Abstract
Radiologists rely on visual search to locate and identify lesions in mammograms. An underlying assumption in radiology is that the perception of each mammogram is free from the influence of preceding images; however, that may not be the case. Serial dependence, the tendency for the visual system to represent images as more similar to those previously viewed, occurs most frequently between ambiguous stimuli just like those found in radiological screening. (Cicchini, et al., PNAS, 2014; Fischer & Whitney, Nature Neuro, 2014; Liberman et al., Curr Bio, 2014; Kiyonaga et al., TiCS, 2017). Recent work has shown radiologists’ perception of simulated tumors is biased toward previously seen stimuli (Manassi et al., Sci Reports, 2019; Ghirardo et. al., VSS, 2020). This serial dependence could cause diagnostic errors; however, previous work on this hypothesis was limited to artificial, unrealistic stimuli. To overcome these limitations, we used a generative adversarial network to create naturalistic simulated mammogram images via interpolation (Ren et al., VSS, 2020), which radiologists misclassified as real mammograms. From these, we created sets of similar simulated mammograms. Using these as stimuli in a standard serial dependence experiment, untrained observers viewed a random simulated mammogram on each trial and subsequently matched the mammogram using continuous report. We found serial dependence with all of the simulated radiographs: the reported mammograms were pulled ~9-12% toward those previously seen. The effect extended back at least two trials (~10 sec). These findings suggest that serial dependence extends to realistic radiographs, and that it may contribute to some of the misdiagnoses found in radiological practice.