Technical Summary: SynVA – A Modular Toolkit for Vessel Generation and Aneurysm Editing

Problem Statement

Intracranial aneurysms (IAs) represent a significant cause of stroke and mortality, yet their unpredictable growth and rupture risk complicate early diagnosis and treatment planning. While digital twins and deep learning offer promising avenues for improving patient-specific prognosis and hemodynamic analysis (e.g., wall shear stress estimation), their efficacy is severely constrained by the scarcity of large-scale, high-quality, labeled 3D medical data. Existing real-world datasets are limited in size, and reconstructing accurate 3D meshes from medical imaging is error-prone and requires expert intervention. Furthermore, current synthetic data generation methods often lack anatomical consistency; they either focus solely on healthy vascular trees without pathology or generate aneurysms as isolated objects rather than as localized deformations of pre-existing vessel structures. There is a critical need for scalable approaches that can generate anatomically plausible, fully labeled 3D vascular meshes with integrated aneurysms to support the training of generative models and downstream vision tasks.

Methodology

The authors present Synthetic VAsculature (SynVA), a modular toolkit designed to generate fully labeled 3D vascular meshes, encompassing both healthy geometries and pathological aneurysm formations. SynVA integrates data-driven learning-based models with a model-driven procedural approach, unified by a specific physiological workflow: aneurysms are not generated in isolation but are synthesized as localized deformations conditioned on a predefined ostium (the intersection between the parent vessel and the aneurysm).

The toolkit comprises three primary components:

1. Healthy Vessel Generation (SynVA-V1)

This component utilizes a flow-matching-based generative model to create healthy vascular structures. It operates in a two-stage process:

• Topology Generation: A discrete vessel topology is learned as a rooted directed tree using a GPT-style autoregressive Transformer. This stage encodes branching structures without continuous geometry.
• Geometry Generation: Given a fixed topology, a TreeFlowNet (a topology-aware Graph Neural Network) predicts continuous node features. These features include parent-relative centerline positions, local cross-section control points, and B-spline knot parameters. The model is trained using optimal-transport conditional flow matching on real data (where aneurysmal regions were manually removed to create healthy ground truth).
• Refinement: The generated spline trees are converted into watertight surface meshes via screened Poisson surface reconstruction, with specific smoothing applied at bifurcations.

2. Ostium-Conditioned Aneurysm Generation

Once a healthy vessel is generated, an ostium is selected either automatically (based on bifurcation regions or local curvature priors) or manually via a GUI. Three distinct methods generate the aneurysm mesh conditioned on this ostium:

• SynVA-A1: An adaptation of the AneuG model. It uses a compact Graph Harmonic Deformation (GHD) latent representation. The model is conditioned on the local vessel geometry, explicit ostium ring geometry, and morphological parameters. It employs a conditional VAE architecture with specific loss terms to enforce geometric consistency at the vessel-aneurysm interface.
• SynVA-A2: An extension of MeshAnything V2. It reformulates mesh generation as an autoregressive task conditioned on partial anatomical context. The model takes a local vessel point cloud and an explicit ostium mesh (tokenized as a forced prefix) to generate the aneurysm mesh as a continuation of the ostium, rather than reconstructing a full shape from a complete point cloud.
• SynVA-P1 (Procedural): A fully procedural model based on physiological principles (e.g., Murray's law for vessel diameters) and statistical priors. It generates a bifurcating vessel, selects an ostium, and synthesizes an aneurysm sac from a geometric primitive (icosphere) that is stretched, perturbed with noise, and optionally augmented with a "bleb" (a localized bulge associated with rupture risk).

3. Post-Processing and Stitching

For all methods, the generated aneurysm mesh is stitched to the parent vessel at the ostium interface. This involves removing invalid elements, recomputing normals, aligning the meshes, and smoothing the transition region to produce a single, continuous pathological vessel mesh with semantic vertex labels (healthy vessel, aneurysm, ostium).

Key Contributions

The paper outlines four primary contributions:

1. SynVA-V1: A flow-matching-based model for healthy vessel generation that outperforms existing baselines (VesselGPT, VesselVAE) in both topology and geometry metrics.
2. Ostium-Conditioned Generators: Two methodologically distinct learning-based models (SynVA-A1 and SynVA-A2) for localized aneurysm synthesis on existing vessels, addressing the gap in modeling aneurysms as surface-consistent processes.
3. Procedural Model (SynVA-P1): A novel procedural approach for generating labeled vascular meshes and aneurysms based solely on biological and mathematical principles, enabling the creation of large-scale datasets without reliance on patient-specific inputs.
4. Large-Scale Synthetic Dataset: The release of a dataset comprising 50,000 fully labeled 3D mesh samples (vessels with aneurysms), including derived point clouds, sub-meshes, and morphological descriptors, intended for training mesh-based generative models and downstream tasks.

Results

The authors conducted extensive quantitative and qualitative evaluations:

• Healthy Vessel Generation: SynVA-V1 achieved superior or competitive performance against baselines. On a curated test set, it demonstrated lower Maximum Mean Discrepancy (MMD) and higher Coverage (COV) and 1-Nearest-Neighbor Accuracy (1-NNA) compared to VesselGPT. Topology metrics (KL divergence of node degrees and Laplacian spectrum distance) were also improved.
• Aneurysm Generation:
  • Quantitative: SynVA-A1 showed slightly better performance on quantitative similarity metrics (Chamfer distance and morphological parameter deviations) compared to SynVA-A2 on a shared subset of successful generations.
  • Qualitative (Expert Study): Two neuroradiologists rated generated samples against ground truth. While ground truth was rated highest, SynVA-A2 received slightly higher realism ratings than SynVA-A1, despite SynVA-A1's better quantitative scores. Experts noted that SynVA-A1 tended to produce overly symmetric shapes, whereas SynVA-A2 generated shapes with more clinically plausible irregularities, though it had a higher failure rate in mesh generation.
• Downstream Task (Semantic Segmentation): To validate the utility of the synthetic dataset, the authors trained a Point Transformer V3 (PTv3) model for aneurysm head identification.
  • Models trained only on synthetic data performed poorly on real test data (mIoU ~36.8%), highlighting the domain gap.
  • However, synthetic pretraining followed by fine-tuning on real data significantly outperformed models trained on real data alone, particularly in low-data regimes (e.g., 10% real data). Pretraining with 40,000 synthetic samples boosted the mIoU from 50.41% to 63.88% when fine-tuned on just 10% of real data.

Significance and Claims

The paper positions SynVA as a critical step toward overcoming data scarcity in cerebrovascular research. By decoupling healthy vessel generation from aneurysm synthesis and connecting them via an explicit ostium, SynVA models aneurysm formation as a localized, vessel-conditioned process, which aligns more closely with physiological reality than isolated shape generation.

The authors claim that SynVA enables:

• The generation of large-scale, diverse, and fully labeled datasets (50,000 samples) that are currently unavailable in the public domain.
• Improved training for downstream tasks, demonstrated by the significant performance gains in semantic segmentation when using synthetic pretraining.
• Controlled editing and analysis, allowing researchers to generate specific pathological variations for studying rupture risks or training diagnostic models.

The authors remain modest regarding the current limitations: the procedural model is limited to single bifurcations and does not yet capture complete patient-level Circle of Willis anatomy; the generated geometries are optimized for morphological plausibility but do not explicitly model hemodynamic quantities (e.g., wall shear stress), requiring further preparation for computational fluid dynamics (CFD) simulations. They emphasize that the synthetic data is intended for research, benchmarking, and pretraining, and should not be used as a standalone substitute for clinical decision-making without extensive validation on real-world data.