WEST LAFAYETTE, Ind. -- Engineers at Purdue University have teamed up with medical experts to develop a computerized system designed to aid in disease diagnosis by matching a patient's CT scans with images in a large data base of scans from previous patients.
Because the images in the data base have been selected by highly trained physicians, the system can be used by less-skilled or less-specialized medical personnel as a tool for diagnosing patients, says Carla Brodley, a computer engineer at Purdue who is leading the research. She stresses that the system is not intended to replace the human element in diagnosis, but to improve the accuracy and speed of diagnosis by completing comprehensive image comparisons within seconds.
"This works through the synergy of human interaction and machine learning and computer vision algorithms," says Brodley, an assistant professor in the Purdue School of Electrical and Computer Engineering.
Sophisticated software enables the user to carry out a visual sort of keyword search, a method referred to as content-based image retrieval. The system views a patient's medical scan and then searches for visually similar images in a data base containing hundreds of CT scans that have been selected carefully by medical experts. The pre-selected scans in the data base represent known examples of specific diseases and conditions, and the four scans that best match the patient's image are retrieved.
The work will be detailed in a paper to appear in the July issue of the journal Computer Vision and Image Understanding, published by Academic Press Inc. A poster paper about the research will be presented June 25 during the 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition -- which is sponsored by the Institute of Electrical and Electronics Engineers -- in Fort Collins, Colo. A paper on the work also will be presented at 2 p.m. July 21 during the 16th National Conference on Artificial Intelligence, sponsored by the American Association for Artificial Intelligence, in Orlando, Fla.
The data base presently contains high-resolution computed tomography -- or CT -- images of human lungs. Data bases containing magnetic resonance images of the liver, knee and possibly the brain soon will be added, says Brodley, who is working with Avi Kak, a professor of electrical and computer engineering; graduate students Chi-Ren Shyu, Jennifer Dy, Mark Flick, Sean MacArthur and Christina Pavlopoulou; and medical doctors Alex Aisen from Indiana University Medical Center and Lynn Broderick from the University of Wisconsin Hospital.
Here is how the system works: A doctor starts out with the patient's diagnostic image, circles the diseased region with a mouse and sends the image as a query to the data base. Computer vision software determines the "mean intensity" of the image -- or how gray it is -- along with other features. That information is used to place the image into a mathematical vector that enables the computer to compare it to other images. The system retrieves the most visually similar images, any of which can be enlarged and further enhanced. Because a CT scan is taken in sections, specific contiguous "slices" within the diseased region can be isolated for more detailed scrutiny. The medical professional then reviews the treatment information of those previous patients, hopefully gaining insight into the current patient's situation.
The research is funded by the National Science Foundation and the National Institutes of Health.
NOTE TO JOURNALISTS: A CT scan image is available from Purdue News Service, 765-494-2096, or from our Web and ftp sites, http://news.uns.purdue.edu and ftp://ftp.purdue.edu/pub/uns/. The photo is named Brodley/CTscan.
STORY AND PHOTO CAN BE FOUND AT:
The large CT image of a patient's lungs at the top left is the query image, which can be compared with the four most visually similar images of lungs at the bottom of the picture. The large image on the top right is an enlargement of the best match.
Photo ID: Brodley.CTscan