Subjects

The DAVE study ( ClinicalTrials.gov identifier NCT01552408 ) was a prospective, randomized, interventional clinical trial of patients with diabetic macular edema (DME) that was aimed at comparing the efficacy and safety of ranibizumab monotherapy vs combination therapy of ranibizumab with UWF FA-guided retinal photocoagulation targeted to areas of peripheral retinal nonperfusion. All subjects signed written informed consent and a Health Insurance Portability and Accountability Act authorization. Inclusion criteria consisted of patients ≥18 years of age with DME caused by diabetes mellitus (type 1 or 2). Participants had Early Treatment Diabetic Retinopathy Study best-corrected visual acuity between 78 and 24 (Snellen equivalent, 20/32 to 20/320), with center macular thickness ≥300 μm caused by DME measured by spectral domain optical coherence tomography (SD OCT; Heidelberg Spectralis+ OCT, Heidelberg, Germany) on clinical examination. Baseline images from 40 eyes of 29 patients with treatment-naïve DME were submitted to the reading center included in this analysis. None of the 40 eyes included in this analysis had evidence of high myopia.

The RECOVERY study ( ClinicalTrials.gov identifier NCT02863354 ) is a prospective, randomized, interventional clinical trial to assess the safety and tolerability of 2-mg intravitreal aflibercept injections given monthly or every 12 weeks for the treatment of retinal nonperfusion associated with proliferative DR (PDR) primarily assessed through retinal capillary nonperfusion. All subjects signed written informed consent and a Health Insurance Portability and Accountability Act authorization. Inclusion criteria consisted of patients ≥18 years of age with PDR secondary to diabetes mellitus (type 1 or 2) in which panretinal photocoagulation can safely be deferred in the presence of antivascular endothelial growth factor intravitreal injections as determined by the investigator. Participants were required to have a best-corrected visual acuity of >19 Early Treatment Diabetic Retinopathy Study letters (Snellen equivalent, >20/400) in the study eye, with substantial nonperfusion (defined as >20 disc areas) as assessed by the investigator. Baseline images from 40 eyes of 40 patients with PDR were submitted to the reading center and included in this post hoc analysis. None of the enrolled patients had evidence of high myopia.

UWF FA images of normal control subjects (30 eyes from 16 subjects) were obtained from a previous study using the Optos 200Tx device. The study was conducted at Medical Center Ophthalmology Associates, San Antonio, Texas, and was also approved by the institutional review board and adhered to the tenets of the Declaration of Helsinki. Written informed consent was obtained from all subjects before imaging.

Image Acquisition

The study protocol has been described in detail in previous studies. Study eyes were dilated with tropicamide 1% and phenylephrine 2.5%, and UWF pseudocolor images were captured using the Optos 200Tx (Optos PLC) centered on the fovea. After intravenous administration of fluorescein dye, UWF FA images were obtained during the early (45 seconds), middle (2 minutes and 30 seconds), and late (5 minutes) phases of the angiogram, and steered peripherally (nasally, temporally, superiorly, and inferiorly) in all 3 phases.

Image Projection and Montage

Uncorrected raw images were collected and then sent to the Doheny Image Reading Center (Doheny Eye Institute, Los Angeles, CA) for analysis. Images were transformed to stereographic projection images using proprietary software available from the manufacturer. This projection technique was accomplished by ray tracing every pixel through a combined optical model of the Optos 200Tx and a Navarro UWF model eye with an axial length of 24 mm. This optical model represented the projection used by the Optos 200Tx scanning laser ophthalmoscopy platform to create the 2-dimensional Optomap. The software also registered the 4 steered images to the on-axis image automatically and created a montage of all images. When the whole retina was not visible on 1 image, a montage image was created. Image registration used their extracted vasculature and aligned this using cross-correlation to find the optimum translation and rotation (ie, an algorithm slightly rotated the peripheral images to align vasculature). Finally, segments were blended together to create a contiguous montage.

