Multimodal MRI classification in vascular mild cognitive impairment

Abstract

Vascular mild cognitive impairment (VMCI) is a disorder in which multimodal MRI can add significant value by combining diffusion tensor imaging (DTI) with brain morphometry. In this study we implemented and compared machine learning techniques for multimodal classification between 58 VMCI patients and 29 healthy subjects as well as for discrimination (within the VMCI group) between patients with different cognitive performances. For each subject, a cortical feature vector was constructed based on cortical parcellation and cortical and subcortical volumetric segmentation and a DTI feature vector was formed by combining descriptive statistical metrics related to the distribution of DTI invariants within white matter. We employed both a sequential minimal optimization and a functional tree classifier, using feature selection and 10-fold cross-validation, and compared their performances in monomodal and multimodal classification for both classification problems (healthy subjects vs VMCI and prediction of cognitive performance). While monomodal classification resulted in satisfactory performance in most cases, turning from monomodal to multimodal classification resulted in an improvement of the performance in the discrimination between VMCI patients with low cognitive performance and healthy subjects by up to 10% in sensitivity (leaving specificity unchanged). We therefore are able to confirm the usefulness of machine learning techniques in discriminating diseased states based on neuroimaging data.