Purpose: To develop and assess an automatic procedure for classifying and staging glaucomatous vascular damage based on optical coherence tomography angiography (OCTA) imaging.
Methods: OCTA scans (Zeiss Cirrus 5000 HD-OCT) from a random eye of 39 healthy subjects and 82 glaucoma patients were used to develop a new classification algorithm based on multilayer and multisector information. The averaged circumpapillary retinal nerve fiber layer (RNFL) thickness was also collected. Three models, support vector machine (SVM), random forest (RF), and gradient boosting (xGB), were developed and optimized for classifying between healthy and glaucoma patients, primary open-angle glaucoma (POAG) and normal-tension glaucoma (NTG), and glaucoma severity groups.
Results: All the models, the SVM (area under the receiver operating characteristic [AUROC] 0.89 ± 0.06), the RF (AUROC 0.86 ± 0.06), and the xGB (AUROC 0.85 ± 0.07), with 26, 22, and 29 vascular features obtained after feature selection, respectively, presented a similar performance to the RNFL thickness (AUROC 0.85 0.06) in classifying healthy and glaucoma patients. The superficial vascular plexus was the most informative layer with the infero temporal sector as the most discriminative region of interest. No significant differentiation was obtained in discriminating the POAG from the NTG group. The xGB model, after feature selection, presented the best performance in classifying the severity groups (AUROC 0.76 0.06), outperforming the RNFL (AUROC 0.67 0.06).
Conclusions: OCTA multilayer and multisector information has similar performance to RNFL for glaucoma diagnosis, but it has an added value for glaucoma severity classification, showing promising results for staging glaucoma progression.
Translational relevance: OCTA, in its current stage, has the potential to be used in clinical practice as a complementary imaging technique in glaucoma management.
Keywords: OCT angiography; glaucoma; machine learning; microvascular density.
Copyright 2020 The Authors.