Dimensionality Reduction of Cortical Thickness Measurements

Neural atrophy patterns in the cerebral cortex  are closely correlated to noticeable cognition decline due to  Alzheimer's disease. These patters of thinning can be analyzed by processing structural magnetic resonance images.  Freesurfer software is used to automatically segment white and gray matter in the brain, label the cortical and subcortical regions, register all of the brains to a common template, perform cortical thickness parcellation and thickness map smoothing. Approximately 300,000 cortical thickness measurements are computed from the whole cortex of every patient. This data has dimension equal to the number of thickness measurements taken on each patient. In order to visualize this high-dimensional data and find features related to Alzheimer's disease we must reduce the dimensionality. We present a method for dimensionality reduction, consisting of anatomically subdividing the brain into patitions and performing principal component analysis to identify a small subset of the original variables that contain the most information about the variance in the data.

Figure 1. Pipeline of Dimensionality Reduction of Cortical Thickness Measurements Using Principal Component Analysis (PCA) of Patches of the Brain Generated by K-Means Clustering of Freesurfer (FS) Labels