Correction of Partial Volume Effects in PET for Alzheimer's Disease Using Unsupervised Deep Learning

点击翻译按钮获取中文摘要

Correction of Partial Volume Effects in PET for Alzheimer's Disease Using Unsupervised Deep Learning Abstract Alzheimer's disease (AD), characterized by memory loss and cognitive impairments, affects approximately 5.8 million people in the United States. It is a burden for patients, families and the healthcare system. Post-mortem studies reveal that hyperphosphorylated Tau protein aggregates are closely related with AD. With the developments of selective Tau tracers, Tau distributions can now be characterized in vivo through Positron emission tomography (PET) imaging. According to the Braak staging, Tau deposits start from the transentorhinal region of the temporal lobe at preclinical AD, then spreading to other cortex regions at late stages. Regional Tau distribution pattern captured through PET Tau imaging can thus provide important staging information, which is vital for early disease diagnosis, progression tracking and treatment monitoring. However, accurate quantification of thin cortex uptake is difficult due to partial volume effects (PVEs) in PET imaging. Besides, for the second- generation Tau tracer, 18F-MK-6240, apart from higher uptakes observed in neocortical and medial temporal brain regions for AD patients, nonspecific uptakes are also observed in areas such as the meninges. Given the thin nature of the cortical ribbon and its close proximity to the meninges, quantitative accuracy and detection precision of 18F-MK-6240 distributions are significantly impacted, which in turn precludes finer localization of early tau accumulation associated with preclinical and prodromal AD. This grant application proposes a novel partial volume correction (PVC) method through unsupervised deep learning for Tau imaging. This new framework does not need high-quality training labels and the network is specifically trained for each subject, with the training objective function formulated based on the Poisson distribution assumption of the sinogram data. To further boost the performance, the transfer learning and the kernel learning are integrated into this PVC framework. The three specific aims of this exploratory proposal are (1) to develop a PET PVC framework based on unsupervised deep learning, (2) to validate the proposed PVC framework using phantom studies and (3) to apply the proposed PVC framework to 18F-MK-6240 imaging datasets. We expect that the integrated outcome of the specific aims will be an efficient, practical and robust PVC method that can better resolve the Tau distribution patterns for the early diagnosis of AD.