1000lungs

Generative Worlds for Adaptive Training of Bronchoscopy SLAM Models

1000 Lungs is a synthetic virtual bronchoscopy dataset built using SynAirG, a customisable model that generates synthetic worlds for bronchoscopy and general endoluminal interventions.

1000 Lungs models are built from MAISI-generated chest CT volumes. Each accepted case keeps the generated volume, paired labels, extracted airway mask, airway mesh, centerline, procedure-style camera path, rendered condition streams, and episode metadata under one case ID.

• Episodes: 1000
• Task: navigate a flexible bronchoscope through synthetic human airways
• Modalities: depth video, normal video, PPS video, composite metadata review MP4, CT NIfTI, mesh geometry, source scope paths, TUM pose trajectory.
• RGB status: photorealistic RGB may be supplied by a later second sweep; geometry-first batches use an explicit unavailable placeholder in the composite review.
• Source: MAISI rflow-ct chest CT plus MAISI-generated paired labels.

MAISI Instantiation

This dataset uses the SynAirG bulk MAISI path:

• Infer template: maisi_thoracic_airway_infer.json.
• Region: chest.
• Anatomy request: trachea, airway, left lung upper lobe, left lung lower lobe, right lung upper lobe, right lung middle lobe, right lung lower lobe, heart.
• Default output size: 512x512x256 voxels.
• Default spacing: 0.75x0.75x1.22656 mm.
• Source mask mode: none; the airway mask is extracted after generation from the paired MAISI label map.
• Airway labels: 57x132, resolved through MAISI output-label mapping when present.

Build Process

For each accepted case, SynAirG:

1. Runs the MAISI-compatible rflow-ct wrapper to emit a chest CT and paired label map.
2. Extracts a binary airway mask from the label map.
3. Builds a bronchoscopy-ready airway mesh, centerline graph, branch metadata, and mesh-to-CT registration sidecar.
4. Samples a procedure-style virtual bronchoscope path.
5. Renders synchronized episode videos and condition streams.
6. Writes LeRobot/Open-H episode parquet, MP4 features, per-case medical metadata, TUM trajectory text, and pose JSONL.

Conditioning for PBR-Based Inference

The exported depth, normal, PPS, and mask streams are aligned frame-for-frame with the bronchoscope pose. They can be used as ControlNet-style conditioning for PBR-based inference systems, including BronchoGen-style pipelines or custom SynAirG renderers. The mesh and material-map artifacts can also drive a geometry-preserving PBR pass before or alongside image-space generation.

Synthetic Use Notice

Synthetic only. Not for clinical interpretation, clinical deployment, autonomous diagnosis, or regulatory submission.

Credits

• Generator: MAISI / MAISI-v2 rflow-ct provides the synthetic chest CT and paired label volumes.
• Format convention: the episode tree follows the Open-H-Embodiment / LeRobot v2.1 layout convention for paired video and kinematic data.

Citation

@dataset{voncsefalvay_1000_lungs_2026,
  author       = {von Csefalvay, Chris},
  title        = {1000lungs: A Dataset of Generative Worlds for Adaptive Training of Bronchoscopy SLAM Models},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/chrisvoncsefalvay/1000lungs}},
  doi          = {10.57967/hf/9209}
}