Blood-brain barrier estimation from DCE MRI using physics-based deep learning approach

Motivation: Pharmacokinetic modelling of DCE MRI faces significant challenges of noisy parameter maps, unreliability, and long processing time. Goal(s): To propose a deep learning approach for estimating blood-brain barrier (BBB) permeability in DCE MRI, comparing its performance with the standard non-linear least squares (NLLS) method. Approach: A physics-derived neural network (DCE-PhysicsNet) was trained on simulated DCE-MRI signals from the extended Tofts-Kety model with patient-specific arterial input function. Results: Our approach outperformed NLLS in the digital phantom with a coefficient of variation up to 50% for the estimated permeability surface area product (PS). Generally, DCE-PhysicsNet parameter estimations correlate with previous studies despite differences in approach. Impact: The proposed deep learning approach demonstrates a reliable estimation of BBB parameters for DCE MRI using only a contrast-enhanced scan. It has the potential to facilitate the wide clinical use of DCE MRI.