Reconstructing cerebral hemodynamics from sparse data using Neural Operator Transformers.

Area of Science:

• Computational fluid dynamics
• Cerebrovascular hemodynamics
• Artificial intelligence in medicine

Background:

• Cardiovascular diseases necessitate advanced methods for analyzing brain hemodynamics.
• Current computational models for the Circle of Willis lack real-time clinical applicability due to accuracy and speed limitations.

Purpose of the Study:

• To develop a novel computational framework for accurate and fast reconstruction of cerebral hemodynamics.
• To enable real-time inference of subject-specific boundary conditions for personalized cerebrovascular diagnostics.

Main Methods:

• Integration of a 1D reduced-order blood flow model with General Neural Operator Transformer and Variational Autoencoding Neural Operator architectures.
• Generation of synthetic data via finite-element simulations for training neural operators.
• Application of an inverse procedure for hemodynamic reconstruction and adaptive grid search for boundary condition estimation.

Main Results:

• Surrogate model achieved <1% mean relative error in velocity and area predictions for major Circle of Willis vessels.
• Global error in reconstructing entire networks from sparse measurements was approximately 3.3%.
• Adaptive boundary condition estimation accurately reproduced outlet pressures, facilitating case-specific calibration.

Conclusions:

• The neural-operator-based framework provides fast and accurate cerebral hemodynamic reconstruction from limited data.
• This approach holds significant potential for real-time clinical integration and personalized digital-twin workflows in cerebrovascular diagnostics.