Developing FAIR practices for cloud-enabled AI deployment for prospective testing

补充项目摘要 目标-家长提案的目标是开发和优化新的深度学习(DL)方法 提高先天性心脏病(CHD)的检测水平。我们正在使用DL和相关方法来提取 来自胎儿超声成像的诊断、生物特征和其他信息。值得注意的是，这项工作 包括对跨越20年、数万名患者、 以及分布在各种医疗保健环境中的多个临床中心。背景-尽管有明显的好处 产前检测冠心病和胎儿超声检测理论上90%以上的冠心病病变的能力，在 实践证明，胎儿CHD的检测率更接近50%。先前的文献表明，这一惊人现象的关键原因 诊断缺口是对胎儿心脏图像的次优采集和解释。初步研究-我们的 冠心病和数据科学领域的多学科团队已成功使用DL区分正常心脏和 合并冠心病者AUC为0.99。进一步的回顾验证表明，我们的模型是一个 适合筛查的异常检测器，并对研究质量和 可以改善临床指南的完整性。下一个合乎逻辑的步骤是使用工作流进行前瞻性测试 足够强大，可以在社区中部署。我们为以下项目开发了临床可行性测试合作伙伴 云部署，执行技术测试，并获得机构批准。补编的目标- 未来的多中心DL算法测试似乎非常适合云部署。然而， 考虑到研究人员、临床医生和预期测试，确定如何以最佳方式优化云平台 是一个悬而未决的问题。云是否可用于实现与最终用户的集成或在其上进行测试 医疗设备(“边缘设备”)也不清楚。我们在本附录中的目标是测试这些方法。 目标-(1)开发用于在云上托管深度学习模型的工作流(用户可以将数据发送到 云，用于直接从边缘设备或通过例如Web应用程序进行推理)。重要的是，我们的方法是 通过利用云服务中提供的工具，实现可查找、可用、可互操作和可重用(公平) 提供程序；开源、文档齐全、单元测试覆盖良好；受版本控制和 完整(“集装箱化”)；便于生物医学研究人员和临床合作伙伴使用和 对于不同的DL型号，可重复使用。(2)我们将在社区胎儿超声诊所进行可行性测试。 环境和影响-这项工作在数据十字路口的出色环境中得到支持 科学、心血管和胎儿成像以及翻译信息学。我们的测试合作伙伴涵盖医疗保健领域 环境和世界，迫使我们的工作流程保持强健。我们将发布我们的工作流程、软件容器、 和研究社区的文档，以及将云引入 边缘设备。建议的工作将提供宝贵的工具和洞察如何最好地使用云服务 对生物医学成像的DL算法进行严格和稳健的前瞻性测试。

Supplemental Project Summary Objective — The goal of the parent proposal is to develop and optimize novel deep learning (DL) approaches to improve detection of congenital heart disease (CHD). We are using DL and related methods to extract diagnosis, biometric characterizations, and other information from fetal ultrasound imaging. Notably, this work includes retrospective evaluation in an imaging collection spanning two decades, tens of thousands of patients, and several clinical centers across a range of healthcare settings. Background — Despite clear benefits to prenatal detection of CHD and an ability for fetal ultrasound to detect over 90% of CHD lesions in theory, in practice the fetal CHD detection is closer to 50%. Prior literature suggests a key cause of this startling diagnosis gap is suboptimal acquisition and interpretation of fetal heart images. Preliminary Studies — Our multi-disciplinary team in CHD and data science has successfully used DL to distinguish normal hearts from those with complex CHD with an AUC of 0.99. Further retrospective validation shows our model to be an anomaly detector appropriate for screening and has generated novel insights into study quality and completeness that can improve clinical guidelines. The next logical step is prospective testing with a workflow robust enough for deployment in the community. We have developed clinical feasibility testing partners for cloud deployment, performed technical testing, and secured institutional approvals. Goals of Supplement — Prospective multi-center testing for DL algorithms seems well-suited to cloud deployment. However, determining how best to optimize cloud platforms with researchers, clinicians, and prospective testing in mind is an open question. Whether the cloud can be used to enable integration with, or testing aboard, end-user medical devices (‘edge devices’) is also unclear. Our goal in this supplement is to test these approaches. Aims — (1) to develop a workflow for hosting deep learning models on the cloud (users can send data to the cloud for inference directly from edge devices or via e.g. web application). Importantly, our approach is Findable, Available, Interoperable, and Reusable (FAIR), by leveraging tools offered across cloud service providers; being open-source, well-documented and well-covered by unit tests; being version-controlled and complete (“containerized”); and being user-friendly for biomedical researchers and clinical partners to use and re-use for different DL models. (2) We will perform feasibility testing in community fetal ultrasound clinics. Environment and Impact — This work is supported in an outstanding environment at the crossroads of data science, cardiovascular and fetal imaging, and translational informatics. Our testing partners span healthcare settings and the world, forcing our workflows to be robust. We will publish our workflow, software container, and documentation for the research community, as well as workflow and best practices for bringing the cloud to edge devices. The work proposed will provide valuable tools and insight into how best to use cloud services for rigorous and robust prospective testing of DL algorithms for biomedical imaging.