Conceived and designed the experiments: BRW AQB AZS BLS J-ML JPC. Performed the experiments: BRW AQB. Analyzed the data: BRW AQB AZS JPC. Contributed reagents/materials/analysis tools: BRW AQB AZS JPC. Wrote the paper: BRW AQB AZS BLS J-ML JPC.
The authors have declared that no competing interests exist.
Functional neuroimaging (e.g., with fMRI) has been difficult to perform in mice, making it challenging to translate between human fMRI studies and molecular and genetic mechanisms. A method to easily perform large-scale functional neuroimaging in mice would enable the discovery of functional correlates of genetic manipulations and bridge with mouse models of disease. To satisfy this need, we combined resting-state functional connectivity mapping with optical intrinsic signal imaging (fcOIS). We demonstrate functional connectivity in mice through highly detailed fcOIS mapping of resting-state networks across most of the cerebral cortex. Synthesis of multiple network connectivity patterns through iterative parcellation and clustering provides a comprehensive map of the functional neuroarchitecture and demonstrates identification of the major functional regions of the mouse cerebral cortex. The method relies on simple and relatively inexpensive camera-based equipment, does not require exogenous contrast agents and involves only reflection of the scalp (the skull remains intact) making it minimally invasive. In principle, fcOIS allows new paradigms linking human neuroscience with the power of molecular/genetic manipulations in mouse models.
The development of functional neuroimaging techniques, particularly functional magnetic resonance imaging (fMRI), has revolutionized human cognitive neuroscience
Neuroimaging of resting-state functional connectivity
An alternative method for functional neuroimaging in small animals is optical intrinsic signal imaging (OIS)
In this paper, we combine resting-state functional connectivity and OIS methods to perform novel functional connectivity optical intrinsic signal imaging (fcOIS) in mice. Using fcOIS, we demonstrate the first functional connectivity maps in mice covering almost the entirety of the convexity (from the olfactory bulb anteriorly to the superior colliculus posteriorly and laterally through primary somatosensory and auditory cortex). Having determined the patterns of functional connections, we show their utility through the use of functional connectivity data to parcellate the mouse cortex into functional areas. This mapping of the spatial arrangement and extent of multiple functional networks yields results in agreement with the expected neuroarchitecture. As fcOIS relies on simple and relatively inexpensive camera-based equipment and requires only the reflection of the scalp (making it minimally invasive), we expect that this method will be a widely useful tool, giving mouse researchers access to functional neuroimaging and allowing human neuroimagers to test hypotheses in standardized mouse models.
We mapped functional connectivity using a custom-built high-speed (30 Hz) OIS system (
(a) Illumination from sequentially flashing LEDs in four different wavelengths (478 nm, 588 nm, 610 nm, and 625 nm) arranged in a ring. Detection by an EMCCD camera is at 120 Hz (30 Hz after decoding of wavelengths). Crossed linear polarizers (not shown for simplicity) prevent artifacts from specular reflection off the skull. (b) A false color image of the mouse cortex generated from the red, yellow, and blue LED channels. The image shows the camera's field-of-view (approximately 1 cm2) of the mouse brain with the cerebral cortex visible through the skull from the olfactory bulb to the superior colliculus and far laterally on the convexity. In the corners, one can see the reflected skin flaps. The brain was manually segmented from the image providing a mask for fcOIS analysis.
