Probabilistic inference in computer vision and medical imaging

医学成像的进步使医学专家能够获得大量的信息。然而，在许多情况下，如果没有进一步的解释，这些信息很难用于指导和诊断。例如，图像引导神经外科系统(IGN)在手术过程中向外科医生提供了丰富的患者特定的术前图像(例如MRI、fMRI)，用于规划和指导。然而，开颅手术中复杂的大脑运动降低了使用这些图像进行指导的有效性。拟议研究计划的目标是为广泛的医学背景开发新的计算机视觉概率框架，在这些背景下，它们有可能导致医疗诊断和患者护理的具体改进。具体地说，我们将从开发概率医学图像配准工具开始，该工具将允许我们量化结果解释中的不确定性。在神经外科的背景下，这将使我们能够将跟踪的术中图像(例如超声)与术前图像进行匹配，以便量化和纠正脑变形，并在系统对解释有信心的情况下与外科医生沟通。这导致了第一个主动图像引导神经外科系统(AIGNS)，其中视觉反馈将允许外科医生优化术中成像系统的放置和使用，以减少解释中的歧义。临床好处包括更快、更准确的手术程序，以及减少患者创伤和医院成本。我们还将把该框架应用于乳腺癌放射治疗，目标是实现更准确的肿块切除(切除肿瘤)部位的放射治疗。最后，这项提议旨在开发新的概率框架，以自动从大脑图像的大型数据集中学习图像模式的局部统计可变性，可能导致我们对健康人脑的自然可变性的理解取得突破，并发现神经疾病(如阿尔茨海默氏症)的新生物标记物。

Advances in medical imaging have permitted a large amount of information to be made available to medical experts. However, in many cases, the information is difficult to use for guidance and diagnosis without further interpretation. Image-guided neurosurgery systems (IGNS), for example, have provided a wealth of patient-specific pre-operative images (e.g. MRI, fMRI) to the surgeon during the procedure to use for planning and guidance. However complex brain movements during open-skull operations reduce the effectiveness of using these images for guidance. The objectives of the proposed research program are to develop new probabilistic frameworks in computer vision for a wide range of medical contexts, where they have the potential to lead to concrete improvements in medical diagnosis and patient care. Specifically, we will begin by developing probabilistic medical image registration tools that will permit us to quantify the uncertainties in the resulting interpretation. In the context of neurosurgery, this will permit us to match tracked intra-operative images (e.g. ultrasound) to pre-operative images in order to quantify and correct for brain deformations, and to communicate to the surgeon where the system has confidence in the interpretation. This leads to the first active image-guided neurosurgery system (AIGNS) where visual feedback will permit a surgeon to optimize placement and use of intra-operative imaging systems to reduce ambiguities in the interpretation. Clinical benefits include faster and more accurate surgical procedures, and reductions in patient trauma and hospital costs. We will also apply the framework to breast cancer radiotherapy, with the goal of attaining more accurate radiation treatment of lumpectomy (resected tumour) sites. Finally, this proposal aims to develop new probabilistic frameworks to automatically learn the local statistical variability of image patterns from large datasets of brain images, potentially leading to breakthroughs in our understanding of the natural variability of healthy human brains and to discoveries of new biomarkers of neurological diseases (e.g. Alzheimer's).