File compression can expand mammography’s power

When it comes to the information in a mammogram, Purdue scientists say less is more – and their findings could bring medical care to many far-flung communities.

A team of researchers, including Bradley J. Lucier, has found that digitized mammograms, the X-ray cross sections of breast tissue that doctors use to search for cancer, are actually interpreted more accurately by radiologists once they have been “compressed” using techniques similar to those used to lessen the memory demand of images in digital cameras. Though compression strips away much of the original data, it still leaves intact those features that physicians need most to diagnose cancer effectively. Perhaps equally important, digitization could bring mammography to many outlying communities via mobile equipment and dial-up Internet connections.

“Any technique that improves the performance of radiologists is helpful, but this also means that mammograms can be taken in remote places that are underserved by the medical community,” said Lucier, who is a professor of mathematics and computer science in Purdue’s College of Science. “The mammograms can then be sent electronically to radiologists, who can read the digitized versions knowing they will do at least as well as the original mammograms.”

The research paper will appear in today’s (Dec. 20) issue of Radiology, the journal of the Radiological Society of North America. Lucier developed the file-compression method used in the study, which was run at the Moffitt Cancer Center at the University of South Florida in Tampa.

Discerning the potential seeds of cancer within the chaff of extraneous detail present in a mammogram requires the expert eye of a radiologist, who must pick out salient features at many different scales within the image. Clues can be very small clusters of tiny calcium deposits, each less than one-hundredth of an inch in diameter. Clues also can range up through the edges of medium-sized objects – which could be benign cysts with smooth edges, for example, or cancerous tumors with rough edges – up to large-scale patterns in tissue fiber.

“The edges of tumors are where growth occurs, and they tell radiologists whether what they see is a tumor or not,” Lucier said. “You have to keep all these features intact when you compress the image if it is to be useful.”

Once a mammogram image has been converted into electronic form, it can contain more than 50 megabytes of data, which makes it prohibitively large for transmission by computer modem over a telephone line. Compounding the issue is that four such images are needed for a complete screening, and though it takes only a few minutes to obtain the X-ray pictures, getting a mammogram can be difficult. A 2001 FDA study showed that the number of mammography facilities has declined in most states, and the population of potential recipients of mammography services has increased. While the study suggests that difficulties obtaining mammograms are localized rather than widespread, Lucier said that telemedicine could potentially mitigate the problem.

“I began experimenting with file-compression algorithms to see if we could shrink files to the point where they could be sent over standard phone lines,” he said. “Some communities do not have easy access to broadband Internet yet, and my colleagues and I wanted to work around that issue.”

Lucier found that one well-tested algorithm – a short set of instructions that can be repeated many times – did the trick after a bit of tweaking. Though the basic mathematics has been around for more than a decade, he said, its finer points required some adjusting.

“I wanted the algorithm to make all the features important to radiologists degrade at the same rate – both the edges of large tumors and the smallest calcium deposits,” Lucier said. “I tried several approaches and eventually got a balance that seemed reasonable, based on what radiologists tell me they want.”

His methods have evidently paid off: On seven of nine measures of diagnostic accuracy, radiologists interpret the compressed images more accurately than they interpret the original images, even though the compressed images contain, on average, only 2 percent of the information in the originals.

“I want to emphasize that this study does not necessarily imply that compression always improves diagnosis,” Lucier said. “It means that radiologists can spot and localize features as well or better than before. The technology filters out the noise, if you will. But so far, there is no question that these radiologists did diagnose better using the compressed images.”

Lucier is optimistic that the technique might be applied to other forms of telemedicine as well, if certain modifications are made.

“There are many forms of medical diagnosis that require an image to be read by a specialist,” he said. “If image compression is applied to other diagnostic situations, you won’t actually have to have that specialist on hand if you can get the equipment to the patient. But this is proof in principle that file compression, if done properly, can confer advantages to both patient and doctor.”

This research was funded in part by the Office of Naval Research, whose Mathematical, Computer and Information Sciences Division supports research on motion and still-image analysis, processing and enhancement.

Writer: Chad Boutin, (765) 494-2081, cboutin@purdue.edu

Source: Bradley J. Lucier, (612) 625-5532 or (765) 494-1979, lucier@math.purdue.edu

Purdue News Service: (765) 494-2096; purduenews@purdue.edu