MPI-NAT Scientific Seminar: Democratizing Deep Learning for Microscopy Image Analysis

Most cutting edge image analysis for microscopy rely on deep learning. While these methods have dramatically improved the quality for challenging tasks such as denoising, cell segmentation or tracking, they are less accessible than previous approaches; hampering broad adoption. They require large datasets with (manually annotated) examples to train deep neural networks. Creating such training data is a significant manual effort. Furthermore, most networks do not generalize: They only perform well for data from a modality similar to the training set and cannot be directly reused if experimental settings change. Even if a suitable network has already been published, the lack of standardized formats and software makes it challenging to apply it.
I will present approaches to overcome these hurdles and democratize access to deep learning for microscopy image analysis: To address the issue of training data, my group is developing domain adaptation methods. These methods can adapt models trained for a certain task (e.g. cell segmentation) on a domain with annotated data to a new domain where no annotations exist. I will show their application to analysis problems in livecell imaging, electron microscopy and more. Furthermore, I will present bioimage.io, a community effort with the goal to make models available in a common format that is supported by popular image analysis tools such as Fiji, ilastik, Icy and others.