PROJECT SUMMARY/ABSTRACT Amyloid and tau are two hallmark pathologies of Alzheimer's disease (AD), which start to accumulate in the brain long before clinical symptom onset. With the recent development of positron emission tomography (PET) tracers, in vivo measurements of these two pathologies became possible and play important roles in improving our understanding of AD pathogenesis and progression, and provide critical diagnostic and prognostic information that can facilitate the treatment development effort now and improve patient outcome when effective treatments become available. However, multiple PET tracers exist for both amyloid and tau imaging and generate images that have different visual appearances and quantitative characteristics. These differences lead to difficulties in comparing different studies, leveraging the large volume of data that have been collected to date, and defining a common criterion for positivity. In this project, we will focus on the development of novel techniques that can harmonize PET imaging data and generate highly compatible imaging derived measurements. We argue that by adopting advanced statistical and deep learning techniques and developing novel quantification strategies that take full use of the available information in imaging data, successful harmonization can be achieved via generalizable approaches. Several lines of research support this argument: 1) we recently demonstrated that using partial volume correction technique to reduce white matter signal contamination, improved agreements in amyloid burden measurements derived from PIB and florbetapir can be obtained; 2) in our preliminary analysis, using machine learning technique, we can substantially improve between-tracer agreements in global amyloid burden; 3) recent advancement in deep learning research demonstrated the ability to “impute” one imaging modality from another by taking advantage of the inherent information in large imaging datasets, and our preliminary results demonstrated its potential in image harmonization tasks. In this project, we will continue this research and develop a set of techniques that can be used to harmonize amyloid imaging data from different tracers and extent this approach for tau PET harmonization. Our specific aims are 1) to develop deep learning models that can generate imputed amyloid PET images from one tracer based on the images from another tracer; 2) to develop statistical learning approaches that can generate harmonized amyloid burden measurements; 3) to acquire head-to-head comparison tau PET imaging data and examine statistical and deep learning approaches to the harmonization of Tau PET imaging data.