Longitudinal Analysis for Early Diagnosis of Alzheimer's Disease

Univ Of North Carolina Chapel Hill, Chapel Hill NC

Investigators

Linked publications & trials

Abstract

Project Summary/Abstract Alzheimer's disease (AD) is the most common cause of dementia in elderly people worldwide. Due to the fact that the longitudinal abnormality associated with AD can be detected in vivo, neuroimaging measures have been playing an increasingly important role in searching for biomarkers of AD that can be used for early diagnosis, progression monitoring, and therapy responses measurement. The goal of this project is to create a set of cutting-edge computational tools for identifying very early biomarkers of AD and alert AD progression via longitudinal image analysis. The key to achieve high sensitivity and specificity AD diagnosis of individuals is the stable longitudinal measurements. In light of this, the candidate will first develop a novel learning-based 4D (3D+t) landmark detector (Aim 1) to find the landmark locations at all time points simultaneously. The trajectories of landmark set form a compact 4D shape representation to achieve temporal consistency in registering longitudinal image sequence to the template (Aim 2). As a result, the consistency of low level features extracted from MR or PET images can be significantly improved. Since the subject-specific patterns of structural/functional changes, although more relevant to AD diagnosis of individuals, is very subtle compared to huge variation across subjects, the candidate will further develop a supervised deep neural network to learn the latent high-level spatial-temporal patterns by coupled stacked auto-encoder and temporal max pooling (Aim 2). The learned spatial-temporal morphological patterns consist of (1) cross- sectional features from MR and PET images, and (2) dynamic short/long term longitudinal patterns w.r.t. different number of historical time points. In a clinical setting, not all patients have a large and complete set of neuroimaging data, or an equal number of imaging scans. In order to eventually translate to clinic arena, a novel AD detector that uses a spatial-temporal hyper-graph learning framework (Aim 3) is proposed to not only provide high sensitivity and specificity in AD diagnosis but also solve the above difficulties. To evaluate the diagnostic value, we will apply our trained AD detector to ADNI dataset and patient data collected in UNC (Aim 3). Finally, we will package all our developed methods into a software package and release it freely to the neuroimaging community, to facilitate other AD-biomarker-exploration projects performed in other institutes.