Revolutionizing Alzheimer’s Detection: The Intersection of Color Fundus Photography and Deep Learning Techniques

Objective:

This report discusses the creation and efficacy of 2 unique deep learning models specifically trained on retinal color fundus images to identify Alzheimer disease (AD).

Patients and methods:

Two separate datasets (UK Biobank and our tertiary educational institution) containing high-quality retinal images from individuals diagnosed with AD and control subjects were utilized to develop the 2 deep learning models, between April 1, 2021, and January 30, 2024. ADVAS employs a U-Net-based framework that incorporates retinal vessel segmentation. ADRET uses a self-supervised learning convolutional neural network modeled after bidirectional encoder representations from transformers, which has been pretrained on a substantial dataset of retinal color images obtained from the UK Biobank. The models’ ability to differentiate AD from non-AD was evaluated by calculating mean accuracy, sensitivity, specificity, and receiving operating curves. The attention heatmaps produced were examined for distinctive characteristics.

Results:

The self-supervised ADRET model demonstrated greater accuracy compared to ADVAS, in both UK Biobank (98.27% vs 77.20%; P<.001) and our institutional testing datasets (98.90% vs 94.17%; P=.04). No significant differences were observed between the original and binary vessel segmentation, nor between models using both eyes versus single-eye models. The attention heatmaps derived from patients with AD indicated areas surrounding minor vascular branches as the most pertinent to the model’s decision-making process.

Conclusion:

A self-supervised convolutional neural network, modeled on bidirectional encoder representations from transformers and pretrained on a large cohort of retinal color photographs, can effectively screen symptomatic AD with high precision, surpassing U-Net-pretrained models. To be applicable in clinical settings, this approach requires additional validation across larger and more diverse populations, as well as integrated methodologies to unify fundus images and mitigate imaging-related noise.