Magnetic Resonance Imaging (MRI) provides three-dimensional anatomical and physiological details of the humanbrain. We describe the Integrated Voxel Analysis Method (IVAM) which, through machine learning, classifies MRI images of brainsafflicted with early Alzheimer’s Disease (AD). This fully automaticmethod uses an extra trees regressor model in which the featurevector input contains the intensities of voxels belonging to corre-sponding regions in the brain, whereby the effect of Alzheimer’son a single voxel can be predicted. The resulting tree is used in thefollowing two steps: a K-nearest neighbor (KNN) algorithm basedon Euclidean distance with the feature vector to classify wholeimages based on their distribution of affected voxels and a voxel-by-voxel classification by the tree of every voxel in the image.Voxel-by-voxel classification is followed by an Ising model filterto remove artifacts and to facilitate clustering of the classificationresults which identifies significant voxel clusters affected by AD.We apply this method to T1-weighted MRI images obtainedfrom the Open Access Series of Imaging Studies (OASIS) usingimages belonging to normal and early AD-afflicted individualsassociated with a Client Dementia Rating (CDR) which we useas the target in the supervised learning. Furthermore, statisticalanalysis of the results is conducted using pre-labeled brain atlasesto automatically pinpoint significantly affected brain regions.While achieving 96% AD classification accuracy on 233 images inthe OASIS dataset, the method reveals morphological differencescaused by the onset of AD.
Armen Aghajanyan, Matthew Hur