Welcome to the QuiltCleaner repository. We labeled 1% of the QUILT_1M dataset for common image impurities that would be deteriorating image generation in a text-conditional image synthesis setting. We provide predictions for the remaining 99% of the QUILT_1M dataset. Additionally, we provide scores for text-image alignment as provided by the CONCH vision-language model.
Paper (accepted for MIDL 2024): Aubreville et al: Model-based Cleaning of the QUILT-1M Pathology Dataset for Text-Conditional Image Synthesis
These annotations and predictions are complimenting the QUILT_1M dataset (Ikezogwo et al., NeurIPS 2023). Please look at their repository as to how to retrieve it.
We annotated a random 1% sample of the QUILT-1M dataset for the following categories:
- Narrator/person: Image with a visible person within the image
- Desktop/Window decorations/Slide viewer: Image where window decorators (e.g., close button, maximize button, title bar) or a complete slide viewer application, or even the windows task bar/desktop can be seen.
- Text/logo: Image contains visible text on top of the main pathology image, or one or multiple logos.
- Image of insufficient quality: Images in this category are mostly non-histology images (endoscopy, macro images, radiology images, only text images) but partially also just images with severe quality issues (blank image, severe blurryness).
- Additional slide overview: A slide overview (thumbnail) is shown on top of the actual slide image.
- Additional buttons/control elements: Additional control elements such as +/- buttons or rotation control are shown on top of the actual slide.
- Multi-panel image: Images consisting of multiple sub-images (panels). These are tricky for text-conditional image synthesis as the text might only describe a single (non-specified) panel.
The annotations are provided in the following files:
- train_annotations.csv Training set (70%)
- val_annotations.csv Validation set (15%, used for model selection)
- test_annotations.csv Hold out test set (15%)
You will need to download the QUILT-1M dataset separately, as this can not be provided in this repository due to licensing reasons. Place all files that were annotated into the images folder of this repository. Then, you will be able to train your own QuiltCleaner using the provided notebook.
@inproceedings{aubreville2024modelbased,
title={Model-based Cleaning of the QUILT-1M Pathology Dataset for Text-Conditional Image Synthesis},
author={Marc Aubreville and Jonathan Ganz and Jonas Ammeling and Christopher C. Kaltenecker and Christof A. Bertram},
booktitle={Medical Imaging with Deep Learning},
url={https://openreview.net/forum?id=m7wYKrUjzV},
year={2024},
eprint={2404.07676},
archivePrefix={arXiv},
}