MedPhyGraph/support-graph-data

MedPhyGraph — Procedural Support-Graph Data

Structured procedural training and evaluation data for:

MedPhyGraph: Counterfactual Support-Graph Maintenance for Dynamic Built-Environment Digital Twins
TwinWorld: Visual Intelligence for Built Environment Digital Twins, ECCV 2026 Workshop

This repository contains the procedural data released with MedPhyGraph, including the candidate-edge data used with CF-SupportNet and the structured procedural scenes used in the public evaluation workflows.

MedPhyGraph is a method contribution, not a dataset or benchmark contribution. This repository is provided to make the released experiments easier to inspect, reuse, and reproduce.

Resource	Link
Dataset	You are here
Model	MedPhyGraph/CF-SupportNet
Project	medphygraph.github.io

At a glance

Item	Public release
Candidate-edge samples	533
Train / validation / test	324 / 101 / 108
Procedural scene states	54
Procedural transfer cases	136
Full expanded suite reported in the paper	217 cases
Isaac for Healthcare cases included here	0
License	CC BY 4.0

The paper evaluates 217 expanded transfer cases in total:

• 136 Procedural
• 81 Isaac for Healthcare

This repository contains the 136 Procedural cases only.
The 81 Isaac for Healthcare cases are not redistributed here.

---

Download

The easiest way to download the complete public release is with the Hugging Face CLI:

hf download MedPhyGraph/support-graph-data \
  --repo-type dataset \
  --local-dir ./support-graph-data

Or with Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MedPhyGraph/support-graph-data",
    repo_type="dataset",
    local_dir="./support-graph-data",
)

Both methods download the complete public release while preserving its directory structure.

Individual artifacts can also be inspected from the Files and versions tab on Hugging Face.

Dataset Viewer

The Hugging Face Dataset Viewer is intentionally disabled. This release preserves the original scientific JSON artifacts, and different directories contain different JSON schemas. We keep those artifacts in their frozen form rather than reshaping them only for a tabular viewer.

---

Contents

Released directory	Contains	Expected MedPhyGraph path
training_data/	Frozen CF-SupportNet candidate-edge dataset and split	outputs/dyphygraph_health/dataset_hard/
procedural_scenes/	Structured Procedural scene states and transition metadata	outputs/dyphygraph_health/scenes_twinworld_phase2/
expanded_transfer_procedural_v1/	136 Procedural transfer cases, targets, and before/after states	outputs/reviewer_response/expanded_transfer_v1/

Note: the release directory is named expanded_transfer_procedural_v1/, while the public evaluation scripts expect its contents under outputs/reviewer_response/expanded_transfer_v1/.

training_data/

Contains:

• dataset.json
• split.json

The frozen dataset contains 533 candidate-edge samples:

• Train: 324
• Validation: 101
• Test: 108

The samples come from procedurally authored scenes and use the analytic AABB host-removal counterfactual evidence protocol used by CF-SupportNet.

The analytic backend is a geometric proxy and should not be interpreted as a full physics-engine simulation.

No Isaac-derived samples are included.

procedural_scenes/

Contains the structured Procedural evaluation states released with MedPhyGraph.

The public release includes:

• 54 scene.json states
• scene indexes
• transition metadata
• procedural AABB proxy geometry

These states are used by the public procedural evaluation workflow and can also be used to regenerate intermediate procedural data required by some evaluation scripts.

expanded_transfer_procedural_v1/

Contains the public Procedural portion of the expanded transfer evaluation:

• 136 transfer cases
• 272 before/after scene.json files
• predeclared_manifest.json
• targets.json
• scenes/procedural/

This directory is not the complete 217-case suite reported in the paper. It contains the 136 Procedural cases only.

For use with the public evaluation scripts, place its contents under:

outputs/reviewer_response/expanded_transfer_v1/

so that the resulting layout contains:

outputs/reviewer_response/expanded_transfer_v1/
├── predeclared_manifest.json
├── targets.json
└── scenes/
    └── procedural/

---

What is not included

This repository does not redistribute:

• NVIDIA Isaac for Healthcare assets
• NVIDIA Isaac Sim USD / USDA / USDC files
• NVIDIA meshes or textures
• Isaac-derived structured scene states
• the 81 Isaac for Healthcare transfer cases reported in the paper

These resources are intentionally kept outside the public MedPhyGraph dataset release.

