CRII: CNS: OCEAN: A Once-for-All Edge Collaboration System for Medical Imaging

鉴于以数据驱动的方式学习复杂表示的能力，深度学习算法极大地影响了医学成像领域。然而，实现可靠的医疗保健人工智能（AI）技术是具有挑战性的，由于缺乏专业标签的成本，人类标签的异质性，以及对医疗保健信息隐私的关注所造成的注释问题。为了克服这种稀缺的注释问题，目前有两种解决方案。一个解决方案是实现多机构合作，以培训一个专用模型。另一个是从训练可共享模型到微调多个任务来获得多尺度特征。第一种方法涉及巨大的计算和通信成本，仅为一种类型的任务训练专用模型;而第二种方法目前仅适用于自动驾驶汽车，其中在单个节点上频繁收集和处理大量数据。该项目旨在通过创新性地训练多个医学成像任务以及缓存机制来填补这一空白，以建立一个高效和有效的多机构协作系统。提出的研究的目标是设计一个多任务学习模型的医学成像任务，设计一个缓存机制，系统激活相关部分时，解释一个特定的任务，并开发一个原型的分布式多任务学习系统的医学成像模型和系统性能的评估所提出的解决方案。建立一个针对所有医学成像的协作系统将是实现临床实践中实用AI的重大技术突破。通过促进机构内部和机构之间的模型共享，所提出的系统可以解决稀缺的注释问题，并加速临床检测、诊断和治疗，以使医疗保健专业人员受益。此外，作为一个通用框架，拟议的系统也将部署到具有类似应用需求的其他领域，例如陆地、空中和海上的互联自动驾驶和其他移动系统，其中多个节点作为一个单元工作，执行一系列任务。从教育方面来看，该系统将被开发为医疗AI的基础实验平台，并可以轻松移植到其他场景，如智能家居和智能交通。该计划将用于本科生和研究生的教育和研究，旨在激发学生对边缘智能的兴趣，扩大对智能计算、网络和系统的参与，提高教育的多样性、包容性和公平性。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Given the ability to learn complex representations in a data-driven manner, Deep learning algorithms have greatly impacted the medical imaging field. However, achieving reliable healthcare artificial intelligence (AI) technology is challenging due to the scarce annotation problem caused by the cost of professional labeling, heterogeneity of human labels, and concerns over healthcare information privacy. To overcome this scarce annotation problem, there are currently two solutions. One solution is enabling multi-institutional collaboration to train a dedicated model. Another is obtaining multi-scale features from training a shareable model to fine-tuning multiple tasks. The first method involves huge computation and communication costs to train a dedicated model for only one type of task; while the second method is currently only applied to the autonomous car where a rich amount of data is collected and processed frequently on a single node. This project seeks to fill this gap by innovatively training multiple medical imaging tasks together with a cache mechanism to build an efficient and effective multi-institutional collaborative system. The goal of the proposed research is to design a multi-task learning model for medical imaging tasks, design a cache mechanism for the system to active relevant portion when interpreting a specific task, and develop a prototype distributed multi-task learning system for medical imaging to evaluate the model and system performance of the proposed solutions. Building a one-for-all collaboration system for medical imaging will constitute a significant technological breakthrough toward achieving practical AI in clinical practice. By facilitating the model sharing within and among institutions, the proposed system can address the scarce annotation problem and accelerates clinical detection, diagnosis, and treatment, to benefit healthcare professionals. Furthermore, as a general-purpose framework, the proposed system will also be deployed to other fields with similar application requirements, such as connected autonomous driving and other mobility systems on land, in the air, and at sea, where multiple nodes work as a unit for a series of tasks. From the education aspect, the proposed system will be developed as a basic experimentation platform for healthcare AI and can be easily transplanted to other scenarios, such as smart homes and smart transportation. The proposed system will be used for undergraduate and graduate education and research with the goal to inspire students' interests in edge intelligence, broaden participation in intelligent computing, networking, and systems, and enhance education diversity, inclusion, and equity.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.