Alzheimer's disease (AD) induces large-scale neuro-degeneration which may underlie various cognitive problems, and Mild cognitive impairment (MCI) is assumed as its prodromal phase. Studies routinely use structural MRI and DTI to map neuroanatomical basis separately in AD while ignoring the relationship between different modalities. In this study, we use sparse canonical correlation analysis (SCCA), an unsupervised multivariate method, to identify mutually predictive regions across structural MRI and DTI, in a cohort of 32 AD, 15 MCI and 16 controls. We found significant correlations between gray matter density and fractional anisotropy (FA) within a distributed network (p < 0.001). Furthermore, multiple regression analysis shows that, within the SCCA identified network, clinical cognitive scores correlate with gray matter and white matter impairment in AD and MCI groups. In sum, SCCA is valuable to fuse information across modalities and reveal a degraded cortical-white matter network in AD.