Multiphysics learning with graph neural networks for thrombosis prediction in intracranial aneurysms

Area of Science:

• Biomedical Engineering
• Computational Fluid Dynamics
• Machine Learning

Background:

• Intracranial aneurysms (IAs) affect 3-5% of the population, with rupture causing life-threatening subarachnoid hemorrhage.
• Predicting thrombus formation in IAs is crucial for endovascular treatments but challenging due to complex hemodynamics and biochemical factors.
• Current Computational Fluid Dynamics (CFD) models are too slow for real-time clinical use.

Purpose of the Study:

• To investigate the efficacy of Transformer-based Graph Neural Networks (GNNs) for predicting thrombus formation in patient-specific IA geometries.
• To explore multitask learning strategies for GNNs applied to multiphysics simulations.
• To develop a computationally efficient model for aiding clinical decision-making in IA treatment.

Main Methods:

• Utilized Transformer-based GNN architectures capable of handling unstructured meshes common in CFD.
• Trained models on multiphysics simulation data for IA thrombus formation.
• Investigated multitask learning by training separate models for different physical fields.

Main Results:

• Transformer-based GNNs achieved state-of-the-art accuracy in predicting thrombus formation.
• The approach significantly reduced computational cost compared to traditional solvers.
• Separate models per physical field demonstrated superior performance and enabled parallelization.
• The model showed robustness to variations in inflow conditions and cardiac cycles.

Conclusions:

• Transformer-based GNNs offer a highly accurate and computationally efficient method for predicting thrombus formation in IAs.
• This approach has significant potential for real-time clinical decision support in endovascular treatment planning.
• Multitask learning with separate field models enhances GNN performance in complex physical simulations.