• AI
  • Molecular Imaging
  • CT
  • X-Ray
  • Ultrasound
  • MRI
  • Facility Management
  • Mammography

Lung CAD faces barriers to widespread adoption


Experts call for full, real-time integration of computer assistance programs with PACS

Widespread adoption of lung computer-aided detection is being stalled by obstacles such as false positives, image overload, and lack of full integration with PACS, according to radiologists reporting at the Stanford Multidetector-Row CT symposium in June.

CAD with multislice CT allows detection of many more nodules than was possible with earlier generation scanners. That ability can be both good and bad. More nodules will be picked up, but radiologists could become inundated with findings to assess. Thin slices make differentiating between vessels and nodules more difficult.

A small new Stanford study also suggests that radiologists have problems distinguishing between true and false positives prompted by CAD, said Dr. Geoffrey Rubin, chief of cardiovascular imaging. Results of the study have not yet been published.

"The new data give cause for pause about anyone using CAD clinically until we understand it better. There is a paucity of studies to assess true- and false-positive detection. We are not using [CAD] in the clinical environment," Rubin said during a symposium question and answer session.

Consequently, he advised caution with lung CAD.

"We have a lot to learn about what is a nodule, what we want to detect, and how effectively we can discipline ourselves to exclude what in the end turns out to be a false positive," he said.

Much less ambiguous is the need for CAD in pulmonary nodule assessment. Extensive research shows that radiologists commonly make interpretation mistakes, missing up to one-third of pulmonary nodules.

"The literature is replete with examples of radiologists' folly in our presumed ability to detect lung nodules. Performance can be strikingly poor in a diverse set of circumstances," Rubin said. "We can benefit from aids to detect lung nodules."

CAD is thought to be most useful after a radiologist has made an initial interpretation and it is used as a follow-up to the primary reading.

In a study published in 2005, Stanford researchers tested CAD with three radiologists and 20 outpatient CT scans. The radiologists made their own interpretations and then checked CAD markings. CAD helped improve detection significantly, spotting 57 additional nodules. But in the new Stanford study discussed at the meeting, researchers asked if the same three radiologists would reject false positives.

Again, individual radiologists made their own interpretations, then looked at the CAD markings. After 10 seconds of looking at CAD, sensitivity improved without an increase in false positives. But as time went on, the radiologists were more likely to accept false positives.

"This is a striking finding and one that we hadn't anticipated. We observed this for all three radiologists and at all size thresholds," Rubin said.

The problem occurred between 60 and 100 seconds of looking at CAD results. The average time for the shift occurred at 90 seconds.

"Can we have too much of a good thing? It appears as though in this patient set, radiologists, given enough time, will relax standards when CAD shows them lung nodules and begin to accept greater numbers of false positives," he said. "We need to understand the nature of false positives better and assess if radiologists can be trained to avoid them."

Other radiologists at the symposium, however, said that the number of false positives is not that problematic. The range of false positives is wide, depending on the vendor, said Dr. U. Joseph Schoepf, an associate professor of radiology at the Medical University of South Carolina. Schoepf added that in his experience, false positives can be quickly dismissed.

At the Stanford meeting, concerns were raised that other barriers to lung CAD adoption are more formidable. As the number of nodules, including incidental findings, detected with CAD increases, radiologists fear being inundated, said Dr. Jeffrey Mendel, chair of radiology at Caritas St. Elizabeth's Medical Center in Boston, during the same symposium session.

"We are seeing more nodules that need follow-up studies," he said.

CAD has great potential in assessment of nodules and is a powerful tool, according to Mendel. At the click of a mouse, it is possible to determine volume.

Lack of full integration with PACS is holding up its implementation, however, he said. Currently, CAD is not available on PACS workstations. Assessment is made at another terminal and sent back to PACS. Even if the software is loaded onto PACS, it is necessary to launch a separate window.

"For CAD to become ready for prime time, it must be integrated into the workflow," he said.

A first step is to have CAD output sent to PACS as an overlay that can be easily used. This is possible with available systems, but the ultimate solution is full integration, Mendel said. It should be possible for PACS to detect when a CT study covers the lung and to automatically send data for CAD analysis.

"When you read a study, CAD tools should be available in the same way ROI and distance measurement are available. You click on a nodule, and the work is done for you," he said.

If radiologists became more vocal about this need, PACS vendors would make integration possible, Mendel said.

"We need to tell PACS vendors that we are missing pulmonary nodules. We need a tool, and they should put CAD on the PACS. They will do it for us if we ask for it," he said.

Related Content
© 2024 MJH Life Sciences

All rights reserved.