Identifying Rodent Resting-State Brain Networks with Independent Component Analysis

Abstract

Rodent models have opened the door to a better understanding of the neurobiology of brain disorders and increased our ability to evaluate novel treatments. Resting-state functional magnetic resonance imaging (rs-fMRI) allows for in vivo exploration of large-scale brain networks with high spatial resolution. Its application in rodents affords researchers a powerful translational tool to directly assess/explore the effects of various pharmacological, lesion, and/or disease states on known neural circuits within highly controlled settings. Integration of animal and human research at the molecular-, systems-, and behavioral-levels using diverse neuroimaging techniques empowers more robust interrogations of abnormal/ pathological processes, critical for evolving our understanding of neuroscience. We present a comprehensive protocol to evaluate resting-state brain networks using Independent Component Analysis (ICA) in rodent model. Specifically, we begin with a brief review of the physiological basis for rs-fMRI technique and overview of rs-fMRI studies in rodents to date, following which we provide a robust step-by-step approach for rs-fMRI investigation including data collection, computational preprocessing, and brain network analysis. Pipelines are interwoven with underlying theory behind each step and summarized methodological considerations, such as alternative methods available and current consensus in the literature for optimal results. The presented protocol is designed in such a way that investigators without previous knowledge in the field can implement the analysis and obtain viable results that reliably detect significant differences in functional connectivity between experimental groups. Our goal is to empower researchers to implement rs-fMRI in their respective fields by incorporating technical considerations to date into a workable methodological framework.

Keywords: BOLD; ICA; MRI; fMRI; protocol; resting-state networks; review; rs-fMRI.

Figure 1

Rodent resting-state network analysis outline. Schematic outlines 14 steps for resting-state network analysis. Note that preprocessing in Step 6 should be performed at the individual level, in contrast to using group-level brain network identification in Step 9. Steps 10 and 11 can be run in any order since they are independent of each other. Note that identification of networks of interest to template networks (Step 10) assumes availability of appropriate species-specific templates. If no species-specific network templates are available for spatial correlation, one should evaluate all components qualitatively and then proceed to Step 11 (Dual Regression).

Figure 2

Awake rat motion assessment. The motion parameters for 2 typical rats (Left column) in the study and the 2 rejected rats (Right column) are displayed. Green Line, translation/rotation X-axis, Blue Y-axis, and Red Z-axis. Figure reprinted with permission from Becerra et al. (2011b) study in adult rats.

Figure 3

Assessment of motion in lightly anesthetized infant rats during imaging. (A,A′) display the rotation (in degrees) and translation (in mm) for an immobile 2-week-old rat during MRI, respectively. Rotation is a rigid body movement and refers to the movement of the head around a center point. Translation is every point on the head moving a constant distance in a specific direction. The immobile rat's head did not rotate more than 0.005 degrees or moved more than 0.02 mm. This is an acceptable amount of movement for the group ICA. (B,B′) illustrate rotation and translation of an infant rat that moved during the scanning, which lead to a motion-related imaging artifact. As a result, data obtained from this animal was excluded from the group ICA. Blue X line, horizontal axis; Green Y line, vertical axis; Red Z line, longitudinal axis of the scanner. Figure reprinted with permission from past study (Bajic et al., 2016) in infant rats.

Figure 4

Examples of registration. (A) Illustrates representative individual animal functional-to-standard registration of the rat. The gray image is individual resting-state fMRI data while the red contour represents the outline of an adult anatomical atlas as reported by FSL output. First 4 columns are in the axial view; the middle 4 are in the sagittal view; last 4 columns are in coronal view. A common expected artifact is seen in the ventral regions of the rat brain [near ear canals (Schwarz et al., 2006)] and is noted in the second coronal section (arrowheads). Distortions noted in the ventral parts of the brainstem were noted in the caudal region of the brainstem (stars). (B) Shows an extreme example of the erroneous registration when the individual resting-state fMRI data image is rotated 180 degrees to the anatomical template. There is an obvious mismatch of registration that clearly implicates flipped data registration as seen in the temporal regions (first axial section; arrow). Obviously, such case of erroneous registration should not be included in subsequent analysis. Numbers below coronal slices represent distance from Bregma (mm). Left hemisphere of the brain corresponds to the right side of the image. Section with Bregma of 0 mm corresponds to Panel 17 of Rat Brain Atlas (Paxinos and Watson, 1998).

Figure 5

Group ICA spatial maps. Figure shows representative group-level component spatial maps extracted from group ICA (Step 9: Network Detection via Group ICA) as they appear in the Melodic report, including pre-determined statistical thresholds (warm colors reflect positive z-scores, while cold colors reflect negative). Putative neural networks (A–C) show coherent BOLD signal predominantly arising from gray matter, while non-neuronal artefactual components (D,E) show a large degree of edges. Left side of image corresponds to right side of brain.

