Abstract
In a prior study we demonstrated the strong performance of convolutional neural networks (CNNs) in distinguishing healthy from glaucomatous eyes and subsequent analysis suggested that CNNs rely on the vessel density (VD) of the perifoveal area for making their predictions. Those findings are of particular interest because current 3 × 3 mm macular optical coherence tomography angiography (OCTA) analysis mostly focuses on the parafoveal area, excluding the perifoveal region. The aim of this study is to further investigate this observation and evaluate whether perifoveal VD offers greater diagnostic accuracy in differentiating healthy from glaucomatous eyes than parafoveal VD. For conducting this study, 3 × 3 mm macular OCTA scans where acquired by the Spectralis OCT II device of 352 eyes, consisting of 198 glaucomatous and 154 healthy eyes. The different layers of the retina were automatically segmented by the manufacturer's software and sectioned into the superficial vascular plexus (SVP), intermediate capillary plexus (ICP) and deep capillary plexus (DCP). Support vector machines (SVMs) were then trained on perifoveal and parafoveal VD for each plexus, whereby parafoveal VD was computed on 12 sectors within a 1.45 mm radius centered at the fovea, while parafoveal VD was trained on 4 quadrants covering the remaining area of the scan. 10-fold cross-validation was used to select the SVMs with the highest mean area under the receiver operating characteristic curve (AUROC) and evaluated on a held-out test set. Differences in AUROC between methods were pairwise statistically tested with Bonferroni correction. It was found that perifoveal VD outperformed parafoveal VD across all plexuses, with the highest improvement observed in the SVP (AUROC perifoveal 0.943 vs. parafoveal 0.897, p = 0.025), followed by the ICP (AUROC perifoveal 0.872 vs. parafoveal 0.851) and DCP (AUROC perifoveal 0.795 vs. parafoveal 0.786). Our results demonstrate that perifoveal VD in 3 × 3 mm macular OCTA scans offers a higher diagnostic performance in distinguishing healthy from glaucomatous eyes than parafoveal VD. These findings suggest that incorporating the perifoveal region into macular OCTA analysis could yield additional information on glaucoma, especially when focusing on the SVP. Furthermore, our findings propose to reconsider the value of 3 × 3 mm macular scans in VD analysis, whose utility appears to be underestimated.