Novel Perceptual and Oculomotor Heuristics for Enhancing Radiologic Performance

用于增强放射学性能的新颖感知和动眼神经启发法

• 批准号: 10220201
• 负责人: Stephen Louis Macknik
• 金额: $ 64.61万
• 依托单位: SUNY DOWNSTATE MEDICAL CENTER
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-06-01 至 2026-05-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10220201
• 关键词: Assessment tool; Benchmarking; Biological Markers; Brain; COVID-19; Cancer Detection; Case Study; Cause of Death; Cessation of life; Characteristics; Clinical/Radiologic; Collection; Conscious; Data; Data Analyses; Databases; Detection; Diagnostic; Dimensions; Disease; Elements; Ensure; Entropy; Exposure to; Eye; Eye Movements; Fatigue; Film; Foundations; Frequencies; Human; Image; Incentives; Individual; Instruction; Knowledge; Lead; Learning; Location; Measurement; Measures; Medical Errors; Medical Imaging; Medicine; Modeling; Nature; North America; Outcome; Participant; Pathway interactions; Perception; Perceptual learning; Performance; Peripheral; Positioning Attribute; Radiologic Finding; Radiology Specialty; Reading; Research; Residencies; Resolution; Rest; Retina; Scanning; Societies; Speed; Sports; Stress; System; Testing; Texture; Thoracic Radiography; Time; Training; Unconscious State; Vision; Visual; Workload; X-Ray Computed Tomography; base; cancer diagnosis; cancer imaging; cohort; deep learning; design; experience; fitness; heuristics; human error; human model; improved; innovation; learning network; lung imaging; meetings; novel; oculomotor; oculomotor behavior; pandemic disease; patient safety; programs; radiological imaging; radiologist; sample fixation; shift work; skills; statistics; tool; tool development

PROGRAM SUMMARY Radiological imaging is often the first step of the diagnostic pathway for many devastating diseases; thus, an erroneous assessment of “normal” can lead to death. Whereas a grayscale object in an image can be described by its first-order image statistics—such as contrast, spatial frequency, position, entropy, and orientation—none of these dimensions, by itself, indicates abnormal vs normal radiological findings. We are a highly diverse team proposing an empirical approach to determine the mixtures of the first-order statistics—the “visual textures”— that radiology experts explicitly and implicitly use to identify the locations of potential abnormalities in medical images. Our innovative approach does not rely on assumptions about which textures may or may not be im-portant to abnormality detection. Instead, we will track the oculomotor behavior of expert radiologists to deter-mine their conscious and unconscious targeting choices, and thus ascertain which textures are empirically in-formative. The ability of expert radiologists to rapidly find abnormalities suggests that they may be able to first identify them in their retinal periphery. Peripheral visual analysis skills are therefore potentially critical to radio-logic performance, despite being understudied. We will measure these skills and leverage the results to develop perceptual learning heuristics to improve peripheral abnormality texture detection. By comparing novices to ex-perts we will determine whether the first are inexpert due to a lack of sensitivity to diagnostically relevant textures (texture informativeness), or to a lack of knowledge about which textures are abnormal, or to a combined lack of both sensitivity and knowledge. Radiology also requires the acquisition of oculomotor skills through practice and optimization. Radiologic expertise thus changes the oculomotor system in predictable and detectable ways, in much the same way that an athlete’s body and brain change as a function of expertise acquisition in their sport. We will therefore analyze both the consistency between experts’ fixation choices in medical images, and the eye movement performance characteristics of experts vs novice radiologists, to create an objective oculomotor bi-omarker of radiological expertise. The differences between novices and experts will train a deep learning (DL) system, which will have human visual and oculomotor performance characteristics. Training the DL with the abnormalities identified by a panel of expert radiologists will allow it to pinpoint the possible solutions in the manner of a simulated human radiologist performing at peak accuracy, precision, and speed. The resulting rank-ordered list of possible optimal and suboptimal image-reading strategies will serve as a benchmarking tool to quantify the performance of actual clinicians and residents who read the same images, rested vs fatigued. Meas-uring the effects of both training and fatigue on radiology expertise will be a major interdisciplinary cross-cutting advance in performance assessment. Our proposal to quantify fatigue in terms of erosion of expertise represents a transformational advance towards objective fitness-for-duty and expertise measures in medicine and beyond.

计划概要 放射成像通常是许多破坏性疾病诊断途径的第一步；因此，对“正常”的错误评估可能会导致死亡。尽管图像中的灰度对象可以通过其一阶图像统计数据（例如对比度、空间频率、位置、熵和方向）来描述，但这些维度本身都不能表明异常与正常的放射学发现。我们是一个高度多元化的团队，提出了一种经验方法来确定一阶统计数据的混合——“视觉纹理”——放射学专家明确或隐含地使用它来识别医学图像中潜在异常的位置。我们的创新方法不依赖于关于哪些纹理对于异常检测可能重要或不重要的假设。相反，我们将跟踪放射科专家的动眼行为，以确定他们有意识和无意识的目标选择，从而确定哪些纹理具有经验信息。放射科医生快速发现异常的能力表明他们可能能够首先在视网膜周边识别异常。因此，尽管尚未得到充分研究，外围视觉分析技能对于放射逻辑性能可能至关重要。我们将测量这些技能并利用结果来开发感知学习启发法，以改进外围异常纹理检测。通过将新手与专家进行比较，我们将确定新手是否不专业是因为缺乏对诊断相关纹理（纹理信息性）的敏感性，或者缺乏关于哪些纹理异常的知识，或者缺乏敏感性和知识。放射学还需要通过练习和优化来获得动眼技能。因此，放射学专业知识以可预测和可检测的方式改变动眼神经系统，就像运动员的身体和大脑随着运动专业知识的获取而变化一样。因此，我们将分析专家在医学图像中注视选择之间的一致性，以及专家与新手放射科医生的眼动表现特征，以创建放射学专业知识的客观动眼生物标记。新手和专家之间的差异将训练一个深度学习（DL）系统，该系统将具有人类视觉和动眼神经性能特征。使用放射科医生专家小组识别的异常情况来训练深度学习，将使其能够以模拟人类放射科医生的方式以最高准确度、精密度和速度执行任务，从而查明可能的解决方案。由此产生的可能的最佳和次优图像读取策略的排序列表将作为基准工具来量化读取相同图像的实际临床医生和住院医师的表现，无论是休息还是疲劳。衡量培训和疲劳对放射学专业知识的影响将是绩效评估中跨学科交叉的重大进展。我们提出的根据专业知识侵蚀来量化疲劳的建议代表了医学及其他领域朝着客观的岗位适应性和专业知识衡量标准迈出的变革性进步。