Exploring visual network connectivity in the mouse brain using DCM fMRI
Arun Niranjan1, Peter Zeidman2, Jack A Wells1, and Mark F Lythgoe1

1Centre for Advanced Biomedical Imaging, University College London, London, United Kingdom, 2Institute of Neurology, University College London, London, United Kingdom

Synopsis

Understanding effective (i.e. causal) connectivity in the brain using fMRI with dynamic causal modelling (DCM) has attracted a large amount of interest in recent years. Applications of fMRI to map brain function in the mouse are on the rise, targeting transgenic mouse models of pathology. However, DCM has not yet been applied to mouse brain fMRI, in part due to the difficulties of acquiring high quality data. In this work we demonstrate the use of DCM fMRI to understand effective connectivity in the healthy mouse visual system, showing results consistent with the underlying biology.

Purpose

To explore effective connectivity in the mouse brain using visual task-based fMRI and dynamic causal modelling.

Introduction

Transgenic mice provide the opportunity to understand the direct effect of genes on brain structure and function. Mouse brain functional magnetic resonance imaging (fMRI) is technically challenging due to the small brain size (and as such the need for small voxels), and the image distortion due to B0 inhomogeneities. The visual pathway represents an important research target for mapping in clinical and basic science applications. In previous work1, we characterised the BOLD response to visual stimulation in the mouse brain, explored the benefits of snapshot GE-EPI2 to reduce image distortion, and demonstrated modulation of the BOLD response with respect to temporal frequency of visual stimulation (Figure 1C)1. Statistical parametric mapping successfully identified the lateral geniculate nuclei (LGN), the superior colliculus (SC) and the visual cortex (VC) as regions associated with a flashing light stimulus. However, standard mapping analysis provides no insight into mechanisms of visual processing. Dynamic causal modelling (DCM)3 is an analysis technique that uses a Bayesian framework to compare models of effective (i.e. causal) connectivity between brain regions, given BOLD time series. Whilst DCM has been used extensively in human fMRI analysis, there is limited existing work in the rat4 and none so far in the mouse. This work explores the use of DCM in control mice, and opens up the possibility of examining the effect of genetic influences on effective network connectivity via transgenic mouse models.

Methods

8 female C57BL6/J mice weighing (20.7 ± 0.7) g were used. Medetomidine anaesthesia (0.4 mg/kg bolus, 0.8 mg/kg constant infusion) was used during functional imaging. Respiration was approximately (170 ± 20) breaths per minute, and core body temperature maintained at (37.0 ± 0.2) °C. Visual stimulation was conducted bilaterally with blue laser light (445 nm, Omicron) transmitted into the scanner bore using a fibre optic cable. A block design was used (40 seconds of rest and 20 seconds of stimulus, 3 repeats per session). During stimulus blocks, the laser was pulsed at 1, 3, 5 or 10 Hz, with 10ms pulse duration. Subjects were scanned using a 9.4T MRI Scanner (Agilent Inc.) with an Agilent 205/120HD gradient set. GE-EPI was used with the following parameters: 4 compressed segments, 12 axial slices each 0.5 mm thick, slice gap 0.1 mm, FOV = 35 x 35 mm2, matrix size = 96 x 96, 83 volumes per session (TE/TR = 19/2500 ms) with 2 sessions per frequency.

Analysis

For each subject, all sessions of pre-processed EPI data were concatenated and modelled using a general linear model (Figure 2). This included regressors modelling the onset of optic stimulation (S) and the parametric effect of stimulus temporal frequency (f). Voxels responding to stimulation were identified by calculating an F-contrast over S and f. Representative BOLD time series were then extracted from each of LGN, SC and VC, within 0.6 mm spheres centered on the strongest clusters of activation (p<0.001 uncorrected). We then created a series of DCMs for each subject, to investigate which connections were modulated by f and which regions were driven by S. Each model had different combinations of connections between regions, modulations by f and driving effects of stimulus. Connections between regions were considered to be bi-directional, f was considered to either modulate nothing, the connection from the LGN to the VC, the connection from the SC to the VC, or the gain of the VC itself. The driving effect of stimulus was considered for all combinations of the three regions. In total this created a space of 168 models to be estimated and compared. A series of 3 random effects model comparisons were conducted, with the model space partitioned into ‘families’ according to the question being addressed.

Results

There was evidence for a fully connected network (Figure 3A), with frequency modulating the LGN->VC connection (Figure 3B), and with stimuli driving LGN and SC directly (Figure 3C). These results are summarized by the network diagram in Figure 3D, which is consistent with studies tracing projections of retinal ganglion cells throughout the mouse brain5 – in particular the mediation of the visual cortical response by mid-brain regions.

Conclusion

We demonstrate the ability of dynamic causal modelling to make inferences of effective connectivity in the mouse brain which are consistent with the underlying biology as evidenced by tracer studies5. This is the first application of DCM analysis to mouse brain fMRI data, and provides a platform for understanding genetic influences on effective connectivity networks.

Acknowledgements

This work is funded by the UK Medical Research Council (MR/K50077X/1)

References

1. Niranjan A., et al. (under review)

2. Guilfoyle D.N., et al. 2006.

3. Friston K.J., et al. 2003.

4. David, Olivier, et al. 2008.

5. Huberman, A.D., Niell, C.M., 2011.

Figures

Figure 1. Previous BOLD response characterisation of the mouse visual pathway (Niranjan et al., under review). A) Group FFX BOLD activation map overlaid on structural image (FWE corrected, two-tailed p < 0.05, 10 Hz stimulus frequency). B) Schematic of the mouse visual system adapted with permission [5]. C) Modulation of BOLD contrast with stimulus frequency in the SC and VC. Standard analyses show information on effect size and location, but not on how brain regions interact with each other.

Figure 2. Design matrix for DCM signal extraction. Column 1 is the convolution of visual stimulation (S) timings with the canonical haemodynamic response function, and column 2 is the parametric modulation of visual stimulation by temporal frequency (f). Columns 3-8 are concatenated motion regressors and columns 9-16 are session specific regressors.

Figure 3. (A-C) RFX Bayesian model selection using family comparisons. A) Family 8 corresponds to fully connected models. B) Family 2 corresponds to temporal frequency f modulating the LGN→VC connection. Families 1, 3 and 4 correspond to no frequency modulation, f modulation of the SC→VC connection, and f modulation of the VC self-connection respectively. C) Family 4 corresponds to models where driving input S enters the LGN and SC. D) Network representation of the resultant characteristics.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1675