A chest CT dataset with
Chinese findings and conclusions

368 COVID-19 chest CT studies and 3,680 axial slices, each paired with the original free-text radiology report written in Chinese by the reporting radiologist.

Guangyi Liu, Yinghong Liao, Fuyu Wang, Bin Zhang, Lu Zhang, Xiaodan Liang, Xiang Wan, Shaolin Li, Zhen Li, Shuixing Zhang, Shuguang Cui
SRIBD · FNii · CUHKSZ-JD Joint AI Lab · CUHK-Shenzhen

368
CT studies
3,680
axial slices
10 per study
~96
patients
repeat scans
368
Chinese reports
free text, untranslated
2
hospitals
Guangzhou · Zhuhai

What is in it

Each row is one CT examination: ten axial lung-window slices, plus the radiologist's findings and conclusion exactly as written, plus sex, age and the RT-PCR result. The reports are not translated and not reduced to labels — they are the raw clinical text, with incidental findings, negative statements about unrelated organs, hedging, and comparisons against the patient's previous scan.

The data was collected between 18 January and 8 March 2020 at the First Affiliated Hospital of Jinan University (Guangzhou) and the Fifth Affiliated Hospital of Sun Yat-sen University (Zhuhai), and was released with our TNNLS paper on Medical-VLBERT, a model that generates these reports from the images.

One dataset row: ten axial CT slices below a table giving the case ID, the Chinese findings and conclusion text, gender, age and RT-PCR result.
One row of the dataset — ten axial slices of a single examination with the radiologist's findings and conclusion.

What the data looks like

Both charts below count studies (examinations), not patients: because patients were re-scanned every few days, the 368 studies come from roughly 96 people.

Age and sex

Studies per ten-year band

Female Male
Show as table

Examinations over time

Studies per week, weeks beginning Monday, 18 Jan – 8 Mar 2020

Show as table

⚠ Do not split this dataset randomly

77% of the reports compare against an earlier scan of the same patient (对比2020-02-08CT检查, 较前吸收), and consecutive studies of one patient have near-duplicate reports. A random study-level split puts a patient on both sides of the train/test boundary and inflates report-generation scores.

Per-study patient identity was not preserved with the data and is no longer recoverable. The release therefore ships patient_block — 67 (sex, age) cells that are guaranteed to contain every study of a given patient — and a recommended_split built from whole blocks, so no patient can straddle the boundary. Use it, or group by patient_block yourself.

Recommended split

SplitStudiesBlocksFemaleMaleAge rangeMean age
train296431601363–7546.8
validation3512152016–7149.8
test3712211620–8051.7

This split is for leakage-free benchmarking and is not the split used in the TNNLS paper, so results are not directly comparable to the published numbers. The paper also used 3 slices per study; this release provides all 10.

Getting the data

The dataset is hosted on the Hugging Face Hub. Access is gated: the dataset card is public, the files are released once you accept the data use terms. This replaces the old arrangement of emailing a signed confidentiality agreement.

  1. Open the dataset at huggingface.co/datasets/guangyil/COVID-19-CT-Chinese and sign in.
  2. Request access. Fill in your name, institutional email, affiliation and intended use, and accept the terms: non-commercial research only, no re-identification, no redistribution, no clinical use.
  3. Load it once the request is approved.
from datasets import load_dataset

ds = load_dataset("guangyil/COVID-19-CT-Chinese", split="train")

ex = ds[0]
print(ex["case_uid"], ex["sex"], ex["age"], ex["rt_pcr"])
print(ex["findings_zh"])      # 检查所见
print(ex["conclusion_zh"])    # 诊断意见
ex["images"][0].show()        # 10 axial slices, PIL images

# use the leakage-free split — do not split randomly
train = ds.filter(lambda x: x["recommended_split"] == "train")
test  = ds.filter(lambda x: x["recommended_split"] == "test")

Terms, in short

Released under CC BY-NC-SA 4.0 — attribution, non-commercial, share-alike. On top of the licence you agree not to re-identify anyone, not to redistribute the data (point people here instead), and not to use the data or any model trained on it for clinical decision-making. The full terms are on the dataset card.

What is not included

The Chinese medical encyclopedia corpus used for the knowledge-pretraining stage of Medical-VLBERT is not redistributed: it was scraped from third-party medical websites and is not ours to relicense. It is not needed to use this dataset. The chest X-ray dataset CX-CHR, also used in the paper, is a third-party dataset and is likewise not distributed here.

Citation

If you use this dataset, please cite:

@article{liu2021medicalvlbert,
  title   = {Medical-VLBERT: Medical Visual Language BERT for COVID-19 CT
             Report Generation With Alternate Learning},
  author  = {Liu, Guangyi and Liao, Yinghong and Wang, Fuyu and Zhang, Bin and
             Zhang, Lu and Liang, Xiaodan and Wan, Xiang and Li, Shaolin and
             Li, Zhen and Zhang, Shuixing and Cui, Shuguang},
  journal = {IEEE Transactions on Neural Networks and Learning Systems},
  volume  = {32},
  number  = {9},
  pages   = {3786--3797},
  year    = {2021},
  doi     = {10.1109/TNNLS.2021.3099165}
}

IEEE Xplore · arXiv:2108.05067 · Hugging Face