Figure 6

Brain network identification via spatial correlation. Figure illustrates representative group-level component spatial maps extracted from group ICA for (A) Sensorimotor, (B) Salience, (C) Autonomic Networks (red-yellow) overlaid on spatial maps of correlated template networks (green). Spatial correlation R > 0.20 between individual components and template network(s) is sufficient to identify potential functional networks of interest amongst the full set of extracted group-level components. Components that do not meet criteria for network classification based on set of template networks (e.g., R-values <0.20 for all templates) may still contain biologically relevant brain activity (as in the example of Autonomic Network). Numbers above each coronal section represent distance from Bregma (in mm). Left side of image corresponds to right side of brain.

Figure 7

Resting-state networks in awake rats. Complete maps for Components (C1–C7). All components have been thresholded according to a mixture model approach. See Methods from Becerra et al. (2011b) for details. The atlas is based on the Paxinos and Watson Atlas (1998). Abbreviations: Ins, Insula; AcB, Nucleus Accumbens; Motor, Motor Cortex; Amyg, Amygdala; Parab, Parabrachial; CPu, Caudate-Putamen; PAG, Periaqueductal Gray; Cereb, Cerebellum; ParA, Parietal Association Cortex; Cnf, Cuneiform Nucleus; Som, Somatosensory Cortex; Ent, Entorhinal Cortex; SupColl, Superior Colliculus; FC, Frontal Cortex; Thal, Thalamus; TpA, Temporal Association Cortex; Hypo, Hypothalamus; cing, Cingulate Cortex (anterior and retrosplenial); InfColl, Inferior Colliculus. Figure reprinted with permission from Becerra et al. (2011b).

Figure 8

Full spatial maps of resting-state networks in awake rats. Components (C1–C7) are ordered according to their reproducibility degree. Component 1 has significant cerebellar structures; Component 2 includes medial and lateral cortical structures resembling the human default mode network; Component 3 includes a basal-ganglia-hypothalamus network; Component 4 encompasses basal-ganglia-thalamus-hippocampus circuitry; Component 5 represents an autonomic pathway; Component 6 represents the sensory network; and Component 7 groups interoceptive structures to form a network. All components have been thresholded according to a mixture model approach. See Methods section for details. The atlas is based on the Paxinos and Watson Atlas (1998). Abbreviations: Ins, Insula; AcB, Nucleus Accumbens; Motor, Motor Cortex; Amyg, Amygdala; Parab, Parabrachial; CPu, Caudate-Putamen; PAG, Periaqueductal Gray; Cereb, Cerebellum; ParA, Parietal Association Cortex; Cnf, Cuneiform Nucleus; Som, Somatosensory Cortex; Ent, Entorhinal Cortex; Sup Coll, Superior Colliculus; FC, Frontal Cortex; Thal, Thalamus; TpA, Temporal Association Cortex; Hypo, Hypothalamus; cing, Cingulate Cortex (anterior and retrosplenial); Inf Coll, Inferior Colliculus. Figure reprinted with permission from Becerra et al. (2011b).

Figure 9

Statistical thresholds identified via Gaussian mixture modeling. Figure illustrates variable appearance of Gaussian mixture modeling (GMM) results using four t-statistic maps from different components (A–D). Histogram of z-scores is shown for each t-statistic map (data line), modeled as the full mixture of Gaussians density (gmm fit line), as well as by distinct Gaussian sub-distributions according to class (i.e., “null,” “activation,” “deactivation”). Probability density (y-axis) is used to determine z-score threshold(s) of statistical significance (black squares). (A) Example of more straightforward results, in which three fit curves are taken to be “deactivation” (fit 1, left-shifted with negative z-scores; dark blue line), “null” (fit 2, centrally localized near zero; green line), and “activation” (fit 3, right-shifted with positive z-scores; turquoise color line). Gray region indicates z-values with statistically insignificant BOLD activity, while yellow regions highlight z-scores beyond identified thresholds exhibiting statistically significant BOLD activity. Intercepts between fit 1 (or fit 3) with “null” (fit 2) curve, is marked with small black square. (B) Example of t-statistic histogram with only one significant positive threshold (black square intercepts), despite presence of both positive (fit 3) and negative (fit 1) distributions. Probability density of the latter never surpasses the null (fit 2), confirmed with zoomed in view (no intercept). (C) Example of “split null” distribution, with the null class modeled by fit 2 and 3. Note, only the z-score furthest from zero is considered the threshold: negative threshold (z-score = −2.645) is described when fit 1 > fit 2 (not when fit 1 > fit 3), while positive threshold (z-score = 2.768) is described when fit 4 > fit 3 (not when fit 4 > fit 2). (D) Example for more challenging interpretation of results when fit 2 (green line) can be described as right-shifted (suggestive of “activation” class) with large volume (suggestive of “null” class).

Figure 10

Example of rodent resting-state network and cluster analysis. (A) Example of a group ICA spatial map. Specifically, this row of images illustrates the Default Mode Network at the group level. (B) Example of group-level differences between adult rat ICA spatial maps. It is an example of group level differences between rats previously treated with morphine or saline in early neonatal period. Melodic report spatial maps are presented as z-scores superimposed on mean functional image in radiological convention (right side of image corresponds to left side of brain). The numbers at the bottom (below each coronal section) refer to distance from Bregma (in mm).