Deep templates for longitudinal morphometry in Alzheimer's disease

/ABSTRACT Alzheimer’s disease (AD) affects one in eight Americans over 65 and is among the leading causes of death in the United States. Despite its prevalence, the accurate clinical diagnosis of AD remains challenging due to non-specific clinical tests and symptom overlap with other types of dementia. Except at top medical centers, the resulting misdiagnosis rates in excess of 15% relative to “gold-standard” neuropathologic analysis at autopsy negatively impact the efficacy of care, patient outcomes, and clinical trial efficiency. Although there are biomarkers that correlate with the neuropathology of AD, many are either invasive or largely inaccessible in routine clinical practice, such as PET imaging. One in-vivo marker that could be largely accessible is morphological metrics derived from non-invasive T1-weighted structural brain magnetic resonance imaging (MRI). These scans are already routinely acquired for individuals presenting with age-related cognitive decline, but integration barriers hamper the clinical accessibility of derived metrics. The fundamental goal of this project is to enable hospitals with limited access to AD biomarkers to achieve the diagnostic accuracy of top-performing institutions. We aim to improve the sensitivity and accessibility of structural MRI metrics by improving longitudinal, within-subject approaches that drive the morphometric analysis of MRI scans using a subject-specific reference image—a template—constructed from time points after rigid registration correcting for differential head positioning. We will build upon recent image-synthesis techniques and cutting- edge deep learning to develop anatomy-aware registration tools that generalize to medical imaging data across different MRI scanners and hospitals, without retraining or preprocessing. First, we will incorporate randomized, synthetic atrophy of AD-specific structures, ventricular expansion, and MRI distortions into a generative model. We will use this model to develop a deep neural network for rigid brain regis- tration that is robust to localized change and irrelevant intensity variations common in clinical imaging. Second, we will develop a deformable registration network that differentiates session-specific MRI distortions from disease effects. We will leverage AD-aware registration networks to construct individualized deformable templates across time to enhance the sensitivity of longitudinal morphometry, making it compatible with the clinical workflow. The project will result in deep-learning tools for longitudinal brain registration and template construction that are fast, accurate, and easy to use without machine-learning expertise or high-end computational resources. They will increase the power to detect disease effects with fewer subjects and dramatically reduced runtimes. Project Number: 1R21AG093037-01A1 | Fiscal Year: 2026 | NIH Institute/Center: National Institute on Aging (NIA) | Principal Investigator: Malte Hoffmann | Institution: MASSACHUSETTS GENERAL HOSPITAL, BOSTON, MA | Award Amount: $660,000 | Activity Code: R21 | Study Section: Imaging Technology for Neuroscience Study Section[ITN] View on NIH RePORTER: https://reporter.nih.gov/project-details/11301393