Content-Based Image Retrieval for Medical Databases

点击翻译按钮获取中文摘要

Increasing computerization of the process of acquiring and storing diagnostic images requires more sophisticated approaches to image retrieval in order to optimally use the images for improved health-care delivery, physician education and research. A content-based image retrieval system is being developed that takes a human-in-the-loop approach, because completely automated approaches are not feasible for radiological images in which the clinically useful information may consist of gray level and texture deviations in highly localized but difficult-to-segment regions of an image. In this approach, a physician manually delineates the pathology-bearing regions in an image. The CBIR system then analyzes these regions for their properties and uses a graph representation, augmented with the spatial relationships of the pathology bearing regions to anatomical landmarks, to index the images into a database and retrieve similar images typically with known diagnoses. This project focuses on the domain of high resolution computed tomography (HRCT) images in patients with a variety of lung diseases. Regarding scientific goals, the project addresses: (1) the computational complexity of retrieval and its mitigation by the use of appropriate heuristics; (2) selection of similarity criteria to use for comparing the pathology-bearing regions of the query and indexed images; and (3) the design of a physician-friendly software architecture that allows for extension to other disease categories.