Computer-aided diagnosis becomes a reality in mammography
For the past several decades, researchers in the medical imaging field have focused on bringing new imaging modalities to clinicians while improving the performance of existing systems. Whether it was introducing a new photon emission tomography system or increasing the sensitivity of phosphor screens in x-ray equipment, optical and imaging scientists have made their mark on the medical world by providing hardware that helps doctors spot disease and identify health problems.
Now, software designers are beginning to take the next step by introducing software improvements, enabling computers to help doctors make sense of an avalanche of noninvasive medical information.
Radiologists in particular have to look at hundreds of x rays, MRIs, and other images each day. Their work requires them to identify a handful of nuances out of hundreds of healthy images -- a job that taxes even the most disciplined mind.
To help solve some of the information overload that researchers have helped to create, dozens of universities around the world and a handful of companies are rushing to develop computer-aided diagnosis (CAD) systems. These systems work with the radiologists by either identifying features that might have been overlooked or by providing a roadmap of suspicious features for radiologists, making their efforts more efficient.
In June 1998, the U.S. Food and Drug Administration approved the first CAD system for mammograms, R2 Technology Inc.'s (Los Altos, CA) ImageChecker M1000 system (Figure 1). The system, which had previously received CE Mark certification from the European Union, can now be found in more than 30 institutions around the world.
The Class III device provides an extra line of defense against breast cancer by supporting radiologists during their examination of breast x rays. The device, which is not designed to be a stand-alone diagnostic tool, checks digitized mammograms for two significant features: microcalcification, which are small bright spots that resemble star constellations, and masses, which appear as larger white blobs in the greyscale image.
The radiology technician feeds the breast x-ray film into the M1000's laser digitizer. The resulting 12-bit greyscale, 6K X 6K-pixel digital image with 50-µm resolution is then sent into ImageChecker's PC-based central brain for processing. While this processing is underway, ImageChecker reads a bar code attached to each x ray identifying the patient, view (right or left breast, etc.), and other relevant information important to the hospital and physician. Attaching the bar code to the original film is not an additional step, explained Bob Wong, R2 Technology's Chairman, because hospitals regularly use bar-code stickers to identify everything from patient records to fluid samples.
The digital images are run through a neural network that uses digital signal processors (DSP) technology to analyze the images. Algorithms developed at the University of Chicago and Sandia National Laboratory, and refined by R2 Technology, search for relevant features in a two-tiered approach.
One half of the process searches for bright points of light that exceed a given threshold of intensity. These spots, which can be as small as 40 µm in diameter, represent microcalcifications. If three or more of these points are within an area approximately 2.5-mm wide, the computer will overlay a triangle on this area when the image is displayed for the physician, drawing his or her attention to the region of interest.
The other half of the search process searches for larger masses. These patches of brightness on a breast x ray can be several hundred microns across. Again the system identifies the blobs by looking for gradient changes in the intensity of pixels within the image, thereby identifying the outline of the mass.
Masses, each of which is unique in shape and size, come in different categories. Sometimes, masses will become spiculated. Spicules are bright spokes emanating from a central point that resemble a thin spider web. The more spicules present, the greater the chance that a given mass is cancerous. When a mass or spicule grouping meets certain criteria, ImageChecker places an asterisk over the area when the image is displayed for the radiologist.
Neural networks are crucial to this kind of analysis because each mass is different in shape. Also, spicules are unique and may not be associated with a mass at all. The changing nature of this organic tissue means the system cannot search using a standard model, as most machine vision systems do. The system has to learn, much as a radiologist does, to extrapolate the likelihood of cancer based on x-ray features. To accomplish this, neural networks are not just programmed, they learn. By evaluating hundreds of mammograms and correlating feature sizes and location with known cancer cases, these neural networks operate like a radiologist, generalizing the likelihood of cancer based on particular findings.
According to company officials, ImageChecker conducts approximately one billion operations per film. A complete evaluation of one set of four images per patient takes about four minutes.
