mocatex/colesion-ai-dataset

CoLesion AI - Dataset

Dermoscopy images of 7 skin-lesion classes, joined with how hard each test image is for humans and how dermatologists answer with and without AI support. Built for human-AI co-learning experiments.

Everything here is rebuilt from public sources by setup_dataset.py in our GitHub repository. BUILD.json stores every source URL and checksum, the split seed and the build date.

Overview

Type	Description
Images	11,427 dermoscopy JPEGs (10,015 train; 1,412 test), unchanged from the source
Masks	10,015 lesion segmentations (train images)
Human data	crowd votes on every test image; 12,260 trials from 89 dermatologists
Classes	akiec bcc bkl df mel nv vasc = dx_idx 0-6, fixed order
Licence	CC BY-NC 4.0, non-commercial research only

Files

tables/items.parquet             one row per image
tables/reader_trials.parquet     one row per reader x trial         (Barata 2023)
tables/support_effects.parquet   one row per test image x support   (Tschandl 2020)
tables/readers.parquet           one row per reader
images/train/ISIC_0024/ISIC_0024306.jpg
images/test/ISIC_0034/ISIC_0034524.jpg
masks/ISIC_0024/ISIC_0024306_segmentation.png
BUILD.json                       sources, checksums, seed, split counts

• Tables join on image_id. reader_trials keeps all trials, so 394 rows (96 images) have no match in items
• The 100 ISIC 2018 test images without crowd votes are left out
• Image folders are split by ID prefix (max 1'000 files each). Use image_path / mask_path from items

Quick start

hf download mocatex/colesion-ai-dataset --repo-type dataset --local-dir data   # ~3 GB

Splits (items.split)

split	from	share	used for
fit	ham_train (no human data)	85 %	train the AI model
val	ham_train	15 %	model selection
pool	isic_test (all human data)	80 %	co-learning rounds: human and AI work through these
holdout	isic_test	20 %	score AI, human (unaided) and team on the same cases; never trained on

Splits are drawn per lesion (a lesion never sits in two splits). They are stratified on dx and, on test, on the d_human quartile, with seed 2026.

Key columns

column	meaning
d_human	share of the crowd that picked the true class (1 = easy for humans, 0 = nobody got it). Test only
h_count_*, n_raters	crowd votes per class. n_raters is taken from the published shares -> null when the crowd was unanimous
h_entropy	normalised entropy of the crowd's answers (0 = all agree, 1 = uniform)
m_prob_*	reference model output shipped with the 2020 study, for sanity checks only
pair_id, phase	each image is shown twice per reader session: unaided, then aided
support	AI shown in the aided trial: SL (supervised) or RL (reinforcement learning); none when unaided
age, sex, localization, dx_type, lesion_id	from the ISIC Archive, one vocabulary for train and test

Sources and licences

source	used for	reference	licence
ISIC 2018 Challenge, Task 3	images, labels	Codella et al. 2019, arXiv:1902.03368	CC BY-NC 4.0
HAM10000	train images, lesion masks	doi:10.1038/sdata.2018.161 · doi:10.7910/DVN/DBW86T	CC BY-NC 4.0
Tschandl et al. 2020	crowd votes, support effects, model output, masks	doi:10.1038/s41591-020-0942-0 · doi:10.7910/DVN/DBW86T	CC BY-NC 4.0
Barata et al. 2023	reader trials	doi:10.1038/s41591-023-02475-5 · doi:10.7910/DVN/PWQMQ7	CC BY-NC 4.0
ISIC Archive, collections 66 + 67	age, sex, site, dx_type, lesion_id	api.isic-archive.com	CC BY-NC

The images are redistributed unchanged under CC BY-NC 4.0 for non-commercial research. Attribution as shipped with the ISIC 2018 files:

ISIC-2018 Challenge: Task 3: Training Input (HAM10000 Dataset: https://arxiv.org/abs/1803.10417 ) (c) by ViDIR Group, Department of Dermatology, Medical University of Vienna

ISIC-2018 Challenge: Task 3: Test Input (HAM10000 Dataset: https://arxiv.org/abs/1803.10417 ) (c) by ViDIR Group, Department of Dermatology, Medical University of Vienna

Citation

If you use this data, cite the original sources:

@article{tschandl2018ham10000,
  title   = {The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions},
  author  = {Tschandl, Philipp and Rosendahl, Cliff and Kittler, Harald},
  journal = {Scientific Data}, volume = {5}, pages = {180161}, year = {2018},
  doi     = {10.1038/sdata.2018.161}
}
@article{codella2019isic2018,
  title   = {Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)},
  author  = {Codella, Noel and others},
  journal = {arXiv preprint arXiv:1902.03368}, year = {2019}
}
@article{tschandl2020collaboration,
  title   = {Human--computer collaboration for skin cancer recognition},
  author  = {Tschandl, Philipp and Rinner, Christoph and Apalla, Zoe and others},
  journal = {Nature Medicine}, volume = {26}, number = {8}, pages = {1229--1234}, year = {2020},
  doi     = {10.1038/s41591-020-0942-0}
}
@article{barata2023rl,
  title   = {A reinforcement learning model for AI-based decision support in skin cancer},
  author  = {Barata, Catarina and Rotemberg, Veronica and Codella, Noel and others},
  journal = {Nature Medicine}, volume = {29}, number = {8}, pages = {1941--1946}, year = {2023},
  doi     = {10.1038/s41591-023-02475-5}
}