Three-Class ROC Analysis for Task-Based Medical Image Quality Assessment

描述（由申请人提供）： 传统的ROC分析已被广泛接受为二元诊断诊断技术评估的标准。然而，许多医学诊断涉及多种诊断选择。例如，使用乳房X线照相术进行乳腺癌诊断，其中诊断类别为正常、良性或恶性肿瘤，以及使用心肌灌注SPECT（MRS）进行心脏病诊断，其中类别为正常、可逆或固定缺陷。为了评估多类诊断技术，需要多类ROC，但自20世纪50年代引入二元ROC分析以来，多类ROC一直是一个未解决的问题。由于MPS优化提出的实际挑战，候选人提出了一种三类ROC分析方法，该方法扩展并统一了三类范式中二进制ROC的决策理论，线性判别分析和概率基础。她对三类ROC分析进行了五项初步研究：（1）推导其决策模型[He，梅斯，et.al IEEE Trans Med Imag（TMI）vol.25（5），2006];（2）研究其决策理论基础[He and Frey，TMI，vol.25（8），2006];（3）探索其线性判别分析（LDA）基础[He and Frey，TMI，in press，2006];（4）建立其概率基础;（5）与传统的三类LDA进行比较，揭示了传统三类LDA的局限性。该候选人于2005年12月获得生物医学工程博士学位，并接受了医学成像方面的强化培训。她增加了她在医学图像质量评估的兴趣，在三个类ROC分析的发展，她的知识，在这项研究中使用的统计和决策理论的原则是自学成才。进一步探索三类ROC分析开辟的新领域需要系统地了解决策理论，统计学习和贝叶斯建模等统计原理，因此，她要求为期两年的指导阶段，重点是正式的生物统计学培训。培训阶段将大大提高候选人作为跨学科研究者的职业发展，并有助于她独立研究，以实现以下具体目标：1）建立三级ROC分析的理论基础;（2）开发三级ROC分析的通用统计方法;（3）将三级方法应用于基于任务的医学图像质量评估。拟议的工作有两方面的意义。首先，它为一个开放的理论问题提供了严格的解决方案，并将在ROC分析和医疗决策方面开辟新的理论研究领域。第二，它使应用程序的任务为基础的评估技术多类诊断。这些技术有可能从根本上改善目前用于疾病检测和表征的成像技术，从而提高医生在疾病诊断方面的表现，这将广泛造福于公共卫生。

DESCRIPTION (provided by applicant): Conventional ROC analysis has been widely accepted as a standard in the assessment of diagnostic techniques for binary diagnoses. Many medical diagnoses, however, involve multiple diagnostic alternatives. Examples are breast cancer diagnosis using mammography, where the diagnostic classes are normal, benign, or malignant tumor, and cardiac disease diagnosis using myocardial perfusion SPECT (MRS), where the classes are normal, reversible or fixed defect. To assess multi-class diagnostic techniques, multi-class ROC is required, but has remained an unsolved problem ever since the introduction of binary ROC analysis in the 1950s. Sparked by a practical challenge raised by MPS optimization, the candidate proposed a three- class ROC analysis method that extends and unifies the decision theoretic, linear discriminant analysis and probabilistic foundations of binary ROC in a three-class paradigm. She has conducted five preliminary studies on three-class ROC analysis: (1) deriving its decision model [He, Metz, et.al IEEE Trans Med Imag (TMI) vol. 25(5), 2006]; (2) investigating its decision theoretic foundation [He and Frey, TMI, vol. 25(8), 2006]; (3) exploring its linear discriminant analysis (LDA) foundation [He and Frey, TMI, in press, 2006]; (4) establishing its probabilistic foundation; and (5) comparing it with conventional three-class LDA and revealing the limitations of conventional three-class LDA. The candidate obtained a PhD in Biomedical Engineering in December 2005 and had intensive training on medical imaging. She increased her interest in medical image quality assessment during the development of three-class ROC analysis; her knowledge of the statistics and decision theory principals used in this research is self-taught. Further exploring new areas opened by three- class ROC analysis requires systematic understanding of the statistical principles in decision theory, statistical learning, and Bayesian modeling, etc. Thus, she requests a two-year mentored phase focusing on formal biostatistics training. The training phase will substantially enhance the candidate's career development as an interdisciplinary investigator and contribute to her independent research to accomplish the following specific aims: 1) to establish the theoretical foundations of three-class ROC analysis; (2) to develop general statistical methods for three-class ROC analysis; (3) to apply the three-class methodologies to task-based medical image quality assessment. The significance of the proposed work is two-fold. First, it provides a rigorous solution to an open theoretical problem and will open new areas of theoretical research in ROC analysis and medical decision making. Second, it enables applications of task-based assessment techniques for multi-class diagnosis. These techniques have the potential to fundamentally improve current imaging techniques for disease detection and characterization, and thus to enhance doctors' performance in disease diagnosis, which will broadly benefit public health.