Content-based image retrieval breaks new ground

November 11, 2004

A new strategy for image retrieval that combines a central database with a distributed system architecture promises to make content-based schemes the preferred method for handling large data efficiently.

A new strategy for image retrieval that combines a central database with a distributed system architecture promises to make content-based schemes the preferred method for handling large data efficiently.

Experts tend to agree that the future of efficient query processing of large image databases lies in content-based, rather than text-based, retrieval. Current content-based image retrieval schemes, however, are task-specific, meaning they are limited to a particular modality, organ, or diagnostic study, according to Thomas Lehmann, Ph.D., of the Institut fur Medizinische Informatik in Aachen, Germany.

In a recent paper, Lehmann presented an alternative content-based approach for medical images combining a database with a distributed system architecture suitable for large image databases, such as those housed within PACS (Methods Inf Med 2004;43:354-361).

This method, called content-based image retrieval for medical applications, or IRMA, is a general framework rather than a specific application. The technique enhances radiology in several ways, including case-based reasoning, Lehmann said.

"Suppose the radiologist detects an abnormal pattern in an image, but pathological diagnosis is difficult," he said. "Instead of consulting text or patient records selected by a certain pathology, IRMA presents images showing similar patterns and offers direct access to the corresponding patient records."

Similarly, for teaching, radiologists can select and retrieve images showing particular artifacts or pathologies by means of a particular pattern rather than textual description.

The scheme also supports rapid prototyping and quick integration of novel image analysis methods.

Compared with standard content-based systems, the IRMA approach employs at least three additional semantic levels of abstraction to cope with image retrieval in medical applications:

  • A low level of medical knowledge is determined by the modality, including technical parameters, orientation of the patient, examined body region, and the functional system under investigation. This information is extracted automatically comparing the current image to a database of more than 10,000 references coded by experienced radiologists.

  • A middle level of knowledge is described by regions of interest. These are extracted automatically by a hierarchical region merging process operating on local as well as global features.

  • A high level is obtained from information regarding the spatial or temporal relationships of relevant objects. Here, prototype images for each category are used to close the semantic gap between the machine-based low-level processing of pixels and the high-level image interpretation of humans.

Thus far, use of the new approach to image retrieval has been limited to answering primitive queries. However, these experiments have proven the validity and applicability of the concept, according to Lehmann.

"IRMA improves current PACS that still rely on alphanumerical descriptions, which are insufficient for image retrieval of high recall and precision," he said.

For more information from the Diagnostic Imaging archives:

Integration bolsters workstation design

Digital case file system collects interdisciplinary 'intellectual capital'

Price is right for stopgap PACS

Grid storage adds intelligence to data archiving task