That's the proposal in this study, to be presented in the scientific sessions on Monday by a team of researchers from Medical College of Wisconsin (MCW) in Milwaukee and Oregon Health and Science University (OHSU) in Portland. They believe that the use of visual information in conducting image searches can help researchers and educators sort and classify large databases of images.
The research group sought to evaluate a hierarchical approach to classify images from radiology journals using "textons" (putative elementary units of preattentive human texture perception), intensity histograms, and filter-convolution feature vectors.
In a study of 1,072 test images, the image-based system correctly classified 903 (84%) of the images. After combining the image-based classifier with text-based modality classifications from the American Roentgen Ray Society's (ARRS) GoldMiner image search engine, its performance improved to 95% (1,015 images accurately classified).
The results offer the promise of faster and more accurate filtering of search results in current resources such as GoldMiner, said co-author Dr. Charles Kahn, a professor of radiology and chief of informatics at MCW.
Also, the study will demonstrate the potential for creating tools to better sort images in teaching files or on the Web, as well as for improved techniques for recognizing other visual features in medical images, Kahn said. Dr. Jayashree Kalpathy-Cramer of OHSU will present the study.
The researchers will also present an educational exhibit on content-based image retrieval, including a description of the ImageCLEF 2009 image retrieval evaluation (part of the international Cross-Language Evaluation Forum (CLEF) research program. You can find all of the details by visiting the exhibit (LL-IN2585) in Lakeside Learning Center.
