Joint detection-estimation of brain activity in fMRI using an autoregressive noise model

Different approaches have been considered so far to cope with the temporal correlation of fMRI data for brain activity detection. However, it has been reported that modeling this serial correlation has little influence on the estimate of the hemodynamic response function (HRF). In this paper, we examine this issue when performing a joint detection-estimation of brain activity in a given homogeneous region of interest (ROI). Following the work of Bullmore et al. (1996), we adopt a space-varying AR(1) temporal noise model and assess its influence, on both the estimation of the HRF and the detection of brain activity, using synthetic and real fMRI data. We show that this model yields a significant gain in detection specificity (lower false positive rate)