Synthetic Brain Imaging A Computational Interface Between Electrophysiology

Synthetic Brain Imaging: A Computational Interface Between Electrophysiology and Neuroimaging James Bonaiuto (work with Michael Arbib, USC) Andersen Laboratory California Institute of Technology Bonaiuto, jimmy@vis. caltech. edu

The Challenge of Data Integration and Interpretation Two of the major problems facing cognitive neuroscientists are data interpretation and comparison across modalities Neuroimaging studies are typically designed to test some conceptual model of the interactions between the brain regions involved in a task. The results are usually evaluated using an ad-hoc verbal analysis and compared to neurophysiological data. ? ≈ Rizzolatti et al (1995) Bonaiuto, jimmy@vis. caltech. edu Buccino et al (2004)

From Neural Activity to BOLD Response Neural activity Signalling Vascular response Vascular tone (reactivity) Autoregulation Synaptic signalling BOLD signal Blood flow, oxygenation and volume arteriole B 0 field glia Metabolic signalling venule (figure by Richard Wise) BOLD signal ≠ Neuron firing Bonaiuto, jimmy@vis. caltech. edu

Synthetic Brain Imaging was developed to address the disconnect between experimental modalities Computational models based on neurophysiological data are used to generate simulated neuroimaging signals: regional cerebral blood flow (r. CBF) blood oxygen level-dependent (BOLD) responses This technique has been used to successfully predict: A projection from the PFC to the anterior intraparietal region AIP (Arbib, Fagg & Grafton, 2002) Populations of sound contour-selective cells in secondary auditory cortex (Husain et al. , 2004) Bonaiuto, jimmy@vis. caltech. edu

What is the Appropriate Neural Model? It depends on the data the model is intended to address Early synthetic brain imaging approaches used firing rate models (Arbib et al. , 1995; Horwitz et al, . 1998) We require the simplest possible model that can produce a wide range of firing patterns Bonaiuto, jimmy@vis. caltech. edu (Izhikevich, 2004)

Neurophysiology and f. MRI Under some conditions spiking activity and CBF dissociate In general, local field potential (LFP) is a better predictor of BOLD than spiking activity LFP most directly reflects synaptic rather than spiking activity We therefore require a model with realistic synaptic activity Bonaiuto, jimmy@vis. caltech. edu (Lauritzen et al. , 2003) (Goense & Logothetis, 2008)

Neurovascular Coupling We use the sum of synaptic conductances as the measure of neural activity Several mechanisms coexist to regulate blood flow neuron-astrocyte pathway (Koehler et al. , 2006) vasomotor GABAergic interneurons (Cauli et al. , 2004) nitric oxide diffusion (Metea & Newman, 2006) We use a generic blood flow-inducing signal that subsumes neurogenic and diffusive components (Friston et al. , 2000) A linear function of neural activity with signal decay and autoregulatory feedback from blood flow e=gain parameter u 0=baseline synaptic activity tb=decay time constant fin=blood flow tf=feedback time constant Bonaiuto, jimmy@vis. caltech. edu

Vascular Signal Generation: Balloon Model We use Friston & Buxton’s balloon model to simulate the vascular response to neural activity and generate simulated PET or f. MRI signals Bonaiuto, jimmy@vis. caltech. edu

A New Synthetic Brain Imaging Model Izhikevich neurons with realistic synaptic dynamics and noise Total synaptic conductance for all synapses in a voxel used to generate to a generic blood flowinducing signal Use normalized blood flow-inducing signal as input to the Balloon model Bonaiuto, jimmy@vis. caltech. edu

Basic Network Architecture Membrane Potential (m. V) 30 { Pyramidal Neuron Firing Rate (Hz) 0 100 Response latency Time (s) 0 -50 2. 0 -80 1 Neuron 100 – Populations of pyramidal neurons and inhibitory interneurons – Center-surround connectivity implements winner-take-all dynamic – Depending on the input, there can be a considerable latency before the network settles on a stable winner Bonaiuto, jimmy@vis. caltech. edu 0

Example 1: Random Dot Motion Discrimination The random dot motion direction discrimination task is commonly used to study perceptual decision-making Task: saccade in the net movement direction of a field of randomly moving dots Stimulus coherence: percentage of dots moving in the same direction This task is useful in demonstrating the power of synthetic brain imaging because A well-defined network of brain regions is involved (MT, LIP, FEF) There exists neural recording, microstimulation, behavioral and imaging data using the task in humans and non-human primates Bonaiuto, jimmy@vis. caltech. edu

