Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2020 May 12.
Published in final edited form as: J Neurosci Methods. 2019 Nov 19;331:108528. doi: 10.1016/j.jneumeth.2019.108528

Estimating regional cerebral blood flow using resting-state functional MRI via machine learning

Ganesh B Chand a,b, Mohamad Habes a,b,c, Sudipto Dolui a,b, John A Detre a,b, David A Wolk b, Christos Davatzikos a,b,*
PMCID: PMC7216635  NIHMSID: NIHMS1580786  PMID: 31756399

Abstract

Background:

Perfusion MRI is an important modality in many brain imaging protocols, since it probes cerebrovascular changes in aging and many diseases; however, it may not be always available.

New method:

We introduce a new method that seeks to estimate regional perfusion properties using spectral information of resting-state functional MRI (rsfMRI) via machine learning. We used pairs of rsfMRI and arterial spin labeling (ASL) images from the same individuals with normal cognition and mild cognitive impairment (MCI), and built support vector machine models aiming to estimate regional cerebral blood flow (CBF) from the rsfMRI signal alone.

Results:

This method demonstrated higher associations between the estimated CBF and actual CBF (ASL-CBF) at the total lobar gray matter (r = 0.40; FDR-p = 1.9e-03), parietal lobe (r = 0.46, FDR-p = 8e-04), and occipital lobe (r = 0.35; FDR-p = 0.01) using rsfMRI signals of frequencies [0.01-0.15] Hertz compared to frequencies [0.01-0.10] Hertz and [0.01-0.20] Hertz. We further observed significant associations between the estimated CBF and actual CBF in 24 regions of interest (p < 0.05), with the highest association observed in the superior parietal lobule (r = 0.50, FDR-p = 0.002). Moreover, the estimated CBF at superior parietal lobule showed significant correlation with the mini-mental state exam (MMSE) score (r = 0.27; FDR-p = 0.04) and decreased in MCI with lower MMSE score compared to NC group (FDR-p = 0.04).

Comparison with existing methods:

Consistent with previous findings, this new method also suggests that rsfMRI signals contain perfusion information.

Conclusion:

The proposed framework can obtain estimates of regional perfusion from rsfMRI, which can serve as surrogate perfusion measures in the absence of ASL.

Keywords: Neuroimaging, Functional MRI, Arterial spin labeling, Machine learning, Support vector machines

1. Introduction

Arterial spin labeled (ASL) perfusion magnetic resonance imaging (MRI) can noninvasively quantify regional cerebral blood flow (CBF), which is an important physiological parameter of brain function and a versatile biomarker in aging and brain disorders, including cerebrovascular disease, Alzheimer’s disease, and neuropsychiatric disorders (Bruandet et al., 2009; Detre et al., 2012; Gupta et al., 2012; Haller et al., 2016; Love and Miners, 2016; MacDonald and Frayne, 2015). Although ASL MRI can be used to measure task activation, most ASL studies are carried out in the “resting” state and are intended to address state- or trait-like aspects of brain function. Task-independent blood oxygenation level dependent (BOLD) functional MRI (fMRI), often termed resting-state fMRI (rsfMRI) is another common strategy for evaluating state- and trait-like properties of brain function.

ASL uses magnetically labeled arterial blood water as an endogenous tracer (Detre et al., 1992) to quantify CBF in physiological units of ml/100 g/min. ASL MRI has been validated against 15O-PET (Fan et al., 2016; Feng et al., 2004; Kilroy et al., 2014; Ye et al., 2000) and has demonstrated sensitivity to regional CBF alterations associated with Alzheimer’s neurodegeneration (Alsop et al., 2010; Dai et al., 2009; Johnson et al., 2005; Wang et al., 2013; Xie et al., 2016). Further, based on the tight coupling between regional CBF and regional brain metabolism, regional hypoperfusion by ASL and regional hypometabolism by 18F-fluorodeoxyglucose positron emission tomography (FDG-PET) display similar patterns (Chen et al., 2011; Verfaillie et al., 2015) suggesting the potential of ASL to provide noninvasive and less costly surrogate marker than FDG-PET in aging, MCI and/or Alzheimer’s disease (Wolk and Detre, 2012). However, robust acquisition procedures and consensus on best practices in ASL MRI have only recently begun to be available (Alsop et al., 2015) and ASL remains technically more challenging than BOLD fMRI.

RsfMRI has emerged as a commonly acquired MRI protocol primarily used in studies of brain functional connectivity (Biswal et al., 1995, 2010; Chand et al., 2018; Fox et al., 2014; Greicius and Kimmel, 2012; Raichle, 2015). BOLD fMRI indirectly captures the regional neurovascular alterations associated with neuronal activations based on associated changes in deoxyhemoglobin reflecting changes in underlying regional CBF, blood volume, and oxygen metabolism (Detre and Wang, 2002; Kwong et al., 1992; Ogawa et al., 1993, 1992). Previous studies suggest a link between BOLD and perfusion as dynamic changes in BOLD contrast are primarily due to changes in CBF (Detre et al., 2012; Li et al., 2012; Raichle, 1998). BOLD contrast primarily detects changes in T2* that indirectly reflect changes in CBF (Ogawa et al., 1992). ASL perfusion contrast is based on alternations in T1 to quantify CBF in physiological units of ml/100 g/min (Detre and Wang, 2002). Both BOLD and ASL showed consistent functional profiles in the premotor/motor cortex during finger tapping task (Detre and Wang, 2002) and were also found tightly coupled during rest and in response to task demands (Liang et al., 2013; Zou et al., 2015). Emerging studies have begun to reveal that perfusion properties may be extracted from the rsfMRI time series (Lv et al., 2013; Tong and Frederick, 2014; Tong et al., 2017; Yan et al., 2018). Those investigations suggest that appropriate filtering and modeling of the rsfMRI signals allow inferences to be made about regional perfusion, which would be of particular value when ASL MRI data is not available. Herein we propose a new method to estimate perfusion properties from the rsfMRI using a machine learning model.

