Features defined on the cortical surface derived from magnetic resonance imaging provide important information to distinguish normal controls from Alzheimer's disease (AD) and mild cognitive impairment (MCI). We adopted cortical thickness and sulcal depth, parameterized by three dimensional meshes, as our feature. The cortical feature is high dimensional and direct use of it is problematic in a modern classifier due to small sample size problem. We applied manifold learning to reduce the dimensionality of the feature and then tested the usage of the dimensionality reduced feature with a support vector machine classifier. A leave-one-out cross-validation was adopted for quantifying classifier performance. We chose principal component analysis (PCA) as the manifold learning method. We applied PCA to a region of interest within the cortical surface. Our classification performance was at least on par for the AD/normal and MCI/normal groups and significantly better for the AD/MCI groups compared to recent studies. Our approach was tested using 25 AD, 25 MCI, and 50 normal control patients from the OASIS database.
Bibliographical noteFunding Information:
This study was supported by Basic Science Research Program through NRF Korea grants 2012005939, 20100023233 , Global Frontier RD Program through NRF Korea grant NRF-M1AXA003-2011-0032035 , and KOSEF NLRL Program grant 2011-0028333 . Image data collection was supported by NIH grants P50AG05681 , P01AG03991 , R01AG021910 , P50MH071616 , U24RR021382 , and R01MH56584 .
All Science Journal Classification (ASJC) codes