Neural network classifies medical images
Reading and analyzing images from magnetic resonance imaging (MRI) scans, Pap smears from optical scans, and tumors from x-rays, and searching for suspicious structures are time-consuming processes. In mammography, for example, images of 4k ¥ 5k ¥ 12 bits are often digitized and analyzed. To analyze such images, researchers at the Kurt Rossman Laboratories for Radiologic Image Research at the University of Chicago (Chicago, IL) have developed an automated scheme to help radiologists classify clustered microcalcifications.
In operation, the classification scheme extracts the thickness, volume, area, shape, and number of individual microcalcifications in the cluster. These features are then used as input to an artificial neural network whose output is the probability that the cluster is malignant. The neural network "learns" the difference between benign and malignant clusters from training on a large number of known cases. According to researchers, both the computerized method and the radiologists correctly identified all the malignant cases. However, the radiologists misclassified 75% of the benign cases, but the computer scheme misclassified just 25% of the benign cases.
To commercialize this technology, R2 Technology (Los Altos, CA) has developed the ImageChecker system, a system based on algorithms the company has licensed from the University of Chicago. Using a 64-processor, CNAPS/PCI board from Adaptive Solutions (Beaverton, OR) to accelerate the neural-network classification algorithm, the system allows 4k ¥ 5k ¥ 12-bit mammograms to be scanned in 14 s. Portions of the algorithm are accelerated as much as 40 times faster than on a 200-MHz Pentium computer, claims James Pell, president of R2 Technology.