In current neuroimaging analysis, feature selection majorly focuses on analysis within single brain regions. However, the fact that brain activities are usually associated with multiple brain regions highlights the importance of the multi-brain-region interaction, which is underexplored. To address this challenge, we propose a multi-stratification feature selection framework for analysing multiple brain regions in Magnetic Resonance Imaging (MRI). This framework consists of two major modules: intra-Region of Interest (ROI) module and inter-ROI module. IntraROI module selects representative features for each brain region by analysing both of the statistical difference of features and the classifier performance of the candidate subset. Inter-ROI module employs the evaluation function to guide the search, sequentially adding features from brain regions based on the corresponding predictive capacity. Only relevant and maximum joint significance features that improve the evaluation performance would be selected in this module. The proposed framework was validated on the diagnostic task of Alzheimer's disease. T1MR images were collected from 196 Alzheimer's disease patients and 259 normal control subjects. The experiments demonstrated that the proposed multi-stratification feature selection outperformed the state-of-the-art single-brain-region analysis and the radiomics early integration methods applied to multiple-brain-region, achieving AUC 0.913.