Automated monitoring of early stage human embryonic cells in time-lapse microscopy images

Abstract

This thesis focuses on automated monitoring of human embryonic cells in time-lapse microscopy images of early stage developing embryos. Our primary biological motivation is to develop an automated system that would assist embryologist to study and analyze the dynamic behavior of developing embryos in an attempt to improve in vitro fertilisation (IVF) outcomes. However, all methods proposed in this thesis are applicable to a wide range of microscopy cellular image analysis applications. Automated analysis tasks involving cellular structures, in general, present significant difficulties (e.g., topological change and deformable objects). These difficulties are even more acute in the context of microscopy images of human embryonic cells. The individual cells in the developing embryos form a complex 3D structure, which, in a 2D projection, overlap immensely. We tackle these difficulties within a principled probabilistic framework and propose methods that can reliably and efficiently analyse growing embryos in a fully automated manner. An important and first step in automated analysis is being able to efficiently and reliably segment the embryo from background clutter. To this end, we propose a framework to segment the developing embryo by estimating the contour around the embryo. We formulate segmentation as an energy minimization problem and solved it efficiently via graph cuts. Next, we propose frameworks to spatially localize embryonic cells and temporally detect their divisions. Predicting the number of cells is a fundamental task in cell biology analysis. In the context of human embryonic cells its importance is prime as current embryo viability biomarkers require accurate cells counts. The number of cells prediction can either be performed directly from the microscopy images or by detecting (localizing) cells. In this thesis, we employ both approaches and propose frameworks that combine both approaches in a conditional random field (CRF) framework. For localization, we model cells as ellipses and derive a data-dependent state space for each time step by applying an ellipse-fitting algorithm with a spatially diverse sampling procedure. We also propose a framework that models the cell division ancestry as a lineage tree. Cell lineage analysis is vital in understanding dynamics of developing embryos and is a fundamental step in cell biology analysis. Our approach generates a lineage tree by measuring cell associations between adjacent frames. We further analyze lineage by annotating the tree with various attributes of the growing embryo such as cell cleavage, abnormal division pattern and blastomeres (cell) symmetry. Our frameworks compactly encode rich contextual and visual information captured by hand crafted features, priors and constraints, which we designed specifically for human embryonic cells. Finally, to further enrich these frameworks we propose a deep learning architecture to count the number of cells directly from the microscopy images. We then incorporate this in our frameworks for cell detection, localization and lineage generation. We evaluate our models against the state-of-the-art methods related to the human embryonic cell automated analysis.