Novel optimization strategies for medical image analysis

解剖学和功能医学成像模式，例如MRI和PET，使医生能够窥视人体内部，并观察对理解，诊断和治疗疾病至关重要的大量数据。获取的医疗数据量正在迅速增长。图像的维数已经从二维标量图像增加到动态三维多值场。这导致图像数据无法用传统的视觉检测进行有效处理。因此，用于医学图像分析（MIA）的自动计算工具在现代医疗保健系统中变得不可或缺。图像分割、配准和形状分析是MIA中三个最重要、最普遍的任务，它们构成了图像解释和量化任务的关键。分割是识别图像中的感兴趣区域的过程（例如，测量心肌的壁厚度），而配准是找到图像之间有意义的对应关系的过程（例如，跨受试者或时间进行比较）。形状分析捕获几何和拓扑特性，并揭示有关疾病阶段，治疗进展或生长速率的关键信息。尽管在过去的几十年里在这些领域取得了许多进展，但精确和自动的MIA算法仍然无法解决问题。这些算法中的绝大多数依赖于解决优化问题。然而，很少的工作一直致力于评估的适当性的目标函数被优化或集成到过程中的高层次的领域知识。该提案将侧重于研究评估目标函数的正式方法，并从一开始就使用严格的数学和计算技术进行设计。这项研究还将补充低层次的优化技术与高层次的，知识驱动的MIA战略，使用一种新的人工生命框架。其目标是确保自动MIA算法的更高准确性，以推进计算机辅助诊断、计算机辅助干预以及依赖医学成像的医疗保健的许多其他方面。

Anatomical and functional medical imaging modalities, e.g. MRI and PET, are allowing physicians to peer inside the human body and observe a wealth of data crucial for understanding, diagnosing and treating diseases. The volume of medical data acquired is growing rapidly. The dimensionality of images has increased from 2D scalar images to dynamic 3D multi-valued fields. This is resulting in image data that cannot be effectively processed with traditional visual inspection. Therefore, automated computational tools for medical image analysis (MIA) are becoming indispensable in modern healthcare systems. The three most important and ubiquitous MIA tasks are image segmentation, registration, and shape analysis, which constitute the crux of image interpretation and quantification tasks. Segmentation is the process of identifying regions of interest in an image (e.g. to measure wall thickness of the myocardium), whereas registration is the process of finding meaningful correspondence between images (e.g. to compare across subjects or time). Shape analysis captures geometric and topological properties and reveals crucial information about disease stages, treatment progress, or growth rates. Despite numerous advances in these areas in the past few decades, accurate and automatic MIA algorithms continue to defy solution. The vast majority of these algorithms rely on solving optimization problems. However, very little work has been devoted to evaluating the appropriateness of the objective functions being optimized or to integrating high-level domain knowledge into the process. This proposal will focus on studying formal approaches for evaluating objective functions and designing them from the outset using rigorous mathematical and computational techniques. This research will also complement low-level optimization techniques with high-level, knowledge-driven MIA strategies using a novel artificial life framework. The goal is to ensure higher accuracy of automated MIA algorithms, in order to advance computer-aided diagnosis, computer-assisted interventions, and the many other aspects of healthcare that rely on medical imaging.