Discriminative analysis of multivariate features from structural MRI and diffusion tensor images.

Imaging markers derived from magnetic resonance images, together with machine learning techniques allow for the recognition of unique anatomical patterns and further differentiating Alzheimer's disease (AD) from normal states. T1-based imaging markers, especially volumetric patterns have demonstrated their discriminative potential, however, rely on the tissue abnormalities of gray matter alone. White matter abnormalities and their contribution to AD discrimination have been studied by measuring voxel-based intensities in diffusion tensor images (DTI); however, no systematic study has been done on the discriminative power of either region-of-interest (ROI)-based features from DTI or the combined features extracted from both T1 images and DTI. ROI-based analysis could potentially reduce the feature dimensionality of DTI indices, usually from more than 10e+5, to 10-150 which is almost equal to the order of magnitude with respect to volumetric features from T1. Therefore it allows for straight forward combination of intensity based landmarks of DTI indices and volumetric features of T1. In the present study, the feasibility of tract-based features related to Alzheimer's disease was first evaluated by measuring its discriminative capability using support vector machine on fractional anisotropy (FA) maps collected from 21 subjects with Alzheimer's disease and 15 normal controls. Then the performance of the tract-based FA+gray matter volumes-combined feature was evaluated by cross-validation. The combined feature yielded good classification result with 94.3% accuracy, 95.0% sensitivity, 93.3% specificity, and 0.96 area under the receiver operating characteristic curve. The tract-based FA and the tract-based FA+gray matter volumes-combined features are certified their feasibilities for the recognition of anatomical features and may serve to complement classification methods based on other imaging markers. [ABSTRACT FROM AUTHOR]