Jody Culham Brain and Mind Institute Department of Psychology Western University http://www.fmri4newbies.com/ Brain Connectivity Last Update: December 1, 2014 Last Course: Psychology 9223, F2014, Western University.
Download ReportTranscript Jody Culham Brain and Mind Institute Department of Psychology Western University http://www.fmri4newbies.com/ Brain Connectivity Last Update: December 1, 2014 Last Course: Psychology 9223, F2014, Western University.
Jody Culham Brain and Mind Institute Department of Psychology Western University http://www.fmri4newbies.com/ Brain Connectivity Last Update: December 1, 2014 Last Course: Psychology 9223, F2014, Western University Networks and Connectivity • In the analyses we have investigated so far, we have been considering brain areas in isolation • More sophisticated statistical techniques have now become available to investigate networks of activation Part I Structural Connectivity White Matter Diffusion Tensor Imaging (DSI) Diffusion = Brownian Motion • cumulative random motion of molecules Robert Brown (1773-1858) Image from Wikipedia Longer Time Larger Diffusion • 2.5 x 2.5 x 2.5 mm cube contains ~ 1020 water molecules • r2 = 6Dt – r2 = squared displacement – D = diffusion coefficient (e.g., 3 x 10-3 mm2/s for water at 37 C) – t = time Restricted Diffusion • diffusion in a particular direction is affected by cell membranes, myelin, microtubules, density of axons, diameter of fibres, consistency of fibre orientation, etc. Isotropic vs. Anisotropic Diffusion isotropic = equal in all directions anisotropic = different in different directions Measure diffusion in at least 6 directions (more directions is better) Slide from Anna Matejko Ellipsoids • eigenvalue = length of one axis of ellipsoid • ranges from 0 to 1 • 1, 2, 3 • fractional anisotropy (FA) = nonuniformity of eigenvalues • ranges from 0 = sphere to 1 = line • reflects multiple factors not just one (e.g., myelination) Ellipsoids • • • • eigenvalue = length of one axis of ellipsoid ranges from 0 to 1 1, 2, 3 Calculated for each voxel Ellipsoids in a Brain Mean Diffusivity (MD) MD= (λ1 + λ2+ λ3)/3 • MD= measures overall water diffusion in a voxel • Insensitive to the orientation of fibers • Often used clinically • High MD often indicates poorer white matter integrity high MD = white low MD = black Slide modified from Anna Matejko Fractional Anisotropy • an + iso + tropy = not + equal + turn • non-uniformity of eigenvalues measure of how elongated an ellipsoid is varies from 0 (sphere) to 1 (line) indicates high white matter integrity • • • high FA = white low FA = black Directions Left-Right Anterior-Posterior Superior-Inferior isotropic DiffusionWeighted Intensity (dark = high diffusion) anisotropic Apparent Diffusion Coefficient (bright = high diffusion) Jones, 2008, Cortex Color Coding of Orientation • red = left-right • green = anterior-posterior • blue = superior-inferior • Note: maps show orientation NOT direction – e.g., you can’t discriminate left right from right left Jones, 2008, Cortex Variety of DTI Maps Mean Apparent Diffusion Coefficient (bright = high diffusion) Fractional Anisotropy (FA) (bright = anisotropic) Color-Coded Orientation Sources of FA • Many microscopic and macroscopic factors can contribute to anisotropy • myelination • ~20% • axon diameter • axon density Tournier et al (2011) Longitudinal vs. Radial Diffusivity High FA can result from • High Longitudinal Diffusivity along λ1 • Low λ2 and/or λ3 • or both along http://www.diffusionimaging.com/2012/10/voxel-based-versustrack-based.html Low FA High FA Deterministic Tractography • assumes largest eigenvector reflects dominant fibre orientation • can set various tracking parameters – e.g., stop tracking if FA < 0.15 – e.g., stop tracking if angle changes > 50 degrees • doesn’t allow branching fibres Jones, 2008, Cortex Major Tracts • based on deterministic tractography Data from: Catani & Ffytche, 2005 Figure from: Jones, 2008, Cortex Visual Tracts Catani et al., 2003, Brain Limitations of Deterministic Tractography deterministic tractography finds medial but not lateral fibres from corpus callosum (red) and cerebro-spinal tracts (green) confidence in deterministic tractography? 