An evaluation of methods for detecting brain activations from PET or fMRI images

Brain activation studies based on PET or fMRI seek to explore neuroscience questions by using statistical techniques to analyze the acquired images, Currently, the predominant viewpoint toward quantifying the detection performance of these statistical methods is to model their output using random field theory, then to ascribe statistical significance (false-positive probability) based on the model. In this paper, we pursue instead an empirical strategy, based on receiver operating characteristics (ROC) analysis, as a first step toward a more-complete evaluation of the performance of brain-activation detection methods, including the power (true-positive probability) of various tests, Using a phantom model derived from parameters measured from PET neuroimaging studies, we compare three methods for detecting brain activation. We consider one method based on pixel-by-pixel image comparisons (the t-test) and two methods based on pixel covariances (correlation thresholding and singular value decomposition thresholding). The simple geometry of our phantom model allows us to construct an optimal detector, the generalized likelihood ratio test (GLRT), for comparison with the simpler detection procedures. In this study the methods based on pixel covariances were found to perform better than the more widely used t-test. Among the covariance-based methods, none was found to be uniformly superior to the others. The performance of the GLRT served as an upper bound against which to compare the other methods. Our results suggest that correlation-based detectors are a promising direction for further investigation.