Image-guided non-invasive tracking for radiotherapy using machine learning

肿瘤患者的生理运动是一个普遍的不确定性，导致传统放射治疗和治疗超声中的辐射输送不准确。通过磁共振成像或超声的术中引导，辅以复杂的图像分析，将在提供可靠、准确和实时的肿瘤运动信息方面发挥重要作用。本项目的目的是通过采用机器学习的最新方法以及通过高精度关键点配准获得的训练数据，推进当前最先进的术中运动估计。而不是依赖于经典的统计运动模型，我们计划学习图像特征和以前看到的运动之间的非线性回归函数的级联。将采用在放射输送开始之前采集的初始患者特定图像序列，以使用准确但更耗时的图像配准生成训练数据。这将产生位于器官表面、血管和肿瘤本身上的相关关键点的对应关系。基于计算机视觉和机器学习领域的最新进展，这些对应关系用于训练一个非线性模型，该模型将运动与使用深度卷积网络为每种模态学习的专门适应的鲁棒图像特征联系起来。通过使用形状增强与人口数据和在线学习程序，我们将能够限制训练序列中所需的帧数。将预先获取的图像的先验知识引入学习模型将使我们不仅能够准确跟踪可见的肿瘤或重要的血管结构，而且重要的是还可以跟踪危险器官的形状和位置，这将有助于避免对健康组织的辐射，从而避免放射治疗的副作用。所学习的模型还可以帮助估计暂时在视野之外或被遮挡的解剖结构的位置。由于计划的回归模型的高计算效率，预期每图像帧几毫秒的处理时间。

Physiological motion of tumour patients is a prevailing uncertainty that leads to inaccurate delivery of radiation in both conventional radiotherapy and therapeutic ultrasound. Intra-operative guidance by magnetic resonance imaging or ultrasound, complemented with sophisticated image analysis is going to play a vital role in providing reliable, accurate and realtime information of tumour motion.The aim of this project is to advance the current state-of-the-art of intra-operative motion estimation by employing recent approaches from machine learning together with training data obtained with highly accurate keypoint registration. Instead of relying on classical statistical motion models, we plan to learn a cascade of nonlinear regression functions between image features and previously seen motion. Initial patient-specific image sequences acquired before the start of radiation delivery will be employed to generate training data using accurate but more time-consuming image registration. This will yield correspondences for relevant keypoints that are located on organ surfaces, vessels and the tumour itself. Building upon recent advances from the field of computer vision and machine learning, these correspondences are used to train a nonlinear model that links motion with specifically adapted robust image features that are learned for each modality using deep convolutional networks. By using shape augmentation together with population data and an online learning procedure, we will be able to limit the required number of frames in the training sequence.Incorporating the prior knowledge of pre-acquired images into the learned model will enable us to not only accurately track visible tumour or important vessel structures, but importantly also the shape and position of organs at risk, which will help to avoid radiation to healthy tissues and therefore side-effects of radiotherapy. The learned model may furthermore help to estimate the position of anatomies that are temporarily out of view or occluded. Due to the high computational efficiency of the planned regression model processing times of few milliseconds per image frame are expected.