Purchase this article with an account.
Julia Schottenhamml, Stefan B Ploner, Lennart Husvogt, Sophia Mardin, Bettina Hohberger, Christian Y Mardin, Andreas K Maier; Deep Learning Glaucoma Classification in enface OCTA scans of the human macular region. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1004.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Glaucoma is one of the leading causes of irreversible vision loss. Several studies have shown a link between vascular damage and glaucoma on optical coherence tomography angiography (OCTA) images. These studies were mostly conducted in the region of the optic nerve head (ONH) or on larger 6x6 mm macula scans. Moreover, all these studies use handcrafted features for the glaucoma classification. We therefore propose a fully automatic classification algorithm on 3x3 mm macula scans based on convolutional neural networks (CNN) that are able to learn features from the images themselves.
Whole retina projection OCTA images (Spectralis OCT II, Heidelberg Engineering, Heidelberg) of 75 eyes of 75 healthy persons (h) and 184 eyes of 125 glaucoma patients (g) were retrospectively identified from the Erlangen glaucoma registry. They were divided into training set (h: 45 eyes/45 patients, g: 110 eyes/76 patients), validation set (h: 15 eyes/15 patients, g: 37 eyes/24 patients) and test set (h: 15 eyes/15 patients, g: 37 eyes/25 patients). Eyes of one person only belong to a single set. Different CNNs were trained and the one performing best (according to the AUROC) on the validation set was chosen for evaluation. For comparison, handcrafted features that were compatible with the available data (only a single enface OCTA image available) were selected. We chose the commonly used vessel density (VD) and a combination of global and local features proposed by Ong et al. (Ong). We then trained a support vector machine (SVM) on each of these traditional feature sets and again chose the best performing one on the validation set.
The ROC curves, AUROC values and confusion matrices can be found in Figure 1 and the accuracies in Figure 2 for the different methods on the test set.
We were able to outperform other handcrafted features on 3x3 mm macula scans and achieved results on par with other algorithms proposed for different ocular regions. These other studies often have shown that their features work better on projections of a single vascular plexus than on a whole retina projection. Therefore, the next step would be to apply the deep learning pipeline presented in this paper to the different retinal plexuses separately.
This is a 2021 ARVO Annual Meeting abstract.
ROC curves, AUROC values and confusion matrices for the different methods.
Accuracies for the different methods for the whole dataset and each class separately.
This PDF is available to Subscribers Only