Altered BOLD signal variation in Alzheimer’s disease and frontotemporal dementia

Recently discovered glymphatic brain clearance mechanisms utilizing physiological pulsations have been shown to fail at removing waste materials such as amyloid and tau plaques in neurodegenerative diseases. Since cardiovascular pulsations are a main driving force of the clearance, this research investigates if commonly available blood oxygen level-dependent (BOLD) signals at 1.5 and 3 T could detect abnormal physiological pulsations in neurodegenerative diseases. Coefficient of variation in BOLD signal (CVBOLD) was used to estimate contribution of physiological signals in Alzheimer’s disease (AD) and behavioural variant frontotemporal dementia (bvFTD). 17 AD patients and 18 bvFTD patients were compared to 24 control subjects imaged with a 1.5 T setup from a local institute. AD results were further verified with 3 T data from the Alzheimer’s disease neuroimaging initiative (ADNI) repository with 30 AD patients and 40 matched controls. Effect of motion and gray matter atrophy was evaluated and receiver operating characteristic (ROC) analyses was performed. The CVBOLD was higher in both AD and bvFTD groups compared to controls (p < 0.0005). The difference was not explained by head motion or gray matter atrophy. In AD patients, the CVBOLD alterations were localized in overlapping structures in both 1.5 T and 3 T data. Localization of the CVBOLD alterations was different in AD than in bvFTD. Areas where CVBOLD is higher in patient groups than in control group involved periventricular white matter, basal ganglia and multiple cortical structures. Notably, a robust difference between AD and bvFTD groups was found in the CVBOLD of frontal poles. In the analysis of diagnostic accuracy, the CVBOLD metrics area under the ROC for detecting disease ranged 0.85 – 0.96. Conclusions The analysis of brain physiological pulsations measured using CVBOLD reveals disease-specific alterations in both AD and bvFTD.


24
The two most common forms of early-onset dementia are Alzheimer's disease (AD) and behavioral 25 variant frontotemporal dementia (bvFTD). Multiple resting-state functional MRI (rs-fMRI) studies 26 concerning AD and bvFTD have been published. In AD the findings have been relatively consistent, 27 with reduced default mode network (DMN) connectivity reported in numerous studies, and it seems to 28 correlate with disease severity [ Structural T1 images were acquired with a TE of 3 ms, TR of 7 ms, FA of 9°, slice thickness of 1.2 131 mm, and a matrix size of 256 x 256 x 170. Preprocessing of the structural T1 images involved bias 132 field correction using a histogram peak-sharpening algorithm (N3) [Sled et al., 1998] and was done 133 already for the datasets downloaded. 134 Functional EPI images were acquired with a TR of 3,000 ms, TE of 30 ms, 140 volumes (7 min), FA 135 of 80°, slice thickness of 3.3 mm, and a matrix size of 64 x 64 x 48. The first three volumes were 136 excluded from the time series due to T1 relaxation effects. 137

Data Preprocessing 138
The BOLD rs-fMRI data were preprocessed with a typical FSL pipeline 139 (http://www.fmrib.ox.ac.uk/fsl, FSL 5.0.8) including: head motion correction (FSL 5.0.8 MCFLIRT, 140 motion estimates were also used in evaluating motion differences between groups), brain extraction (f 141 = 0.5 and g = 0), spatial smoothing (Gaussian kernel 5-mm full width at half maximum), and high-142 pass temporal filtering by using a cutoff of 100 seconds. Multi-resolution affine co-registration within 143 FSL FLIRT software was used to co-register mean, non-smoothed fMRI volumes to 3D FSGR 144 volumes of corresponding subjects, and to co-register anatomical volumes to the Montreal 145 Neurological Institute's (MNI152) standard space template. 146 CV BOLD maps 147 CV was used as a metric for the variation of fluctuations in the BOLD signal. The same method has 148 been used in a study by [Jahanian et al., 2014] and is similar to the method used by [Makedonov et al.,149 where X is voxel time series, ߪ is standard deviation and ߤ is mean. 153 Calculations were done using Matlab (version R2014b). Representative maps from one AD patient, 154 one bvFTD patient and one control subject, as well as the group mean CV BOLD maps, are shown in Fig.  155 1 for both local institute and ADNI data. map, thresholded to 50-100 %) were used as gray matter (GM), WM and CSF templates for the 159 regions-of-interest (ROI) analysis. This approach was used in a study by [Jahanian et al., 2014] where 160 they also showed that results are somewhat independent of the precise GM, WM and CSF 161 segmentation strategy employed. From these ROIs, mean CV BOLD was calculated subject-wise. 162

