CAREER: Estimation Methods for Image Registration

生物医学图像分析中最核心的操作之一是图像配准：图像配准。过去几十年来的大量努力已经产生了复杂的配准方法，适用于从纳米级显微成像到器官级成像的各种应用。然而，关于变换模型的基本假设既没有受到实质性的挑战，也没有开发出量化可变形配准结果的不确定性的实用方法。例如，现有的方法不适合于许多对象间的配准或涉及病理学的配准，而病理学是许多成像研究的核心要求。这项研究的目标是设计先进的图像配准估计方法，从动力系统的角度出发，(1)改进图像变形的估计，(2)改进图像相似性度量以考虑图像变化(如组织生长或退化)，(3)联合建模特定于对象的变形和基于群体的变形，(4)考虑配准不确定性的估计，(5)提供适当的验证策略，以及(6)以非专家可以理解的直观方式呈现方法。更好的图像配准方法将改善图像分析结果，从而有助于临床研究和最终患者护理。应用领域包括创伤性脑损伤研究、放射治疗计划、肿瘤进展评估或骨关节炎的人群研究、正常脑发育、阿尔茨海默病？S或亨廷顿？S病，仅举几例。我们将专注于神经成像和肺部运动的捕获，以突出图像配准问题的鲜明特征。智力价值建议的研究将显著推动最新技术的发展，因为它将提供在图像配准中制定估计问题时使用领域知识的迫切需要的方法。假设通过添加这样的信息，将改善对空间变形的估计，从而改善严重依赖于图像配准的分析结果的质量。特别是，开发的方法将解决以下高度相关的问题：如何处理受病理影响的图像、如何在捕获数据趋势的同时执行基于总体的分析，以及如何评估配准中的估计不确定性。这项研究将对当前的成像研究产生直接影响，并将为未来的应用奠定基础。虽然研究将专注于神经成像和肺运动分析的应用，但所开发的方法将普遍适用于任何需要估计从生物学到动画、目标跟踪和神经成像的空间对应关系的图像分析问题。特别是，由于这些方法将明确处理对象水平(包括病理或老龄化引起的变化)和人群水平的变化，它们将导致对疾病进展的更好估计，将能够提供特定于患者的信息，并有望改善基于人群的成像研究(例如临床药物测试)的结果。为了允许其他人创建定制的图像分析解决方案，所有开发的方法都将以开源的形式提供给社区。1

One of the most central operations in biomedical image analysis is image alignment: image registration. Substantial efforts over the last decades have produced sophisticated registration approaches for applications ranging from microscopy imaging at the nanometer scale to organ-scale imaging. However, neither the underlying assumptions regarding transformation models have been substantially challenged nor have practical methods to quantify uncertainties for deformable registration results been developed. For example, the existing methods are not appropriately designed for many inter-subject registrations or for registrations involving pathologies which are core requirements for many imaging studies. The goal of this research program is to devise advanced estimation methods for image registration, motivated from a dynamical systems point of view which (1) improve estimation of image deformations, (2) improve image similarity measures to account for image changes (such as tissue growth or degeneration), (3) model subject-specific deformations and population-based deformations jointly, (4) allow for the estimationof registration uncertainties, (5) provide appropriate validation strategies, and (6) to present the methods in an intuitive way understandable for the non-expert. Better image registration methods will improve image analysis results and hence will help clinical studies and ultimately patient care. Application areas include the study of traumatic brain injury, radiation treatment planning, assessment of tumor progression or population studies of osteoarthritis, normal brain development, Alzheimer?s or Huntington?s disease, to name but a few. We will focus on neuroimaging and the capturing of lung motions to highlight distinct characteristics of image registration problems.Intellectual Merit The proposed research will significantly advance the state-of-the art, because it will provide much needed ways of using domain knowledge when formulating estimation problems in image registration. The hypothesis is that by adding such information, estimations of space deformations will be improved, hence improving the quality of the analysis results which critically depend on image registration. In particular, the developed methods will address the highly relevant problems of how to deal with images subject to a pathology, how to perform population-based analysis while capturing data-trends, and how to assess estimation uncertainty in registration. The research will have immediate impact on current imaging studies and will form the basis for future applications.Broader Impact While the research will focus on applications in neuroimaging and lung motion analysis, the developed methods will be generally applicable to any image analysis problem requiring an estimate of spatial correspondence ranging from biology, to animation, to object tracking, and to neuroimaging. In particular, since the methods will explicitly address changes on the subject level (including changes caused by pathology or aging) and the population level, they will lead to better estimates of disease progression, will be able to provide patient-specific information, and are expected to improve results for population-based imaging studies (for example for clinical drug testing). To allow others to create customized image analysis solutions all developed methods will be made available to the community in open-source form.1