This dataset contains de-identified human chest CT images and the corresponding radiology reports, contributed by two Chinese hospitals during the early COVID-19 outbreak. Access is granted for non-commercial academic research only.
By requesting access you agree that you will: (1) use the data solely for non-commercial research or educational purposes; (2) make no attempt to re-identify any individual, or to link this data with any other dataset for the purpose of re-identification; (3) not redistribute the data, in whole or in part, to any third party — refer others to this repository instead; (4) not use the data, or any model trained on it, for clinical decision-making or any other patient-care purpose; (5) cite the paper below in any publication that uses this data; (6) comply with the CC BY-NC-SA 4.0 license.

COVID-19 CT Dataset — Chest CT with Chinese Findings and Conclusions

A chest CT dataset of 368 studies / 3,680 axial slices paired with the original Chinese radiology reports (free-text findings and conclusion) written by attending radiologists during the early COVID-19 outbreak in China.

This is the dataset released with the TNNLS paper Medical-VLBERT: Medical Visual Language BERT for COVID-19 CT Report Generation With Alternate Learning (arXiv:2108.05067). Project page: https://covid19ct.github.io/

What changed: the dataset was previously distributed by email on request. It has now been re-curated and is hosted here on the Hugging Face Hub, with a machine-readable schema, English field names alongside the original Chinese text, and a leakage-free recommended split. Access remains gated — see Access and intended use.

To the best of our knowledge this is one of the few publicly available chest CT datasets that pairs images with native, untranslated, free-text Chinese reports rather than with categorical labels, which makes it suitable for medical report generation, Chinese clinical NLP and vision-language grounding.

One row of the dataset: the 10 axial slices of a single examination, together with the radiologist's findings and conclusion, sex, age and RT-PCR result.

Quick start

from datasets import load_dataset

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

ex = ds[0]
print(ex["case_uid"], ex["sex"], ex["age"], ex["rt_pcr"])
print(ex["findings_zh"])
print(ex["conclusion_zh"])
ex["images"][0].show()          # 10 axial slices per study, PIL images

Use the leakage-free split shipped with the data — do not split randomly, see Longitudinal structure:

train = ds.filter(lambda x: x["recommended_split"] == "train")
val   = ds.filter(lambda x: x["recommended_split"] == "validation")
test  = ds.filter(lambda x: x["recommended_split"] == "test")

Dataset structure

One row = one CT study (one examination of one patient), holding all of its slices and its single report.

Field	Type	Description
`case_uid`	`string`	Original study identifier, `YYYYMMDD` + in-hospital sequence number.
`patient_block`	`string`	Grouping key for leakage-free splitting (`B001`…`B067`). A superset of patient identity, not a patient id — see below.
`exam_date`	`string`	Examination date, `YYYY-MM-DD` (derived from `case_uid`).
`sex`	`string`	`female` / `male`.
`age`	`int32`	Age in years at examination.
`rt_pcr`	`string`	RT-PCR result: `positive` / `negative`.
`findings_zh`	`string`	Findings section of the report, verbatim Chinese. Mean 147 characters.
`conclusion_zh`	`string`	Conclusion / impression section, verbatim Chinese. Mean 77 characters.
`num_slices`	`int32`	Number of slices (10 for every study).
`references_prior_exam`	`bool`	Whether the report compares against an earlier examination.
`prior_exam_dates`	`list[string]`	Prior examination dates explicitly cited in the report text, `YYYY-MM-DD`.
`recommended_split`	`string`	`train` / `validation` / `test` — block-disjoint, see Splits.
`images`	`list[Image]`	10 axial lung-window slices, JPEG, ordered by ascending slice index.

Images are 512×512 (2,850 slices) or 1024×1024 (830 slices), RGB-encoded grayscale, reconstructed in a lung window and cropped to the body field of view.

The reports are not translated and not structured into labels. They are the raw clinical text, including negative statements about unrelated organs, incidental findings, hedging ("考虑…可能性大"), and comparisons to prior scans.

Statistics


Studies (rows)	368
Slices	3,680 (exactly 10 per study)
Patients	~96 (see below — not recoverable per-study)
Patient blocks (published grouping key)	67
Sex	196 female / 172 male studies
RT-PCR	357 positive / 11 negative
Age range	3 – 80 years
Reports referencing a prior exam	285 / 368 (77%)
Exam dates	2020-01-18 – 2020-03-08

Source institutions: The First Affiliated Hospital of Jinan University, Guangzhou and The Fifth Affiliated Hospital of Sun Yat-sen University, Zhuhai.

Longitudinal structure (read this before splitting)

The 368 studies come from roughly 96 patients, not 368 patients. 77% of the reports explicitly compare against an earlier scan ("对比2020-02-08CT检查", "较前吸收"), because patients were re-scanned every few days to track resolution. Consecutive studies of one patient have near-duplicate reports.

Per-study patient identity is not recoverable. The original patient identifiers were not preserved alongside the released images and reports, and they no longer exist. The ~96 figure comes from the patient-level demographics reported when the dataset was originally described; it cannot be reconstructed from the released fields.

