Computational Methods for Medical Image Interpretation

数字图像数据正在以惊人的速度生成和收集，用于社交媒体通信、国防和安全、工业制造、农业和环境、科学和生物、医疗保健和医学等众多应用。如今，静态和动态图像(视频)约占所有大数据的80%。这种巨大的“可视化”数据隐藏着大量的信息，这些信息对于提供新的见解、建立预测模型、指导决策、执行搜索和管理等至关重要。这就需要为图像分析设计新的计算工具，而不再可能通过视觉检查手动执行。*拟议研究的目标是开发计算机方法，以应对自动、准确、健壮和快速图像分析的关键挑战。具体地说，我将为以下基本图像解释任务开发新的数学模型和计算技术，这些任务对于利用原始图像数据的视觉信息是必要的：(I)图像分割，将图像分割成有意义的部分，用于随后的量化和决策；(Ii)图像配准，用于配对/分组图像配准，从而能够进行比较研究并构建对象的概率模型；以及(Iii)图像分类，用于发现区别视觉模式和特征，以便为视觉数据分配量化值或分类标签。从技术上讲，我将重点介绍基于优化的公式来解决上述三项任务。我将通过将领域专家知识与机器学习技术(来自训练数据库)相结合来开发构建基本目标函数的方法。我将通过为未知数(例如对象形状几何)开发新的表示法来解决这些公式中的最佳性和保真度之间的权衡，以促进有效的优化和推理。*生物医学图像数据的复杂性和多样性，其自动化分析面临的挑战，以及它们为推进医疗保健提供的机会，使它们成为拟议研究的理想应用领域。图像解释结果将为许多临床应用中的诊断和治疗提供宝贵的支持。然而，考虑到它们的社会和经济负担，将强调两个应用：肿瘤学(例如，用于放射治疗的器官和肿瘤的描绘)和神经学(例如，发现神经发育和神经退化的成像生物标记物)。

Digital image data is being generated and collected at an astounding rate for numerous applications such as social media communication, defence and security, industrial manufacturing, agriculture and environment, science and biology, healthcare and medicine, etc. Today, static and dynamic images (video) constitute about 80% of all big data. This big `visual' data hides within it extraordinary amount of information that is critical for providing new insights, building predictive models, guiding decision making, performing search and curation, etc. This is necessitating the design of new computational tools for image analysis that no longer is possible to carry out manually via visual inspection.******The objective of the proposed research is to develop computer methods that address the key challenges towards automated, accurate, robust, and fast image analysis. Specifically, I will develop novel mathematical models and computational techniques for the following fundamental image interpretation tasks, which are necessary for harnessing visual information from raw image data: (i) image segmentation, to partition images into meaningful parts for subsequent quantification and decision making; (ii) image registration, for pair/group-wise image alignment enabling comparative studies and constructing probabilistic models of objects; and (iii) image classification, for discovering discriminatory visual patterns and features for assigning quantitative values or categorical class labels to visual data. More technically, I will focus on optimization-based formulations to solving the above three tasks. I will develop methods to construct the underlying objective functions by combining domain expert knowledge with machine learning techniques (from training databases). I will address the optimizabilty-fidelity tradeoff in these formulations by developing new representations for the unknowns (e.g. object shape geometry) that facilitate efficient optimization and inference.******The complexity and variety of biomedical image data, the challenges facing their automated analysis, and the opportunities they provide for advancing healthcare, make them an ideal application domain for the proposed research. The image interpretation results will be invaluable for supporting diagnostics and therapeutics in many clinical applications. However, two applications will be emphasized given their societal and economic burden: oncology (e.g. organ and tumour delineation for radiation therapy) and neurology (e.g. discovering imaging biomarkers for neuro-development and neuro-degeneration).**