Radiologists embrace CAD, with an eye on its limitations

Does software that flags malignancies on medical images help, hinder, or make no difference to patient management? That question has dogged radiology for years. Automated detection systems are undoubtedly becoming smarter, strengthening arguments for their use. But, as with most emerging technologies, there is room for improvement.

"CAD methods are here, and at least some of them are ready for clinical use. The ones that are not quite ready for clinical use show substantial promise," said Dr. James Thrall, radiology chair at Massachusetts General Hospital, who moderated a refresher course at the 2007 European Congress of Radiology in March.

CAD software's long-standing weakness is its tendency to flag apparent abnormalities that are not actually malignancies. Progress is being made in this area as the technology evolves, particularly in lung CT, said Dr. Catherine Beigelman-Aubry, a radiologist at Pitie-Salpetriere Hospital in Paris. She noted, however, that the workflow for CAD systems is not optimized.

"The best CAD systems are not always available on clinical workstations used for diagnosis," she said.

Systems requiring that CT data sets be transferred to a separate workstation waste too much time. Lung CAD software will remain underused until it can be integrated into a reading workstation or PACS, Beigelman-Aubry said. A number of different lung imaging problems can be addressed with CAD, including detection of pulmonary embolism and quantification of emphysema. Most systems, however, have been designed to pick up lung nodules. The best systems have a sensitivity of around 85% for solid nodules, whereas the sensitivity of a human observer will be around 60%.

"Without a CAD system, screening for lung nodules is a long and tedious task," she said. "Some CAD systems may also depict nonsolid nodules with a higher sensitivity than human observers. This is important to consider because the prevalence of malignancy in these nodules is around 10 times higher than in the solid nodules."

The task of hunting for abnormalities on screening mammograms is fraught with similar problems. Images are complex, and the changes indicative of a cancer can be extremely subtle. Readers faced with a large pile of screening mammograms to report will inevitably become fatigued. Little wonder, then, that there is a reasonably steady miss rate of 20%.

Breast CAD packages are now prevalent in the U.S. Radiologists who previously worked alone can double-check their findings and be reimbursed for doing so. Uptake in Europe, however, has been considerably slower. This is largely due to the long-standing tradition of double reading for breast screening in Europe, said Dr. Rosalind Given-Wilson, a radiologist at St. George's Hospital in London.

As healthcare costs increase, breast screening providers in Europe will undoubtedly look at replacing their second human pair of eyes with a software package. Studies examining the benefits of breast CAD have produced contradictory findings, however. The switch may not even be cost-effective if the number of women called for follow-up rises but the rate of cancer detection stays reasonably constant.

One area where breast CAD could be used in Europe with relatively little controversy is training. The "show me one like it" feature, which brings up similar-looking abnormalities and their correct diagnoses, could be a useful tool for trainees, Given-Wilson said. What happens tomorrow will depend on how the technology evolves to reduce false positives and minimize unwarranted recalls.

"At the moment, if given a choice, I would rely on a second reader for screening mammograms. But that's not to say that in a few years' time I won't be relying on CAD," she said. "The decision-making interface is where things need to improve. It's not that breast CAD systems are bad, or that radiologists are bad, it is just how they work together."

Ms. Gould is a contributing editor of Diagnostic Imaging.