Predicting flow velocity from fMRI inflow signals using physics-informed deep learning

Motivation: fMRI has been used to measure large scale cerebrospinal fluid (CSF) flow dynamics with high sensitivity and temporal resolution, however the measured signal is not quantitative. Goal(s): Our goal is to develop a physics-based neural network framework for flow quantification directly from fMRI flow-enhanced signals. Approach: We designed a neural network that can use fMRI data as input to predict flow velocity. We then trained the model on a simulated dataset generated using a physics-based model. Results: Validation on phantom and human data showed accurate predictions of flow velocity when using measured fMRI signals as input into the neural network. Impact: Here, we significantly increase quantitative information obtainable from fMRI which will enable neuroimaging researchers studying fluid flow dynamics to take advantage of the high sensitivity and temporal resolution of fMRI to obtain flow signals that are physically interpretable.