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.

Log in or Sign Up to review the conditions and access this dataset content.

MIMIC-CXR image embeddings for QML-MedImage

Pre-extracted image embeddings of MIMIC-CXR-JPG chest X-rays, organised by coreset-selection strategy and random seed, used to study quantum and classical SVMs on medical-imaging embeddings (label: new_insurance_type, derived from MIMIC-IV).

Licence and access. These data derive from MIMIC-CXR-JPG 2.0.0 and MIMIC-IV, which are distributed by PhysioNet under the PhysioNet Credentialed Health Data License 1.5.0. Use requires PhysioNet credentialing and signing the data use agreement (DUA). Access to this repository is granted manually, only to credentialed users. Do not redistribute any file or row of this repository.

References. Paper: arXiv 2604.24597. Basis paper: C.-Y. Lin et al., "Validating large-scale quantum machine learning: Efficient simulation of quantum support vector machines using tensor networks", Machine Learning: Science and Technology (2025), https://doi.org/10.1088/2632-2153/adb4ba (code: https://github.com/Tim-Li/cuTN-QSVM). The public MNIST and Fashion-MNIST counterpart is sebasmos/QuantumEmbeddings. The code repository of the MIMIC study is currently private; this card and the arXiv paper are the public descriptions.

Which folder should I use? (read this first)

Folder Encoder Model id and revision Pooling Dim Seeds Rows per seed (types 5 / 9 / 11) Provenance Status
medsiglip-448-embeddings/ MIXED, see note seed 0: google/medsiglip-448 (rev 9cea28a1); seeds 1-19: generic google/siglip-so400m-patch14-384 pooler_output (attention-pool head) 1152 0-19 1999 / 2371 / 2371 (seed 0); about 1999-2372 other seeds original extraction (2026-02), audited 2026-11 MIXED. Seeds 1-19 are NOT MedSigLIP. For the real model use reextracted-2026-11/medsiglip Types 5 and 11 of seeds 1 to 19 now also re-extracted, see reextracted-2026-11/medsiglip/.
vit-base-patch32-224-embeddings/ timm ViT-B/32 timm vit_base_patch32_224.augreg_in21k_ft_in1k (not the Google checkpoint google/vit-base-patch32-224-in21k) _cls_embedding: CLS after final norm; _gap_embedding: mean over the last 4 blocks' patch tokens, then over patches 768 0-19 1999 / 2371 / 2371 original, audited (reproduced at cosine 1.0000 for seeds 0, 1, 5) Correct as long as it is cited as the timm checkpoint
vit-base-patch16-224-embeddings/ ViT-B/16 not audited (checkpoint not verified) _cls_embedding and _gap_embedding, as above 768 0-19 1999 / 2371 / 2371 original extraction Not covered by the audit; cite with care
rad-dino-embeddings/ RAD-DINO microsoft/rad-dino (rev 110cbc18) pooler_output (CLS after final LayerNorm) 768 0-19 1998 / 2370 / 2370 (one image missing per type in seed 0) original, audited (reproduced, cosine 1.0) Correct
clip-biomed-embeddings/ unidentified unknown (no extraction code or model id in history) none: spatial feature map of 8 x 8 x 1376 per row (88,064 values) n/a 0-19 (plus pickles and data-cleaned-pca-*) 1999 / 2371 / 2371 original, 78 GB Not BiomedCLIP, not used by any experiment. An identical copy is at legacy-2026-02/clip-biomed-embeddings-unidentified/; this top-level folder is kept until the owner confirms its removal
reextracted-2026-11/ audit re-extractions, see below see below see below see below see below about 2369-2372 (type 9 only) extracted on MIT ORCD, 2026-10-31 and 2026-11-01 Use these for a model whose identity is documented
coreset-ids/ n/a n/a n/a n/a 0-19 one id list per strategy selection output Correct
legacy-2026-02/ see subfolders backups and unidentified material consolidated on 2026-11-01 from the private repo sebasmos/mimic-cxr-quantum (Feb 2026) and from this repo See the section legacy-2026-02/ below. Nothing here is used by the paper
tests/ n/a n/a n/a n/a n/a n/a verification scripts n/a

Audit finding for the stored files (2026-11)

Stored folder / seeds Model that reproduces the stored vectors Agreement on all rows
medsiglip-448-embeddings, seed 0 google/medsiglip-448 cosine 1.0000
medsiglip-448-embeddings, seeds 1-19 google/siglip-so400m-patch14-384 (generic SigLIP, not MedSigLIP) cosine 1.0000 (seeds 1, 5); cosine versus seed 0: 0.14
rad-dino-embeddings, all seeds microsoft/rad-dino cosine 1.000000
vit-base-patch32-224-embeddings, all seeds timm vit_base_patch32_224.augreg_in21k_ft_in1k cosine 1.0000

reextracted-2026-11/

Independently re-extracted embeddings of the same images (dicom_id), so that each stored file can be checked. Existing files elsewhere in this repo are untouched. Details, recipes and validation are in reextracted-2026-11/README.md and the README of each subfolder.

