Automated Retinal Imaging Analysis for Alzheimer’S Disease Screening

Screening Alzheimer's Disease (AD) with retinal vascular imaging is an innovative research area. Even though there is no standardized canonical biomarker to diagnose AD in an early stage, specific retinal vascular abnormalities have shown potential linkage to early AD diagnosis and monitoring. Quantitative retinal vascular feature extraction however, faces the challenges of laboriousness and subjectiveness. Building on the concept of weakly supervised learning, we propose a deep learning-based framework to classify AD versus Normal Control (NC) and automatically localize potential AD retinal vascular imaging biomarkers. The proposed method was evaluated on a real-world AD dataset from Mayo Clinic and NC from the Eyepacs dataset, and achieved an AUC-ROC of 0.938. The medium-size or distal retinal vascular branches were identified as key features that distinguish AD from NC.