We propose the Ising model as a spatial prior on assessments of neural activity, to penalize disagreement in these assessments for nearby voxels. We show how to this prior can be integrated in principled manner with both classical and information theoretic measurements of neural activity. In this approach, reguarlization can proceed without spatial smoothing, which further compromises the already low spatial resolution of fMRI. Furthermore, we show that the maximum a posteriori (MAP) activation map under this spatial prior can be computed in polynomial time, by reduction to the min cut/ max flow problem.
The above figure shows the effect of varying the strength of the spatial prior on assessments of protocol-related neural activity measured using a General Linear Model (GLM). Voxels declared active in a word-association task are colored white and displayed overlaying two axial slices (at the level of the Sylvian fissure). The strength of the spatial prior increases from left to right, and the voxel-independent activitation map (with test size 1e-7) is shown in the leftmost column. Note that as the strength of the prior increases, voxels which might be declared inactive in a voxel-independent test, may be declared active due to their proximity to other strongly active voxels. Of course, since most voxels are inactive for typical test sizes, the primary effect of the Ising prior is to control the number of false detections by removing spatially-isolated activations. The fMRI data were not pre-processed or pre-smoothed, so that the effect of the spatial prior can be observed in isolation.
Eric Cosman, John Fisher, and William Wells. "Exact MAP Activity Detection In fMRI Using A GLM with an Ising Spatial Prior." MICCAI 2004. In Press.
Junmo Kim, John W. Fisher III, Andy Tsai, Cindy Wible, Alan S. Willsky, William M. Wells III: Incorporating Spatial Priors into an Information Theoretic Approach for fMRI Data Analysis. MICCAI 2000: 62-71
Cindy Wible cindy@bwh.harvard.edu
Eric Cosman ercosman@ai.mit.edu
Back to the MIT AI Lab page.