tensor imaging studies in amyotrophic lateral sclerosis. Neurobiol Aging. 2012;33:1833–8. 32. Youden WJ. Index for rating diagnostic tests. Cancer. 1950;3:32–5. 33. Müller HP, Grön G, Sprengelmeyer R, Kassubek J, Ludolph AC, Hobbs NZ, et al. Evaluating multicentre DTI data in Huntington’s disease on site-specific effects: an ex post facto approach. Neuroimage Clin. 2013;2:161–7.

Supplementary material available online Supplementary Table I, Figures 1–2.

101

34. Schimrigk SK, Bellenberg B, Schlüter M, Stieltjes B, Drescher R, Rexilius J, et al. Diffusion tensor imaging-based fractional anisotropy quantification in the corticospinal tract of patients with amyotrophic lateral sclerosis using a probabilistic mixture model. AJNR Am J Neuroradiol. 2007;28: 724–30. 35. Turner MR, Modo M. Advances in the application of MRI to amyotrophic lateral sclerosis. Expert Opin Med Diagn. 2010;4:483–96. 36. Foerster BR, Dwamena BA, Petrou M, Carlos RC, Callaghan BC, Pomper MG. Diagnostic accuracy using diffusion tensor imaging in the diagnosis of ALS: a metaanalysis. Acad Radiol. 2012;19:1075–86. 37. Foerster BR, Dwamena BA, Petrou M, Carlos RC, Callaghan BC, Churchill CL, et al. Diagnostic accuracy of diffusion tensor imaging in amyotrophic lateral sclerosis: a systematic review and individual patient data meta-analysis. Acad Radiol. 2013;20:1099–1106. 38. Moraal B, Meier DS, Poppe PA, Geurts JJ, Vrenken H, Jonker WM, et al. Subtraction MR images in a multiple sclerosis multicentre clinical trial setting. Radiology. 2009; 250:506–14. 39. Simmons A, Westman E, Muehlboeck S, Mecocci P, Vellas B, Tsolaki M, et al. The AddNeuroMed framework for multicentre MRI assessment of Alzheimer’s disease: experience from the first 24 months. Int J Geriatr Psychiatry. 2011;26:75–82. 40. Filippini N, Douaud G, Mackay CE, Knight S, Talbot K, Turner MR. Corpus callosum involvement is a consistent feature of amyotrophic lateral sclerosis. Neurology. 2010; 75:1645–52. 41. Pfefferbaum A, Adalsteinsson E, Sullivan EV. Replicability of diffusion tensor imaging measurements of fractional anisotropy and trace in brain. J Magn Reson Imaging. 2003; 18:427–33. 42. Walker L, Curry M, Nayak A, Lange N, Pierpaoli C, Brain Development Cooperative Group. A framework for the analysis of phantom data in multicentre diffusion tensor imaging studies. Hum Brain Mapp. 2013;34:2439–54. 43. Zhang Y, Schuff N, Woolley SC, Chiang GC, Boreta L, Laxamana J, et al. Progression of white matter degeneration in amyotrophic lateral sclerosis: a diffusion tensor imaging study. Amyotroph Lateral Scler. 2011;12:421–9. 44. Prudlo J, Bißbort C, Glass A, Grossmann A, Hauenstein K, Benecke R, et al. White matter pathology in ALS and lower motor neuron ALS variants: a diffusion tensor imaging study using tract based spatial statistics. J Neurol. 2012; 259:1848–59. 45. Iwata NK, Aoki S, Okabe S, Arai N, Terao Y, Kwak S, et al. Evaluation of corticospinal tracts in ALS with diffusion tensor MRI and brainstem stimulation. Neurology. 2008; 70:528–32. 46. Prokscha T, Guo J, Hirsch S, Braun J, Sack I, Meyer T, et al. Diffusion tensor imaging in amyotrophic lateral sclerosis: increased sensitivity with optimized region-of-interest delineation. Clin Neuroradiol. 2014;24:37–42. 47. Blain CR, Williams VC, Johnston C, Stanton BR, Ganesalingam J, Jarosz JM, et al. A longitudinal study of diffusion tensor MRI in ALS. Amyotroph Lateral Scler. 2007;8:348–55. 48. Agosta F, Rocca MA, Valsasina P, Sala S, Caputo D, Perini M, et al. A longitudinal diffusion tensor MRI study of the cervical cord and brain in amyotrophic lateral sclerosis patients. J Neurol Neurosurg Psychiatry. 2009;80:53–5. 49. Keil C, Prell T, Peschel T, HartungV, Dengler R, Grosskreutz J. Longitudinal diffusion tensor imaging in amyotrophic lateral sclerosis. BMC Neurosci. 2012;13:141.

Copyright of Amyotrophic Lateral Sclerosis & Frontotemporal Degeneration is the property of Taylor & Francis Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.

Ex post facto assessment of diffusion tensor imaging metrics from different MRI protocols: preparing for multicentre studies in ALS.

Diffusion tensor imaging (DTI) for assessing ALS-associated white matter alterations has still not reached the level of a neuroimaging biomarker. Sinc...
3MB Sizes 0 Downloads 6 Views