Although simple averaging and Gaussian spatial smoothing of neighboring time series can suppress the noise of fMRI, but they may degrade the activated areas. As an alternative approach, the canonical correlation analysis (CCA) performs a weighted averaging of time series data such that the resulted time series has maximum correlation with the bases of a signal subspace. In this paper, we select only the most similar neighbors of each voxel for further adaptive averaging via CCA. Thus for an inactive central voxel, the surrounding active voxels are eliminated from weighted averaging. This intelligent selection prevents the false spreading of activated areas. After spatial filtering, we used the results of CCA (maximum cross correlation) for activation detection. We applied our method on simulated and experimental fMRI data and compared it with the conventional CCA (without intelligent selection) and match (spatial) filter. The ROC curve obtained from simulated data shows the superior performance of our proposed method.