Cerebrovascular reactivity mapping using breath-hold BOLD-fMRI: Comparison of signal models combined with voxelwise lag optimization

Abstract

Cerebrovascular reactivity (CVR) can be mapped noninvasively using blood oxygenation level dependent (BOLD) fMRI during a breath-hold (BH) task. Previous studies showed that the BH BOLD response is best modeled as the convolution of the partial pressure of end-tidal CO2 (PetCO2) with a canonical hemodynamic response function (HRF). However, previous model comparisons employed a global bulk time lag, which is now well accepted to provide only a rough approximation of the heterogeneous distribution of response latencies across the brain. Here, we investigate the best modeling approach for mapping CVR based on BH BOLD-fMRI data, when using a lagged general linear model approach for voxelwise lag optimization. In a group of 14 healthy participants, we compared two types of regressors (PetCO2 and Block), and three convolution models (no convolution; convolution with a single gamma HRF; and convolution with a double gamma HRF), as well as a range of HRF delays and dispersions (for models with convolution). Convolution with a single gamma HRF yielded the greatest CVR values in PetCO2 models, while a double gamma HRF performed better for block models. Although PetCO2-based regressors generally outperformed block-based regressors, as expected, the latter may be an appropriate alternative in cases of poor CO₂ recordings. Overall, our results support the use of specific modeling approaches for CVR mapping based on end-expiration BH BOLD-fMRI, including the voxelwise optimization of the lag.

Keywords: BOLD-fMRI; PetCO2; breath-hold; cerebrovascular reactivity; hemodynamic response function; lag optimization.

Fig. 1.

Illustration of the breath-hold (BH) task paradigm, with the respective parameters. Each BH was performed following an expiration, and it was followed by an exhalation and a free breathing recovery period. The baseline consisted in cued breathing at the subject’s breathing rate (as assessed in a calibration task before scanning). Adapted from Pinto et al. (2021).

Fig. 2.

Illustration of the ∆PetCO2 calculation: CO₂ signal for one representative subject, overlaid on the BH paradigm.

Fig. 3.

Breath-hold BOLD-fMRI data analysis pipeline, including the different models tested.

Fig. 4.

Hemodynamic response functions (HRFs) considered in the analysis: single-gamma (left) and double-gamma (right), for a range of delays/dispersions.

Fig. 5.

Individual responses to the breath-hold paradigm: average GM BOLD signals overlaid with the PetCO2 and Block regressors for all subjects. Both regressors were convolved with a single gamma HRF (delay = 6 s) and plotted considering the respective bulk lag that maximizes the correlation (with the value indicated in the figure by the corr_p,black and corr_b,blue, for PetCO2 and block, respectively, and an average correlation across subjects equal to corr_p = 0.78 and corr_b = 0.83) between them and the average GM BOLD signal. The percent signal change was computed for the BOLD signal. The PetCO2 and the Block signals were normalized by demeaning the signal, dividing by the standard deviation, and multiplying by 100%.

Fig. 6.

Subject example (top) and group median (bottom) maps of cerebrovascular reactivity (CVR) (left) and relative lag (Lag_rel) obtained using the two regressor types (PetCO2 and Block) and the three convolution models (without convolution (WoC), convolution with single gamma (CSg), and convolution with double gamma (CDb), using the canonical delay of 6 s), for three representative slices in the MNI space.

Fig. 7.

(A) ROI analysis of F-values, averaged across GM (left) and WM (right), for each of the models tested: two regressor types (PetCO2, top, and Block, bottom); three convolution models (without convolution (WoC), with convolution with single gamma (CSg), and with convolution with double gamma (CDb)); and different HRFs delays for the models with convolution (3-11 s) and (B) respective detailed statistical differences in model fitting. Boxplots represent the interquartile range of the distributions across subjects with + denoting outliers, and significant pairwise differences between convolution models are indicated with *. Purple box indicates the without HRF convolution and the convolution with an HRF delay = 6 s, for the convolution with single and double gamma.

Fig. 8.

ROI analysis of CVR (first row) and Lag_rel (second row) values, averaged across GM (left) and WM (right), for the two regressor types (PetCO2 and Block) and the three convolution models (without convolution (WoC), with convolution with single gamma (CSg) and with convolution with double gamma (CDb), CDb)), with an HRF delay of 6 s for the convolution models). Boxplots represent the interquartile range of the distributions across subjects with + denoting outliers, and significant pairwise differences between convolution models are indicated with *. For CVR, all pairwise differences between regressor types (PetCO2 and Block) were statistically different (indicated in brown *).

Fig. 9.

Voxelwise group-level permutation testing between regressors type (PetCO2 vs. Block) for the three convolution methods (without convolution (WoC), with convolution with single gamma (CSg), and with convolution with double gamma (CDb)), in terms of F-value (first column), CVR (second column), relative lag (Lag_rel) (third column), and HRF delay (fourth column). Results are shown as 1-p-value maps for each comparison (both positive and negative changes) overlaid on MNI space T1-weighted image for four representative slices (colors representing significant differences placed over the MNI template). Permutation testing was performed using FSL’s Randomise, and the color bar represents the p-value with Family-Wise Error (FEW) correction; the p-value was thresholded after correction of the three multiple comparisons (p < 0.017, i.e., 1-0.017, p > 0.983).

Fig. 10.

Voxelwise group-level permutation testing between convolution methods for the two regressor types (PetCO2, top, and Block, bottom) for the three convolution methods (without convolution (WoC), with convolution with single gamma (CSg), and with convolution with double gamma (CDb)), in terms of F-value (first column), CVR (second column), relative lag (Lag_rel) (third column), and HRF delay (fourth column). Results are shown as 1-p-value maps for each comparison (both positive and negative changes) overlaid on MNI space T1-weighted image for four representative slices (colors representing significant differences placed over the MNI template). Permutation testing was performed using FSL’s Randomise, and the color bar represents the p-value with Family-Wise Error (FEW) correction; the p-value was thresholded after correction of the six multiple comparisons (p < 0.008, i.e., 1-0.008, p > 0.992).