Questions linger on clinical value of breast CAD

July 22, 2004

Computer-aided detection software may help radiologists spot more cancers, but its clinical potential remains limited by high false-positive rates.Two studies presented at the ECR highlighted CAD's ability to boost single-reporting sensitivity and

Computer-aided detection software may help radiologists spot more cancers, but its clinical potential remains limited by high false-positive rates.

Two studies presented at the ECR highlighted CAD's ability to boost single-reporting sensitivity and nearly match independent double-reading detection rates. Speakers failed to reach consensus, however, on the effect that CAD's generally low specificity would have in clinical practice.

Researchers at Friedrich-Schiller University in Jena, Germany, used CAD in a retrospective analysis of 39 mammograms from patients whose breast cancer was initially overlooked. The software detected all microcalcifications and 21 of the 35 masses on the prior mammograms. This translated to an overall detection rate of 22 out of 32 malignancies (68.9%), with a false-positive rate of 0.44 for masses and 0.31 for microcalcifications.

The study demonstrates the ability of CAD to detect malignancies previously missed on breast screening, said Dr. Ansgar Malich, a radiologist at Friedrich-Schiller. A shortage of radiologists in Germany makes independent double-reading difficult, but using CAD and a single reader could improve cancer detection rates, especially when younger radiologists report on their own.

"About two-thirds of overlooked malignancies are in principle detectable by CAD, but a couple of factors are still limiting its use in clinical practice," he said. "Rates of false-positive CAD markers are still much too high. There are also questions whether further use of CAD would increase the number of patients recalled for further examinations. However, in our own study, we did not see this result."

In another study, researchers at the Center for Studies into the Prevention of Cancer (CSPO) in Florence, Italy, tested early versions of competing commercial breast CAD systems against single- and double-read conventional mammography. Six radiologists experienced in reporting screening mammograms assessed 120 selected films (89 negatives, 31 interval cancers). Two weeks later, they evaluated the same films with the help of printouts from R2 and CADx software that had been produced from digitized films.

The results showed noticeable differences in the nature of the two software systems. The R2 software marked many more calcifications than the CADx system (218 versus 132), while the CADx printouts highlighted more masses (208 versus 105). Taken overall, however, the two algorithms had similar sensitivity (70.9%) and a higher sensitivity than single reading without CAD (58.6%).

The idea behind breast CAD is to reduce the number of false negatives, which essentially means cutting the number of interval cancers, said Dr. Stefano Ciatto, principal investigator in diagnostic imaging at CSPO. Interval cancers generally develop from masses missed at breast screening, not from calcifications, which are easier for radiologists to see.

"So theoretically, a perfect CAD system should be dedicated to masses more than microcalcifications," Ciatto said. "But the CAD companies should pool their expertise to produce an algorithm that does well when detecting both microcalcifications and masses."

Both programs on trial also flagged a large number of false alerts, translating to a small but significant increase in recall rate: from 18% to between 24% and 30%. Many radiologists are concerned about a possible escalation in recall rates with the use of CAD, but this is unlikely to occur, according to Ciatto.

"It is more important to rely on the overall effect of the reading than on what CAD displays on the screen as marks," he said. "Radiologists are able to interpret false positives as false positives, and it is my impression that the increased recall rate is not that dramatic."

Results of all single reads in the Florence study were additionally combined to simulate the effect of double-reading, with and without CAD. This comparison yielded no significant difference in sensitivity between single-reading with CAD and independent double-reading of unmarked studies.

"Many retrospective studies show little difference in accuracy between CAD-assisted reading and double-reading," Ciatto said. "But like ours, these are retrospective studies on prepared film sets that tend to enhance sensitivity and reduce specificity. What we need is prospective studies from a breast screening environment, though this would involve a lot of work to set up."

If such prospective studies were to confirm a close similarity between results from single-view CAD-assisted reading and double-reading, it would be feasible to recommend the use of breast CAD in clinical practice instead of a second reader, he said.