Robust constrained nonGaussian fMRI detection

For fMRI detection, it is desirable to have sensitive detectors for enhanced performance in low SNR environment. This sensitivity , usually captured through learning of associated models, comes at the price of increased false alarms. In this paper, we address the issue of robustness to false alarm while maintaining sensitivity by providing the analytical framework for incorporating prior information in the form of constraints in Gaussian and non-Gaussian settings. We show that the impact on the decision statistic of incorporating constraints is simply captured through a simple modification of the unconstrained detector´s statistic. The computational burden of the constrained and unconstrained detectors are thus similar. The performance of the new constrained detector is shown on fMRI data to provide superior performance when compared the conventional CFAR detector