Grading of the Retinal Nonperfusion Area

A group of trained reading center–certified UWF FA graders (K.W. and K.G.F for DAVE study; W.F. and M.N. for Recovery study) were masked to the study groups and independently analyzed the images according to previously reported standardized grading protocols at the Doheny Image Reading Center. Graders were allowed to adjust the contrast and brightness to optimize visualization of the areas of nonperfusion. NPA was defined by the absence of retinal arterioles or capillaries with overall hypofluorescence relative to the overall background on a midphase FA frame. Using Image J version 1.49b (US National Institutes of Health, Bethesda, MD, USA), the graders manually delineated the border of the NPA and the grading results were exported as a binary mask and subsequently calculated in millimeters squared automatically by summing the size of all pixels using a software provided by the manufacturer (Optos PLC). To be segmented or graded as a separate foci of nonperfusion, the minimum region size was 3 pixels. This threshold, however, rarely came into play (if ever), because most potentially small separate regions of nonperfusion were connected to larger regions of nonperfusion, even if by small bridges. The size of a pixel was defined by its location in the image and was calculated using spherical trigonometry after projecting it back onto a sphere. In cases where a difference in NPA of >20% was present between graders, the graders met in open adjudication to agree on a single consensus result for each case. When smaller differences were present, the results of the 2 graders were averaged to yield a final result for subsequent correlative analyses.

Extraction of Retinal Vasculature

The early phase images of the UWF FA angiogram (4000 × 4000 pixels) was selected to segment the retinal vasculature. In addition, we also excluded the images if the retinal vessels were low contrast relative to the background choroidal fluorescence. For this study, we adopted a previously reported protocol that was particularly designed for vessel segmentation in UWF view scanning laser ophthalmoscope images. In brief, the image was processed using scaled filtering operations to increase vessel contrast with respect to the background. The filters exploited the elongated nature of the vasculature and their cross-sectional intensity profile. A machine learning technique was then used to classify image pixels as vessel or nonvessel. The segmented results were exported as binary masks. Figure 1 shows the extent of vessels that could be extracted using this method—in particular, while there is a clear delineation of the large, medium, and small caliber retinal vessels, there is little visualization of the capillaries. Microaneurysms, intraretinal microvascular abonormalities, and neovascularization elsewhere could not be extracted ( Figure 1 ).

Binarized and skeletonized imaging of retinal vasculature in diabetic retinopathy on ultrawide-field fluorescence angiography. The segmentation of retinal vasculature takes advantage of the elongated nature of the vessels and enables a clear delineation of the vessels with large, medium, and small caliber but not microvasculature. (A-C) Microvasculature is little visualized. (D-F) Microanuerysms, especially those that appear not to be connected to the vascular branches, could not be extracted. (G-I) Neovascularization could not be extracted. Therefore, the fractal dimension metric produced by our analysis protocol represents the fractal features of the larger retinal vessel branches but not the capillary circulation.

Calculation of FD

The binarized image was then skeletonized by reducing all the continuous white segments to a line with a single pixel’s width using ImageJ software ( https://imagej-nih-gov.easyaccess2.lib.cuhk.edu.hk/ij/ ). The obtained images were then subjected to image analysis using ImageJ software with FracLac plugin ( https://imagej-nih-gov.easyaccess2.lib.cuhk.edu.hk/ij/plugins/fraclac/fraclac.html ) to calculate FDs. The basic box counting algorithm was originally modified from ImageJ’s box counting algorithm. Theoretically, the skeletonized image was covered with square boxes of side length (L) in pixels. The number of boxes (N) of side length L are counted and denoted as N(L). The size of the box is then incrementally increased and the process is repeated. N(L) is tabulated versus the size of the box (L). A plot of log of N(L) versus (L) is generated, and a linear least squares regression is made to determine the slope of the plot. The slope of the line relating these 2 variables, with the inverted signal, is the FD.

Division of Retinal Field

A prespecified custom grid consisting of 2 rings or circles centered on the fovea with radius of 10 mm and 15 mm was applied to the images, dividing the UWF FA image into 3 zones: a posterior zone (within a radius of 10 mm), a midperipheral zone (10~15 mm), and a far peripheral zone (>15 mm).

Statistical Analysis

Baseline FD in the DR eyes was compared with age- and sex-matched healthy control subjects and then correlated to NPA. Statistical analysis was performed using the R statistical analysis package (version 3.3.0). A Shapiro–Wilk normality test was used to explore the distribution normality of continuous variables. When the numeric data were normally distributed, an independent t test was used to compare means between 2 independent groups, otherwise the nonparametric analysis (Wilcoxon rank sum test) was applied. Linear modeling was applied to calculate person correlation coefficients and describe the relationship between FD and NPA. Statistical results were expressed as P values. P < .05 was considered statistically significant.