September 2011
Volume 11, Issue 11
Free
Vision Sciences Society Annual Meeting Abstract  |   September 2011
Why don't Computer Aided Detection (CAD) algorithms help experts as much as they should?
Author Affiliations
  • Trafton Drew
    Brigham and Women's Hospital
    Harvard Medical School
  • Corbin Cunningham
    Brigham and Women's Hospital
  • Jeremy M. Wolfe
    Brigham and Women's Hospital
    Harvard Medical School
Journal of Vision September 2011, Vol.11, 1337. doi:https://doi.org/10.1167/11.11.1337
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Trafton Drew, Corbin Cunningham, Jeremy M. Wolfe; Why don't Computer Aided Detection (CAD) algorithms help experts as much as they should?. Journal of Vision 2011;11(11):1337. https://doi.org/10.1167/11.11.1337.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Radiologists are extremely good at difficult medical visual search tasks, but far from perfect. Computer Aided Detection (CAD) programs have been developed to improve radiologists' performance. However, though the CAD systems perform well, adding CAD does not produce the gains in radiologist performance that one would expect. In some studies, CAD does not improve performance (d′) (e.g. Gur et al., 2004). In others, there is evidence for improvement, but the effects are surprisingly small (e.g. Birdwell et al., 2005). Traditional CAD systems mark areas that exceed some threshold: One point on the CAD ROC, representing a specific tradeoff between CAD misses and false alarms. Locations generating very high CAD signals produce the same marks as near-threshold locations. Suppose the CAD signal reflected the computer's confidence. Would this analog CAD signal improve observers' performance more than traditional binary (on/off) signals? We created stimuli defined by two noisy signals: a visible color signal (targets are redder) and an “invisible” signal that informed our CAD system. Observers were tested in four blocks: visible signal alone, visible plus binary CAD, visible plus analog CAD, or analog CAD alone. Set size was one in Experiment 1. Binary CAD was slightly but not significantly better than no CAD. However, analog CAD performed significantly better (p < .02). In Experiment 2, observers searched for a target amongst six items. Both traditional binary and our analog CAD significantly improved performance (p < .01) and again we found that analog CAD performance was better than binary CAD. There are sizable individual differences in performance on this task and other factors are important in the efficacy of CAD in clinical settings. Nevertheless, our data suggest that the form of the CAD signal can directly influence performance and that analog CAD signals may allow the computer to be more helpful to the searcher.

Ey017001. 
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×