Spies Vs. Breast Cancer
THE PICTURES BEGIN AS STREAMS OF DIGI tal data beamed from space down to receiving stations on the ground. Computers transI late the data into images on high-resolution screens for analysis by highly trained CIA and military photo interpreters. It’s a synergy of satellites, computers, and human eyes and minds that has been honed to a keen edge since the relatively primitive early days of photoreconnaissance. Now this technology is confronting a new enemy that kills about 46,000 Americans every year: breast cancer.
The leading cause of death among women of ages 35 to 54 in the United States, breast cancer can be insidious in avoiding detection until it has threatened a life by spreading to other organs. Just as with other forms of cancer, the earlier it’s caught, the greater the chance of a complete cure. A powerful weapon against breast cancer is screening mammography. The National Cancer Institute says screening mammography can reduce the breast cancer mortality rate by up to 30 percent, potentially saving thousands of lives a year.
Yet studies have shown that perhaps as many as 30 percent of women with breast cancer have supposedly normal screening exams. The cancer is there, but it’s so subtle that even the experienced eyes of a radiologist can’t pick it up. Along with the danger of false negatives, mammography can also result in false positives, suspicious findings that later turn out to be perfectly innocent. In its earliest stages, breast cancer often manifests itself in the form of microcalcifications, tiny deposits of calcium that may be smaller than a tenth of a millimeter. However, overlapping normal tissue can mimic microcalcifications, and microcalcifications can be benign. Either way, a careful radiologist will generally err on the side of caution and order further imaging studies or even a biopsy. More often than not the results are benign, but the cost in time, money, and especially patient anxiety can be high.
Although X rays were discovered by Wilhelm Roentgen more than a century ago and quickly established their usefulness in medicine, mammography is a relatively recent development. Only in the 1950s did doctors realize that breast tumors could be found on X rays, and not until the 1960s and 1970s did clinical studies prove that mammography could save lives, bringing the technology into general use. It has improved since the earliest days, but it is still essentially little more than an X-ray picture, with all the limitations that this implies: the resolution and contrast range of film, a susceptibility to over- and underexposure, and the possibility that tumors can be missed because part of the breast is hidden or poorly positioned.
An additional problem is the nature of the screening process itself. “The task of screening for breast cancer can be likened to looking for 5 needles in 1,000 haystacks,” explained Dr. Robert M. Nishikawa and Dr. Robert A. Schmidt, radiologists at the University of Chicago, in a 1994 paper. “It is a demanding and fatiguing job.… It is difficult for human observers to consistently maintain the level of attention that this requires on a daily basis.” Dr. Mitchell Schnall, of the University of Pennsylvania Hospitals, agrees: “You sit in front of a computer workstation or a reading board looking at films all day, there’s no human around who’s not going to miss something.” Having a different doctor read a mammogram independently may improve cancer detection rates by 5 to 15 percent, but double-reading is a staffing and logistical nightmare, especially with the steadily increasing volume of screening studies performed each year. There are simply too many patients and too few doctors for most hospitals and clinics to afford the luxury of reading routine studies more than once.
Aside from the needle-in-a-haystack challenge, radiologists examining the complexities of breast architecture are dealing with a moving target. Kathryn Evers, M.D., director of mammography at Fox Chase Cancer Center, near Philadelphia, explains: “Each woman’s breast tissue is different and varies for the same woman over time or even at different times of the month.” As Nishikawa and Schmidt noted, “No other medical imaging technique demands such fine resolution as mammography.”
Concerns such as these led Dr. Susan Blumenthal, deputy assistant secretary for women’s health and assistant surgeon general in the Department of Health and Human Services under President Clinton, to organize an effort in the mid-1990s to find emerging technologies developed by government agencies that might have medical applications. Blumenthal found a gold mine in the Department of Defense, particularly the National Reconnaissance Office (NRO), which for most of its life has been one of the most secret agencies in the intelligence community (its very existence wasn’t publicly acknowledged by the U.S. government until September 1992).
Under the aegis of the NRO is the National Information Display Laboratory, which resides at the Sarnoff Corporation, in Princeton. Established in 1990, NIDL facilitates the transfer of advanced imaging techniques between government agencies and the private sector. Under Blumenthal’s urging, NIDL began a collaboration with the National Cancer Institute and various medical research groups across the country to develop defense technologies for medical purposes. More study groups, alliances, and coalitions among doctors, scientists, and Department of Defense personnel followed, most bearing unwieldy labels such as the NDMDG (National Digital Mammography Development Group) of the National Cancer Institute. With government and corporate funding, a major effort to bring breast imaging into the twenty-first century began.
