Large Study Casts Doubt on Value of CAD

May 1, 2007

Computer software used to help decipher screening mammograms reduces interpretation accuracy, increases the rate of unnecessary biopsies, and offers no clear improvement in the detection of invasive breast cancer, the largest and most comprehensive community-based study of the technology has found.

SACRAMENTO, California—Computer software used to help decipher screening mammograms reduces interpretation accuracy, increases the rate of unnecessary biopsies, and offers no clear improvement in the detection of invasive breast cancer, the largest and most comprehensive community-based study of the technology has found.

Researchers examined the use of computer-aided detection (CAD) in interpreting routine mammograms as part of a survey of 43 facilities in Colorado, New Hampshire, and Washington. The study, which ran from 1998 to 2002, included 429,345 mammograms from 222,135 women, and 2,351 patients who received a breast cancer diagnosis within 1 year after screening. The results showed that women screened at facilities using CAD were more likely to be told their mammograms showed abnormalities, which resulted in an increased number of biopsies (N Engl J Med 356:1399-1409, 2007).

CAD software marks suspicious areas on a mammogram for further review by radiologists, with the intent of aiding the detection of potential or actual invasive tumors. FDA approved CAD software for interpreting mammograms in 1998, and many medical imaging practices now have incorporated the technique for use by their radiologists. Indeed, according to Joshua J. Fenton, MD, the study's lead author, 10% of US mammography facilities adopted the technology within 3 years of its approval.

"There had been no large-scale community-based review of CAD efficacy despite the rapid adoption of this technology, so we did this study to see if CAD was proving to be beneficial," said Dr. Fenton, assistant professor of family and community medicine, University of California, Davis. "It's unfortunate that the use of the software has proliferated so widely before we are certain of its benefits. We need studies to determine if the benefits of the software outweigh its harms and costs. There is also the potential for new studies to improve the performance of CAD software."

The study was conducted by researchers at the University of California Davis Health System, Sacramento, and the Breast Cancer Surveillance Consortium, an NCI-supported effort to assist research by linking mammogram registries to population-based cancer registries. All the facilities surveyed contributed data to consortium registries. The study measured things that can affect mammogram interpretation, such as procedures used in reading the images, use of CAD, and a radiologist's years of experience and the number of mammograms interpreted in the previous year.

Key Findings

During the 5-year survey, seven (16%) of the facilities began using CAD and served as the study's comparison group. The team tracked newly diagnosed invasive breast cancers and ductal carcinomas in situ through the end of 2003. Key findings include:

• Diagnostic specificity dropped from 90.2% prior to the use of CAD to 87.2% after the incorporation of CAD into the interpretation process (P < .001).

• The positive predictive value declined from 4.1% to 3.2% (P = .01).

• The biopsy rate jumped 19.7% (P < .001).

• Sensitivity rose from 80.4% to 84.0% following the introduction of CAD, but was not significant (P = .32).

• The cancer detection rate, which included invasive breast tumors and ductal carcinomas in situ, increased. However, the rise from 4.15 cases per 1,000 screening mammograms to 4.2 cases also did not reach statistical significance (P = .90).

• CAD use was associated with a significantly lower overall accuracy, compared with its nonuse (P = .005).

The study also found that sites using CAD recalled 32% more women for further testing than those that did not adopt the technology, and 20% more women who underwent CAD scrutiny had biopsies. The researchers estimated that for every woman diagnosed with breast cancer as the result of CAD, 156 women were called back for additional testing and 14 underwent unnecessary biopsies.

CAD software also failed to demonstrate a clear advantage in detecting breast tumors. "The study suggests that, if anything, the software may promote the detection of the least dangerous breast cancers, such as localized, in situ breast cancers," NCI noted in a media release. "The effect of in situ cancer on breast cancer mortality remains unknown, and some evidence suggests that not all develop into invasive cancers."

Should all mammography facilities in the United States install and use CAD software, the authors estimated that the annual cost of mammograms nationwide would increase an additional $550 million, or 18%.

"This study points out the need for the use of other techniques to find cancer at its earliest stages," said NCI director John E. Niederhuber, MD. "NCI is incorporating techniques for imaging at the molecular level into many of its studies and is also conducting studies to improve the use of CAD and conventional mammography."