Generalized automatic classification of cancer cell migratory patterns

Abstract: Automated fluorescence microscopy is an emerging technique that enables researchers to generate large amounts of data which in turn allows for large scale quantitative studies. In previous small scale quantitative studies, migration modes found by qualitatively observing migrating lung cancer cells have been described. Since such a method starts from qualitative observations it might not generalize to other modes of migration or cells. Furthermore it is not obvious that this approach captures the full complexity of the system. In order to perform a fully quantitative study of cancer cell migration modes large data sets are needed which requires a method that allows us to detect cells of interest in low magnification. This way, larger samples can be scanned in an achievable time frame and high magnification data collected only on relevant cells. The aim of this thesis project was therefore to create a pipeline that extracts features from large data sets of low magnification live fluorescence imaging data and uses machine learning to cluster the cells into subpopulations that could be of interest for further study. Additionally, it would be of interest to see if such a method could recover the migration modes found in the previously mentioned small scale quantitative studies. To run the pipeline a server and a database for storing metadata and analysis results were set up. I developed scripts for automated analyses such as cell tracking and extraction of morphology data. For the purpose of interaction with the data, a web server serving both a graphical and a programmatic interface was created. The created infrastructure was successfully used for post-acquisition analysis. The clustering methods that I have evaluated tend to cluster dead cells. Further analysis of the other clusters and more data acquired by real-time classification is required for a generalized automatic method.