WHILE EVERYONE INVOLVED HAD strong professional motivations for wanting to improve the state of the mammographie art, the quest was more personal for some, including Steve Rogers, a former Air Force officer with 20 years of experience in military intelligence, smart weapons, and sophisticated image analysis. Rogers was startled at the relatively primitive realities of medical imaging when his mother developed breast cancer in the early 1990s. “The radiologists put films on a light box and looked at them with a magnifying glass,” he remembers. “That’s where we had been many years ago in the spysatellite business. I thought to myself, this is not acceptable.”
Even before Blumenthal began her crusade, scientists had started to develop digital mammography to replace the traditional film X-ray method. Fully digital mammographie imaging is expensive and has yet to be widely adopted, but it promises images of much higher image quality and resolution. Still, Rogers and his colleagues in the medical and defense community knew that digital technology could do more than just take pictures. Enter the spies.
As it happens, looking for the telltale signs of a new secret chemical weapons factory and searching for the elusive changes in breast patterns that herald cancer aren’t all that different. Both involve something called serial change detection, which basically means identifying the presence of new objects in an image by removing (or ignoring) old ones. It’s a sophisticated version of the children’s puzzle of looking at what’s different between two pictures. This was one of several techniques demonstrated by NIDL at a 1994 symposium organized by Blumenthal that was attended by a highly diverse cadre of doctors, scientists, and engineers, including Robert Nishikawa. One particularly interesting possibility involved the advanced computer algorithms designed to find subtle changes in spy satellite images. “They had technology they developed to look for objects, tanks or buildings, which would show up as bright objects on the image,” Nishikawa says. “It looked at the context. For example, buildings are going to be near roads. So it basically analyzes the background, the texture of the image, looking for things that might be roads, and whether those roads are associated with bright spots that are buildings. The idea was that in mammograms, calcifications may be associated with some linear structure in the breast that could be used to look for calcifications.”
The intelligence community’s serialchange detection techniques were combined with artificial intelligence programs for image analysis to create a powerful new medical tool, computer-aided detection, interchangeably and somewhat confusingly also known as computer-aided diagnosis. CAD employs highly sophisticated computer algorithms to analyze radiographie images, giving the radiologist the equivalent of a highly trained and very sharp-eyed second reader. As the 1990s progressed, successful clinical studies of CAD led several companies to develop commercial systems for sale to hospitals and clinics.
This alliance of spy technology and medicine isn’t as unlikely as it might seem. In an address to imaging professionals at the Rochester Institute of Technology, Jeffrey K. Harris, a former NRO director, explained what intelligence and medical professionals have in common. “Both of us gather information, analyze, evaluate, and process data under conditions of uncertainty, looking for clues to difficult problems, and the penalties for being wrong are very high—especially for false negatives.”
One challenge in adapting the technology was that while man-made objects, once recognized, are fairly unchanging in appearance, the biological phenomenon of cancer can vary greatly. “Finding Scuds in the desert was relatively easy compared to finding cancer in breasts,” says Rogers, whose CADx Medical Systems of Beavercreek, Ohio, developed one of the first PDA-approved commercial CAD mammography systems. “A Scud is a Scud, but there are lots of manifestations of cancer.” CAD algorithms have to be sensitive to all the nuances of possible cancerous appearance, from large calcified masses to slight changes in tissue density. To achieve this, the CAD software incorporates a set of definitive past cases. Rogers explains: “What you do is go out and acquire a bunch of examples, thousands of cases, of women who have had screening and diagnostic mammograms, and in some subset of that data you have women in which a human found cancer in the images.” The CAD algorithms are trained on these test studies, and as they process further cases, their knowledge base continues to grow and becomes more discriminating. The systems have the capacity to learn from experience, so that their performance improves over time.
First, the mammograms have to be converted from physical film images into digital data sets. This is done by a digital scanner into which the doctor inserts the films, just as you might scan a photograph to post it on your Web page. A digitizer for a CAD system offers higher resolution than a personal scanner, however, breaking down an image into pixel sizes as tiny as 40 microns.
Then the CAD algorithms go to work. Employing a suite of different imageprocessing techniques with various artificial intelligence tools, the computer may juggle a billion operations per case (a typical screening mammogram consists of four films, two views of each breast). Despite its sophistication, the CAD software doesn’t require enormous computer power, no more than a good desktop system. Processing by the CAD program takes only a few minutes. When it’s done, the CAD system generates a printed or onscreen report, noting areas on the mammogram that might warrant closer evaluation. These will be marked on the image—with an asterisk or a box or circle around them—and the radiologist can then decide whether these flagged abnormalities warrant further examination. If a patient has older studies available, the radiologist will compare them with the new exam and thus can tell whether the computer has found something new and potentially dangerous.
