PUB - Publikationen an der Universität Bielefeld

This dissertation delves into the utilization of microfluidic systems for single-cell analysis within controlled environmental conditions, presenting novel prospects for both fundamental and applied biotechnology. Yet, the intricate nature of data resulting from microfluidics experiments renders manual assessment unfeasible, while conventional image processing techniques prove inadequate. Consequently, deep learning methodologies, such as convolutional networks, present a versatile solution, encompassing automated cell quantification. Nevertheless, procuring the essential label information for supervised learning proves to be a resource-intensive and time-consuming endeavor. To confront this concern, innovative machine learning architectures and specialized training procedures of different learning paradigms are introduced, facilitating the creation of highperforming regression models with minimal labeled data, and in certain instances, even in the absence of labeled data entirely. Through the training of generative models on both genuine and synthetic data concurrently, the architectures learn shared representations capable of accurately estimating a target variable, the cell count. Furthermore, transfer learning strategies built upon deep networks trained on limited image collections demonstrate significant success in the domain of single-cell cultivation within microfluidics experiments. Nevertheless, acquiring even a limited number of labeled examples might prove unfeasible in specific biotechnological applications, compounded by data traits that could markedly differ from existing domains, thereby engendering challenges in transfer learning. To surmount this obstacle, unsupervised learning methodologies have been explored, encompassing a VAE-based Siamese architecture that is training in a cyclic fashion. Facilitated by shared embeddings of differing data types, this approach permits the generation of pseudonatural images from synthetic counterparts, simulating the existence of labeled natural data and furnishing dependable estimates for multiple microscopy technologies and previously unencountered datasets without necessitating manual labels. These pioneering machine learning techniques and architectures are agnostic to platforms, openly available, and freely accessible, presenting a promising avenue for the analysis of microfluidics data and single-cell cultivation experiments within biotechnology and other realms of image processing research.