1.IntroductionClinical diagnosis based on radiographs is not always perfect because of misperceptions and misinterpretations.1,2 Some sources of interpretive error have been identified and characterized; these include search and recognition errors,3,4 cognitive biases,2,5 search satisfaction,6,7 subsequent search misses,8–10 and low prevalence.11–17 However, some other errors in cancer image interpretation are still without explanation.18–20 Thus a great deal of research has been carried out in the last several decades to identify and characterize the sources of these errors in order to mitigate them. Radiologists often read dozens or hundreds of radiographs in batches,21 sometimes looking at several related images one after the other. Their job is to localize the lesions (if present) and then to recognize them by judging their size, class, and so on. A main underlying assumption here is that radiologists’ perceptual decisions about the current radiograph are independent of prior perceptual experience. Recent theoretical and empirical research suggests that this assumption is not true. For example, the human visual system is characterized by visual serial dependency, a type of sequential effect in which what was previously experienced influences (captures) what is seen and reported at this moment.22,23 Serial dependencies can manifest in several domains, such as perception,23–26 decision making,27,28 and memory,29–31 and they occur with a variety of features and objects, including orientation,23,32 position,26,33 faces,34,35 attractiveness,36–38 ambiguous objects,39 ensemble coding of orientation,32 and numerosity.24,40 Serial dependence is characterized by three main kinds of tuning. First is feature tuning: serial dependence occurs only between similar features and not between dissimilar ones.23,26,32,41 Second is temporal tuning: serial dependence gradually decays over time.23,26,39 Third is spatial tuning: serial dependence occurs only within a limited spatial window; it is strongest when previous and current objects are presented at the same location, and it gradually decays as the relative distance increases.23,26,33,42 In addition, attention is a necessary component for serial dependence.23,43,44 Because our visual world is stable—objects that were present a moment ago tend to still be present at this moment—we benefit from serial dependence most of the time. This is because it is more efficient to simply recycle perceptual history,23,24,26 using the past to predict the present. However, this recycling is not always beneficial. When stimuli are randomly ordered or in unnatural situations—such as when the visual world is not autocorrelated or stable—serial dependence can negatively impact perceptual decisions.23,34,45 For example, visual search in clinical settings, such as reading randomly ordered radiographs or pathology slides, is a striking example in which stimuli may not be autocorrelated. In this case, the past may not be a good predictor of the present, and serial dependence in perceptual decisions would be problematic. In fact, empirical experiments have found that clinicians’ perceptual decisions can be biased toward the previous images that they have seen.46,47 A drawback of previous work46,47 is that serial dependence was measured with unrealistic stimuli, such as random geometric shapes superimposed on a mammogram section [Fig. 2(a)]. Although well-controlled, these images are clearly inauthentic and are therefore far from naturalistic mammograms.46,47 Unfortunately, because serial dependence has only been measured with unrealistic stimuli, it remains unclear whether serial dependence in perceptual judgments would even occur for truly realistic radiographs. In this study, we aim to measure the presence of sequential effects in the perceptual decisions of observers who view controlled, realistic generative adversarial network (GAN)-generated radiographs. To accomplish this, we created authentic-looking medical images generated by a computer vision model. The model allows precise control over the stimulus space, while simultaneously ensuring that the simulated radiographs are realistic. In fact, a previous study found that these images are indistinguishable from (i.e., metameric with) down-sampled real radiographs, even to many professional clinicians.48,49 We hypothesize that even with authentic-looking simulated mammograms, perceptual decisions about any given current image will be biased toward the previously seen images due to serial dependence. 2.Methods2.1.Mammogram GenerationIn computer vision, generative models50,51 have been utilized for authentic image generation for years. In particular, GANs are a promising method to create authentic images in different modalities of human faces, places, animals, cars, etc.52–54 Similar approaches have also been applied for medical image generation.48,55–57 In this study, we adopted a controllable medical image generation method48,49 to create all stimuli used in our experiments. Because of the GAN generation paradigm, the generated samples share the similar data distribution as the real samples maintaining the variety of the real ones. A comparison between downsampled real samples and generated samples is shown in Fig. 1. Once the GAN was pretrained, we randomly sampled points on the straight lines connecting every pair of 3 anchor points in the latent space, and generated the images corresponding to these points. Each anchor point and the corresponding interpolations are latent vectors with the size of 512 (see the publications48,49 for thorough details about the model and latent space.). Then we passed the anchors as well as the interpolations through the pretrained generator to generate the corresponding images, forming a circular continuum [Fig. 2(b)]. One hundred and forty-seven images (48 between each anchor) were generated on the circular continuum with size . (The reason we used was for proof of concept and because the training takes exponentially longer with higher-resolution images.). In the experiment, 20 circular continuums such as this were generated by creating 20 sets of anchors and passing them along with the corresponding interpolations to the pretrained generator. In total, we generated 2940 images. Four example continua are shown in Fig. 3. Because the images within a given continuum were generated based on interpolations in the latent space, nearby images on the circular continuum tend to be more similar, whereas distant images on the circular continuum tend to differ from each other. With random picking, the generated image sequence can represent a certain variety of the real samples. Moreover, moving around the circular continuum, the tissue texture, tumor size, tumor location, and other semantic properties gradually change and return to the same place when looping through all of the GAN-generated mammograms on this circular continuum. 