Automated Deep Learning Models for the Analysis of Biological Microscopy Images

Recent advances in artificial intelligence (AI) and deep learning are having a tremendous impact on diverse application domains. It would be surprising if the domain of digital microscopy in biology would not benefit from this. This thesis addresses problems that are encountered when analyzing microscopic images with current methods from artificial intelligence, in particular detecting, segmenting and classifying cells, organelles, etc., in microscopic images. It appears that specialized knowledge is needed to ensure success with AI. For example, current AI only works well with a massive amount of samples, for which the desired answer is given. Individual cells would need to be outlined using the mouse and obtain a label (annotation) in the process, for each new type of biomedical research goal. This is too impractical. We present a method from AI that tries to circumvent this stumbling block: Self-supervised learning (SSL) to learn organoid segmentation. The data consists of microscopy images of organoid images provided by our academic hospital. We compared U-net/ResNet with our SSL and traditional supervised learning on labeled images. Surprisingly, the SSL results were even better than relying on the human-labeled samples only. Other problems addressed in this project concern the detection of overlapping objects in an organoid dataset and the classification of organelles within yeast cells using Mask-RNN and Yolo4. In order to help the biological researchers, an 'e-Science' website was developed where they can submit images for segmentation or classification. This website has been extended with other functions related to biological data analysis, e.g., the classification of FASTA codes for protein molecules.