Voxel-level statistical analysis of CV BOLD maps 163
Differences between study groups in the CV BOLD maps were statistically tested using permutation-164 based nonparametric testing incorporating threshold-free cluster enhancement (TFCE) implemented in 165 the FSL randomise tool with 10,000 random permutations [Smith and Nichols, 2009]. Resulting 166 statistical maps were thresholded at p<0.05, 0.005 and 0.0005. The effect of GM atrophy on the 167 CV BOLD maps was also evaluated repeating the randomize analysis using the GM volume as a 168 regressor. The resulting statistic maps were spatially correlated to the ones without a GM regressor 169 using the fslcc tool from FSL. 170

Effect of motion 171
Motion estimates computed by the MCFLIRT algorithm in the preprocessing step was used to assess 172 the effect of motion on the CV BOLD values. Subject-wise absolute displacement vectors (in mm) were 173 extracted, which describes the amount of movement in all directions over the whole scan as a marker maximum motion value and the number of peaks in the subject-wise motion data were calculated 177 using max and findpeak functions implemented in Matlab R2014b. 178 A univariate linear model analysis of covariance (ANCOVA) was conducted to determine a 179 statistically significant difference between different study groups on the mean CV BOLD values in 180 different ROIs (GM, WM and CSF) controlling for motion parameters. This was performed by using 181 SPSS for Windows statistical software (version 24.0; SPSS, Chicago, Illinois). 182 Furthermore, the effect of removal of residual motion was assessed using the additional preprocessing 183 step of spike removal from the time-series with the AFNI 3dDespike tool using default threshold 184 settings. After 3dDespike, the CV BOLD maps were calculated and tissue-template-based mean CV BOLD 185 values were compared to the ones calculated without despiking. 186

Effect of gray matter atrophy 187
Structural data were analyzed with FSL-VBM, a voxel-based morphometry-style analysis [Ashburner 188 and Friston, 2000;Good et al., 2001]. Structural images were brain-extracted using BET [Smith, 189 2002]. This procedure was verified with visual inspection of the extraction results. Tissue-type 190 segmentation into GM, WM and CSF was carried out using FAST4 (55). The resulting GM partial 191 volume images were then aligned to the Montreal Neurological Institute's (MNI152) standard 192 structural space template using the affine registration tool FLIRT [Jenkinson et al., 2002;Jenkinson 193 and Smith, 2001], followed optionally by nonlinear registration using FNIRT 194 (www.fmrib.ox.ac.uk/analysis/techrep), which uses a b-spline representation of the registration warp 195 field [Rueckert et al., 1999]. 196 To analyze the between-group differences on GM atrophy patterns, these resulting images were 197 averaged to create a study-specific template, to which the native GM images were then nonlinearly re-198 registered. The registered partial volume images were then modulated to correct for local expansion or 199 contraction by dividing by the Jacobian of the warp field. The modulated segmented images were then 200 smoothed with an isotropic Gaussian kernel with a sigma of 4 mm. Finally, GM differences between 201 different study groups were statistically tested using permutation-based nonparametric testing incorporating TFCE implemented in the FSL randomise tool with 10,000 random permutations [Smith 203 and Nichols, 2009]. Resulting statistical maps were thresholded at p<0.05 (TFCE-corrected for 204 familywise errors). 205 To analyze the effect of GM atrophy on CV BOLD values GM volume in voxels was correlated 206 subjectwise with mean CV BOLD values (within GM, WM and CSF) using Spearman's rank correlation 207 coefficient. Scatter plots were used to visualize the results. The effect of GM atrophy was also 208 evaluated repeating the FSL randomise analysis but using this time the GM volume as a regressor. The 209 resulting statistic maps were spatially correlated to the ones without GM regressor using fslcc tool 210 from FSL. 211

Receiver operating characteristic curves 212
We plotted receiver operating characteristic (ROC) curves to evaluate whether CV BOLD could be used 213 to separate healthy controls from patients with either AD or bvFTD, or patient groups from each other. 214 The mean CV BOLD was calculated using different ROIs: GM, WM, CSF (Fig. 2) and disease-specific 215 templates (Fig. 3A). Area under the curve (AUC) was calculated as a measure of classification 216 accuracy. The bootstrap approach was used to estimate the 95% confidence interval of AUC in SPSS 217

