In order to evaluate the different acquisition areas we will subsequently describe the data used in this study, followed by the design of our experiments.DataIn total, 219 eyes from 110 subjects (117 eyes from 59 patients with glaucoma and 102 eyes from 51 healthy control cohorts) were identified from the Erlanger Glaucoma Registry (Erlangen Glaucoma Registry, ISSN 2191-5008, CS-2011; NTC00494923). All subjects received a standardized ophthalmological examination including automated visual field testing, fundus photography and measurement of intraocular pressure (IOP) by Goldmann tonometry.The patients in each group were selected based on the following inclusion criteria: the control cohort was defined as eyes showing no systemic disease with ophthalmological involvement or ophthalmological dysfunction neither having had any ophthalmic surgery. Glaucoma suspects were defined as having a normal visual field, and an elevated IOP (above 21 mmHg, ocular hypertension, OHT) or showed additive glaucomatous optic disc damage classified by Jonas et al.25 (preperimetric glaucoma). Glaucoma patients showed perimetric field defects and alterations of the optic nerve head according to Jonas et al.25. This group was further subdivided into patients having an elevated IOP (above 21 mmHg, open angle glaucoma, OAG)) and those not having an increased IOP (normal tension glaucoma). All glaucoma patients enroled in this study had bilateral glaucoma and showed signs of the disease in both eyes. In the glaucoma cohort are 20 eyes with OHT, 18 eyes with preperimetric glaucoma and 32 eyes of patients with normal tension glaucoma. The open angle glaucoma group consists of 35 eyes with primary open angle glaucoma and 12 eyes with secundary open angle glaucoma.Exclusion criteria were an age below 18 years and any further eye disorders and/or systematic disorders with ocular involvement at the time of enrolement. All patients enrolled in this study were caucasian. The mean and standard deviation of the age distribution and refraction sphere for the healthy controls and the glaucoma patients and the perimetric data (mean deviation (MD) and square root of loss variance (sLV)) for the glaucoma patients is given in Table 1.Table 1 Mean and standard deviation of the age distribution and refraction sphere for the healthy controls and the glaucoma patients and the perimetric data (mean deviation (MD) and square root of loss variance (sLV)) for the glaucoma patients.OCT and OCTA imaging was done using a SOLIX fullrange device (Optovue Inc, Fremont, California, USA). Images were acquired with a lateral resolution of 15 \(\upmu \textrm{m}\) and an axial resolution of 5 \(\upmu \textrm{m}\) in tissue. Three different acquisition areas were recorded:
\(\sim \) 3 \(\times \) 3 \(\textrm{mm}\) scan centered at the fovea consisting of 400 A-scans per B-scan and 400 consecutive B-scans
\(\sim \) 6.4 \(\times \) 6.4 \(\textrm{mm}\) scan centered at the fovea consisting of 512 A-scans per B-scan and 512 consecutive B-scans
\(\sim \) 6 \(\times \) 6 \(\textrm{mm}\) scan centered at the ONH consisting of 512 A-scans per B-scan and 512 consecutive B-scans
For each OCTA scan acquisition area the different retinal plexus projections were defined according to the approach presented by Campbell et al.26 based on pixel offsets of the retinal layer boundary segmentation of internal limiting membrane (ILM), nerve fibre layer (NFL), inner nuclear layer (INL) and outer plexiform layer (OPL) of the manufacturer’s software:
NFLVP (nerve fibre layer vascular plexus): [ILM, NFL]
SVP (superficial vascular plexus): [NFL, IPL-9]
ICP (intermediate capillary plexus): [IPL-9, IPL+25]
DCP (deep capillary plexus): [IPL+25, OPL+9]
Retina (SVP + ICP + DCP): [ILM, OPL+9]
SVP, ICP, DCP and Retina projections were computed by the manufacturer’s software for all acquisition areas. The NFL plexus was only generated for the 6 \(\times \) 6 \(\textrm{mm}\) ONH scans, since this capillary plexus is barely visible in the macular region. A depiction of the layer boundaries used for the plexus projection is provided in Fig. 1, while a visual impression of the appearance of the different plexuses in the different acquisition areas is provided in Fig. 2.Figure 1Visual impression of the projection boundaries of the different plexuses in the different acquisition areas from a right eye of a healthy control. The upper and lower boundaries are indicated in red or green.Figure 2Visual impression of the appearance of the different plexuses in the different acquisition areas from a right eye of a healthy control. The scale bars in the lower left corners indicate 1\(\textrm{mm}\).When acquiring OCTA measurements, the SOLIX fullrange device (Optovue Inc, Fremont, California, USA) also gives access to the co-registered OCT measurements. Consequently, commonly used biomarkers for glaucoma diagnosis were extracted from the manufacturer’s software as well. One set of biomarkers were the ganglion cell complex (GCC) thickness. Here, 8 values were obtained from the 6.4 \(\times \) 6.4 \(\textrm{mm}\) macula scans