Visual-search ideal observers for modeling reader variability

项目概要/摘要 该项目的目标是开发新的方法来预测人类的决策与诊断 图像.预期的项目成果包括对放射科医师变异性来源的新见解， 先进的工具，以加速临床研究中的成像试验。这些试验与专家读者和 已知事实的病例是用于评价成像技术的公认但繁琐的金标准。 许多临床研究人员无法获得必要的试验资源。虚拟试验与表面- 已经提出了罗盖特模型观测器，但重要的限制，包括主要相关 估计和持续依赖人类数据进行训练的模型，阻止了它们的广泛应用， 是的。对人类输入依赖最小的定量模型将大大改善临床 先进的成像技术。我们开发这种“低资源”模型的方法将 探索目标检测和估计任务中的读者可变性。理想观测器（Ideal Observers，IO） gist-processing和极值理论将是出发点。这些IO最适合于去- 最大化提取的特征值集的决策过程，这是任务的共同前提 包括视觉搜索其结果将是自适应观测器模型， 与现有模型相比，这些新模型将测试读者的可变性 可以归因于候选池和认知阈值机制，这些机制定义了图像结构， 真正的兴趣。将开发诊断性视觉搜索任务的分析价值。我们 将测试模型在放射学模态、任务、成像模型（例如，同时， 患者/患者数据）和读者类别（外行/临床医生），所有这些都与研究人员相关。这些任务将 包括位置已知、定位和联合检测-估计格式。联合任务迫使 更精确的信息提取比目标检测单独;我们假设，检测性能- mance与估计技能相关，后者有助于解决结构问题。我们将利用我们的 研究发现，设计多读者虚拟试验方案，以提高统计学的严谨性。增强随机 2D和3D图像研究的目标建模将是支持目标。IO还将允许 检查个体读者的非线性行为。项目研究涉及减少剂量 以及用于X射线和核医学模态的重建方法，但是这些方法可以应用于 更普遍地说。通过加速先进成像技术的临床应用，我们的模型 观察员将对临床操作和病人护理产生直接和广泛的影响。

Project Summary/Abstract The goal of this project is to develop novel methods for predicting human decisions with diagnostic images. Expected project outcomes include new insights into sources of radiologist variability and advanced tools to accelerate imaging trials in clinical research. Such trials with expert readers and known-truth cases are an accepted but burdensome gold standard for evaluating imaging technology. The necessary trial resources are not available to many clinical researchers. Virtual trials with sur- rogate model observers have been proposed, but important limitations, including primarily correlative estimates and persistent model reliance on human data for training, prevent their widespread adop- tion. Quantitative models with minimal dependence on human input will substantially improve clinical access to advanced imaging technology. Our approach to develop such “low-resource” models will explore reader variability in target detection and estimation tasks. Ideal observers (IOs) derived from gist-processing and extreme-value theories will be the starting point. These IOs are optimal for de- cision processes that maximize over sets of extracted feature values, a common premise for tasks involving visual search. The result will be adaptive observer models that produce tighter bounds on human performance compared to existing models. These new models will test if reader variability can be attributed to candidate pooling and cognitive threshold mechanisms that define image struc- ture of interest. Analytic figures of merit for diagnostic visual-search tasks will be developed. We will test model generalizability across radiological modalities, tasks, imaging models (e.g., simula- tion/patient data), and reader classes (lay/clinician), all of relevance for researchers. The tasks will include location-known, localization, and joint detection-estimation formats. The joint task compels more precise information extraction than target detection alone; we hypothesize that detection perfor- mance correlates with estimation skill, with the latter helping to resolve structure. We shall leverage our findings to devise multireader virtual trial protocols for improved statistical rigor. Enhanced stochastic target modeling for studies with 2D and 3D images will be supporting aims. The IO will also allow examination of nonlinear behaviors for individual readers. The project studies relate to dose reduction and reconstruction methods for x-ray and nuclear medicine modalities, but the methods can apply more generally. By accelerating the clinical adoption of advanced imaging technology, our model observers will have a direct and widespread impact on clinical operations and patient care.