Statistical Analysis 219
Statistical analyses were performed with the SPSS and Matlab software, and p values of less than 0.05 220 were considered to indicate a significant difference for all analyses. Between-group differences were 221 assessed using Kruskal-Wallis, two-tailed t and Χ 2 tests, as appropriate. Subject-wise mean motion 222 and GM volume values were correlated with the mean CV BOLD values using Spearman's rank 223 correlation coefficient. 224 A total of 64 healthy controls, 47 AD and 18 bvFTD patients were eligible for analysis. Demographics 227 and clinical data are summarized in Table 1 for the local institute and in Table 2 for the ADNI data. To analyze this further, template-based ROI-analysis was conducted using ICBM152 tissue-template 233 for WM, GM and CSF (Fig. 2). The CV BOLD values on average were higher in both patient groups than 234 in the control group (p-values ranging from 0.008 to 0.00001). This difference was confirmed using 235 data from the ADNI study. The lowest CV BOLD values were detected in the WM in all of the study 236 groups. In the bvFTD group, mean CV BOLD values were higher in WM and GM compared to both AD 237 and the control group (Fig. 2). In the AD group, the mean CV BOLD values were higher in the CSF 238 compared to bvFTD and control group. However, the difference between bvFTD and AD groups in 239 the large-scale ROI-analysis did not reach statistical significance (p=0.27). 240 The CV BOLD demonstrates disease-specific changes 241 More-detailed voxel-level differences of CV BOLD values between study groups were analyzed using 242 permutation-based nonparametric testing incorporating TFCE with 10,000 random permutations. This 243 revealed distinct differences between the patient and control groups. In the AD group, significantly 244 higher CV BOLD values compared to control group were located closer to the center in periventricular 245 WM; in GM higher CV BOLD values are located in the parietal, occipital and posterior part of frontal 246 lobes as well as in frontal pole. In the bvFTD group differences extend more to external parts of the 247 WM and towards the frontal and temporal lobes as well as the middle occipital gyri. There are also 248 higher CV BOLD values in the cerebellum near the 4th ventricle in both diseases. In AD higher CV BOLD values located more towards the horizontal fissure in posterior lobe. In order to pinpoint the most 250 significant changes, the most statistically significant differences are illustrated with p < 0.005 and p < 251 0.0005 in Fig. 3 and in supplementary Fig. S1. 252 Statistically significant (p < 0.005) voxel-wise differences between AD patients and controls showed 253 increased CV BOLD accumulated in a circular area around the CSF ventricles centered on WM. The 254 increased CV BOLD values are found symmetrically in corpus callosum, thalamus, putamen, sagittal 255 stratum, insula and also amygdala and anterior hippocampi areas as well as cerebellum. In the GM 256 increased CV BOLD are found in Broca's areas, somatosensory, supplementary and sensorimotor SM1 257 cortices, paracingulate gyri, and also in visual V1-V3 cortices. Notably, there was no statistically 258 significant difference between AD and control group in the parts of the DMN (posterior cingulate, 259 angular and medial prefrontal gyri). The spatial localizations of the statistically significant changes in 260 CV BOLD values were markedly similar to the results of the ADNI data (Fig. 4). 261 The most prominent changes in bvFTD patients showed increased CV BOLD values more towards the 262 frontal areas and lateral periventricular structures, and towards the temporal pole, premotor cortex and 263 temporal fusiform cortex, and also in visual areas V3-V5 in lateral occipital cortex. Bilateral 264 amygdala, putamen, insula, hippocampus, and areas in the cerebellum showed also higher CV BOLD 265 values (Fig. 3). 266 The statistically significant differences between the AD and bvFTD (bvFTD>AD) patients on voxel-267 level CV BOLD were located bilaterally in anterior part of the frontal lobe (Fig. 3C). 268 The supplementary Tables S1-3 show the most significant group difference clusters and their 269 anatomical labeling in the local institute data. 270

