A Google-like approach to mammograms

Knowledge-based computer-assisted detection (KB-CAD) systems, which are increasingly being used in clinical settings, compare a mammogram image with images in a database of known cases of breast cancer. The results of the comparison are then used by a radiologist to aid in the diagnosis. Although they are diagnostically accurate, detailed comparisons become increasingly inefficient as image databases grow. At the recent meeting of the American Association of Physicists in Medicine (AAPM), held in Orlando, Florida, Georgia Tourassi (Duke University) presented a way to speed up the process using the information-theoretic idea of image entropy. An image that is all black or all white has zero entropy; a complex image with a more uniform distribution of different pixel intensity levels has higher entropy. Image entropy can be easily and quickly determined, and that’s what the Duke system does first. In a way similar to an internet search engine, the new KB-CAD system first returns a list of database images whose entropies are similar to that of the one under study, then does a detailed comparison on only that subset of images. The database of Tourassi and her colleagues uses images of both normal and cancerous tissue for comparisons. In their tests, a database of 2318 images was quickly whittled down to only 600, and achieved the same performance as the standard KB-CAD method in one-fourth the time. The researchers expect to follow up their pilot study with a larger clinical investigation. (AAPM Meeting Paper TU-D-330A-8.)