Shape Exploration for Medical Applications --- From Representation, Correspondence, Deformation to Image Segmentation

项目摘要（Song Wang） EIA-0312861Song WhangUSC研究基金会本次拟开展的研究是开发一个完整实用的形状建模框架，并利用该框架实现更可靠的医学图像分割。可靠地提取、表示和结合一些先验形状信息可以大大降低图像噪声带来的分割误差，这是本研究的动机。为了有效地探索医学图像分割中的形状信息，将解决并整合三个重要问题：形状表示、形状对应和形状变形。形状表示法是用一组界标点紧凑地描述一个形状。形状对应是在一组形状样本之间建立基于地标的对应关系。形状变形是首先构造一个形状模板，该模板捕获特定对象类的具有代表性的形状特征，然后利用它从新图像中勾画出相同类别的对象。智力优势：本研究以独特的系统视角研究形状建模方法。以往的研究通常集中在形状建模中的一个特定问题上，而我们将在一个集成和统一的框架中研究形状表示、形状对应和形状变形这三个相互关联的关键问题，特别是针对医学图像分割。在这个统一的框架下，我们开发了新的方法来解决这三个问题中的最重要的问题。在形状表示中，我们将结合与分辨率无关的特征提取方法来更准确地估计关键形状参数，如位置，切线方向和曲率。这将在描述一组形状实例时提供一种准确、紧凑和灵活的表示格式。在形状对应中，我们将基于以下标准开发全局和局部对应方法来选择地标：(1)形状表示简单，(2)表示误差小，(3)形状间地标对应最优。将采用先进的网络流技术将这三种措施整合成一个新的统一公式。在形状变形中，我们将基于一组良好对应的训练样本开发有效的形状学习技术。与简单的高斯分布相比，该技术采用了更一般、更精确的概率分布模型。在此基础上，我们将进一步发展我们的形状变形方法，以解决算法鲁棒性和拓扑保持等重要问题。广泛影响：医学成像行业的快速发展为我们提供了大量的数据，如MRI， CT， PET，冷冻切片，显微镜等。迫切需要发展先进的信息技术，将这些医疗数据转化为有用的信息。其中，几何形状信息在临床诊断中应用广泛，尤其重要。该研究将极大地促进医学图像形状信息的探索，为医生和放射科医生提供新的强大的计算工具。从长远来看，这项研究将进一步加强目前的努力，弥合信息技术与医疗应用之间的差距。此外，所提出的研究可以应用于许多其他应用，如视频跟踪和生物识别，这些应用与从医学图像中进行形状探索具有相同的问题。这项研究也提供了许多教育贡献。我们将整合医学图像处理到一系列的图像处理和计算机视觉课程在南卡罗来纳大学（USC）。这项研究还将通过协助授予该大学的NSF工程教育桥梁项目，以及参与南卡罗来纳州州长科学与数学学院的暑期导师项目，积极为该州的K-12教育做出贡献。生物医学工程被选为南加州大学研究的三个主要重点领域之一。因此，我们的研究工作将直接促进整个大学的努力。特别是，我们的努力可以与心理学系的神经成像倡议和新的结肠癌研究中心相结合。

Project Summary (Song Wang) EIA-0312861Song WhangUSC Research Foundation The proposed research is to develop an integrated and practical framework for shape modeling and use it to achieve more reliable segmentation of medical images. This research is motivated by the fact that reliable extraction, representation, and incorporation of some prior shape information can greatly reduce the segmentation error resulting from image noise. Three important problems will be addressed and integrated to effectively explore shape information for medical image segmentation: shape representation, shape correspondence, and shape deformation. The shape representation is to compactly describe a shape in terms of a set of landmark points. The shape correspondence is to establish landmark-based correspondence among a set of shape samples. The shape deformation is to first construct a shape template that captures representative shape characteristics of a particular object class, and then use it to delineate an object of the same class from a new image. Intellectual Merits: This research studies the shape modeling method from a unique systematic perspective. While previous researches usually focus on a particular problem in shape modeling, we will investigate all three interrelated key problems, i.e., shape representation, shape correspondence, and shape deformation, in an integrated and unified framework, specifically for medical image segmentation. Under this unified framework, we develop novel methods to address most important issues in each of these three problems. In the shape representation, we will incorporate our resolution-independent feature extraction method to estimate key shape parameters such as position, tangent direction, and curvature more accurately. This will provide a representation format that is accurate, compact, and flexible in describing a group of shape instances. In the shape correspondence, we will develop global and local correspondence methods to choose landmarks based on the following criteria: (1) simple shape representation, (2) small representation error, and (3) optimal inter-shape landmark correspondence. An advanced network-flow technique will be used to integrate these three measures into a new unified formulation. In the shape deformation, we will develop effective shape-learning techniques based on a group of well-corresponded training samples. The techniques employ more general and accurate probability distribution models than simple Gaussian distributions. Based on the models, we will further develop our shape deformation method to deal with important issues like algorithm robustness and topology preservation. Broad Impacts: Fast growth of medical imaging industry provides us tremendous amount of data in many forms like MRI, CT, PET, cryosection, microscopic, etc. It is urgent to develop advanced information technology to convert those medical data into useful information. Among them, geometric shape information is of particular importance as they are widely used in clinical diagnoses. The proposed research will greatly facilitate the shape information exploration from medical images and provide physician and radiologists new and powerful computational tools. In the long term, this research will further intensify the current efforts to bridge the gap between information technology and medical applications. Furthermore, the proposed research can be applied to many other applications like video tracking and biometrics that shares the same problems as in shape exploration from medical images. This research also provides many educational contributions. We will integrate the medical image processing into a series of image processing and computer vision courses in the University of South Carolina (USC). This research will also actively contribute to the K-12 education in the state by assisting the NSF Bridges for Engineering Education Program awarded to the University and participating in the summer mentor program by the South Carolina Governor's School for Science and Mathematics. Biomedical engineering has been selected as one of three main focus areas for research at USC. Thus, our research effort will directly contribute to the university wide efforts. In particular, our effort can be integrated with the neuro-imaging initiative at the Psychology Department and the new Center for Colon Cancer Research.