• AI
  • Molecular Imaging
  • CT
  • X-Ray
  • Ultrasound
  • MRI
  • Facility Management
  • Mammography

CAD's role in breast screening shifts from quantity to quality


In just seven years, computer-aided detection has become an indispensable tool for breast imaging centers. As experience accrues, CAD's role in cancer detection with mammography is evolving. Some longtime users find that initial surges in cancer detection rates eventually return to pre-CAD baselines in stable screening populations. Now the focus is on CAD's ability to find cancers at an ever earlier stage. Research is shifting toward optimizing CAD in practice and developing its potential for tumor classification.

In just seven years, computer-aided detection has become an indispensable tool for breast imaging centers. As experience accrues, CAD's role in cancer detection with mammography is evolving. Some longtime users find that initial surges in cancer detection rates eventually return to pre-CAD baselines in stable screening populations. Now the focus is on CAD's ability to find cancers at an ever earlier stage. Research is shifting toward optimizing CAD in practice and developing its potential for tumor classification.

Clinical benefits aside, CAD is also a revenue generator for cash-strapped breast centers. With an upfront cost ranging from $50,000 to $175,000, even high-end systems pay for themselves within three years. The few remaining barriers to third-party reimbursement are falling fast. Aetna, which formerly considered CAD an experimental technology, issued a new coverage policy in April. The insurer now considers CAD a "medically necessary" adjunct to mammography.

CAD algorithms are not perfect. While the tool excels at picking up microcalcifications, it stumbles at identifying masses. Moreover, CAD is only as good as the mammographic data on which it is based. And it tends to err on the side of caution, presenting images with a forest of potentially false-positive marks. Facing an average of two CAD-generated marks per study, a radiologist reviewing 1000 mammograms must investigate 2000 marks, said Dr. Michael Linver, director of mammography for X-Ray Associates of New Mexico in Albuquerque.

"That's a lot of marks, particularly when you consider that on average, there are five cancers for every 1000 patients," Linver said at the National Consortium of Breast Centers meeting in March. "For each cancer, I am looking at 400 marks. That needs to be improved."

Research conducted in Norway and the U.S. suggest that CAD's clinical virtues are most accessible to radiologists who read screening mammograms only occasionally. That isn't necessarily a bad thing, Linver said.

"In the U.S. today, more mammograms are being interpreted by low-volume readers, and this is an issue that CAD can address," he said.

Even with caveats, CAD is considered a boon to mammography screening and interpreting physicians. The worst kept secret in radiology is that many diagnosed cancers can be seen in retrospect on prior screening exams. Double reading, which increases cancer detection at a rate that ranges between 4% and 32%, according to Linver, is not economically feasible for most sites.

Depending on which retrospective or prospective study one favors, CAD initially increases detection 7% to 20%.

At least as important, however, is CAD's ability to focus a radiologist's attention, said Dr. Stamatia Destounis, a radiologist at Elizabeth Wende Breast Clinic in Rochester, NY.

"CAD does not get tired. It predictably marks the same things," she said. "It misses the same things all the time, too. But it does mark distortion and calcifications and does it consistently. CAD makes you stop to take another look."

With so much attention paid to CAD's clinical assets, its contribution to workflow often goes unnoticed.

"Interpreting screening mammography is a tedious job, particularly when dealing with high volumes. There's no doubt that at the end of a day using CAD, radiologists are less tired," said David Gur, Sc.D., a professor of radiology at the University of Pittsburgh.

CAD has also helped standardize workflow, he said, which may lead to reduced variability in cancer detection among radiologists.

CAD's popularity has spawned several options for sites that can't afford the upfront capital cost. PlusCAD, headquartered in Pittsburgh, provides CAD services for mammography screening exams on a fee-per-study basis for its 23 clients, which include hospitals, radiology practices, and breast centers.

"Even though the cost of CAD systems has declined, many hospitals are cash poor," said PlusCAD CEO Herta Klaman. "When it comes to deciding whether to replace a piece of critical clinical equipment or enhance a service, it can be difficult to get capital allocation for CAD."

A similar service, Mammassist, began in 2003 in Statesville, NC. The company also charges a per-exam fee that is less than Medicare's technical reimbursement payment for CAD. And in February, R2 Technologies, one of three FDA-approved CAD device vendors, released a tabletop version of its ImageChecker system tailored to low-volume film-based mammography sites.

"All breast centers should have CAD," Destounis said. "Even if you only pick up one cancer in 1000, you are going to save someone's life. Some systems are more expensive than others, but even those that lack the bells and whistles can still do the job."