---

NVIDIA Isaac for Healthcare

The MedPhyGraph paper also evaluates the method in NVIDIA Isaac for Healthcare environments.

To work with the Isaac portion of MedPhyGraph, users should obtain NVIDIA Isaac Sim / Isaac for Healthcare and the required assets directly from NVIDIA under the applicable NVIDIA license terms.

Official NVIDIA resources

Resource	Official link
NVIDIA Isaac for Healthcare	developer.nvidia.com/isaac/healthcare
Getting Started	NVIDIA documentation
Setup Guide	NVIDIA setup instructions
Loading Assets	NVIDIA asset-loading guide
Isaac for Healthcare GitHub	github.com/isaac-for-healthcare
i4h Workflows	isaac-for-healthcare/i4h-workflows

MedPhyGraph does not repackage or redistribute NVIDIA-authored assets. Users obtain those resources directly from NVIDIA and work with them locally.

---

Reproducibility

The public release supports several procedural evaluation workflows, but it is important to distinguish between re-running the released evaluations and regenerating every paper artifact from scratch.

Procedural evaluation

The following public artifacts are available:

• the frozen 533-sample candidate-edge dataset and split
• the 54 structured Procedural scene states
• the 136-case Procedural expanded-transfer subset
• the paper-frozen CF-SupportNet checkpoints
• the corresponding public MedPhyGraph evaluation code

With these artifacts, users can re-run evaluation using the released candidate-edge dataset and paper-frozen CF-SupportNet checkpoints, and re-evaluate the public 136-case Procedural expanded-transfer subset.

Some core procedural paper tables additionally require an intermediate file:

outputs/dyphygraph_health/dataset_twinworld_phase2/dataset.json

That intermediate dataset is not included directly in this Hugging Face release.

The corresponding structured Phase-2 procedural scenes are included under procedural_scenes/, and the public MedPhyGraph code contains the data-generation components needed for this workflow.

For this reason, this repository should be understood as the frozen public data release supporting the Procedural evaluation workflows, rather than as a one-command reproduction package for every paper table.

Isaac for Healthcare evaluation

The Isaac portion additionally requires:

1. NVIDIA Isaac Sim / Isaac for Healthcare
2. the required NVIDIA assets obtained separately from NVIDIA
3. MedPhyGraph Isaac-facing scene-generation and evaluation components

The public MedPhyGraph code includes relevant components for working with locally installed Isaac environments. However, the complete paper-frozen Isaac scene-generation workflow is not currently packaged as a standalone end-to-end public pipeline.

The structured states used for the paper's 81 Isaac cases are also not redistributed in this repository.

Therefore, the exact Isaac results reported in the paper should not be expected to reproduce from the currently released public artifacts alone.

---

Relationship to CF-SupportNet

CF-SupportNet is the learned candidate-edge scorer used within MedPhyGraph.

The model repository contains five paper-frozen checkpoints:

• seed 0 — primary paper checkpoint
• seeds 1–4 — multi-seed reproducibility checkpoints

MedPhyGraph itself is broader than CF-SupportNet.

In addition to the learned scorer, MedPhyGraph applies deterministic State Consistency and Union-Based Transition-Aware Consistency to maintain support relations across adjacent digital-twin states.

These deterministic components are part of the MedPhyGraph inference pipeline and are not separate learned checkpoints.

---

Integrity

Two files are provided for release verification:

• MANIFEST.json — index and metadata for the released public artifacts
• SHA256SUMS.txt — per-artifact SHA256 hashes for integrity checking

These files can be used to verify the released scientific artifacts in a local copy.

---

License

The MedPhyGraph-owned content in this repository is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

NVIDIA Isaac for Healthcare assets are not included in this repository and are not being relicensed by MedPhyGraph. NVIDIA resources remain subject to NVIDIA's own license terms.

---

Citation

If you use MedPhyGraph or these released artifacts, please cite:

@inproceedings{gholizadeh2026medphygraph,
  title     = {MedPhyGraph: Counterfactual Support-Graph Maintenance for Dynamic Built-Environment Digital Twins},
  author    = {Gholizadeh HamlAbadi, Kamran and Vahdati, Monica and El Saddik, Abdulmotaleb},
  booktitle = {TwinWorld: Visual Intelligence for Built Environment Digital Twins, ECCV 2026 Workshop},
  year      = {2026}
}