AneumoBench: A Source-Linked Benchmark for Synthetic-Geometry Transfer in Aneurysm CFD

arXiv:2505.14717 · eess.IV, cs.AI, cs.CV, cs.LG · Submitted 2025-05-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "AneumoBench: A Source-Linked Benchmark for Synthetic-Geometry Transfer in Aneurysm CFD".

Tom: This research introduces Aneumo, a large-scale,

Jane: First, who's behind it and why it matters.

Title and authors: Tom: Let's talk about the title and who put this paper together; "AneumoBench: A Source-Linked Benchmark for Synthetic-Geometry Transfer in Aneurysm CFD." The name itself tells you exactly what the study is about—it's a benchmark, which means they are setting up a standard for others to follow when they want to test their methods.

Jane: And the authors listed include people from institutions like Fudan University and Huashan Hospital, which suggests this work has strong connections to clinical medical research and high-level AI development. It shows a collaborative effort between different fields.

Lu: The authors clearly bring together expertise in computational fluid dynamics, geometry generation, and machine learning integration. They are the ones who understand how to translate complex physical simulations into data that an AI can actually learn from effectively.

Meng: When you look at the title again, "Source-Linked Benchmark," it implies that the synthetic geometries aren't just random shapes; they are derived directly from real patient models, which is a big deal for ensuring clinical relevance.

Lalam: I think what's important here is that this paper establishes a standard way to compare different approaches for using these types of multimodal datasets in aneurysm research. It gives the community a common yardstick to measure progress.

The paper's summary: Tom: So, if we look at the summary, they’re essentially describing how they took real patient models and used controlled deformation to create synthetic aneurysms across different growth stages. Then, they ran eight different steady-state mass flow conditions through these shapes to get a massive amount of hemodynamic data.

Jane: That means the core of their contribution is combining those high-fidelity CFD simulations—the velocity and pressure fields—with the geometric models and segmentation masks into one big resource for machine learning.

Lu: They are generating eighty-five thousand two hundred eighty hemodynamic data samples across ten thousand six hundred sixty synthetic geometries derived from four hundred twenty-seven real patient geometries, which is a substantial scale for this kind of medical simulation work.

Meng: The summary highlights that they provide segmentation masks similar to medical images and integrate these simulation results under eight physiological flow conditions, which sounds like a very complete picture for training purposes.

Lalam: This dataset is structured to be multimodal, offering binary ROI images for structure identification alongside the VTK files and NumPy versions of the fields, which makes it very easy for AI frameworks to ingest all necessary information at once.

The paper's improvements: Tom: Now, moving onto what they suggest as improvements in their approach, they are essentially proposing how this dataset can be used to advance data-driven modeling by testing different AI architectures. They show that using a DeepONet-SwinT model outperforms the standard DeepONet baseline when analyzing these hemodynamic predictions.

Jane: That’s interesting; they aren't just presenting data; they are showing how a specific, more sophisticated AI architecture can leverage this data better to make more accurate predictions about aneurysm evolution and rupture risk.

Lu: The improvements focus on validating the generalization capability of these deep learning models by testing them across different training set flow condition diversity and geometric scales, which is crucial for ensuring the models work in diverse real-world scenarios.

Meng: From an engineering standpoint, this validation step is vital because it proves that the model isn't just memorizing the training data but actually learning the underlying physics of blood flow dynamics.

Lalam: The paper suggests a rigorous benchmarking process using specific loss functions and metrics like MNAE and MSE to objectively compare models, which moves the discussion from "this model is good" to "this model performs better under these controlled conditions."

Conclusion: Tom: So, wrapping up, the main point of this AneumoBench paper is that they’ve created a comprehensive source-linked dataset that integrates real patient geometry with detailed CFD data across various flow conditions, setting a new standard for how we approach training AI for aneurysm analysis.

Jane: Essentially, they are giving researchers a powerful tool where they can move past relying only on morphology and start incorporating the actual physics of blood flow into their predictions, which is what this work aims to do by providing that multimodal resource.

Lu: The scale of eighty-five thousand two hundred eighty samples and the controlled generation process really underscore the potential for advancing simulation studies in this field by providing a realistic training environment.

Meng: For us, this means we have a much richer environment to test our generative design tools and see how they perform when simulating actual physiological states, rather than just idealized ones.

Lalam: I feel that the impact here is in making aneurysm risk assessment more predictive by moving it from a purely visual or morphological assessment toward one deeply informed by simulated fluid dynamics.

