Fundus10K

Fundus10K, containing 10,861 expert-labeled color fundus images, is so far the largest image collection for training and evaluating laser scar detection algorithms. Concerning data sources, the dataset consists of 9,864 images from the Kaggle Diabetic Retinopathy Detection challenge and 997 images from our hospital partners.

Basic statistics of Fundus10K are summarized as:

	Training (70%)	Validation (10%)	Testing (20)%
# Images	7,602	1,086	2,173
# Images from Kaggle	6,903	987	1,974
# Images with laser scars	282	42	80

Downloads

• Version 1: image size 448x448, released on Nov-28-2018: We provide 1) binary labels indicating whether a fundus image has laser scars visible, and 997+80 images with a resolution of 448x448. For the kaggle images, please download them from the Kaggle website.

State-of-the-art

Performance on Test2k

Model	Sensitivity	Specificity	Precision	AP	AUC
DenseNet-Ensemble	0.950	0.999	0.974	0.988	0.999

Performance on Test2k+

Model	Sensitivity	Specificity	Precision	AP	AUC
DenseNet-Ensemble	0.925	0.999	0.987	0.983	0.998

Reference

If you find the dataset useful, please consider citing the following paper:

@inproceedings{accv2018-laser-scar-detection,
title = {Laser Scar Detection in Fundus Images using Convolutional Neural Networks},
author = {Qijie Wei and Xirong Li and Hao Wang and Dayong Ding and Weihong Yu and Youxin Chen},
year = {2018},
booktitle = {Asian Conference on Computer Vision (ACCV)},
}