Developing Enabling PET-CT Image Analysis Tools for Predicting Response in Radiation Cancer Therapy

摘要 集成正电子发射断层扫描和计算机断层扫描(PET-CT)已成为一种 现代癌症治疗中不可缺少的工具。准确的目标勾画是不可避免的第一步 充分发挥PET-CT的潜力。然而，在目前的临床实践中，这 重要的任务通常是在一个切片一个切片的基础上以可视方式执行，对 自动分割工具。最先进的PET-CT分割技术依赖于 单一模式或融合的PET-CT数据，可能不能充分利用这两种模式，因此 影响了分割的准确性。此外，最先进的治疗反应 预测方法高度依赖于手工制作的图像特征和参数，这构成了一个 限制其在临床上广泛应用的因素。本研究建议发展快速客观的正电子发射计算机断层扫描 促进双模式成像在两种大规模临床应用中的分析方法 试验研究和日常临床护理。所提出的方法的新颖之处在于首次 引入共分割技术用于PET-CT肿瘤的勾画，识别肿瘤的轮廓差异 PET中的肿瘤与CT中的肿瘤不同。将探索和纳入新的PET-CT特定先例 该分割框架，进一步提高了分割的准确性。建议数 反应预测方法是建立在我们的PET-CT合作的准确的肿瘤定义之上的 分段方法，用一种创新的卷积神经网络设计用于自动 直接从PET-CT扫描中学习分层特征，从而实现高精度的预测 回应。开发的方法将与使用的最先进的方法进行比较测试 今天。这些方法的性能将在足够的数据样本中进行统计评估 大小。

ABSTRACT The integrated Positron Emission Tomography and Computed Tomography (PET-CT) has become an indispensable tool in modern cancer therapy. Accurate target delineation is an inevitable first step towards fully making use of the potentials of PET-CT. However, in current clinical practice, this important task is typically performed visually on a slice-by-slice basis with very limited support of automated segmentation tools. The state-of-the-art PET-CT segmentation techniques rely on either a single modality or the fused PET-CT data, which may not fully take advantage of both modalities, thus compromising the segmentation accuracy. In addition, the state-of-the-art therapeutic response prediction methods highly rely on the handcrafted image features and parameters, which poses a limiting factor for their wide use in clinic. This research proposes to develop fast and objective PET-CT analysis methods to facilitate the utilization of the dual modality imaging for both large-scale clinical trial research and daily clinical care. The novel feature of the proposed methods is the first time to introduce co-segmentation for PET-CT tumor delineation, which recognizes the contour difference of tumors in PET from those in CT. New PET-CT specific priors will be explored and incorporated into the segmentation framework, further improving the accuracy of segmentation. The proposed response prediction method is built on the accurate tumor definition from our PET-CT co- segmentation approach, with an innovative design of a convolutional neural network for automatically learning hierarchical features directly from the PET-CT scans, leading to highly accurate prediction of response. The developed methods will be tested in comparison with state-of-the-art methods utilized today. The performance of the methods will be statistically assessed in data samples of sufficient sizes.