Visual search of experts in medical image reading: the effect of training, target prevalence, and expert knowledge

Determinants of Face Recognition: The Role of Target Prevalence and Similarity

Studies of facial identity processing typically assess perception (via matching) and/or memory (via recognition), with experimental designs differing with respect to one important aspect: Target Prevalence. Some designs include "target absent" (TA) among "target present" (TP) trials. In visual search tasks, TA trials shift an observer's decisional criterion towards a stricter one, increasing misses. However, decisional biases will differ between individuals and across an individual's decisions as well. In this way, excluding TA trials ensures comparable levels of expectation and thus a more controlled decisional bias both within and between observers by not considering correct rejections and false alarms. However, TA trials may occur, e.g., in police line-ups, where it is important to consider observers' face recognition ability net of the potential biases introduced by TA and TP trials. And, while these have been investigated in numerous other stimulus domains, their effects have not yet been extended to face recognition. We therefore sought to fill this void by testing different versions of the previously established Models Memory Test, which measures old/new recognition of experimentally learned facial identities. Our study found significant expectation effects, driven by target prevalence that persist even given prevalence changes. This implies that face recognition - even measured with naturalistic changes - is influenced by prior perceptual decisions.

Graded prioritisation of targets in search: reward diminishes the low prevalence effect

In many real-life contexts, observers are required to search for targets that are rarely present (e.g. tumours in X-rays; dangerous items in airport security screenings). Despite the rarity of these items, they are of enormous importance for the health and safety of the public, yet they are easily missed during visual search. This is referred to as the prevalence effect. In the current series of experiments, we investigate whether unequal reward can modulate the prevalence effect, in a multiple target search task. Having first established the impact of prevalence (Experiment 1) and reward (Experiment 2) on how efficiently participants can find one of several targets in the current paradigm, we then combined the two forms of priority to investigate their interaction. An unequal reward distribution (where lower prevalence items are more rewarded; Experiment 3) was found to diminish the effect of prevalence, compared to an equal reward distribution (Experiment 4) as indicated by faster response times and fewer misses. These findings suggest that when combined with an unequal reward distribution, the low prevalence effect can be diminished.

Hip fracture or not? The reversed prevalence effect among non-experts' diagnosis

Despite numerous investigations of the prevalence effect on medical image perception, little research has been done to examine the effect of expertise, and its possible interaction with prevalence. In this study, medical practitioners were instructed to detect the presence of hip fracture in 50 X-ray images with either high prevalence (N_signal = 40) or low prevalence (N_signal = 10). Results showed that compared to novices (e.g., pediatricians, dentists, neurologists), the manipulation of prevalence shifted participant's criteria in a different direction for experts who perform hip fracture diagnosis on a daily basis. That is, when prevalence rate is low (p_{fracture-present} = 0.2), experts held more conservative criteria in answering "fracture-present," whereas novices were more likely to believe there was fracture. Importantly, participants' detection discriminability did not vary by the prevalence condition. In addition, all participants were more conservative with "fracture-present" responses when task difficulty increased. We suspect the apparent opposite criteria shift between experts and novices may have come from medical training that made novices to believe that a miss would result in larger cost compared to false positive, or because they failed to update their prior belief about the signal prevalence in the task, both would suggest that novices and experts may have different beliefs in placing the optimal strategy in the hip fracture diagnosis. Our work can contribute to medical education training as well as other applied clinical diagnosis that aims to mitigate the prevalence effect.

Oral presentation assessment and image reading behaviour on brain computed tomography reading in novice clinical learners: an eye-tracking study

BACKGROUND

To study whether oral presentation (OP) assessment could reflect the novice learners' interpretation skills and reading behaviour on brain computed tomography (CT) reading.

METHODS

Eighty fifth-year medical students were recruited, received a 2-hour interactive workshop on how to read brain CT, and were assigned to read two brain CT images before and after instruction. We evaluated their image reading behaviour in terms of overall OP post-test rating, the lesion identification, and competency in systematic image reading after instruction. Students' reading behaviour in searching for the target lesions were recorded by the eye-tracking technique and were used to validate the accuracy of lesion reports. Statistical analyses, including lag sequential analysis (LSA), linear mixed models, and transition entropy (TE) were conducted to reveal temporal relations and spatial complexity of systematic image reading from the eye movement perspective.

RESULTS

The overall OP ratings [pre-test vs. post-test: 0 vs. 1 in case 1, 0 vs. 1 in case 2, p < 0.001] improved after instruction. Both the scores of systematic OP ratings [0 vs.1 in both cases, p < 0.001] and eye-tracking studies (Case 1: 3.42 ± 0.62 and 3.67 ± 0.37 in TE, p = 0.001; Case 2: 3.42 ± 0.76 and 3.75 ± 0.37 in TE, p = 0.002) showed that the image reading behaviour changed before and after instruction. The results of linear mixed models suggested a significant interaction between instruction and area of interests for case 1 (p < 0.001) and case 2 (p = 0.004). Visual attention to the target lesions in the case 1 assessed by dwell time were 506.50 ± 509.06 and 374.38 ± 464.68 milliseconds before and after instruction (p = 0.02). However, the dwell times in the case 2, the fixation counts and the frequencies of accurate lesion diagnoses in both cases did not change after instruction.