2.2.DatasetTraining data for the GAN are from the digital database for screening mammography (DDSM).58 It contains 2620 normal, benign, and malignant cases with verified pathology information. The images were first center cropped and then resized to for training. To generate stimuli containing tumors for the visual search task, only benign and malignant cases were utilized for training. Several downsampled real samples are shown in Fig. 1. 2.3.Participants and ApparatusAll experimental procedures were approved by and conducted in accordance with the guidelines and regulations of the UC Berkeley Institutional Review Board. Participants provided informed consent in accordance with the IRB guidelines of the University of California at Berkeley. All participants had normal or corrected-to-normal vision and were all naïve to the purpose of the experiment. 80 nonexpert participants (28 males, aged 18 to 72, and 52 females, aged 18 to 62) participated in the experiment. They were students and affiliates at UC Berkeley. Experiments were coded with PsychoPy and published on Pavlovia. Participants were able to access the experiment by themselves through the Internet. Sets of 4 participants were assigned to the same circular continuum, and there were 20 circular continuums in total (for a total of 80 observers). Participants used a keyboard for all responses. 2.4.Experiment DesignThe 20 circular morph continua mentioned in Sec. 2.1 were used to test the perceptual decisions of the participants. Each simulated mammogram of any continuum contains a particular pattern of lesions and texture, and these characteristics gradually change along the circular continuum. On each trial, participants viewed a random simulated mammogram, which was randomly extracted from one of the 20 circular continua, mimicing the randomness in real diagnostic scenarios. The simulated mammogram was presented for 500 ms. Next, we presented a mask composed of random Gaussian noise for 1000 ms (to avoid the possibility of afterimages). After the mask, a random simulated mammogram drawn from the same morph continuum appeared at the fixation point location, and participants were asked to adjust the simulated mammogram to match the perceived simulated mammogram using the left/right arrow keys (continuous report, adjustment task; left–right arrow keys to adjust the simulated mammogram along the circular morph continuum). The starting simulated mammogram was randomized on each trial. Participants were allowed to take as much time as necessary to respond and pressed the space bar to confirm that the chosen simulated mammogram was the correct match. Following the response and a 250 ms delay, the next trial started. A concise experiment pipeline can be found in Fig. 4(b). During the experiment, participants were asked to continuously fixate a black dot in the center. In total, each participant performed 3 blocks of 85 trials. Between each block, participants were allowed to take a break. In a preliminary session, observers completed a practice block of 10 trials to familiarize themselves with this experiment. Among the 80 participants, 3 participants were removed from data analysis because they hit the space bar all of the time during the experiment without any adjustments. 2.5.Data AnalysisThe response error was computed as the smallest difference along the morph continuum between the match morph and the target morph (current match morph–current target morph). For each participant’s data, trials were removed if the response error was 3 standard deviations away from the mean response error or if the response time was longer than 20 s. The average reaction time was . Previous research shows that individual observers can have idiosyncratic biases in object recognition and localization, which are unrelated to serial dependence.59,60 For example, observers may make a consistent error in reporting a simulated lesion of 20 morph units as being 10, thus creating a systematic error of units. Conversely, if there was no systematic error, all error would approximate zero. For this reason, we conducted an additional data processing strategy to remove such potential unrelated biases before further analyses. We modeled observers’ response error as a function of the target morph presented by fitting a radial basis function in which 30 Gaussian kernels are utilized. This allowed us to quantify the idiosyncratic bias for each observer. We then regressed out the bias quantified by the radial basis fit by subtracting it from the observer’s error. This subtraction left us with residual errors that did not include the idiosyncratic biases unrelated to serial dependence. 2.5.1.Feature tuning analysisThe difference in morphs between the current and previous trial is computed as the smallest difference along the morph continuum between the previous target morph (-back) and the current target morph (previous target morph–current target morph). To quantify the feature tuning characteristic of serial dependence, we fit a derivative of von Mises distribution to each observer’s data points. The derivative of von Mises distribution is expressed by the following equation: where parameter is the response error on each trial, is the relative orientation of the previous trial, is the amplitude modulation parameter of the derivative-of-von-Mises (DoVM), is the symmetry axis of the derivative of von Mises distribution, is the concentration of the derivative of von Mises distribution, and is the modified Bessel function of order 0. In our experiments, is set to 0. We fitted the derivative of von Mises using constrained nonlinear minimization of the residual sum of squares. As a measure of serial dependence, we reported half the peak-to-trough amplitude of the DoVM.Additionally, for each observer, we computed the running