Intelligent Interfaces for Interactive Analysis of High-Content Cellular Images

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Cell-based High-Content Screening (HCS) has recently led to high-throughput image-based studies of cellular phenotypes under various external treatments such as chemical compound or or RNA interference (RNAi). Such studies will significantly advance our understanding of gene functions, shed new light on the underlying biological networks, and have direct impact on cancer research and drug discovery/development. However, due to the inadequacies of existing image analysis tools, most HCS screens only relied on analyses of simple marker readouts and left the most informative and profound aspects of cellular morphology unexplored. Domain knowledge is yet to be accumulated for developing image analysis tools to effectively and thoroughly analyze highly diverse cellular images generated by the HCS technology, which are relatively new to image processing research. Nonetheless, building up domain knowledge requires human experts to visually explore a prohibitively large number of images. Therefore, it calls for a new computing paradigm that facilitates teamwork between experimental and computational biologists to overcome this dilemma. We propose to develop a novel computing paradigm that integrates unsupervised pattern mining techniques, visual data exploration interfaces and content-based image retrieval with relevance feedback techniques to facilitate the application of the HCS technology to biomedical research. This paradigm will be realized as a system called imCellPhen, which will be evalutated and tested in the context of two morphological screens of Drosophila neurodisease models using the HCS technology. The main features of imCellPhen are its intelligent interfaces that allow users to (a) effectively and efficiently navigate large-scale HCS image databases, (b) reliably detect novel cellular phenotypes, and (c) teach the system to recognize cellular phenotypes by interactively training computational models. The model training procedure is in fact an implicit, seamless, and effective process for accumulating domain knowledge. The scheme and techniques developed in this research will benefit any HCS screens and thus will be valuable tools for the biomedical research community.