Microscopy video analysis for breast cancer research: an automatic pipeline for image processing and cell tracking

To analyze the progression of a disease, the response to a treatment, or the effects of drugs, it is essential to study the dynamic behavior of cells and the morphological changes they may undergo over time. For this reason, the analysis of microscopy videos is a fundamental practice in biology and medicine. This thesis work has been developed in collaboration with the REAP project, an initiative of nine multidisciplinary European teams dedicated to developing advanced imaging tools for studying Drug Tolerant Persister cells (DTP) in breast cancer. The most common type of cancer  worldwide and the primary cause of cancer-related deaths in women is breast cancer. In some cases, patients become resistant to medications and treatments due to DTP cells, which often survive therapy and can cause metastasis and disease recurrence. As a result, these cells have become an essential focus for research on disease recurrence prevention. The Medical University of Vienna used Incucyte system, a device that combines microscopy imaging with cell incubation, to acquire the images analyzed in this thesis. Specifically, phase contrast and fluorescence microscopy are used to obtain the images. To study the cells dynamic behaviour from microscopy videos, it is necessary to track their movements and morphological changes over time. The gold standard for cell tracking is manual tracking, which is error-prone, slow, and highly operator-dependent. Moreover, microscopy videos may produce large amounts of images and data that are difficult and laborious to analyze manually. For these reasons, in recent years more and more automated image processing, machine learning, deep learning, and computer vision techniques have been introduced to the microscopy video analysis pipeline. These techniques enable faster and more reliable extraction of information, image processing, and cell tracking. This thesis is divided into two main parts. The first focuses on preparing the dataset, identifying and correcting errors in the frames to ensure accurate tracking. The creation of a gold standard through an image study allowed for the identification of five types of errors. Across the entire dataset of approximately 8100 images, 1064 images contain at least one of these errors. An algorithm was developed that can correct more than 85% of errors (around 900 images), although it also introduced approximately 17% false positives and false negatives relative to the total detected errors. In the second part of the work the frames were then divided into two groups: pre- and post-treatment. Through Voronoi segmentation and tracking with TrackMate, an attempt was made to compare these groups and extract different features for their characterization. The last part presents the results obtained and the information extracted about the cells, as well as possible improvements and future work.