Meeting Banner
Abstract #3633

Prediction & Correction of Physiological Noise in fMRI using Machine Learning

Tom Ash1, John Suckling2, Martin Walter3, Cinly Ooi2, Claus Tempelmann4, Adrian Carpenter1, Guy Williams1

1Wolfson Brain Imaging Centre, University of Cambridge, Cambridge, United Kingdom; 2Brain Mapping Unit, University of Cambridge, Cambridge, United Kingdom; 3Department of Psychiatry, University of Magdeburg, Magdeburg, Germany; 4Department of Neurology, Otto v. Guericke University, Magdeburg, Germany


We present a support vector machine based technique for recreation of partially or fully absent physiological recording data, to allow detrending of physiological noise to occur even in the absence of complete recordings of the physiological cycles. The technique uses a multi-class SVM to predict phase of each physiological cycle from fMRI image data, after training on prior data. Using these predicted phase values as inputs to physiological detrending tool RETROICOR show similar impact on Fourier transforms of the data as using recorded values, showing that they are accurate enough for use as inputs to detrending tools.

Keywords