Trafton Drew PhD, Presenter: Nothing to Disclose

Melissa Le-Hoa Vo PhD, Abstract Co-Author: Nothing to Disclose

Steven E. Seltzer MD, Abstract Co-Author: Nothing to Disclose

Francine L. Jacobson MD, MPH, Abstract Co-Author: Research support, Toshiba Corporation

Jeremy Michael Wolfe PhD, Abstract Co-Author: Research grant, Toshiba Corporation Consultant, The Procter & Gamble Company

We understand a great deal about how visual search is carried out in 2D scenes and medical images (e.g. Kundel et al., 2007). Very little is known about how search is accomplished in stacks of 2D images representing a 3D volume. How do radiologists search such stimuli? More importantly, how does that search behavior relate to errors?

5 trained radiologists searched lung CT scans while we monitored their eye-movements. Each radiologist had 3 minutes to read each of 12 abbreviated lung CT cases. We inserted an average of 10 artificial nodules into the otherwise clean cases. Radiologists marked nodules with a mouse click. Eye-position in X/Y space was recorded at 1000Hz and co-registered with slice/depth plane as the radiologist scrolled through the lung, allowing a 3D representation of eye position.

Radiologists varied widely in the pattern of movements through slices with some making a single pass while others repeatedly moved up and down through the stack. Regardless, all tended to hold their eyes fixed in the XY plane while moving in Z. In typical studies of errors in 2D, it is reported that a roughly equal proportion of missed lesions that are never fixated (search errors), fixated briefly (recognition errors), or fixated extensively (decision errors) (c.f. Kundel). In 3D, we found that 46% of missed nodules were never fixated, 44% were briefly fixated, while 9% were long-fixation, decision errors. 

Understanding the source of errors is critical to improving radiologist performance. While low rates of decision error may simply reflect the ease with which our inserted nodules could be identified, the relatively high rate of unfixated nodules may reflect a form of satisfaction of search error specific to the 3D stack. It is possible that viewers who pause on a slice while moving in depth, may overestimate the degree to which they have searched that slice. If so, feedback about eye position could serve as a useful warning.

Our data suggest that, when searching volumetric data, misses are frequently the result of never fixating the lesion. Completeness of search may be a useful point to emphasize when teaching radiologis

Drew, T, Vo, M, Seltzer, S, Jacobson, F, Wolfe, J, Searching in Three Dimensions: How Do Radiologists Move Their Eyes When Viewing Chest CTs?.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11009645.html