PHYSICAL LIMITS OF QUANTITATIVE SPECT

这项拟议的研究将比较几种散射和衰减 基于它们对性能的影响的校正方法 使用SPECT数据的定量成像任务。要完成的成像任务 所考虑的包括对球体的活动和大小的估计 脑深部结构内活动的估计使用 从磁共振数据中提取的边界信息和大脑分类 使用皮质活动将图像转换为正常组和阿尔茨海默病组 大脑深层结构中的模式和活动。一种衡量 将获得每个定量成像任务的性能 正在评估校正方法。这些措施的范围将从 理论量，即关于方差的Cramer-Rao下界 参数估计，到实验测量的量，即， 从模拟或模型估计的活度值的标准误差 使用最大似然或贝叶斯技术的图像，以临床 量，即，在统计或神经的ROC曲线下 网络分类器在这些定量成像任务中的性能可以 被视为图像质量的衡量标准，因此是一个相关的 用于比较散射校正方法和衰减校正方法的基础。 拟议的研究将实现几个超越主要目标的目标。 校正方法比较的目标。首先，我们将调查 非均匀投影抽样对提高任务绩效的潜力 通过降低中心位置的噪音水平。第二，我们将解决 使用先验信息的内部估计，并将评估一个 利用磁共振法估算深部构造活动性的实用方法 边界信息。前两个领域的进展预计将 从而改善大脑深层结构中的活动量化。 最后，我们将应用神经网络，这才刚刚开始 用于放射学和核医学，以及阿尔茨海默氏症- 相关分类任务。

The proposed research will compare several scatter- and attenuation- correction methods on the basis of their effects on performance in quantitative imaging tasks using SPECT data. The imaging tasks to be considered include estimation of activity and size of spherical structures, estimation of activity within deep brain structures using boundary information derived from MR data, and classification of brain images into normal and Alzheimer's disease groups using cortical activity patterns and activity within deep brain structures. A measure of performance in each quantitative imaging task will be obtained for each correction method being evaluated. These measures will range from theoretical quantities, i.e., the Cramer-Rao lower bound on variance of parameter estimates, to experimentally measured quantities, i.e., standard error of activity values estimated from simulated or phantom images using maximum-likelihood or Bayesian techniques, to clinical quantities, i.e., are under the ROC curve of a statistical or neural network classifier Performance in these quantitative imaging tasks can be viewed as a measure of image quality and is, therefore, a relevant basis on which to compare scatter- and attenuation-correction methods. The proposed research will achieve several objectives beyond the major goal of comparison of correction methods. First, we will investigate the potential of nonuniform projection sampling to improve task performance by reducing noise levels at central locations. Second, we will address the use inn estimation of a priori information, and will evaluate a practical method of estimating deep structure activity using MR-derived boundary information. Progress in these first two areas is expected to lead to improved quantitation of activity in deep brain structures. Finally, we will apply neural networks, which are only beginning to be used in radiology and nuclear medicine, to the Alzheimer's disease- related classification tasks.