IAVS Dataset: Intracranial Aneurysm Segmentation

• The IAVS dataset is a curated collection of clinical and synthetic imaging data specifically designed for intracranial aneurysm and vessel segmentation.
• It features diverse modalities and precise annotation protocols that enable cross-modality benchmarking using metrics like Dice, IoU, and HD95.
• The dataset integrates CFD-ready workflows with automated preprocessing and mesh generation to bridge segmentation outputs with hemodynamic simulations.

The Intracranial Aneurysm Vessel Segmentation (IAVS) dataset encompasses a family of curated data resources central to algorithmic development and benchmarking in the segmentation of aneurysms and parent vessels from medical images. These datasets arise from both multi-center clinical imaging studies and advanced synthetic simulations, supporting applications ranging from deep learning–based segmentation to downstream computational fluid dynamics (CFD) analyses. The following sections describe the composition, annotation protocols, benchmarks, methodological frameworks, and CFD-centric evaluation paradigms across major IAVS datasets, notably including the recent multi-center initiative, synthetic datasets, and mesh-based segmentation collections.

1. Dataset Composition and Modalities

IAVS datasets provide comprehensive coverage of the anatomical and imaging variability required for intracranial aneurysm research across both real-patient and synthetic domains:

• Multi-center Clinical Dataset: The IAVS dataset presented by AbsoluteResonance consists of 641 time-of-flight (TOF) MRA volumes with 587 cases annotated for both aneurysm (IA) and parent-vessel masks. Sourcing integrates ADAM (113 volumes), INSTED (191), the Royal dataset (63), and 278 in-house cases. Image spacing spans min (0.21, 0.21, 0.30) mm to max (0.47, 0.47, 1.20) mm, with volume size ranging from (348×384×44) to (1024×1024×368) (Xiao et al., 1 Dec 2025).
• Synthetic Data Augmentation: The synthetic IAVS dataset (VaMos) leverages spline-based modeling of the Circle of Willis, controlled perturbations, and elastic deformation to generate 998 patches (64³ voxels; ~0.5 mm spacing) paired with 190 real TOF-MRA cases. Each synthetic sample includes vessels and aneurysm masks in one-hot format (Nader et al., 27 Mar 2024).
• Surface- and Point-Based Datasets: The IntrA collection features 103 TOF-MRA–derived surface models, from which 1,909 vessel "segments" are extracted (1,694 healthy, 215 aneurysmal). Detailed mesh annotations exist for 116 aneurysm-bearing segments, with point clouds (512–2,048 points/segment), triangular meshes (.PLY/.OBJ), voxelizations ({24³–40³}), and geodesic matrices (Yang et al., 2020).
• CFD-Ready Segmentation Subsets: Aneumo’s IAVS subset provides 10,660 three-dimensional binary mask volumes (NIfTI, [0.15 mm]³ spacing), derived from both 427 real aneurysm geometries and >10,000 synthetic deformations, matching the geometric/mesh context used for CFD (Li et al., 19 May 2025).

The diversity of imaging representations (volumetric, mesh, point cloud, mask) enables cross-modality algorithm development and benchmarking.

2. Annotation Protocols and Ground Truth Definition

Annotation workflows are dataset-specific, balancing manual expert intervention, semi-automated methods, and synthetic ground truth:

• Clinical Mask Annotation: In the multi-center IAVS dataset, IA masks were sourced from existing public datasets or annotated by two board-certified radiologists (with consensus) for in-house cases. Parent-vessel masks were pre-segmented via a COSTA vessel model (test Dice = 0.9204) and manually refined in 3D Slicer, with adaptive branch truncation to preserve physiological realism (Xiao et al., 1 Dec 2025).
• Quality Assurance for CFD: Each segmented mask undergoes geometric post-processing (median filtering, hole fill, removal of artifacts) and conversion to STL, meshing, centerline extraction, and an end-to-end CFD simulation to ensure solution convergence—the failure of which triggers manual correction (Xiao et al., 1 Dec 2025).
• Mesh-based Labeling: IntrA’s mesh annotation uses a custom mesh-based annotator, in which annotators mark sac boundaries and perform region growing from an interior seed. Labels are per-vertex binary (aneurysm sac vs. vessel), with the sac label including the neck region (Yang et al., 2020).
• Synthetic Ground Truth: The VaMos synthetic pipeline explicitly encodes aneurysm and parent vessel geometry via procedural, noise-driven deformations, yielding “perfect” binary masks for training and benchmarking, with vessel radius and intensity sampled from empirical data (Nader et al., 27 Mar 2024). Annotation in Aneumo is performed by automated mesh voxelization with pipeline-level expert review (Li et al., 19 May 2025).

