A spatial template independent component analysis model for subject-level brain network estimation and inference
share
ICA is commonly applied to fMRI data to extract independent components (ICs) representing functional brain networks. While ICA produces highly reliable results at the group level, single-subject ICA often produces inaccurate results due to high noise levels. Previously, we proposed template ICA (tICA), a hierarchical ICA model using empirical population priors derived from large fMRI datasets. This approach is fast and results in more reliable subject-level IC estimates than dual regression, a standard alternative. However, this and other hierarchical ICA models assume unrealistically that subject effects, representing deviations from the population mean, are spatially independent. Here, we propose spatial template ICA (stICA), which incorporates spatial process priors into the tICA framework. This results in greater estimation efficiency of subject ICs and subject effects. Additionally, the joint posterior distribution can be used to identify areas of significant activation using an excursions set approach. By leveraging spatial dependencies and avoiding massive multiple comparisons, stICA can achieve more power to detect true effects. Yet stICA introduces computational challenges, as the likelihood does not factorize over locations. We derive an efficient expectation-maximization (EM) algorithm to obtain maximum likelihood estimates of the model parameters and posterior moments of the ICs and subject effects. The stICA framework is applied to simulated data and fMRI data from the Human Connectome Project. We find that stICA produces ICs and subject effects that are more accurate and reliable than tICA or dual regression and identifies larger and more reliable areas of activation. The algorithm is quite tractable for realistic datasets: with approximately 6000 cortical locations per hemisphere and 1200 time points, convergence was achieved within 7 hours.
READ FULL TEXT
