Probabilistic Inference in Computer Vision and Medical Imaging

提出的多学科研究计划的目标是为计算机视觉问题开发新的概率公式，这些问题应用于特别具有挑战性的医学成像环境，例如神经病学和神经外科，在这些领域，它们有可能在鲁棒性、准确性或速度方面比传统技术取得重大进步。首先，新的复杂的，多层次的概率图形模型将被设计用于医学成像领域的各种开放和困难的病理检测和分割任务。新的，高层次的，不规则的网格（即非晶格）随机场将被开发，以适当地表示上下文（空间和纹理）信息。推理将遍历体素、损伤和上下文级别的随机场，直到收敛。为这些领域开发新的图形模型将需要脱离流行的邻域模型，在节点和边缘信息的适当模型设计上进行创新，以及适当的推理技术。除了对机器学习、医学图像分析和计算机视觉的理论贡献外，所得到的模型还应显著改善现实、临床领域中各种困难和重要的病理检测和分割任务，如多发性硬化症病变分割、脑肿瘤亚类分割和乳腺癌检测。通过与当地公司的合作，我的研究团队将可以访问世界上任何医学图像分析小组的最大的真实，多中心，临床试验，手动注释的患者MRI图像数据集，以测试这些方法。

The objectives of the multidisciplinary research program proposed are to develop new probabilistic formulations for problems in computer vision applied in particularly challenging medical imaging contexts, such as those presented in neurology and neurosurgery, where they have the potential to lead to significant advancements in terms of robustness, accuracy or speed over traditional techniques. First, new sophisticated, multi-level probabilistic graphical models will be devised for a wide variety of open and difficult pathology detection and segmentation tasks in the domain of medical imaging. New, high level, irregular grid (i.e. non-lattice) random fields will be developed to properly represent contextual (spatial and textural) information. Inference will iterate across voxel, lesion and context level random fields until convergence. Development of new graphical models for these domains will require departures from popular neighbourhood models, innovation in design of appropriate models for node and edge information, as well as appropriate inference techniques. In addition to theoretical contributions machine learning, medical image analysis and to computer vision, the resulting models should significantly improve a wide variety of difficult and important pathology detection and segmentation tasks in real, clinical domains, such as Multiple Sclerosis lesion segmentation, the segmentation of sub-classes of brain tumours, and breast cancer detection. Through collaborations with a local company, my research team will have access to the largest dataset of real, multicenter, clinical trial, manually annotated, patient MRI images of any medical image analysis group in the world in which to test the methods. The research program also includes the development of new fast multi-modal image registration techniques for time-sensitive domains (e.g. image-guided neurosurgery), where, by embedding learned context-specific, probabilistic feature/pixel selection into the registration strategy, we hope to show significant speedups without compromising accuracy. Probabilistic techniques will be developed for the analysis of neurological images, focusing on learning the variability of cortical folding patterns in different sub-populations (e.g. age, IQ) and to automatically discover biomarkers of neurological disorders, such as MS and Alzheimer's patients. This will lead to innovations in terms of robust feature learning and clustering in contexts with large variability.