CV BOLD alterations are not explained by head motion or gray matter atrophy 271
There were no significant differences in the absolute or relative head motion parameters between any 272 of the study groups in the local institute data (Fig. 5). In the ADNI dataset, the AD patients moved 273 more before the motion correction (absolute: 0.20 mm for the control group and 0.30 mm for the AD group, p=0.03; relative: control 0.15 mm and AD 0.21 mm, p=0.02). Head motion did not exceed half 275 a voxel size. 276 Mean CV BOLD values did correlate to motion (Fig. 5A, correlation coefficient R ranging from 0.22 to 277 0.50). However, there is still a highly statistically significant effect of study group on the mean 278 CV BOLD values after controlling for motion parameters, c.f. Table 3 (ANCOVA). 279 The number of peaks in the motion signal was also analyzed as markers of sudden head movement. No 280 statistically significant differences were found between groups and there was no correlation to the 281 CV BOLD values (data not shown, R ranging from -0.12 to 0.22, p>0.05 [0.17 to 0.45]). 282 We further verified the effects of motion by performing scrubbing of residual motion spikes (AFNI 283 3dDespike) and repeated the analysis of CV BOLD group differences. The results were not affected by 284 despiking, and there was no significant difference between mean CV BOLD values calculated before and 285 after 3dDespike (p>0.05 [0.68 -0.98]). 286 The effect of GM atrophy on CV BOLD values was analyzed using the local institute dataset (Fig. 6). 287 There was no correlation between the mean CV BOLD values and volume of GM (R=-0.01, p=0.91). The 288 use of GM maps as a regressor in voxel-level analysis implementing FSL randomise resulted statistical 289 maps that were 99 % the same as those without a regressor (spatial correlation coefficient R=0.99). 290 The accuracy of separating controls from patients with CV BOLD 291 The mean CV BOLD calculated from the disease-specific templates showed excellent diagnostic 292 accuracy. Both AD and bvFTD can be differentiated from the controls in local data with 0.96 ROC 293 AUC values. The method also enables differentiation between AD and bvFTD, AUC being 0.806 (Fig.  294 7, Table 4). 295

Discussion 296
The goal of this study was to investigate if physiological signal contributions of BOLD data measured 297 using CV BOLD are altered in different types of dementia. We found that CV BOLD is markedly increased 298 in both AD and bvFTD compared to age-matched controls (p<0.0005), and that the CV BOLD changes most profound changes in the CV BOLD involve areas surrounding CSF, extending to WM, basal ganglia 302 and multiple cortical structures. Suiting the disease pathology of the bvFTD, the most significant 303 differences in CV BOLD comparison to AD were detected in frontolateral GM areas. Mean CV BOLD in 304 the disease-specific templates was able to discern AD patients from controls with the receiver In the present study, analysis of CV BOLD was not limited to WM as in previous study by [Makedonov 349 et al. In the present study the significant difference between the AD and bvFTD was shown to be in bilateral 378 frontal poles. The known neuropsychological differences between AD and bvFTD can be in part The changes in CV BOLD in present study are not explained by difference in age, gender, motion or GM 397 atrophy. There were no statistically significant differences in age, gender or motion parameters 398 between different study groups in the local institute data. As expected, there was a positive correlation 399 with CV BOLD values and motion parameters in all groups alike. However, the patient groups had 400 increased CV BOLD values with the same amount of motion. The effect of motion was also evaluated 401 using an ANCOVA, where differences between groups prevailed as statistically significant after 402 motion parameters were used as covariates. Also, the effect of sudden motion "peaks" was evaluated 403 and this did not explain the group differences in the CV BOLD values. maps, as a regressor in voxel-level analysis did not affect the results. 408 The previous literature and our results suggest that the changes in CV BOLD are not only due to motion, 409 but rather the changes may be due to yet unknown intrinsic properties of the degenerated brain tissue. 410 ADNI data may be more sensitive to hemodynamically coupled BOLD signal changes due to the 411 higher magnetic field strength (3 T) than in local institute data (1. However, the increased motion in ADNI data may partly mask the CV BOLD differences between 413 groups. Furthermore, the local institute data is also nearly two times faster in sampling rate (TR 1.8 vs.     ROC curve for distinguishing patients with AD from control subjects on the basis of CV BOLD values 930 within AD template as ROI. Red lines represent the template created using the most significant 931 differences (p<0.0005) between groups in local data (Fig. 3B) and in blue the significant differences 932 (p<0.005) common in both ADNI and local data (Fig. 4B). B, Same as in A for the ADNI data. C, 933 ROC curve for distinguishing patients with bvFTD from control subjects on the basis of CV BOLD 934 values within the bvFTD template as ROI. The template for ROI was created using the most 935 significant differences (p<0.0005) between groups in local data (Fig. 3B). D, ROC curve for 936 distinguishing patients with bvFTD from those with AD on the basis of AD or bvFTD template. 937 Confidence intervals are shown in Table 4. 938 939