Objective: To discriminate Parkinson’s disease (PD) patients and controls in the Parkinson’s Progression Markers Initiative (PPMI) cohort according to MRI features using a novel a data-driven computational framework.

Background: Neuroimaging studies have provided invaluable insights in underlying mechanisms of PD, and many imaging-based biomarkers have been demonstrated to be closely related to presence and progression of PD, for example, MRI volumetric loss in the olfactory bulb and of the substantia nigra (SN) in PD, or changes in Fractional Anisotropy (FA) and other modalities. Existing studies have been hypothesis-driven with a focus on analysis of one or limited specific sequences.

Methods: We developed a data-driven computational framework for analyzing neuroimaging data from the PPMI study based on Graph Convolution Networks (GCN), which can extract nonlinear features from graphs in different frequency bands. We constructed brain networks using regions of interest (ROIs) from T1-weighted MRI and used DTI images to construct connectivity matrices using multiple tractography algorithms. We then used GCN to extract features from them. Those features were then further passed through a pooling layer and fed into a softmax function to classify PD cases and controls.

Results: Image acquisitions on 754 subjects from PPMI included 596 PD cases and 158 controls. Classification performance was evaluated with Area Under the receiver-operating characteristic Curve (AUC) value. Our method achieved 0.9537 AUC for 5-fold cross validation, while the traditional approach based on multilinear principal component analysis can only achieve 0.6443 AUC value. The best performance for GCN with a single graph modality was 0.9390. With 2-D visualization using student t-distributed Stochastic Neighbor Embedding, we can clearly observe our method is able to learn very compact representations on both PD case and control group. Our method also located the discriminative ROIs in the brain network, including the caudate and putamen areas.

Conclusions: GCN can effectively extract nonlinear and discriminative features from the graph/networks. Information fusion from multiple imaging modalities can further improve the classification performance. Discriminative features can also be effectively identified with our model.

« Back to 2018 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/discriminating-idiopathic-pd-patients-and-controls-in-ppmi-using-integrative-multi-modal-neuroimaging-analysis-with-graph-convolutions/