In the current study, we hypothesized that CBF metrics can be estimated from rsfMRI signals at an array of different frequencies, using machine learning models constructed from individuals scanned with both protocols and therefore serving as a training set. We further hypothesized that the predicted regional CBF metrics correlate with global cognitive performance.

2. Materials and methods

2.1. Study sample

Subjects of this study were recruited from the Penn Memory Center, Penn Alzheimer’s Disease Core Center. The study protocol was reviewed and approved by the Institutional Review Board of the University of Pennsylvania. The clinical diagnosis of mild cognitive impairment (MCI) was determined following the criteria outlined by Petersen and others (Petersen, 2004; Petersen et al., 2009; Winblad et al., 2004). MCI subjects had memory problem, generally intact cognitive functioning and activities of daily living, objective evidence of memory impairment on cognitive testing, and did not qualify for a diagnosis of dementia. MCI subjects generally performed more than 1.5 standard deviations below age-adjusted means on verbal and/or non-verbal memory tests. Control subjects had no significant cognitive complaints, normal performance on age-adjusted cognitive measures. Study exclusion criteria were a history of clinical stroke, significant traumatic brain injury, alcohol or drug abuse, or any medical or psychiatric condition that thought to impact cognition. Out of 78 subjects who were scanned for both rsfMRI and ASL, 7 subjects were excluded due to artifacts and/or incomplete data points in their imaging data sets and thus final sample size turned out to be 71 subjects. In our final sample, 45 individuals were cognitively normal while 26 individuals were with MCI. Sample demographics are presented in Table 1.

Table 1.

Sample demographics.

Variable Normal controls (NC) (n = 45) MCI (n = 26) P value
Age, mean years (SD) 70.56 (7.21) 70.50 (6.11) 0.23
Sex, female (%) 27 (60.00 %) 11 (42.31 %) 0.15
Education, mean years (SD) 16.18 (2.78) 16.38 (2.77) 0.76
MMSE, min-max (SD) 26-30 (0.87) 22-30 (2.68) 3.09e-09
Open in a new tab

2.2. Image acquisition

Images were acquired at the University of Pennsylvania using Siemens 3 T Prisma. 3D T1 images were obtained in sagittal using repetition time (TR) = 2400 ms, echo time (TE) = 2.24 ms, inversion time (TI) = 1060 ms, flip angle = 8 °, matrix = 320 × 320, slices per slab = 208, and slice thickness = 0.80 mm. The rsfMRIs were recorded using TR = 720 ms, TE = 37 ms, flip angle = 52 °, field of view (FOV) = 208 × 208 mm2, matrix = 104 × 104, number of slices = 72, slice thickness = 2 mm, and bandwidth = 2290 Hz/pixel. ASL images were acquired using 1D-accelerated 4-shot 3D stack-of-spirals acquisition with 90 % background suppression, TR = 4250 ms, TE = 9.8 ms, TI = 150 ms, flip angle = 90 °, FOV = 240 × 240 mm2, matrix = 96 × 96, slices per slab = 52, and voxel size = 2.5 × 2.5 × 2.5 mm3. Unbalanced pseudo-continuous labeling with a labeling time of 1.8 s, and post labeling delay of 1.8 s was performed at an optimal location perpendicular to straight segments of the internal carotid and vertebral arteries as determined by time-of-flight angiography. Ten label/control pairs were acquired for signal averaging. Two volumes of M0 images acquired with TR = 6 s and without background suppression were averaged and used to normalize the control-label difference for CBF quantification.

2.3. RsfMRI preprocessing, time series extraction, and spectral features

The rsfMRI preprocessing analyses followed our previous study (Chand et al., 2018) that were slice time correction, motion correction, coregistration to T1, normalization to MNI space, and spatial smoothing. These preprocessing analyses were carried out using SPM12 (Wellcome Trust Centre for Neuroimaging, London, UK). Total 117 anatomical gray matter regions of interest (ROIs) defined by a multi-atlas segmentation method (MUSE) (Doshi et al., 2013, 2016) were considered. RsfMRI time series were extracted from MUSE ROIs and the spectral features were computed using Fourier transform at those ROIs.

2.4. ASL preprocessing and CBF extraction

ASL processing followed the method of Dolui et al. (Dolui et al., 2019). Briefly, the ASL stack-of-spirals time series was motion corrected using the method by Wang (Wang, 2012). Thereafter the control-label images were pairwise subtracted, temporally averaged, divided by the M0 image, and then converted to CBF maps using a single compartment model with parameters recommended by Alsop et al. (Alsop et al., 2015). The mean stack-of-spiral images were coregistered to the high resolution T1 images using FSL boundary-based registration (BBR). Additionally, the T1 images were normalized to the MNI152 template using FMRIB’s non-linear image registration tool (FNIRT) (Andersson et al., 2010). Each subject’s CBF maps were normalized to the MNI152 space using the ASL-T1 BBR coregistration parameters and T1-MNI152 nonlinear warp. The CBF values were finally extracted using MUSE-defined anatomical ROIs and lobes.

