New Class of Numerical Observers for Nuclear Cardiology

描述（由申请人提供）：单光子发射计算机断层扫描（SPECT）成像已成为现代心脏病学的标准组成部分。在SPECT研究中（和一般的医学成像一样），人们普遍认为成像硬件和算法的进步应该以所谓的基于任务的评估标准为指导，即反映成像技术如何影响临床决策的措施。一般来说，人类观察者研究是衡量基于任务的标准的黄金标准；然而，这种研究的费用和复杂性阻碍了它们的常规使用。因此，数值观测者——模拟人类观测者表现的算法——现在被广泛用作人类观测者的替代品。在SPECT中，一种特殊的数值观测器，被称为信道化霍特林观测器（CHO），已经成为该领域的主导。CHO是一种检测算法，用于近似人类观察者在检测病变方面的表现；在心脏SPECT的情况下，感兴趣的病变是灌注缺陷。成像系统或算法可以通过CHO根据生成的图像准确检测缺陷的能力来判断。SPECT研究人员现在严重依赖（有时完全依赖）像CHO这样的数值观察者，不仅要验证他们的最终结果，而且要作为指导硬件或算法优化的价值指标。由于它所发挥的核心作用，CHO及其扩展本身已成为一个重要的研究课题。在提议的项目中，我们的目标是创建一套数值观察者，这将比CHO更广泛地阐明临床任务，我们将追求一种我们假设比CHO更准确的方法。因此，拟议的研究具有重要意义，因为它将产生一种评估方法，这种方法可能被研究界广泛使用，为成像硬件和软件的发展奠定基础。我们将使用提议的NO方法开发一个用于图像质量评估的软件包，并将其免费分发给研究社区。作为研究的副产品，拟议的项目还将对主要图像重建算法进行彻底的基于任务的评估，并将回答以下问题：哪种类型的数据（从简单的幻影到临床数据）是数值观察者按预期执行的充分基础。这项研究也将为潜在的心脏病学计算机辅助诊断系统奠定基础。研究的具体目的如下：1)创建一套完整的成像数据和人类观察者评分；2)基于我们新颖的基于学习的方法，以及更传统的统计决策理论原理，开发一套数值观察者；3)比较和评价数值观测者；4)开发和传播研究项目的副产品。

DESCRIPTION (provided by applicant): Single-photon emission computed tomography (SPECT) imaging has become a standard component of modern cardiology. In SPECT research (as in medical imaging generally), it has become widely accepted that advances in imaging hardware and algorithms should be guided by so-called task-based evaluation criteria, i.e., measures that reflect how the imaging technique will impact clinical decision making. In general, a human observer study is the gold standard for measuring task-based criteria; however, the expense and complexity of such studies precludes their routine use. Therefore, numerical observers-algorithms that emulate human observer performance-are now widely used as surrogates for human observers. In SPECT, one particular numerical observer, known as the channelized Hotelling observer (CHO), has come to dominate the field. The CHO is a detection algorithm that is used to approximate the human observer's performance in detecting lesions; in the case of cardiac SPECT, the lesions of interest are perfusion defects. An imaging system or algorithm can be judged by the ability of the CHO to accurately detect defects based on the images produced. SPECT researchers now rely heavily (and sometimes exclusively) on numerical observers such as the CHO, not only to validate their final results, but also as a figure of merit that guides optimization of hardware or algorithms. Because of the central role it has come to play, the CHO and its extensions have become a major research topic in their own right. In the proposed project, our goal will be to create a suite of numerical observers that will shed light on a much wider set of clinical tasks than the CHO, and we will pursue an approach that we hypothesize will be more accurate than the CHO. Therefore, the proposed research is significant because it will yield an evaluation methodology that could potentially be used very widely by the research community, underpinning the development of imaging hardware and software. We will develop a software package for image quality assessment using the proposed NO approach and distribute it freely to the research community. As a by-product of the research, the proposed project will also yield a thorough task-based evaluation of major image reconstruction algorithms, and will answer the question of which sorts of data (on the spectrum from simple phantoms to clinical data) are a sufficient foundation for numerical observers to perform as desired. The research will also yield the basis for a potential computer aided diagnostic system for cardiology. The specific aims of the research will be as follows: 1) Create a comprehensive set of imaging data and human observer scores; 2) Develop a suite of numerical observers based on our novel learning-based approach, as well as more-conventional statistical decision theory principles; 3) Compare and evaluate the numerical observers; and 4) Develop and disseminate by-products of the research program.