Leigh A. Johnston1,2, Maria Gavrilescu2, Eugene P. Duff3, Gary F. Egan2,4
1Electrical and Electronic Engineering and NICTA Victorian Research Laboratory, University of Melbourne, Parkville, VIC, Australia; 2Howard Florey Institute, Melbourne, VIC, Australia; 3FMRIB, Oxford University, UK; 4Centre for Neuroscience, University of Melbourne, Australia
A stochastic linear model (SLM) of the BOLD signal is presented in which neurovascular dynamics are modelled by an autoregressive exogenous input signal, embedded in parametrically modelled noise. The unknown SLM states and parameters are estimated by an iterative coordinate descent algorithm based on the Kalman smoother, from which novel activation weights are calculated. We demonstrate, through application to a motor task fMRI dataset, that the SLM produces more robust and consistent activation estimates than the general linear model. The stochasticity of the SLM embodies sufficient flexibility to account for observed variations in the BOLD signal.