AI-Based Physics Geometry Adaptation Framework for Accurate and Generalizable Hemodynamic Modeling

Heye Zhang, Weiyuan Lin, Zhifan Gao, Xiujian Liu, Guang Yang, Lingyun Zu, Ge Guo, Zhihui Zhang, Dan Deng, Hui Liu, Rui Chen, Lei Xu, Zhen ZhouResearch Square 2025 · DOI

Abstract

Three-dimensional Hemodynamic modeling and application demonstrate significant value. In the paper, we propose a 3D hemodynamic computation framework (PGAF) integrating variational inference and optimal transport theory. Variational inference enables adaptive inference of reasonable initial solutions for unseen cases by linking vascular features and boundary conditions with latent hemodynamic patterns. Optimal transport theory constructs a metric between predicted flow fields and established hemodynamic solution space, which is established through analysis of large-scale computational fluid dynamics (CFD) simulation data, to constrain global physical deviation. PGAF implements a dynamic coupling correction mechanism where the hemodynamic results are iteratively refined through the optimization, so it ensures global physical consistency in vessels and calculation accuracy during evolution. We collect 2500 different vessels from 7 public image datasets and perform 8000 CFD simulations under various boundary conditions. PGAF is ultimately validated and evaluated in 500 cases containing invasive catheter measurements. The results demonstrate the ability of our method to accurately predict the velocity and pressure distributions within the vessels, demonstrating high generalizability, robustness, and precision under different boundary conditions.

Introduction

Understanding hemodynamics—the study of blood flow through the vascular system—is essential for the diagnosis and treatment of cardiovascular diseases. Traditionally, researchers have relied on Computational Fluid Dynamics (CFD) to calculate pressure and velocity fields. While CFD is considered the "gold standard" for accuracy, it is computationally expensive, often requiring over eight hours per case and significant manual intervention for complex patient geometries.

Recent advancements in deep learning, such as Physics-Informed Neural Networks (PINNs) and UNets, offer faster alternatives. However, these models frequently struggle with physical inconsistency—often "hallucinating" flow patterns that violate mass conservation—and poor generalization when encountering anatomical structures or pathological conditions (e.g., severe stenosis) not present in the training data. This paper introduces the Physics-Geometry Adaptation Framework (PGAF) to address these limitations by bridging the gap between data-driven efficiency and physics-based rigor.

Methodology

The PGAF framework utilizes a three-component architecture to ensure that AI-generated predictions remain physically consistent and highly accurate across diverse vascular geometries.

  1. Geometric Encoding (PointNet): Unlike traditional models that use rigid grids, PGAF employs PointNet to process 3D point clouds of blood vessels. This approach captures the intricate nuances of complex shapes, such as bifurcations and sharp curves, more effectively than voxel-based methods.
  2. Conditional Variational Inference (CVAE): To handle the variability of unseen cases, the model uses a probabilistic approach. It links vascular features and boundary conditions (such as inlet pressure) to a latent space of hemodynamic patterns, providing a robust initial solution for any new geometry.
  3. Optimal Transport (OT) Refinement: This component acts as a dynamic coupling correction mechanism. By using the Wasserstein distance, the framework compares the initial AI prediction against a manifold of physically valid solutions derived from thousands of CFD simulations. The solution is then iteratively "transported" toward a state that satisfies the Navier-Stokes equations, ensuring global physical consistency.
The PGAF architecture integrates geometric encoding, variational inference, and optimal transport refinement to ensure physical consistency.

Experimental Setup and Data

To ensure the robustness and generalizability of the framework, the study utilized a massive multi-fidelity dataset:

  • Vascular Reconstruction: 2,500 distinct vessels were reconstructed from seven public image datasets, including ImageCAS, SegTHOR, and TotalSegmentator.
  • Physics Manifold: The research team performed 8,000 CFD simulations under varying boundary conditions to establish a comprehensive solution space for the Optimal Transport refinement.
  • Clinical Validation: The model was validated against 500 real-world cases involving invasive catheter measurements—the clinical gold standard—and Doppler ultrasound data.

Key Results and Discussion

The PGAF framework demonstrated significant improvements over existing deep learning models and traditional CFD methods:

  • High Accuracy and Generalization: In aortic cases, PGAF achieved an Area Under the Curve (AUC) of 0.91, closely approaching the CFD gold standard of 0.95. It maintained the lowest Root Mean Square Error (RMSE) across five different pathological scenarios, including aortic coarctation and coronary artery disease, outperforming UNet, PINN, and DeepONet.
  • Clinical Utility: Predictions for Fractional Flow Reserve (FFR), a critical metric for heart disease assessment, showed a strong correlation ($r = 0.84$) with invasive pressure-wire measurements in coronary arteries.
  • Efficiency: By reducing the analysis pipeline from several hours to just a few minutes, PGAF makes real-time clinical monitoring and rapid diagnostic assessment a practical reality.
  • Uncertainty Quantification: The use of variational inference allows the model to provide confidence assessments, a vital feature for high-stakes clinical decision-making.

Conclusion

PGAF represents a significant step forward in hemodynamic modeling, successfully decoupling physics from geometry to allow for high generalizability. While the current model assumes laminar flow, future iterations will aim to incorporate explicit physical constraints for transitional and turbulent flows to improve accuracy in cases of extreme stenosis. By slashing computation time without sacrificing physical rigor, this framework paves the way for the integration of advanced hemodynamic modeling into routine clinical workflows.

Summary updated

Abstract

<title>Abstract</title> Three-dimensional Hemodynamic modeling and application demonstrate significant value. In the paper, we propose a 3D hemodynamic computation framework (PGAF) integrating variational inference and optimal transport theory. Variational inference enables adaptive inference of reasonable initial solutions for unseen cases by linking vascular features and boundary conditions with latent hemodynamic patterns. Optimal transport theory constructs a metric between predicted flow fields and established hemodynamic solution space, which is established through analysis of large-scale computational fluid dynamics (CFD) simulation data, to constrain global physical deviation. PGAF implements a dynamic coupling correction mechanism where the hemodynamic results are iteratively refined through the optimization,so it ensures global physical consistency in vessels and calculation accuracy during evolution. We collect 2500 different vessels from 7 public image datasets and perform 8000 CFD simulations under various boundary conditions. PGAF is ultimately validated and evaluated in 500 cases containing invasive catheter measurements. The results demonstrate the ability of our method to accurately predict the velocity and pressure distributions within the vessels, demonstrating high generalizability, robustness, and precision under different boundary conditions.
BibTeX
@inproceedings{zhang2025ai,
  author = {Heye Zhang and Weiyuan Lin and Zhifan Gao and Xiujian Liu and Guang Yang and Lingyun Zu and Ge Guo and Zhihui Zhang and Dan Deng and Hui Liu and Rui Chen and Lei Xu and Zhen Zhou},
  title = {{AI-Based Physics Geometry Adaptation Framework for Accurate and Generalizable Hemodynamic Modeling}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-6779296/v1},
}
Figure