**The dataset** contains 256x256 tiles extracted from Whole Slide Images (WSI) of stained tissue that was introduced in in [Zingman et al., "Learning image representations for anomaly detection: application to discovery of histological alterations in drug development, 2022.](https://arxiv.org/abs/2210.07675)
**The train** subfolder contains tiles of normal tissue samples used to train image representations on an auxiliary classification task. It contains tiles from 16 categories. Each category is a combination of specie (mouse or rat), organ (liver, brain, lung, heart, pancreas, spleen, kidney), and staining (H&E and Masson's Trichrome). Thre are around 6920 tiles for each category.
Liver tissue was also used to train one-class classifier.
**Put attention**, the data in the train subfolder was not manually curated; the tiles were automatically extracted from WSI, which may contain various artifacts.
**The test** subfolder contains the data used to evaluate performance of anomaly detection (AD) method described in the paper, compare it to other AD methods, and perform ablation studies. The data contain normal mouse liver tissue image tiles, 2170 and 2390 for H&E and Masson's Trichrome stained tissue respectively and tissue with Non-Alcoholic Fatty Liver Disease (NAFLD), 2150 and 2372 for H&E and Masson Trichrome stained tissue respectively.
The data labels (Normal versus abnormal NAFLD) in the test subfolder were manually verified by an experienced pathologist.
**The trained models** folder contains trained Boehringer Ingelheim Histological Network (BIHN) PyTorch models (for Masson's Trichrome and H&E images), pickled training configuration files, and pickled One-class SVM classifiers (scikit-learn).
**The code** that reproduces the paper results and runs these models can be found here [https://github.com/Boehringer-Ingelheim/anomaly-detection-in-histology][1]
[1]: https://github.com/Boehringer-Ingelheim/anomaly-detection-in-histology