What can be guaranteed from the released fields is a superset of patient identity. A patient's sex and age are constant across the seven-week collection window, so every study of one patient necessarily falls inside a single (sex, age) cell. Those 67 cells are published as patient_block:

Studies per block	1	2	3	4	5	6	7	8	9	10	11	12	13	15
Blocks	7	7	11	11	6	2	2	6	2	6	1	3	2	1

patient_block over-merges: 67 blocks pool ~96 patients, so a block sometimes contains two or three different patients who happen to share a sex and an age. That is the safe direction — splitting on blocks cannot put one patient on both sides of the boundary. Do not use patient_block as a patient identifier, and do not count blocks as patients.

If you need finer structure than blocks, prior_exam_dates gives the dates each report cites. Resolving those citations recovers 123 unambiguous study-to-study follow-up links, but 59 citations are ambiguous (several studies share the cited date, sex and age), which is why it is not published as an identity column.

Splits (use these)

recommended_split assigns whole blocks, walking them in age order and always filling the split with the largest relative shortfall — so ages stay interleaved across splits instead of clustering. Verified: 0 blocks straddle a split boundary.

Split	Studies	Blocks	Female / Male	Age range	Mean age
`train`	296 (80.4%)	43	160 / 136	3 – 75	46.8
`validation`	35 (9.5%)	12	15 / 20	16 – 71	49.8
`test`	37 (10.1%)	12	21 / 16	20 – 80	51.7

Two consequences of blocking by (sex, age) worth knowing:

All 11 RT-PCR-negative studies fall in train. You cannot evaluate RT-PCR-negative behaviour on the recommended validation / test splits.
Because a whole age cell goes to one split, an age present in test may be absent from train. This is inherent to any leakage-free split of this data.

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

Access and intended use

Access is gated. Request access on the repository page; the terms are shown in the access form and reproduced in LICENSE_TERMS.md.

Intended use. Non-commercial academic research: medical report generation, vision-language pretraining, Chinese clinical NLP, radiology text mining, and teaching.

Out of scope. This dataset must not be used for clinical decision-making, diagnosis, triage, or any patient-care purpose. It is small, drawn from two hospitals over a seven-week window in early 2020, and overwhelmingly RT-PCR positive (357/368). A model trained on it will not generalize to a screening population, to other scanners or reconstruction kernels, or to other languages.

De-identification

Patient names, medical record numbers and other direct identifiers were removed at the source; they are not present in the images or the report text. Per-patient identifiers were removed so thoroughly that patient grouping is no longer recoverable — see above.
The images were cropped to the body field of view, which removes the DICOM corner annotation burned into the original reconstructions. All 3,680 released slices were re-checked programmatically for burned-in text (bright connected components of glyph size and aspect ratio isolated on a dark background) and every flagged region was reviewed by eye; all were anatomy or ECG electrodes, none was text.
case_uid retains the examination date. Dates were kept deliberately: the longitudinal structure is only visible because of them, and they carry no identifying power on their own once names and record numbers are gone.
Report text may mention incidental findings and comorbidities (e.g. cirrhosis, renal cysts, thyroid nodules). These are clinical, not identifying.

If you nonetheless believe you have identified residual identifying information, please contact the authors rather than disclosing it publicly.

What is not included

The knowledge-pretraining stage of Medical-VLBERT used a separate corpus of Chinese medical encyclopedia articles (3,052 diseases × 9 fields: overview, cause, symptom, examination, complication, prevention, treatment, diet). That corpus is not redistributed here: it was scraped from third-party Chinese medical websites and is not ours to relicense. It is not required to use this dataset; it is only needed to reproduce the paper's pretraining stage. Contact the authors if you need details on its provenance.

The large-scale Chinese chest X-ray dataset CX-CHR, also used in the paper, is a third-party dataset and is not distributed here either.

Citation

If you use this dataset, please cite the paper:

@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}
}

Paper (IEEE): https://doi.org/10.1109/TNNLS.2021.3099165
Paper (arXiv): https://arxiv.org/abs/2108.05067
Project page: https://covid19ct.github.io/

License

Released under CC BY-NC-SA 4.0: attribution required, non-commercial use only, derivatives must be shared under the same license. The additional access terms above apply on top of the license.

Acknowledgements

The dataset was constructed by the Shenzhen Research Institute of Big Data (SRIBD), the Future Network of Intelligence Institute (FNii) and the CUHKSZ-JD Joint AI Lab at The Chinese University of Hong Kong, Shenzhen, with data contributed by the two hospitals named above.

Supported in part by the Key Area R&D Program of Guangdong Province (No. 2018B030338001), the National Key R&D Program of China (No. 2018YFB1800800), the Shenzhen Outstanding Talents Training Fund, and Guangdong Research Project No. 2017ZT07X152.

We thank the radiologists of the two contributing hospitals, who wrote these reports under the conditions of early 2020.

Contact

Guangyi Liu · Yinghong Liao · Zhen Li — see the project page for current addresses.

Downloads last month: 13

Total file size:

374 MB

Paper for guangyil/COVID-19-CT-Chinese

Medical-VLBERT: Medical Visual Language BERT for COVID-19 CT Report Generation With Alternate Learning

Paper • 2108.05067 • Published Aug 18, 2021