Satisfaction of Search in Breast Cancer Detection

项目摘要/摘要 乳腺癌是美国和世界上女性癌症发病率最高的地方。尽管 技术进步-从胶片屏幕图像到全场数字乳房摄影(FFDM)，再到现在 数字乳房断层合成(DBT)-每年的错失率一直顽固地保持稳定，范围在 筛查时10-30%。技术本身并不能减少遗漏的错误；我们需要了解具体的 人类读者在解读这些图像时面临的挑战，以及它们导致的具体错误。 搜索满意度(SOS)是指在检测到病例中的第一个病变后， 同一病例中其他病变的漏诊率显著升高。这种特定类型的错误一直是 在研究过它的放射学领域，包括胸部，显示出30%的遗漏 射线照相术和计算机断层摄影术。然而，它从来没有在乳腺癌领域被研究过。 因此，迫切需要确定SOS如何导致乳腺癌筛查中的错误。 在本项目中，我们将确定SOS的发生率和潜在原因 FFDM和DBT。我们设计了一种新的方法，克服了以前方法的局限性，那就是 已针对在FFDM和DBT中使用进行了优化。以前研究SOS的方法包括摄影相加 人工病变的图像，这对于乳房成像是不可行的。相反，我们将构建一个数据库 为研究SOS而构建的自然发生的案例。这将涉及多个病变的收集 病例和对照的单病变病例，其中前者与后者在关键诊断上相匹配 尺寸，如病变类型、病变大小和乳房密度。在两个主要实验中(一个使用FFDM，另一个使用 一个患有DBT)，放射科医生将从实验组中读取病例，标记位置和诊断 良性和恶性病变。对双病变和单病变病例的信号检测分析将用于 估计SOS的比率。当参与者阅读每个案例时，眼睛的位置和瞳孔直径将被跟踪。 这些数据将使我们能够评估导致SOS的不同已知原因的流行率：(A)过早 终止，其中，与单一病变相比，在第一个病变检测之后进行搜索不那么全面 对照病例；(B)知觉集合，其中，在检测到第一个病变之后，参与者偏向于发现 具有相似感知特征的后续病变，导致检测到的敏感性降低 感知上不同的目标；以及(C)资源枯竭，在这种情况下，维护信息的需求 关于记忆中第一次检测到的损伤会减少可用的感知/认知资源，从而减少 后续搜索的效率。了解乳腺癌中SOS的发生率和潜在原因 检测将为计划的未来工作奠定基础，以制定培训方案和最佳做法，以 减少SOS错误的特定原因，从而减少乳腺癌筛查中的漏检率。

PROJECT SUMMARY/ABSTRACT Breast cancer has the highest incidence of cancer for women in the U.S. and across the world. Despite advances in technology—from film-screen images to Full Field Digital Mammography (FFDM) and now to Digital Breast Tomosynthesis (DBT)—the yearly miss rate has remained stubbornly stable, ranging between 10-30% at screening. Technology alone is not reducing errors of omission; we need to understand the specific challenges faced by the human readers interpreting the images, and the specific errors that they lead to. Satisfaction of Search (SOS) refers to the fact that, after having detected a first lesion in a case, the miss rate for additional lesions in the same case is substantially elevated. This specific type of error has been shown to account for 30% of misses in the domains of Radiology where it has been studied, including chest radiography and Computed Tomography. And yet, it has never been studied in the domain of breast cancer. Thus, there is a critical need to determine how SOS contributes to errors in breast cancer screening. In the present project we will determine the rates of occurrence and the underlying causes of SOS in FFDM and DBT. We have devised a novel method that overcomes limitations of previous methods and that is optimized for use in FFDM and DBT. Previous approaches to studying SOS involved the photographic addition of artificial lesions to images, which is not feasible for breast imaging. Instead, we will construct a database of naturally occurring cases that is structured for studying SOS. This will involve the collection of multiple-lesion cases and controlled single-lesions cases, where the former are matched with the latter on key diagnostic dimensions, such as lesion type, lesion size, and breast density. In two main experiments (one with FFDM and one with DBT), radiologists will read cases from the experimental set, marking the locations and diagnoses for benign and malignant lesions. Signal-detection analyses over dual- and single-lesion cases will be used to estimate the rate of SOS. Eye position and pupil diameter will be tracked as participants read each case. These data will allow us to assess the prevalence of different known causes of SOS: (a) premature termination, in which search following first lesion detection is less comprehensive compared with single-lesion control cases; (b) perceptual set, in which, after having detected a first lesion, participants are biased to find subsequent lesions with similar perceptual features, leading to reduced sensitivity in the detection of perceptually dissimilar targets; and (c) resource depletion, in which the demands of maintaining information about a first-detected lesion in memory reduce available perceptual/cognitive resources, thereby reducing the efficiency of subsequent search. Understanding the rates and underlying causes of SOS in breast cancer detection will lay the foundation for planned future work to develop training programs and best practices that mitigate the specific causes of SOS errors and thereby reduce miss rates in breast cancer screening.