Patient Group

SD OCT raster scans (512 × 128) (Cirrus; Carl Zeiss Meditec, Dublin, California, USA; software version 3.0.0.64 or later) and corresponding fluorescein angiograms from 142 patients with treatment-naïve DME in the study eye were analyzed. Images in the baseline data sets of a large multicenter clinical study were processed by the Vienna Reading Center. The data used were from patients with DME and no other confounding retinal disease. Ethical approval was obtained from the institution’s review board. Each patient gave informed consent before inclusion in the study. The study adhered to the tenets of the Declaration of Helsinki and the standards of Good Scientific Practice of the Medical University of Vienna. The images were taken at each study site according to the same predetermined scanning protocol to guarantee a standardized procedure.

Inclusion criteria for the study were age over 18 years, treatment naivety (for at least 3 months before scanning with OCT), and DME attributable to nonproliferative diabetic retinopathy. At least 1 eye had to be affected by clinically significant DME (as diagnosed by the investigator, an ophthalmologist trained in diagnosis of clinically significant macular edema) with a central retinal thickness (CRT) >300 μm. Clinically significant macular edema was defined as “retinal thickening that involves or threatens the center of the macula (even if visual acuity is not yet reduced)” according to the Early Treatment Diabetic Retinopathy Study Report 1. Exclusion criteria were ocular opacity attributable to cataract, vitreous opacification or corneal opacities that might lead to poor image quality, history of any other vitreoretinal disease or treatment, and high myopia ≥8 diopters. The control group comprised healthy volunteers of whom at least 50% were aged 50 years or older. The healthy control and patient groups had no history of ocular disease and no ocular opacity attributable to cataract or vitreous opacification that might lead to poor image quality. Study patients’ eyes were divided into 3 groups: study eyes with DME, non-study fellow eyes with DME, and non-study fellow eyes without DME.

Spectral-domain Optical Coherence Tomography Analysis of Retinal and Choroidal Alteration

OCT images underwent automatic analysis at the University of Iowa and manual evaluation by certified graders of the Vienna Reading Center. As a first step, the retinal layer segmentation of 10 intraretinal surfaces and Bruch’s membrane were detected automatically using the graph-based multilayer segmentation approach, called the Iowa Reference Algorithm. Errors in the segmentation of the Bruch’s membrane were systematically evaluated and large deviations were corrected manually by the reading center ( Figure 1 , Top). Afterwards the fully automated choroidal volume assessment could be performed. After flattening the image along the Bruch’s membrane (to reduce geometric distortion; Figure 1 , Upper center), an automated vessel segmentation algorithm was applied to identify vessel shadowing effects and to smoothen the images to prevent shadows being identified as choroidal vessels ( Figure 1 , Lower center).

Choroidal volume assessment on standard spectral-domain optical coherence tomography by a 4-step process of automated choroidal thickness calculation. Images underwent automatical analysis. The figure shows the segmentation on a Cirrus 512 × 128 macular cube scan by a single B-scan. The entire segmentation was performed 3-dimensionally. As a first step, the retinal layer segmentation including Bruch’s membrane (blue line) was detected automatically. Errors in the segmentation of the Bruch’s membrane were corrected manually (Top). This was followed by the fully automated choroidal volume assessment. After flattening the image along the Bruch’s membrane (Upper center), an automated vessel segmentation algorithm was applied to identify vessel shadowing effects and to smoothen the images to prevent shadows being identified as choroidal vessels (Lower center). A pair of thin-plate splines was fitted on each side of the choroidal vasculature (red and green lines), which allowed quantification of choroidal thickness for each A-scan (Bottom). The thickness between the red and blue lines shows the choriocapillaris equivalent thickness.

As a next step, the entire choroidal vasculature was segmented using an automated method for 3-dimensional segmentation of choroidal vasculature that includes Hessian detection of 3-dimensional cylindrical features and probabilistic identification of vasculature seed points. Region-growing segmentation to determine choroidal vasculature followed ( Figure 2 ). A pair of thin-plate splines was fitted on each side of the choroidal vasculature, which allowed quantification of choroidal thickness for each A-scan ( Figure 1 , Bottom, thickness between the red and green lines). Similarly, the distance between the inner surface (thin-plate spline) of the choroidal vasculature and the Bruch’s membrane was defined as the choriocapillaris equivalent, and its thickness was calculated automatically for each A-scan ( Figure 1 , Bottom, thickness between the blue and red lines). The sum of the 2 choroidal layers was defined as choroidal thickness in this study.

Choroidal vessel detection illustrated on the en face choroidal image from the macular cube scan without segmentation, with the projection of the segmentation into the en face image and the 3-dimensional reconstruction of the vessels. The entire choroidal vasculature was segmented using an automated method for 3-dimensional segmentation of choroidal vasculature that includes Hessian detection of 3-dimensional cylindrical features and probabilistic identification of vasculature seed points. Region-growing segmentation to determine choroidal vasculature followed. The figure shows the en face choroidal image before segmentation (Left), the choroidal vessel segmentation (Center), and the 3-dimensional segmentation projection image as a result (Right).

