README.md

metadata

license: cc-by-nc-sa-4.0
language:
  - en
pretty_name: CT-RATE-Refined
tags:
  - medical
  - radiology
  - chest-ct
  - ct-rate
  - label-cleaning
size_categories:
  - 10K<n<100K
configs:
  - config_name: majority_vote
    default: true
    data_files:
      - split: train
        path: labels/volume_level/train_predicted_labels.csv
      - split: validation
        path: labels/volume_level/valid_predicted_labels.csv
  - config_name: report_level
    data_files:
      - split: majority_vote
        path: labels/report_level/majority_vote.csv
      - split: positive_votes
        path: labels/report_level/positive_votes.csv
      - split: gpt_5.4
        path: labels/report_level/gpt-5.4.csv
      - split: gpt_5.4_mini
        path: labels/report_level/gpt-5.4-mini.csv
      - split: deepseek_v3.2
        path: labels/report_level/deepseek-v3.2.csv
  - config_name: confirmatory_reference
    data_files:
      - split: reference_standard
        path: reference/confirmatory/reference_standard.csv
      - split: preconsensus_reader_1
        path: reference/confirmatory/preconsensus_reader_1.csv
      - split: preconsensus_reader_2
        path: reference/confirmatory/preconsensus_reader_2.csv
      - split: design
        path: reference/confirmatory/design.csv

CT-RATE-Refined

Refined report-aligned abnormality labels for the CT-RATE chest CT dataset, generated by multi-LLM majority voting (GPT-5.4, GPT-5.4 mini, and DeepSeek-V3.2), together with the model-specific outputs and the radiologist reference annotations used for validation.

This dataset accompanies the paper "Large Language Model–Assisted Cleaning of Report-Derived Labels in a Large-Scale Chest CT Dataset", accepted for publication in the Journal of Medical Systems. Methods and results are described in the paper (arXiv:2606.22382). Code: https://github.com/yamagishi0824/ct-rate-llm-label-cleaning.

Note. All labels were derived from report text. They reflect report alignment, not image-level truth.

Relationship to CT-RATE

This repository contains labels only. CT volumes and radiology reports are available from ibrahimhamamci/CT-RATE (access must be requested there). All files are keyed by the CT-RATE VolumeName (e.g., train_1_a_1.nii.gz).

• labels/volume_level/{train,valid}_predicted_labels.csv have the same rows, row order, and columns as CT-RATE's dataset/multi_abnormality_labels/{train,valid}_predicted_labels.csv and can be used as drop-in replacements.
• Labels were assigned per unique report (24,446 reports after deduplication on Findings and Impressions text) and propagated to every volume sharing that report. mapping/volume_to_report.csv gives this correspondence.

The 18 label columns follow CT-RATE: Medical material, Arterial wall calcification, Cardiomegaly, Pericardial effusion, Coronary artery wall calcification, Hiatal hernia, Lymphadenopathy, Emphysema, Atelectasis, Lung nodule, Lung opacity, Pulmonary fibrotic sequela, Pleural effusion, Mosaic attenuation pattern, Peribronchial thickening, Consolidation, Bronchiectasis, Interlobular septal thickening.

Files

labels/ — full dataset

File	Description
volume_level/train_predicted_labels.csv, volume_level/valid_predicted_labels.csv	Majority-vote labels per CT volume, in CT-RATE format (1 = present, 0 = absent).
report_level/majority_vote.csv	Majority-vote labels per unique report. VolumeName is the representative volume of the report; n_models_available is the number of models with valid output. A label is assigned when at least two models selected the same class and is empty otherwise (11 reports).
report_level/positive_votes.csv	Number of models assigning 1 to each label (0–n_models_available), as an indicator of intermodel agreement.
report_level/{gpt-5.4,gpt-5.4-mini,deepseek-v3.2}.csv	Labels from each model. extraction_failed = True for reports without valid output (GPT-5.4: 12; GPT-5.4 mini: 10; DeepSeek-V3.2: 0); their labels are empty.

mapping/

File	Description
volume_to_report.csv	VolumeName, split, and report_VolumeName (the representative volume used in report_level/).

reference/development/ — development-phase radiologist annotations

File	Description
random_set.csv	Development reference set: 100 randomly sampled reports × 18 labels, binary annotations by two readers (reader_1, reader_2).
discordance_general.csv	General discordance review set: CT-RATE and GPT-5.4 labels and the label supported by each reader (gpt-5.4, ct-rate, or both).
discordance_lymphadenopathy.csv	Targeted lymphadenopathy discordance review set; same columns.

reference/confirmatory/ — confirmatory reference set (250 reports, validation split)

File	Description
design.csv	Sampled reports with reading order, stratum, stratum size (stratum_N), sample size (stratum_n), and design weight.
preconsensus_reader_{1,2}.csv	Independent four-level annotations (present, uncertain, negated, not_mentioned) before consensus. temporality (current, prior_only, or resolved) is filled only for present and uncertain.
reference_standard.csv	Both readers' annotations, the consensus four-level label, source (agreed or adjudicated), and the binary label under Rule A (present and uncertain = 1).
llm_runs/{model}_run{1,2,3}.csv	Repeated binary runs of each model on the confirmatory set.
llm_runs/{model}_4level_run1.csv	Four-level run of each model on the confirmatory set.

License

CC BY-NC-SA 4.0, following the license of CT-RATE.

Citation

If you use these labels, please cite our paper and the original CT-RATE paper.

The paper has been accepted for publication in the Journal of Medical Systems. Until the journal version is available, please cite the arXiv preprint below.

@article{yamagishi2026llmlabelcleaning,
  title   = {Large Language Model-Assisted Cleaning of Report-Derived Labels in a Large-Scale Chest CT Dataset},
  author  = {Yamagishi, Yosuke and Takamatsu, Atsushi and Sato, Mototsugu and Kikuchi, Tomohiro and Hanaoka, Shouhei and Yoshikawa, Takeharu and Abe, Osamu},
  journal = {arXiv preprint arXiv:2606.22382},
  year    = {2026},
  doi     = {10.48550/arXiv.2606.22382}
}

@article{hamamci2026ctrate,
  title   = {Generalist foundation models from a multimodal dataset for 3D computed tomography},
  author  = {Hamamci, Ibrahim Ethem and Er, Sezgin and Wang, Chenyu and Almas, Furkan and Simsek, Ayse Gulnihan and Esirgun, Sevval Nil and others},
  journal = {Nature Biomedical Engineering},
  volume  = {10},
  number  = {8},
  pages   = {1610--1628},
  year    = {2026},
  doi     = {10.1038/s41551-025-01599-y}
}