2.5. CBF prediction

We binned the spectral features of rsfMRI into 8 bins, which covered 0.01-0.20 Hz with 0.0268 Hz separation between consecutive bins. To examine the prediction with varying frequency range, the spectral features were further divided into [0.01-0.10] Hertz, [0.01-0.15] Hertz, and [0.01-0.20] Hertz. The frequency intervals were chosen for the following reasons. First, the spectral signals [0.01-0.10] Hertz were selected based on a typical frequency range included in rsfMRI studies (Fox et al., 2005; Fransson, 2005). Second, the signals [0.01-0.15] Hertz were selected as recent studies have begun to show that rsfMRI signals with this frequency range might have higher perfusion information (Tong et al., 2017). The frequency range [0.01-0.20] Hertz was chosen to compare with the former two ranges. However, frequency ranges above 0.2 Hz were not included as the signals might pick up physiological noise, such as cardiac cycles (> 0.5 Hz) and respiratory cycles (~0.2-0.3 Hz) (Dagli et al., 1999; Limbrick-Oldfield et al., 2012; Raj et al., 2001). These rsfMRI signals were used as features in the kernel-based support vector machines (Cortes and Vapnik, 1995) to predict the CBF. The support vector machines regression was performed using libSVM package (Chang and Lin, 2011; Drucker et al., 1997) based on 10 folds cross-validation. Briefly, the data was divided into 10-subsets and the method was repeated 10-times. Each time, 9-subsets were used as training set and the remaining 1-subset as a testing set. Gaussian kernel function was used in the model and model parameters were optimized within each training set using 10 folds cross-validation grid search.

2.6. Statistical analysis

To compare MCI and control groups, two-sample t-tests were used for age, education, and mini-mental state exam (MMSE), while a chi-square test was used for sex. Both age and sex variables were adjusted in regional CBF using a linear model. All the correlation analyses were performed using Pearson’s correlation. The relative CBFs at superior parietal lobule and precuneus (i. e., normalized by lobar CBF) were compared between controls, MCI with higher cognitive score (MMSE ≥ 28), and MCI with lower cognitive score (MMSE < 28) using two-sample t-tests. Notably, the subjects with MCI had a measure of MMSE between 26-30. In a sensitivity analysis and to better understand whether those subjects with worse MMSE would show difference in our prediction model, we stratified the MCI population by a cut-off value of 28. This cutoff was arbitrary, but it gave a way of assessing the impact of cognitive severity on these effects. To account for multiple comparisons in statistical significance tests, a false discovery rate (FDR) was used. A MATLAB software version R2018a (https://www.mathworks.com) was used for analyzing the data.

3. Results

3.1. Lobar CBF prediction

The CBF predictions were carried out at the lobar gray matter by using rsfMRI signals with frequencies [0.01-0.10] Hertz, [0.01-0.15] Hertz, and [0.01-0.20] Hertz. Association between the actual CBF and predicted CBF at the total lobar gray matter was higher using frequencies [0.01-0.15] Hertz (r = 0.40; FDR-p = 1.9e-03) than using frequencies [0.01-0.10] Hertz (r = 0.20; FDR-p = 0.09) and [0.01-0.20] Hertz (r = 0.26, FDR-p = 0.04), respectively, as shown in Fig. 1. The correlation coefficients between predicted CBF and actual CBF varied as a function of frequency (Supplementary materials: Figure S1). In each subject, the absolute difference between actual CBF and predicted CBF showed fluctuating behavior with frequency (Supplementary materials: Figure S2). The mean (standard deviation) of absolute difference over all subjects versus frequency value is shown in Figure S3 (Supplementary materials).

Fig. 1.

Fig. 1.

Open in a new tab

Associations between the actual CBF and predicted CBF at total lobar gray matter with varying frequency information of rsfMRI: A) frequency [0.01-0.10] Hertz (Hz) (r = 0.20, FDR-p = 0.09), B) frequency [0.01-0.15] Hz (r = 0.40, FDR-p = 1.9e-03), and C) frequency [0.01-0.20] Hz (r = 0.26, FDR-p = 0.04). Note that the association is higher in [0.01-0.15] Hz compared to that of other frequency ranges.

Prediction analyses were also performed at the frontal, temporal, parietal, and occipital lobes using rsfMRI with frequencies [0.01-0.10] Hertz, [0.01-0.15] Hertz, and [0.01-0.20] Hertz. Table 2 lists the correlations between the predicted and actual CBF values for each lobe and each frequency range. The highest and most significant correlations were observed in the parietal lobe with frequencies [0.01-0.15] Hertz (r = 0.46, FDR-p = 8e-04) (Fig. 2). The correlations between actual CBF and predicted CBF were lower at the occipital lobe compared to that of the parietal lobe. Moreover, associations between actual CBF and predicted CBF were not statistically significant at the frontal lobe and temporal lobe (FDR-p > 0.05) as presented in Table 2.

Table 2.

Associations between the actual CBF and predicted CBF at individual lobar level using rsfMRI signals with frequencies [0.01-0.10] Hertz (Hz), [0.01-0.15] Hz, and [0.01-0.20] Hz (p values are FDR-corrected).

Lobe Frequency [0.01-0.10] Hz Frequency [0.01-0.15] Hz Frequency [0.01-0.20] Hz
Frontal r = 0.19; p = 0.16 r = 0.09; p = 0.41 r = 0.14; p = 0.26
Temporal r = 0.19; p = 0.16 r = 0.16; p = 0.24 r = 0.14; p = 0.26
Parietal r = 0.25; p = 0.09 r = 0.46; p = 8e-04 r = 0.36; p = 0.01
Occipital r = 0.25; p = 0.09 r = 0.35; p = 0.01 r = 0.21; p = 0.16
Open in a new tab

Fig. 2.

Fig. 2.

Open in a new tab

Associations between the actual CBF and predicted CBF at the parietal lobe with varying frequency information of rsfMRI: A) frequency [0.01-0.10] Hertz (Hz) (r = 0.25, FDR-p = 0.09), B) frequency [0.01-0.15] Hz (r = 0.46, FDR-p = 8e-04), and C) frequency [0.01-0.20] Hz (r = 0.25, FDR-p = 0.01). Note that the association is higher in [0.01-0.15] Hz compared to that of other frequency ranges.

