Model 3D nodes
Represent the spatial layout of vessels and airways as node coordinates and sparse graph connectivity.
MICCAI 2025
Semantically Consistent Discrete Diffusion
Semantically Consistent Discrete Diffusion for 3D Biological Graph Modeling
Figure preview is unavailable.View the original figure ↗
Overview of the graph diffusion framework, from the official code repository. · Original figure ↗
Overview
Why this work matters
Biological networks such as cerebral vessels and lung airways are sparse, spatially embedded graphs with strong structural constraints. This work introduces a semantically consistent graph diffusion method that uses edge-deletion noising and a stochastic projection operator to enforce plausibility during sampling.
Approach
How the method works
Diffuse sparse edges
Use an edge-deletion-based noising process suited to biological networks, followed by discrete edge generation.
Project to valid graphs
Apply a stochastic projection operator during sampling to correct semantic or structural inconsistencies.
Research directions
Applications and evaluation
Circle of Willis
Generate graphs of major cerebral vessels with anatomically meaningful connectivity.
Lung airways
Model realistic branching airway trees under relevant structural constraints.
Downstream utility
Evaluate graph labeling and link prediction in addition to graph quality.
Results
Visual comparison
Figure preview is unavailable.View the original figure ↗
Representative qualitative comparison from the official repository. · Original image ↗
Resources
Paper and code
The official repository provides model code, sampling configurations, evaluation scripts, and setup guidance for the Circle of Willis and airway experiments.
Reference
Citation
@inproceedings{prabhakar2025semantically,
title = {Semantically Consistent Discrete Diffusion for 3D Biological Graph Modeling},
author = {Prabhakar, Chinmay and Shit, Suprosanna and Amiranashvili, Tamaz and Li, Hongwei Bran and Menze, Bjoern},
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
pages = {594--604},
year = {2025},
publisher = {Springer},
doi = {10.1007/978-3-032-05162-2_57}
}