Color Fundus Photography and Deep Learning Applications in Alzheimer Disease.

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Alzheimer disease (AD) diagnosis relies on clinical assessment and biomarkers.
• Retinal imaging offers a non-invasive window into systemic health, potentially revealing AD-related changes.
• Deep learning (DL) models are increasingly explored for medical image analysis.

Purpose of the Study:

• To develop and evaluate two distinct DL models for classifying Alzheimer disease (AD) using only retinal color fundus photographs.
• To compare the performance of a U-Net-based model (ADVAS) with a self-supervised transformer-based model (ADRET).

Main Methods:

• Two independent datasets (UK Biobank and institutional) of retinal photographs from AD patients and controls were used.
• ADVAS: U-Net architecture with retinal vessel segmentation.
• ADRET: Bidirectional encoder representations from transformers (BERT) style self-supervised learning CNN, pre-trained on UK Biobank data.
• Performance metrics included accuracy, sensitivity, specificity, and receiver operating characteristic curves.

Main Results:

• The self-supervised ADRET model demonstrated superior accuracy compared to ADVAS in both datasets (UK Biobank: 98.27% vs 77.20%; institutional: 98.90% vs 94.17%).
• Attention heatmaps from AD patients highlighted perivascular regions as critical for model decision-making.
• No significant differences were observed based on vessel segmentation type or eye laterality.

Conclusions:

• A self-supervised BERT-style CNN trained on retinal images can accurately screen for symptomatic Alzheimer disease.
• This approach shows higher accuracy than U-Net-based models for AD detection via retinal imaging.
• Further validation in diverse populations and harmonization of imaging techniques are necessary for clinical translation.