Deep learning methods for the analysis of biological electron microscopy images

The analysis of biological structures in electron microscopy (EM) images is an important task that can lead to a deeper understanding of cellular processes and disease development, as it is widely used when researching and diagnosing various diseases such as viral infections and genetic disorders. However, large-scale manual analysis and quantification of EM images is timeconsuming, subjective, and error-prone. Deep learning (DL) techniques have the potential to automate image analysis workflows for objective quantification of biological structures in EM images. Despite this, deep learning has not been fully explored in biological EM, as it often requires large amounts of labelled ground truth datasets. The availability of such data is rather scarce in the biological field, which in turn limits its applications. To overcome this challenge, this thesis focuses on developing DL methods that can classify, synthesize, detect, and segment biological structures in EM images in an effective manner using only small amounts of labelled ground truth data. The methods developed are intended to provide biologists with an automated quantification and analytical workflow for routine use. To this end, we leveraged transfer learning techniques to detect human cytomegalovirus (HCMV) particles in transmission electron microscopy (TEM) images. Two different transfer learning techniques were investigated to assess their effectiveness and suitability for this task. Our study provided the proof of principle that transfer learning can be applied to the development of DL models that allow for the automatic detection of particles in EM images. This work was then extended to detect and classify HCMV secondary capsid envelopment stages in TEM images. Since lack of large high-quality labeled ground truth datasets hampers model performance, we introduced a technique to generate synthetic TEM images and self-labelling data as an augmentation method. We could show, that the addition of synthetic data greatly improves the DL models' learning ...