CONCLUSION

Our results showed OP performance may change concurrently with the medical students' reading behaviour on brain CT after a structured instruction.

Controllable Medical Image Generation via GAN

Medical image data is critically important for a range of disciplines, including medical image perception research, clinician training programs, and computer vision algorithms, among many other applications. Authentic medical image data, unfortunately, is relatively scarce for many of these uses. Because of this, researchers often collect their own data in nearby hospitals, which limits the generalizabilty of the data and findings. Moreover, even when larger datasets become available, they are of limited use because of the necessary data processing procedures such as de-identification, labeling, and categorizing, which requires significant time and effort. Thus, in some applications, including behavioral experiments on medical image perception, researchers have used naive artificial medical images (e.g., shapes or textures that are not realistic). These artificial medical images are easy to generate and manipulate, but the lack of authenticity inevitably raises questions about the applicability of the research to clinical practice. Recently, with the great progress in Generative Adversarial Networks (GAN), authentic images can be generated with high quality. In this paper, we propose to use GAN to generate authentic medical images for medical imaging studies. We also adopt a controllable method to manipulate the generated image attributes such that these images can satisfy any arbitrary experimenter goals, tasks, or stimulus settings. We have tested the proposed method on various medical image modalities, including mammogram, MRI, CT, and skin cancer images. The generated authentic medical images verify the success of the proposed method. The model and generated images could be employed in any medical image perception research.

Diagnostic Accuracy of Point-of-Care Fluorescence Imaging for the Detection of Bacterial Burden in Wounds: Results from the 350-Patient Fluorescence Imaging Assessment and Guidance Trial

Objective: High bacterial load contributes to chronicity of wounds and is diagnosed based on assessment of clinical signs and symptoms (CSS) of infection, but these characteristics are poor predictors of bacterial burden. Point-of-care fluorescence imaging (FL) MolecuLight i:X can improve identification of wounds with high bacterial burden (>10⁴ colony-forming unit [CFU]/g). FL detects bacteria, whether planktonic or in biofilm, but does not distinguish between the two. In this study, diagnostic accuracy of FL was compared to CSS during routine wound assessment. Postassessment, clinicians were surveyed to assess impact of FL on treatment plan. Approach: A prospective multicenter controlled study was conducted by 20 study clinicians from 14 outpatient advanced wound care centers across the United States. Wounds underwent assessment for CSS followed by FL. Biopsies were collected to confirm total bacterial load. Three hundred fifty patients completed the study (138 diabetic foot ulcers, 106 venous leg ulcers, 60 surgical sites, 22 pressure ulcers, and 24 others). Results: Around 287/350 wounds (82%) had bacterial loads >10⁴ CFU/g, and CSS missed detection of 85% of these wounds. FL significantly increased detection of bacteria (>10⁴ CFU/g) by fourfold, and this was consistent across wound types (p < 0.001). Specificity of CSS+FL remained comparably high to CSS (p = 1.0). FL information modified treatment plans (69% of wounds), influenced wound bed preparation (85%), and improved overall patient care (90%) as reported by study clinicians. Innovation: This novel noncontact, handheld FL device provides immediate, objective information on presence, location, and load of bacteria at point of care. Conclusion: Use of FL facilitates adherence to clinical guidelines recommending prompt detection and removal of bacterial burden to reduce wound infection and facilitate healing.

Dual-target hazard perception: Could identifying one hazard hinder a driver's capacity to find a second?

Low-level cognitive processes like visual search are crucial for hazard detection. In dual-target searches, subsequent search misses (SSMs) are known to occur when the identification of one target impedes detection of another that is concurrently presented. Despite the high likelihood of concurrent hazards in busy driving environments, SSMs have not been empirically investigated in driving. In three studies, participants were asked to identify safety-related target(s) in simulated traffic scenes that contained zero, one, or two target(s) of low or high perceptual saliency. These targets were defined as objects or events that would have prevented safe travel in the direction indicated by an arrow preceding the traffic scene. Findings from the pilot study (n = 20) and Experiment 1 (n = 29) demonstrated that detecting one target hindered drivers' abilities to find a second from the same scene. In Experiment 2 (n = 30), explicit instructions regarding the level of risk were manipulated. It was found that search times were affected by the instructions, though SSMs persisted. Implications of SSMs in understanding the causes of some crashes are discussed, as well as future directions to improve ecological and criterion validity and to explore the roles of expertise and cognitive capabilities in multi-hazard detection.

