An Accelerated Training Framework for Physics-Informed Neural Networks: Applications in Ultrafast Ultrasound Blood...

Area of Science:

• Biomedical Engineering
• Computational Fluid Dynamics
• Artificial Intelligence

Background:

• Ultrafast ultrasound enables high-volume blood flow data acquisition.
• Physics-informed Neural Networks (PINNs) solve Navier-Stokes equations but are slow for ultrafast imaging.
• Current PINN methods are computationally prohibitive for real-time applications.

Purpose of the Study:

• To develop accelerated PINN training frameworks for ultrafast ultrasound blood flow imaging.
• To enable efficient and accurate solutions to the Navier-Stokes equations for complex blood flow dynamics.
• To facilitate real-time, imaging-based blood flow assessment in clinical settings.

Main Methods:

• Introduced SeqPINN: a sequential time-domain discretization and test-time adaptation approach.
• Developed SP-PINN: a parallel training scheme using averaged constant stochastic gradient descent initialization.
• Utilized Stochastic Weight Averaging Gaussian for uncertainty estimation and generalizability assessment.

Main Results:

• SeqPINN and SP-PINN demonstrated significant speedups compared to standard PINN.
• Achieved low Root Mean Square Errors in blood flow velocity recovery: 0.63-0.81 cm/s (straight vessel) and 1.07-1.41 cm/s (trifurcate vessel).
• Validated performance through finite-element simulations and in vitro phantom studies.

Conclusions:

• SeqPINN and SP-PINN offer computationally efficient solutions for Navier-Stokes equations in ultrafast ultrasound.
• These methods pave the way for real-time PINN training and clinical blood flow assessment.
• The developed algorithms significantly advance the application of AI in medical imaging and hemodynamics.