Existing analytical techniques for functional magnetic resonance imaging (fMRI) data always need some specific assumptions on the time series. In this article, we present a new approach for fMRI activation detection, which can be implemented without any assumptions on the time series. Our method is based on a region growing method, which is very popular for image segmentation. A comparison of performance on fMRI activation detection is made between the proposed method and the deconvolution method and the fuzzy clustering method with receiver operating characteristic (ROC) methodology. In addition, we examine the effectiveness and usefulness of our method on real experimental data. Experimental results show that our method outperforms over the deconvolution method and the fuzzy clustering method on a number of aspects. These results suggest that our region growing method can serve as a reliable analysis of fMRI data.