Feature-based methods for 3D medical image analysis

医学成像手段，如MRI、超声波和CT，使我们能够在体内可视化3D人体，做出关于解剖学的定量陈述，或者诊断病理或疾病。医学图像数据集合的规模迅速增长，为大规模分析量化解剖变异和了解疾病过程提供了前所未有的机会，例如脑MRI中的阿尔茨海默病或肺部CT图像中的慢性阻塞性肺疾病(COPD)。鉴于医疗保健系统面临的压力，迫切需要能够分析大量不同医学图像数据的健壮、可伸缩的计算工具。 我研究的长期目标是开发最先进的算法来分析体积医学图像数据，特别是大型医学图像集。短期目标将针对主要的计算任务和临床应用。结果将包括对医学图像数据进行大规模计算机辅助分析和诊断的计算工具，这将帮助医疗卫生从业者提供更准确、更循证的决策。 医学图像分析专注于从图像数据中解决与生物有机体的结构和功能有关的临床研究问题的计算算法。主要的研究挑战来自于稳健地处理图像外观和几何形状的变化，可能是由于病理或损伤等异常，在计算处理效率方面扩展到大量数据。主要计算任务包括： *配准：对齐同一底层物体或组织的不同图像。 *分割：描绘感兴趣的组织或对象。 *分类和回归：从图像数据中预测感兴趣的未知临床参数，如疾病状态。 *发现：识别与感兴趣参数相关的图像结构。 我的研究为医学图像分析开发了一个通用的计算框架，名为基于特征的分析(FBA)。FBA将医学图像数据建模为通用图像块或特征的拼贴，例如斑点或角状结构，其以与图像几何(例如，由于扫描仪中的患者位置)和外观(例如，由于成像形态、噪声)的全局变化不变的方式自动提取。因此，基于局部不变特征数据的FBA算法对从不同站点、扫描仪和对象获取的图像数据中的滋扰变化具有高度的鲁棒性。此外，高效的特征索引/匹配算法(例如，近似最近邻方法)用作机器学习方法的基础，所述机器学习方法可扩展到任意大的图像数据集，从而打开了随着训练样本的数量增加而改进的“大数据”式医学图像分析的大门。

Medical imaging modalities such as MRI, ultrasound and CT allow us to visualize the 3D human body in-vivo, to make quantitative statements regarding anatomy or to diagnose pathologies or disorders. Collections of medical image data are growing rapidly in size and offer unprecedented opportunities for large-scale analyses to quantify anatomical variability and understand disease processes, for example Alzheimer's disease in brain MRI or chronic obstructive pulmonary disease (COPD) in lung CT images. Given pressures on health care systems, an urgent need exists for robust, scalable computational tools capable of analyzing large quantities of diverse medical image data. The long term goal of my research is to develop state-of-the-art algorithms for analyzing volumetric medical image data, in particular large sets of medical images. The short term goals will target primary computational tasks and clinical applications. The results will include computational tools for large-scale computer-assisted analysis and diagnosis of medical image data, which will help medical health practitioners provided more accurate, evidence-based decisions. Medical image analysis focuses on computational algorithms for addressing clinical research questions pertaining the structure and function of biological organisms from image data. Major research challenges stem from robustly coping with variations in image appearance and geometry, potentially due to abnormalities such as pathology or injury, scaling to large numbers of data in terms of efficiency in computational processing. Primary computational tasks include the following: * Registration: aligning different images of the same underlying object or tissue. * Segmentation: delineating tissues or objects of interest. * Classification and Regression: predicting unknown clinical parameters of interest from image data, such as disease state. * Discovery: identifying image structure correlated with parameters of interest. My research develops a general computational framework for medical image analysis, entitled feature-based analysis (FBA). FBA models medical image data as a collage of generic image patches or features, such as blob- or corner-like structures, that are automatically extracted in a manner invariant to global variations in image geometry (e.g. due to patient positioning in the scanner) and appearance (e.g. due to imaging modality, noise). FBA algorithms based on local invariant feature data are thus highly robust to nuisance variations in image data acquired from different sites, scanners and subjects. Furthermore, efficient feature indexing/matching algorithms (e.g. approximate nearest neighbor methods) serve as the basis for machine learning methods that scale to arbitrarily large image datasets, opening the door to “Big Data” style medical image analysis that improves as the number of training examples increases.