Skip to content
Open Imaging Index
X-ray Chest

CheXthought

CheXthought: A global multimodal dataset of clinical chain-of-thought reasoning and visual attention for chest X-ray interpretation

Spoken step-by-step reasoning of radiologists reading chest X-rays, with synchronized click points on the image, collected to train and test vision-language models that explain where and why they see findings. 103,592 reasoning traces from 501 readers in 71 countries.

Go to dataset Edit Report

Overview

CheXthought records how radiologists think while they read chest X-rays. Readers talked through each image in a browser-based DICOM viewer and clicked on the regions they were describing at the same time. Every reasoning trace therefore pairs free text with numbered image points. The traces were collected to study clinical reasoning and to train vision-language models whose explanations are grounded in the image. The images and their original reports come from CheXpert Plus at Stanford, and the release is hosted by Stanford AIMI on Redivis.

Composition

Version 1.0 contains 103,592 reasoning traces with 6,609,082 click points, made by 501 attending radiologists, residents and fellows from 71 countries. They cover 50,312 chest X-rays from 45,272 studies of 18,046 patients. About a fifth of the patients have two or more studies, which allows comparison of serial images. Many images were read by several people independently, up to 36 per image. The paper describes another 10% of traces held back as a test set for future benchmarking, which is not part of the public release.

The download is one CSV file of about 0.5 GB with the image id, the reasoning text, the click coordinates and the CheXpert Plus report for each trace. The images themselves are fetched from CheXpert Plus by image id.

Annotations

Readers followed a structured outline: image quality and search pattern, lines and tubes, normal and abnormal findings, uncertainty, final findings, comparison with earlier images and follow-up advice. They reviewed and corrected the automatic transcripts before submitting. Coordinates are stored relative to the image so that zoom and pan do not shift them. A typical trace has about 117 words and 66 points.

Known limitations

  • The click points track where a reader pointed, not where they looked. They are not eye-tracking data and they are not lesion outlines.
  • Finding counts in the paper are given per trace, not per image or patient, so they are not listed here.
  • Access needs a Redivis account, AIMI membership and a Stanford research agreement whose text is visible only after sign-in. Images need separate access to CheXpert Plus.
  • The dataset DOI shown on Redivis (10.57761/j56p-e576) did not resolve on 2026-11-01.

License and access

Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer

Access
Free registration

Create an account and accept terms online

Sign in to Redivis, join the Stanford AIMI organization and complete the per-dataset Stanford research agreement requirement. Metadata are public; the file is visible only after both steps. The release holds annotations and reports only. The images are downloaded separately from CheXpert Plus by image id.

Access page

Stanford University School of Medicine Dataset Research Use Agreement (AIMI)

Free for personal, non-commercial research only. You may not share the data or the download link, may not create derivative works, and must not try to re-identify patients. Stanford sells a separate commercial license.

The Redivis access panel names the requirement "Stanford_Research_Agreement"; its text is shown only after sign-in and was not read for this entry. The images come from CheXpert Plus, which has its own access terms.

Original license text Version read: Text shown in the Stanford AIMI shared-datasets portal, read 2026-10-07 (no version number given) Checked 2026-10-07

What you can do

  • No
  • Not stated
  • Not stated
  • No
  • Not stated

What you can share

  • No
  • No
  • Share trained models Not stated

What you must do

  • Yes
  • Share alike No
  • Yes
  • Ethics approval No
  • Manuscript review No
  • Release code No
  • Return results No
  • No

Limits

  • Yes
  • Location limits No

Commercial license: Yes Companies can apply for a one-year license per dataset, reviewed by a Stanford committee. The page lists an annual fee of USD 70,000 per dataset for agreements contracted in FY25. Details

Citation

Sharma S, Long J, Shih G, et al. CheXthought: A global multimodal dataset of clinical chain-of-thought reasoning and visual attention for chest X-ray interpretation. arXiv:2604.26288 (2026). https://arxiv.org/abs/2604.26288

Sources

Every number on this page comes from one of these documents. Each chart names the table or page it is taken from. The raw numbers are in stats.csv.