Physics-Informed Deep Learning for Fast Cerebral Perfusion Assessment in Stroke Digital Twins

Sebastiaan Broos, Thijs P. Kuipers, Henk Marquering

Keywords: Digital twins, Physics-informed Neural Networks (PINNs), Implicit neural representations (SIREN), Neural PDE surrogates, Ischemic stroke, Cerebral perfusion, Geometric deep learning, Medical imaging (MRI)

TL;DR: Physics-informed SIREN surrogate learns Laplacian perfusion on patient brain meshes, matches finite element method accuracy, and runs about 500x faster, moving stroke digital twins toward real-time use.

Abstract: In acute ischemic stroke, reduced blood flow decreases cerebral perfusion. This lowers tissue oxygenation and triggers cell death that can result in severe neurological impairment. Finite element methods (FEM) are used to study cerebral perfusion and simulate blood-flow patterns in stroke, but their long computation times limit clinical use. We propose a physics-informed neural surrogate based on a sinusoidal representation network (SIREN) that reproduces Laplacian perfusion physics. The SIREN uses a lightweight encoder that embeds mesh geometry and estimates the normalized harmonic distance from the ventricles to the cortex and the corresponding blood-flow directions. In held-out patients, it achieved FEM-level accuracy (MAE $0.034$ on a $[0,1]$ scale, $9.5^{\circ}$ mean angular deviation) while 500 times reduction in computation time from $19$s to $0.036$s.

Serve As Reviewer: ~Sebastiaan_Broos1

Submission Number: 33