Staff Presentations

This page contains information about presentations made by staff. Some of these presentations are available as PPT/PDF upon request. Contact here.

Clinical DICOM image visualization for Interactive Publications

Tian Shen, Glenn Ford, Sameer Antani, Dina Demner-Fushman, George Thoma

National Library of Medicine, National Institutes of Health, Bethesda, MD 20894
Presentation given at NLM on August 15, 2008.

Abstract

An interactive publication (IP) is a self-contained multimedia document that enables reader control over media objects and reuse of media content for further analysis. Biomedical publications often contain lots of clinical images such as CT, MRI and ultrasound converted from DICOM to PNG, JPEG or BMP format, since none of the traditional publication viewers, such as Adobe Acrobat and Microsoft Word, provides native support for DICOM. We have developed a module for the IP viewer Panorama by using Java Eclipse RCP and ImageJ, which allow the reader to interact with and analyze DICOM clinical images. The user can zoom in/out, change contrast, and cycle through an opened stack of DICOM images. Furthermore, our module provides access to DICOM header information, which is hidden in a traditional publication. The module also allows a user to observe the 2D projection of the stack from three orthogonal directions and to render a 3D volume. The user can observe the volume from arbitrary viewpoints by doing a rotation, and see the detailed information by scaling.

Algorithm Developed to Make More Reliable Links to Better Photographic Views of Nursing Homes for Prototype "Nursing Home Screener" Web Site

Wen Cheng and Glenn Pearson

National Library of Medicine, National Institutes of Health, Bethesda, MD 20894
Presentation given at NLM on August 15, 2008.

Abstract

To facilitate screening American nursing homes by location and quality, we are developing a website that uniquely shows multiple homes as rank markers on a Google Map. We want also to see photos of the homes, ones that are more informative than provided by the standard three Views. Our prototype offers a hyperlink to Microsoft Maps Live from each home marker, so that "bird's-eye" views (i.e., moderately close-up "45 degree " aerial views) can be shown if available.

Such linking is often unsuccessful in reaching a home's image, due to format mismatch of either the home name or address between the two mapping systems. We have developed a preprocessing algorithm that recognizes such a failure and attempts to overcome it. If a reliable link is found, it then detects whether a bird's-eye view is available. (If not, the hyperlink is suppressible.)

Four types of unsuccessful links are detected: "Blank", "Pink Box", "Yellow Box", and "Multiple Addresses". We present the nature of each failure type, our solutions, and the degree of improvement found. How we determine bird's-eye availability is described, and the coverage of bird's-eyes for all US homes estimated. Finally, open issues and further research are discussed.

Towards Image and Text Integration - Understanding Image Caption

Emilia Apostolova, Sameer Antani, Dina Demner-Fushman, George Thoma

National Library of Medicine, National Institutes of Health, Bethesda, MD 20894
Presentation given at NLM on August 15, 2008.

Abstract

Biomedical images are one of the main tools in establishing diagnosis, facilitating training and research. A vast amount of the "image data" available is embedded in biomedical articles and thus not directly accessible. Our goal is to automatically annotate images extracted from scientific publications with respect to their usefulness for clinical decision support and instructional purposes.

In this study, text analysis was applied to the captions of 67,115 figures from biomedical articles with the goal of automatic image annotation. An algorithm was developed to segment multiple figure image captions (92% accuracy) and identify in the text three types of concepts (1) the object of the image, (2) findings demonstrated in the image, and (3) the medical device/diagnostic procedure used to produce the image. The concepts were mapped to the National Library of Medicine Unified Medical Language System terms and were correctly identified with an F-measure of 0.78. In addition, it was recognized that pointers (arrows, lines, letters, etc) are commonly used as an overlay in images and described in captions to emphasize the main image finding. An algorithm was developed to extract from the text such "image pointers" (F-measure of 0.81) together their referents.

Return to top Return to top