Table 1 summarizes annotation characteristics:

Dataset	Source	Annotation Mode	Label Structure
IAVS (Xiao et al., 1 Dec 2025)	TOF-MRA	Manual/Model-assisted + QA	Aneurysm + Parent Vessel
IntrA (Yang et al., 2020)	Surface Mesh	Mesh-based, per-vertex, expert	Sac vs. Parent Vessel (mesh)
VaMos (Nader et al., 27 Mar 2024)	Synthetic/Real	Procedural (synthetic)/manual (real)	Vessel, Aneurysm
Aneumo (Li et al., 19 May 2025)	CFD Mesh/STL	Automated (voxelization, end-to-end QA)	Vessel (binary)

3. Data Splits, Preprocessing, and Augmentation Strategies

Best practices for reproducibility and benchmarking emphasize careful data splitting, augmentation, and normalization:

• Recommended Data Splits:
  • The multi-center IAVS dataset employs splits at the geometry (case) level, separating train/val/test to prevent leakage of correlated features (e.g., synthetic deformations of the same geometry in multiple sets). Typical distributions: 80% train, 10% validation, 10% test (Li et al., 19 May 2025); four- or five-fold cross-validation in IntrA and VaMos (Yang et al., 2020, Nader et al., 27 Mar 2024).
• Preprocessing Steps:
  • For volumetric data: normalization to zero-mean/unit-variance per patch, optional resampling to user-defined spacing, morphological smoothing (median filters), and artifact removal (Nader et al., 27 Mar 2024, Xiao et al., 1 Dec 2025).
  • For mesh data: Gaussian smoothing, hole-filling, non-manifold removal, retriangulation, per-vertex normal computation, geodesic distance precomputation (Yang et al., 2020).
• Augmentation:
  • Synthetic datasets incorporate geometric, intensity, and anatomical variability at the data-generation layer (random control-point perturbation, elastic fields, noise simulation) (Nader et al., 27 Mar 2024).
  • On-the-fly: spatial jitter, rotations, flips, and scaling. Additional augmentations (contrast/gamma) are avoided to preserve modality realism (Yang et al., 2020, Nader et al., 27 Mar 2024).

4. Benchmarking Tasks, Metrics, and Baseline Architectures

IAVS datasets underpin a range of algorithmic evaluations, with segmentation, detection, and classification tasks, standardized metrics, and comprehensive baselines:

• Tasks:
  • Global aneurysm localization (center detection), IA+vessel joint segmentation, binary vessel segmentation for CFD, and mesh-based part segmentation (Xiao et al., 1 Dec 2025, Yang et al., 2020, Nader et al., 27 Mar 2024, Li et al., 19 May 2025).
• Metrics:
  • Dice coefficient: $2|P\cap G|/(|P|+|G|)$
  • Intersection-over-Union (IoU): $|P\cap G|/|P\cup G|$
  • Hausdorff distance (HD95): 95th-percentile symmetric surface distance
  • clDice (centerline Dice), BIoU (boundary IoU), and, for detection, precision, recall, accuracy, and F1 (Xiao et al., 1 Dec 2025)
• Baselines:
  • IAVS: Two-stage pipeline—Stage I: heatmap-based IA localizer; Stage II: nnUNet+clDice for topology-aware segmentation (Xiao et al., 1 Dec 2025). Example results—Set A: Dice=0.8563, HD95=3.28 mm, clDice=0.8629.
  • VaMos: 3D U-Net with Dice+BCE loss; mean Dice=0.7613 ± 0.12 (real+synthetic), lesion-level sensitivity improves substantially for small aneurysms with synthetic augmentation (Nader et al., 27 Mar 2024).
  • IntrA: SO-Net, PointNet++, PointConv, SpiderCNN (point-based); MeshCNN (mesh-based); SSCN-F/U (voxel-based). SO-Net achieves best aneurysm IoU (81.4%), PointConv best vessel IoU (94.65%) (Yang et al., 2020).
  • Aneumo: No baseline reported, but even a trivial 3D U-Net achieves Dice > 0.95 on synthetic masks (Li et al., 19 May 2025).