circular average within a 20 morph units window. Figure 5 (blue line) shows the average of the moving averages across all observers and the corresponding DoVM fit. 2.5.2.Temporal tuning analysisIn this study, we report half the peak-to-trough amplitude of the DoVM as a measure of serial dependence (Fig. 5). Sequentially, we can get the strength of 1-back, 2-back, and 3-back serial dependence effects by fitting the derivative of von Mises distribution on the data points, where the difference in morphs between the current and previous trial is computed as the smallest difference along the morph continuum between the 1-, 2-, and 3-trial back target morph and the current target morph. Additionally, as a control analysis, we explored the effect of future trials on the current response to check for potential unrelated biases and artifacts that might be lurking in the data.30,61 In particular, we calculated whether the current trial response error depended in some fashion on the difference in stimuli between the current and 1-forward (following) trials. Because observers have not seen the future trial stimulus, their current response in a given trial should not be influenced by the future morph stimuli. If there are artifacts in the data, however, (for example, observers perseverate on a particular response from trial to trial), there might appear to be an effect of future stimuli on the current response. This analysis reveals and serves as a control for such artifacts.23,47 BootstrappingFor each result that we obtained, we resampled the data with replacement, processed the sampled data recursively for 5000 times, and reported the mean result with 95% confidence intervals. Permutation testSignificance testing was done through permutation tests. Data were randomly shuffled and processed 5000 times. The 97.5% upper bound of the permuted null distribution was compared with the error bar from bootstrapping to confirm the significance of the result. In an additional analysis, to more intuitively convey the magnitude of the serial dependence effect, we analyzed the percentage difference between pro-SD (pulling effect due to serial dependence) and anti-SD (repelling effect against serial dependence) for 1, 2, and 3 trials back. Stimuli on the circular continuum were categorized into three types according to the nearest anchor images. Trials in which the response image was not within the same category as the target image were considered classification errors, which are misjudgments of the image category. Classification errors that are in a direction consistent with the previously seen stimulus are pro-SD errors, and those that are in a direction opposite the previously seen stimulus are anti-SD errors. In principle, classification errors should be randomly distributed, not biased in either direction. As a sanity check, we also analyzed the percentage difference between pro-SD and anti-SD for 1 trial forward because future trials naturally are not correlated with current trials. 3.ResultsThe goal of this experiment was to test whether perceptual decisions on consecutive realistic GAN-generated images of mammograms were biased toward the previously seen images. Here the observers’ response error in a particular trial was computed as the shortest distance along the morph continuum between the actual observed shape and the chosen answer shape. The average response error was units, and the average reaction time was . To test whether there are sequential effects in observers’ judgments of realistic GAN-generated mammograms, we first analyzed the response error in relation to the difference in stimulus shape between the current and previous trials for each continuum separately. Then we fitted a DoVM function to this data (Fig. 5). We operationalized serial dependence, the pull toward previous stimuli, as the half amplitude of the DoVM curve of each continuum. We bootstrapped the half amplitude and reported the average bootstrapped half amplitude for each continuum: all continua showed a positive half amplitude [Fig. 6(a)]. Importantly, the average half amplitude across all continua was significant (average bootstrapped 1-back half amplitude = 2.77 morph units, , permutation analysis), which suggests an influence of the simulated radiograph in the previous trial on the current response. The influence of previous stimuli extended to two trials back (average bootstrapped 2-back half amplitude = 1.38 morph units, , permutation analysis). By contrast, the stimuli presented three trials prior had no influence on the current response (average bootstrapped 3-back half amplitude = 0.09 morph units, , permutation analysis). To control for artifacts, we calculated the influence of the stimuli presented in the next trial on the current response. We found a modest bias, as found in previous studies of sequential effects,61,62 but, importantly, the 1-back and 2-back effects were significantly larger than this 1-forward baseline (1-back versus 1-forward: ; 2-back versus 1-forward: . This confirms that there are sequential effects in perceptual decisions about realistic GAN-generated mammograms. To quantify the serial dependence effect in an alternative manner, we also analyzed the percentage difference between pro-SD and anti-SD classification errors [Fig. 6(b)]. Overall, the classification error rate is 28.35%. The 1-back and 2-back percentage differences were 6.85% and 3.96%, respectively, indicating the dominance of serial dependence in the sequential effects in perceptual decisions of participants. Essentially, when there are classification errors, these are much more likely to be in the direction of previous stimuli. The 3-back percentage difference was 0.1%. Overall, serial dependence dominated the sequential effects for 1 and 2 trials back. In addition, the sanity check of 1 trial forward, 0.03%, shows no influence of future trials on classification errors in the current trial. This is expected and confirms that there were no artifacts masquerading as serial dependence. 