Recurrent Neural Networks for Spatiotemporal Dynamics of Intrinsic Networks from fMRI Data

Functional magnetic resonance imaging (fMRI) of temporally-coherent blood oxygenization leveldependent (BOLD) signal provides an effective means of analyzing functionally coherent patterns in the brain [6, 5, 13]. Intrinsic networks [INs, 3] and functional connectivity are important outcomes of fMRI studies and are central to understanding brain function and making diagnoses [4, 1, 10]. The most popular method for separating INs, independent component analysis [ICA, 2], 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 [11]. 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.