0<p<1 Jones, 2008, Cortex Cones of Uncertainty Jones, 2008, Cortex Probabilistic Tractography • propagate a large number of pathways from the seed point • pathways sample from the distribution of directions • output: proportion of pathways from seed point reach a given voxel • high probability does not guarantee that the tract exists • false positives and false negatives are still a big problem • accumulated error problem: the longer the tract, the more small errors add up Probabilistic Tractography LGN seed optic radiations Data from: Geoff Parker Figure from: Jones, 2008, Cortex Probabilistic Tractography finds missing fibres left motor strip seed 3% 7% Jones, 2008, Cortex 20% Ambiguity of Overlapping Fibres Crossing Fibres Kissing Fibres Multifibre Models Jones, 2008, Cortex One- vs. Multi-Fibre Models • acoustic radiations (MGN-primary auditory cortex) Using DTI to Define Areas • Strictly speaking, “Areas” in the formal anatomical sense are defined by Function, Architectonics, Connectivity and Topography, yet imagers typically (and erroneously) only consider Function functional boundaries Connectional fingerprints of dorsal premotor (PMd) and ventral premotor (PMV) cortex define areas with excellent correspondence to functionally determined boundaries Data from: Tommasini et al., 2007, J Neurosci Figure from: Johansen-Berg & Behrens, 2009, Ann Rev Neurosci Stats vs. Tracts • while pictures of tracts can be very pretty, we’ve seen many problems gauging their validity • don’t underestimate the utility of basic stats on mean ADC, FA, etc. FA histograms in patients with traumatic brain injury FA histograms in controls Benson et al., 2007, J Neurotrauma correlation between mean FA and posttraumatic amnesia FA and Age Can you change white matter? Scholz et al. (2009) Combining DTI and fMRI Van Eimeren et al (2010) Atlas-based Parcellation Navas-Sanchez et al. (2014) Diffusion Spectrum Imaging (DSI) DSI vs. DTI • Diffusion Tensor Imaging – find main direction and FA within each voxel – cannot image crossing fibers • Diffusion Spectrum Imaging – find distribution of fiber orientations within each voxel – can image crossing fibers – other techniques (HARDI, Q-BALL) are similar in spirit Fiber Distributions Within A Voxel Seunarine & Alexander, 2009, In Johanssen-Berg & Behrens (Eds.), Diffusion MRI FA vs. Distributions fODF = fiber orientation distribution function Seunarine & Alexander, 2009, In Johanssen-Berg & Behrens (Eds.), Diffusion MRI Example pons, where cerebellar peduncle crossses corticospinal tract Hagmann et al., 2006, RadioGraphics DSI vs. DTI of the optic chiasm DSI Wedeen et al., 2008, NeruoImage DTI DSI vs. DTI of Callosal Fibres So why isn’t everyone using DSI vs. DTI? • Despite the clear advantages of DSI, most diffusion-based tractography still relies on DTI • DSI scans are very long (min ~40 min) • Rapid improvements are being made in scanning technology and postprocessing that should make DSI easier to do Part II Functional and Effective Connectivity Resting State Scan • a scan in which the subject relaxes without falling asleep and is told not to think about anything in particular while activation is measured throughout the brain Critique of fMRI Critique of Resting State? OUR RESTING STATE CONNECTIVITY STUDY IDENTIFIED THE BRAIN NETWORKS IMPLICATED IN DROWSINESS, BACK PAIN, AND HAVING TO PEE REALLY BADLY http://xkcd.com/1453/ Functional Connectivity • Areas show correlations in activation • Those areas may or may not be directly interconnected Step 1: Extract time course from area of interest = “seed”. Filter out high frequencies, leaving low frequencies < ~0.1 Hz (~1 cycle/10 s). MT+ motion complex resting state scan (10 mins) Step 2: Look for other areas that are show correlated activity in the same scan V6 (another motion selective area correlation with MT+: r > .8 Default Mode Network Fox & Raichle, 2007, Nat Rev Neurosci Fox and Raichle, 2007, Nat. Rev. Neurosci. • During resting state scans, there are two networks in which areas are correlated with each other and anticorrelated with areas in the other network Default Mode in Anesthetized Monkeys saccade task Monkey default mode network posterior cingulate seed Human default mode network Data from: Vincent et al., 2007, Nature Figure from: Fox & Raichle, 2007, Nat Rev Neurosci LIP tracer • suggests that the default mode network does not just reflect uncontrolled cognition ICA and Resting State Connectivity • ICA can be used to examine resting state connectivity ICA Identifies RS Subnetworks Data from: Beckman et al., 2005, J Neurosci Figure from: Huettel et al., 2nd ed. Correlation ≠Causation Figure from: Huettel et al., 2nd ed. Partial Least Squares (PLS) • data-driven approach developed by Randy McIntosh & co. • identifies components (latent variables) whose amplitude is affected by the experimental manipulation (unlike ICA) • output = set of weights applied to experimental conditions and set of voxels where activation was influenced by those weights • components can be evaluated statistically through permutation tests – resample original data to determine probability of a given effect size Psychophysiological Interactions (PPI) • identify the effect of an experimental manipulation on the functional connectivity between two regions • Subjects watched a moving pattern passively or paid attention to its speed • With attention, there was a steeper slope in the relationship between the primary