SUMMARY Cerebral blood flow (CBF) is a fundamental physiological parameter reflecting cerebrovascular integrity and brain function. Regional CBF alterations are well documented in Alzheimer's disease and related dementias (ADRD) and can contribute to ADRD prognosis, diagnosis, and differentiation. Arterial Spin Labeled (ASL) perfusion magnetic resonance imaging (MRI) is the only non-invasive method for imaging regional CBF and has been increasingly adopted in multi-site clinical ADRD research. However, ASL signals are prone to artifact from physiological noise and other nonidealities. Additionally, while multi-site ASL data can provide the statistical power needed to assess regional CBF changes in the ADRD continuum, they are often acquired with different ASL protocols from different scanner vendors and software versions leading to bias and variance in site specific CBF values. Both factors can obscure biological effects of interest, even in large datasets. The purpose of this R21/R33 study is to develop an ASL research framework for multisite clinical research studies aimed at I) developing automated indices to rate the quality of CBF maps that can be used reproducibly in place of more subjective visual inspection for quality control (QC) as well as for identifying the likely source of artifacts, and II) harmonizing multisite ASL data to remove site effects. The R21 phase will focus on method prototyping and validations using ASL data from a single vendor platform. Expert ratings for quality of ASL CBF maps and predetermined features of CBF maps will be used to train a deep learning (DL) algorithm that can automatically generate a quality index. The DL algorithm will be validated using synthetic data and expert reading results as well as cross-validations. Another DL algorithm will be used to identify the probable source of artifacts in input CBF maps. Prototype harmonization will be based on approaches previously used for structural MRI, but with ASL specific modifications involving mathematical manipulations to preserve more biological effects and to avoid under-harmonization. The R33 phase will substantially expand the methods developed in the R21 phase to be generalizable to multiple vendor platforms. A data driven DL network, as opposed to predetermined features used in the R21 phase, will provide both a composite and a voxel-wise quality index for the input ASL CBF images; the latter allowing regional CBF data to be retained instead of discarding the whole volume. We will also expand the capability of harmonization to cover data from unseen scanners. Finally, we will demonstrate the benefit of these methods in ADRD research by testing hypothesis that use of these research strategies will increase sensitivity for differentiating patients with mild cognitive impairment from cognitively normal older subjects. This novel project leverages our decades of expertise in ASL technologies and ASL-based ADRD research and our access to multi-site ASL...