Detection of Glaucoma Progression on Longitudinal Series of Macular Optical Coherence Tomography Angiography Maps with a Deep Learning Model

Abstract

Purpose : To design a deep learning (DL) model for detection of glaucoma progression with longitudinal series of macular optical coherence tomography angiography (OCTA) images.

Methods : Two-hundred and two eyes of 134 patients with open angle glaucoma were included. Eligible eyes were required to have 4 visits and 2 years of follow up of OCTA. Glaucoma progression was defined as having 24-2 visual field (VF) mean deviation (MD) rates of < 0 and P value < 0.05 during the follow-up. The data were split with 80% in the training dataset and 20% in the testing dataset. The baseline and final macular OCTA images were aligned according to the center of fovea avascular zone automatically, by checking the correlation of the two images on vector space (Figure 1). To improve the generalizability of the model, data augmentation such as random cropping and resizing were used to increase the training data. For the classification model, a customized convolutional neural network (CNN) with a multi-layer perceptron (MLP) classifier was used. The baseline and final images for each eye were concatenated along the color channel and then sent into the CNN for classification. The cross entropy was used as the loss function for training and a weight of 0.13 was assigned to the non-progressing samples and 0.87 for progressing samples. The performance of the model was evaluated using the area under receiver operating characteristics (AUC), sensitivity, specificity and accuracy. As a comparison method, a logistic regression model was performed to evaluate the performance of whole image vessel density (wiVD) loss on detection of glaucoma progression.

Results : The average (range) follow-up time was 3.5 (2.4-5.5) years and average (standard deviation) baseline VF MD was –3.4 (±5.0) dB. Twenty-eight (14%) eyes demonstrated glaucoma progression. The AUC for detection of progression from macular OCTA images of the DL model was 0.81and for the logistic regression model 0.66 (Figure 2). The sensitivity, specificity and accuracy of the DL model on the testing dataset was 67%, 83% and 80%, respectively.

Conclusions : The optimized DL model detected glaucoma progression based on longitudinal macular OCTA images with high performance. Implementation of this DL model, after external validation, could enhance detection of glaucoma progression.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.