4.DiscussionSerial dependence in medical image perception has been studied for years.46,47 However, none of the previous research used realistic medical images. In previous studies, the stimuli incorporated simple geometric shapes and artificial backgrounds consisting of either healthy tissue texture or simple noise patterns. Although the prior empirical results indicate the existence of serial dependence in the perception of those unrealistic stimuli, whether serial dependence extends to and occurs for realistic medical images remained unknown. In this study, we tested whether there is serial dependence in perceptual judgments of more realistic GAN-generated radiographs. We utilized authentic-looking simulated medical image stimuli created with a GAN.48,49 The magnitude of serial dependence found in the current study was similar to that found in previous studies. Prior studies found that perceptual judgments were pulled toward the stimulus presented in the previous trial, and the pull effect was around 15% for 1-back trials. Moreover, this effect lasted up to 10 s or more in the past.47 The results in this study were comparable. For example, the half amplitudes of the DoVM curve in Fig. 6 show a similar effect size as that previously reported. This indicates that serial dependence affects untrained observers’ judgments of the simulated radiographs. The fact that clinicians show serial dependence in other domains,46,47 and the fact that serial dependence can increase with expertise63 hints at the possibility that clinicians may not be immune from serial dependence. Nevertheless, whether serial dependence influences clinician judgments of the more realistic GAN-generated radiographs here remains an important question for future research. In addition to replicating and extending the presence of serial dependencies in perceptual judgments of realistic medical images, our study also highlights the broader point that computer vision tools can be used in concert with psychophysical experiments to isolate and shed light on human performance limits. Computer vision models, in this approach, are not employed with the goal of replacing human readers. Rather, computer vision is used to create controlled stimuli that allow human performance to be more accurately assessed, controlled, and potentially enhanced. Computer vision models are in the service of human behavior. There are several caveats and concerns that readers may have noted. It may be argued, for example, that the presentation duration of the simulated mammogram was too short (500 ms) or too low resolution () in our study, whereas clinicians typically have longer periods of time to process higher-resolution radiographs. In fact, the average fixation duration when targeting the first mass has been reported as 1.8 to 2 s, which is surprisingly brief.4,64 Moreover, when scrolling through volumetric images, the viewing time in any given slice can be a fraction of a second. In addition, peripheral viewing and effectively lower resolution images can be sufficient for detecting abnormalities.15,65,66 Conversely, images viewed for a sufficiently long exposure duration can lead to negative aftereffects. For example, it was found that adapting normal observers to image samples of dense or fatty tissues caused a subsequent image to appear less dense (and vice versa; a type of negative aftereffect).67–69 Sequential effects (either repulsive or attractive) can therefore emerge across many different exposure durations. In addition to the fixed duration of the stimuli in this experiment, this study has some additional limitations. First, we chose a continuous report matching task in our experiments as it provides precise trial-wise errors and has proven to be very reliable in measurements of serial dependence in the past.23,24,34,41,43,70 However, the actual task of the typical radiologist is far more complicated and involves detecting, locating, and classifying the lesions. Future studies should therefore implement more realistic tasks. Second, we only tested untrained observers in this study. Future studies should also recruit clinician observers. Third, the simulated mammograms were only presented briefly in our experiment to mimic the brevity of images viewed in quick succession. To generalize the results here, it will be necessary to test which biases arise with longer presentation durations. Fourth, even though we utilized both benign and malignant images for training, we did not consider the malignancy of the stimuli in the GAN model and experiments. Future studies can investigate how malignancy can be disentangled in the GAN model and how malignancy may influence the diagnostic tasks. Our goal in this study was to test the presence of sequential effects in judgments of more realistic and controlled GAN generated medical images, and we found evidence for this. However, the caveats and concerns described here prevent us from concluding that serial dependence impacts clinical image interpretation in real clinical practice. The results raise the possibility, though, and if there are serial dependencies in clinical interpretations, then the consecutive similarity between images from one or more patients could matter. Future work is needed to test this. 5.ConclusionIn this study, we utilized a GAN to produce authentic-looking GAN-generated mammograms. These realistic stimuli were used in a psychophysical experiment that tested for serial dependence in perceptual judgments. We found that the perception of the current simulated mammogram was biased toward the previously seen mammograms. On average, perceptual judgments of naturalistic GAN-generated mammograms had 7% categorization errors that were pulled in a direction consistent with serial dependence, and this pulling effect lasted up to 10 s in the past. Our study provides evidence that serial dependence may contribute to the decision errors in the perception of realistic-looking medical images. AcknowledgmentsThis work was supported by the National Institutes of Health (Grant No. R01CA236793). ReferencesL. Berlin et al.,
“Accuracy of diagnostic procedures: has it improved over the past five decades?,”
Am. J. Roentgenol., 188
(5), 1173
–1178 https://doi.org/10.2214/AJR.06.1270 AJROAM 0092-5381
(2007).
Google Scholar
P. Croskerry,
“The importance of cognitive errors in diagnosis and strategies to minimize them,”
Acad. Med., 78
(8), 775
–780 https://doi.org/10.1097/00001888-200308000-00003 ACMEEO 1040-2446
(2003).
Google Scholar
D. P. Carmody, C. F. Nodine and H. L. Kundel,
“An analysis of perceptual and cognitive factors in radiographic interpretation,”
Perception, 9
(3), 339
–344 https://doi.org/10.1068/p090339 PCTNBA 0301-0066
(1980).
