Physics-Informed Neural Network for Quantifying Time-Encoded Arterial Spin Labeling: A Simulation Study

Alessandro Giupponi, Chiara Da Villa, Mattia Veronese, Marco Castellaro

Keywords: Arterial Spin Labeling, Physics-Informed Neural Network, Hadamard Encoding, Cerebral perfusion

TL;DR: This study presents a physics-informed neural network approach for estimating cerebral perfusion parameters—CBF and ATT—from simulated time-encoded ASL MRI data, demonstrating improved robustness to noise.

Abstract: Arterial Spin Labeling (ASL) MRI enables non-invasive quantification of cerebral perfusion. Hadamard time-encoding improves acquisition efficiency and allows the simultaneous estimation of cerebral blood flow (CBF) and arterial transit time (ATT) via the Buxton model. Physics-informed neural networks (PINNs) integrate physical laws into neural networks, improving parameter estimation under noisy and sparse data conditions. We propose a two-stage PINN framework trained on synthetic ASL data from the Boston ASL Template and Simulator. Leveraging coupled neural networks and differential equation constraints, our method produces smoother and more robust CBF and ATT maps compared to regularized nonlinear least squares (NLLS), demonstrating its potential for clinical ASL quantification. While this work focuses on simulation data, it represents a first step toward extending such models to in vivo applications using a similar architecture.

Submission Number: 73