| |||||||||||||||
|
Home
|
For other articles and previous issues click here. November 8, 2004 New
Order? Recently approved “joint-read” software moves away from the second-read computer-aided detection (CAD) standard. Many radiologists wonder whether or not a computer reading an exam before the doctor will bias his or her interpretation. It’s an intriguing issue as CAD moves into the mainstream. In many hospitals and imaging facilities today, computers are serving as a “second pair of eyes” for the radiologist. The radiologist conducts a standard review of the images. Then the computer “reads” the scans as a second check and marks any potential area or areas of concern. But this past August, the FDA approved London-based Medicsight’s Lung CAR (computer-assisted reader), which is designed to read spiral CT scans at the same time the radiologist does. Medicsight Lung CAR displays two scans side by side: one the “unfiltered” image and the other the enhanced image, where the computer has marked regions that may be a potential threat. The company is promoting its product as a “joint-reader.” Medicsight expects to soon receive FDA marketing clearance for a similar software product that detects polyps in the colon. Mary Roddie, MD, a consultant radiologist at Charing Cross and Hammersmith Hospitals NHS Trust in London and radiology director at Medicsight, believes that joint-read software is a necessary step forward in today’s information age and that its use can only help improve detection of cancers in their earliest stages when they are most treatable. When evaluating CT lung scans, radiologists are bombarded with information. They have to scroll through hundreds of images. Catching More However, some radiologists have the same concerns about using computer technology as a joint-reader as they do about using it as first reader. They are fearful that radiologists could rely too heavily on the computer to find suspect areas and ignore, or at least give less weight to, their years of knowledge and experience. “I think that at the present time, radiologists are more comfortable with the idea that they read the case first and that the computer-assisted system serves a backup to check their work,” says Ronald M. Summers, MD, PhD, a senior investigator and staff radiologist chief in the diagnostic radiology department at the Clinical Center, National Institutes of Health (NIH) in Bethesda, Md. “That may change, but I don’t think that, as a professional culture, we are ready yet to turn over our primary responsibility to a computer program.” Developed over the past 20 years, CAR technology has been proving helpful in detecting lung nodules. Those nodules can indicate early lung cancer, but it’s often difficult for radiologists to distinguish between benign and malignant nodules. Medicsight’s Lung CAR works by deploying a series of filters against the image data derived from spiral CT scans. The filters help radiologists by highlighting the areas of the image that contain potential nodules. The software can extract the boundaries of suspect nodules and show them in three dimensions with a volume measurement. Those measurements also allow the radiologist to review and track any growth in the nodule. The Lung CAR software also has a number of automatic and manual measurement tools to aid diagnosis, including the ability to review follow-up scans and determine the length of time it took for the nodule’s volume to double. Additional filters reduce background noise on the image and enhance the boundary of the nodule. “CT is fantastic technology for picking up tiny lung cancers,” Roddie says. “The problem is that there are many tiny nodules in people’s lungs that are not cancerous. In a screening situation, you want to follow up [on] all the nodules you find. If you find suspicious lung nodules and can measure accurately their growth pattern over six to eight weeks, you can work out whether what you have seen is a small lung cancer or a benign nodule of no importance to the patient at all.” Saving Time Medicsight’s Lung CAR is one of the most accurate measurement tools available, Roddie says. Accurate boundary identification is key, she explains. Because of the level of detail Lung CAR provides, radiologists can follow up on suspicious CT scans in a shorter period of time. When an area of potential concern is identified, standard clinical practice calls for a follow-up scan to be done in six to 12 months. With Lung CAR, radiologists can schedule follow-up scans within weeks rather than months. More than 7 million chest CT scans are performed in the United States each year. Up to 25% of them reveal nodules in the lungs that at least require further investigation. Lung CAR already falls under the existing $129 Medicare and Medicaid reimbursement code for the use of 3-D CT detection software. According to the American Cancer Society, lung cancer is the leading cause of cancer deaths in the United States, outnumbering deaths from breast, prostate, and colon cancer combined. Lung cancer represents 28% of all diagnosed cancers. More than 160,000 deaths are projected for this year alone. The average five-year survival rate for all lung cancers is estimated to be less than 13%. If lung cancer is caught before it has spread, however, the five-year survival rate improves to 70% or better. Medicsight, which has also received European marketing approval for the software, expects to have one or more distributors for Lung CAR by the end of the year. While Roddie can understand the radiology community’s concern about relying too heavily on computers to do their job, she believes most radiologists would welcome the opportunity to joint-read images. Turf Issue The joint-reader