Novel computer vision algorithms for automated cell event detection and analysis

Live cell imaging is the study of living cells using microscope images and is used by biomedical researchers to provide a novel way to analyze biological functions through cell behavior and motion studies. Cell events are seen as morphological changes in image sequences, and their analysis has great potential for the study of normal/abnormal phenotypes and the effectiveness of drugs. While current quantitative cell analysis typically focuses on measuring whole populations of cells, we need to be able to recognize cell events at the single cell level, identify these events automatically, and analyze these events over time. For this reason, we developed and evaluated several novel automatic single cell event detection and analysis methods based on a detailed knowledge of the cell cycle and other cell event characteristics. The first method detects significant events within the temporal sequence using a machine learning method to use features derived from segmented cell images. We used a Neural Network (NN) algorithm to classify cell events to pre-defined categories. The second and third methods apply statistical and econometric techniques originally developed for time-series analysis of financial markets to facilitate the identification of cell entry into mitosis. We developed graph trend analysis and paired graph analysis methods from trend analysis and pairs trading to determine significant data points in cell feature data. The final method determines the position of cells in order to associate daughter cells with their parent cells after mitosis using Kalman filter techniques. By using the Kalman filter approach, we estimated future cell border centroid positions and successfully associated daughter cells with their parent cells after mitosis. In this study, the performance of these novel computer vision algorithms for automatic cell event detection and analysis were evaluated and verified by applying models to different image sequences from the Large Scale Digital Cell Analysis System (LSDCAS). The results show that the approaches developed can yield significant improvements over existing algorithms.