3607

Resting-state diffusion fMRI bears strong resemblance and only subtle differences to BOLD fMRI
Ileana Ozana Jelescu1

1Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland

Synopsis

Diffusion fMRI (dfMRI) is a presumably non-BOLD technique sensitive to transient microstructural changes underlying neural activity. Previous task-fMRI studies have assessed the characteristics of the dfMRI signal and potential BOLD contamination, with conflicting results. Here we acquired resting-state fMRI data with five protocols with incrementally reduced BOLD contributions and analyzed the characteristics of resting-state networks and functional connectivity using model-free approaches. We report dfMRI data does not contain fundamentally different information to BOLD-fMRI, with the exception of a few regions that switch from anti- to positively-correlated. Future work will focus on removing any remaining BOLD contribution from the dfMRI acquisition.

Introduction

Diffusion functional MRI (dfMRI) has been proposed as a non-BOLD technique that relies on subtle microstructural changes related to neural activity (e.g. cell swelling) to induce changes in the overall apparent diffusion coefficient (ADC)1. This method has been suggested as more time- and region-specific to neural activity than BOLD fMRI which relies on neurovascular coupling. A body of conflicting evidence has been published since, with studies showing improved temporal and spatial specificity of the dfMRI signal compared to BOLD2,3, and others a vanishing of such signal once potential BOLD-related confounds – residual T2-weighting, perfusion component, cross-terms between blood susceptibility gradients and diffusion gradients – have been corrected4,5. One major limitation of dfMRI studies reported so far is the model-based approach of task-fMRI, assuming ad hoc response functions convolved with an external paradigm, appropriate modeling of physiological noise, and testing one neural pathway at a time. Here, we compare resting-state dfMRI and BOLD fMRI experiments for the first time, and use independent component analysis (ICA) to identify neural networks and compute functional connectivity in an entirely data-driven approach.

Methods

All experiments were approved by the local Service for Veterinary Affairs and performed on a 14T Varian system using a home-built surface quadrature transceiver. Sprague-Dawley rats (N=3) were anesthetized using isoflurane for initial setup and promptly switched to medetomidine sedation (bolus: 0.1mg/kg, perfusion: 0.1mg/kg/h), which preserves neural activity and vascular response better than isoflurane. An anatomical whole-brain scan was acquired for atlas-registration (0.15x0.15x0.50mm3 resolution). Five fMRI protocols, designed to contain gradually reduced sources of BOLD contrast, were compared (Fig. 1): gradient-echo EPI (GE-EPI, #1), spin-echo EPI (SE-EPI, #2), diffusion-weighted SE-EPI (DW SE-EPI, #3), ADC derived from alternating b=0 and b=1ms/μm2 images (ADC b=0/1, #4), and ADC derived from alternating b=0.4 and b=1ms/μm2 images (ADC b=0.4/1, #5). Each block of five protocols was acquired twice on each rat.

Data pre-processing included denoising6, slice-timing correction and spatial smoothing7 (Gaussian kernel: 0.36x0.36x1mm3). For the protocols alternating two b-value acquisitions, ADC timecourse was calculated as $$$ADC=\ln\left(S_{2}/S_{1}\right)/\left(b_{1}-b_{2}\right)$$$. Each signal timecourse (protocols #1-3) and ADC timecourse (protocols #4-5) was analyzed using FSL’s MELODIC8 with high-pass temporal filtering (f>0.01Hz) and 30 independent components (IC’s). Group-ICA was further performed for each protocol, using time-concatenation of all runs from all rats (N=6). Anatomically-relevant resting-state networks (RSNs) were compared between protocols. In parallel, noise-related IC’s were identified and cleaned from each individual fMRI run9 and the “cleaned” fMRI data were used to compute individual functional connectivity (fc) matrices between 18 atlas-defined ROIs, using global signal regression. Mean fc matrix for each protocol was calculated.

Results

Using group-ICA, three bilateral RSN’s – S1/S2, S1 barrel field and CPu – were identified across all protocols (Fig. 2) and could also be detected at the individual level (Fig. 3). Mean fc matrices for each protocol are shown in Fig. 4, along with ANOVA analysis across spin-echo based protocols (#2-5).

Discussion and Conclusions

We identified at least three RSN’s based on ICA analysis of resting-state dfmri data, whose spatial extent and temporal characteristics were highly comparable to GE-EPI or SE-EPI BOLD rs-fmri. In the low b-value regime, the overall ADC can be expressed as a weighted sum of local ADC’s, which justifies the use of ICA analysis on quantitative ADC timecourses. Functional connectivity matrices were also very similar between all spin-echo based protocols (with or without diffusion weighting).

Protocol #5 (ADC b=0.4/1) should have limited contributions to functional contrast from: T2-weighting (both images are weighted similarly) and perfusion (fast-flowing spins should be suppressed at b≥0.4ms/μm2) and should be closest to potential non-BOLD dfmri signal. It is noteworthy that the significant differences in fc matrices between SE-EPI and ADC b=0.4/1 were driven not by the magnitude of the correlation, but by the change in sign – with anti-correlated ROIs in SE-EPI being positively correlated in terms of ADC changes. This could suggest a genuinely different source of functional contrast in dfmri.

The similarity to SE-EPI in terms of RSN’s and fc matrix raises nonetheless the question of residual BOLD contrast. BOLD could stem from variations in T2 between the acquisition of successive b=0.4 and b=1 images – which could be minimized using an INDI scheme10 – or cross-terms between venous susceptibility gradients and diffusion gradients, potentially significant at ultra-high field but which could be reduced using a bipolar gradient design. Future work will include minimization of aforementioned confounds and suppression of vascular response using nitroprusside2.

