MedOtter/ISIC2017

ISIC 2017 — Skin Lesion Segmentation (Task 1)

Dermoscopic RGB images of skin lesions (melanoma / seborrheic keratosis / nevus) with expert binary lesion-boundary masks, from the ISBI 2017 challenge "Skin Lesion Analysis Toward Melanoma Detection" hosted by the International Skin Imaging Collaboration (ISIC).

Split	Images	Masks	Melanoma	Seborrheic keratosis	Nevus
train	2000	2000	374	254	1372
validation	150	150	30	42	78
test	600	600	117	90	393

All three splits carry public ground truth. The test split is the official Test_v2 release — the final public revision of the challenge test set (v1 was withdrawn upstream). Resolutions vary widely (~540×722 up to 4499×6748 px).

Schema

Column	Type	Notes
image_id	string	Stable ISIC Archive ID (ISIC_XXXXXXX) — cross-references all ISIC challenge editions
image	Image	Original challenge JPEG, bytes unmodified
mask	Image	Original expert mask PNG: 0 = background, 255 = lesion
age_approximate	float	From the official challenge metadata CSV; null when unknown
sex	string	male / female; null when unknown
in_isic2018_train	bool	True if this image also appears in ISIC 2018 Task 1's training split (see below)

Masks were created by expert clinicians via manual polyline tracing or a supervised flood-fill workflow (one published mask per image; the per-image method is not disclosed). The *_superpixels.png scaffolding files and the Part-2 (dermoscopic features) / Part-3 (classification) ground truth belong to other challenge tasks and are not mirrored here.

⚠️ Overlap with ISIC 2018 (leakage warning)

ISIC challenge editions draw from the same growing archive. Measured by exact image_id intersection against ISIC 2018 Task 1:

• 2,450 / 2,750 images (89.1%) of ISIC 2017 reappear in ISIC 2018's training split — train 1800/2000, validation 121/150, test 529/600.
• ISIC 2018's own validation/test splits share zero IDs with ISIC 2017.
• Consequence: a model trained on ISIC 2018 Task 1 training data has already seen 650 of the 750 ISIC 2017 validation+test images. Filter on in_isic2018_train (or use isic_2017_2018_split_ids.json at the repo root) before treating the two datasets as independent benchmarks.
• Filename-level matching is a lower bound: pixel-level near-duplicates (rescaled/re-encoded variants) exist across ISIC editions and against ISIC 2016 — see Cassidy et al., Medical Image Analysis 75:102305 (2022), https://github.com/mmu-dermatology-research/isic_duplicate_removal_strategy.

Provenance

Official author-hosted S3 bucket (https://isic-archive.s3.amazonaws.com/challenges/2017/), zips verified byte-exact against Content-Length; image/mask counts match the challenge paper (2000/150/600). Original encoded bytes are embedded unmodified.

License

CC-0 (public domain), per the 2017 section of the ISIC Challenge data page.

Citation

@inproceedings{codella2018skin,
  title     = {Skin lesion analysis toward melanoma detection: A challenge at
               the 2017 International Symposium on Biomedical Imaging (ISBI),
               hosted by the International Skin Imaging Collaboration (ISIC)},
  author    = {Codella, Noel C. F. and Gutman, David and Celebi, M. Emre and
               Helba, Brian and Marchetti, Michael A. and Dusza, Stephen W. and
               Kalloo, Aadi and Liopyris, Konstantinos and Mishra, Nabin and
               Kittler, Harald and Halpern, Allan},
  booktitle = {2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI)},
  pages     = {168--172},
  year      = {2018},
  doi       = {10.1109/ISBI.2018.8363547}
}