ACVD Processed Lung Nodule CT Data

Processed data used for Anatomy-Constrained Voxel Diffusion for Controllable Synthesis of Complex Lung Nodules, accepted at IEEE BIBM 2026 (publication forthcoming).

Code: lakelk/ACVD.

Contents

Subset	CT series with each anatomical mask	Paired 64³ crops
LUNA16	888	1186
LUNA25	4069	6155

The release provides one ZIP per subset. Extract both ZIPs into the same data directory. Each archive preserves this structure:

Dataset_LUNA16/  # Dataset_LUNA25 uses the same structure
  annotations.csv
  airway_masks/*.nii.gz
  vessel_masks/*.nii.gz
  lung_masks/*.nii.gz
  bone_masks/*.nii.gz
  ControlNet_Data/*.npy

Full-volume anatomical masks

In addition to the paired 64³ crops, we release four anatomical channels covering the full CT grid: airway, pulmonary artery (named vessel), lung parenchyma and bone. These masks support research on CT morphology, anatomical structure and spatial relationships over larger regions or entire scans, beyond localized nodule synthesis.

Masks are binary NIfTI volumes and retain the corresponding CT grid and physical geometry, including anisotropic spacing. They are automatically generated predictions. Original CT images and original full-volume nodule masks must be obtained separately from the sources below; original CT images are not included in this release.

Sources

Component	Source
CT and coordinates	Official LUNA16/LUNA25 challenge data
LUNA16 nodule masks	LIDC-IDRI
LUNA25 nodule masks	WangLab LUNA25-MedSAM2
Lung and bone	TotalSegmentator predictions consolidated by the authors
Airway	Author-trained nnUNet on ATM22, applied to LUNA CT
Vessel	Author-trained nnUNet on PARSE22 (pulmonary artery segmentation), applied to LUNA CT

See SOURCES.md for links and citations and LICENSE.md for component conditions. The ATM22/PARSE22 original training data and segmentation model weights are not included.

Paired crop format

Each .npy contains a dictionary:

Key	Meaning
uid	CT series UID
nodule_idx	Zero-based row index in the included annotation table
gt_image	float32 (1,64,64,64), normalized to [-1,1]
conditions	Binary float32 (5,64,64,64)

Condition order: nodule, vessel, airway, lung, bone. Spatial array order: Z, Y, X.

import numpy as np
sample = np.load("path/to/sample.npy", allow_pickle=True).item()
ct = sample["gt_image"]
conditions = sample["conditions"]

CT values are clipped to [-1000,400] HU before normalization. The dictionaries do not include a physical affine. Preserve annotation-table row order when interpreting nodule_idx. Preprocessing and experimental settings are described in the code repository.

Use and download

After downloading the repository, extract:

python -m zipfile -e archives/Dataset_LUNA16.zip ./ACVD-data
python -m zipfile -e archives/Dataset_LUNA25.zip ./ACVD-data

For training, evaluation and raw-volume data requirements, see the code repository.

Citation

Until the proceedings are published, cite the paper as:

Anatomy-Constrained Voxel Diffusion for Controllable Synthesis of Complex Lung Nodules. Accepted at IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2026. Publication forthcoming.

The author list and final proceedings/DOI citation will be added when available. Please also cite the upstream datasets and annotation sources relevant to your use, as listed in SOURCES.md, which also contains copyable BibTeX entries.

License

The authors' generated four-channel anatomical masks are released under CC BY-NC 4.0: attribution is required and commercial use is not permitted. Third-party CT, coordinates and nodule masks retain their source-specific conditions; see LICENSE.md. ACVD original code is licensed separately under MIT.