visual cortex and motion-selective area MT+/V5 Friston et al., 1997, NeuroImage A Clear Explanation of PPI 57 Key Idea of PPI • If two areas are interacting, their activity will go up and down in synch • This effect may be task dependent • It should be more than can be explained by the shared main effect of task 58 Based on O’Reilly et al., 2012, SCAN PPI Example • Task: Participants actively navigate through VR maze • Control: Participants passively travel through VR maze • Standard fMRI analysis – Task – Control • Activation in prefrontal cortex (PFC) and hippocampus (HC) • Hypotheses: H1: PFC and HC are independently activated during active navigation H2: PFC and HC work together interactively during active navigation • Prediction – If PFC and HC interact, their activity should be more correlated during active navigation than passive control 59 Based on O’Reilly et al., 2012, SCAN Regressors Task Regressor Task Regressor + ROI Activity Task Regressor x ROI Activity = PPI Regressor O’Reilly et al., 2012, SCAN 60 Logic Baseline: Red and Task: Red and blue blue are antiare positively correlated correlated Task Regressor x ROI Activity = PPI Regressor • Regions that are correlated because of inherent connectivity (as one would see in resting state) don’t show up because correlation during task and anticorrelation during baseline cancel each other out • Regions that interact more during task than baseline show up because correlation outweighs anticorrelation 61 Based on O’Reilly et al., 2012, SCAN Regressors Task Regressor x ROI Activity = PPI Regressor • Even though we are only interested in the PPI regressor, we must include the Task Regressor and ROI Activity as covariates of no interest • This ensures we are only looking at interactions over and above the task activation (which may have been the basis for selecting the region) and the inherent correlation • Because the PPI Regressor is highly correlated with the other two regressors, PPI has relatively low statistical power • As in any analysis, it is beneficial to include other regressors of no interest to soak up known sources of noise (e.g., error trials, head motion) 62 Based on O’Reilly et al., 2012, SCAN Structural Equation Modelling (SEM) • statistical approach for inferring causal relationships amongst variables • derived from econometrics and applied to fMRI Activity in PPC (posterior parietal cortex) Structural Equation Modelling Example PFC high PFC low PFC = prefrontal cortex Activity in V5 (= MT+ = motion-selective area) • Participants viewed moving stimuli • Connectivity between V5 and PPC was modulated by activation in PFC Büchel & Friston, 1997, Cerebral Cortex 66 Dynamic Causal Modelling (DCM) • create model of connections (perhaps based on known structural connections) • examine how experimental manipulations affect connectivity Grol et al., 2007, J Neurosci Granger Causality Modelling (GCM) • identifies how the past history in one voxel affects the activation in other voxels • doesn’t require a priori models of networks • need to demonstrate that it’s not an artifact of different HRF latencies – show that effect occurs in some but not all conditions Does A:red improve prediction of B:blue relative to prediction from other info alone (e.g., B:green and Z:purple) Warning: Many reviewers are highly skeptical of GCM applied to fMRI data. Use at your own risk! Graph Theory: Small World Networks Hubs Six Degrees of Kevin Bacon Sex Degrees of Copulation Matthew Perry HIV/AIDS hub • “Patient Zero”: Gaetan Dugas • Canadian flight attendant • 250 partners/year • 40 of 248 people diagnosed with AIDS in 1982 had had sex with him or someone who had 9-11 Terrorist Links Internet nodes in 1998: 800 million Average degrees of separation: 19 Hubs Graph Theory: Nodes, Edges and Hubs The whole diagram is called a “graph” Global hub Node Edge 2 4 2 1234 2 2 4 1 Degree 2 2 4 Cluster 2 1 Provincial hub Provincial hub Undirected vs. Directed Edges Undirected Directed Thresholding of Edges threshold Length and Clustering highly clustered long paths highly clustered short paths weakly clustered short paths L = path length C = clustering coefficient Motifs very common very common Example: Monkey Anatomical Connections The Brain: It’s a Small World After All Bullmore & Sporns, 2009, Nat Rev Neurosci Brain Hubs Area 46 (DLPFC) = global hub Example 2: Human Anatomical and Functional Connections Bullmore & Sporns, 2009, Nat Rev Neurosci DSI-based Hubs in Humans Hagmann et al., 2008, PLoS Biology Resting State Connectivity-based Hubs in Humans van den Heuvel et al., 2008, NeuroImage Example 3: It’s a Small Worm After All C. elegans ~1 mm 302 neurons Combinations of Connectivity Measures Data from: Andrews-Hanna et al., 2007, Neuron Figure from: Huettel et al., 2nd ed. Want to Learn More? Book Recommendation 2010 MIT Press