[Paper Review] Recurrent Neural Networks for Spatiotemporal Dynamics of Intrinsic Networks from fMRI Data

Devon Hjelm, Sergey Plis+1 more|arXiv (Cornell University)|Nov 3, 2016

Functional Brain Connectivity Studies5 references5 citations

TL;DR

This paper proposes a novel RNN-ICA model that integrates recurrent neural networks with independent component analysis to capture both temporal dynamics and subject-specific spatial variability in intrinsic brain networks from fMRI data. By training via gradient descent, the model overcomes ICA’s limitations in handling time-ordered data and enables joint learning of dynamic spatial maps and temporal patterns.

ABSTRACT

Functional magnetic resonance imaging (fMRI) of temporally-coherent blood oxygenization level-dependent (BOLD) signal provides an effective means of analyzing functionally coherent patterns in the brain. Intrinsic networks and functional connectivity are important outcomes of fMRI studies and are central to understanding brain function and making diagnoses. The most popular method for separating INs, independent component analysis, begins with the assumption that the data is a mixture of maximally independent sources. ICA is trainable through one of many relatively simple optimization routines that maximize non-Gaussianity or minimize mutual information. Although fMRI data is a time series, ICA, as with other popular linear methods for separating INs, is order-agnostic in time: each multivariate signal at each time step is treated as i.i.d.. ICA in its common use in the field employs the same parameterization across subjects, which allows for either temporal or spatial variability, but not both. In order to overcome shortcomings of temporal ICA in lack of dynamics and subject-wise/temporal variability of spatial maps, but without abandoning the fundamental strengths of ICA, we combine recurrent neural networks (RNNs) with an ICA objective. The resulting model naturally represents temporal and spatial dynamics---having subject-wise and temporally variable spatial maps---and is easily trainable using gradient descent and back-propagation.

Motivation & Objective

• Address the limitation of traditional ICA in fMRI analysis, which treats time series as i.i.d. and ignores temporal dynamics.
• Overcome the inability of standard ICA to model subject-specific and time-varying spatial maps simultaneously.
• Preserve the strengths of ICA—such as statistical independence and interpretability—while enhancing its capacity for dynamic modeling.
• Develop a trainable, end-to-end framework that jointly optimizes spatial maps and temporal dynamics using gradient descent.

Proposed method

• Combine recurrent neural networks (RNNs) with an ICA objective function to model time-ordered fMRI data while preserving statistical independence of components.
• Use RNNs to learn time-varying spatial maps by encoding temporal dependencies in the BOLD signal across time points.
• Train the model using back-propagation through time and stochastic gradient descent to optimize the ICA objective (e.g., maximizing non-Gaussianity or minimizing mutual information).
• Allow subject-specific parameterization of spatial maps while maintaining shared temporal dynamics across subjects.
• Formulate the objective as a differentiable function that combines ICA principles with RNN-based temporal modeling.
• Ensure the model is end-to-end trainable and avoids the need for separate, fixed preprocessing steps common in traditional ICA.

Experimental results

Research questions

• RQ1 Can an RNN-based model effectively capture dynamic spatial maps in intrinsic brain networks from fMRI data while preserving ICA’s statistical independence principles?
• RQ2 How does the proposed RNN-ICA model compare to standard ICA in modeling temporal dynamics and subject-specific variability?
• RQ3 To what extent can gradient-based training of an RNN-ICA model improve the representation of spatiotemporal brain network dynamics compared to static ICA?
• RQ4 Does the integration of RNNs with ICA enable joint learning of time-varying spatial maps and temporal patterns without sacrificing interpretability?

Key findings

• The RNN-ICA model successfully captures time-varying spatial maps that reflect subject-specific and dynamic functional organization in intrinsic networks.
• The model outperforms standard ICA in representing temporal dynamics due to its inherent sequential modeling capability.
• Gradient descent training enables end-to-end optimization of both spatial and temporal components, avoiding the limitations of fixed, order-agnostic ICA.
• The model maintains the interpretability of ICA while extending its applicability to dynamic, subject-specific brain network analysis.