DC-AIDE - Dedicated Clinical Artificial Intelligence Deployment Equipment

深度学习已经成为生物医学成像分析的一项关键技术，但由于对计算和数据管理的要求很高，对于非专家来说，使用它很复杂。该项目将开发一个平台，旨在通过尖端的深度学习方法简化多模态生物医学成像和患者数据的大规模统计分析。拟议的基础设施将提供对最先进算法的访问，并定义一个标准化的数据基础设施，易于在异构环境中部署。我们的原型将提供一个有效的机制来共享预训练的人工智能算法和先进的分析工具。该平台以生物医学研究界为目标，将为科学家提供新颖、强大和经过验证的工具，以应对诸如基于图像的疾病表型和预测建模等挑战。最新的分析管道将被实现并打包成易于使用的工具箱，可直接部署到临床工作流程中，允许提取成像生物标志物和定量测量。我们的方法基于三个基本原则：数据链接（跨系统）、数据治理（维护患者隐私和法律/道德合规）和数据互操作性（使用公共api和开放标准）。为了实现这一点，我们将建立在现有模型的基础上：数据将保留在安全的环境中，并引入人工智能算法，在我们合作诊所的防火墙内对敏感的患者数据进行培训。将支持两种方法：一种是安全的学习编排服务器，用于在我们的合作医院大学埃尔兰根诊所（UKER）内为安全数据enclave执行学习协调；另一种是安全沙箱，允许在FAU大学托管的安全环境中进行模型开发。就像在联邦学习范例中一样，大多数模型将通过我们的基础设施移动，而不是数据。我们将在这些环境中放置（并支持）基础设施，并得到FAU生物医学工程人工智能部门（AIBE）、UKER放射学部门和Erlangen区域研究中心（RRZE）的支持，以提供所需的能力。我们将把HL7 FHIR标准中的多模式机器学习和数据链接功能与PACS数据的XNAT相结合，结合其他开源工具，为成像和患者记录数据提供完整的解决方案。这使得所提出的解决方案具有高度的互操作性和可扩展性，可用于其他诊所，并将实现与医疗信息学倡议等的集成。这种方法将提供安全且符合法规的访问PACS和诊所的电子患者记录，以及前瞻性同意的研究数据集，这将允许临床医生和科学家以最有效的方式对大型患者队列进行可重复的研究。

Deep learning has emerged as a key technology in biomedical imaging analytics, but it is complex to use for non-experts due to demanding computational and data management requirements. This project will develop a platform that aims to simplify large-scale statistical analysis of multi-modal biomedical imaging and patient data with cutting-edge deep learning methods. The proposed infrastructure will provide access to state-of-the-art algorithms and define a standardised data infrastructure with easy deployment in heterogeneous environments. Our prototype will provide an effective mechanism for sharing pre-trained AI algorithms and advanced analysis tools. The platform targets the biomedical research community and will equip scientists with novel, powerful and validated tools to tackle challenges such as image-based disease phenotyping and predictive modelling. Latest analysis pipelines will be implemented and packaged into easy-to-use toolboxes that are directly deployable into clinical workflows, allowing extraction of imaging biomarkers and quantitative measurements. Our approach is based on three fundamental principles: data linkage (across systems), data governance (maintaining patient privacy and legal/ethical compliance), and data interoperability (using public APls and open standards). To deliver this, we will build on an existing model: data will be retained within a secure environment, with Al algorithms brought in for training on sensitive patient data inside our partner clinic’s firewall. Two approaches will be supported: a secure learning orchestration server to perform learning coordination for secure data enclaves within our partner hospital University clinic Erlangen (UKER), and secure sandboxes allowing model development within a university-hosted secure environment at FAU. Like in a federated learning paradigm, mostly models will move through our infrastructure, not the data. We will place (and support) infrastructure within these environments, with the support of the Department Artificial Intelligence in Biomedical Engineering (AIBE) at FAU, the Radiology Department at UKER, and Regionales Rechenzentrum Erlangen (RRZE), to provide the capabilities required. We will combine the capabilities of multi-modal machine learning and data linkage within the HL7 FHIR standard with XNAT for PACS data to provide a complete solution for imaging and patient record data, combined with additional open-source tools. This makes the proposed solution highly interoperable and scalable to other clinics and will enable integration with, e.g., the Medical Informatics Initiative. This approach will provide secure and regulatory compliant access to the PACS and electronic patient records of clinics, in addition to prospectively consented research datasets, which will allow clinicians and scientists to conduct reproducible research in the most efficient way on large patient cohorts.