Quantitative Water Permeability Mapping using Biophysical-modeling-based Deep Learning

Motivation: In diffusion-weighted arterial spin labeling (DW-ASL) images, quantification of the water exchange rate $$$k_{w}$$$ uses a single-pass approximation (SPA) which introduces systematic error while fitting the non-linear model is difficult. Goal(s): Our goal was to reduce the blood-brain-barrier (BBB) water exchange rate ($$$k_{w}$$$) quantification errors in DW-ASL images. Approach: We introduced the biophysical-modeling-based deep learning method (QTMNet) and tested both the simulated and in vivo data. Results: On simulated data, QTMNet has 90% less normalized root mean square error (NRMSE) compared to the traditional kinetic model. Impact: The improvement in evaluation accuracy by QTMNet may benefit Alzheimer’s Disease detection where $$$k_{w}$$$ has significant reduction.