Isolating and Mitigating Sequentially Dependent Perceptual Errors in Clinical Visual Search

Vision Clinical Trial

Official title: Isolating and Mitigating Sequentially Dependent Perceptual Errors in Clinical Visual Search

Status Recruiting

Clinical Trial Summary

When looking at an x-ray, radiologists are typically asked to localize a tumor (if present), and to classify it, judging its size, class, position and so on. Importantly, during this task, radiologists examine on a daily basis hundreds and hundreds of x-rays, seeing several images one after the other. A main underlying assumption of this task is that radiologists' percepts and decisions on a current X-ray are completely independent of prior events. Recent results showed that this is not true: perception and decisions are strongly biased by past visual experience. Although serial dependencies were proposed to be a purposeful mechanism to achieve perceptual stability of otherwise noisy visual input, serial dependencies play a crucial and deleterious role in the everyday task performed by radiologists. For example, an x-ray containing a tumor can be classified as benign depending on the content of the previously seen x-ray. Given the importance and the impact of serial dependencies in clinical tasks, in this proposal, the investigators plan to (1) establish, (2) identify and (3) mitigate the conditions under which serial effects determine the participants' percepts and decisions in tumor search tasks. In Aim 1, the investigators will establish the presence of serial effects in four different clinically relevant domains: tumor detection, tumor classification, tumor position and recognition speed. In Aim 2, the investigators plan to identify the specific boundary conditions under which visual serial dependence impacts tumor search in radiology. In Aim 3, once the investigators fully understand these boundary conditions in Aim 2, they will propose a series of task and stimulus manipulations to control and mitigate the deleterious effects of visual serial dependence on tumor search. As a result of these manipulations, visual search performance should improve in measurable ways (detection, classification, position, speed). Aim 3 is particularly crucial because it will allow the investigators to propose new guidelines which will greatly improve tumor recognition in x-ray images, making this task even more effective and reliable. Taken together, the proposed studies in Aim 1, 2, and 3 will allow the investigators to establish, identify, and mitigate the deleterious effect of serial dependencies in radiological search tasks, which could have a significant impact on the health and well-being of patients everywhere.

Clinical Trial Description

SPECIFIC AIMS Radiological scans are critically important for the health and well-being of millions of patients [1]. Despite advances in machine learning and computer vision, reading and searching in radiological scans still depends on the visual system of human observers (radiologists), who receive continuous training to best detect tumor masses and anomalies. Radiologists can examine hundreds and hundreds of x-rays on a daily basis, seeing several images one after the other. A main underlying assumption of their work is that radiologists' percepts and decisions on a current x-ray are completely independent from prior viewings. However, participants recently demonstrated that the visual system has visual serial dependencies (VSDs) at many levels, from perception to decision making [2, 3]. These sequential dependencies mean that what was seen in the past influences (and captures) what is seen and reported at this moment. Theoretically, VSDs would be helpful in an autocorrelated natural world, but they are suboptimal in search tasks conducted in artificial situations where images may not always be related. Importantly, serial dependences in perceptual processing could thus produce significant errors during diagnostic searches through radiological scans. The investigators' central hypothesis is that VSD can have a disruptive effect in radiologic searches that impairs accurate detection and recognition of tumors or other structures. This hypothesis is supported by the investigators' robust pilot data, which show that VSD strongly biases object classification in naïve observers and expert radiologists. The rationale for the proposed research projects is that once it is known how serial dependence arises and how it impacts visual search, we can understand how to control for it. Hence, accuracy of tumor detection and diagnosis can significantly improve. The specific objectives of this proposal are to establish (Aim 1), identify (Aim 2) and mitigate (Aim 3) the impact of VSD on visual search tasks in a clinical setting. Aim 1: Establish the presence of sequential effects in perceptual decision making in clinically relevant tasks. The investigators' previous theoretical work on sequential dependencies, based on simplified laboratory experiments, predicts that there should also be sequential effects in clinically relevant tasks such as tumor search and recognition in images. The investigators will test tumor recognition in four different domains: tumor detection, tumor classification, tumor localization, and search times. Pilot data in both naïve observers and expert radiologists supports the investigators' hypothesis that there are sequential effects in clinically relevant perceptual judgments. Aim 2: Identify that the sequential effects in Aim 1 are, in fact, VSD. The investigators will test our prediction that the sequential effects in Aim 1 are, in fact, VSD, which is operationally defined by a set of diagnostic criteria: VSDs are spatially, temporally, and feature-similarity tuned. Aim 2 will therefore identify the tuning and boundary conditions of VSD in tumor search. This is crucial to develop strategies to mitigate the harmful consequences of VSD in Aim 3. Pilot data supports the feasibility of this aim. Aim 3: Improve visual search performance in radiologists by mitigating VSD. Aim 2 will define the boundary conditions under which VSD operates in tumor search. Aim 3 will develop task manipulations to mitigate the impact of VSD. As a result of these manipulations, visual search performance should improve in measurable ways in several domains (detection, classification, localization, and search times). Aim 3 is particularly crucial because it will allow the investigators to propose new guidelines that will improve tumor recognition in radiological scans, making radiologists' judgments more effective and reliable. More globally, taken together, the proposed studies in Aims 1, 2 and 3 will allow the investigators to establish and work toward mitigating the deleterious role of VSD in a critically important clinical task. ;

Related Conditions & MeSH terms

• Vision

• NCT number: NCT04332783
• Study type: Interventional
• Source: University of California, Berkeley
• Contact: Valerie Ekko, BA
• Phone: 510-642-5797?
• Email: vekko@berkeley.edu
• Status: Recruiting
• Phase: N/A
• Start date: April 1, 2019
• Completion date: March 30, 2026