Resting-state functional connectivity methods evaluate spatio-temporal correlation patterns in spontaneous brain activity (here viewed indirectly through the neurovascular response)
fcOIS correlation analysis revealed distinct resting-state networks. Comparison of seed time traces showed both high correlatations (
(a) Time traces (ΔHbO2) for three cortical locations: left retrosplenial (blue), right retrosplenial (green), and right motor (red). The right and left retrosplenial are functionally related and show time traces with high correlation (
Correlation maps for seeds chosen manually using the expected cortical positions of various functional areas (Mouse 1). Seed positions and sizes are shown with black circles. The scale for all correlation maps is from
Similar patterns are visible in all five mice scanned (
Seed-based functional connectivity maps are biased by the particular choice of the seed location. To address this issue, we also evaluated a method that is independent of user input. First we constructed the full correlation matrix, which contains correlation values between each pixel and every other pixel in the entire image. Because this produces many patterns that are essentially redundant, we used singular value decomposition (SVD) to find the predominant orthogonal functional connectivity patterns (the first four singular vectors are shown in
Having found the resting-state connection patterns, we wanted to use this information to divide the surface of the cortex into its component regions. Such a method will group cortical pixels based on similarities in their functional connectivity and will, ideally, regenerate the map of the expected neuroarchitecture. From the above functional connectivity maps, one can already observe borders around highly correlated regions. However, we desired a method to identify the borders from the maps without need for user input. Thus, we devised an automated parcellation scheme to take intuitive interpretations of the connectivity patterns and recreate them in a data-driven manner (see
The parcellation method divides up the cerebral surface into functional zones (
(a) The results of iterative parcellation using the first twenty singular vectors from the correlation matrix as an initial condition. We see clear delineation of a frontal/olfactory/cingulate (limbic) network (oranges), a motor network (reds), a somatosensory network (greens), a visual network (blue), the retrosplenial cortex (magenta), and the superior colliculus (light blues). Numbers on the parcels are arbitrary designations from the initial condition. (b) Dendrogram showing clustering of the parcels from their correlations. Each terminal branch is a parcel (numbered to match the parcellation image and color-coded based on functional assignments); parcels that are more closely related (i.e., that share similar correlation maps have branches that meet lower on the tree. Note the tight correlations within the frontal network, in turn connecting to first medial and then lateral motor areas. In total, there are main branches for all of the main networks we expect. (c) The Paxinos atlas applied to this mouse brain for comparison with the functional parcellation. (For the names of the different cytoarchitectural regions shown in the atlas, see
Each row and column corresponds to a parcel (labeled with a functional assignment, anatomic location, and a number that matches the scheme in
Performing parcellation analysis on resting-state functional connectivity data from multiple mice yields similar maps (
Different networks have been color-coded (green for somatosensory; red, motor; orange, frontal/cingulate/olfactory; magenta, retrosplenial; blue, visual; gray-blue, parietal; light blue, superior colliculus; purple, inferior colliculus; pink, auditory). Note that overall the patterns are similar across all the mice though there are slight individual differences in borders of the functional areas.
Although the exact shape of each region's borders differed between mice, these minor variations are consistent with the slight differences seen in the seed-based correlation maps (
We have shown (to our knowledge) the first results using optical intrinsic signal imaging to measure functional connectivity, and the first published mapping of functional connectivity in mice with resting-state hemodynamics. The findings of fcOIS were repeatable in time and also robust across multiple mice. These results satisfy our original goals of determining functional connections within the mouse brain in the resting state and of using the patterns of connections to generate a map of functionally distinct parcels. The functional neuroarchitecture found with fcOIS matches our expectations from previous studies in rats, primates, and humans as well as expectations that distinctions between functional regions should correspond to histological patterns
Bilaterally symmetric functional connectivity is a prominent feature of our mapping results in visual, somatosensory, motor, frontal, cingulate, and retrosplenial cortices, as well as the olfactory bulb and the superior colliculus (these being all of the major parts of the brain within our field-of-view; for an equivalent human result see Salvador
Once we were able to demonstrate the presence of resting-state functional connectivity networks in the OIS data, our goal was to use this data to recreate the functional divisions within the mouse cortex and to recreate parcellations found in histological atlases. Our iterative parcellation scheme followed by clustering is able to divide the brain into networks in a data-driven manner. This method robustly parcellates the brain into similar functional regions as are found in the histological atlas
In addition to the sensory and motor cortices, we also found functional connectivity (and associated parcellations) of higher-order cortical areas. Identifying these networks with resting-state neuroimaging is particularly noteworthy as developing task-paradigms to activate “cognitive” regions is difficult in the mouse. The olfactory, frontal, and cingulate cortices are all limbic areas
While, in the present analysis, we have focused on comparisons of large-scale functional distinctions (e.g., between retrosplenial and somatomotor regions), future methodological development could provide robust finer distinctions (e.g., between subdivisions of visual cortex). That such further differentiation might be possible is suggested by the interesting finding that the multiple parcels in somatosensory cortex (as in
Several potential improvements of fcOIS correlation mapping technique can be identified. For example, in this paper we used only ΔHbO2 as a contrast. While previous functional connectivity studies with optical techniques have shown similar mapping results using different hemoglobin species as contrasts
Additionally, while we used the same functional connectivity frequency band as in previous human and rat studies, the dependence of murine fcOIS on temporal filtering remains a question for future investigation. The use of different frequency bands could potentially capture fast vs. slow correlations that reveal the structure of the brain's information processing, as has been recently attempted in human fcMRI
Numerous studies have also shown that functional connectivity persists, albeit in modified form, under anesthesia
Future work is also needed to address fcOIS accuracy through direct comparisons to histology. The most direct comparison would be with activation studies using somatosensory (e.g., whisker, forelimb, hindlimb), auditory, or visual stimuli, as in the seminal report of Biswal
Alternatively, comparison could be made to histological staining in order to comprehensively define all functional brain regions, which would be the gold standard for quantifying differences in brain organization between mice. In such an analysis, if fcOIS maps were to differ from the histological atlas, steps would be need to be taken to determine whether the divergence was due to variation in the arrangement of cytoarchitecture, noise in the imaging method, or functional connectivity borders differing from histological borders. Although the parameter space for both the optimization and validation of fcOIS is large, these studies will be critical in establishing a firm foundation for fcOIS as a tool for routine mouse neuroscience.