Brain region abbreviations

ACC: Anterior cingulate cortex; RSC: Retrosplenial cortex; S1/S2: primary/secondary somatosensory cortex; M: motor cortex; Tha: thalamus; HTh: hypothalamus; CPu: dorsal striatum; NAc: ventral striatum; L/R: left/right.

Acknowledgements

The author thanks Olivier Reynaud for insightful discussions, and Analina da Silva, Mario Lepore and Stefan Mitrea for technical assistance with animal setup and monitoring. This work was supported by Centre d'Imagerie Biomédicale (CIBM), and the Leenaards and Jeantet Foundations.

References

1. Le Bihan, D., Urayama, S.-i., Aso, T., Hanakawa, T., Fukuyama, H., 2006. Direct and fast detection of neuronal activation in the human brain with diffusion MRI. Proceedings of the National Academy of Sciences 103, 8263.

2. Tsurugizawa, T., Ciobanu, L., Le Bihan, D., 2013. Water diffusion in brain cortex closely tracks underlying neuronal activity. Proceedings of the National Academy of Sciences 110, 11636.

3. Nunes, D., Ianus, A., Shemesh, N., 2019. Layer-specific connectivity revealed by diffusion-weighted functional MRI in the rat thalamocortical pathway. Neuroimage 184, 646-657.

4. Miller, K.L., Bulte, D.P., Devlin, H., Robson, M.D., Wise, R.G., Woolrich, M.W., Jezzard, P., Behrens, T.E.J., 2007. Evidence for a vascular contribution to diffusion FMRI at high b-value. Proceedings of the National Academy of Sciences 104, 20967.

5. Jin, T., Kim, S.-G., 2008. Functional changes of apparent diffusion coefficient during visual stimulation investigated by diffusion-weighted gradient-echo fMRI. Neuroimage 41, 801-812.

6. Veraart, J., Novikov, D.S., Christiaens, D., Ades-Aron, B., Sijbers, J., Fieremans, E., 2016. Denoising of diffusion MRI using random matrix theory. Neuroimage 142, 394-406.

7. Friston, K.J., Penny, W., Ashburner, J., Kiebel, S., Nichols, T.E., 2007. Statistical Parametric Mapping. Elsevier.

8. Beckmann, C.F., Smith, S.M., 2004. Probabilistic independent component analysis for functional magnetic resonance imaging. IEEE Trans Med Imaging 23, 137-152.

9. Griffanti, L., Douaud, G., Bijsterbosch, J., Evangelisti, S., Alfaro-Almagro, F., Glasser, M.F., Duff, E.P., Fitzgibbon, S., Westphal, R., Carone, D., Beckmann, C.F., Smith, S.M., 2017. Hand classification of fMRI ICA noise components. Neuroimage 154, 188-205.

10. Ianus, A., Shemesh, N., 2018. Incomplete initial nutation diffusion imaging: An ultrafast, single-scan approach for diffusion mapping. Magn Reson Med 79, 2198-2204.

11. Schwarz, A.J., Gass, N., Sartorius, A., Risterucci, C., Spedding, M., Schenker, E., Meyer-Lindenberg, A., Weber-Fahr, W., 2013. Anti-correlated cortical networks of intrinsic connectivity in the rat brain. Brain connectivity 3, 503-511.

Figures

Fig.1: Sequence parameters for the five protocols tested. GE-EPI (#1) yields T2* BOLD contrast. SE EPI (#2) yields T2 BOLD contrast. DW SE-EPI (#3) combines T2 contrast with diffusion contrast but eliminates perfusion contrast. Interleaved b=0 and b=1 ms/μm2 acquisitions for ADC estimation (ADC b=0/1, #4) largely eliminate T2 contrast but preserve perfusion and diffusion contrast. Interleaved b=0.4 and b=1 ms/μm2 acquisitions (ADC b=0.4/1, #5) largely eliminate T2 and perfusion contrast, retaining diffusion contrast. Protocols #3-5 can however display pseudo-BOLD contrast due to cross-terms between diffusion gradients and venous blood susceptibility gradients, particularly for a PGSE sequence.

Fig.2: Example S1/S2 bilateral RSN identified by group ICA in each of the five protocols. This network is spatially consistent across all protocols. The maximum z-score is lower for the ADC-based timecourses due to the lower SNR and temporal resolution (3.2s instead of 1.6s). This could potentially be addressed by using a sliding window of two consecutive images (instead of separate blocks of two images) to reconstruct ADC with the same temporal resolution as protocols #1-3, but might introduce temporal smoothing as each image would be used for the ADC computation of two timepoints.

Fig.3: The S1/S2 bilateral RSN was identifiable at the level of individual-run ICA analysis in each of the five protocols (leftmost column). The statistical power is however gradually reduced from protocols #1-2 down to #5, due to lower SNR, lower temporal resolution and reduced sources of functional contrast. The timecourses (middle column) and power spectra (rightmost column) of the independent component specific to this RSN are similar between protocols and essentially free from physiological artifacts (with the exception of the SE-EPI one that shows residual cardiac or respiratory artifact aliased around 0.01 and 0.05 Hz).

Fig.4: A-E: Average functional connectivity matrices estimated from each of the five protocols. A: GE-EPI shows strongest positive correlations between left and right hemispheres and between functional sub-units (ACC and RSC; Tha and HTh; CPu and NAc), as well as anti-correlations between S2 and motor cortex11. B-E: Matrices are very similar to each other. F: Differences in fc between protocols #2-5 (B-E) assessed using ANOVA. These differences involve mainly S2 and the dorsal striatum, and remained significant in paired comparison of protocol #3 (C) and #5 (E) with Tukey-Cramer correction. *: p<3.3E-04 (Bonferroni corrected threshold for 153 comparisons).

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
3607