Robust and Accurate Multi-Tumor, Multi-Species, Multi-Laboratory and Multi-Scanner Mitosis Detection with Large-Scale Datasets and Artificial Intelligence (mu ROMI)

Neoplasms are one of the most common causes of death in humans and animals. The decision for appropriate therapy is based, among other things, on the histological examination of the tumor samples, capturing various prognostic parameters. One of the most relevant histological parameters for assessing the prognosis of tumors is the number of mitotic figures (Mitotic Count). However, various studies have shown that manual measurement by a pathologist are associated with high inter- and intra-observer variability. Therefore, in recent years, computerized measurement methods using artificial intelligence (especially deep learning-based approaches) have gained great interest. In various studies of our research group, we have shown how accurate and reproducible (compared to the pathologist) these algorithms can be when adequate datasets are used to train the deep learning models. However, these models extract the relevant features of the patterns of interest based on the training data, and the "decision criteria" of the algorithms are often very specific to the training dataset. Since digital images of histological tumor preparations often have a very large variance in image properties (so-called "domains", for example, based on the tumor type examined, the staining protocol applied, and the whole slide image scanner used to digitize the samples), algorithms are mostly not applicable to datasets of different domains due to a so-called "domain shift" if these domains were not considered during training. The consequence is that algorithms, which must be produced with great effort, cannot be used in different laboratories or for different tumor types.
 

This research project has the primary goal of creating a large dataset for mitotic figures in histological tumor preparations, including a large number of different domains. Established methods of database generation are used, which enable maximum quality of labels and an efficient workflow. A fundamental approach is the combination of immunohistochemical staining (with antibodies against Phosphohistone H3) as a decision aid for mitotic figures in histological specimens stained routine with Hematoxylin and Eosin. This dataset is intended to serve scientists and diagnostic laboratories as a reference corpus for the development of algorithms. Furthermore, it is a goal of our own research project to find methodological approaches that improve the transferability of algorithms between the numerous domains and thus enable a broad applicability of the algorithms (as computer-assisted prognosis). Finally, a learning platform for pathologists will be developed, which should provide practical exercise opportunities for proper mitosis recognition and the use of image analysis algorithms as decision support in the histological assessment of tumor samples.

Project staff:
Ass.-Prof. Dr. Christof Bertram

Cooperation partners:
Prof. Dr. Marc Aubreville
Prof. Dr. Robert Klopfleisch 
Ing. Christopher Kaltenecker, PhD
MedUni Vienna - Digital Pathology

Project period: 04/2024 – 03/2027

Funding:
Austrian Science Fund (FWF) with the project number: I 6555