Novel search engine targets radiology data in archives

April 30, 2003

Images searched rather than textQualitative characteristics, embedded in digital images and squirreled away on archives, could be ripe pickings for a search engine being developed by engineers at Washington University in St. Louis.

Images searched rather than text

Qualitative characteristics, embedded in digital images and squirreled away on archives, could be ripe pickings for a search engine being developed by engineers at Washington University in St. Louis. The researchers claim the method could eventually search image archives for certain anatomic or pathological properties, such as specific types of fractures or vascular features.

"This technique is not restricted to flat text or genomic data," said Ronald Indeck, Ph.D., director of the Center for Security Technologies in the electrical engineering department at Washington University in St. Louis. "The data can be radiological images."

This project has grown from efforts at Washington University to develop advanced signal processing techniques for increasing the capacity of magnetic recording systems. Preliminary research conducted using the resulting technology has demonstrated its potential for searching medical images, according to Indeck. Encouraging results have been found when examining medical databases, such as the American College of Radiology Imaging Network, which holds images obtained during multi-institution clinical trials related to cancer.

The technology evolved initially to assess and report on flat text, such as genetic sequences or written documents. Its speed-200 times faster than generally available search engines used on the Internet-and its ability to search unstructured data, however, provide the basis for migrating this technology to radiology.

Searches can and will become increasingly complex. Soon the system will be able to assess chest images, for example, for specific parameters, such as the presence of a solitary pulmonary nodule, Indeck said. Accomplishing such a search is currently beyond the scope of commercially available technologies.

"Today images are tagged with metadata, which are extremely limited," he said. "These data might pertain to pathology, or the person's name, but once you're through the metadata, you're done. You can't search anymore."

Conventional search technology indexes data, he said, and medical images do not lend themselves to such indexing. Even if a means for developing such an index could be found, the index would require too much memory to process effectively.

The system being developed at Washington University, however, works quite differently. The St. Louis engineers have applied so-called reconfigurable hardware, which makes use of existing computing components but puts them to work in novel ways.

Indeck recruits the high-speed parallel magnetic sensing systems already present in modern magnetic storage devices to facilitate searches. With these, he and his colleagues search databases directly, without processor, memory, or bandwidth limitations. Instead of indexing data in memory, data are searched where they reside in the archive. The results improve the speed and reduce the cost of performing approximate matches within large spaces by several orders of magnitude.

The applications for such a search engine include the analysis of archived data for epidemiological purposes and public health trends. Teaching files might be rapidly assembled on specific topics. There may also be routine medical applications, such as the assembly of reference images against which suspicious but inconclusive pathologies can be compared for qualitative analysis.