The most significant impact of this project is to propose for the first time a novel generative adversarial network (GAN), as one kind of deep learning architecture, to automatically generate synthetic PET images reflecting tau deposition, from brain DTI images. If successful, this framework will become the most state-of-the-art approach to simulate the stereotypical pattern of intracerebral tau accumulation and distribution in vivo.

Synthetic tau-PET images via DTI, possessing overwhelming superiority in radiation-free, non-invasiveness and cost-effectiveness, will potentially serve as one of alternative modalities of PET in detecting tau-load and probably outperform PET on accessibility, generalizability, and availability in future, making it much more attractive in clinical application. A big conceptual shift may occur preferring a fire-new tau-PET simulated via DTI.

The DTI data-driven deep learning framework to be created in this project will constitute an accurate, robust, clinically applicable and explainable tool to efficiently categorize the subjects into tau-burden positive and tau-burden negative cases, which will undoubtedly contribute to both clinical and research activities.

Inclusion Criteria:

• With the age of 55 years and above
• With brain MRI taken within ±6 months from the date of clinically confirmed diagnosis of AD, MCI or normal cognition.

Exclusion Criteria:

• AD with mixed dementia
• Non-AD dementia
• History of severe traumatic brain injury, severe depression, stroke, brain tumors, and incident major systemic illness

The data will be collected from a public database, i.e, Alzheimer's Disease Nueroimaging Initiative (ADNI). According to their recruitment protocol at http://adni.loni.usc.edu/methods/documents/, the study enrolls men and women aged 55-90 years across MCI and mild AD dementia participant groups, and control normal aged 65 years with exceptions granted for minority participants.