Google Scholar
C. F. Nodine et al.,
“Nature of expertise in searching mammograms for breast masses,”
Acad. Radiol., 3
(12), 1000
–1006 https://doi.org/10.1016/S1076-6332(96)80032-8
(1996).
Google Scholar
C. S. Lee et al.,
“Cognitive and system factors contributing to diagnostic errors in radiology,”
Am. J. Roentgenol., 201
(3), 611
–617 https://doi.org/10.2214/AJR.12.10375 AJROAM 0092-5381
(2013).
Google Scholar
C. J. Ashman, J. S. Yu and D. Wolfman,
“Satisfaction of search in osteoradiology,”
Am. J. Roentgenol., 175
(2), 541
–544 https://doi.org/10.2214/ajr.175.2.1750541 AJROAM 0092-5381
(2000).
Google Scholar
K. S. Berbaum and Jr. E. A. Franken,
“Satisfaction of search in radiographic modalities,”
Radiology, 261
(3), 1000
–1001 https://doi.org/10.1148/radiol.11110987 RADLAX 0033-8419
(2011).
Google Scholar
R. L. Birdwell et al.,
“Mammographic characteristics of 115 missed cancers later detected with screening mammography and the potential utility of computer-aided detection,”
Radiology, 219
(1), 192
–202 https://doi.org/10.1148/radiology.219.1.r01ap16192 RADLAX 0033-8419
(2001).
Google Scholar
B. Boyer et al.,
“Retrospectively detectable carcinomas: review of the literature,”
J. Radiol., 85
(12 Pt 2), 2071
–2078 https://doi.org/10.1016/S0221-0363(04)97784-0
(2004).
Google Scholar
J. A. Harvey, L. L. Fajardo and C. Innis,
“Previous mammograms in patients with impalpable breast carcinoma: retrospective vs blinded interpretation. 1993 ARRS president’s award,”
Am. J. Roentgenol., 161
(6), 1167
–1172 https://doi.org/10.2214/ajr.161.6.8249720 AJROAM 0092-5381
(1993).
Google Scholar
J. M. Wolfe, T. S. Horowitz and N. M. Kenner,
“Rare items often missed in visual searches,”
Nature, 435
(7041), 439
–440 https://doi.org/10.1038/435439a
(2005).
Google Scholar
J. M. Wolfe et al.,
“Low target prevalence is a stubborn source of errors in visual search tasks,”
J. Exp. Psychol. Gen., 136
(4), 623 https://doi.org/10.1037/0096-3445.136.4.623 JPGEDD 1939-2222
(2007).
Google Scholar
A. N. Rich et al.,
“Why do we miss rare targets? Exploring the boundaries of the low prevalence effect,”
J. Vis., 8
(15), 15
–15 https://doi.org/10.1167/8.15.15 1534-7362
(2008).
Google Scholar
T. Menneer et al.,
“High or low target prevalence increases the dual-target cost in visual search,”
J. Exp. Psychol. Appl., 16
(2), 133 https://doi.org/10.1037/a0019569 1076-898X
(2010).
Google Scholar
K. K. Evans, R. L. Birdwell and J. M. Wolfe,
“If you don’t find it often, you often don’t find it: why some cancers are missed in breast cancer screening,”
PLoS One, 8
(5), e64366 https://doi.org/10.1371/journal.pone.0064366 POLNCL 1932-6203
(2013).
Google Scholar
T. S. Horowitz,
“Prevalence in visual search: from the clinic to the lab and back again,”
Jpn. Psychol. Res., 59
(2), 65
–108 https://doi.org/10.1111/jpr.12153
(2017).
Google Scholar
M. A. Kunar et al.,
“Low prevalence search for cancers in mammograms: evidence using laboratory experiments and computer aided detection,”
J. Exp. Psychol. Appl., 23
(4), 369 https://doi.org/10.1037/xap0000132 1076-898X
(2017).
Google Scholar
M. A. Bruno, E. A. Walker and H. H. Abujudeh,
“Understanding and confronting our mistakes: the epidemiology of error in radiology and strategies for error reduction,”
Radiographics, 35
(6), 1668
–1676 https://doi.org/10.1148/rg.2015150023
(2015).
Google Scholar
S. Waite et al.,
“Interpretive error in radiology,”
Am. J. Roentgenol., 208
(4), 739
–749 https://doi.org/10.2214/AJR.16.16963 AJROAM 0092-5381
(2017).
Google Scholar
S. Waite et al.,
“Analysis of perceptual expertise in radiology–current knowledge and a new perspective,”
Front. Hum. Neurosci., 13 213 https://doi.org/10.3389/fnhum.2019.00213
(2019).
Google Scholar
R. J. McDonald et al.,
“The effects of changes in utilization and technological advancements of cross-sectional imaging on radiologist workload,”
Acad. Radiol., 22
(9), 1191
–1198 https://doi.org/10.1016/j.acra.2015.05.007
(2015).
