DMR-CNN: A CNN Tailored For DMR Scans With Applications To PD Classification

Convolutional neural networks are ubiquitous in Machine Learning applications for solving a variety of problems. They however can not be used in their native form when the domain of the data is commonly encountered manifolds such as the sphere, the special orthogonal group, the Grassmannian, the manifold of symmetric positive definite matrices and others. Most recently, generalization of CNNs to Riemannian homogeneous spaces have been reported in literature. In this work, we propose an end-to-end CNN architecture for classification of diffusion MRI (dMRI) signals, dubbed dMR-CNN. In each voxel of the dMRI scan, the signal is acquired as a real number along each diffusion sensitizing magnetic field direction over a hemisphere of directions in 3D. Hence, in each voxel, we have a function $f : \mathbf { S } ^ { 2 } \times P _ { 1 } \rightarrow \mathbf { R }$. We formulate a definition of correlation on this space to extract intra-voxel features and then use standard CNN model to capture the spatial interactions between the intra-voxel features. Our proposed framework comprises of architectures to extract these intra- and inter- voxel features. We present an experimental setup to classify dMRI scans acquired from a cohort of 44 Parkinson Disease patients and 50 control$/$normal subjects.