Estimating blood-brain barrier exchange rate using diffusion-weighted arterial spin labeling and physics-guided neural networks.

This study proposes physics-guided neural network (PgNN) approaches for estimating perfusion and blood-brain barrier (BBB) water permeability parameters from multi-slice, multi-delay diffusion-weighted arterial spin labeling (DW-ASL) MRI. To enable simultaneous estimation of perfusion and permeability-related parameters, a water-impermeable arterial compartment was incorporated into a two-compartment perfusion kinetic model, allowing estimation of the water exchange rate (kw), cerebral blood flow (CBF), and arterial transit time (ATT). DW-ASL data were acquired using pseudo-continuous arterial spin labeling with 2D gradient-echo-planar imaging and bipolar diffusion gradients in sixteen subjects (nine young healthy adults, three elderly healthy adults, and four elderly patients with Alzheimer's disease). Multiple post-labeling delays were achieved by permuting the slice acquisition order. Two PgNN frameworks were developed. First, PgNN based on a multi-layer perceptron (MLP) was trained using synthetic DW-ASL signals generated from the physical model and adapted to each test dataset. Second, a 3D U-Net-based PgNN was devised to adapt to test data using the physical model without pretraining. Neither approach required training with real in vivo data. The proposed methods were compared with conventional alternating direction method of multipliers (ADMM) and nonlinear least squares (NLS). Both PgNN approaches demonstrated improved accuracy and noise robustness compared with ADMM and NLS. The estimated kw was lower and ATT higher in older than young subjects, with the kw reduction being more pronounced in the Alzheimer's disease subgroup. These results indicate that the proposed frameworks provide stable quantitative estimation of perfusion and BBB permeability parameters from multi-slice, multi-delay DW-ASL.