Background: Multimodal MRI may support single-subject differentiation diagnosis of Alzheimer's disease (AD) and behavioral variant frontotemporal dementia (bvFTD). Yet, what MRI modality combinations maximize classification accuracy is not yet known. We analyzed various MRI modality combinations to determine what combinations suit single-subject classification of AD patients, bvFTD patients, and control subjects best. Methods: Thirty-seven patients with probable AD, 28 patients with bvFTD, and 40 cognitively normal subjects (Controls) (Table 1) underwent diffusion tensor imaging (DTI), resting-state functional, and T1-weighted structural MRI at 3T. From each T1-w image, probabilistic gray matter (GM) maps were calculated and used to estimate local GM density (GMD) within 96 Harvard-Oxford cortical and 14 deep GM regions. 20 tracts of the Johns-Hopkins-University tractography atlas were used to determine local white matter density (WMD) values and, after projection on the study-specific skeleton calculated with tract-based spatial statistics, local measures of fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity, and radial diffusivity. Seventy rs-fMRI components, derived from a temporally concatenated independent component analysis, were used to calculate the components' pairwise correlations and partial correlations. These features were used to establish classification accuracy using nested cross-validation classification analyses with elastic net classifiers. Individual and combinations of MRI modalities were step-wise evaluated to determine classification accuracy, expressed as area under the receiver operating characteristic curves (AUC). A three-group classification model was used to determine the accuracy for differentiating between AD, bvFTD, and Control subjects. Three two-group models were calculated to differentiate AD from Controls, bvFTD from Controls, and AD from bvFTD subjects. Results: Classification accuracy for three-group and AD versus Control classifications were highest with GMD-based features ((AUC [min-max]) 0.7±5[0.711-0.847] and 0.933[0.886- 0.963]) (Table 2). bvFTD versus Control and AD versus bvFTD classification models were most accurate using DTI-based features (FA: 0.860[0.807-0.913], MD: 0.693[0.614-0.821]). Overall, rsfMRI features did not improve classification accuracy. Conclusions: MRI-based differentiation of dementia forms may depend on dementia- type and MRI modality. T1-w-based measures particularly supported AD classifications, whereas DTI-based measurements facilitated bvFTD classifications. Functional connectivity measures and multimodal combinations did not substantially improve dementia- type discrimination. For MRI-based differential dementiatype diagnosis careful selection of MRI modalities is warranted.