Radiologists give small polyps only cursory glances, leading to lots of false-negatives

Tracking the eye movements of 42 radiologists viewing CT colonography videos, British researchers found the physicians quite capable of spotting even very small polyps—but not so good at paying them sufficient attention to recognize them as true polyps.

The radiologists tended to spot the deceptively inconspicuous polyps only to quickly move on, incorrectly labeling them negative even when computer-aided detection (CAD) flagged the polyps for closer inspection.

The small yet intriguing study, lead-authored by Andrew A. Plumb, MRCP, FRCR, of University College London, is presented in the October edition of the American Journal of Roentgenology.

Participants eyeballed 30 endoluminal CT colonography videos, all of which contained footage of a polyp and half of which were outfitted with CAD.

Eye-tracking technology showed that the radiologists saw—but then ignored or dismissed—some 97 percent of small polyps that CAD identified as positive.

The problematic polyps were much smaller than the obvious ones (mean diameter, 5.4 vs 8.2 mm, respectively p = 0.014) and were subjectively less conspicuous (median score, 4 vs 2; p = 0.0032).

Regardless of the presence of CAD, radiologists’ eyes lingered longer on the easy-to-identify polyps (42.6 percent of total viewing time) than on the difficult ones (29 percent).

From these findings, the authors suggested that efforts to improve sensitivity in CT colonography ought to focus on classification decisions rather than on image search alone.

In their study discussion, they note that their findings are surprising because “it is widely believed that small and inconspicuous polyps are missed because they are not seen rather than because they are seen but were then categorized incorrectly."

Plumb and colleagues conclude, “In essence, despite looking at these polyps, radiologists dismissed them swiftly even when they were marked correctly by CAD—suggesting that these brief inspections were too cursory to allow correct characterization.”