concept capitalizes on the radiologists’ ability to multitask and could be a real and important time-saver, she says. “As radiologists, we’re already used to doing this [joint-reading],” Roddie says. “When we look at a CT scan of the chest, there are different structures we are examining—the chest wall, the lung, the heart—and so we are quite used to flicking our eyes back and forth between an image displaying lung detail and the same image processed to emphasize other organs. We are quite used to the concept of looking at two screens side by side with different features being enhanced. That’s what Medicsight is trying to capitalize on here.” Also, Roddie says, reading the scans with the computer means patients do not have to wait until the scans have gone through a second read to hear the results—good or bad. “We have found that when in practice, radiologists are confidently flicking backward and forward between the two screens—it’s a genuine joint-read—and that has potential for time savings,” Roddie says. Joint-read, Roddie says, should address a lot of concerns radiologists have with the computer technology. Research has also shown that when the radiologist and computer-aided detection (CAD) or CAR systems function as a team, it increases physician accuracy by decreasing observational oversights. A study by Timothy W. Freer, MD, director of the Women’s Diagnostic & Breast Health Center in Plano, Tex., in Radiology in 2001, found that using CAD to scan mammograms for abnormalities improved early detection of breast cancer by almost 20%. Freer and his research team studied 12,860 patients in a community breast center whose scans were read by radiologists first and then with the R2 Technology ImageChecker CAD System, the first FDA-approved system for use with film screening, diagnostic, and digital mammography. R2 Technology, Inc., headquartered in Sunnyvale, Calif., is developing CAD systems for a variety of imaging modalities and disease states. The Freer team found not only a 19.5% increase in the number of cancers detected when a computer does a second read but also an increase from 73% to 78% in the proportion of early-stage (0 and 1) malignancies detected. One of the biggest concerns with CARs is that they increase the number of false positives and false negatives and thus the recall and biopsy rate. However, the Freer study found that mammography CAD had only a slight increase on the recall rate—from 6.5% to 7.7%—and no change in the 38% predictive value for biopsy. Roddie says radiologists can address the false-negative or false-positive issue by adjusting the sensitivity of Medicsight’s colon and lung readers. For example, when reading lung scans, the radiologists can determine the shapes—such as spherical or elongated—they would be most concerned about and have the computer look only for them. By adjusting the settings, the radiologist can feel most comfortable with the areas the computer highlights. Roddie believes that as the algorithms used to detect abnormalities become even more sophisticated, computer-assisted technology will play an even greater role in the detection and treatment of disease. However, she says, it can never replace the radiologist, who will always be the final arbitrator. Bias Concerns Abraham H. Dachman, MD, FACR, professor of radiology at the University of Chicago, agrees with Summers, of the NIH, that radiologists should work completely independently first and use the computer only as confirmation or as an additional review. “My gut feeling,” Dachman says, “is that you’re going to give it your best read if you look at it without any bias.” Dachman doesn’t believe a joint-read is even possible. “You may have the enhanced and the unfiltered images up on the screen at the same time, but you will have to look at one first and then the other. Joint-read seems to imply that there is some in-between. By my logic, there isn’t any. If you have both in front of you, how do you decide which one to look at first?” Dachman has no doubt that if the computer-enhanced image were in front of the radiologist at the time he was reviewing the unfiltered scans, it would bias his decision making, whether or not he even realizes it. It’s the same when a doctor knows a patient’s history, he says. He has often tested his theory with his residents. “Just for fun, when I am with my residents, as we read, I will tell them not to give me the patient’s history. Then I will say, ‘Now give me the history,’ and see if that changes what I want to say about the images we are looking at. We all know that we do a different kind of read when we’re directed by a clinical history. I think the computer markings are the same notion. Once you are guided or directed by something, it changes your thinking.” Another issue, Dachman says, is that CAD is not yet designed to find everything. For example, when doing a virtual colonoscopy, “you need to know how it is validated for polyps, masses, and flat lesions. Not all CAD is created equally.” Patient Expectations The patient may be treated for the lesion found
on the virtual colonoscopy, but neither the Such dangers could be inherent, Dachman says, when radiologists use the computer as their first or coreaders and rely, intentionally or unintentionally, too heavily on its results. “My patients expect that I am going to look at their whole colon,” he says, “and that is what they can expect. They don’t want to hear, ‘I found one lesion, and I’m too busy to look at the rest of your colon.’” — Beth W. Orenstein is freelance writer based in Northampton, Pa. She is a frequent contributor to Radiology Today. |
 |
|
3801 Schuylkill Rd • Spring City, PA 19475 Publishers of Radiology Today All rights reserved. |