Radiologists puzzle over alternatives in chest CAD

Thoracic imaging is attracting considerable interest and investment as a prime application for computer-aided detection. But incorporating lung CAD into the clinical radiology workflow requires careful planning to realize its true potential, according to speakers at the Computer Assisted Radiology and Surgery meeting.

Many radiologists distrust the ability of chest CAD to improve their performance or speed reporting throughput, said Dr. Dag Wormanns, a radiologist at the University of Munster in Germany, during the June CARS/EuroPACS conference in Berlin.

Common complaints include the software's tendency to highlight clinically irrelevant findings and the extra time required to load CAD applications or move to another workstation to run the package.

"If you talk with general radiologists, they are quite skeptical. They do not use chest CAD for the most part," Wormanns said.

CAD should be applied to every chest CT scan as a standard procedure, he said. This strategy would be more acceptable if CAD software was readily available on radiologists' desktops along with the CT data.

More information is needed about the best use of the software, Wormanns said. Using chest CAD as a prereading filter saves time but leaves radiologists reliant on the software's sensitivity. Looking at CAD marks after viewing a series of images should improve radiologists' sensitivity, but it would inevitably increase reporting time. Alternatively, a radiologist could use CAD while actively reporting, although the clinical and workflow impacts of that strategy require further clarification.

"CAD can either save time as a filter or increase sensitivity as a second reader. I am not too optimistic that there is a trade-off," Wormanns said. "The role of concurrent reading is still to be defined."

Addressing skeptics' concerns about chest CAD is only the first step. Departments may face a more complex choice than they had expected when purchasing a system, said Dr. Matthew Freedman, a radiologist with the Lombardi Cancer Center and ISIS Imaging Science Center at Georgetown University in Washington, DC.

One way of rating a CAD system's performance is to evaluate machine sensitivity combined with the number of false positives per image. But these measurements will vary according to the content of the trial database, making it difficult to compare competing systems unless they have been run through exactly the same caseload.

A less frequently used measure is radiologists' absolute improvement in disease detection rate, Freedman said. But this also is problematic because the impact of CAD prompts on individuals' performance is contingent on the caseload content. In addition, different CAD packages can aid some radiologists more than others.

For example, when 15 radiologists were asked to report chest x-ray films from a group of male heavy smokers, their initial variance in sensitivity ranged from 57.5% to 77.5%. Addition of CAD prompts caused their performance to change by -2% to +15%. The impact of CAD on the group would have been difficult to predict, however. The radiologists whose performance changes fell at the two extreme ends of the range both started with a baseline sensitivity of 57.5%.

In a second study, two radiologists reviewed a number of chest radiographs containing non-small cell lung cancers measuring 5 to 30 mm, first without and then immediately with the chest CAD software. Addition of CAD boosted the first radiologist's sensitivity by 3.1% for one group of the cases and by 9.4% for a different group. The second radiologist recorded performance improvements of 7.4% and 8% on another two collections of images.

"If you are trying to compare two CAD systems, you really have a difficult problem, because the best system for one radiologist is not necessarily the best for another, and the best CAD system for one practice or caseload may not be the best for another," Freedman said.