ISIC 2020
SIIM-ISIC 2020 Melanoma Classification Challenge Dataset
Dermoscopic images of skin lesions grouped by patient, released for a melanoma classification challenge so models can compare a lesion with others on the same person. 44,108 images of 2,746 patients from six centres, 584 confirmed melanomas, with age, sex and body site.
Overview
The ISIC 2020 dataset was built for the SIIM-ISIC Melanoma Classification challenge, which ran on Kaggle from May to August 2020. Unlike earlier ISIC challenge sets, it groups lesions by patient, so a model can compare a suspicious lesion with the other lesions on the same person ("ugly duckling" reasoning). It is used for melanoma classification and for studying patient-level context.
Composition
The labelled training part holds 33,126 dermoscopic images of 2,056 patients; 584 images show melanoma and 428 patients have at least one. The remaining lesions are benign, either confirmed by biopsy or by at least six months of follow-up without change. A separate test part of 10,982 images from 690 patients is downloadable, but its labels were never released. Each image comes with an anonymised patient id, approximate age (rounded to five years), sex and a coarse anatomic site; version 2 of the training table adds a lesion id. Images are offered as JPEG and as DICOM with embedded metadata.
Acquisition
Images come from Memorial Sloan Kettering Cancer Center, Hospital Clínic de Barcelona, the Medical University of Vienna, Melanoma Institute Australia and the Sydney Melanoma Diagnosis Centre, and The University of Queensland, mostly high-risk or referral clinics. The test set also includes cases from a hospital in Athens. Contact and non-contact, polarised and non-polarised dermoscopy are mixed; the capture devices were not recorded.
Annotations
Malignant labels were checked against histopathology reports and visually by a dermoscopy expert. Melanoma in situ and invasive melanoma share one label. Most benign lesions carry the diagnosis "unknown"; nevus, seborrheic keratosis, lentigo and a few other benign diagnoses are given where available.
Known limitations
- 425 training images are pixel-identical duplicates; a list is provided, and lesion ids in version 2 help remove them.
- Darker skin types are under-represented, and melanoma is far more frequent than in the general population.
- Each lesion is one image with one type of dermoscopy.
- Age bins on this page are approximate because ages are rounded to five years.
- Condition counts cover training labels only.
Cohort
Aggregate numbers from the sources below. Bars are relative to the 2,746 subjects.
Sex
- Female 1,303 47%
- Male 1,441 52%
- Unknown 2 0%
Condition
Groups can overlap
- Melanocytic nevus 878 32%
- Melanoma 428 16%
- Seborrheic keratosis 104 4%
Table 1 in Rotemberg et al. 2021, Scientific Data 8, 34 (Table 1); From ISIC_2020_Training_GroundTruth_v2.csv and ISIC_2020_Test_Metadata.csv (open download, CC BY-NC 4.0), counted per image and per unique patient id; joint cells below 10 dropped
Age
mean 51.3
From ISIC_2020_Training_GroundTruth_v2.csv and ISIC_2020_Test_Metadata.csv (open download, CC BY-NC 4.0), counted per image and per unique patient id; joint cells below 10 dropped; Table 1 in Rotemberg et al. 2021, Scientific Data 8, 34 (Table 1)
Anatomy
- Trunk 2,614 95%
- Lower limb 2,174 79%
- Upper limb 1,906 69%
Sex by split
Reported cross table. A dot marks a cell the source does not give.
| Training | Test | |
|---|---|---|
| Female | 977 | 326 |
| Male | 1,077 | 364 |
| Unknown | 2 | · |
From ISIC_2020_Training_GroundTruth_v2.csv and ISIC_2020_Test_Metadata.csv (open download, CC BY-NC 4.0), counted per image and per unique patient id; joint cells below 10 dropped; Table 1 in Rotemberg et al. 2021, Scientific Data 8, 34 (Table 1)
Age by sex
Reported cross table. Missing cells were not published (fewer than 10 or not reported).
- 2180-8951
- 9070-79195
- 20460-69254
- 32050-59313
- 34340-49298
- 21530-39205
- 8620-2997
- 1710-1921
License and access
Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer
Download without an account
Creative Commons Attribution-NonCommercial 4.0 International
Use, share and adapt the data with credit, but only for non-commercial purposes.
The challenge data page lists the 2020 data under CC-BY-NC and asks that the aggregate data be cited as "International Skin Imaging Collaboration. SIIM-ISIC 2020 Challenge Dataset", DOI 10.34970/2020-ds01.
What you can do
- No
- Not stated
- Conditional
- Yes
- Yes
What you can share
- Conditional
- Conditional
- Conditional
What you must do
- Yes
- Share alike No
- No
- No
- Manuscript review No
- Release code No
- Return results No
- Delete after use No
Limits
- No
- Location limits No
Citation
International Skin Imaging Collaboration. SIIM-ISIC 2020 Challenge Dataset. https://doi.org/10.34970/2020-ds01 (2020). Rotemberg V, Kurtansky N, et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Sci Data 8, 34 (2021).
Sources
Every number on this page comes from one of these documents. Each chart names the table or page it is taken from. The raw numbers are in stats.csv.