Deep Learning-Based Retinal Imaging Screening Tool for Preclinical Alzheimer’s Disease - PROJECT SUMMARY/ABSTRACT Alzheimer's disease (AD) imposes a substantial burden on the aging population, with projected costs surpassing $1 trillion by 2050. Mounting evidence supports the notion that targeting AD in its preclinical phase holds the greatest potential for effective interventions and substantial social benefits. However, identifying individuals at risk for preclinical AD in routine clinical practice or for inclusion in trials currently relies on diagnostic modalities with limited availability and affordability. While PET, cerebrospinal fluid analysis, structural MRI, and emerging blood-based biomarkers (BBBM) for AD pathology (amyloid (A), tau (T), neurodegeneration (N), and inflammation (I)) have advanced the field, they remain constrained by high cost, invasiveness, or the need for further validation. An easily accessible extension of the central nervous system, the retina, exhibits amyloid and tau deposition, vascular alterations, inflammation, and other neurodegenerative changes that mirror brain pathology. Non-mydriatic retinal color fundus photography (CFP) is a low-cost, non-invasive modality that enables repeated, large-scale imaging. Despite its promise, CFP analysis in preclinical AD remains underexplored, partly due to challenges in image quality and the lack of validated automated diagnostic tools. Meanwhile, deep learning (DL) enabled automated detection of retinal CFP biomarkers for conditions such as diabetic retinopathy, leading to FDA-cleared algorithms deployed in primary care settings. This project will develop and validate a suite of DL methods to enhance CFP image quality, automate retinal AD biomarker identification, and enable scalable, cost-effective preclinical AD risk assessment. Specifically, we will: (1) develop an unsupervised method using optimal transport-guided generative adversarial networks with domain adaptation to enhance low quality CFPs; (2) build nn-MobileNet++, a lightweight DL model combining attention, dynamic convolution, and hybrid modules for AD retinal biomarker detection; and (3) evaluate predictive performance of CFP alone, BBBM alone, and integrated CFP+BBBM models to predict PET- and BBBM-defined central nervous system (CNS) amyloid positivity. We will also explore DL-based retinal age gap and cognitive prediction as novel AD biomarkers. We will leverage three large-scale datasets: UK Biobank, Canadian Longitudinal Study on Aging, and Mayo AD databases. The Mayo Preclinical AD cohort, a cohort of 100 preclinical AD patients (cognitively unimpaired (CU), amyloid PET positive) and their age- and gender-matched 230 controls (CU, amyloid PET negative), will serve as a primary testbed, leveraging retinal imaging data alongside genetic, brain imaging, BBBM, and clinical information. External validation will be performed in the rural, point-of-care MindCrowd MobileLab cohort (n > 1,000 with retinal and brain imaging, genetics, cognitive testing, BBBM). By integrating cutting-edge DL models with rich multimodal datasets, this project aims to create a robust, accessible platform for non-invasive preclinical AD detection, facilitating earlier diagnosis, enabling large-scale trial recruitment, and ultimately helping reduce clinical trial costs and accelerating the development of effective AD therapies.
