Reconstructing Temporal Dynamics of fMRI Time Series via Encoded Contextual Information

To uncover the rich temporal dynamics from the noisy fMRI time series, we propose a data-driven time series model: temporal reconstruction model, targets on reconstructing subsequences of fMRI time series via encoded contextual representations. This novel reconstruction model is further perfected by learning the commonality among multiple reconstructed temporal resolutions. Through an empirical validation on a synthetic noisy time series, we demonstrate the superior denoising capacity of our reconstruction model. Implementing this reconstruction model on a real fMRI dataset, it reveals the rich temporal dynamics of reconstructed fMRI time series are revealed, assisting the subsequent neuroscientific analysis on discovering more consistent subject-level temporal independent functional modes.