Model Derivation We used three connected WTA networks to simulate the MTLIP-FEF network Neural parameters were set using values from experimental data Network parameters were set using a genetic algorithm that used the model’s fit to neural recording and behavioral data as the fitness function Bonaiuto, jimmy@vis. caltech. edu (Gold & Shadlen, 2007)

Results: Neural Activity 3. 2% Stimulus Coherence 12. 8% 51. 2% Pyramidal neurons in LIP converge on a population code centered on the chosen saccade direction Response time was interpreted as the time taken for max firing rate in FEF to reach 100 Hz (Hanes & Schall, 1996) Bonaiuto, jimmy@vis. caltech. edu

Results: Behavioral Measures Model Performance Human Behavioral Data (Palmer et al. , 2005) Response time and accuracy were fit to the same psychometric and chronometric functions used to analyze human data Fitted parameters were within the range of human performance Bonaiuto, jimmy@vis. caltech. edu

Results: Microstimulation Simulations Simulation Results Monkey Data Control MT Stim LIP Stim (Hanks et al. , 2006) MT microstimulation biased decision process and reaction time LIP microstimulation had the same effect, but to a lesser extent The same results are found in monkey microstimulation experiments Bonaiuto, jimmy@vis. caltech. edu

Results: Synthetic f. MRI Human f. MRI MT LIP FEF (Rees, Friston & Koch 2000) The model replicated human f. MRI data that only found a positive correlation between BOLD response and stimulus coherence in MT In the model this is because intraregional processing (WTA) dominates LIP and FEF activity and is roughly the same at each coherence level Bonaiuto, jimmy@vis. caltech. edu

Example 1: Summary A basic neural microcircuit was connected in a network based on anatomical considerations Neural parameters were set using values from experimental data. Network parameters were set using a genetic algorithm that fit the firing rate and model behavior to neural recording and psychophysical data The model was validated by replicating microstimulation and f. MRI studies Bonaiuto, jimmy@vis. caltech. edu

Example 2: Reach Target Selection We developed a model of the parieto-frontal reach circuit with each region based on macaque neurophysiological data and interregional connections constrained by tract-tracing studies The output of the dorsal premotor region was decoded and used to control a simulated arm/hand Bonaiuto, jimmy@vis. caltech. edu

Synthetic PET: Comparison with Experimental Data Signal % Change 0. 8 0. 4 0. 0 F 2 F 6 LIP V 6 A PFC % V 4 F 2 F 6 LIP V 6 A PFC V 4 change in r. CBF in F 2, F 6, V 6 a, and PFC matches published neuroimaging data (Savaki et al. , 1997, left) Differences in activity in V 4 highlight a PFC→V 4 connection overlooked in model construction Updated model activity closely matches published data (right) Bonaiuto, jimmy@vis. caltech. edu

Example 2: Summary A virtuous cycle of models and experiments We used available connectivity and neural recording data to develop a model of reach target selection Synthetic brain imaging was used to compare the global model activity to metabolic signals in the monkey brain This comparison was used to update the model to include feedback connections from PFC to V 4 Bonaiuto, jimmy@vis. caltech. edu

Example 3: Imaging of a Cognitive Model of Apraxia Peigneux et al (2004) Looked at familiar vs novel gesture imitation Found that brain areas associated with gesture recognition and production were not more active for familiar vs novel imitation Red: visuo-gestural coding Green: input praxicon Blue: output praxicon Bonaiuto, jimmy@vis. caltech. edu

Synthetic PET on a Cognitive Model of Apraxia Bonaiuto, jimmy@vis. caltech. edu

Example 3: Summary A model of gesture recognition and imitation was used to generate PET predictions that conflict with those generated in an ad-hoc manner using a conceptual model The model is simple, but nonetheless more complex than the conceptual one used by experimentalists Data necessary to constrain the model do not exist – the model demonstrates one possibility In this situation multiple competing models should be developed and used to determine an experiment that could disambiguate them Bonaiuto, jimmy@vis. caltech. edu

Summary Synthetic imaging can bridge the gap between electrophysiology and neuroimaging We gave three examples using this technique to Validate a neural model Refine a neural model Offer a novel interpretation of experimental data Future studies could use this technique to generate predictors for f. MRI analysis Bonaiuto, jimmy@vis. caltech. edu

Thank you Arbib Lab Rob Schuler Itti Lab David Berg Farhan Baluch Funding NSF Sloan Foundation Bonaiuto, jimmy@vis. caltech. edu