Digital Holographic Microscopy for High-throughput Analysis of Tumour Cells

visible, and in time-lapse imaging, light levels can be harmful for cells. These routines can affect the behaviour of cells and introduce unwanted artefacts in the captured images. Digital holographic microscopy(DHM)enables stain-free single-shot imaging of living cells, uses low light intensities, and requires low data storage capacity. Due to advantages over many other methods, DHM is a strong candidate for high-throughput analyses. Use of DHM for imaging complex three-dimensional organoids has not been reported before. The major limitation preventing widespread-use of DHM is the lack of analysis algorithms and limitations with those that are available. There are no algorithms to detect the presence of objects encoded in digital holograms. Existing algorithms used to find the in-focus depths of objects encoded in digital holograms require multiple reconstructions that increase processing time making them unsuitable for high-throughput analyses. There are no algorithms that could be used to segment digital hologram reconstructions of biological objects to multiple distinctive regions. To the best of our knowledge, using digital hologram reconstructions for classification of cysts has not been reported before. This thesis introduces novel approaches for efficient analyses of holograms of celllines and real patient samples. By analysing and interpreting different features of organoids, cancer-specific signatures are identified. In this thesis, a number of novel contributions are reported. A model-based object presence detection approach exploiting information extracted from a CNN is reported. CNNs are trained to find in-focus depths of organoids encoded in digital holograms without any numerical propagation. Reconstructions from digital holograms of organoids are segmented to multiple discrete regions using a CNN, allowing novel quantitative analyses. Different classifiers using either extracted feature vectors or phase reconstructions are trained to discriminate healthy and tumorigenic organoids. A large-scale experiment is conducted for finding a CNN model with sufficient classification accuracy and minimum number of learning parameters;hand-crafted features are added to these shallow networks to improve the classification accuracy. Organoids derived from tissue samples of thirteen prostate cancer patients are shown to introduce additional challenges. Based on the existing data,there is an indication that prostate cancer is unique for each patient thus complicating detection of cancer.