Software that compares a patient’s mammogram to those in a database has become a valuable tool to help physicians more reliably detect breast cancer. However, as the size of these databases grows, the time it takes to evaluate a new image increases.
Now researchers from Duke University have developed a method for sorting through thousands of mammogram images, and picking out the most informative ones, in a matter of seconds.
Spotting a problem in mammograms can be difficult. For one thing, cancerous lesions can blend into the background, making them difficult for doctors to pick out. Software-based detection systems assist by comparing a new mammogram to a database of already analyzed ones, flagging cancer indicators that a physician might have missed.
The first such software-based system was approved by the Food and Drug Administration in 1998. Currently, more than 1,600 such systems are used in clinical practice in the United States. While the systems have been shown to increase the rate of breast cancer detection, they’re still far from perfect, and researchers are constantly trying to improve their accuracy, says Georgia Tourassi, professor of radiology at Duke University and lead developer of the new software.
Newer “knowledge-based” software programs, such as the ones used by the Duke researchers, allow doctors to interact with the system – asking questions and getting answers based on previous cancer cases in the database. However, churning through all the images in a knowledge-based system, which may contain thousands of mammograms, is time consuming. “As the knowledge database increases in size, this brute force method doesn’t make sense anymore,” says Tourassi.
To speed up the image-searching process, Tourassi and her team divided it into two steps. First they looked for the most useful mammograms in the database by using “image entropy” – the amount of grayscale variation in the pixels – on suspected mammographic regions. An all-black or all-white image has zero entropy, while more complex images have higher degrees of entropy, produced by patches of pixels with varying intensity. These high-entropy cases occur around images of cancerous lesions and are the most useful in evaluating new mammograms.
Comparing the entropy in images is a particularly attractive tactic, says Tourassi, because these values are automatically calculated for mammograms when they’re submitted into the Duke database. Thus, no extra image-processing computation is needed for the technique.
In a pilot study, the Duke researchers showed that by comparing the image entropy of a suspicious region in a new mammogram to the entropy of all known cancerous regions in the database, they were able to slash the number of analyzed mammograms from about 2,300 to 600.
From there, says Tourassi, a more fine-grain analysis is used to compare the region in question with the known regions in database images. Since the system has to fully process only around 600 images, the computation effort is reduced by 75 percent, and the search can be done in seconds. Their results were presented this week at the annual meeting of the American Association of Physicists in Medicine in Orlando, FL.
In addition to speeding up the search process, the Duke technique could also improve the detection rate of cancerous lesions, says Maryellen Giger, professor of radiology at the University of Chicago. Image entropy searches, in particular, are well suited to detecting lesions. Current systems have an 80 percent accuracy rate in detecting this type of cancer indicator; Giger says the Duke technique could significantly improve that rate. “It’s very promising,” says Giger.
Within a year, the Duke researchers will launch a study to evaluate the clinical impact of their new technique.