Content-based retrieval for PACS arrives

PACS should provide a more intelligent retrieval mechanism - one based on image and video content - according to some experts.

While PACS already allow physicians to search traditional SQL-based image databases using textual keys such as patient name, medical record number, and study date, a content-based retrieval architecture, or COBRA, would be valuable for several reasons.

A sophisticated COBRA could be used as a second opinion diagnosis tool, said Dr. Essam El-Kwae, an assistant professor in the department of computer science at University of North Carolina, Charlotte. El-Kwae has developed a prototype content-based retrieval system for MR brain images.

The ability to compare images will allow physicians to retrieve cases with similar etiology and review the diagnosis, treatment plan, and outcomes, according to El-Kwae.

Content-based retrieval could also be used to:

?collect statistics about specific image features, such as the size of a certain heart chamber in a particular age range;
?help train residents;
?prepare conference presentations by searching for particular cases; and
?apply image data-mining techniques.

Although still in its infancy, data mining allows pearls of information to be extracted from large databases.

"The idea of the COBRA is to suggest an architecture that can allow existing PACS to incorporate retrieval by content, based on medical and technological standards, without having to rebuild the system from scratch," El-Kwae said.

Within El-Kwae's prototype, an anatomy classification algorithm is used to automatically classify PACS studies based on their anatomy. Such a classification allows the use of different segmentation and image-processing algorithms for different anatomies. His COBRA uses primitive retrieval criteria such as color, texture, and shape and more complex criteria including object-based spatial relations and regions of interest.

The COBRA, which El-Kwae hopes to implement in collaboration with some Charlotte-area hospitals, features two major components: information extraction and querying.

Some information can be extracted from images while they are being inserted into the database. At the same time the system is extracting the DICOM header, for example, another part of the system can extract the raw pixel values; segment the image; and get the color, texture, shape, and spatial information from the image - all of which is then saved in a logical database.

The query component then allows the user to interrogate the logical database employing techniques that search by image example, sketch, region of interest, characteristics, or spatial relations.