DermaLit-140k / README.md
robvanvolt's picture
Card: link to the fine-tuned CLIP models (DermaLit-CLIP ViT-B/16)
a74dd34 verified
|
Raw History Blame Contribute Delete
8.39 kB
metadata
license: other
license_name: mixed-creative-commons-per-item
license_link: LICENSE.md
language:
  - en
  - de
pretty_name: DermaLit-140k (open tier and metadata release)
size_categories:
  - 100K<n<1M
task_categories:
  - image-to-text
  - zero-shot-image-classification
  - image-classification
tags:
  - dermatology
  - medical
  - image-text
  - provenance
  - fitzpatrick
configs:
  - config_name: open
    data_files: open/samples.parquet
  - config_name: metadata
    data_files: metadata/samples.parquet

DermaLit-140k — open tier and metadata release

DermaLit-140k is a provenance-tracked dermatology image–text corpus: 139,642 image–text samples (134,623 unique images) from 7 dermatology journals (1997–2025), 14 web atlases and 2 textbooks. More than 100,000 images are dermatological (clinical photographs, dermoscopy or histopathology).

Most of the source material is under copyright and cannot be redistributed. This repository therefore contains two parts:

Part Rows What you get
open/ 7,639 samples (7,282 images) figures with images and captions from 2,182 Creative Commons–licensed journal articles, with attribution
metadata/ 129,848 samples identifiers and derived labels only: DOI and figure label or image URL, hashes, image size, image type, Fitzpatrick estimate, diagnosis label, split, duplicate groups. No images and no caption text

The full corpus is kept as an internal research copy under the German text-and-data-mining exception for scientific research (§60d UrhG). The two textbooks are not part of this release.

Not for clinical use. This is a research resource, not a diagnostic tool.

Content warning

The open tier contains medical photographs of skin disease, including wounds, tumours and infections, as published in the source articles.

The open tier

  • What it is. Figures, captions and up to three citing sentences (context) from articles that the publisher page marked with a Creative Commons licence. 169 figures whose caption credits a third party ("reproduced with permission", "courtesy of", "©") were excluded, because the article licence does not cover them.
  • Sources (samples). JDDG 2,295; Journal of Investigative Dermatology 1,856; Dermatology and Therapy 1,383; JAAD 1,006; Die Dermatologie 554; American Journal of Clinical Dermatology 492; JAMA Dermatology 53.
  • Languages. English 5,953; German 1,686.
  • Image types (predicted). Non-photographic figures (charts, diagrams) 5,112; clinical photographs 1,269; histopathology 530; other microscopy 470; other photographs 108; radiology 89; dermoscopy 61. Most open-access figures are charts, not photographs. Filter on image_type if you need clinical images.
  • Licences (samples). CC BY-NC-ND 3,194; CC BY-NC 2,876; CC BY 1,560; CC BY-ND 9. The licence of each sample is in the license column, and open/ATTRIBUTION.csv lists DOI, title, authors, journal, year and licence for every article.

You must follow the licence of each item. NC items may only be used non-commercially. ND items may only be shared unmodified. If you need the most permissive part only, keep the 1,560 samples whose licence is plain CC BY.

The metadata tier

Use it to rebuild the same sample from your own lawful copy of a source, to check your data for overlap with public benchmarks, or to study composition and bias.

  • 21,474 rows have terms_restrict_ai = True: the source's terms explicitly restrict AI use (DermNet NZ, DermIS). They are listed for transparency. Do not use them to train a model you intend to release.
  • llm_vision_description and llm_vision_most_likely_diagnosis were written by a vision–language model (Claude 3 Opus) from the image. They are unvalidated.
  • fst_llm is a model estimate that was too light on average (exact agreement with expert labels 38%). Prefer fst_pred, and use it only at the level of distributions.

Columns

Column Description
sample_id <source>-<number>, one row per image–text occurrence
image_id, sha256, phash image hashes (image_id = first 16 hex characters of the SHA-256)
source, source_type source key; journal or atlas
split train / val / test, grouped by near-duplicate cluster, case, article and same-patient links
release_tier open or metadata
license detected article licence (journals) or the source's terms (atlases)
terms_restrict_ai the source's terms explicitly restrict AI use
language, figure_label, doi, url, image_url, year provenance
diagnosis, body_site, age, sex as given by the source, not normalised
image_type, subset predicted image type; derm_core = clinical, dermoscopy or histopathology
fst_gold_f17k expert Fitzpatrick type from Fitzpatrick17k, where the image is also in that benchmark
fst_pred Fitzpatrick estimate of a CLIP-based regressor (clinical photographs only)
near_dup_cluster, case_group duplicate and same-case groups
bench_overlap, bench_overlap_test public benchmarks that contain a near-duplicate of this image
width, height image size in pixels
open tier only: file, caption, context, title, authors, journal_name image path in this repository, verbatim caption, citing sentences, article data

Classifier heads

classifiers/ holds two small linear heads for L2-normalised OpenAI CLIP ViT-B/16 image embeddings (512 dimensions):

File Task Validation
image_type_head_clip_vitb16.npz seven image types (multinomial logistic: coef, intercept, classes) 93.8% accuracy on a random sample of 288 hand-labelled images (5-fold cross-validation); 89.3% on all 383 labelled images, which include deliberately hard cases. Dermoscopy recall is weak (0.55, 11 positives)
fitzpatrick_head_clip_vitb16.npz Fitzpatrick skin type as a number 1–6 (ridge regression: coef, intercept) grouped cross-validation on 14,862 expert-labelled Fitzpatrick17k images: mean absolute error 0.71, 91% within ±1 type, exact 43%. Dark skin (V–VI) is under-detected (recall 0.63)

Known limitations

  • Benchmark overlap. 23,502 images of the corpus are near-duplicates of images in public benchmarks (Fitzpatrick17k, SkinCAP, Daffodil, HAM10000, PAD-UFES-20), mainly because Fitzpatrick17k was built from two of the atlases. Exclude rows with bench_overlap before you evaluate on these benchmarks.
  • Skin type. Only about 6% of clinical photographs are estimated as Fitzpatrick V–VI (11% after correcting the estimator's bias). Thirteen diagnoses with at least 100 images have no type V–VI image.
  • Labels. Atlas diagnoses are as published and partly noisy. Diagnoses are not normalised.
  • Composite figures. Journal figures often combine several panels and are typed by their dominant content.
  • Licence detection is time-bound. It reflects the article pages as retrieved in 2025.
  • Same patient across splits. Splits respect 24,786 detected same-patient links, but up to a few percent of atlas test photographs may still have a photo of the same patient in the training split.
  • Links decay. Some atlas domains have changed owner. Affected rows carry an archived URL or none.

Removal requests

If you hold rights in an item and want it removed or corrected, open a discussion on this repository.

Models

CLIP ViT-B/16 models fine-tuned on this corpus: https://huggingface.co/laion/DermaLit-CLIP-ViT-B-16 (access on request, non-commercial research licence). The default model was trained without the rows with terms_restrict_ai = True.

Code

https://github.com/LAION-AI/derm-datasets

Funding and citation

Built in the project "The future of medical datasets: large-scale analysis and training of medical multimodal models through an iterative ML approach", funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Walter Benjamin Programme, project number 526052741.

@misc{kaczmarczyk2026dermalit,
  author = {Kaczmarczyk, Robert},
  title  = {DermaLit-140k: a provenance-tracked dermatology image--text corpus (open tier and metadata release)},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/laion/DermaLit-140k}}
}