Google Scholar
G. M. Cicchini, G. Anobile and D. C. Burr,
“Compressive mapping of number to space reflects dynamic encoding mechanisms, not static logarithmic transform,”
Proc. Natl. Acad. Sci. U. S. A., 111
(21), 7867
–7872 https://doi.org/10.1073/pnas.1402785111
(2014).
Google Scholar
J. Fischer and D. Whitney,
“Serial dependence in visual perception,”
Nat. Neurosci., 17
(5), 738
–743 https://doi.org/10.1038/nn.3689 NANEFN 1097-6256
(2014).
Google Scholar
G. M. Cicchini, K. Mikellidou and D. Burr,
“Serial dependencies act directly on perception,”
J. Vis., 17
(14), 6 https://doi.org/10.1167/17.14.6 1534-7362
(2017).
Google Scholar
G. M. Cicchini, K. Mikellidou and D. C. Burr,
“The functional role of serial dependence,”
Proc. R. Soc. B, 285
(1890), 20181722 https://doi.org/10.1098/rspb.2018.1722 PRLBA4 0080-4649
(2018).
Google Scholar
M. Manassi et al.,
“Serial dependence in position occurs at the time of perception,”
Psychonom. Bull. Rev., 25
(6), 2245
–2253 https://doi.org/10.3758/s13423-018-1454-5 PBUREN 1069-9384
(2018).
Google Scholar
A. Abrahamyan et al.,
“Adaptable history biases in human perceptual decisions,”
Proc. Natl. Acad. Sci. U. S. A., 113
(25), E3548
–E3557 https://doi.org/10.1073/pnas.1518786113
(2016).
Google Scholar
S. W. Fernberger,
“Interdependence of judgments within the series for the method of constant stimuli,”
J. Exp. Psychol., 3
(2), 126 https://doi.org/10.1037/h0065212 JEPSAK 0022-1015
(1920).
Google Scholar
J. Barbosa and A. Compte,
“Build-up of serial dependence in color working memory,”
Sci. Rep., 10
(1), 10959 https://doi.org/10.1038/s41598-020-67861-2 SRCEC3 2045-2322
(2020).
Google Scholar
M. Fornaciai and J. Park,
“Attractive serial dependence between memorized stimuli,”
Cognition, 200 104250 https://doi.org/10.1016/j.cognition.2020.104250 CGTNAU 0010-0277
(2020).
Google Scholar
A. Kiyonaga et al.,
“Serial dependence across perception, attention, and memory,”
Trends Cogn. Sci., 21
(7), 493
–497 https://doi.org/10.1016/j.tics.2017.04.011 TCSCFK 1364-6613
(2017).
Google Scholar
M. Manassi et al.,
“The perceived stability of scenes: serial dependence in ensemble representations,”
Sci. Rep., 7
(1), 1971 https://doi.org/10.1038/s41598-017-02201-5 SRCEC3 2045-2322
(2017).
Google Scholar
D. P. Bliss, J. J. Sun and M. D’Esposito,
“Serial dependence is absent at the time of perception but increases in visual working memory,”
Sci. Rep., 7
(1), 14739 https://doi.org/10.1038/s41598-017-15199-7 SRCEC3 2045-2322
(2017).
Google Scholar
A. Liberman, J. Fischer and D. Whitney,
“Serial dependence in the perception of faces,”
Curr. Biol., 24
(21), 2569
–2574 https://doi.org/10.1016/j.cub.2014.09.025 CUBLE2 0960-9822
(2014).
Google Scholar
J. Taubert, D. Alais and D. Burr,
“Different coding strategies for the perception of stable and changeable facial attributes,”
Sci. Rep., 6
(1), 32239 https://doi.org/10.1038/srep32239 SRCEC3 2045-2322
(2016).
Google Scholar
J. Taubert, E. Van der Burg and D. Alais,
“Love at second sight: sequential dependence of facial attractiveness in an on-line dating paradigm,”
Sci. Rep., 6
(1), 22740 https://doi.org/10.1038/srep22740 SRCEC3 2045-2322
(2016).
Google Scholar
Y. Xia, A. Y. Leib and D. Whitney,
“Serial dependence in the perception of attractiveness,”
J. Vis., 16
(15), 28 https://doi.org/10.1167/16.15.28 1534-7362
(2016).
Google Scholar
A. Kondo, K. Takahashi and K. Watanabe,
“Sequential effects in face-attractiveness judgment,”
Perception, 41
(1), 43
–49 https://doi.org/10.1068/p7116 PCTNBA 0301-0066
(2012).
Google Scholar
M. Wexler, M. Duyck and P. Mamassian,
“Persistent states in vision break universality and time invariance,”
Proc. Natl. Acad. Sci. U. S. A., 112
(48), 14990
–14995 https://doi.org/10.1073/pnas.1508847112
(2015).
