chest2vec/chest2vec_labels

CT-RATE Findings — Chest Imaging Leaf Labels

Chest-CT findings from the CT-RATE dataset. Each row maps an original findings report → a section-structured refined version, plus a 137-label ternary multi-label annotation over a chest-imaging taxonomy.

• Rows: 23,614 unique CT-RATE findings reports (one row per report)
• Splits (report-text-level de-duplicated): train 20,648 / valid 1,483 / test 1,483
• Labels: 137 leaf labels (106 clinical + 31 other). Full taxonomy, definitions and per-split counts in LABEL_HIERARCHY.md.

Columns

Column	Type	Description
original_report	string	Original CT-RATE findings text (input)
refined_report	string	Section-structured / cleaned findings (target; empty for 281 rows)
split	string	train / valid / test
137 label columns	int8	Per-label status — see encoding below

Label encoding (ternary)

Value	Meaning
1	positive
0	negative
-1	uncertain
null	not assessed for this report

Notes

• Splits are de-duplicated across each other at the report-text level: no original_report or refined_report text appears in more than one split.
• Breast & Axilla were originally intended as their own top-level section, but because such findings are relatively infrequent in CT-RATE they were folded into the Others section. The individual breast/axilla leaf labels are still present (under Others).
• 281 rows have an empty refined_report (labels are still provided).
• 601 reports (2.5%) have no positive on any label.
• IVC filter is present for taxonomy completeness but is entirely unlabeled (all null — no positive, negative, or uncertain in any row).
• Label names exactly match the hierarchy in LABEL_HIERARCHY.md.

Radiologist validation (test set)

A radiologist manually reviewed 966 of the 1,483 test reports (65.1%), comparing the assigned labels against the report:

Verdict	Reports	Share
Fully accepted	857	88.7%
Imperfect / uncertain but acceptable	60	6.2%
Failed	49	5.1%
Reviewed	966	100%

Acceptable (accepted + borderline): 917 / 966 = 94.9%. This is a spot-check of the weak-label quality on the held-out test split, not a re-annotation — the published labels are the original pipeline output, unchanged.

Provenance & caveats

• Weak labels, not radiologist ground truth. Labels were generated by an LLM labeling pipeline from the report text (not from images), then validated against the fixed taxonomy.
• refined_report is an LLM-cleaned, section-structured rewrite of original_report.

License & attribution

This dataset is a derivative of CT-RATE and is released under CC-BY-NC-SA-4.0, the same license as CT-RATE. The report text (original_report) originates from CT-RATE; the refined_report rewrite and the 137 leaf labels are added by this work.

If you use this dataset you must cite the original CT-RATE paper (a requirement of the CC-BY-NC-SA attribution terms), in addition to this dataset:

@misc{hamamci2024ctrate,
  title         = {A foundation model utilizing chest CT volumes and radiology reports
                   for supervised-level zero-shot detection of abnormalities},
  author        = {Hamamci, Ibrahim Ethem and others},
  year          = {2024},
  eprint        = {2403.17834},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

Please confirm the canonical CT-RATE citation on the official CT-RATE page.

Usage

from datasets import load_dataset
ds = load_dataset("chest2vec/chest2vec_labels")
ds["test"][0]["original_report"]
[k for k, v in ds["test"][0].items() if v == 1]   # positive labels for the first test report