Convolutional neural network to identify symptomatic Alzheimer's disease using multimodal retinal imaging

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) diagnosis relies on clinical assessments and neuroimaging, often late in disease progression.
• Retinal imaging offers a non-invasive window into neurodegenerative changes associated with AD.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for detecting symptomatic Alzheimer's disease (AD).
• To evaluate the efficacy of multimodal retinal imaging and patient data in predicting AD diagnosis.

Main Methods:

• A CNN was trained using multimodal retinal images including ganglion cell-inner plexiform layer (GC-IPL) thickness maps, optical coherence tomography angiography (OCTA) of the superficial capillary plexus (SCP), and ultra-widefield (UWF) fundus autofluorescence (FAF) and color images.
• The model incorporated quantitative data from OCT and OCTA, alongside patient data.
• The CNN's diagnostic performance was assessed on an independent test set using area under the receiving operating characteristic curve (AUC).

Main Results:

• The CNN achieved an AUC of 0.836 when using all multimodal retinal images, quantitative data, and patient data.
• GC-IPL maps alone provided the highest predictive value among single imaging modalities, with an AUC of 0.809.
• Models incorporating only retinal images performed comparably to those including additional patient and quantitative data.

Conclusions:

• A CNN effectively predicts symptomatic Alzheimer's disease using multimodal retinal imaging.
• GC-IPL maps represent a highly informative biomarker for AD detection via retinal imaging.
• AI-driven analysis of retinal images shows significant potential for non-invasive AD diagnosis.