Abstract
Radiologists rely on visual perception to diagnose breast diseases via mammograms. A basic underlying assumption in medical image perception is that clinician visual perception and decision at any moment is largely independent of recently seen radiographs; however, that may not be true. Researchers have found that serial dependence, the tendency for the visual system to bias representations toward recent history, occurs more frequently between ambiguous stimuli just like those found in radiological screening (Cicchini, et al., 2014; Fischer & Whitney, 2014; Liberman et al., 2014; Kiyonaga et al., 2017). In particular, recent work shows that radiologist perception of simulated tumors is biased toward previously seen stimuli (Manassi et al., 2019; Ghirardo et. al., 2020). However, previous work was limited to unrealistic stimuli. To address the limitation, a generative adversarial network was developed to produce naturalistic mammogram stimuli (Ren et al., 2020). In this study, we aimed to investigate if radiologists also experience serial dependence with these realistic stimuli. In each trial, radiologists viewed a random simulated mammogram and subsequently matched the mammogram by picking the image from a morph continuum. The reported mammograms were pulled 9% towards those seen in the previous trials. These findings suggest that serial dependence extends to realistic radiographs and that it occurs even for radiologists. Serial dependence could therefore have a negative impact on the diagnostic accuracy of practicing clinicians.