Constrained Disentanglement (CODE) Network for CT Metal Artifact Reduction in Radiation Therapy

 摘要 世界卫生组织报告说，癌症是全球第二大死亡原因， 2018年导致960万人死亡。大约50%的癌症患者接受放射治疗 （RT）。他们中的许多人有金属植入物，这会在治疗计划CT图像中引起图像伪影， 在所有放射治疗患者中，估计有15%的患者会危及或妨碍治疗。尽管做了大量CT 金属伪影减少（MAR）研究仍然是CT领域长期存在的挑战之一，没有一个 临床满意度。 这个项目的总体目标是开发尖端的深度学习成像方法和软件解决方案 用于商业CT扫描仪，以消除CT金属伪影，特别是改善RT。我们提出 一个三管齐下的方法，以系统地应对这一挑战，在三个具体目标：（1）对抗性学习 估计正弦图缺失数据和金属痕迹的技术;（2）约束解缠（CODE） 在图像重建过程中，通过后处理，以及在两种数据中， 和图像域;和（3）系统评价我们提出的CT MAR技术和临床翻译 转化为强大的RT计划方法，以最大限度地提高RT治疗计划的准确性，从而提高患者的 成果。我们在CT MAR研究中的协同跟踪记录，特别是与深度成像方法相比， 在过去的三年里，为CT MAR的全新解决方案提供了前所未有的机会。 我们将在数据预处理，图像重建，后处理， 观察者研究和治疗计划在统一的数据驱动框架中协同作用， 项目独特的消除金属伪影及其并发症的放射治疗。 该项目将通过长期的学术和工业合作伙伴关系进行，其中王戈博士在仁- sselaer Polytechnic Institute（RPI）、GE Research Center（GRC）的Bruno De Man博士和Harald Paganetti博士 在马萨诸塞州总医院。虽然我们的团队将在整个项目中密切合作， GRC拥有CT研究和翻译的历史，包括直接原始数据处理，并将专注于目标1。 RPI是断层重建领域的先驱小组，尤其是基于深度学习的CT成像， 瞄准2号。MGH团队处于放射治疗研究的最前沿，将负责Aim 3。 在这个项目完成后，我们将重新定义CT MAR的最新技术，在很大程度上消除CT 金属伪影，并大大提高放射治疗计划和输送的准确性。与 上述建议的网络CT MAR，金属伪影将已基本消除，针对残留 光子和质子治疗计划的误差<10 HU，目的是将临床直径误差减少到 ±3%，金属伪影引起的质子范围误差<2 mm。由于我们的方法是基于软件的， 开放源代码，技术转移和临床翻译的路径是明确定义的，之前也经过了测试。

 ABSTRACT The World Health Organization reported that cancer is the second leading cause of death globally and is re- sponsible for 9.6 million deaths in 2018. Approximately 50% of all cancer patients receive radiation therapy (RT). Many of them have metal implants, which induce image artifacts in the treatment planning CT images and compromise or preclude treatment in an estimated 15% of all radiation therapy patients. Despite extensive CT metal artifact reduction (MAR) research it remains one of the long-standing challenges in the CT field, without a clinically satisfactory solution. The overall goal of this project is to develop cutting-edge deep learning imaging methods and software solutions for commercial CT scanners to eliminate CT metal artifacts in general and improve RT in particular. We propose a three-pronged approach to systematically tackle this challenge in three specific aims: (1) adversarial learning techniques for estimation of sinogram missing data and metal traces; (2) constrained disentanglement (CODE) networks to remove CT image artifacts during image reconstruction, through post-processing, and in both data and image domains; and (3) systematic evaluation of our proposed CT MAR techniques and clinical translation into robust RT planning methods to maximize the RT treatment planning accuracy and thus improve patient outcomes. Our synergistic track records in CT MAR research, especially with deep imaging methods over the past three years, promises an unprecedented opportunity for a brand-new solution to CT MAR. For the first time we will integrate contemporary AI innovations in data preprocessing, image reconstruction, post-processing, observer studies and treatment planning synergistically in a unified data-driven framework, positioning this project uniquely to eliminate metal artifacts and their complications in radiation therapy. This project will be pursued through the long-term academic-industrial partnership among Dr. Ge Wang at Ren- sselaer Polytechnic Institute (RPI), Dr. Bruno De Man at GE Research Center (GRC), and Dr. Harald Paganetti at Massachusetts General Hospital (MGH). While our teams will collaborate closely through the whole project, GRC has a history of CT research and translation, including direct raw data processing, and will focus on Aim 1. RPI is a pioneering group in tomographic reconstruction, especially deep-learning-based CT imaging, and will lead Aim 2. The MGH team is at the forefront of radiation therapy research and will be responsible for Aim 3. Upon completion of this project, we will have redefined the state of the art of CT MAR, largely eliminating CT metal artifacts and substantially improving radiation therapy planning and delivery accuracy. With the above-proposed networks for CT MAR, metal artifacts will have been basically eliminated, targeting residual errors <10 HU for photon and proton therapy planning, with the goal of reducing the clinical diametric error to ±3% and the proton range error due to metal artifacts to <2mm. Since our approach is software-based and open-source, the path for technology transfer and clinical translation is clearly defined, as well tested before.