3.2. ROI CBF prediction

We further carried out the prediction analyses at 117 MUSE ROIs. At each ROI level, the association between actual CBF and predicted CBF was highest using rsfMRI with frequencies [0.01-0.15] Hertz. Out of those ROIs, the correlations were statistically significant (uncorrected for multiple comparisons; p < 0.05) at 24 ROIs, which were predominantly the cortical ROIs (Supplementary materials: Table S1). After correcting for multiple comparisons, there were 7 statistically significant ROIs (FDR-p < 0.05) (Supplementary materials: Table S1). Predicted CBF at the subcortical ROIs, including putamen, hippocampus etc., was not statistically significant with actual CBF. Among 117 ROIs, the association was highest at the superior parietal lobule. Fig. 3 shows the association between actual CBF and predicted CBF at the superior parietal lobule using rsfMRI signals with frequencies [0.01-0.10] Hertz (r = 0.33, FDR-p = 0.09), [0.01-0.15] Hertz (r = 0.50, FDR-p = 0.002), and [0.01-0.20] Hertz (r = 0.34, FDR-p = 0.08) demonstrating better estimates in [0.01-0.15] Hertz.

Fig. 3.

Fig. 3.

Open in a new tab

Associations between the actual CBF and predicted CBF at superior parietal lobule with varying frequency information of rsfMRI: A) frequency [0.01-0.10] Hertz (Hz) (r = 0.33, FDR-p = 0.09), B) frequency [0.01-0.15] Hz (r = 0.50, FDR-p = 0.002), and C) frequency [0.01-0.20] Hz (r = 0.34, FDR-p = 0.08). Note that the association is higher in [0.01-0.15] Hz as compared to that of other frequency ranges.

3.3. Predicted CBF comparison between controls and MCI

We further examined the predicted CBF at superior parietal lobule (i. e., normalized by predicted CBF of lobar) between controls, MCI with higher cognitive performance (MMSE ≥ 28), and MCI with lower cognitive performance (MMSE < 28). Similar to the actual CBF, the predicted CBF significantly decreased in MCI with lower cognitive performance (FDR-p < 0.05) compared to the controls (Fig. 4). At precuneus, the similar trends were present between controls and MCI, but the p-values were not statistically significant.

Fig. 4.

Fig. 4.

Open in a new tab

Comparing relative CBF between normal controls (NC), MCI with higher cognitive performance (MMSE ≥ 28), and MCI with lower cognitive performance (MMSE < 28): A) Actual and B) Predicted relative CBF at superior parietal lobule demonstrating a significant difference between NC and MCI with lower cognitive performance (MMSE < 28); C) Actual and D) Predicted relative CBF at precuneus demonstrating a trend of difference between NC and MCI with lower cognitive performance (MMSE < 28) (FDR-corrected p values within each region).

3.4. Associations between predicted CBF and cognition

We tested an association between predicted CBF and global cognitive performance assessed by MMSE (Fig. 5). Similar to the actual CBF, the predicted CBF at superior parietal lobule correlated positively with MMSE (FDR-p < 0.05).

Fig. 5.

Fig. 5.

Open in a new tab

Associations between relative CBF at superior parietal lobule (SPL) and MMSE score: A) Correlation between actual relative CBF and MMSE (r = 0.23; FDR-p = 0.05), and B) Correlation between predicted relative CBF and MMSE (r = 0.27; FDR-p = 0.04).

4. Discussion

In the present study, we proposed a novel multi-folded framework that demonstrated that regional CBF metrics can be estimated using the spectral information of rsfMRI. In this approach, we acquired rsfMRI and ASL from same individuals, trained support vector machine regression models on those data using standard cross-validation techniques, predicted regional CBF measures, and finally illustrated an association of the predicted CBF with original (actual) CBF from the ASL data itself. Our analyses revealed that the predicted CBF at the lobar level and/or individual cortical ROI level associates higher with the respective actual CBF using rsfMRI signals of frequencies [0.01-0.15] Hertz. CBF estimates were better in the parietal lobe or region(s) than in other brain regions. Estimated CBF at the superior parietal lobule decreased in MCI group with lower MMSE compared to control group and associated positively with MMSE.

4.1. Better prediction using rsfMRI signals with frequencies [0.01-0.15] Hertz

The predicted CBF metrics varied as a function of frequency of rsfMRI signals. These results were in line with the reports that rsfMRI signal amplitude shows fluctuating behavior with the frequency (Buzsaki, 2004; Zuo et al., 2010). Our results provided strong evidence that the rsfMRI signals with frequencies [0.01-0.15] Hertz had higher associations between the predicted CBF and actual CBF at global lobar, individual lobe, and individual ROI compared to the signals of frequencies [0.01-0.10] Hertz and [0.01-0.20] Hertz, respectively. Perfusion estimations from rsfMRI were broadly in line with recent reports that rsfMRI signals may carry perfusion properties (Lv et al., 2013; Tong and Frederick, 2014; Yan et al., 2018). It has also been shown that the rsfMRI signals with [0.01-0.15] Hertz to a larger extent resembled perfusion measurement with dynamic susceptibility contrast compared to the signals with [0.01-0.10] Hertz (Tong et al., 2017). Therefore, there are mounting evidence that the rsfMRI signal carries perfusion properties, which may be extracted using the novel analytic and modeling approaches.