Additionally, a focus of current fMRI research is how closely functional and structural connectivity are connected. In humans, structural connectivity can be assessed only indirectly using diffusion tensor imaging (DTI). While studies comparing fcMRI and DTI have shown reasonable agreement
We expect that advances in MRI technology and methods will eventually allow fMRI-based functional connectivity mapping in mice. However, the need for high-field MRI scanners will most likely restrict its use to dedicated neuroimaging researchers and centers. In contrast, fcOIS provides a combination of high resolution, low cost, and ease of use (a simple intraperitoneal injection of anesthetic and no thinning of the skull) that should enable many laboratories that previously did not consider functional neuroimaging to connect with on-going studies of human disease. One physical limitation of OIS (due to light scattering) is the restriction of the field-of-view to the cortical surface (<1 mm), which precludes direct mapping of deep brain structures (e.g., the thalamus and hippocampus). Thus, we expect the two methods to eventually play a complementary role where interesting results can be found “at the benchside” using fcOIS, and then a subsequent fcMRI study could be done to visualize deep brain structures and compare with high-resolution anatomic scans
In summary, we have demonstrated functional connectivity mapping with OIS in mice. Because we have determined that fcOIS is able to map both functional regions and their connections, this methodology should be a powerful tool for detecting when functional connectivity networks are disrupted (either in the distribution of the neuroarchitecture or in the pattern of connections). Thus, one could examine the functional consequences of disease models including genetic
All procedures were approved by the Washington University School of Medicine Animal Studies Committee (protocol # 20080216). Male Swiss Webster mice (6–10 weeks of age, 23–32 g, Harlan Laboratories) were anesthetized with a Ketamine/Xylazine mixture (86.9 mg/kg Ketamine, 13.4 mg/kg Xylazine) and allowed 30 minutes for anesthetic transition. Anesthetic effect was verified by ensuring that the animal was not responsive to a hind paw pinch. Once induced, the animal was placed on a heating pad maintained at 37°C (mTCII, Cell Microcontrols) and its head secured in a stereotactic frame using a nose cone and ear bars. The scalp fur was shaved and prepped, and a midline incision was made along the top of the head and the scalp was reflected, exposing approximately 1 cm2 of the skull. The skull was kept moist with an application of mineral oil before each scan. Arterial blood pressure from the left femoral artery was monitored using a blood pressure analyzer (Digi-Med, BPA 400a) and measured to be 95±10 mmHg (mean blood pressure +/− standard deviation averaged across mice 2–4; blood pressure data were not available for mouse 1). Scan times were 15, 30, 20, 30, and 30 minutes for mice 1–5 respectively.
