Purchase this article with an account.
Clayton Ellis Wisely, Dong Wang, Ricardo Henao, Dilraj S Grewal, Stephen Paul Yoon, Bryce Polascik, Atalie C. Thompson, James R Burke, Lawrence Carin, Sharon Fekrat; Deep learning algorithm for diagnosis of Alzheimer’s disease using multimodal retinal imaging. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1461.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a deep learning algorithm to diagnose Alzhiemer’s disease (AD) using multimodal retinal images including OCT angiography (OCTA) and ultra-widefield (UWF) scanning laser ophthalmoscopy (SLO) color and autofluorescence (FAF) images.
AD and control subjects aged ≥50 were enrolled in this cross-sectional study (Clinicaltrials.gov NCT03233646). AD was diagnosed by an experienced neurologist. Controls were cognitively intact healthy volunteers. Exclusion criteria included history of non-AD dementia, diabetes, uncontrolled hypertension, demyelinating disorders, glaucoma, macular degeneration, other vitreoretinal pathology, or corrected visual acuity <20/40. Subjects were imaged with Zeiss Cirrus HD-5000 AngioPlex OCTA. Non-mydriatic color and FAF images were captured using Optos California UWF fundus SLO. Images with poor quality, low resolution, poor centration, or motion artifacts were excluded. A deep learning binary classifier model was built to distinguish between AD and control subjects. Our classifier was built on top of the first five layers from a ResNet-18 network, pre-trained on a public image dataset, CIFAR-10. We adopted batch normalization after each convolution and before rectified linear unit (ReLU) activation.
222 eyes of 117 controls and 63 eyes of 36 AD subjects were used to develop the algorithm. After exclusions UWF color and FAF images were included for both eyes of 105 control and 27 AD subjects, with partial sets included for others. 3x3 mm OCTA images centered on the fovea were included for 125 control eyes and 28 AD eyes. The model was trained and tested with 9-fold cross-validation. In each fold the training set included 198 control and 57 AD eyes, and test set included 24 control and 6 AD eyes. With OCTA and UWF color and FAF images, the model predicted the probability of AD, P(AD), for each AD or control subject (Figure 1). The area under the receiver operating characteristic curve (AUC) and mean accuracy describe the model’s precision-recall and trade-off. AUC (Figure 2) was 0.74 (SD = 0.014) and accuracy was 0.79 (SD = 0.012).
Our deep learning algorithm demonstrates the ability to use multimodal retinal imaging for automated AD diagnosis with nearly 80% accuracy. Larger databases containing additional imaging modalities may improve diagnostic accuracy of this algorithm and enable clinical utility for early AD diagnosis.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only