Commercially available CAD mammography systems are self-contained console units about the size of a washer/dryer that consist of a digitizing scanner, a computer, proprietary software, an input device, such as a keyboard or touch-sensitive screen, for entering patient information, and some kind of output equipment, either a printer or display screens. The heart of the technology is the software, the algorithms that scrutinize images (or rather, their digitized equivalents) for minute alterations and irregularities. First, imageprocessing software breaks down an image into smaller parts, organizing them according to similarities and differences in their component features. Then the CAD system decides what features of the image may warrant further investigation, drawing on its database of known signs of malignancy. Artificial neural networks that simulate the operation of nerve cells work on different parts of a problem simultaneously, comparing various possibilities and outcomes from multiple inputs. (Perhaps the most famous demonstration of this multitasking power of neural networks was the defeat of the chess master Carry Kasparov by IBM’s Deep Blue computer.) Coupled with fuzzy logic, which can judge shades of difference according to context, these neural networks and other artificial intelligence techniques are the key to CAD.
Unlike a doctor studying a CT scan, an ultrasound image, or even a simple chest X ray, who is hunting for many different kinds of disease, a radiologist looking at a mammogram is pretty much seeing only one thing. Because breast cancer most frequently manifests itself by alterations in breast tissue, it’s a perfect quarry for a computer system dedicated to hunting for subtle pattern changes.
A PROBLEM ARISES WHEN CAD IS too sensitive, flagging too many false positives. Yet one doesn’t want a system to be too cavalier in its detection procedures. Current work in perfecting CAD is chiefly focused on solving the false-positive problem. As Robert Nishikawa explains, “If there are too many false positives in the image, the radiologist won’t use the technology, because it’s basically just saying to look at the whole image again.” As the computer algorithms become even more sophisticated, such difficulties are expected to decline.
Most radiologists, however, don’t mind a few extra false positives. Steve Rogers says, “I actually had a radiologist tell me recently that he wanted more false positives. He wanted a dial on the CAD where he could crank it up so it would never miss a cancer. He said, ‘That’s what I want, because I can discount your false positives.’” Kathryn Evers agrees: “If you’re going to have a couple of extra false positives, it’s certainly preferable to a couple of extra false negatives.”
In any event, patients shouldn’t fear that their cancer might be missed because of a computer error. No matter how advanced the intelligence imagery of the CIA and the military, human analysts still make the final interpretations, just as no President would declare war simply because a computer thought it had found a suspicious clue on a satellite picture. “The computer is a second opinion for the radiologist,” says Nishikawa. “It doesn’t have to be perfect, but just point out suspicious areas to look at again.” Dr. Timothy Freer, director of the Women’s Diagnostic and Breast Health Center in Piano, Texas, who conducted the first large-scale clinical trial of CAD for screening mammography, is adamant that “the radiologist must be the final arbiter on the merits of any finding, regardless of how it was detected.” The skill, judgment, and instincts of the human observer, acquired and finely honed over years of training and experience, far surpass the abilities of the computer, though the computer can complement and augment those unique human qualities. Rogers sums it up this way: “We spent a lot of time in the government trying to do AI, artificial intelligence, where you try to make a computer replace a person. What I finally concluded after many years of doing that is that a more important application for computing is IA, intelligence amplification, helping the human make a more intelligent decision. What’s a more intelligent radiologist? If you don’t increase the workup rate significantly yet you find statistically more cancer, you’ve made a more intelligent radiologist.”
Might the day come when computers read screening mammograms and other routine medical studies exclusively? Mitchell Schnall, for one, doubts it. “No one is bold enough to say that yet,” he states. “The number of false positives is enough that you need human intervention.” Still, there’s a definite consensus that CAD not only is here to stay but will become nearly universal as its algorithms become even more advanced. Steve Rogers says, “In my opinion it will become the standard of care. It has proven too valuable not to be included in the processing of every screening mammogram.”
One factor smoothing the way for its widespread adoption is the rapid advance of digital mammography technology. As digital mammography drops in cost and eventually replaces the classic technique in most hospitals, CAD will naturally follow. Government reimbursement programs for hospitals using digital mammography and CAD systems are proving a strong incentive for smaller hospitals with lower budgets to buy into the technology. At present, two major companies have FDA approval to market their proprietary CAD mammography systems, R2 Technology of Sunnyvale, California, and Rogers’s CADx Systems. Both firms have already sold hundreds of systems worldwide, and more companies are entering the field. Some of them are developing CAD systems for other imaging modalities, such as CT and MRI. These will be able to use other intelligencecommunity technology, such as the ability to visualize and compare three-dimensional patterns.
Much as digital technology has taken over such diverse fields as filmmaking, publishing, music, and photography, its eventual prevalence in medical technology is all but inevitable. In the form of CAD, it’s an example of a sword beaten into a lifesaving plowshare by some highly imaginative and dedicated people. As Timothy Freer says, “If future studies confirm improved detection, CAD will prove to be a true technological leap, perhaps the greatest single advance in breast cancer detection in the last 20 years.” After all the billions of dollars spent during the Cold War on ever newer and better weapons, it’s not such a bad legacy for the spooks.
