Computing with Trajectories: Characterizing Dynamics and Connectivity in Spatiotemporal Neuroimaging Data
Publication or External Link
Human functional Magnetic Resonance Imaging (fMRI) data are acquired while participants engage in diverse perceptual, motor, cognitive, and emotional tasks. Although data are acquired temporally, they are most often treated in a quasi-static manner. Yet, a fuller understanding of the mechanisms that support mental functions necessitates the characterization of dynamic properties. Firstly, we describe an approach employing a class of recurrent neural networks called reservoir computing, and show their feasibility and potential for the analysis of temporal properties of brain data. We show that reservoirs can be used effectively both for condition classification and for characterizing lower-dimensional "trajectories" of temporal data. Classification accuracy was approximately 90% for short clips of "social interactions" and around 70% for clips extracted from movie segments. Data representations with 12 or fewer dimensions (from an original space with over 300) attained classification accuracy within 5% of the full data. We hypothesize that such low-dimensional trajectories may provide "signatures" that can be associated with tasks and/or mental states. The approach was applied across participants (that is, training in one set of participants, and testing in a separate group), showing that representations generalized well to unseen participants.
In the second part, we use fully-trained recurrent neural networks to capture and characterize spatiotemporal properties of brain events. We propose an architecture based on long short-term memory (LSTM) networks to uncover distributed spatiotemporal signatures during dynamic experimental conditions. We demonstrate the potential of the approach using naturalistic movie-watching fMRI data. We show that movie clips result in complex but distinct spatiotemporal patterns in brain data that can be classified using LSTMs (≈90% for 15-way classification), demonstrating that learned representations generalized to unseen participants. LSTMs were also superior to existing methods in predicting behavior and personality traits of individuals. We propose a dimensionality reduction approach that uncovers low-dimensional trajectories and captures essential informational properties of brain dynamics. Finally, we employed saliency maps to characterize spatiotemporally-varying brain-region importance. The spatiotemporal saliency maps revealed dynamic but consistent changes in fMRI activation data. Taken together, we believe the above approaches provide a powerful framework for visualizing, analyzing, and discovering dynamic spatially distributed brain representations during naturalistic conditions.
Finally, we address the problem of comparing functional connectivity matrices obtained from temporal fMRI data. Understanding the correlation structure associated with multiple brain measurements informs about potential "functional groupings" and network organization. The correlation structure can be conveniently captured in a matrix format that summarizes the relationships among a set of brain measurements involving two regions, for example. Such functional connectivity matrix is an important component of many types of investigation focusing on network-level properties of the brain, including clustering brain states, characterizing dynamic functional states, performing participant identification (so-called "fingerprinting"), understanding how tasks reconfigure brain networks, and inter-subject correlation analysis. In these investigations, some notion of proximity or similarity of functional connectivity matrices is employed, such as their Euclidean distance or Pearson correlation (by correlating the matrix entries). We propose the use of a geodesic distance metric that reflects the underlying non-Euclidean geometry of functional correlation matrices. The approach is evaluated in the context of participant identification (fingerprinting): given a participant's functional connectivity matrix based on resting-state or task data, how effectively can the participant be identified? Using geodesic distance, identification accuracy was over 95% on resting-state data and exceeded the Pearson correlation approach by 20%. For whole-cortex regions, accuracy improved on a range of tasks by between 2% and as much as 20%. We also investigated identification using pairs of subnetworks (say, dorsal attention plus default mode), and particular combinations improved accuracy over whole-cortex participant identification by over 10%. The geodesic distance also outperformed Pearson correlation when the former employed a fourth of the data as the latter. Finally, we suggest that low-dimensional distance visualizations based on the geodesic approach help uncover the geometry of task functional connectivity in relation to that during resting-state. We propose that the use of the geodesic distance is an effective way to compare the correlation structure of the brain across a broad range of studies.