Abstract
Serial Dependence is a perceptual phenomenon in which the visual system tends to bias representations toward recent visual history (Fischer & Whitney, 2014; Cicchini, et al., 2014). Studies under both lab and naturalistic scenarios have shown that radiologists’ perception is influenced by serial dependence, which could have a negative impact on diagnostic accuracy (Manassi et al., 2019; Manassi et. al., 2021; Ren et. al., 2022). However, those studies often suffer from a dearth of expert participants; typically, less than 20 experts are recruited per experiment. In this study, we analyzed a large volume of diagnostic data from a commercially available mobile app. All of the 7,798 skin cancer images used were genuine and were retrieved from either public databases or medical clients and were randomly selected and randomly ordered for each participant. In total, 756,001 diagnostic judgments from 1,137 medical students and residents were analyzed. We found an effect of serial dependence, such that judgments on the current image were pulled toward the observer's previous experience. Importantly, this effect displayed both feature-tuning and temporal-tuning, two of the diagnostic criteria of serial dependence. With respect to feature-tuning, although the sequential images were completely random, there was stronger serial dependence when two medical images happened to be similar in severity (estimated based on popularity vote). Utilizing a semantic similarity metric in deep learning, we also found larger serial dependence effects when two skin cancer images were, by chance, semantically similar. We also found that shorter inter-trial intervals led to stronger serial dependence, suggesting a kind of temporal tuning of the effect. This study suggests that serial dependence may negatively impact skin cancer diagnoses in a realistic diagnostic scenario. It also hints at the possible methods in which these biases can be mitigated.