Visual search in breast imaging

Breast cancer is the most common cancer among females worldwide and large volumes of breast images are produced and interpreted annually. As long as radiologists interpret these images, the diagnostic accuracy will be limited by human factors and both false-positive and false-negative errors might occur. By understanding visual search in breast images, we may be able to identify causes of diagnostic errors, find ways to reduce them, and also provide a better education to radiology residents. Many visual search studies in breast radiology have been devoted to mammography. These studies showed that 70% of missed lesions on mammograms attract radiologists' visual attention and that a plethora of different reasons, such as satisfaction of search, incorrect background sampling, and incorrect first impression can cause diagnostic errors in the interpretation of mammograms. Recently, highly accurate tools, which rely on both eye-tracking data and the content of the mammogram, have been proposed to provide feedback to the radiologists. Improving these tools and determining the optimal pathway to integrate them in the radiology workflow could be a possible line of future research. Moreover, in the past few years deep learning has led to improving diagnostic accuracy of computerized diagnostic tools and visual search studies will be required to understand how radiologists interact with the prompts from these tools, and to identify the best way to utilize them. Visual search in other breast imaging modalities, such as breast ultrasound and digital breast tomosynthesis, have so far received less attention, probably due to associated complexities of eye-tracking monitoring and analysing the data. For example, in digital breast tomosynthesis, scrolling through the image results in longer trials, adds a new factor to the study's complexity and makes calculation of gaze parameters more difficult. However, considering the wide utilization of three-dimensional imaging modalities, more visual search studies involving reading stack-view examinations are required in the future. To conclude, in the past few decades visual search studies provided extensive understanding about underlying reasons for diagnostic errors in breast radiology and characterized differences between experts' and novices' visual search patterns. Further visual search studies are required to investigate radiologists' interaction with relatively newer imaging modalities and artificial intelligence tools.

Getting satisfied with "satisfaction of search": How to measure errors during multiple-target visual search

Visual search studies are common in cognitive psychology, and the results generally focus upon accuracy, response times, or both. Most research has focused upon search scenarios where no more than 1 target will be present for any single trial. However, if multiple targets can be present on a single trial, it introduces an additional source of error because the found target can interfere with subsequent search performance. These errors have been studied thoroughly in radiology for decades, although their emphasis in cognitive psychology studies has been more recent. One particular issue with multiple-target search is that these subsequent search errors (i.e., specific errors which occur following a found target) are measured differently by different studies. There is currently no guidance as to which measurement method is best or what impact different measurement methods could have upon various results and conclusions. The current investigation provides two efforts to address these issues. First, the existing literature is reviewed to clarify the appropriate scenarios where subsequent search errors could be observed. Second, several different measurement methods are used with several existing datasets to contrast and compare how each method would have affected the results and conclusions of those studies. The evidence is then used to provide appropriate guidelines for measuring multiple-target search errors in future studies.

Temporal Characteristics of Radiologists' and Novices' Lesion Detection in Viewing Medical Images Presented Rapidly and Sequentially

Although viewing multiple stacks of medical images presented on a display is a relatively new but useful medical task, little is known about this task. Particularly, it is unclear how radiologists search for lesions in this type of image reading. When viewing cluttered and dynamic displays, continuous motion itself does not capture attention. Thus, it is effective for the target detection that observers' attention is captured by the onset signal of a suddenly appearing target among the continuously moving distractors (i.e., a passive viewing strategy). This can be applied to stack viewing tasks, because lesions often show up as transient signals in medical images which are sequentially presented simulating a dynamic and smoothly transforming image progression of organs. However, it is unclear whether observers can detect a target when the target appears at the beginning of a sequential presentation where the global apparent motion onset signal (i.e., signal of the initiation of the apparent motion by sequential presentation) occurs. We investigated the ability of radiologists to detect lesions during such tasks by comparing the performances of radiologists and novices. Results show that overall performance of radiologists is better than novices. Furthermore, the temporal locations of lesions in CT image sequences, i.e., when a lesion appears in an image sequence, does not affect the performance of radiologists, whereas it does affect the performance of novices. Results indicate that novices have greater difficulty in detecting a lesion appearing early than late in the image sequence. We suggest that radiologists have other mechanisms to detect lesions in medical images with little attention which novices do not have. This ability is critically important when viewing rapid sequential presentations of multiple CT images, such as stack viewing tasks.

Experimental system for measurement of radiologists' performance by visual search task

PURPOSE

Detective performance of radiologists for "obvious" targets should be evaluated by visual search task instead of ROC analysis, but visual task have not been applied to radiology studies. The aim of this study was to set up an environment that allows visual search task in radiology, to evaluate its feasibility, and to preliminarily investigate the effect of career on the performance.

MATERIALS AND METHODS

In a darkroom, ten radiologists were asked to answer the type of lesion by pressing buttons, when images without lesions, with bulla, ground-glass nodule, and solid nodule were randomly presented on a display. Differences in accuracy and reaction times depending on board certification were investigated.

RESULTS

The visual search task was successfully and feasibly performed. Radiologists were found to have high sensitivity, specificity, positive predictive values and negative predictive values in non-board and board groups. Reaction time was under 1 second for all target types in both groups. Board radiologists were significantly faster in answering for bulla, but there were no significant differences for other targets and values.

CONCLUSION

We developed an experimental system that allows visual search experiment in radiology. Reaction time for detection of bulla was shortened with experience.