Nathan Churchill1, Stephen Strother2
1Medical Biophysics, University of Toronto, Toronto, Ontario, Canada; 2University of Toronto / Rotman Research Institute, Toronto, Ontario, Canada
We have developed a multivariate, data-driven framework to estimate and control physiological noise in functional MRI. This model (1) identifies and down-weights sources of pulsatile flow (e.g. arteries, sinuses and ventricles), based on a map of high-frequency spectral power, correlated with an atlas of potential artifact regions. It then (2) regresses out physiological noise present in grey matter, using an adaptation of Canonical Correlation Analysis and the spatial weights derived from Step (1). This denoising procedure consistently reduces activation false positives, and increases both prediction accuracy and reproducibility of activation maps in subsequent analyses.