FIGURE 2.

Pipeline for the proposed experimental approach. On the left, an overview of the structural connectome generation is shown. The PD25 subcortical atlas was nonlinearly registered with the T1 skull-stripped volumes. T1 brain extractions were visually inspected. If large parts of the brain were missing, a dilated T2 skull-stripped mask was used to compute the T1 brain extraction. Next, the T1 volume was registered to the b0 dMRI volume space using a standard nonlinear registration. Then, the atlas was transformed onto the b0 dMRI space to parcellate the diffusion volumes into 16 subcortical areas, and a probabilistic whole brain streamline tractography and its associated structural connectome were created. Averaged fractional anisotropy and averaged mean diffusivity connectomes were also computed. On the right, the various steps to create the longitudinal connectomes and output the Parkinson’s disease-relevant progression metric are shown. An L1-Norm distance metric was used to build the longitudinal connectomes between baseline and year-1 follow-up for the structural, fractional anisotropy, and mean diffusivity connectomes. Finally, the three longitudinal connectomes of each Parkinson’s disease (PD) and Control participant were concatenated into a single feature vector and used to train an L1-norm regularized logistic regression classifier. Further, the obtained progression metric from the trained datasets was generalizable to discriminate PROD from Controls.