Artificial Intelligence from Chest CT to Assess COVID-19 Clinical Trials

在大流行期间的短时间内，NIH、合作伙伴和扩展团队通过TCIA部署了公共多国CT数据集（迄今为止有650个CT），并建立了数据共享和国际合作伙伴关系。NIH合作伙伴已经让第三方部署了一个具有拖放功能的网站，用于CT扫描。计算机行业合作伙伴共同开发了深度学习的算法和工具包，用于COVID AI分类的免费“免费软件”公共管道。部署一个管道来促进深度学习，而没有隐私风险或限制，已经被证明用于学术和研究使用共享模型，通过公共/私人合作伙伴关系的联合学习。这允许共享AI模型权重，而实际数据本身不会从其私有控制源移动。NIH团队已经验证了来自成像行业合作伙伴的用于COVID-19特定分析和量化的预商用beta软件。 免疫和炎症分子动力学与CT AI成像特征的相关性在2020财年过去几个月的临床试验中得到了验证。目前正在开发标准化工具，用于通过CT表征指标统一量化COVID-19疾病：1. COVID-19感染率，2. %毛玻璃样混浊，3. %合并组分。这些指标被假定为响应的相关性。与此同时，NIH团队帮助开发和验证来自大型中央注释源数据的公共和私人商业软件工具。 迫切需要标准化的工具来测量和与临床结果的相关性，例如单克隆抗体或抗病毒治疗反应。非深度学习工具可以测量简单的密度统计，这可能不太具体。尽管机器学习或基于放射学的CT特征评估可能执行类似的功能，但对特定成像特征的依赖可能导致数据点较少以及标准化和可重复性较差的工具。深度学习和人工智能工具与所提出的方法应该提供更可重复和标准化的模型，这些模型更有可能被推广到非常广泛和异质的社区环境中，具有患病率和患者人群的固有可变性，以及动态演变的人口动态。 开发的CT AI分类模型可检测COVID-19，并将COVID-10与流感和其他非COVID-19诊断区分开来（Nature Communications 2020）。这种检测和区分/分类模型在秋季北方流感季节可能是有用的。此外，该模型可能在护理点和紧急情况下有用，以便快速识别和隔离COVID-19的典型浸润。在一种使用方法中，在无症状患者被允许离开CT套件之前，将存在护理点警报以用于随后的快速放射科医师检查。 症状前CT AI混浊的时间动态与COVID-19症状前病毒动态相关。使用CT AI工具对无症状CT扫描进行量化。在具有早期疾病的患者中的系列/顺序CT扫描（平均> 40天，每个患者4次扫描）在广义曲线中统计学地显示，并且与血清实验室（诸如CRP、降钙素原、LDH、WBC等）相关。这些在轻度和早期疾病中的广义曲线已被证明提供了参考，使得能够预测偏离该曲线作为不良结局或更高水平干预的风险因素。还开发了基于CT AI的肺部疾病自动化和标准化量化模型。 已经完成了使用商业预发布测试版软件对1000次CT扫描的回顾性验证，并评估了观察者间性能。与国家卫生研究院跨机构工作组建立了伙伴关系，包括NIAID IRF、NIBIB MIDCR、NCI、NCATS、NIDDK、N3 C、RSNA、RICORD、ACR、AAPM和MITA。涉及CT等医学成像的COVID-19相关数据科学的集中通信和发现途径是一个共同的主题和目标。

In the short time during the pandemic, NIH, partners, and extended teams have deployed public multinational CT dataset via TCIA (650 CT's to date), with data sharing and international partnerships. NIH partners have had 3rd parties deploy a website with drag-and-drop functionality for CT scans. Computer industry partner has jointly developed algorithms and toolkits for deep learning for no-cost "freeware" public pipeline for COVID AI classification. Deployment of a pipeline to promote deep learning without he privacy risk or restrictions has been demonstrated for academic and research use with shared models, via federated learning via a public/private partnerships. This allows sharing of AI model weights, without the actual data itself moving around from its private controlled source. NIH team has validated pre-commercial beta software for COVID-19 specific analysis and quantification, from an imaging industry partner. Correlation of immune and inflammatory molecular dynamics with CT AI imaging profiles was validated in clinical trials during the past few months in FY 2020. Development is underway for standardized tools for uniform quantification of COVID-19 disease via CT characterization of metrics: 1. % COVID-19 involvement, 2. % ground glass opacities, 3. % consolidation components. These metrics are postulated to be correlates of response. In parallel, NIH teams helped develop and validate public and private commercial software tools from large central annotated source data. There is a critical need for standardized tools for measurement and correlation with clinical outcomes, such as monoclonal antibody or anti-viral therapy responses. Non-deep learning tools may measure simple density statistics, which is likely to be less specific. Although machine learning or radiomics-based CT feature assessment might perform a similar function, the reliance upon specific imaging features may result in fewer data points and a less standardized and reproducible tools. Deep learning and AI tools with the methodology proposed should provide more reproducible and standardized models, which have better chances for being generalized to a very broad and heterogeneous community setting, with inherent variabilities in prevalence and patient populations, and dynamically evolving population dynamics. The CT AI classification model developed detects COVID-19 and differentiates COVID-10 from influenza and other non-COVID-19 diagnoses (Nature Communications 2020). This detection and differentiation/classification model might be useful during the fall Northern hemisphere influenza season. In addition, the model may be useful at point of care and emergency settings in order to rapidly identify and isolate typical infiltrates of COVID-19. In one method of use, there would be point of care alarming for subsequent rapid radiologist review before an asymptomatic patient were allowed to leave the CT suite. Temporal dynamics of pre-symptomatic CT AI opacities correlated with COVID-19 pre-symptomatic viral dynamics. Asymptomatic CT scans were quantified with CT AI tools. Serial / sequential CT scans in patients with early disease (dating average of > 40 days and 4 scans per patient) were statistically displayed in generalized curves and correlated with serum labs such as CRP, pro-calcitonin, LDH, WBC, etc). These generalized curves in mild and early disease have been shown to provide a reference, enabling prediction of deviation from this curve as a risk factor for poor outcome or higher level interventions. Models have also been developed for automated and standardized quantification of lung disease based on CT AI. Retrospective validation of 1000 CT scans with a commercial pre-release beta software has been accomplished, with interobserver performance assessed. Partnerships with the Trans-NIH working group has been forged, including NIAID IRF, NIBIB MIDCR, NCI, NCATS, NIDDK, N3C, RSNA, RICORD, ACR, AAPM, and MITA. Centralized communication and discovery pathways for COVID-19-related data science that involves medical imaging like CT is a common theme and goal.