Preclinical Alzheimer?s disease (the presymptomatic phase of Alzheimer?s disease) is characterized by pathophysiological changes without measurable cognitive decline and begins decades before the onset of cognitive symptoms. Preclinical Alzheimer?s disease research is in pressing need of new biomarker endpoints to enable disease monitoring before traditional cognitive endpoints are measurable. The overarching research objectives of this R03 Small Project Grant are to develop a super-resolution (SR) positron emission tomography (PET) imaging framework for tau (a pathophysiological hallmark of Alzheimer?s disease) and to assess the clinical utility of localized outcome measures obtained from SR PET images. Studies show that tau pathology in the medial temporal lobe is an important marker of cognitive decline in Alzheimer?s disease. Cohorts focused on preclinical Alzheimer?s now incorporate serialized 18F-flortaucipir PET scans for longitudinal tracking of tau accumulation in key anatomical regions-of-interest (ROIs). The quantitative accuracy of tau PET, however, is degraded by the limited spatial resolution capabilities of PET, which lead to inter-ROI spillover and partial volume effects. The problem is further compounded in studies spanning several decades, many of which were commenced on legacy scanners with even lower resolution capabilities than the current state of the art. Additionally, many longitudinal studies began on older scanners and later transitioned to newer models posing a multi-scanner data harmonization challenge. The proposed SR framework will perform a mapping from a low- resolution scanner?s image domain to a high-resolution scanner?s image domain and enable PET resolution recovery and data harmonization. Underlying the proposed framework is a neural network model that can be adversarially trained in self-supervised mode without requiring paired input/output image samples for training. This critical feature ensures practical clinical utility of the method as the need for paired low-resolution and high- resolution datasets from the same subject with similar tracer dose and scan settings is a major barrier for the clinical translatability of simpler supervised alternatives for SR. The proposed network, although trained using unpaired clinical data, receives guidance from an ancillary neural network separately pretrained using paired simulation datasets. For this purpose, we will synthesize paired low- and high-resolution images from a series of digital tau phantoms that will be created for this project. Training and validation of the self-supervised SR framework will be performed via secondary use of de-identified 18F-flortaucipir PET scans from the Harvard Aging Brain Study, a longitudinal cohort focused on preclinical Alzheimer?s disease. We will evaluate SR performance using a variety of image quality metrics. To assess the clinical utility of localized super-resolution measures, we will perform cross-sectional statistical power analyses that estimate sample sizes per arm needed to power clinical trials. Accurate localized measures of tau generated by this project could enable early diagnosis of Alzheimer?s disease and facilitate ongoing clinical trials by reducing sample sizes required for a given effect size.
The objective of this R03 Small Project Grant is to develop methods for generating high-resolution images of abnormal tau protein tangles, which are a hallmark of Alzheimer?s disease. This will be achieved by building a self-supervised super-resolution framework based on a deep neural network for positron emission tomography (PET) imaging of tau. The proposed imaging technique can facilitate early diagnosis and accurate monitoring of Alzheimer?s disease.
