Improving our understanding of machine learning robustness – in microscopy and beyond

A central goal when applying machine learning in the biomedical research and medical imaging is to answer relevant interdisciplinary questions robustly and reliably across various setups and ideally modalities; however, this is challenged by different imaging systems, differences between a laboratory setting and the real world as well as a multitude of other factors. This talk will explore the challenge of domain generalization in digital pathology, microscopy, and beyond. Specifically, we will discuss the Mitosis Domain Generalization (MIDOG) Challenge that took place at the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in 2021 and 2022 as one example how we can work toward more robust machine learning approaches. Furthermore, we will explore how we can potentially transfer the corresponding machine learning insights from imaging very tiny structures close-by to structures far, far away.