Tom: Exactly; this AneumoBench paper gives us a much better starting point for developing AI tools that can actually predict rupture risk based on what's happening inside the vessel, not just what the vessel looks like. What an exciting step forward for the whole field.

Artificial Intelligence Innovation and Incubation Institute, Fudan University

eess.IV, cs.AI, cs.CV, cs.LG

Submitted: 2025-05-19

Updated: 2026-09-30

Code: https://github.com/Xigui-Li/Aneumo

Project page: http://ecm2.mathcs.emory.edu/aneuriskweb

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 88/100

The gist: This research introduces Aneumo, a large-scale, multimodal dataset designed to advance data-driven approaches in intracranial aneurysm research by integrating high-fidelity computational fluid

Key concepts

Aneumo
A large dataset integrating high-fidelity CFD simulations with geometric models. It provides 85,280 hemodynamic samples across 10,660 synthetic aneurysm shapes based on real patient anatomy. This allows researchers to train and test data-driven algorithms for analyzing aneurysm behavior.
Multimodal Data Representation
The dataset organizes information into four formats: NIfTI masks for segmentation, STL meshes for geometry, VTK files for pressure/velocity fields, and NumPy arrays. This structure is designed to allow machine learning models to fuse spatial and flow data effectively during training.
DeepONet-SwinT
A proposed deep learning model architecture that uses a Swin Transformer-based geometric encoder. Benchmarking showed this model significantly outperforms the standard DeepONet baseline in predicting hemodynamic parameters, demonstrating superior accuracy in analyzing aneurysm evolution.

Terminology

Summary

This research introduces Aneumo, a large-scale, multimodal dataset designed to advance data-driven approaches in intracranial aneurysm research by integrating high-fidelity computational fluid dynamics (CFD) simulations with geometric models and segmentation masks. This dataset addresses the critical gap in publicly available resources by providing 85,280 hemodynamic data samples generated under eight steady-state mass flow conditions across 10,660 synthetic aneurysm geometries derived from 427 real patient geometries. By offering this comprehensive multimodal resource, the work aims to facilitate the development of efficient machine learning algorithms for analyzing aneurysm evolution and rupture risk in a data-driven manner.

Dataset Curation and Generation

The Aneumo dataset is constructed by synthesizing high-fidelity 3D models to simulate aneurysm evolution across different stages. This process involves systematically expanding the dataset based on real patient anatomy using controlled deformation techniques. Specifically, the methodology includes:

  1. Generating 10,660 high-quality 3D models via controlled deformation applied to a base set of 427 real aneurysm geometries to capture geometric evolution.

  2. Applying stochastic deformation and geometric optimization to ensure topological integrity and physiological plausibility after removing the aneurysm region.

  3. Converting these optimized 3D vascular models into segmentation masks using a voxelization process in 3D Slicer, resulting in NII.GZ format files (binary image files).

CFD Simulation and Data Acquisition

The authenticity of the synthetic shapes is confirmed by neurosurgeons, and subsequent CFD computations are performed on each shape under specific conditions to generate the hemodynamic data. The simulation parameters are rigorously controlled:

  1. CFD meshes are created for the deformed geometries using an unstructured polyhedral mesh in ANSYS Fluent Meshing 2023 R1, with a minimum mesh size of 0.15 mm selected for optimal balance between accuracy and cost.

  2. The fluid is modeled as an incompressible Newtonian fluid with a density of 1050 kg/m3 and a dynamic viscosity of 0.00345 Pa·s, simulating laminar flow behavior under steady-state conditions.

  3. Eight steady-state mass flow conditions are simulated, ranging from 0.001 to 0.004 kg/s, with the solver being OpenFOAM v2312 using the icoFoam solver and PISO algorithm for pressure-velocity coupling.

  4. A total of 85,280 CFD simulations were completed across all models under these conditions, generating key hemodynamic parameters such as velocity components (u, v, w) and pressure (p).

Multimodal Data Representation

The dataset is designed to support tasks requiring the fusion of different data modalities. The final output files are organized into standardized formats compatible with machine learning frameworks:

  1. Binary region-of-interest (ROI) images in NIfTI format (.nii.gz) serve as segmentation masks for vascular structures, enabling spatial alignment of simulation results with imaging data.

  2. 3D surface mesh models in STL format (.stl) capture the vascular geometry itself, used for 3D shape reconstruction tasks.

  3. VTK files in.vtp and.vtu formats represent pressure and velocity fields across different domains (inlet, internal, outlet, wall), providing the raw CFD data.

