The Preponderance of Evidence Supports Computer-aided Detection for Screening Mammography

Complex relationships exist between humans and computers. The Association for Computing Machinery Special Interest Group on Computer-Human Interaction Curricula for Human-Computer Interaction defines this interaction in the following statement: "Human-computer interaction is a discipline concerned with the design, evaluation and implementation of interacting computing systems for human use and with the study of major phenomena surrounding them" (42). Such major phenomena include, on the machine side, techniques in graphics and programming languages and, on the human side, cognitive psychology and human performance. For some radiologists, CADe marks are annoying, and for some the marks appear to be threatening, suggesting in some way that the individual may be incompetent to interpret the image by him- or herself. These beliefs are compounded by the fact that most CADe marks do not indicate a recognizable cancer and must be dismissed. It has been estimated that as many as 2000 marks will be dismissed in the time period of a true-positive marking of a cancer. The result of this barrage of false-positive marks limits the daily perception of the reader as to any tangible benefit afforded by the CADe system. It is only from the carefully performed clinical studies (prospective sequential read or historical control studies) described above that the true impact of CADe be measured. At this time, CADe is an adjunct to the radiologist in reading mammograms - the radiologist is the final decision maker. Thus, one must not recall a patient simply because there is a CADe mark when no suspicious lesion at the marked location is interpreted by the imager. It is equally important, maybe more so, not to dismiss any suspicious finding simply because it is not marked by the CADe system. Current prospectively obtained data show that most readers are assisted by CADe, in that more cancers are detected with the use of this software tool, with a concordant reasonable increase in the recall rate. Wolfe et al (43) tell us that "low target prevalence is a stubborn source of errors in visual search tasks" and is in fact related to the prevalence effect in visual search. The effect of reducing target prevalence is to make the decision criterion more conservative; "one is less likely to call something a target if the a priori probability of a target is lower." The prevalence effect is not mere human frailty. Bond and Kamil (44) trained blue jays to hunt for artificial moths on a computer screen and found that moths with a rare wing pattern survived better than those with more common markings. As Wolfe et al (43) postulate, ". . . the prevalence effect is a sensible behavior unless the goal is to minimize misses, as in medical and airport screening tasks." Having a system to aid the human eye that does not take vacations, is not vulnerable to fatigue or environmental distractions, is without emotion, and is designed specifically to assist the very human eye to "look over here" seems like a good idea. Speaking for myself, in this busy and emotional area of imaging, I am happy to take all the help I can get. © RSNA, 2009.