4.2. Better prediction at the parietal lobe or regions

Interestingly, CBF prediction using our approach was better in the parietal lobe and/or regions compared to other areas. The apparent stronger link between rsfMRI activity and perfusion in the parietal area (s) might be related to the performance of our modeling method. However, there could be other reasons as well. We used spectral information of rsfMRI as input features in the model. In fMRI literature, it has been shown that the parietal cortex and/or its sub-regions support a wide range of cognitive processes in association with the prefrontal/frontal cortex (Miller and Cohen, 2001; Petrides and Pandya, 1999). Specifically, studies consistently argue that parietal regions are the key hubs of brain processes and the information is encoded there to communicate with other regions, including prefrontal/frontal regions, and to regulate the cognitive processes (Culham and Valyear, 2006; Fogassi et al., 2005). The parietal lobe is densely vascularized structure (Clark et al., 2017; Duvernoy et al., 1981; Liebeskind, 2003), which might provide a stronger signal to be estimated. In ASL literature, regions within the parietal lobe are consistently reported to be associated with hypoperfusion in MCI or Alzheimer’s disease (Alsop et al., 2010; Dolui et al., 2017). The parietal regions thus might be well-positioned for functional measures captured via fMRI and/or ASL perfusion, which eventually could reflect in our machine learning-based predictions. It is interesting to note that our method was not performing well in estimating CBF of the frontal lobe/regions compared to other lobes. The frontal lobe has been consistently suggested as higher order area and receives information from relatively lower order occipital, parietal, and temporal cortices during complex cognitive processes and/or external tasks (Miller and Cohen, 2001; Petrides and Pandya, 1999). Whether CBF information available in rsfMRI signals are weakly represented there so that the prediction was poor, or our method was not sensitive enough to capture the available perfusion information remains to be elucidated.

4.3. CBF comparisons between controls and MCI

In our findings, the estimated relative CBF at superior parietal lobule was significantly lower in more impaired MCI patients based on MMSE (< 28) compared to the cognitively normal group. This result was in agreement with prior studies that have reported decreased relative CBF in MCI or Alzheimer’s pathology (Dolui et al., 2016; Huang et al., 2018). Interestingly, MCI patients with higher MMSE (≥ 28) had higher estimated perfusion, albeit not statistically significant. This increased perfusion might be related to compensatory mechanism, which has been observed in early stages during the course of Alzheimer’s disease (Alsop et al., 2008), and which might be captured more by the rsfMRI.

4.4. Implications

Former studies have suggested that ASL captures regional perfusion alterations in patients with Alzheimer’s disease (Alsop et al., 2008; Dai et al., 2009). Importantly, ASL has been envisioned as noninvasive and less costly surrogate marker of FDG-PET in the Alzheimer’s research (Wolk and Detre, 2012). If ASL data is missing in those populations but rsfMRI is available within with appropriate frequency resolution (i.e., [0.01-0.15] Hertz), we showed a proof-of-concept analysis that a surrogate ASL marker can be estimated at the lobes and/or cortical regions. This approach will be useful especially for studies that have large sample size (Habes et al., 2016), where reacquiring missing imaging modalities from the same subject is not practical. CBF estimation was better at the posterior regions, where many previous studies (Alsop et al., 2008; Dolui et al., 2017; Huang et al., 2018) and our results showed altered CBF in MCI or Alzheimer’s pathology, suggesting the potential use of proposed method in those populations.

4.5. Limitations and future work

Although we were able to estimate CBF at the lobar and cortical ROI levels, especially in the posterior regions, it is worth to note that our method did not offer very robust perfusion estimates, which were less accurate in the subcortical regions. Previous rsfMRI studies have reported that the midbrain and/or nearby subcortical structures are prone to physiological noises such as cardiac cycles (> 0.5 Hz), respiratory cycles (~0.2–0.3 Hz or lower) (Dagli et al., 1999; Limbrick-Oldfield et al., 2012; Raj et al., 2001). With these cautions, future studies should focus on optimizing perfusion estimates at the cortical and subcortical structures. The future studies may further experiment on implementing the feature selection methods to investigate whether CBF predictions can be improved.

5. Conclusions

In summary, we provided a novel framework to estimate the regional CBF metrics via machine learning algorithm incorporating spectral information of the rsfMRI. This method demonstrated that the perfusion properties can be estimated better using rsfMRI signals with frequencies [0.01-0.15] Hertz compared to signals with frequencies [0.01-0.10] Hertz and [0.01-0.20] Hertz. The perfusion estimation was better in the parietal lobe and/or regions compared to other lobes and/ or regions. Furthermore, estimated CBF at the superior parietal lobule significantly associated with cognitive score and it was decreased in MCI with lower cognitive scores. In light of growing attentions of ASL perfusion in aging and Alzheimer’s disease pathology, we believe that the proposed approach has the potential to shed light on imputing CBF measures when ASL data are missing.

Supplementary Material

2

Footnotes

Appendix A. Supplementary data

Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.jneumeth.2019.108528.

