What’s Out There
← Back to swiping
Medicine & clinicalFreeAcademic

CheXpert

Stanford ML Group and Stanford Center for Artificial Intelligence in Medicine and Imaging (AIMI)

224,316 chest X-rays with labels for 14 findings, including when the radiologist was uncertain.

A large chest radiograph data set from Stanford Hospital, with 14 observations such as pneumonia, edema, and cardiomegaly labeled positive, negative, or uncertain by an automated reader of radiology reports. It became a standard benchmark for deep learning on medical images, with a radiologist-labeled test set. Data are distributed through Stanford AIMI.

Access

Cost
FreeFree download through Stanford AIMI after agreeing to its research use agreement.
License
Open, non-commercial onlyStanford University research use agreement (non-commercial)
Sign-up
Free account
How you get it
Bulk downloadJPEG, CSV

What's in it

One row is
One row per chest radiograph image
Covers
Stanford Hospital, California
Years
2002-2017
Updated
One-time release
Size
224,316 radiographs of 65,240 patients
image pathpatient IDview (frontal or lateral)sexageobservation labelsuncertainty labels

Quality

4 of 5

Large and well described, with expert test labels, but training labels come from automated report parsing and are noisy.

  • Training labels are machine-extracted from reports
  • Single institution
  • Images are downsampled from originals

Connects to

  • MIMIC-IV: Its sister set MIMIC-CXR uses the same 14 labels, enabling cross-hospital tests

Questions it could answer

  • Do models trained on Stanford X-rays hold up on other hospitals' images?
  • How does label uncertainty vary by finding?

Similar data sets

  • MedMNIST+: 18 standardized biomedical image datasets

    Eighteen medical imaging classification tasks shrunk to MNIST-size, in several resolutions.

    Medicine & clinicalYang, Shi, Wei, et al. (Shanghai Jiao Tong University)Free

  • NCT-CRC-HE-100K: histological images of colorectal cancer and healthy tissue

    100,000 stained tissue patches, labeled into nine tissue types, for training cancer classifiers.

    Medicine & clinicalKather, Halama, Marx (NCT Heidelberg)Free

  • OpenMind

    114,000 head-and-neck 3D MRI scans pooled from 800 public data sets for self-supervised learning.

    Medicine & clinicalGerman Cancer Research Center (MIC-DKFZ)Free

  • CT-RATE

    Chest CT scans paired with radiology reports, built to train 3D vision-language models.

    Medicine & clinicalCT-RATE team (VLM3D Challenge organizers)Free

  • MIMIC-IV

    Real, de-identified records of 364,000 Boston hospital patients, free to anyone who completes ethics training.

    Medicine & clinicalMIT Laboratory for Computational Physiology and Beth Israel Deaconess Medical Center (via PhysioNet)By application

  • NYU Langone and FAIR fastMRI Dataset

    Raw MRI scanner data for over 1,500 knees and 6,970 brains, built for faster scans.

    Medicine & clinicalfastMRI (NYU Langone Health and Meta FAIR)Free