Clustering of fMRI data for activation detection using HDR models

The aim of this work is to perform parametric estimation and detection of activation in fMRI data. The proposed procedure is a segmentation algorithm where clustering is based on estimated parameters for a chosen hemodynamic response (HDR) model. These parameters are estimated for each individual voxel by performing a weighted least-squares nonlinear curve fit to its time series. The parameters are used to decide which voxels are candidates for activation. A segmentation algorithm is executed on a subset of the image voxels, selected based upon fitting parameters. Our procedure will yield activation maps constructed from possibly 2-D (i.e., multi-voxel) regions of activation as opposed to identifying voxels based on individual voxel statistical significance, followed by merging into regions. The approach is intended to reduce false detections, producing "cleaner" activation results without resorting to filtering techniques that may sacrifice spatial resolution.