The International Skin Imaging Collaboration (ISIC) 2018 Challenge comprises 10,015 dermoscopic skin lesion images, annotated with one of seven possible skin disease types.
Dataset homepage: https://challenge.isic-archive.com/landing/2018/47
Paper: https://arxiv.org/abs/1902.03368
The bash script download_data.sh provided in this directory can be used to download the actual image data and original annotations.
Doing so will result in a folder ISIC2018_Task3_Training_Input containing the images and a file ISIC2018_Task3_Training_GroundTruth.csv with the original annotations.
The original dataset is imbalanced, so we created the following non-exhaustive but balanced split:
| Split | Total Images | Images / Class |
|---|---|---|
| train | 350 | 50 |
| val | 210 | 30 |
| trainval | 560 | 80 |
| test | 1,944 | 35-400 |
Note that the test set is still imbalanced. Therefore, balanced accuracy should be used to assess performance instead of plain accuracy.
We achieved a baseline performance of 66.19% in terms of balanced accuracy (averaged over 10 runs) using a ResNet-50 trained on the trainval split.