Parkinson's Disease Detection on DaTscan Images Using Random Forest and Graph Deep Learning

Parkinson's disease is characterized by motor disorders caused by the progressive death of neurons producing dopamine. DaTscan imaging, namely Single-Photon Emission Computed Tomography (SPECT) with ioflupane (<tex>${ }^{123} I$</tex>), allows to quantify this loss of neurodopaminergic neurons, more particularly observed in the striatum. In this article, a two-step classification method is proposed to detect Parkinson's disease on DaTscan images focusing on this key structure. A graph is first created from the segmentation of the striatum, relying on the creation of the Max-Tree of the DaTscan image and a prior segmentation. Then, a Random Forest algorithm is applied to classify the nodes of the striatal graph, computing features extracted locally or related to subject's properties. Finally, the classification is made on the whole graph using Deep Learning, and adding the Random Forest predictions as new features. The combination of these two approaches offers promising results on the Parkinson's Progression Markers Initiative (PPMI) database, achieving an average balanced accuracy of <tex>$91.1\%$</tex> using a nested cross-validation strategy to increase the robustness of the results.