To assure grading of the same position at the posterior pole, the fovea was checked and, if necessary, manually replotted by experienced graders from the Vienna Reading Center to ensure that the Early Treatment Diabetic Retinopathy Study (ETDRS) grids were in the same position in each image stack ( Figure 3 , Top). Image stacks with obvious segmentation errors in the Bruch’s membrane were corrected manually or excluded if manual repositioning of this surface’s segmentation was not possible (eg, owing to poor image quality). All images with an insufficient quality of Bruch’s membrane or where parts of the data set were missing (eg, fluorescein angiograms) were excluded from statistical analysis. The retinal and choroidal thicknesses between the internal limiting membrane and Bruch’s membrane and between the Bruch’s membrane and the outer choroidal limiting membrane were calculated automatically in all 9 subfields. A second grid was centered at the peak of edema to allow comparison of the values from the grids to each other and to retinal and choroidal values in the peak region ( Figure 3 , Bottom). All values were calculated in micrometers (μm). Finally, all insufficient segmentation results (defined as segmentation where >50% of choroidal vessels in 6 mm × 6 mm could not be detected automatically) were identified manually ( Figure 4 ).

Process of Early Treatment Diabetic Retinopathy Study (ETDRS) grid positioning distinguishing between fovea-centered and edema-centered grid positions. The fovea was checked and, if necessary, manually replotted by experienced graders from the Vienna Reading Center to ensure that the ETDRS grids were in the same position in each image stack and to assure grading of the same position at the posterior pole (Top). A second grid was centered at the peak of edema (thickest point of edema) to allow comparison of the values from the grids to each other (Bottom). The figure shows the central B-scan of each grid position (right; fovea or peak of edema; yellow, center point; light blue, 1 mm subcircle; blue, 3 mm subcircle; dark blue, 6 mm subcircle of grid) and the corresponding retinal surface image (Left) with the retinal thickness map and the ETDRS grid plotted on it (thick-normal-thin equals red-yellow-green on the thickness map; see above for grid subcircle colors).

Examples of segmentation failure cases after choroidal volume assessment on standard spectral-domain optical coherence tomography by automated choroidal thickness calculation. After segmentation of choroidal thickness in all scans, all insufficient segmentation results, where >50% of choroidal vessels could not be detected automatically, were identified manually. These segmentation failure cases were excluded. The figure shows 1 sufficient segmentation result (Left) and 2 insufficient segmentation results of different severity (Center, Right). For each example 1 B-scan (Top) with the resulting 3-dimensional segmentation result (Bottom) of the entire macula cube is displayed.

Angiographic Delineation of Leakage Areas

Analysis of fluorescein angiograms was confined to measurement of the maximum area of leakage in the late phase. The timing of the late phase is strictly set within a standardized reading center trial and was therefore the same for each angiogram. The measurements were performed within a reading center setting. The pixel number and size were standardized in all images before inclusion of study sites in the multicenter trial using the Vienna Reading Center’s test eye procedure to allow measurement of the area in mm ² with reference to the SISPOT test eye. To define the maximum area of leakage, the total area of hyperfluorescence of the macula in the late-phase image compared with the early-phase image was marked manually with ImageJ (Version 1.46r) by experienced and certified graders of the Vienna Reading Center under continuous supervision by senior retinal specialists. Staining areas and window defects were not included.

Correlation of Retinal and Choroidal Disease

Retinal and choroidal (choriocapillaris) mean thickness values were calculated in all ETDRS subcircles of eyes with DME manually centered on the fovea or on the peak of edema, and in non-DME fellow eyes, DME fellow eyes, and healthy eyes. The correlation between each retinal and choroidal thickness value in the same patient group at the same position was calculated. As a next step, choroidal thickness values were compared between the patient groups, that is, values of DME fovea-centered vs DME peak of edema–centered ETDRS grids, DME fovea-centered vs non-DME fellow eyes, DME fovea-centered vs DME fellow eyes, DME fovea-centered vs healthy eyes, DME peak of edema–centered vs non-DME fellow eyes, DME peak of edema–centered vs DME fellow eyes, DME peak of edema–centered vs healthy eyes, non-DME fellow vs DME fellow eyes, non-DME fellow vs healthy eyes, and DME fellow vs healthy eyes. Furthermore, mean leakage area was calculated and compared with choroidal thickness values to assess differential effects of diabetic conditions on the retinal and the choroidal vasculature.

Statistical Analysis

Statistical comparisons were done by nonparametric methods: Mann-Whitney tests for comparison of independent groups and Wilcoxon matched-pairs tests for within-participant comparisons. Relations between continuous variables were inspected by scatterplots and, if found, linear Pearson correlation coefficients were computed. In all tests, P values below .05 were considered significant. Multiple endpoint correction according to Bonferroni-Holm was applied, and P values were adjusted for significance.