Different modalities of neuroimaging data can provide complementary information about neurodegenerative diseases such as Alzheimer's Disease (AD). In this study, we propose to use multiview learning method - Deep Generalized Canonical Correlation Analysis (DGCCA) to learn correlated components from multimodal imaging data. DGCCA can project multiview features to a latent space by applying non-linear transformation using neural networks and is able to explain more variance than it's linear counterpart, Generalized Canonical Correlation Analysis (GCCA), with low dimensionality. Subsequent clustering analysis on the learned components from DGCCA shows that it's able to identify population structure - namely case control relationship in an AD cohort. Further genetic association analysis on clustering results demonstrates that DGCCA can yield a population structure with stronger genetic basis than GCCA and single modality features.