Deep-Learning Enhanced ASL MRI For Early AD Assessment

Summary Alzheimer’s Disease (AD) is a fatal disease without a cure. Understanding its neuromechanisms is a top research priority as it may provide regional targets for early intervention or for monitoring disease progression. Cerebral blood flow (CBF) is a fundamental physiological parameter indicating regional brain function as well as brain vascular integrity and has been increasingly adopted to evaluate AD related functional and neurovascular alterations. Arterial spin labeling (ASL) perfusion MRI is the only non-invasive non-radioactive approach to quantify regional CBF. It can be incorporated into multi-modal MRI protocols routinely acquired in patients with cognitive symptoms and can be repeated many times without risk, which is very appealing to AD and other basic, translational, and clinical research. Although ASL has shown a great promise in AD characterization and differential diagnosis, its full potential and impact are still hampered by three major limitations of current ASL techniques: the relatively low signal-to-noise ratio (SNR), low spatial resolution, and long scan time. Theoretically these problems are in direct conflict and there is no analytical way to address them in a single scheme. In this project, we aim to solve this analytically insurmountable challenge using the state-of-art artificial intelligence technique, deep machine learning (DL). We will use a divide-and-conquer strategy to address each of the three limitations separately to delineate the effects of different learning strategies and ASL data properties. Domain expert knowledge about ASL MRI will be used to guide DL algorithm design or to augment the learning and prediction process to achieve better performance and reliability. Using these methods, we expect to see a significant increase of SNR and spatial resolution, and a substantial reduction of ASL MRI data acquisition time. To show clinical importance, we will apply the proposed methods to existing ASL MRI data from subjects with clinical and preclinical AD. We expect to see increased sensitivity of ASL CBF for detecting the AD specific hypoperfusion patterns, monitoring disease progression, and detecting the CBF versus memory associations. The proposed suite of DL methods will not only benefit the AD research community but also have a wide range of basic, translational, and clinical neuroscience applications to which robust and high-resolution CBF mapping can contribute. To disseminate those benefits, we will release the code through our widely used toolbox for processing ASL data: ASLtbx. Feasibility of this highly novel but very important project is ensured by our decades of research experience in ASL MRI and AD and substantial preliminary data.