Early studies of CAD focused on its performance in retrospectively detecting cancers on screening mammograms. These data propelled CAD to FDA approval in 1998, and it received the reimbursement nod from Medicare in 2001.

A prospective 2001 study based on 13,000 screening exams performed in a community clinic bolstered CAD's favorable statistics. Researchers found that CAD increased cancer detection by 20% during the one-year duration of the study (Radiology 2001;220: 781-786).

Users should expect to see an increased detection rate when CAD is initially implemented, but rates vary. Most users report increases of 10% to 15%, said Dr. Murray Rebner, director of the Vattikuti Digital Breast Diagnostic Center at William Beaumont Hospital in Royal Oak, MI.

"At our site, we are probably seeing increases in the 5% to 10% range across all radiologists and levels of experience," he said. "It has saved me on a number of occasions."

That has not been the experience everywhere, however. In a prospective study at the University of Pittsburgh involving 115,571 mammograms, researchers found no statistically significant difference in breast cancer detection or patient recall rates with the use of CAD among all radiologists, including highly experienced breast imagers (J Natl Cancer Inst 2004;96:185-190).

The study prompted debate as well as a subsequent editorial in Radiology about the hazards of evaluating new technologies, penned by chief investigator Gur (Radiology 2005; 234:659-660).

"When you look at any new technology, you have to be careful how you assess its impact before incorporating it into practice. In the case of CAD, reimbursement came early, before we had a chance to carefully assess it outside of the laboratory," he said.

Gur emphasizes that he is a strong proponent of CAD, but users must have a realistic outlook. Most expect CAD to spur a continual increase in cancer detection rates. But over time, as a population is screened with CAD, detection rates will return to pre-CAD figures.

"We can't generate new cancers. But we can, hopefully, measure a shift in the stage and detection of cancer with CAD," Gur said.

Dr. Tommy Cupples, director of breast imaging at South Carolina Comprehensive Breast Center in Columbia, arrived at the same conclusion earlier this year. In analyzing results from a two-year CAD study involving 27,000 women, he found that the cancer detection benefits reaped in the study's first year eroded in year two.

The study, to be published in the American Journal of Roentgenology, showed no statistically significant increase in the detection rate of cancers with CAD but a nearly statistically significant increase in detection of small, invasive cancers 1 cm or less in size.

"Cancer detection rates have to return to the baseline, depending on the rate of growth of the cancer and the sojourn time," he said. "Ultimately, detection rates are not the best measure of CAD performance. But our experience does show that CAD resulted in better detection of smaller, earlier stage tumors."

As more sites track their years-long experience with CAD, they might expect similar results, Cupples said.

"This parallels screening mammography, where we saw that there was no benefit to tracking how many cancers we found. Instead, we had to go to another standard, which was mammography's impact on mortality, outcomes, and longevity," he said. "If your detection rate hasn't changed at all with CAD, go back and look at the cancers found, their type and size."

Elizabeth Wende Breast Clinic relies on both CAD and double readers in reviewing screening exams. In a retrospective study, double reading provided superior cancer detection compared with a solo radiologist and CAD (Radiology 2004;232:578-584). But CAD did prompt radiologists to work up 7% more exams that resulted in cancer diagnoses, Destounis said.

In a prospective study being readied for publication, Destounis has also documented CAD's help in identifying invasive cancers.

"The criticism of CAD is that it picks up tiny microcalcifications that don't turn out to be serious. In the study we are writing, CAD prompted us to work up invasive cases, not just intraductal cancers. And those are the killers," she said.


One of CAD's biggest selling points is its ability to reveal malignant calcifications, with sensitivities ranging from 86% to 99%, according to a host of published studies. But radiologists must remain vigilant, particularly when it comes to amorphous calcifications.

This subset harbors a 20% likelihood of malignancy and is frequently overlooked on mammography. In one recent study, CAD sensitivity proved markedly lower for amorphous calcifications than for malignant calcifications overall (AJR 2005; 184:887-892).

Users should also recognize that a CAD report is only as good as the image it is based on. Moreover, the segmentation process by which CAD organizes digital data for image analysis may exclude tissue that harbors critical findings.

In 2020 screening mammograms evaluated at Duke University Medical Center, segmentation was nearly perfect or acceptable in 96.8% of images, according to Dr. Jay Baker, chief of mammography (Radiology 2005; 235:31-35). Among those considered unacceptable, segmentation defects were more common in mammograms with heterogeneously dense tissue than in those with fatty replaced tissue, scattered tissue, or extremely dense breast tissue.

Images with unacceptable segmentation were found to exclude up to 25% of breast tissue. The bottom line, Baker reported, is that incomplete or inaccurate segmentation may affect lesion detection.

