Our dataset for flatfish images provides lesion bounding boxes and pathogen labels (presence and type), and includes generated images alongside the originals. This enables unified lesion detection, pathogen-type classification, and studies on data augmentation and domain generalization within a single corpus.
💬 Comments
(last updated 25-10-11)
This also looks like a lesion
This was reviewed by a human. If you believe it was mis-annotated (e.g., you think it is a lesion), please feel free to contact us anytime.
When a fish is taken out of water, its gills naturally flare; this is not a lesion. ↓
Occasionally unnatural backgrounds
Because the data came from real aquaculture sites, it contained a lot of company/identifying information, which we removed.
Backgrounds were removed using a generation model.
Data volume
We added more normal (non-pathological) data.
We also expanded the annotations (e.g., additional fields/items).
Planned data additions
We have access to more data provided by partner companies. We plan to add more normal data and broaden the dataset so it can support a wider range of studies.
We will also add segmentation annotations.
Additionally, we plan to explore and include more diverse generation approaches (e.g., trying different generation models and settings).
🖇️ Citation
@article{hwang2025flatfish,
title={Flatfish lesion detection based on part segmentation approach and lesion image generation},
author={Hwang, Seo-Bin and Kim, Han-Young and Heo, Chae-Yeon and Jeong, Hie-Yong and Jung, Sung-Ju and Cho, Yeong-Jun},
journal={Journal of the World Aquaculture Society},
volume={56},
number={3},
pages={e70031},
year={2025},
publisher={Wiley Online Library} }