A deep learning computer vision system for image based cytometry

Abstract

A Complete Blood Count (CBC) is one of the most commonly deployed tests for checking the overall medical condition of a human body. It results in a summary of blood cell related measures allowing for a quick preliminary detection of common diseases. Image based cytometry is a modern way of overcoming common efficiency issues related to current gold standard method of Coulter Counter analysis. This thesis describes a computer vision pipeline for measuring main components of CBC based on methods with varied level of human influence in feature design. The low quality input images are acquired from a prototype lensless microscope. Rule-based approaches proved sufficient in cases of red blood cells and platelets analysis and their configuration is explored in more details at the segmentation stage. Leukocyte differentiation is a subproblem of CBC requiring more sophisticated classification approach. The comparison of results based on rule-based, classical and deep learning methods is presented. Support Vector Machines and k-Nearest Neighbours were chosen as representatives of classical approaches. The deep learning scheme has the advantage of finding quick solutions to complex problems involving highly dimensional and hierarchical structure of underlying solutions. It also alleviates the amount of work needed by human experts to design appropriate descriptors. The results discussed here include promising enumeration and feature description of the objects of interest. Their linear dependence with gold standard clinical results proved their usability for point of care blood analysis. Classification results showed superiority of convolutional neural networks in several important aspects. Further analysis explored the learned parameters and analyzed wrongly classified cases in order to understand better what kind of cytological objects are most fragile for particular classifier and why.