Data mining digs deeper

Facilities are generating vast amounts of data every year, and collecting and storing all of that information isn't easy. To help meet the challenge of data mining, new visual information systems in medicine (VISIM) are emerging that will search and retrieve items from large image collections.

Many hospitals easily generate 1 TB or more of digital images per year in the acquisition of radiographs, ultrasounds, CT, and MR scans. In addition, PET scans and genomic and other research pictures will soon begin to flood image databases.

"When images are archived, they are typically stored with a limited text-based description of their content," said Dr. Hemant Tagare, an assistant professor of diagnostic radiology at Yale University. "Such a description inherently fails to quantify the properties of an image and refers to an extremely small fraction of its information content."

Several papers presented at the VISIM workshop held in the Netherlands last October offered novel content-based image retrieval (CBIR) ideas. Computer scientist Tatjana Zrimec, Ph.D., of the University of Sydney in Australia described a prototype medical image information system that enables quick and efficient image access based on semantic content.

"The system makes use of intermediate levels of descriptions in images and metalevel information about relations among images," Zrimec said.

The organization of the images enables a content-based search to be applied hierarchically, resulting in a major increase in retrieval speed.

"In contrast to systems that perform image retrieval using only image features, our system uses a combination of text-based and content-based image retrieval," Zrimec said.

Another approach, called IRMA for image retrieval in medical applications, combines high query completion on medical image materials with small user interaction.

"This is made possible by a structured and successive abstraction and image interpretation," said Dr. Thomas Lehmann, leader of the medical image processing department at Aachen University in Germany. "In contrast to other approaches, the IRMA concept is truly general and not restricted to a certain modality or diagnostic question."

Previously, textual index entries had been mandatory to retrieve medical images from PACS or other sources, Lehmann said. Information contained in medical images, however, differs considerably from that residing in alphanumeric format.

"Although all CBIR systems provide nontextual indexing, the IRMA approach is specially designed to handle primitive and semantic queries as well as browsing of medical images," Lehmann said.