Traditional fMRI analysis has focused on modeling temporal changes in BOLD signals on a voxel-by-voxel basis to infer brain activation. To incorporate spatial information, we have previously proposed to model the spatial changes in the distribution of BOLD signals within regions of interest. Specifically, we examined changes in the spread of BOLD signals and the distribution of signals along the three coordinate dimensions of the MR scanner to detect activation. However, the scanner-based coordinate directions do not have any particular functional significance. Hence, in this paper, we demonstrate how canonical correlation analysis can be used to combine directional spatial features for detecting spatial distribution changes along the direction of maximal spatial response and validate this method on synthetic data. In addition, we apply our proposed method to real fMRI data collected from Parkinson's disease patients, which detected significant medication-induced effects that were neglected with traditional intensity-based measures.