CAD's greatest weakness is its failure to consistently and accurately identify masses. Dr. Rachel Brem, director of breast imaging and intervention at George Washington University, has studied the pros and cons of CAD extensively. In two new studies, she found that while breast density, mammographic appearance, and lesion size have little impact on CAD's ability to detect cancers, its sensitivity in masses lags behind its sensitivity in calcifications (AJR 2005;184;439-444,893-896).

CAD's poor reputation in this area has influenced the way radiologists interpret marks, Gur said.

"Some pretty much ignore the CAD unless it marks both views in the same location, and the radiologist did not see the mass before viewing the CAD results," he said. "Others mentally discount the mass results altogether. Thoughts about how seriously one should take mass marks on CAD vary substantially among radiologists."

Thus radiologists may discount CAD findings of malignant masses even when the report is right. In a paper presented at the 2004 RSNA meeting, Dr. Dianne Georgian-Smith, a breast imaging specialist at Massachusetts General Hospital, reported that two malignancies-one an architectural distortion and the other involving punctate calcifications-were identified using CAD but were dismissed by a baseline reader.

There are bound to be disparities between what CAD detects and what radiologists detect. In the majority of cancers, CAD and the interpreting physician will concur, Gur said. But the more subtle the cancer, the more likely that CAD will mark it on only one mammographic view. That presents a problem for breast imagers.

"If it shows on only one view, the question is, should the radiologist discard it, or does the radiologist take it seriously enough to pursue? Right now, it depends very much on the individual," he said.

Experience has been the teacher at William Beaumont Hospital, where CAD has been in use for about three years, Rebner said.

"The more experienced people quickly learned how to ignore the irrelevant marks, whereas the newer people were calling those patients back," he said. "Now most people are pretty comfortable knowing what to call normal and not normal."


CAD's success with mammography has spurred pairings with MRI and ultrasound. But in terms of breast cancer detection and characterization, the technology is far from mature.

Users' wish list for CAD includes better detection of masses, and, if not tumor classification capability, at least more reliable marks. Rebner would like to see refined CAD techniques for detecting subtle ductal carcinoma in situ and, potentially, a way to measure lymph node data in the axilla.

He predicts that the next big development for CAD will be in breast MRI, pending clinical data that verify its sensitivity and specificity.

"One way the industry can help breast imagers is by enabling them to look at a particular lesion and say, 'This has a 92% chance of being malignant based on kinetic and morphologic data, compared to a lesion that has only a 20% chance,'" Rebner said.

Destounis would like to see that capability applied to mammography-based CAD.

"If there was a ratio or percentage of concern, we would pay more attention. A percentage system or graph would show radiologists why one mark is more important than others," she said.

Lesion classification capability is on the horizon. A system for computer-aided diagnosis has been developed and tested at the University of Chicago (Radiology 2002; 224: 560-568). Existing CAD systems also show promise for tumor classification, a development that will be valuable in reducing unnecessary biopsies, Gur said.

"It has gotten easier and easier to sample breast tissue with fine-needle aspiration and core biopsy. Computerized classification is potentially a great alternative to sampling tissue. But the question remains how much it will affect biopsy rates," he said.

Another key development for CAD will be the ability to compare data from all current images and priors-an image and data processing advance that may be beyond reach. Today's systems analyze each image largely independently of the other views, Gur said. And, of course, prior exam data are not factored into the mix. As a result, CAD is unable to take advantage of the complete data set available for cancer detection and analysis.

"It's an extremely difficult problem because of differences in soft breast tissue each time it is compressed," he said. "We don't yet have a good solution."

Cupples envisions a futuristic scenario involving CAD as a quality assurance tool, with algorithms embedded in mammography devices that optimize breast image acquisition before data are ever analyzed for cancer.

"One of the real tasks of CAD is to find a good mammogram," he said. "So many things can affect mammogram quality and hide subtle cancers: motion blur, underexposed film, suboptimal technique. These are the sorts of parameters that can be measured pretty reproducibly by CAD programs."

Such beyond-the-box thinking about CAD and how it can best be used is needed if the technology is to fulfill its potential, Gur said.

"We are seeing the opening of a chapter rather than the closing of one," he said. "Just because CAD is reimbursed shouldn't stop the curiosity that we have in understanding and optimizing it."

Ms. Dakins is feature editor of Diagnostic Imaging.

Related Videos
Nina Kottler, MD, MS
The Executive Order on AI: Promising Development for Radiology or ‘HIPAA for AI’?
Related Content
© 2024 MJH Life Sciences

All rights reserved.