High Dimensional Indexing in Medical Image Databases

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): NHANES II, which is maintained by the National Library of Medicine, contains a database of 17,000 cervical and lumbar spine x-ray images. The aim of this proposal is to create graphical retrieval mechanisms for NHANES II so that images may be retrieved on the basis of anterior osteophyte severity and malalignment. Two retrieval mechanisms are proposed: a graphical query mechanism, where an example image is used to indicate the osteophyte severity and malalignment, and a graphical category mechanism, where osteophyte severity and malalignment are categorized using expert classification and the categories are used for retrieval. Graphical queries for osteophyte severity and malalignment require retrieval by similarity of shape. Shape belongs to high-dimensional shape spaces and indexing for retrieval in such spaces is an open problem. The P.I. has developed a mathematical framework for indexing in such spaces. This framework is unique in that it guarantees that the indexing tree that is best adapted to the given data distribution in high dimensional spaces will be found. The extension of this framework and its application to NHANES H is one of the main goals of the proposed research. Graphical categories for osteophyte severity and malalignment can also be constructed from shape indexing trees. The idea is to approximate the category by a union of node covers in the indexing tree. Such a category creation mechanism with controllable error rates is proposed and its application to NHANES II is suggested. Expert classification available from NHANES H and from the co-investigator (who has expertise in interpreting spine x-rays) will be used to train the graphical category. Validation for graphical queries and categories with NHANES II images is proposed. Validation will be carried out with the help of the co-investigator who is an expert in this domain.