CAD helps spot smaller cancers but may not change detection rate

May 19, 2008

Computer-aided detection will probably help radiologists detect small cancers earlier but will not increase the cancer detection rate, according to a CAD expert.

Computer-aided detection will probably help radiologists detect small cancers earlier but will not increase the cancer detection rate, according to a CAD expert. Yulei Jiang, Ph.D., of the University of Chicago, a leading CAD research center, presented the results of an analysis of multiple trials that have examined the effect CAD has had on the diagnosis of cancer. He suggested that over time, CAD will become known more for helping detect early, smaller cancers and less for reducing the cancer rate. That wasn't always the case, Jiang said. Early on, CAD was promoted as a way to spot cancers that radiologists would otherwise miss. As research has evolved, however, logic and the nature of the studies suggest that cancer detection rates may not change, but that cancers will be spotted at an earlier stage. Suppose a cancer is spotted in year one. It's removed from the pool for years two and three, and in those years, fewer cancers are waiting to be found, Jiang said. Over time, the cancer detection rate should hit a steady state. If CAD helps, it will be in spotting cancer sooner. This idea has been modeled at the University of Chicago and found to be valid.

Jiang's offered a primer on the types of studies being used to evaluate CAD. He broke them into four groups:

  • Observer studies take place in a lab setting with blinded results early in the development of a CAD approach. They tend to show statistically significant improvements from the use of CAD and often examine complex issues, such how to present the images and sequences.
  • Potential benefits studies also take place in the lab and look at the ability of CAD to spot false-negative or missed interval cancers. The advantage of these studies is that they pinpoint the benefits of CAD. The disadvantage is that they show potential, not real-world, results. Among the cases Jiang reviewed, CAD systems were able to spot about half of 1000 missed cancers.
  • Head to head studies more closely match clinical experience and compare radiologists reading without CAD with those reading with CAD. The advantage is that they show when CAD helps. The disadvantage is that radiologists behave differently in the study, perhaps being more or less vigilant when participating in either arm. Among the studies Jiang reviewed there were about 60,000 cases. Radiologists alone spotted 319 cancers. Those with CAD got 31 more, an increase of 9.7%. In the CAD arm, the recall rate went up 12.5%, a recurring theme in CAD studies.
  • Historical controlled trials calculate cancer and recall rates before and after the adoption of CAD and calculate cancer detection and recall rates. The advantage is that these reflect actual clinical practice and provide a relatively large number of cases. The disadvantage is the studies fail to pinpoint the effect of CAD and the confounding effects of longitudinal studies with staff changes or changing skill levels.

Four large studies fit this last category, and the results were not clear-cut, Jiang said. One by Cupples et al found a 16.3% increase in the cancer detection rates and an 8.1% increase in the recall rate. Others by Gur, Fenton, and Gromet found increases in the cancer detection rate ranging from 1.2% to 1.9% and increases in the recall rate ranging from 0.1% to 30.7%.

The Fenton study was highly controversial when it was published in

The New England Journal of Medicine

a year ago. One problem with that study was that the numbers in the CAD arm were 10 times higher than in the no-CAD arm, Jiang said.Worse, though, was that it used BI-RADS categories to set up an ROC curve. They won't work for that purpose, he said.