High resolution embryonic images, e.g., the data set BDGP (Berkley Drosophila Genome Project), have been introduced as an important tool for the discovery of gene-gene interaction. These images contain not only temporal information of a gene but also precise spatial information of expression regions of genes. So the biologic problem of the discovery of gene-gene interaction can be characterized as a computational problem of matching expression patterns of embryos at the same developmental stage. It is, however, very challenging to design a fully automatic computational system due to severe imaging and artificial variations in embryonic images. Current research on embryonic image processing involves significant manual manipulation or addresses only a small subset of variations. In this proposal, I propose a comprehensive automatic framework to achieve the three fundamental tasks: image standardization, stage determination, and expression pattern modeling. The proposed project will essentially advance the integration of biologic, image processing and pattern recognition, and machine learning. The PI will develop a series of analysis modules to standardize the variation across images, provide for inpainting, and provide estimates of the boundaries of embryos. The expression pattern modeling will develop discriminate features to address issues of distinguishing specific pixels in varied refraction circumstances. A key concept is to develop an imbalance point detection scheme that will minimize the occurrences of edge points and provide a measure of the imbalance degree.
These methods should be adaptable for analysis of images from other model species, e.g., mouse. The proposed work will directly facilitate basic and applied research on image processing crucial to biological image analysis. The challenges in analyzing developmental biological images as compared to natural images have created increasing demands on and opportunities for developing novel image processing techniques. The algorithms and tools developed in this project will have made available to the community. This project will also facilitate the development of new courses and laboratory infrastructure for knowledge discovery from biological data.
Drosophila is a representative model to discover the nature of animals, and the study of gene expression pattern of Drosophila is helpful to discover mechanisms of organ growth. The development of high-throughput imaging technology helped biologists collect a large number of images that contain rich information of gene expression patterns of Drosophila, and arose the necessity of computational methodologies to analyze data quantitatively and efficiently. The methods developed in this project advance computational methodologies of Drosophila embryonic data in the following two major aspects: 1) improve the successful rate of automatic contour extraction of a Drosophila embryo in images and ii) improve the accuracy of the localization of gene expression regions. The developed methods can be conveniently adapted to analyze images of other model species, e.g., mouse. Besides advancing biological image processing methods, the project also advances theories on how to optimize parameters of algorithms that are very common challenges in applications of computer vision. In this project, four WKU undergraduate students participated algorithm implementation for contour extraction or gene expression region extraction. One of them pursued a Master degree with a thesis research on Drosophila image processing under my supervision. There are totally three graduate students who successfully completed their thesis research on Drosophila image processing. A graduate course on Computer Vision (CS 568) was proposed, and results on Drosophila image processing were also engaged as a homework topic in this course.
