In an attempt to reduce radiologist uncertainty in identifying suspicious pulmonary nodules, a team of university researchers has created an open source content-based CT image retrieval framework. "Our system provides a way of performing a look-up on a query image to return similar images from a collection," said Michael O. Lam, a University of Maryland computer science graduate student. Lam is one of a team of software engineers and imaging researchers led by Daniela Stan Raicu, Ph.D., an assistant professor of computer science at DePaul University, and University of Northwestern radiologist Dr. David S. Channin. When presented with a nodule image, the system, called BRISC (a recursive acronym for Content-Based Image Retrieval System), retrieves images of similar nodules from a collection maintained by the National Cancer Institute's Lung Image Database Consortium.

BRISC functions in a process with four steps:

• extract images of individual nodules from the LIDC collection based on LIDC expert annotations
• store the extracted data in a flat XML database
• calculate a set of quantitative descriptors for each nodule that provide a high-level characterization of its texture
• use various measures to determine the similarity of two nodules and perform queries on a selected query nodule

"The precision of image retrieval can be very high, so this technique has the potential to be useful as an adjunct to radiologist decision making in the context of pulmonary nodules in CT images," Lam said.

The researchers compared three feature extraction methods: Haralick co-occurrence, Gabor filters, and Markov random fields. "Gabor and Markov descriptors perform better at retrieving similar nodules than do Haralick co-occurrence techniques, with best retrieval precisions in excess of 88%," Lam said. The researchers speculate the reason for this disparity is that the co-occurrence method encodes texture information at the global (image) level, whereas both Gabor and Markov are calculated at the local (pixel) level, allowing for more robust comparison. Their results show that as the number of experts in agreement increases so does the precision of the retrieval, Lam said. "This supports the hypothesis that as experts agree on the nature of a lesion, the computable descriptors of the lesion become more homogeneous," he said. He cautions, however, that features computed by humans and by the software to make similarity decisions are not the same. This is an active are of research. BRISC software is available free online. Other details of the system can be found in the Journal of Digital Imaging (2007;20 Suppl 1:63-71).