dc.contributor.advisorChellappa, Ramaen_US
dc.contributor.authorNandy, Kaustaven_US
dc.contributor.departmentComputer Scienceen_US
dc.contributor.publisherDigital Repository at the University of Marylanden_US
dc.contributor.publisherUniversity of Maryland (College Park, Md.)en_US
dc.description.abstractRecent advances in the field of optical microscopy have enabled scientists to observe and image complex biological processes across a wide range of spatial and temporal resolution, resulting in an exponential increase in optical microscopy data. Manual analysis of such large volumes of data is extremely time consuming and often impossible if the changes cannot be detected by the human eye. Naturally it is essential to design robust, accurate and high performance image processing and analysis tools to extract biologically significant results. Furthermore, the presentation of the results to the end-user, post analysis, is also an equally challenging issue, especially when the data (and/or the hypothesis) involves several spatial/hierarchical scales (e.g., tissues, cells, (sub)-nuclear components). This dissertation concentrates on a subset of such problems such as robust edge detection, automatic nuclear segmentation and selection in multi-dimensional tissue images, spatial analysis of gene localization within the cell nucleus, information visualization and the development of a computational framework for efficient and high-throughput processing of large datasets. Initially, we have developed 2D nuclear segmentation and selection algorithms which help in the development of an integrated approach for determining the preferential spatial localization of certain genes within the cell nuclei which is emerging as a promising technique for the diagnosis of breast cancer. Quantification requires accurate segmentation of 100 to 200 cell nuclei in each patient tissue sample in order to draw a statistically significant result. Thus, for large scale analysis involving hundreds of patients, manual processing is too time consuming and subjective. We have developed an integrated workflow that selects, following 2D automatic segmentation, a sub-population of accurately delineated nuclei for positioning of fluorescence in situ hybridization labeled genes of interest in tissue samples. Application of the method was demonstrated for discriminating normal and cancerous breast tissue sections based on the differential positioning of the HES5 gene. Automatic results agreed with manual analysis in 11 out of 14 cancers, all 4 normal cases and all 5 non-cancerous breast disease cases, thus showing the accuracy and robustness of the proposed approach. As a natural progression from the 2D analysis algorithms to 3D, we first developed a robust and accurate probabilistic edge detection method for 3D tissue samples since several down stream analysis procedures such as segmentation and tracking rely on the performance of edge detection. The method based on multiscale and multi-orientation steps surpasses several other conventional edge detectors in terms of its performance. Subsequently, given an appropriate edge measure, we developed an optimal graphcut-based 3D nuclear segmentation technique for samples where the cell nuclei are volume or surface labeled. It poses the problem as one of finding minimal closure in a directed graph and solves it efficiently using the maxflow-mincut algorithm. Both interactive and automatic versions of the algorithm are developed. The algorithm outperforms, in terms of three metrics that are commonly used to evaluate segmentation algorithms, a recently reported geodesic distance transform-based 3D nuclear segmentation method which in turns was reported to outperform several other popular tools that segment 3D nuclei in tissue samples. Finally, to apply some of the aforementioned methods to large microscopic datasets, we have developed a user friendly computing environment called MiPipeline which supports high throughput data analysis, data and process provenance, visual programming and seamlessly integrated information visualization of hierarchical biological data. The computational part of the environment is based on LONI Pipeline distributed computing server and the interactive information visualization makes use of several javascript based libraries to visualize an XML-based backbone file populated with essential meta-data and results.en_US
dc.subject.pqcontrolledComputer scienceen_US
dc.subject.pquncontrolledBioimage informaticsen_US
dc.subject.pquncontrolledCancer Detectionen_US
dc.subject.pquncontrolledOptical Microscopyen_US


Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
4.47 MB
Adobe Portable Document Format