CVPR26-3DCTFMCompetition Dataset

This repository contains data prepared for the CVPR 2026 Workshop Challenge: Foundation Models for General CT Image Diagnosis.

Data origin

The datasets included here are from previously published works:

1. COVID-CT

   • Rahimzadeh, M., Attar, A., & Sakhaei, S. M. (2021). A fully automated deep learning-based network for detecting COVID-19 from a new and large lung CT scan dataset. Biomedical Signal Processing and Control, 102588. https://doi.org/10.1016/j.bspc.2021.102588

2. LUNA25

   • Peeters, D., Obreja, B., Antonissen, N., & Jacobs, C. (2025). Benchmarking of Artificial Intelligence and Radiologists for Lung Cancer Screening in CT: The LUNA25 Challenge. Medical Image Computing and Computer Assisted Intervention 2025 (MICCAI). Zenodo. https://zenodo.org/records/14223624; https://zenodo.org/records/14673658

3. AMOS

   • Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al. (2022). AMOS: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in Neural Information Processing Systems, 35, 36722–36732.

4. STOIC2021

   • Revel, M.-P., Boussouar, S., de Margerie-Mellon, C., Saab, I., Lapotre, T., Mompoint, D., et al. (2021). Study of Thoracic CT in COVID-19: The STOIC Project. Radiology, 301(1), E361–E370. https://doi.org/10.1148/radiol.2021210384
   • Challenge page: https://stoic2021.grand-challenge.org/

Modifications

The data released in this repository are from the original sources and were prepared for challenge use. This may include file format standardization, and train/validation split preparation.

This repository does not claim ownership of the original datasets. Users should refer to the original sources for full dataset descriptions, intended use, and any official updates.

STOIC2021

Derived from the public STOIC2021 training set (2000 CT scans). Preparation was limited to format standardization and split assignment; voxel values are unmodified Hounsfield Units.

• Format — the original .mha volumes were converted to .nii.gz. This was a pure container conversion: voxels remain int16 raw HU, and voxel spacing, origin, and orientation are carried over unchanged. No resampling, windowing, cropping, or intensity normalization was applied.
• Labels — labels/covid.csv (columns case_id,covid,split) holds the binary COVID-19 presence label, taken from probCOVID in the original metadata/reference.csv. The original probSevere severity label is not included in this release.
• Split — a stratified 60/20/20 train/val/test split over the covid label (seed 42). Class balance is preserved at 60.2% positive in every split; see subset_stats.csv.
• Held-out test set — only the train and val scans (1600 of 2000) are published here. The 400 test scans and their labels are withheld for evaluation.

split	positive	negative	total	positive %	published
train	723	477	1200	60.2	yes
val	241	159	400	60.2	yes
test	241	159	400	60.2	no (withheld)
all	1205	795	2000	60.2

The STOIC2021 data are licensed CC BY-NC 4.0; the full license text is included at stoic2021/LICENSE. Non-commercial use only, with attribution to the original authors.

License and attribution

The redistributed data remain subject to the license terms of their respective original sources. Users must preserve appropriate attribution to the original creators and sources.

Please consult the original dataset pages and publications for the exact licensing terms that apply to each component dataset before reuse or redistribution.