Physics-Informed Neural Networks for Brain Hemodynamic Predictions Using Medical Imaging

Area of Science:

• Biomedical Engineering
• Computational Fluid Dynamics
• Medical Imaging

Background:

• Accurate brain hemodynamics are crucial for diagnosing cerebrovascular diseases.
• Transcranial Doppler (TCD) ultrasound offers noninvasive velocity measurements but is spatially limited.
• Conventional physics-based models struggle with unknown boundary conditions.

Purpose of the Study:

• To develop a physics-informed deep learning framework for high-resolution brain hemodynamic mapping.
• To integrate sparse Transcranial Doppler (TCD) data with reduced-order model (ROM) simulations.
• To overcome limitations of conventional methods in estimating subject-specific hemodynamics.

Main Methods:

• A deep learning framework combining TCD velocity measurements and 3D angiography data.
• Integration of one-dimensional (1D) reduced-order model (ROM) simulations.
• Validation against four-dimensional (4D) flow magnetic resonance imaging (MRI) data.
• Application to diagnose cerebral vasospasm (CVS) using synthetic stenosis data.

Main Results:

• Generation of physically consistent, high spatiotemporal resolution hemodynamic maps (velocity, area, pressure).
• Accurate prediction of hemodynamic variables despite sparse input data and unknown boundary conditions.
• Successful validation against 4D flow MRI.
• Demonstrated capability in diagnosing cerebral vasospasm (CVS) by predicting vessel diameter changes.

Conclusions:

• The physics-informed deep learning approach provides accurate, subject-specific brain hemodynamic quantification.
• This method enhances diagnostic capabilities for cerebrovascular diseases, including cerebral vasospasm (CVS).
• The framework overcomes key limitations of traditional computational modeling in hemodynamics.