Datasets
We strongly believe in the importance of democratizing data access in order to drive advancements in our research domain. We therefore made available the following data sets:
- The CATCH dataset is a tumor tissue segmentation dataset, comprising of 350 whole slide images of seven different canine cutaneous tumors complemented by 12,424 polygon annotations for 13 histologic classes. [paper] [github] [data]
- The MITOS_WSI_CMC dataset is the currently largest dataset for breast cancer that is annotated on complete whole slide images. It contains 21 cases of canine breast cancer and was labeled as the consensus of three experts. [paper] [github] [data]
- The MITOS_WSI_CCMCT dataset is the largest dataset for mitosis on canine cutaneous mast cell tumor and the largest dataset with verified mitotic figures (around 44k cells). Each cell has been labeled as a consensus of three pathologists. [paper] [github] [data]
- The TUPAC16_AL dataset is a set of alternative labels for the TUPAC16 challenge auxiliary mitosis dataset. Our research paper shows that it’s more complete and has a lower level of label noise. [paper] [github]
- The EIPH dataset is an Inter-species cell detection – dataset aimed at detecting pulmonary hemosiderophages in equine, human and feline specimens. [paper] [github] [data]
- The pan-tumor T-lymphocyte detection dataset is an immunohistochemistry dataset stained positive for CD3. It comprises 92 ROIs from four tumor indications with cell-level annotations for CD3+ cells, tumor cells, and other cells. [paper] [data]
- The multi-scanner SCC dataset comprises a subset of the CATCH dataset digitized with five different scanning systems. The images provide local correspondences useful for domain shift experiments and WSI registration. [paper] [github] [data]
- The MIDOG++ dataset extends the MIDOG2021 and MIDOG2022 challenge datasets. It comprises 503 ROIs of seven different tumor types with variable morphologic, laboratory, and scanner origins with in total 11,937 annotated mitotic figures. [paper] [github] [data]
- The QuiltCleaner dataset is a representative selection of impurity annotations for images from the QUILT-1M dataset and complimented with predictions for the remainder of the dataset. [paper] [github]
Microscopy Machine Learning Approaches
Our mission is to not improve insignificantly over previous work but to define new and interesting approaches, leading to new insights in the field of microscopy and pathology. Some of our most recent work is:
- Jonas Ammeling: Attention-based Multiple Instance Learning for Survival Prediction on Lung Cancer Tissue Microarrays [paper]
- Jonathan Ganz: Deep Learning-based Automatic Assessment of AgNOR-scores in Histopathology Images [paper]
- Marc Aubreville: Deep Learning-based Subtyping of Atypical and Normal Mitoses using a Hierarchical Anchor- free Object Detector [paper]
- Frauke Wilm: Mind the Gap: Scanner-induced domain shifts pose challenges for representation learning in histopathology [paper]
Challenges
We organized the Mitosis Domain Generalization Challenge (MIDOG) on MICCAI in the years 2021 and 2022. For both challenges, we published a challenge report paper in the journal Medical Image Analysis:
There are some rumors ( 😉 ) that there might be a MIDOG 2025 – but this is still unconfirmed. Stay tuned.