Quantitative mapping of cerebrovascular reactivity using resting-state BOLD fMRI: Validation in healthy adults

Abstract

In conventional neuroimaging, cerebrovascular reactivity (CVR) is quantified primarily using the blood-oxygenation level-dependent (BOLD) functional MRI (fMRI) signal, specifically, as the BOLD response to intravascular carbon dioxide (CO2) modulations, in units of [%ΔBOLD/mmHg]. While this method has achieved wide appeal and clinical translation, the tolerability of CO2-related tasks amongst patients and the elderly remains a challenge in more routine and large-scale applications. In this work, we propose an improved method to quantify CVR by exploiting intrinsic fluctuations in CO2 and corresponding changes in the resting-state BOLD signal (rs-qCVR). Our rs-qCVR approach requires simultaneous monitoring of PETCO2, cardiac pulsation and respiratory volume. In 16 healthy adults, we compare our quantitative CVR estimation technique to the prospective CO2-targeting based CVR quantification approach (qCVR, the "standard"). We also compare our rs-CVR to non-quantitative alternatives including the resting-state fluctuation amplitude (RSFA), amplitude of low-frequency fluctuation (ALFF) and global-signal regression. When all subjects were pooled, only RSFA and ALFF were significantly associated with qCVR. However, for characterizing regional CVR variations within each subject, only the PETCO2-based rs-qCVR measure is strongly associated with standard qCVR in 100% of the subjects (p≤0.1). In contrast, for the more qualitative CVR measures, significant within-subject association with qCVR was only achieved in 50-70% of the subjects. Our work establishes the feasibility of extracting quantitative CVR maps using rs-fMRI, opening the possibility of mapping functional connectivity and qCVR simultaneously.

Keywords: Cerebrovascular reactivity; Deconvolution; End-tidal CO(2); Heart-rate variability; Hemodynamic response function; Non-invasive vascular mapping; Respiratory-volume variability; Resting-state fMRI.

Figure 1. Experimental set up

In addition to rs-fMRI data, our proposed CVR quantification methodology also relies on the recording of end-tidal CO₂ (PETCO₂), heart-rate and respiratory-volume traces during the rs-fMRI acquisition. These inputs are processed through our physiological modeling algorithm to produce quantitative CVR values.

Figure 2. Schematic of methodology and validation approach

Validation of the rs-fMRI-based quantitative CVR estimates are the focus of this work, and are shaded in red. The gold-standard measurement step, which serve as reference for the validation, is shaded in blue. Alternative rs-fMRI-based CVR metrics, primarily non-quantitative, are highlighted in orange.

Figure 3. Sample fits–cross-correlation method

A section of the time traces for the BOLD (single voxel), ΔPETCO₂ and the optimally shifted ΔPETCO₂ measurements are shown in (a). A sample linear fit of a single-voxel BOLD time series to the shifted ΔPETCO₂ time course is shown in (b), with the slope of the fit designated as the CVR. The preprocessed data were binned into equally spaced bins of 100 points each, in the interest of visualization. The error bars represent the variability in the BOLD signal across each of these bins.

Figure 4. Sample fits–parametric deconvolution method

A single-voxel BOLD time course (blue) overlaid by the fitted time course resulting form the convolution of ΔPETCO₂ and HRF_CO2 are shown in (a). The corresponding linear fit is shown (b), and CVR is determined as the slope of the fit. Again, The preprocessed data were binned into equally spaced bins of 100 points each, in the interest of visualization. The error bars represent the variability in the BOLD signal across each of these bins.

Figure 5. CVR maps obtained using various approaches

(a) Average across all 16 subjects; (b) a sample subject.

Figure 6. The correspondence between rs-fMRI and “gold-standard” CVR estimates, based on global grey-matter averages

Each symbol represents the grey-matter average from one of 16 subjects. All methods examined achieved significant agreement with the standard CVR values, although the strength of the correlations varied. The average from each subject excludes respective outlier parcellations for each method.

Figure 7. The correspondence between rs-fMRI and “gold-standard” CVR estimates, based on within-subject agreement in different brain regions, for a representative subject

Each symbol represents an ROI average CVR, with the number conveying the ROI definition (refer to Table 1). In this case, as is representative across the entire data set, while significant agreement was achieved using the cross-correlation and parametric-deconvolution methods, it was not the case for RSFA and ALFF.

Figure 8. Summary comparison of different rs-fMRI CVR methods

These results do not include data from the 3 subjects exhibiting pronounced motion. Results based on method-specific outlier ROI removals (Table 2) are shown in (a) and (b), while results associated with an identical ROI set for all methods (Table 3) are shown in (c) and (d). The common ROI set is described in Table 1. As seen in both (a) and (c), the parametric-deconvolution method was the only approach to yield quantitative CVR estimates that were in significant or near-significant agreement with the standard CVR measurements in 100% of the subjects, while the global-signal regression method is ranked second. Moreover, as seen in (b), when examining the scaling factor (slope) that related the rs-fMRI based CVR measurements to standard CVR measurements, the global-signal regression method demonstrated the least variability in scaling factors, followed by our parametric deconvolution method.

Figure 9. Reproducibility of resting-state CVR estimates

In a small group of subjects (N = 5), we compared test-retest reproducibility of resting-state CVR estimates through the intra-class correlation coefficient (ICC) and coefficient of variation (CV). For both accelerated and conventional rs-fMRI acquisitions, the ICC associated with the parametric-deconvolution exceeded 0.6, indicating reproducibility. In addition, the CV for our method ranges between 0.2 and 0.35, indicating that between-session variability amounts to up to 35% of the mean values. These CV values are rather higher than for the alternative methods, only to be exceeded by those of the global-signal regression approach.

Figure 10. Estimating CVR using accelerated vs. conventional rs-fMRI acquisitions

We compare HRF_CO2 estimated from our current multi-slice accelerated rs-fMRI acquisition scheme (TR = 0.38 s) to those based on conventional rs-fMRI acquisitions (TR = 2 s). (a) The shapes of the whole-brain average HRF_CO2 estimates are similar (group average, N = 8). The amplitude (b) and time of onset (TTP) (b) of the estimates are also found to be similar across 2 sessions for both conventional and accelerated acquisitions in various brain regions, as quantified through the between-session cosine similarity index (similarity threshold = 0.9, shown in green). Finally, as shown in (d), the cosine similarity between rs-qCVR estimates based on accelerated and conventional acquisitions is moderately high in all regions examined.