Abstract
Computer-aided detection (CAD) is widely used in medical screening tasks. While the technology holds great promise and computer algorithms continue to improve, the benefits of CAD are remarkably small in practice. In fact, recent work has shown that the presence of CAD in mammography screening clinics resulted in no net benefit (Lehman et al., 2015). Previous work investigated this surprising result by creating a visual search task that emulates important aspects of screening mammography: difficult to detect 'targets' embedded in 1/f noise (Drew et al, 2012). They found that while CAD led to a small behavioral benefit, it also led to significant costs. In particular, targets that were missed by the CAD algorithm were missed at a very high rate, and eye-tracking measures suggested that this may have been driven by less complete search in the presence of CAD. However, targets occurred in this study at a high prevalence while targets in screening mammography are extremely rare. It not currently known how target prevalence influences how CAD is used. In order to investigate this gap in our understanding, observers completed blocks of trials at both high (50%) and low (10%) prevalence. Replicating previous work, low prevalence led to shorter response times and lower hit rates. Our eye-tracking measures of coverage demonstrated reduced coverage in the CAD condition, irrespective of target prevalence. There was a significant interaction between CAD presence and prevalence on hit rate such that targets that were missed by the CAD were much more likely to be missed in the low prevalence block of trials. While helpful in theory, the benefits for CAD in practice are mixed. This work shows that this may be a result of the costs of CAD being exacerbated at low target prevalence.
Meeting abstract presented at VSS 2017