Super-Resolved q-Space Deep Learning

Abstract

q-Space deep learning (q-DL) enables accurate estimation of tissue microstructure from diffusion magnetic resonance imaging (dMRI) scans with signals undersampled in the q-space. However, in many scenarios, such as clinical settings, the quality of tissue microstructure estimation is limited not only by q-space undersampling but also by low spatial resolution. Therefore, in this work, we extend q-DL to super-resolved tissue microstructure estimation, which is referred to as super-resolved q-DL. In super-resolved q-DL, low resolution (LR) image patches of diffusion signals are mapped directly to high resolution (HR) tissue microstructure patches with a deep network. Specifically, inspired by the successful integration of sparse representation into q-DL, we have designed an end-to-end deep network that comprises two functional components. The first component computes a sparse representation of diffusion signals at each voxel via $1^{3}$ convolutions, where the network structure is constructed by unfolding an iterative optimization process. In the second component, convolutional layers with different kernel sizes are used to compute HR tissue microstructure patches from the LR patches of sparse representation. The weights in the two components are learned jointly. Experiments were performed on brain dMRI data with a reduced number of diffusion gradients and a low spatial resolution, where the proposed approach outperforms competing methods.