Abstract
Purpose :
Migraines are complex neurovascular disorders affecting more than a billion people worldwide, typically diagnosed through patient history and subjective symptoms. Evidence suggests that they may also be associated with objective, widespread neurovascular degeneration. This project assessed the diagnostic accuracy of using convolutional neural networks (CNNs) to predict migraines from retinal imaging.
Methods :
Single eyes from 594 individuals were consecutively sampled across all visits from the Centre for Eye Health, divided into 68 single eyes with and 526 without self-reported history of migraines. Participants had otherwise normal, healthy maculae and optic nerves. Colour fundus photography (CFP) images (partitioned as type 1 [macula and optic nerve head (ONH)] and type 2 [ONH]) were each concatenated with optical coherence tomography (OCT) default ONH measures, then inputted through three popular CNNs – Vgg-16, ResNet-50, and Inceptionv3 (Figure 1A-C). The primary outcome was diagnostic performance for classifying migraines/no migraines. Class Activation Maps (CAMs) were also generated to visualise regions of interest.
Results :
Diagnostic accuracy (AUC) was good (0.84 [0.8, 0.88] to 0.87 [0.84, 0.91]) and non-different (all P > 0.2) between all three CNNs using type 1 CFP [macula and ONH] data (Figure 1D). Using type 2 CFP [ONH] data, diagnostic accuracy was reduced (all P < 0.05) and poor-to-fair (0.69 [0.62, 0.77] to 0.74 [0.67, 0.81]; Figure 1E). CAMs highlighted that the paravascular arcades were regions of interest (Figure 2).
Conclusions :
From a sample of 594 individuals, CNN analyses of retinal imaging can provide objective and good diagnosis of self-reported history of migraines. Further study with expanded classification of migraines and OCT imaging data is underway.
This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.