from the inner and outer ring of the ETDRS grid. The second set of biomarkers were the retinal nerve fibre layer (RNFL) thickness. Again, 8 values were extracted from the 6 \(\times \) 6 \(\textrm{mm}\) ONH scans in a ring around the ONH. A visual example for the RNFL and GCC thickness parameter is fiven in Fig. 3. The third set of biomarkers contained three disc parameters, namely the disc area, the rim area and the cup area. The disc area is defined as the area inside the disc margin which is automatically detected based at the Bruch’s Membrane Opening (BMO). The rim and the cup are then measured within the BMO plane: the area above the BMO plane is rim, while the portion below the plane is cup. An overview of these parameters for the healthy control group and the glaucoma patients is given in the supplementary material S1.Figure 3Depiction of the RNFL and GCC thickness parameters. For the RNFL thickness, 8 values were extracted from the 6 \(\times \) 6 \(\textrm{mm}\) ONH scans in a ring around the ONH, while for the GCC thickness 8 values were obtained from the 6.4 \(\times \) 6.4 \(\textrm{mm}\) macula scans from the inner and outer ring of the ETDRS grid. The scale bars in the lower left corners indicate 1\(\textrm{mm}\).All images were visually reviewed by ophthalmology experts who excluded images that they considered to be of insufficient quality for use in clinical routine.The study was performed in accordance with the tenets of the Declaration of Helsinki and was approved by the local ethics committee of the University of Erlangen (3458, 295_20 B). Informed written consent was obtained from each participant.This data was split into 60% training set, 20% validation set and 20% test set, with all eyes from one patient belonging exclusively to only one set. This leads to a distribution of 66/21/21 patients in the training/validation/test set.ExperimentsIn order to examine the impact of the acquisition area on the performance of deep learning-based approaches, we trained 13 CNNs on the different regions and plexuses displayed in Fig. 2. We chose the same hyperparameters that have worked best in our previous work23, namely a DenseNet161. Since the task to solve was the binary classification problem of distinguishing between glaucoma patients and healthy controls, binary cross-entropy was chosen as loss function. The networks were trained using stochastic gradient descent (SGD) with a momentum of 0.9, a learning rate of 0.001 and a batch size of 16. Moreover, the mean and standard deviation of the pixel intensities of all images have been calculated before the training and used to normalize the input images to have zero mean and a standard deviation of one. Class weights were assigned to be inversely proportional to the total amount of images per class in order to deal with the class imbalance inherent to the dataset.For comparison, we also trained support vector machines with the VD as features for the same 13 regions and plexuses. Again the VD was computed as in our previous work23 by first matching the histogram of the input image to the one of a pre-selected scan. Afterwards the images were contrast enhanced using contrast limited adaptive histogram equalization (CLAHE) followed by a Frangi vesselness filter27 to highlight the vessel structures. In order to obtain a binary vessel map, hysteresis thresholding was applied. The VD was then computed as the fraction of vessel pixel to all pixel in the image.Since biomarkers obtained from OCT scans are more established in the clinical routine in diagnosing and tracking progress of the disease, we also wanted to examine the performance of these biomarkers in comparison to our deep learning-based approach using OCTA. One set of biomarkers were the 8 values obtained from the GCC thickness maps. The second set of biomarkers were the 8 values obtained from the RNFL thickness maps. The third set of biomarkers contained three disc parameters, namely the disc area, the rim area and the cup area. For each of these three sets, SMVs were trained as well to distinguish between the glaucoma patients and the healthy controls.In order to evaluate the performance of the different approaches the area under receiver operating characteristics (AUROC) was chosen as metric. Moreover 10 runs were carried out were in each run 20% of the data was randomly assigned to the test set and 20% of the data was randomly assigned to the validation set. The remaining 60% form the training set. It was taken care that all eyes from one patients exclusively belonged to only one of the sets.A statistical analysis of the AUROC values of the 10 runs was performed in order to investigate which of the results are statistically significant different. This was done using an analysis of variance (ANOVA) model and Bonferroni correction was employed to account for multiple comparisons. The statistical analysis was done using SPSS version 28 (IBM Corp. Released 2021. IBM SPSS Statistics for Windows, Version 28.0. Armonk, NY: IBM Corp.) and a p-value less than 0.05 was considered statistically significant.