Generalizable Deep Learning Networks for Dual-tracer Amyloid/Tau PET/MRI Imaging of Alzheimer's Disease

项目摘要 阿尔茨海默病(AD)是一种毁灭性的神经退行性疾病，是一种重大的公共卫生危机， 目前影响着超过580万美国人，随着人口老龄化，预计还会增加。正电子发射 断层扫描(PET)成像可以识别AD的标志性蛋白病变，包括淀粉样蛋白斑块 和神经原纤维缠结(主要由tau蛋白组成)在大脑中积累。尽管有明显的证据 例如，需要更多的PET神经成像来阐明淀粉样蛋白和tau沉积的序列。 临床前阿尔茨海默病，其在大研究人群的纵向成像研究中增加的实用性受到以下限制 招聘和成本。特别是，对参与者来说，多次访问扫描站点将是困难的 居住在遥远的地方，注射放射性示踪剂的高昂成本将限制正电子发射计算机断层扫描研究的可扩展性。 在这个项目中，我们建议使用基于深度学习的卷积神经网络(CNN)来增强 用于AD成像的超低剂量淀粉样蛋白和tau PET。我们的具体目标是(1)验证其诊断价值 在实际的超低剂量淀粉样蛋白和tau成像过程中，注射剂量低至1%的CNN 原始的和实际的超低剂量数据，以验证用于后续AIMS和未来的模拟 研究；(2)将超低剂量CNN应用于在其他PET系统和示踪剂上收集的数据，以便 论证CNN的普适性；以及(3)评估深度学习辅助超低剂量的价值 淀粉样蛋白和tau PET用于追踪临床前阿尔茨海默病患者的认知功能下降。 这项工作的创新之处在于除了使用高级机器学习外，还使用了多模式成像 能够在极低剂量水平下获取诊断级PET图像的技术。表演 实际的超低剂量PET采集本身也是非常新颖的。这项提议的结果是 限制大规模临床纵向成像的因素，缩短跨越多天的采集 以及使用连续的超低剂量和全剂量双示踪剂扫描方案在一次访问中访问几个小时。 还可以显著减少剂量，允许更频繁地进行淀粉样蛋白/tau PET扫描。这 灵活性不仅将增加PET的实用性，有助于痴呆症的纵向研究，还将使未来 随着这些示踪剂的开发，对多个基于PET的生物标记物进行全面成像。

Project Summary Alzheimer’s Disease (AD) is a devastating neurodegenerative disorder and a major public health crisis, currently affecting over 5.8 million Americans and expected to rise as the population ages. Positron emission tomography (PET) imaging can identify the hallmark proteinopathies of AD, including amyloid protein plaques and neurofibrillary tangles (composed primarily of tau protein) accumulating in the brain. While there is evident need for more PET neuroimaging, for example, to elucidate the sequence of amyloid and tau deposition in preclinical AD, its increased utility in longitudinal imaging studies with large study populations is limited by recruitment and cost. In particular, making multiple visits to the scanning site will be difficult for participants living far away, and the high cost of injected radiotracers will limit the scalability of PET studies. In this project we propose using deep learning-based convolutional neural networks (CNNs) to enhance ultra-low-dose amyloid and tau PET for imaging AD. Our specific aims are (1) to validate the diagnostic value of the CNNs in actual ultra-low-dose amyloid and tau imaging sessions, with the injected dose as low as 1% of the original, and with actual ultra-low-dose data, to validate simulations for use in subsequent aims and future studies; (2) to apply the ultra-low-dose CNN to data collected on other PET systems and tracers, in order to demonstrate the CNN’s generalizability; and (3) to evaluate the value of deep learning-aided ultra-low-dose amyloid and tau PET for tracking cognitive decline in a preclinical AD population. The innovation of this work lies in using multimodal imaging in addition to advanced machine learning techniques to enable acquisition of diagnostic-level PET images at extremely low dose levels. Performing actual ultra-low-dose PET acquisitions is also highly novel in itself. The outcome of this proposal is removing the limiting factors to large-scale clinical longitudinal imaging, shortening acquisitions spanning multiple days and visits to several hours in one visit with a successive ultra-low-dose and full-dose dual-tracer scan protocol. Significant dose reduction can also be achieved, allowing for more frequent amyloid/tau PET scanning. This flexibility will not only increase the utility of PET, aid longitudinal studies in dementia, but enable future comprehensive imaging of multiple PET-based biomarkers as these tracers are being developed.