Once the digital image is analyzed, the technician places the original film on an x-ray reader conveyor belt. The radiologist sits down at the reader, which is the second component of the ImageChecker system. Digital images with triangles and asterisks (Figure 2) overlaid on the images are passed through a Local Area Network when the reader is within 150 feet of the processor, or via modem for telemedicine applications where the two units are separated by more than 150 feet.
A pair of monitors mounted above the film reader display subsampled, lower-resolution digital images corresponding to the film on the reader; a second bar code reader on the motorized viewer ensures that the digital images are the same as the film on the reader.
After the radiologist looks at the film, he or she can glance up at the computer screen to see if the computer found additional suspicious features. According to Bob Wong, chairman and CEO of R2 Technologies, ImageChecker is not intended to replace radiologists, but only to support them. "During a standard viewing, a physician looks at 30 to 50 cases," Wong said. "A majority of these cases are pretty normal. Physicians actually find three to six cancers for every 1000 patients. We found that physicians missed about 20 percent of cancers that could have been detected. Our system has the potential to recover about 80 percent of what they missed."
Wong expects that CAD machines will become more important as digital imaging techniques gain commercial acceptance. R2 Technology has signed an agreement with GE Medical (Waukesha, WI), giving GE exclusive rights to ImageChecker in conjunction with their digital x-ray equipment. GE is undergoing clinical trials for its digital x-ray equipment in anticipation of FDA approval. Elsewhere, companies such as Sterling Diagnostic Imaging (Greenville, SC) and Siemens (New York, NY) are also developing digital x-ray equipment for breast screening and other clinical applications.
According to Wong, R2 has spent more than five years developing ImageChecker. With the expertise that has come with this development, he hopes that ImageChecker will expand to other applications such as chest x rays and computer-aided tomography (CAT) screenings.
Kunio Doi, professor at the University of Chicago and one of the developers of the R2 algorithms, believes there is incredible potential for CAD systems. "Today, the systems identify microcalcifications and masses. In the future, it could be the calcification of lesions and telling the difference between malignant and benign cancers," Doi said. Ongoing research has been very successful in discriminating dangerous cancers from benign tumors. "The performance of these schemes is very high. I have seen CAD provide better performance than radiologists," Doi said.
Despite their advantages in performance and endurance (computers never get tired and their minds never wander), Doi sees a difficult road ahead for commercial deployment of next-generation systems.
"This is a delicate situation," Doi said. "If a medical device detects something suspicious, it can be used as a second opinion. However, if the device can say malignant or not for 100 or 90 percent cancers, the system enters a delicate domain of physicians. I'm not sure what the FDA's position on this type of device would be."
Doi estimates that convincing the FDA of the sensitivity of a next-generation CAD system would be very expensive, which is likely to deter commercial companies from developing these kinds of systems. "But once the issues are resolved, I think the CAD concept is applicable to all kinds of imaging devices and modalities," he said. "In the past, the U.S. market has been concerned solely with hardware. We believe in the future, it will be software that leads the way."
Regardless of political or other concerns associated with new CAD technology, it seems likely that any solution that increases the chances of early breast cancer detection will be given a fair shot. According to the Susan G. Komen Breast Cancer Foundation, breast cancer is the leading cause of death among women between the ages of 35 and 54. Staggeringly, one out of eight women will develop breast cancer in the U.S. According to R2 Technology, for every 100,000 breast cancers detected by mammographic screening, ImageChecker could identify another 12,800 each year, leading to earlier detection and greater chances of survival.
325 Distel Circle
Los Altos, CA 94022
Phone: (1) 650/254-8988
Fax: (1) 650/254-8958
20975 S Swenson Dr
Waukesha, Wisconsin,
Phone: (1) 414/798-4401
Fax: (1) 414/798-4573
PO Box 19048
10 South Academy St
Greenville, SC 29602-9048
Phone: (1) 864/421-1700 or: (1) 888/557-6100 x1700
Fax: (1) 864/421-1612
1301 Avenue of the Americas
New York, NY 10019-6022
Phone: (1) 212/258-4000
Fax: (1) 212/767-0580
