Reconstruction of diffusion anisotropies using 3d deep convolutional neural networks in diffusion imaging

Abstract

The reconstruction of neural pathways is a challenging problem in case of crossing or kissing neuronal fibers. High angular resolution diffusion imaging models are required to identify multiple fiber orientations in a voxel. Disadvantage of those models is that they require a multitude of acquired gradient directions, otherwise these models become inaccurate. We present a new approach to derive the fiber orientation distribution function using a Deep Convolutional Neural Network, which remains stable, even if less gradient directions are acquired. In addition, the Convolutional Neural Network is able to improve the signal in a voxel by extracting useful information of surrounding neighboring voxels. Subsequently, the functionality of the network is evaluated using 100 different brain datasets from the Human Connectome Project.