Sequential illumination was provided at four wavelengths by a ring (diameter = 7 cm) of light emitting diodes (LEDs; 478 nm, 588 nm, 610 nm, and 625 nm; RLS-5B475-S, B5B-4343-TY, B5B435-30S, and OSCR5111A-WY, respectively, Roithner Lasertechnik) placed approximately 10 cm above the mouse's head. For image detection, we used a cooled, frame-transfer EMCCD camera (iXon 897, Andor Technologies) set to acquire via external triggering. The LED ring and the camera were time-synchronized and controlled via computer using custom-written software (MATLAB, Mathworks). To acquire images at a frame-rate well above the heart and respiration rates (∼10 and 2.5 Hz, respectively), we used a full frame rate of 30 Hz, which, with four temporally encoded wavelengths, required running the camera at a frame rate of 120 Hz. To prevent specular reflection from the surface of the mouse skull, crossed linear polarizers were placed just in front of the LEDs and the camera lens. A simplified diagram of the system is shown in
The secured mouse was placed at the focal plane of both the camera and the LED ring and held in place with a stereotactic holder. The field-of-view was adjusted to be approximately 1 cm2 square resulting in a field-of-view that covers the majority of the convexity of the cerebral cortex with anterior-posterior from the olfactory bulb to the superior colliculus (
Image light intensity was interpreted using the Modified Beer-Lambert Law: Φ(t) = Φ0*exp(-Δμa(t)*L). Here Φ(t) is the measured light intensity, Φ0 is the baseline light intensity (with no hemodynamic perturbation), Δμa(t) is the change in absorption coefficient due to changes in blood volume, and L is the path length of the photons in the tissue. With resting-state activity there is no pre-stimulus baseline and instead we normalized relative to the average light intensity: ΔΦ (t) = -ln(Φ(t)/<Φ0(t)>) = Δμa(t)*L. If we are only interested in intensity changes at a single wavelength, then there is no need to correct for the multiplicative constant, L. In order to perform spectroscopy, and recover Δμa(t), we used path length factors calculated using the analytical formula given by Arridge
We then converted absorption coefficient data to hemoglobin concentration changes using the spectroscopy matrix: the system of equations, Δμa,λ (t) = Eλ,i Δ[Hbi](t) (where E is the extinction coefficient matrix and
To create a false color “white light” image of the mouse brain, the first images from the red (625 nm), yellow (588 nm), and blue (478 nm) LED channels were normalized to a maximum value of one and then stored in the red, green and blue channels of an RGB image (
To guide seed placement, an atlas of the locations of cortical functional regions (as viewed from a superior projection of the convexity) was constructed using a histological atlas
In the atlas, we noted the position of the junction between the olfactory bulb and cerebrum along the midline and the position of the fissure between the superior colliculus and the cerebrum along the midline (which is also the position of lambda). These two points also were found in the “white light” mouse brain images. Using these two points, the atlas was affine-transformed to brain coordinates; this transform used only one stretch component (the anterior-posterior stretch was also used for the medial-lateral stretch). Then, every pixel in the mouse brain (as defined by the earlier mask) could be assigned to segmented cortical polygons from the atlas (as in
Since previous functional connectivity studies
Using the atlas as a reference, seed locations were chosen at coordinates expected to correspond to the left and right visual, motor, somatosensory, frontal, cingulate, and retrosplenial cortices as well as the right and left superior colliculi and olfactory bulbs. A 0.5 mm diameter circle at each seed location was averaged to create a seed time trace. These seed traces were correlated against every other brain pixel to create functional connectivity maps. Because seed-based methods are dependent on the seed location, we also used seed-independent methods for determining connectivity patterns. The time traces in every pixel were correlated against every other pixel to create an
With the goal of regenerating the atlas divisions in a data-driven manner, we parcellated the brain into functional regions using the resting-state brain signals and an iterative strategy. Similar methodology has been used to parcellate the human cingulate cortex on the basis of anatomical connectivity assessed by diffusion tensor tractography
Results obtained by iterative techniques potentially depend on the details of initialization. Accordingly, we explored three initial parcellation conditions. The first initial condition was derived from the first ten singular modes of the connectivity matrix (after SVD) with pixels assigned to whichever singular vector with which they had the highest positive or negative coefficient. The pixels that had high positive or negative coefficients with a given parcel were then split into two different parcels, yielding twenty total parcels. While a relatively large amount of high-frequency noise is present, this initial parcellation shows the expected structure of cingulate, retrosplenial, motor, somatosensory, visual and superior colliculus (
Once we had stable parcellations, we investigated the network membership of the obtained regions (the numerical labels assigned to each region being completely arbitrary) using a clustering algorithm. First, we correlated every parcel against every other parcel to create a parcel-to-parcel correlation matrix. Clustering was then performed using a linkage function (Matlab™) with the distance between any two regions defined as 1-
(EPS)
(EPS)
(EPS)
(EPS)
(EPS)
(EPS)
(EPS)
(EPS)
(EPS)
(EPS)
We thank Ernesto Gonzales and Ronaldo Perez for help with animal surgery, Ralph Nothdurft for help with instrumentation, and Timothy Holy and Eric Herzog for thoughtful reading of the manuscript.