References

  1. Alsop DC, Casement M, de Bazelaire C, Fong T, Press DZ, 2008. Hippocampal hyperperfusion in Alzheimer’s disease. Neuroimage 42, 1267–1274. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Alsop DC, Dai W, Grossman M, Detre JA, 2010. Arterial spin labeling blood flow MRI: its role in the early characterization of Alzheimer’s disease. J. Alzheimers Dis 20, 871–880. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Alsop DC, Detre JA, Golay X, Gunther M, Hendrikse J, Hernandez-Garcia L, Lu H, Macintosh BJ, Parkes LM, Smits M, van Osch MJ, Wang DJ, Wong EC, Zaharchuk G, 2015. Recommended implementation of arterial spin-labeled perfusion MRI for clinical applications: a consensus of the ISMRM perfusion study group and the European consortium for ASL in dementia. Magn. Reson. Med 73, 102–116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Andersson JLR, Jenkinson M, Smith SM, 2010. Non-linear registration, aka spatial normalization. FMRIB. [Google Scholar]
  5. Biswal B, Haughton VM, Hyde J, 1995. Functional connectivity in the motor cortex resting human brain using echo-planar MRI. Magn. Reson. Med 34, 537–541. [DOI] [PubMed] [Google Scholar]
  6. Biswal BB, Mennes M, Zuo XN, Gohel S, Kelly C, Smith SM, Beckmann CF, Adelstein JS, Buckner RL, Colcombe S, Dogonowski AM, Ernst M, Fair D, Hampson M, Hoptman MJ, Hyde JS, Kiviniemi VJ, Kotter R, Li SJ, Lin CP, Lowe MJ, Mackay C, Madden DJ, Madsen KH, Margulies DS, Mayberg HS, McMahon K,, Monk CS, Mostofsky SH, Nagel BJ, Pekar JJ, Peltier SJ, Petersen SE, Riedl V, Rombouts SA, Rypma B, Schlaggar BL, Schmidt S, Seidler RD, Siegle GJ, Sorg C, Teng GJ, Veijola J, Villringer A, Walter M, Wang L, Weng XC, Whitfield-Gabrieli S, Williamson P, Windischberger C, Zang YF, Zhang HY, Castellanos FX, Milham MP, 2010. Toward discovery science of human brain function. Proc. Natl. Acad. Sci. U. S. A 107, 4734–4739. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Bruandet A, Richard F, Bombois S, Maurage CA, Deramecourt V, Lebert F, Amouyel P, Pasquier F, 2009. Alzheimer disease with cerebrovascular disease and vascular dementia: clinical features and course compared with Alzheimer disease. J. Neurol. Neurosurg. Psychiatry 80, 133–139. [DOI] [PubMed] [Google Scholar]
  8. Buzsaki G, 2004. Neuronal oscillations in cortical networks. Science 304, 1926–1929. [DOI] [PubMed] [Google Scholar]
  9. Chand GB, Hajjar I, Qiu D, 2018. Disrupted interactions among the hippocampal, dorsal attention, and central-executive networks in amnestic mild cognitive impairment. Hum. Brain Mapp 39, 4987–4997. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Chang CC, Lin CJ, 2011. LIBSVM : a library for support vector machines. ACM Trans. Intell. Syst. Technol 2. [Google Scholar]
  11. Chen Y, Wolk DA, Reddin JS, Korczykowski M, Martinez PM, Musiek ES, Newberg AB, Julin P, Arnold SE, Greenberg JH, Detre JA, 2011. Voxel-level comparison of arterial spin-labeled perfusion MRI and FDG-PET in Alzheimer disease. Neurol 77, 1977–1985. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Clark LR, Berman SE, Rivera-Rivera LA, Hoscheidt SM, Darst BF, Engelman CD, Rowley HA, Carlsson CM, Asthana S, Turski P, Wieben O, Johnson SC, 2017. Macrovascular and microvascular cerebral blood flow in adults at risk for Alzheimer’s disease. Alzheimers Dement. Amst. (Amst) 7, 48–55. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Cortes C, Vapnik V, 1995. Support-vector networks. Mach. Learn 20, 273–297. [Google Scholar]
  14. Culham JC, Valyear KF, 2006. Human parietal cortex in action. Curr. Opin. Neurobiol 16, 205–212. [DOI] [PubMed] [Google Scholar]
  15. Dagli MS, Ingeholm JE, Haxby JV, 1999. Localization of cardiac-induced signal change in fMRI. Neuroimage 9, 407–415. [DOI] [PubMed] [Google Scholar]
  16. Dai W, Lopez OL, Carmichael OT, Becker JT, Kuller LH, Gach HM, 2009. Mild cognitive impairment and Alzheimer disease: patterns of altered cerebral blood flow at MR imaging. Radiol 250, 856–866. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Detre JA, Leigh JS, Williams DS, Koretsky AP, 1992. Perfusion imaging. Magn. Reson. Med 23, 37–45. [DOI] [PubMed] [Google Scholar]
  18. Detre JA, Rao H, Wang DJJ, Chen YF, Wang Z, 2012. Applications of arterial spin labeled MRI in the brain. J. Magn. Reson. Imaging 35, 1026–1037. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Detre JA, Wang J, 2002. Technical aspects and utility of fMRI using BOLD and ASL. Clinical Neurophysiol 113, 621–634. [DOI] [PubMed] [Google Scholar]
  20. Dolui S, Tisdall D, Vidorreta M, Jacobs DR Jr., Nasrallah IM, Bryan RN, Wolk DA, Detre JA, 2019. Characterizing a perfusion-based periventricular small vessel region of interest. Neuroimage Clin. 23, 101897. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Dolui S, Vidorreta M, Wang Z, Nasrallah IM, Alavi A, Wolk DA, Detre JA, 2017. Comparison of PASL, PCASL, and background-suppressed 3D PCASL in mild cognitive impairment. Hum. Brain Mapp 38, 5260–5273. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Dolui S, Wang Z, Wang DJJ, Mattay R, Finkel M, Elliott M, Desiderio L, Inglis B, Mueller B, Stafford RB, Launer LJ, Jacobs DR Jr., Bryan RN, Detre JA, 2016. Comparison of non-invasive MRI measurements of cerebral blood flow in a large multisite cohort. J. Cereb. Blood Flow Metab 36, 1244–1256. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Doshi J, Erus G, Ou Y, Davatzikos C, 2013. Ensemble-based medical image labeling via sampling morphological appearance manifold In: MICCAI Challenge Workshop on Segmentation: Algorithms, Theory and Applications. Nagoya, Japan.. [Google Scholar]
