Quantifying inter-hemispheric differences in Parkinson’s Disease using siamese networks

Abstract

Classification of medical imaging is one of the most popular application of intelligent systems. A crucial step is to find the features that are relevant for the subsequent classification. One possibility is to compute features derived from the morphology of the target region in order to check its role in the pathology under study. It is also possible to extract relevant features to evaluate the similarity between different regions, in addition to compute morphology-related measures. However, it can be much more useful to model the differences between regions. In this paper, we propose a method based on the principles of siamese neural networks to extract informative features from differences between two brain regions. The output of this network generates a latent space that characterizes differences between the two hemispheres. This output vector is then fed into a linear SVM classifier. The usefulness of this method has been assessed with images from the Parkinson’s Progression Markers Initiative, demonstrating that differences between the dopaminergic regions of both hemispheres lead to a high performance when classifying controls vs Parkinson’s disease patients.