CAD’s growing up, as a technology, but has it matured to the point of being diagnostically useful? That was the question speakers undertook to answer at this morning’s session on “CAD in Radiology: The Hope, the Hype, and the Hard Truth.”
First, the hope and hype: since the first study on CAD appeared in JACR in 2000, vendors have been clamoring about the software’s ability to improve upon radiologists’ diagnostic power, increasing their confidence while helping them detect subtle lesions. Nicholas Petrick, PhD, of the FDA noted that the first CAD product was approved in 1998 – the veritable dark ages – and that today, radiologists are using CAD in 74% of screening mammograms. “We can expect that CAD’s utility is going to grow,” he said. “It’s out there being heavily utilized in the marketplace.”
CAD for colonography and lung nodules is still relatively new, but CAD for mammography has been heavily studied – to varying degrees of effectiveness, noted Robert Nishikawa, PhD. Faithful Stat Readers will perhaps be aware that their humble narrator is not much of a statistician, so I’m ill-equipped to explain exactly how to construct the ideal study of CAD’s effectiveness, a topic covered by Nico Karssemeijer, PhD.
However, I can tell you that according to Nishikawa, most CAD studies are fundamentally flawed in that they use as their keystone cancer detection rates when they ought to be using cancer size and stage at time of detection. The real benefit of CAD, he argued, is its ability to increase diagnoses of early-stage cancers -- which the human eye is likely to pick up at some point, but maybe not until the disease has advanced.
Nonetheless, he provided an interesting breakdown on the available research on CAD’s effectiveness at detecting disease at any stage. Study results oscillate wildly, with some finding CAD to be 20% more effective than a radiologist alone and others setting that figure at more like 2%. “An objective person looking at these data would say CAD is not working so well,” he noted. “The bigger studies tend to show very small increases in cancer detection rates.”
Nishikawa did note that in independent double reading studies, CAD scores favorably when compared with radiologist double reads (and even more favorably, of course, when compared with single reads). In the largest study of this kind, the sensitivity rate and recall rate increased by the same amount when using CAD. But is this a good thing?
Not necessarily, because it ignores – as did the 2000 study alluded to at the beginning of this entry – the issue of false positives, which increase patient anxiety and lead to unnecessary (not to mention costly) biopsy procedures. Still, Nishikawa said, “We conclude from this that CAD can be an effective tool to assist radiologists in screening mammography. Cancer size and stage are more appropriate endpoints than cancer detection rate, and this is important when we apply CAD to other screening modalities.”