Learning to Register, Segment and Retrieve Medical Images

由放射科医生等专家对医学图像进行分析是诊断和治疗许多疾病(如癌症)的重要任务。软件程序经常被用来帮助专家。医学图像分析的三个主要软件操作是配准(对齐不同图像中的解剖结构)、分割(标记器官或肿瘤的边界)和检索(在大型数据库中查找图像)。医学图像的配准、分割和检索不仅是非常具有挑战性的问题(要达到准确和高效对所有这些问题来说都是非常困难的)，而且它们之间是相互关联的。例如，配准可以用于分割，分割可以用于图像检索。此外，在大多数情况下，当使用其中一种方法时，通常也需要其他两种方法。目前的技术忽略了这一关系，把配准、分割和检索分开来解决。考虑到这些技术之间的关系的统一解决方案预计将提供更准确的结果并节省资源。本研究的目的是调查多个问题。首先，我们将更密切地研究医学图像分析中这些任务之间的关系，其次，我们将开发一个统一的框架，可以通过图像的表征来学习这些任务。在这项研究中，将审查几种学习方案，如进化规则和强化权重，以设计一类新的算法来解决这三个任务。我们将尝试为所有三种手术定义一个统一的框架，该框架可以通过学习形态和解剖规范来自主调整自己，以适应手头的任务。为此，图像将以局部和全局特征以及编码解剖结构为特征。统一的医学图像配准、分割和检索框架将提高临床工作流程的效率，并通过纳入专家反馈来帮助建立高质量的结果。这将为临床医生(减少繁琐的任务)、医院(节省时间和资源)和患者(高质量护理)带来好处。

Analysis of medical images by an expert such as a radiologist is an important task for both diagnosis and treatment of many diseases such as cancer. Software programs are regularly used to assist the experts. Three main software operations of medical image analysis are registration (aligning anatomical structures in different images), segmentation (marking the boundaries of an organ or tumour) and retrieval (finding an image in a large database). Registering, segmenting and retrieving medical images are not only very challenging problems (achieving accuracy and efficiency is very difficult for all them), but also they are interrelated. For instance, registration can be used in segmentation, and segmentation can be used for image retrieval. Besides, in most cases, when one of these methods is used, the other two are generally required too. The current technologies neglect this relationship and spend a lot of effort to solve the problems of registration, segmentation and retrieval separately. A unified solution, that considers the relationship between these techniques, is expected to provide more accurate results and save resources. The purpose of this research is to investigate multiple questions. Firstly, we will examine the relationship of theses tasks in medical image analysis more closely, and secondly, we will develop a unified framework that can learn these tasks by characterization of images. In this research, several learning schemes, such as evolving rules and reinforced weights, will be examined to design a new class of algorithms for these three tasks. We will attempt to define a unified framework for all three operations that can autonomously adjust itself to the task at hand by learning from the modality and anatomy specifications. For this purpose, images will be characterized with local and global features as well as with encoding anatomical structures. A unified framework for registering, segmenting and retrieving medical images will increase the efficiency of clinical workflow and help to establish high quality results by incorporating expert feedback. This will have benefits for clinicians (less tedious tasks), hospitals (saving time and resources) and patients (high-quality care).