Abstract Details

Title Retinal Color Fundus Photography and Deep Learning can Discriminate Alzheimer’s Disease

Topic Aging, Dementia, and Behavioral Neurology

Presentation(s) P8 - Poster Session 8 (5:30 PM-6:30 PM)

Poster/Presentation
Number 9-006

Objective

To develop a retinal imaging-based deep neural network that discriminates Alzheimer’s disease (AD) from cognitively intact individuals and identifies AD-specific retinal markers.

Background

AD has rising incidence and vast socio-economic impact. Retinopathy features predict AD-specific cognitive and neuroimaging measures. To overcome the subjectivity and inefficiency of color fundus photographs (CFP) manual analysis, we employed a neural network to classify AD and extract AD biomarkers from CFPs. We further enhanced our automated AD discrimination tool by comparing the accuracy of neural networks trained using various inputs.

Design/Methods A pre-trained U-Net-based architecture was used to segment 45-degree macula and optic disc-centered CFPs (118 AD; 129 cognitively intact). The segmented vessel results, and binary vessel segmentation results were inputted into the U-Net encoder for feature extraction. Weakly supervised localization and Gradient-weighted Class Activation Mapping were used to calculate the AD classification performance (AUC-ROC). We compared the accuracy of neural networks built on CFPs from single versus both eyes and employing vessel segmentation versus binary vessel segmentation. AD-specific biomarkers were extracted from the CFPs-derived heatmaps.

Results After data curation, we obtained 318 binary vessel segmentation images and 338 vessel segmentation images (116 AD subjects). The control group included 259 images (129 subjects). The overall vessel segmentation group accuracy was 95% (120 testing, 476 training images) and binary vessel segmentation group accuracy was 88% (116 testing, 460 training). The accuracy of the segmented and binary segmented images was similar in the right (97.7% and 90.7%) and left eye (86.4% and 97.7%). Vessel segmentation for both eyes had an accuracy of 97.7%. The generated heatmaps identified vascular changes in the retinal mid-periphery in AD.

Conclusions A deep neural network using CFPs and vessel segmentation exhibited 97.7% accuracy to discriminate AD from intact cognition. No significant differences were noted between single versus both eyes. The framework highlighted AD-specific retinal vascular changes.

Authors/Disclosures
Oana M. Dumitrascu, MD, FAAN (Mayo Clinic) PRESENTER	Dr. Dumitrascu has nothing to disclose.
Xin Li (Arizona State University)	Xin Li has received personal compensation for serving as an employee of Arizona State University.
Jacob M. Sobczak	Mr. Sobczak has nothing to disclose.
Wenhui Zhu (Arziona state university)	No disclosure on file
Richard J. Caselli, MD, FAAN (Mayo Clinic)	The institution of Dr. Caselli has received research support from NIA. The institution of Dr. Caselli has received research support from State of Arizona.
Bryan K. Woodruff, MD (Mayo Clinic)	The institution of Dr. Woodruff has received research support from National Institute on Aging.
Yalin Wang (Arizona State University)	Yalin Wang has nothing to disclose.

��ɫ��

��ɫ��