APTOS 2019
APTOS 2019 Blindness Detection (Kaggle competition)
Retinal fundus photographs from Aravind Eye Hospital screening in rural India, each graded by a clinician for diabetic retinopathy severity on a 0 to 4 scale, released for a 2019 Kaggle competition.
Overview
APTOS 2019 Blindness Detection was a Kaggle competition run from June to September 2019 for Aravind Eye Hospital in Madurai, India, together with the Asia Pacific Tele-Ophthalmology Society. Its fundus photographs come from eye screening in rural areas, where technicians capture images that doctors later review. The labelled training images are widely used to train and compare models that grade diabetic retinopathy.
Composition
Each image shows the retina of one eye and carries a single severity grade from 0 (no diabetic retinopathy) through mild, moderate and severe to 4 (proliferative). The download contains a training set with labels and a public test set without them. Winning models were scored on a hidden private test set of roughly 13,000 images, which was never released. The number of training images and of patients is not stated on the readable competition pages, and no patient identifiers, ages or sex are provided.
Acquisition
According to the organisers, images were collected at several clinics with different cameras over a long period. Quality varies: some images are out of focus, under- or overexposed or contain artifacts.
Annotations
A clinician assigned each image one diabetic retinopathy grade. The organisers warn that labels contain noise. The competition metric was quadratic weighted kappa between predicted and clinician grades.
Known limitations
- Use is restricted to non-commercial purposes, and the data may not be shared with anyone who has not accepted the competition rules.
- Label noise and varying image quality, with no information on graders, cameras or clinics per image.
- The private test set is not available, so published results on it cannot be reproduced.
Cohort
Aggregate numbers from the sources below. Bars are relative to the largest value.
Split
images
- Test ~13,000
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
Create an account and accept terms online
APTOS 2019 Blindness Detection Official Competition Rules (Kaggle)
Kaggle competition rules accepted when joining the competition. The data may be used only for non-commercial purposes, including the competition and academic research and education, and must not be passed on to anyone who has not accepted the rules. Residents of several sanctioned countries are excluded. The rules say nothing about citation, derived data or publications.
Kaggle lists the data license as "Subject to Competition Rules". Kaggle's own Terms of Service also apply to every account holder.
What you can do
- No
- Conditional
- Not stated
- Not stated
What you can share
- No
- Not stated
- Not stated
What you must do
- Not stated
- No
- Yes
- Ethics approval No
- Manuscript review No
- Conditional
- Conditional
- Delete after use No
Limits
- No
- Yes
Citation
APTOS 2019 Blindness Detection. Kaggle competition hosted for Aravind Eye Hospital and the Asia Pacific Tele-Ophthalmology Society, 2019. https://www.kaggle.com/competitions/aptos2019-blindness-detection
All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Images | split=test Hidden private test set ("approximately" 13000 images, 20 GB); training and public test counts are not stated in readable public documentation | ~13,000 | kaggle-aptos2019-data Files |
Sources
The keys used in the table above.
- kaggle-aptos2019-data Kaggle competition data description page (read through Kaggle's page API) website