Integration of artificial intelligence in medical image analysis has already had a large impact on e.g. diagnostics of various diseases. Additionally, it offers a potentially transformative understanding of pathology. In this project our aim is to take advantage of these benefits in the context of Alzheimer's Disease (AD). One hallmark of AD is abnormal accumulation of hyperphosphorylated tau proteins forming neurofibrillary tangles in the brain. The spread of this pathology often mirrors the progression of clinical symptoms, making it crucial for understanding the disease in general and vital as a target for therapeutic interventions and progression monitoring in particular. Recent clinical trials have shown that AD patients with low to intermediate tau loads appear more responsive to novel anti-amyloid AD therapies (e.g. Lecanemab and Donanemab), nominating this subpopulation as prime candidates for such trials. However, the gold standard for measuring tau load, tau-PET, is not widely accessible, which necessitates alternative methods. This research proposes an AI-driven approach to estimate tau-PET images using more accessible patient data, including (but not limited to) plasma biomarkers, multimodal MRI sequences, clinical information and demographics. A supervised deep learning approach will be used to synthesise tau-PET images from more readily-available clinical tools, based on data from an aggregated multi-site tau-PET imaging dataset of unprecedented size (n=15,000). Our preliminary data using machine learning and much smaller sample sizes (n~2000) suggests this approach is feasible and promising. One region specifically implicated in tau accumulation is the medial temporal lobe, encompassing the hippocampus, amygdala and parahippocampal regions, a pivotal brain area for diverse cognitive processes. Accurate segmentation/parcellation of these subregions is essential for deciphering its role in both cognition and disease. Despite its importance, there are no automated methods for segmentation, while manual segmentation suffers from inter- intra-rater variability and is extremely labour intensive. We propose using an unsupervised deep learning approach for delineating the MTL using a large, rare and extremely high-resolution histology dataset. By basing our model on Variational Autoencoders (VAEs), we aim to capture “hidden” properties of MTL slices to delineate boundaries between regions. However, to limit learning of irrelevant features the full model architecture will be more complex and therefore computationally intensive. Efforts will be made to integrate our histology data (which is co-registered to postmortem MRI) with our in-vivo PET data for further analysis. These projects will uniquely use AI to gain insight and potentially clinical utility from large and rare human datasets, and have both recently been funded by research grants and industry partnerships. However, these projects can only be executed using extensive computational and memory resources.