A microscopy-trained model to predict super-resolution fibre orientations from diffusion MRI

Motivation: Microscopy provides fibre orientations at much higher resolutions than MRI. Diffusion MRI (dMRI) and microscopy in the same brain offer the opportunity to train a machine-learning model to super-resolve fibre orientation distributions (FODs). Goal(s): To super-resolve FODs in dMRI based on paired MRI-microscopy data, using a model that can be applied to MRI where microscopy is unavailable. Approach: We construct a microscopy-informed network that takes structural and single-shell dMRI, and outputs super-resolved FODs. Through domain adaptation, our network is applicable to both in-vivo and postmortem MRI. Results: Our super-resolution FODs from macaque and in-vivo human MRI demonstrate detailed neuroanatomy. Impact: We develop a microscopy-informed network that provides super-resolved FODs from single-shell dMRI, doubling the resolution. Notably, our network can be applied to in-vivo MRI where microscopy is unavailable and offers the possibility of more precise fibre tracking in widespread applications.