Task	Metric(s)	Baseline (if reported)	Reference
Aneurysm Detection	F1, PR, RE, ACC	0.7632–0.8507 (F1)	(Xiao et al., 1 Dec 2025)
IA-Vessel Segment.	Dice, HD95, clDice	Dice 0.8368–0.8563 (Stage II)	(Xiao et al., 1 Dec 2025)
Mesh Segmentation	IoU, Dice	SO-Net IoU 81.4% (sac)	(Yang et al., 2020)
Volume Segmentation	Dice, IoU, H95	>0.95 Dice (Aneumo, U-Net)	(Li et al., 19 May 2025)

5. Workflow Integration for Computational Fluid Dynamics

A primary distinguishing feature of recent IAVS datasets is the explicit linkage to CFD, enabling direct evaluation of segmentation outputs in hemodynamics pipelines:

• Hemodynamic Outputs: For each annotated mask, IAVS (Xiao et al., 1 Dec 2025) supplies velocity $u(x)$, pressure $p(x)$, and wall shear stress $\tau_w(x)$ (obtained with OpenFOAM icoFoam/PISO, μ: viscosity, grid spacing 0.15 mm). CFD solutions rely on incompressible, Newtonian assumptions and grid-independence validation.
• Applicability Evaluation Pipeline (IAVS): Automated steps include Betti number topology check, STL conversion, centerline extraction (VMTK), cross-section definition, mesh generation in Fluent Meshing, surface fitting, boundary labeling in ANSYS SpaceClaim, and flow computation. CFD applicability (AS_CFD) is assessed via:
  • VTA: topological validity
  • MGA: mesh generation success
  • BFA: CFD convergence
  • The CFD-Applicability Score is computed as $AS_{CFD} = \widehat{TP} / (TP + FP + FN)$, where $\widehat{TP}$ includes only predictions passing all CFD checks. Reported scores: 57.45% (Set A), 54.76% (Set B) (Xiao et al., 1 Dec 2025).

6. Access, Licensing, and Usage Guidelines

• Public Release: IAVS (Xiao et al., 1 Dec 2025) and IntrA (Yang et al., 2020) are accessible via GitHub under permissive licenses. VaMos synthetic generator and data is available by request under CC BY-NC terms (Nader et al., 27 Mar 2024). Aneumo is downloadable at its public repository (Li et al., 19 May 2025).
• Usage Protocols: Consistent recommendations include reporting metrics on held-out sets (no patient overlap across train/val/test), using standardized splits and augmentations, and benchmarking against published baselines to facilitate direct comparison across studies (Yang et al., 2020, Xiao et al., 1 Dec 2025, Nader et al., 27 Mar 2024, Li et al., 19 May 2025).

7. Significance and Current Directions

IAVS datasets collectively form a critical platform for algorithm development in neurovascular image analysis:

• They address limitations of legacy datasets by supporting joint segmentation of aneurysms and parent vessels with ground truth suitable for both machine learning and CFD.
• They provide synthetic data to combat class imbalance, particularly for small aneurysms where training data are scarce and detection rates are limited (Nader et al., 27 Mar 2024).
• Integrated end-to-end CFD validation introduces a new paradigm in which segmentation model assessment is rigorously tied to downstream clinical and physiological applications, with proposed metrics (e.g., AS_CFD) that better reflect clinical utility compared to conventional image-space metrics (Xiao et al., 1 Dec 2025).
• A plausible implication is that further advances may arise from integrated pipelines exploiting these multi-modal, multi-source benchmarks and their CFD-aware evaluation systems.

Key published resources include IAVS (AbsoluteResonance, (Xiao et al., 1 Dec 2025)), IntrA (Yang et al., 2020), VaMos synthetic data (Nader et al., 27 Mar 2024), and the Aneumo CFD-multimodal benchmark (Li et al., 19 May 2025).