Volume-guided diffusion
A conditional diffusion formulation uses volumetric context to generate the complete aortic centerline tree.
IEEE VIS 2025
AortaDiff: Volume-Guided Conditional Diffusion Models for Multi-Branch Aortic Surface Generation
An end-to-end path from volumetric medical images to analysis-ready, multi-branch aortic surface geometry.
Illustrative transition · stylized volume, subject005 V2 surface mesh
01 · What’s next
AortaDiff+
AortaDiff+ is under active development.
02 · Overview
From an image volume to a simulation-ready surface.
AortaDiff conditions a diffusion model on CT or MRI volume information to generate multi-branch aortic centerlines. The resulting contours are reconstructed as a smooth surface suitable for downstream analysis.
Why it matters
High-quality vascular geometry is a prerequisite for computational fluid dynamics, yet manual and semi-automatic mesh preparation can be time-consuming and sensitive to image quality.
Data provenance
All subject cases and interactive meshes displayed on this homepage are derived from the AortaSeg24 dataset: Multi-Class Segmentation of Aortic Branches and Zones in CTA.
03 · Contributions
Volume-aware generation, built for vascular geometry.
Multi-branch reconstruction
Generated centerlines guide contour extraction before a smooth NURBS surface is reconstructed across the aorta and its branches.
CFD-oriented evaluation
Surface quality and OpenFOAM simulations are compared with reference geometry using velocity, pressure, and wall shear stress.
04 · Interactive comparison
One difficult junction.
Two generations.
Subject005 exposes a V1 failure case near the junction of the right subclavian artery (RSA) and right common carotid artery (RCCA). V2 (AortaDiff+) no longer shows this unintended merge in this case.
V1 · failure case
RSA / RCCA regionsubject005
The warm red vessels and ROI marker call attention to the unintended V1 connection.
V2 · updated result
RSA / RCCA regionsubject005
The corrected region is highlighted in teal. Camera movement and zoom are synchronized across both viewers.
Drag to rotate · Scroll to zoom · Press R to reset
05 · AortaDiff+
Four anatomies.
Every angle.
Explore V2 (AortaDiff+) meshes directly in the browser. Subjects are identified only by anonymized IDs; no clinical category is inferred here. Data source: AortaSeg24.
V2 mesh
subject005
V2 mesh
subject033
V2 mesh
subject057
V2 mesh
subject073
06 · Method
An end-to-end geometry generation pipeline.
07 · Results & CFD
Geometry evaluated all the way to hemodynamics.
The paper evaluates generated meshes against reference surfaces and carries them into OpenFOAM simulations, comparing velocity, pressure, and wall shear stress.
08 · Video
See the complete AortaDiff story.
A paper presentation covering the problem, the volume-guided diffusion pipeline, mesh evaluation, and downstream CFD analysis.
04:57 · presentation video09 · Open collaboration
Help us evaluate the next 1,000+ aortic meshes.
We plan to generate and evaluate 1,000+ aortic meshes and welcome collaborations with vascular imaging, cardiovascular mechanics, CFD, and clinical research teams.
da568@cornell.edu10 · Citation
Build on AortaDiff.
@article{an2026aortadiff,
title = {AortaDiff: Volume-Guided Conditional Diffusion Models for Multi-Branch Aortic Surface Generation},
author = {An, Delin and Du, Pan and Wang, Jian-Xun and Wang, Chaoli},
journal = {IEEE Transactions on Visualization and Computer Graphics},
volume = {32},
number = {1},
pages = {922--932},
year = {2026},
doi = {10.1109/TVCG.2025.3634652}
}