Geometry-Aware Abdominal Aortic Aneurysm Digital Twin for Patient-Specific Wall Stress Mapping

Area of Science:

• Biomedical Engineering
• Computational Mechanics
• Artificial Intelligence in Medicine

Background:

• Accurate abdominal aortic aneurysm (AAA) wall stress estimation is crucial for predicting rupture risk, moving beyond diameter-based criteria.
• High-fidelity finite element analysis (FEA) provides precise patient-specific stress data but is computationally intensive, limiting clinical application.
• A need exists for rapid, accurate methods to assess AAA biomechanics for improved patient management.

Purpose of the Study:

• To develop and evaluate a graph-based deep learning framework for rapid prediction of patient-specific AAA wall stress distributions.
• To compare the performance of different graph neural networks (GNNs) including Gated GraphConvolutional Network (GGCN), Equivariant Graph Neural Network (EGNN), and Graph Transformer (GT).
• To assess the potential of this framework for near real-time clinical decision-making in AAA management.

Main Methods:

• A graph-based deep learning framework using GGCN, EGNN, and GT was developed to predict wall stress from AAA surface meshes.
• Computed tomography angiography (CTA) data from 202 AAA patients were used to generate patient-specific meshes and extract geometric/biomechanical features.
• Node-specific features included wall thickness, intraluminal thrombus thickness, wall strength, curvatures, and normalized local diameter, with local information from nearest nodes.

Main Results:

• All GNNs demonstrated high fidelity in reproducing patient-specific FEA wall stress fields.
• The Graph Transformer (GT) achieved the best performance, accurately capturing spatial stress patterns and biomechanical trends (node-specific R2 > 0.97, graph-level R2 > 0.98).
• Inference time was reduced from hours (FEA) to seconds per case on a single GPU, enabling near real-time analysis.

Conclusions:

• The proposed graph-based deep learning framework, particularly the GT model, offers a rapid and accurate alternative to FEA for AAA wall stress prediction.
• This approach has the potential to significantly enhance AAA rupture risk assessment and facilitate routine patient-specific biomechanical analysis in clinical workflows.
• Further refinement and validation on larger patient cohorts are warranted for broader clinical integration.