Conclusions

We have proposed a technique for firstly choosing a basis set, and then the means to constrain the subspace spanned by the basis set to only include sensible HRF shapes within a GLM framework. We can carry out inference using Variational Bayes which also performs adaptive spatial regularisation of temporal autocorrelation. Constraining the subspace spanned by the basis set allows for far superior separation of activating voxels from non-activating voxels in FMRI data. We used spatial mixture modelling to produce final probabilities of activation, and demonstrated on FMRI data the increased sensitivity produced.