Google Scholar
J. E. Corbett, J. Fischer and D. Whitney,
“Facilitating stable representations: serial dependence in vision,”
PLoS One, 6
(1), e16701 https://doi.org/10.1371/journal.pone.0016701 POLNCL 1932-6203
(2011).
Google Scholar
M. Fritsche, P. Mostert and F. P. de Lange,
“Opposite effects of recent history on perception and decision,”
Curr. Biol., 27
(4), 590
–595 https://doi.org/10.1016/j.cub.2017.01.006 CUBLE2 0960-9822
(2017).
Google Scholar
T. Collins,
“The perceptual continuity field is retinotopic,”
Sci. Rep., 9
(1), 18841 https://doi.org/10.1038/s41598-019-55134-6 SRCEC3 2045-2322
(2019).
Google Scholar
M. Fritsche and F. P. de Lange,
“The role of feature-based attention in visual serial dependence,”
J. Vis., 19
(13), 21 https://doi.org/10.1167/19.13.21 1534-7362
(2019).
Google Scholar
S. Kim et al.,
“Serial dependence in perception requires conscious awareness,”
Curr. Biol., 30
(6), R257
–R258 https://doi.org/10.1016/j.cub.2020.02.008 CUBLE2 0960-9822
(2020).
Google Scholar
I. Fründ, F. A. Wichmann and J. H. Macke,
“Quantifying the effect of intertrial dependence on perceptual decisions,”
J. Vis., 14
(7), 9 https://doi.org/10.1167/14.7.9 1534-7362
(2014).
Google Scholar
M. Manassi, Á. Kristjánsson and D. Whitney,
“Serial dependence in a simulated clinical visual search task,”
Sci. Rep., 9
(1), 19937 https://doi.org/10.1038/s41598-019-56315-z SRCEC3 2045-2322
(2019).
Google Scholar
M. Manassi et al.,
“Serial dependence in the perceptual judgments of radiologists,”
Cogn. Res. Princ. Implic., 6
(1), 65 https://doi.org/10.1186/s41235-021-00331-z
(2021).
Google Scholar
Z. Ren, S. X. Yu and D. Whitney,
“Controllable medical image generation via generative adversarial networks,”
in Proc. IS&T Int. Symp. Electronic Imaging: Human Vision and Electronic Imaging,
112-1
–112-6
(2021). https://doi.org/10.2352/ISSN.2470-1173.2021.11.HVEI-112 Google Scholar
Z. Ren, S. X. Yu and D. Whitney,
“Controllable medical image generation via GAN,”
J. Percept. Imaging, 5 1
–15 https://doi.org/10.2352/J.Percept.Imaging.2022.5.000502
(2022).
Google Scholar
D. P. Kingma and M. Welling,
“Auto-encoding variational Bayes,”
in 2nd Int. Conf. Learn. Represent.,
(2014). Google Scholar
I. Goodfellow et al.,
“Generative adversarial nets,”
in Adv. Neural Inf. Process. Syst.,
(2014). Google Scholar
I. Gulrajani et al.,
“Improved training of Wasserstein GANs,”
in Adv. Neural Inf. Process. Syst.,
(2017). Google Scholar
T. Karras et al.,
“Progressive growing of GANs for improved quality, stability, and variation,”
in Int. Conf. Learn. Represent.,
(2018). Google Scholar
T. Karras, S. Laine and T. Aila,
“A style-based generator architecture for generative adversarial networks,”
in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit.,
4401
–4410
(2019). https://doi.org/10.1109/CVPR.2019.00453 Google Scholar
C. Han et al.,
“GAN-based synthetic brain MR image generation,”
in IEEE 15th Int. Symp. Biomed. Imaging (ISBI 2018),
734
–738
(2018). https://doi.org/10.1109/ISBI.2018.8363678 Google Scholar
M. Frid-Adar et al.,
“GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification,”
Neurocomputing, 321 321
–331 https://doi.org/10.1016/j.neucom.2018.09.013 NRCGEO 0925-2312
(2018).
Google Scholar
Z. Ren et al.,
“Improve image-based skin cancer diagnosis with generative self-supervised learning,”
in IEEE/ACM Conf. Connected Health: Appl., Syst. and Eng. Technol. (CHASE),
23
–34
(2021). https://doi.org/10.1109/CHASE52844.2021.00011 Google Scholar
K. Bowyer et al.,
“The digital database for screening mammography,”
in Third Int. Workshop on Digit. Mammogr.,
27
(1996). Google Scholar
Z. Wang, Y. Murai and D. Whitney,
“Idiosyncratic perception: a link between acuity, perceived position and apparent size,”
Proc. R. Soc. B, 287
(1930), 20200825 https://doi.org/10.1098/rspb.2020.0825 PRLBA4 0080-4649
(2020).
Google Scholar
A. Kosovicheva and D. Whitney,
“Stable individual signatures in object localization,”
Curr. Biol., 27
(14), R700
–R701 https://doi.org/10.1016/j.cub.2017.06.001 CUBLE2 0960-9822
(2017).