Subfolder Model id Pooling Dim Seeds Notes
medsiglip/ google/medsiglip-448, rev 9cea28a1a1195f665105faa6e8544c112fd960a4 pooler_output 1152 0-19 (seeds 1 and 5 also in seed_1_v2/, seed_5_v2/) The real MedSigLIP for all 20 seeds. Data types 5, 9 and 11 (60 files)
siglip384/ google/siglip-so400m-patch14-384 pooler, tok0, mean variants 1152 0, 1, 5 Reproduces the stored seeds 1-19 "MedSigLIP"
rad-dino/ microsoft/rad-dino, rev 110cbc18 pooler, mean, tok0_raw 768 0, 1, 5 Reproduces rad-dino-embeddings
vit32-timm/ timm vit_base_patch32_224.augreg_in21k_ft_in1k cls, gap 768 0, 1, 5 Reproduces vit-base-patch32-224-embeddings
vit32/ google/vit-base-patch32-224-in21k cls, pooler, gap, gap_lastlayer, cls_prenorm 768 0, 1, 5 Does NOT match the stored files (cosine 0.0); kept as the record of the model-card recipe
vit32-in21k/ google/vit-base-patch32-224-in21k, rev 98cb6f70 cls, mean 768 0-19 Type 9
dinov2-base/ facebook/dinov2-base, rev f9e44c81 cls, mean 768 0-19 Type 9
biomedclip/ microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224, rev 9f341de2 encode_image (projected, not normalised) 512 0-19 Type 9
xraysiglip/ StanfordAIMI/XraySigLIP__vit-b-16-siglip-512__webli, rev e2ad64ee pooler_output 768 0-19 Type 9

legacy-2026-02/ (consolidated 2026-11-01)

Everything old or unidentified is kept here so it is easy to find. Nothing in this folder is used by any experiment of the paper.

Subfolder Files Size What it is
baseline-embeddings/ 162 20.78 GB Pickles data_type<k>_insurance*.pkl and data-cleaned-pca-{100,500,1000,1500,1999}/ (PCA pickles, models/*.npz, pca_processing_summary.csv) from the Feb 2026 backup that were not already in this repo. 158 files (19.93 GB) are new; 4 files (0.85 GB) share a name with a file elsewhere in this repo but have different content. 15 files (2.45 GB) identical to files already present in clip-biomed-embeddings/ were not copied again
unidentified-pickles-8x8x1376/ 60 47.59 GB seed_<s>/data_type{5,9,11}_n*.pkl, seeds 0-19. UNIDENTIFIED: about 353 KB per row, i.e. an 8 x 8 x 1376 float32 map, the same shape as the unidentified clip-biomed-embeddings; probably a pickled copy of those maps. Not BiomedCLIP (BiomedCLIP gives a 512-d vector), producing model unknown, not used by any experiment. Pickles were not opened for this card
clip-biomed-embeddings-unidentified/ 106 77.97 GB Identical server-side copy of the top-level clip-biomed-embeddings/. UNIDENTIFIED: 8 x 8 x 1376 per row (88,064 values), a spatial feature map rather than an image embedding; not BiomedCLIP; no extraction code or model id known; not used by any experiment

The private source repo sebasmos/mimic-cxr-quantum stays as a backup until the owner confirms.

Coreset strategies and splits

Each seed has three coreset strategies (data_type); counts are for seed 0 and vary slightly by seed.

data_type Name Samples (seed 0)
5 PathologyStratifiedClean 1,999
9 GradMatch 2,371
11 Uncertainty 2,371

Split convention of the paper: stratified 80/10/10 train/validation/test split of each coreset on new_insurance_type with random_state = seed, seeds 0 to 19. The coreset id lists are in coreset-ids/seed_<s>/.

File format and loading

Parquet files hold a pandas DataFrame with the columns:

  • embedding: feature vector (list of floats)
  • dicom_id, subject_id, study_id: MIMIC-CXR identifiers
  • new_insurance_type: label (insurance type from MIMIC-IV)
  • metadata: further MIMIC-CXR / MIMIC-IV metadata columns (up to 49 columns in total in the seed-0 files; 15 in some seeds)

clip-biomed-embeddings/ also contains legacy pickles (.pkl, pandas DataFrames) with the same structure.

import pandas as pd
from huggingface_hub import hf_hub_download

# requires: PhysioNet credentialing, approved access to this repo, and `huggingface-cli login`
path = hf_hub_download(
    repo_id="MITCriticalData/qml-mimic-cxr-embeddings",
    filename="rad-dino-embeddings/20-seeds/seed_0/data_type9_n2370_seed0_rad_dino.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)
X = df["embedding"].tolist()
y = df["new_insurance_type"]

Stored files with two variants (ViT-B/16 and ViT-B/32) are ..._cls_embedding.parquet (CLS token) and ..._gap_embedding.parquet (patch tokens pooled over the last 4 blocks and over patches), both 768-dim.

Folder structure

qml-mimic-cxr-embeddings/
  coreset-ids/seed_0 ... seed_19/                 id lists per strategy
  medsiglip-448-embeddings/                       seed-0 files + 20-seeds/seed_<s>/ (MIXED, see above)
  vit-base-patch32-224-embeddings/                seed-0 files + 20-seeds/seed_<s>/ (cls and gap)
  vit-base-patch16-224-embeddings/                seed-0 files + 20-seeds/seed_<s>/ (cls and gap)
  rad-dino-embeddings/20-seeds/seed_<s>/
  clip-biomed-embeddings/                         unidentified 8x8x1376 maps, pickles, data-cleaned-pca-*, 20-seeds/
  reextracted-2026-11/                            see above
  tests/                                          verification scripts

Verification status (seed 0)

All seed-0 coreset ids were checked against the extracted embeddings: 100 % coverage for every folder except RAD-DINO, which lacks one image per strategy (1,998 of 1,999; 2,370 of 2,371 for strategies 9 and 11). The missing dicom_ids are not listed on this public card.

Citation

arXiv 2604.24597 (QML-MedImage paper), and C.-Y. Lin et al., Mach. Learn.: Sci. Technol. (2025), doi:10.1088/2632-2153/adb4ba, as the basis paper. Please also cite MIMIC-CXR-JPG and MIMIC-IV as required by PhysioNet.

Downloads last month
873