By requesting access to this repository, you acknowledge that MedConcept-23M is a research resource derived from publicly available PMC-OA materials and metadata processing. You are responsible for complying with any applicable upstream source terms when reconstructing, redistributing, or using source-derived content. You also agree not to use this resource to identify individuals, and not to use it as a substitute for clinical judgment or patient care.

Dataset Card for MedConcept-23M

Dataset Summary

MedConcept-23M is a large-scale biomedical image-text-concept resource associated with the ConceptCLIP project.

It is designed to support research on:

biomedical vision-language pretraining
concept-grounded medical image understanding
explainable medical AI
multimodal retrieval and downstream evaluation

In the associated paper, MedConcept-23M is described as a large-scale resource of biomedical image-text-concept triplets used to pretrain ConceptCLIP.

What This Repository Contains

This repository is intended to host the released artifacts associated with MedConcept-23M, such as processed captions, concept annotations, metadata, and reconstruction-related resources, depending on the release version.

Users should inspect the repository file tree carefully to understand exactly which assets are included in the current release.

Data Provenance

MedConcept-23M is derived from biomedical figure-caption data collected from the PubMed Central Open Access Subset (PMC-OA) and further enriched with standardized biomedical concepts.

At a high level:

biomedical image-text pairs are collected from PMC-OA
captions are processed to identify candidate biomedical entities
entities are linked to standardized concepts
the resulting image-text pairs are enriched into image-text-concept triplets

This provenance matters because downstream users may interact with both:

files released directly in this repository, and
reconstructed or linked source-derived content obtained from upstream resources

Gated Access Notice

This repository uses gated access with contact sharing and an additional acknowledgment form.

The gate is intended to make users explicitly acknowledge that:

this is a research resource, not a clinical product
provenance and upstream-source compliance matter
users remain responsible for lawful and ethical use
source-derived content may carry obligations beyond the metadata hosted here

By requesting access, users confirm that they understand these points.

Intended Uses

Direct Use

Biomedical vision-language pretraining
Research on medical concept grounding
Retrieval benchmarking and representation learning
Data-centric studies of biomedical multimodal learning
Reproducibility work for the ConceptCLIP project

Downstream Use

Fine-tuning or continued pretraining of biomedical foundation models
Studying concept extraction or concept alignment in medical AI
Building explainable medical image analysis pipelines
Internal research evaluation under appropriate governance

Out-of-Scope Use

Re-identification or privacy attacks
Use as a medical device or as a substitute for clinician judgment
Clinical decision-making without task-specific validation and oversight
Redistribution or reuse in ways that violate upstream source terms, local law, or institutional policy

Responsible-Use and Compliance Notice

Using this repository may involve more than simply downloading files.

Depending on your workflow, you may also:

reconstruct content from upstream sources
link back to upstream materials
build derived datasets or training corpora

If you do so, you are responsible for:

checking applicable upstream terms
preserving provenance where needed
following your institution's data-governance requirements
complying with applicable law and publisher / repository rules

This dataset should not be used for any attempt to identify individuals or for any privacy-invasive purpose.

Recommended Good Practice

Keep clear records of which files come from this repository versus upstream reconstruction.
Preserve source provenance and versioning information.
Document all filtering, cleaning, and reconstruction choices in derived work.
Validate downstream models before any real-world use.

Relationship to Other Project Repositories

Model weights: JerrryNie/ConceptCLIP
Held-out retrieval benchmark: JerrryNie/pmc9k
Code and reconstruction utilities: https://github.com/JerrryNie/ConceptCLIP

Example Loading Pattern

from datasets import load_dataset

# Replace the split / config names with the actual ones present in the repo.
dataset = load_dataset("JerrryNie/MedConcept-23M")
print(dataset)

Limitations

The released repository contents may differ from a fully reconstructed end-to-end corpus.
Upstream biomedical figures and captions can contain publication bias, modality imbalance, and domain skew.
Rare diseases or rare concepts may still be underrepresented.
Benchmark performance achieved with models trained on this resource does not by itself establish clinical validity.

Citation

If you use this resource, please cite:

@article{nie2025conceptclip,
  title={An Explainable Biomedical Foundation Model via Large-Scale Concept-Enhanced Vision-Language Pre-training},
  author={Nie, Yuxiang and He, Sunan and Bie, Yequan and Wang, Yihui and Chen, Zhixuan and Yang, Shu and Cai, Zhiyuan and Wang, Hongmei and Wang, Xi and Luo, Luyang and Wu, Mingxiang and Wu, Xian and Chan, Ronald Cheong Kin and Lau, Yuk Ming and Zheng, Yefeng and Rajpurkar, Pranav and Chen, Hao},
  journal={arXiv preprint arXiv:2501.15579},
  year={2025}
}

Contact

For questions about the dataset release, please use the repository discussion page or the main project repository.

Downloads last month: 8

Total file size:

44.9 GB

Paper for JerrryNie/MedConcept-23M

ConceptCLIP: Towards Trustworthy Medical AI via Concept-Enhanced Contrastive Langauge-Image Pre-training

Paper • 2501.15579 • Published Jan 26, 2025