  4. NumPy-formatted (.npy) versions of these data fields are provided for each flow case to ensure seamless integration into machine learning pipelines.

Machine Learning Benchmarking

The dataset serves as a foundation for benchmarking deep learning architectures in hemodynamic prediction tasks, specifically comparing standard DeepONet against the proposed DeepONet-SwinT model. The evaluation process involves:

  1. Training models using the Sum of Square Error (SSE) loss function, weighted by factors such as λp = 1 × 105 and λu = λv = λw = 1 × 103 to minimize the total loss function L.

  2. Employing various evaluation metrics, including Mean Normalized Absolute Error (MNAE), Mean Squared Error (MSE), and Mean Absolute Error (MAE).

  3. Systematically testing the model's generalization capability by performing experiments across different training set flow condition diversity, batch sizes, and data scales to analyze scaling performance. The study demonstrated that DeepONet-SwinT consistently exhibits significantly lower median error and superior prediction accuracy compared to the baseline DeepONet, particularly when leveraging its Swin Transformer-based geometric encoder.

Conclusion and Future Directions

The Aneumo dataset provides a comprehensive foundation for advancing research in aneurysm modeling by integrating rich geometric morphology with high-resolution hemodynamic information.

Improvements for AI systems

Based on the scientific paper Aneumo: A Large-Scale Multimodal Aneurysm Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks, here are specific, actionable improvements for AI systems derived from this work:

  1. Incorporate a high-fidelity, multimodal data pipeline into medical AI frameworks for aneurysm risk assessment.

  2. Develop and deploy a DeepONet-SwinT architecture specifically optimized for the prediction of complex hemodynamic parameters (pressure and velocity fields) within 3D aneurysm geometries.

  3. Enhance the generalization capability of fluid dynamics deep learning models by systematically training them on diverse flow conditions and geometric variations derived from the Aneumo dataset.

  4. Create a standardized benchmark for assessing SciML methods in cardiovascular fluid dynamics by utilizing the Aneumo dataset to compare established architectures (DeepONet) against novel, geometrically-enhanced architectures (DeepONet-SwinT).

Specific capabilities of these improved AI systems:

  1. The system can perform virtual hemodynamic simulations for any given aneurysm geometry and flow condition in near real-time by using the trained DeepONet-SwinT model to predict pressure and velocity fields.

  2. The system can accurately segment vascular regions from medical images (using the provided segmentation masks) and simultaneously predict the internal blood flow dynamics within those segmented structures.

  3. The system can provide quantitative risk scores for aneurysm rupture by predicting key hemodynamic metrics (like pressure drop or maximum velocity) that are strongly correlated with rupture risk, moving beyond simple morphological assessment.

  4. The system can be deployed in a generative design environment to rapidly evaluate the hemodynamic performance of synthetic aneurysm shapes generated via controlled deformation, accelerating the discovery of optimal surgical targets or intervention strategies.

  5. The system can be robust across different clinical scenarios (e.g., varying mass flow rates), as validated by its training on diverse flow conditions, allowing for reliable prediction even when encountering novel physiological states.

Abstract

Scientific machine learning uses simulation data to train surrogate models for fast physical-field prediction across geometries. Local shape editing can expand limited geometry collections, but whether its variants improve prediction on unseen geometries, and how to allocate them across sources, require controlled evaluation. We introduce AneumoBench, a dataset and benchmark linking 401 source aneurysm geometries to 9,693 locally edited descendant records, with computational fluid dynamics (CFD) fields computed on both. It contains 80,752 steady velocity-pressure cases across eight inlet conditions and 9,715 transient sequences of velocity, pressure, and wall shear stress (WSS). Each sequence contains 100 frames sampled at 0.01-s intervals from a 1-s cardiac cycle. Mesh, point, and voxel interfaces support steady field prediction and WSS forecasting from four observed frames. With family-disjoint splits, we compare source-only training, descendant training, and descendant pretraining followed by source fine-tuning across nine architectures on 79 held-out sources. Under the reported schedules, two-stage training lowers steady-field and reset-window WSS errors relative to source-only training. With the number of sampled fields and training updates fixed within each comparison, GraphSAGE benefits from descendant training and from distributing a fixed number of descendants across more sources. For WSS, reset-window gains do not consistently persist through 96-step rollout, and lower trajectory error need not improve cycle-level shear metrics or hotspot localization. These data and protocols enable researchers to compare descendant selection and training strategies on the same unseen source geometries.

Related papers