Super-Resolution of Manifold-Valued Diffusion MRI Refined by Multi-modal Imaging

Abstract

Diffusion MRI is the foundation for understanding the structures and disorders in the human connectome, but low spatial resolution fundamentally limits this understanding. Methods for increasing DTI resolution post hoc must carefully utilize all available information to reduce bias and uncertainty in this ill-conditioned inverse problem. Previous machine learning approaches have largely cast this problem as a standard super-resolution task without taking advantage of domain knowledge surrounding DTIs. Outside this domain, recent work in super-resolution has given attention to preserving an input’s fine-scale information as it is passed through a network. Our contribution consists of a novel deep learning model for DTI super-resolution with three important advancements: 1) a novel procedure for refining DTI predictions with high-resolution T2-weighted images, 2) interpolation over log-Euclidean tensors that is immune to the “swelling effect,” and 3) the effective use of densely-connected residual networks that preserve detail from the input. Through experiments on HCP data, we show that our model achieves the best performance in the literature for increasing the resolution of DTIs. We further analyze the effect of each proposed component with thorough model ablation tests.