CONFLUX Chest-CT
200,000 synthetic 3D chest CT volumes with structured abnormality and demographic labels, generated by CONFLUX.
Released with the paper CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training.
Paper (arXiv) • Model • Code — coming soon
About
CONFLUX is a conditional 3D latent generative model for chest CT: a VAE tokenizer compresses each volume into a compact 16-channel latent, a single-stream rectified-flow transformer generates in that latent space, and a reinforcement-learning stage sharpens label faithfulness. Every volume in this release is generated from a clinical profile — 18 abnormality findings, sex, age group, and reconstruction kernel — so the cohort is fully labeled and covers realistic clinical variation. See the paper for the model, training, and evaluation.
| Samples | 200,000 volumes |
| Modality / region | CT / chest (lung-focused) |
| Grid | 216 × 176 × 200 voxels, 1.5 mm isotropic |
| Format | int16 Hounsfield units (NIfTI) |
| Labels | 18 abnormality findings + sex + age group + reconstruction kernel |
| License | CC BY-NC-SA 4.0 (non-commercial research) |
Dataset structure
conflux-chest-ct/
├── metadata.csv one row per volume (schema below)
└── data/
├── 000/ sample_000000.nii.gz … sample_000999.nii.gz
├── 001/ …
└── 199/ … sample_199999.nii.gz (bucketed by index // 1000)
metadata.csv columns:
| column | description |
|---|---|
file_name | relative path to the volume (join key) |
sex | M / F |
age_group | decade band, e.g. 60-69, 70+ |
kernel | reconstruction-kernel label |
| 18 finding columns | binary presence of each abnormality |
Each .nii.gz is a 216 × 176 × 200 int16 array in Hounsfield units; voxel
geometry is stored in the NIfTI affine.
Usage
from huggingface_hub import hf_hub_download
import pandas as pd, nibabel as nib
REPO = "gevaertlab/conflux-chest-ct"
meta = pd.read_csv(hf_hub_download(REPO, "metadata.csv", repo_type="dataset"))
# e.g. all cardiomegaly cases
cardio = meta[meta["Cardiomegaly"] == 1]
path = hf_hub_download(REPO, cardio.iloc[0]["file_name"], repo_type="dataset")
vol = nib.load(path).get_fdata() # (216, 176, 200) int16 Hounsfield units
Intended use
Research use — pre-training, augmentation, and benchmarking for chest-CT models where real labeled data is scarce or access-restricted. Not for clinical use. The volumes are synthetic: generated by a model, they do not correspond to real patients. The cohort was screened for training-data memorization (no copies of training scans).
Citation
@article{vanpuyvelde2026conflux,
title = {CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training},
author = {Van Puyvelde, Max and Gulluk, Halil Ibrahim and Van Criekinge, Wim and Gevaert, Olivier},
journal = {arXiv preprint arXiv:2607.02998},
year = {2026}
}
Acknowledgements
CONFLUX was trained on CT-RATE (Hamamci et al., 2024).
License
CC BY-NC-SA 4.0 — non-commercial research use.