Limits of Fundus-Based Deep Learning for Alzheimer’s Disease Prediction: Anatomical Ablation and Clinical Benchmarking in UK Biobank

To evaluate retinal fundus imaging for predicting Alzheimer’s Disease and Related Dementias (ADRD) using a pretrained retinal foundation model (RETFound), including assessment of model performance from baseline models, retinal region contributions, and added value beyond demographic risk factors. 

• The retina reflects brain neurobiology through neuromodulator release while retinal ganglion cell-inner plexiform layer (GC-IPL) and retinal nerve fiber layer (RNFL) thinning contributes to blood vessel inflammation and impaired exchange of nutrients, and wastes linking retinal changes to systemic neurodegeneration and cognitive decline.
• Instead of using costly and labor intensive traditional MRI scans for Alzheimer’s Disease diagnosis, fundus images and OCT scans in the eye can potentially provide a cheap and efficient alternative for providing accessible and equal care among people and communities.
• The retina is a non-invasive biomarker for ADRD. Optical coherence tomography (OCT) studies show thin RNFL and GC-IPL are associated with higher dementia risk and poorer cognition in large cohorts (e.g. UK Biobank), establishing retinal structure as a marker of cognitive decline.
• Advances in machine learning (ML) have shifted ADRD research from traditional models to deep learning. Neural networks extract complex image features, and models using retinal and vascular biomarkers achieve strong performance (average AUC to ~0.85)
• RETFound is a pretrained retinal foundation model trained on 1.6M+ retinal images. Its self-supervised framework enables efficient fine-tuning. We apply RETFound to UK Biobank fundus images and clinical variables to evaluate ADRD prediction against baseline models, trained using nature images, and custom sigmoidal multimodal fusion model.
• OCT Logistic Regression (LR) acts as clinical baseline using age, sex and human derived GC-IPL and RNFL numerical values from fundus images. Human derived variables are very labor intensive and are not ideal for mass data.