  24. Doshi J, Erus G, Ou Y, Resnick SM, Gur RC, Gur RE, Satterthwaite TD, Furth S, Davatzikos C, 2016. Alzheimer’s Neuroimaging I. MUSE: MUlti-atlas region Segmentation utilizing Ensembles of registration algorithms and parameters, and locally optimal atlas selection. NeuroImage 127, 186–195. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Drucker H, Burges CJ, Kaufman L, Smola A, Vapnik V, 1997. Support vector regression machines Advances in Neural Information Processing Systems 9. MIT Press, Cambridge, Mass, pp. 155–161. [Google Scholar]
  26. Duvernoy HM, Delon S, Vannson JL, 1981. Cortical blood vessels of the human brain. Brain Res. Bull 7, 519–579. [DOI] [PubMed] [Google Scholar]
  27. Fan AP, Jahanian H, Holdsworth SJ, Zaharchuk G, 2016. Comparison of cerebral blood flow measurement with [15O]-water positron emission tomography and arterial spin labeling magnetic resonance imaging: a systematic review. J. Cereb. Blood Flow Metab 36, 842–861. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Feng CM, Narayana S, Lancaster JL, Jerabek PA, Arnow TL, Zhu F, Tan LH, Fox PT, Gao JH, 2004. CBF changes during brain activation: fMRI vs. PET. Neuroimage 22, 443–446. [DOI] [PubMed] [Google Scholar]
  29. Fogassi L, Ferrari PF, Gesierich B, Rozzi S, Chersi F, Rizzolatti G, 2005. Parietal lobe: from action organization to intention understanding. Science 308, 662–667. [DOI] [PubMed] [Google Scholar]
  30. Fox MD, Buckner RL, Liu H, Chakravarty MM, Lozano AM, Pascual-Leone A, 2014. Resting-state networks link invasive and noninvasive brain stimulation across diverse psychiatric and neurological diseases. Proc. Natl. Acad. Sci. U. S. A 111, E4367–75. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME, 2005. The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proc. Natl. Acad. Sci. U. S. A 102, 9673–9678. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Fransson P, 2005. Spontaneous low-frequency BOLD signal fluctuations: an fMRI in-vestigation of the resting-state default mode of brain function hypothesis. Hum. Brain Mapp 26, 15–29. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Greicius MD, Kimmel DL, 2012. Neuroimaging insights into network-based neurodegeneration. Curr. Opin. Neurol 25, 727–734. [DOI] [PubMed] [Google Scholar]
  34. Gupta A, Nair S, Schweitzer AD, Kishore S, Johnson CE, Comunale JP, Tsiouris AJ, Sanelli PC, 2012. Neuroimaging of cerebrovascular disease in the aging brain. Aging Dis. 3 (5), 414–425. [PMC free article] [PubMed] [Google Scholar]
  35. Habes M, Erus G, Toledo JB, Zhang T, Bryan N, Launer LJ, Rosseel Y, Janowitz D, Doshi J, Van der Auwera S, von Sarnowski B, Hegenscheid K, Hosten N, Homuth G, Volzke H, Schminke U, Hoffmann W, Grabe HJ, Davatzikos C, 2016. White matter hyperintensities and imaging patterns of brain ageing in the general population. Brain 139, 1164–1179. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Haller S, Zaharchuk G, Thomas DL, Lovblad KO, Barkhof F, Golay X, 2016. Arterial spin labeling perfusion of the brain: emerging clinical applications. Radiol 281 (2), 337–356. [DOI] [PubMed] [Google Scholar]
  37. Huang CW, Hsu SW, Chang YT, Huang SH, Huang YC, Lee CC, Chang WN, Lui CC, Chen NC, Chang CC, 2018. Cerebral perfusion insufficiency and relationships with cognitive deficits in alzheimer’s disease: a multiparametric neuroi-maging study. Sci. Rep 8, 1541. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Johnson NA, Jahng GH, Weiner MW, Miller BL, Chui HC, Jagust WJ, Gorno-Tempini ML, Schuff N, 2005. Pattern of cerebral hypoperfusion in Alzheimer disease and mild cognitive impairment measured with arterial spin-labeling MR imaging: initial experience. Radiol 234, 851–859. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Kilroy E, Apostolova L, Liu C, Yan L, Ringman J, Wang DJJ, 2014. Reliability of two-dimensional and three-dimensional pseudo-continuous arterial spin labeling perfusion MRI in elderly populations: comparison with 15o-water positron emission tomography. J. Magn. Reson. Imaging 39, 931–939. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Kwong KK, Belliveau JW, Chesler DA, Goldberg IE, Weisskoff RM, Poncelet BP, Kennedy DN, Hoppel BE, Cohen MS, Turner R, Cheng HM, Brady TJ, Rosen BR, 1992. Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proc. Natl. Acad. Sci. U. S. A 89, 5675–5679. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Li Z, Zhu Y, Childress AR, Detre JA, Wang Z, 2012. Relations between BOLD fMRI-derived resting brain activity and cerebral blood flow. PLoS One 7, e44556. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Liang X, Zou Q, He Y, Yang Y, 2013. Coupling of functional connectivity and regional cerebral blood flow reveals a physiological basis for network hubs of the human brain. Proc. Natl. Acad. Sci. U. S. A 110, 1929–1934. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Liebeskind DS, 2003. Collateral circulation. Stroke 34, 2279–2284. [DOI] [PubMed] [Google Scholar]
  44. Limbrick-Oldfield EH, Brooks JC, Wise RJ, Padormo F, Hajnal JV, Beckmann AF, Ungless MA, 2012. Identification and characterisation of midbrain nuclei using optimised functional magnetic resonance imaging. Neuroimage 59,1230–1238. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Love S, Miners JS, 2016. Cerebrovascular disease in ageing and Alzheimer’s disease. Acta. Neuropathol 131 645–658. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Lv Y, Margulies DS, Cameron Craddock R, Long X, Winter B, Gierhake D, Endres M, Villringer K, Fiebach J, Villringer A, 2013. Identifying the perfusion deficit in acute stroke with resting-state functional magnetic resonance imaging. Ann. Neurol 73, 136–140. [DOI] [PubMed] [Google Scholar]
