Physics-Informed Neural Operators for Parameter Inference in Multi-vessel Cardiovascular Networks | JoVE Visualize

Area of Science:

Cardiovascular Physiology
Computational Fluid Dynamics
Artificial Intelligence in Medicine

Background:

Accurate patient-specific hemodynamic assessment is crucial but limited by computational costs and sparse clinical data.
Traditional computational solvers are expensive, hindering real-time clinical applications.
Developing efficient and accurate methods for hemodynamic modeling is essential.

Purpose of the Study:

To demonstrate the efficacy of physics-informed neural operator surrogates for emulating multi-vessel 1D hemodynamics.
To enable Bayesian parameter inference and non-invasive pressure estimation in a clinically relevant setting.
To overcome the limitations of cost and data sparsity in hemodynamic assessment.

Main Methods:

Constructed neural operator surrogates (DeepONet, POD-DeepONet, FNO) for a 17-vessel arterial network.
Incorporated conservation laws and bifurcation conditions via physics-informed loss terms.
Utilized a Wasserstein autoencoder for generating inflow boundary conditions and embedded surrogates in a Bayesian pipeline for parameter inference and pressure prediction in Fontan patients using 4D flow MRI data.

Main Results:

The physics-informed Fourier Neural Operator (FNO) architecture achieved superior accuracy, reducing median relative and maximum errors compared to data-driven methods.
Physics-informed neural operators (PINO) demonstrated more accurate vascular parameter recovery than other surrogate models in inverse tests.
The framework successfully reproduced measured flow waveforms and provided non-invasive brachial pressure predictions consistent with clinical measurements for Fontan patients, including uncertainty quantification.

Conclusions:

Physics-informed neural operators offer a highly accurate and computationally efficient alternative to traditional hemodynamic solvers.
The proposed framework facilitates practical, uncertainty-aware estimation of vascular parameters and non-invasive pressure waveforms from sparse clinical flow data.
This approach holds significant potential for advancing patient-specific cardiovascular analysis and clinical decision-making.