CT and CXR Phenotyping Platform for Assessing COVID-19 Susceptibility and Severity

摘要 新冠肺炎于3月11日被世卫组织宣布为大流行。自那以来，已有815万人 全球确诊病例，病死率从16.3%到0.1%不等。在美国，已经有了 截至2020年6月16日，2187,202例，病死率为5.4%。这种传染病的严重程度 强调需要开发新的方法来定义谁处于最高的感染风险 急性症状。X射线(CXR)和计算机体层摄影术(CT)在检测和 新冠肺炎肺损伤的随访观察。它还提供了一个独特的机会来定义定量生物标记物 可以利用感染前和早期感染来识别疾病急性期的易感对象 放射学检查。 这项建议的广泛目标是提供对急性新冠肺炎易感标志物的更好了解 基于人工智能的放射检查方法，包括CT和CXR。CT提供了一种独特的方法 肺的表型及其变化。正常组织的细微变化与 CT可以检测到的全身性炎症。我们假设急性冠状病毒感染的易感对象- 19个疾病的演变表现为炎症的正常实质征象，可以在CT扫描上测量到 在感染之前或病毒感染的早期阶段。我们将开发新的计算 识别与炎症的正常实质以及早期一致的放射学模式的方法 新冠肺炎损伤和计算机放射组学特征，可捕捉X线图像的异质性 每种肺型的表达。我们将定义基于CT的急性新冠肺炎易感性的新生物标记物 使用梯度提升决策树和特征重要性。然后我们将把量化的 利用基于深度神经网络的图像转换方法对CXR图像中最相关的特征进行处理。 最后，我们将使用临床友好的端到端将这些自动化工具集成到CIP工作站中 支持全球临床研究的工作流程。我们将继续支持和传播 这个工具在整个研究社区中的应用。在过去的15年里，我们团队开发了胸部成像 Platform(CIP)，美国国立卫生研究院资助的用于胸部CT扫描自动表型的开源软件工具 这在慢性肺病研究界被广泛使用。自大流行开始以来，CIP 已经习惯于使用现有的密度测量指标来表征新冠肺炎。我们致力于 以开放工具包的形式免费分发的开放科学是催化应用 在这场大流行的背景下，人工智能和成像。

Abstract COVID-19 was declared a pandemic by WHO on March 11. Since then, there have been 8.15 million confirmed cases worldwide with a case fatality rate ranging from 16.3% to 0.1%. In the US, there have been 2,187,202 cases with a 5.4% case fatality rate as of June 16, 2020. The magnitude of this infectious disease has stressed the need to develop novel methodologies to define who are at the highest risk of developing acute symptoms. X-Ray (CXR) and Computed Tomography (CT) play a fundamental role in the detection and follow-up of the COVID-19 lung injury. It also provides a unique opportunity to define quantitative biomarkers that may identify susceptible subjects to the acute phase of the disease using pre-infection and early infection radiological exams. This proposal's broad objective is to provide a better understanding of acute COVID-19 susceptibility markers based on artificial intelligence approaches on radiological exams, both CT and CXR. CT offers a unique way to phenotype the lung and its changes. Subtle changes of normal parenchyma have been associated with systemic inflammation that can be detected on CT. We hypothesize that susceptible subjects for acute COVID- 19 disease evolution will express inflamed normal parenchymal signatures that can be measured on CT scan prior to the infection or in the early phases of the viral infection. We will develop new computational approaches to identify radiographic patterns consistent with inflamed normal parenchyma as well as early COVID-19 injury and compute radiomics signature that can capture the heterogeneity of the radiographic expression for each lung pattern. We will define new CT-based biomarkers for acute COVID-19 susceptibility using Gradient Boosting decision trees and feature importance. We will then translate the quantification of the most relevant features in CXR image using image translation approaches based on deep neural networks. Finally, we will integrate these automated tools in the CIP workstation using clinically friendly end-to-end workflows to empower clinical investigations across the world. We will continue the support and dissemination of this tool across the research community. Over the last 15 years, our group has developed the Chest Imaging Platform (CIP), an NIH-funded open-source software tool for the automated phenotyping of chest CT scans that is widely used in the chronic lung disease research community. Since the beginning of the pandemic, CIP has been used to the characterization of COVID-19 using existing densitometric metrics. Our commitment to open science in the form of open toolkits that are freely distributed is fundamental to catalyze the application of AI and imaging in the context of this pandemic.