  47. MacDonald ME, Frayne R, 2015. Cerebrovascular MRI: a review of state-of-the-art approaches, methods and techniques. NMR Biomed. 28, 767–791. [DOI] [PubMed] [Google Scholar]
  48. Miller EK, Cohen JD, 2001. An integrative theory of prefrontal cortex function. Annu. Rev. Neurosci 24, 167–202. [DOI] [PubMed] [Google Scholar]
  49. Ogawa S, Menon RS, Tank DW, Kim SG, Merkle H, Ellermann JM, Ugurbil K, 1993. Functional brain mapping by blood oxygenation level-dependent contrast magnetic resonance imaging. A comparion of signal characteristics with a biophysical model. Biophys. J 64, 803–812. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Ogawa S, Tank DW, Menon R, Ellermann JM, Kim SG, Merkle H, Ugurbil K, 1992. Intrinsic signal changes accompanying sensory stimulation: functional brain mapping with magnetic resonance imaging. Proc. Natl. Acad. Sci. U. S. A 89, 5951–5955. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Petersen RC, 2004. Mild cognitive impairment as a diagnostic entity. J. Intern. Med 256, 183–194. [DOI] [PubMed] [Google Scholar]
  52. Petersen RC, Roberts RO, Knopman DS, Boeve BF, Geda YE, Ivnik RJ, Smith GE, Jack CR Jr., 2009. Mild cognitive impairment. Arch. Neurol 66 (12), 1447–1455. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Petrides M, Pandya DN, 1999. Dorsolateral prefrontal cortex: comparative cytoarchitectonic analysis in the human and the mamaque brain and corticocortical connection patterns. Eur. J. Neurosci 11, 1011–1036. [DOI] [PubMed] [Google Scholar]
  54. Raichle ME, 1998. Behind the scenes of functional brain imaging: a historical and physiological perspective. Proc. Natl. Acad. Sci. U.S. A 95, 765–772. [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Raichle ME, 2015. The restless brain: how intrinsic activity organizes brain function. Philos. Trans. R. Soc. Lond., B, Biol. Sci 370. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Raj D, Anderson AW, Gore JC, 2001. Respiratory effects in human functional magnetic resonance imaging due to bulk susceptibility changes. Phys. Med. Biol 46, 3331–3340. [DOI] [PubMed] [Google Scholar]
  57. Tong Y, Frederick B, 2014. Tracking cerebral blood flow in BOLD fMRI using recur-sively generated regressors. Hum. Brain Mapp 35, 5471–5485. [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Tong Y, Lindsey KP, Hocke LM, Vitaliano G, Mintzopoulos D, Frederick BD, 2017. Perfusion information extracted from resting state functional magnetic resonance imaging. J. Cereb. Blood Flow Metab 37, 564–576. [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Verfaillie SC, Adriaanse SM, Binnewijzend MA, Benedictus MR, Ossenkoppele R, Wattjes MP, Pijnenburg YA, van der Flier WM, Lammertsma AA, Kuijer JP, Boellaard R, Scheltens P, van Berckel BN, Barkhof F, 2015. Cerebral perfusion and glucose metabolism in Alzheimer’s disease and frontotemporal dementia: two sides of the same coin? Eur. Radiol 25, 3050–3059. [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Wang Z, 2012. Improving cerebral blood flow quantification for arterial spin labeled perfusion MRI by removing residual motion artifacts and global signal fluctuations. Magn. Reson. Imaging 30, 1409–1415. [DOI] [PMC free article] [PubMed] [Google Scholar]
  61. Wang Z, Das SR, Xie SX, Arnold SE, Detre JA, Wolk DA, 2013. Arterial spin labeled MRI in prodromal Alzheimer’s disease: a multi-site study. Neuroimage Clin. 2, 630–636. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Winblad B, Palmer K, Kivipelto M, Jelic V, Fratiglioni L, Wahlund LO, Nordberg A, Backman L, Albert M, Almkvist O, Arai H, Basun H, Blennow K, De Leon M, Decarli C, Erkinjuntti T, Giacobini E, Graff C, Hardy J, Jack C, Jorm A, Ritchie K, Van Duijn C, Visser P, Petersen RC, 2004. Mild cognitive impairment — beyond controversies, towards a consensus: report of the international working group on mild cognitive impairment. J. Intern. Med 256, 240–246. [DOI] [PubMed] [Google Scholar]
  63. Wolk DA, Detre JA, 2012. Arterial spin labeling MRI: an emerging biomarker for Alzheimer’s disease and otherneurodegenerative conditions. Curr Open Neurol 25, 421–428. [DOI] [PMC free article] [PubMed] [Google Scholar]
  64. Xie L, Dolui S, Das SR, Stockbower GE, Daffner M, Rao H, Yushkevich PA, Detre JA, Wolk DA, 2016. A brain stress test: cerebral perfusion during memory encoding in mild cognitive impairment. Neuroimage Clin. 11, 388–397. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Yan S, Qi Z, An Y, Zhang M, Qian T, Lu J, 2018. Detecting perfusion deficit in Alzheimer’s disease and mild cognitive impairment patients by resting-state fMRI. J. Magn. Reson. Imaging. [DOI] [PubMed] [Google Scholar]
  66. Ye FQ, Berman KF, Ellmore T, Esposito G, van Horn JD, Yang Y, Duyn J, Smith AM, Frank JA, Weinberger DR, McLaughlin AC, 2000. H(2)(15)O PET validation of steady-state arterial spin tagging cerebral blood flow measurements in humans. Magn. Reson. Med 44, 450–456. [DOI] [PubMed] [Google Scholar]
  67. Zou Q, Miao X, Liu D, Wang DJJ, Zhuo Y, Gao J-H, 2015. Reliability comparison of spontaneous brain activities between BOLD and CBF contrasts in eyes-open and eyes-closed resting states. NeuroImage 121, 91–105. [DOI] [PubMed] [Google Scholar]
  68. Zuo X-N, Di Martino A, Kelly C, Shehzad ZE, Gee DG, Klein DF, Castellanos FX, Biswal BB, Milham MP, 2010. The oscillating brain: complex and reliable. NeuroImage 49, 1432–1445. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

2
NIHMS1580786-supplement-2.docx (752.3KB, docx)

RESOURCES