Google Scholar
G. W. Maus et al.,
“The challenge of measuring long-term positive aftereffects,”
Curr. Biol., 23
(10), R438
–R439 https://doi.org/10.1016/j.cub.2013.03.024 CUBLE2 0960-9822
(2013).
Google Scholar
A. Chopin and P. Mamassian,
“Predictive properties of visual adaptation,”
Curr. Biol., 22
(7), 622
–626 https://doi.org/10.1016/j.cub.2012.02.021 CUBLE2 0960-9822
(2012).
Google Scholar
K. Turbett et al.,
“Individual differences in serial dependence of facial identity are associated with face recognition abilities,”
Sci. Rep., 9 18020 https://doi.org/10.1038/s41598-019-53282-3
(2019).
Google Scholar
E. A. Krupinski,
“Visual scanning patterns of radiologists searching mammograms,”
Acad. Radiol., 3
(2), 137
–144 https://doi.org/10.1016/S1076-6332(05)80381-2
(1996).
Google Scholar
K. K. Evans et al.,
“The gist of the abnormal: above-chance medical decision making in the blink of an eye,”
Psychonom. Bull. Rev., 20 1170
–1175 https://doi.org/10.3758/s13423-013-0459-3 PBUREN 1069-9384
(2013).
Google Scholar
P. C. Brennan et al.,
“Radiologists can detect the ‘gist’ of breast cancer before any overt signs of cancer appear,”
Sci. Rep., 8
(1), 8717 https://doi.org/10.1038/s41598-018-26100-5 SRCEC3 2045-2322
(2018).
Google Scholar
E. Kompaniez et al.,
“Adaptation aftereffects in the perception of radiological images,”
PLoS One, 8
(10), e76175 https://doi.org/10.1371/journal.pone.0076175 POLNCL 1932-6203
(2013).
Google Scholar
E. Kompaniez-Dunigan et al.,
“Adaptation and visual search in mammographic images,”
Atten. Percept. Psychophys., 77
(4), 1081
–1087 https://doi.org/10.3758/s13414-015-0841-5
(2015).
Google Scholar
E. Kompaniez-Dunigan et al.,
“Visual adaptation and the amplitude spectra of radiological images,”
Cogn. Res. Princ. Implic., 3
(1), 3 https://doi.org/10.1186/s41235-018-0089-4
(2018).
Google Scholar
G. M. Cicchini, A. Benedetto and D. C. Burr,
“Perceptual history propagates down to early levels of sensory analysis,”
Curr. Biol., 31
(6), 1245
–1250.e2 https://doi.org/10.1016/j.cub.2020.12.004 CUBLE2 0960-9822
(2021).
Google Scholar
BiographyZhihang Ren received his BS degree in electrical and electronic engineering from the University of Electronic Science and Technology of China and his MS degree in electrical and computer engineering from UC San Diego. He is a PhD candidate in vision science at UC Berkeley. He has been working on computer vision aided psychophysical research. His research interests also include generative models, deep learning, and machine learning. Teresa Canas-Bajo received her BS degree in psychology from the Universidad Autonoma de Madrid, her MSc degree in psychological research from the University of Edinburgh, and her PhD in vision science at UC Berkeley. Her work focuses on visual perception with a special emphasis in face recognition. Cristina Ghirardo received her BS degree in psychology from UC Berkeley. After completing her degree, she worked as a laboratory manager conducting research on the study of serial dependence and its implications in medical imaging. Mauro Manassi received his BA and MA degrees in clinical psychology from Padua University, Padua, Italy, and his PhD in neuroscience from the Swiss Federal Institute of Technology in Lausanne. He is a lecturer (assistant professor) at the School of Psychology, University of Aberdeen, UK. In his career, he has served as a postdoctoral fellow at the University of California, Berkeley. His research interests include spatial vision, serial dependence, and medical image perception. Stella X. Yu received her PhD from Carnegie Mellon University, where she studied robotics at the Robotics Institute and vision science at the Center for the Neural Basis of Cognition. Before she joined the University of Michigan faculty in Fall 2022, she has been the Director of Vision Group at the International Computer Science Institute, a senior fellow at the Berkeley Institute for Data Science, and on the Faculty of Computer Science, Vision Science, Cognitive and Brain Sciences at UC Berkeley. She is interested not only in understanding visual perception from multiple perspectives, but also in using computer vision and machine learning to automate and exceed human expertise in practical applications. David Whitney received his BA degree in psychology, economics, and philosophy from Boston University and his MA degree and PhD in psychology from Harvard University. He is a professor in the Department of Psychology and the Helen Wills Neuroscience Institute at the University of California, Berkeley. In addition, he serves as the director of cognitive sciences at the University of California, Berkeley. In his career, he has served as a postdoctoral fellow at the University of Western Ontario and an associate professor at the University of California, Davis. |
Mammography
Gallium nitride
Radiography
Medical imaging
Error analysis
Education and training
Interpolation