A recursive algorithm suitable for functional magnetic resonance imagi
ng (FMRI) calculations is presented. The correlation coefficient of a
time course of images with a reference time series, with the mean and
any linear trend projected out, may be computed with 22 operations per
voxel, per image; the storage overhead is four numbers per voxel. A s
tatistical model for the FMRI signal is presented, and thresholds for
the correlation coefficient are derived from it. Selected images from
the first real-time functional neuroimaging experiment (at 3 Tesla) ar
e presented. Using a 50-MHz workstation equipped with a 14-bit analog-
to-digital converter, each echo planar image was acquired, reconstruct
ed, correlated, thresholded, and displayed in pseudocolor (highlightin
g active regions in the brain) within 500 ms of the RF pulse.