October 9 - 12, 2012
Building 10 & Natcher Conference Center
- General Schedule of Events
- Opening Plenary Session
- Concurrent Symposia Sessions
- Poster Sessions
- National Graduate Student Research Conference
- FARE Award Ceremony
- Special Exhibits on Resources for Intramural Research
- TSA Research Festival Exhibit Show
- Clinical Center Tours
- Research Festival Committees
- Past Research Festivals
An Interactive Region-Of-interest (ROI)-Based Image Retrieval Approach of Biomedical Articles in a Local Concept-Based Feature Space
| Friday, November 08, 2013 — Poster Session IV | |||
|---|---|---|---|
2:00 p.m. – 4:00 p.m. |
FAES Academic Center (Upper-Level Terrace) |
NLM |
COMPBIO-13 |
Authors
- M. M. Rahman
- S.K. Antani
- D Demner-Fushman
- G.R. Thoma
Abstract
Authors of biomedical publications frequently use images to illustrate various medical concepts and highlight special cases as a region-of-interest (ROI) within the image. Users of state-of-the-art image retrieval systems are often seeking images that are similar with respect to ROI, but are limited to similarity of the entire image. At the National Library of Medicine, we are developing an image retrieval system that usues a local concept-based feature representation scheme to retrieve images containing patterns similar to an interactively marked ROI on a query image. Here, “concept” refers to perceptually distinguishable image patches obtained from a codebook, i.e. a glossary of visual swatches that correspond to biomedical imaging terms. Our classification methods are capable of automatically mapping the appearance of visual entities within the selected query image ROI those in the codebook, and then use these concepts to search images indexed with similar concepts. In addition, spatial layout information is introduced as a post-processing step to re-rank the retrieved images. The retrieval effectiveness is validated through experiments on a dataset of 450 lung CT images appeared in journal articles from four different collections. We achieved 10% improvement in precision compared to search on entire images without using any spatial information.
