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

• Applicants: Professor Dr.-Ing. Marc Aubreville; Professor Dr. Robert Klopfleisch
• Subject Area: Pathology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 520330054

Neoplasia is a common cause of death in humans and animals alike. Cell division rate (number of mitoses) is an important measure of tumor proliferation and has a strong correlation with tumor patient outcome. However, mitotic counting is subject to considerable inter-rater bias. We have shown that computer-assisted mitotic counting using artificial intelligence (deep learning) methods can significantly improve reproducibility and accuracy. However, deep learning models are susceptible to changes in tissue processing and image acquisition, as well as differences between tumor types, also known as domain shift. This domain shift complicates the application of the algorithms in different clinical settings. The main goal of this research project is to generate a large multi-domain dataset with unprecedented data quality and quantity. This dataset will be made publicly available and thus serve as a reference for method development and solution validation of algorithms and further as a learning platform for pathology education. Together with diverse pathology laboratories, we are able to generate a multi-scanner, multi-laboratory, multi-species, and multi-tumor type dataset that is representative of numerous clinical settings (first project goal). In contrast to previous approaches, our interdisciplinary team will not only rely on expert decisions on hematoxylin and eosin-stained images to generate the dataset, but complementarily on immunohistochemical staining for mitoses (phospho-histone H3 on de-stained specimen) and supported by machine learning. For this, we will improve our advanced methods of dual-stain image registration and annotation (second project goal). Different image registration methods (depth of field, resolution, image planes) will be performed for a subset of samples to determine the impact on mitosis detection by algorithms and pathologists (third project goal). Furthermore, it is planned to explore methods for robust mitosis detection under a variety of imaging conditions, incorporating methods for uncertainty detection of AI solutions (fourth project goal). Finally, the impact of the developed algorithms on diagnostic mitosis counting will be investigated (fifth project goal). On the one hand, the influence of a computer-assisted diagnosis system on expert decision-making will be examined, and on the other hand, the prognostic informativeness of the computer-assisted diagnosis system for a human and a canine tumor type will be explored.

• DFG Programme: Research Grants
• International Connection: Austria
• Partner Organisation: Fonds zur Förderung der wissenschaftlichen Forschung (FWF)
• Cooperation Partners: Dr. Christof Albert Bertram; Christopher Kaltenecker, Ph.D.