You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access requires PhysioNet credentials for MIMIC-CXR. State your PhysioNet username; approval is manual after the requester emails proof to sebasmos@mit.edu.
Log in or Sign Up to review the conditions and access this dataset content.
ShortKit-ML Benchmark Data
This folder holds the inputs read by scripts/revision/common.py, the loader behind every result in the revised paper. Nothing under data/ is committed to GitHub except this README, .gitignore and .gitattributes. The derived files (embeddings, manifests, metadata, race_mapping.csv) are hosted on the gated Hugging Face dataset MITCriticalData/ShortKit-ML-data. Access is granted to users who hold the matching credentials (Stanford for CheXpert, PhysioNet for MIMIC-CXR).
hf download MITCriticalData/ShortKit-ML-data --repo-type dataset --local-dir data/
Files not on the hub
The Stanford files and the CelebA images are never redistributed; download them from the original source and place them at the paths below. The two MIMIC-CXR files in the last row are rebuilt locally.
| File | Source | Used for |
|---|---|---|
data/train.csv |
CheXpert-v1.0-small train.csv, Stanford research use agreement |
sex, age, view, AP/PA and the 14 labels, joined to the embeddings by image path through race_mapping.csv |
data/CHEXPERT DEMO.xlsx |
CheXpert demographics file from Stanford | PRIMARY_RACE per patient |
data/celeba_real/celeba/ |
CelebA via torchvision.datasets.CelebA (identity_CelebA.txt, list_attr_celeba.txt, img_align_celeba/) |
celebrity identity for grouping; images for scripts/extract_celeba_embeddings.py |
mimic_cxr/vit32_cls_embeddings_realigned.npy, mimic_cxr/mimic_race_raw.csv |
Rebuilt by scripts/revision/prepare_mimic_source.py from the coreset parquet files in mimic_cxr/source/ |
the realigned ViT-B/32 array and the raw MIMIC race strings |
Layout
data/
├── train.csv # from Stanford (see above)
├── CHEXPERT DEMO.xlsx # from Stanford (see above)
│
├── chexpert_multibackbone/ # CheXpert, 2,000 images, 477 patients
│ ├── chexpert_manifest.csv # row order, task_label (Pneumothorax), age_bin
│ ├── race_mapping.csv # image_idx, path, patient_id, PRIMARY_RACE
│ ├── resnet50_embeddings.npy # (2000, 2048)
│ ├── densenet121_embeddings.npy # (2000, 1024)
│ ├── vit_b_16_embeddings.npy # (2000, 768)
│ ├── {backbone}_metadata.csv # one per backbone, same row order
│ └── public/ # re-extracted from pinned public checkpoints
│ ├── vit_b_32__vit_b_32__class.npy # torchvision ViT_B_32_Weights.DEFAULT
│ ├── dinov2_base__facebook_dinov2-base__pooler_output.npy
│ ├── rad_dino__microsoft_rad-dino__pooler_output.npy
│ └── medsiglip__google_medsiglip-448__vision.npy # (2000, 1152)
│
├── mimic_cxr/ # MIMIC-CXR, 1,491 subjects, one image each
│ ├── rad_dino_embeddings.npy # (1491, 768)
│ ├── vit16_cls_embeddings.npy # (1491, 768)
│ ├── vit32_cls_embeddings_realigned.npy # (1491, 768), built by prepare_mimic_source.py
│ ├── medsiglip_embeddings.npy # (1491, 1152), google/siglip-so400m-patch14-384
│ ├── {backbone}_metadata.csv # image_path, task_label, race, sex, age, age_bin
│ ├── mimic_race_raw.csv # image_path, race_raw (built by prepare_mimic_source.py)
│ ├── mimic_cxr_manifest.csv
│ ├── mimic-cxr-2.0.0-chexpert.csv # PhysioNet label file
│ └── mimic-cxr-2.0.0-metadata.csv # PhysioNet DICOM metadata
│
├── celeba/ # CelebA, 10,000 images
│ ├── celeba_real_embeddings.npy # (10000, 2048), ResNet-50 ImageNet avgpool
│ ├── celeba_real_attributes.csv # image_id plus the 40 CelebA attributes
│ └── celeba_real_metadata.csv
│
└── celeba_real/celeba/ # torchvision CelebA download (see above)
CheXpert (chexpert_multibackbone/)
The 2,000 images are the first 2,000 rows of the official train.csv, from 477 patients. Seven backbones are audited. ResNet-50, DenseNet-121 and ViT-B/16 are read from <backbone>_embeddings.npy. ViT-B/32, DINOv2, RAD-DINO and MedSigLIP-448 are read from public/, which was extracted from the same images with scripts/revision/extract_missing_chexpert_backbones.py and pinned public checkpoints; the same code reproduces the stored ViT-B/16 array to cosine 1.000. The older vit_b_32, dinov2_base, rad_dino and medsiglip arrays at the top level have no extraction code and are kept only to replay the first submission (stored=True in the loader).
Sex, age, view and the labels come from train.csv, joined by image path through race_mapping.csv, which also gives patient_id. The sex column in the shipped _metadata.csv files is inverted and is not used. Race comes from CHEXPERT DEMO.xlsx (PRIMARY_RACE, joined by patient): images without a demographics record, or recorded as Unknown or Patient Refused, are missing; Native Hawaiian or Pacific Islander is its own group. The primary analysis compares White against non-White. Cross-validation folds and the permutation null are grouped by patient.
MIMIC-CXR (mimic_cxr/)
1,491 subjects with one image each: the coreset in mimic_cxr/source/ (strategy 5, PathologyStratifiedClean, 1,999 images, seed 0) minus rows without an in-hospital mortality label. Four backbones: RAD-DINO, ViT-B/16, ViT-B/32 and the array labelled MedSigLIP, which was extracted with google/siglip-so400m-patch14-384 and is reported in the paper as SigLIP (So400m). The stored vit32_cls_embeddings.npy does not match its metadata; the loader reads vit32_cls_embeddings_realigned.npy instead. Race is coded from mimic_race_raw.csv, with Unknown, Unable to Obtain and Patient Declined treated as missing.
CelebA (celeba/)
10,000 images embedded with an ImageNet ResNet-50 by scripts/extract_celeba_embeddings.py, grouped by celebrity identity from celeba_real/celeba/identity_CelebA.txt. The attribute is Male; the tasks are Blond_Hair, Heavy_Makeup and Attractive.
First-submission files
chest_embeddings.npy, chest_labels.npy, chest_group_labels.npy, chexpert_manifest.csv and the chexpert/ folder on the hub belong to the first submission (MedCLIP and the stored backbones) and are only needed to replay it with scripts/run_dataset2_benchmark.py.
Provenance
- CheXpert: Stanford ML Group, CheXpert. Images from the
danjacobellis/chexpertmirror; labels, sex and age from the officialtrain.csv. - MIMIC-CXR: Johnson et al., MIMIC-CXR-JPG embeddings and metadata in
mimic_cxr/, with the source parquet files inmimic_cxr/source/. Diagnosis labels from PhysioNet; demographics from MIMIC-IV joined bysubject_id. - CelebA: Liu et al., Large-scale CelebFaces Attributes Dataset.
- Downloads last month
- 49