Study Design.
A retrospective review of OCTA and ICGA images was conducted of patients diagnosed with posterior and panuveitis and evaluated by the uveitis service at the National Eye Institute (NEI), National Institutes of Health (NIH) Bethesda, MD between December 2016 to August 2021. All research was conducted under an Institutional Review Board (IRB) approved protocol where informed consent for the inclusion for the use of data was obtained by all study participants. This study adhered to the tenets of the Declaration of Helsinki.
Anatomical diagnosis was made based on the Standardization of Uveitis Nomenclature (SUN) criteria.14 Demographic and clinical data including age, sex, best corrected visual acuity (BCVA) in LogMAR, spherical equivalent (S.E.), disease duration, disease activity were obtained.
Inclusion and exclusion criteria.
Eyes with posterior and panuveitis and with hypocyanescent lesions present on mid-phase ICGA were included. OCTA images obtained on the same date were compared with ICGA. Exclusion criteria were as follows: 1) OCTA image with signal strength < 7; 2) substantial motion artifact or segmentation error; 3) co-existent retinal and choroidal diseases other than uveitis including age-related macular degeneration or central serous chorioretinopathy, and 4) eyes with previous history of macular laser. Structural en-face OCTA images were reviewed to exclude false flow deficits.
Image acquisition.
The ICGA and OCTA images were obtained using the Spectralis HRA, (Heidelberg, Germany) and the CIRRUS AngioPlex Model 5000, (Carl Zeiss Meditec, Germany), respectively. On OCTA, only 6 X 6-mm scans centered on the fovea were included. For quantitative and qualitative interpretation of OCTA images, maximum projection was used to construct the en-face representations of the choriocapillaris slab. We defined the choriocapillaris slab at the default setting (29–49 µm below the retinal pigment epithelium (RPE)-fit segmentation line, an estimate of Bruch’s membrane).15 If improper automated segmentation was noted, manual corrections were performed accordingly.
Image processing and analysis.
To quantify the agreement between choriocapillaris flow deficits on OCTA and hypocyanescent lesions on ICGA, a novel semi-automatic algorithm was developed. Each image was graded independently from the corresponding imaging modality. Throughout, the grader was blinded to the patient’s diagnosis.
First, images were imported into ImageJ (version 1.51g, National Institutes of Health, Bethesda, MD). Next, the en-face superficial capillary plexus (SCP) OCTA image was opened alongside the ICGA image for feature-based image registration. Image registration is the process of aligning two images to determine corresponding points.16 Points are often utilized for the basis of registration and were placed on distinctive vessel junctions at corresponding locations.17,18 A single point was placed in each quadrant of the image for a total of four per image (Fig. 1a,b). Moving Least Squares (MLS) is an automatic, non-rigid, point-based technique in which two images are aligned based on feature points extracted from them.19 Using the similarity MLS deformation technique (Plugins > Registration > Moving Least Squares > Similarity method), the OCTA image was registered to the ICGA image (Fig. 1c). Image registration was validated by careful visual inspection.
Using the freehand selection tool in ImageJ, hypocyanescent lesions on ICGA images (Fig. 1d,e) or were meticulously outlined (in magenta) and added to the region of interest (ROI) manager. The lesion outlines on each image were filled, and the image was binarized (Fig. 1f). The process was repeated for choriocapillaris flow deficits on OCTA images (Fig. 1g,h,i). To determine the spatial overlap between the lesion areas, the Dice Similarity coefficient (DSC) was computed (Fig. 2). The DSC measures the spatial overlap between two segmentations.20 A value closer to 0 indicates no spatial overlap between two sets of binary images, and a value of 1, indicates complete overlap.21 Good agreement was considered if the DSC ≥ 0.7, moderate agreement if the DSC was between 0.5 and 0.7, and poor agreement if the DSC ≤ 0.5, as similarly reported in the literature.22 The DSC was calculated using the CLIJ2 plugin on ImageJ.23
To characterize lesion morphology the “Analyze particles” function on ImageJ was applied to measure the lesion number (LN), mean lesion size (MLS), lesion density (LD%), and the lesion circularity index (LCi). LN is the number of distinct lesions in each image. MLS describes the average lesion size in each image. LD% is the proportion of lesion area to the total image area. LCi quantifies the mean circularity of the lesions in the image, where for any lesion, a value of 1.0 indicates a perfect circle and a value approaching 0.0 indicates an increasingly elongated shape.
Statistical methods.
Statistical analysis was performed using GraphPad Prism version 9.1.2 (GraphPad Software Inc, La Jolla, CA) and. The Anderson-Darling test was performed to detect the normality of distribution. Given their distribution, lesion morphology measurements (LN, MLS, LD% and LCi) obtained from both modalities were compared using the non-parametric Clustered Wilcoxon signed rank test.24 At the eye level, Bland-Altman plots were used to compare the differences in lesion morphology between OCTA and ICGA against the average of each measurement. Limits of agreement were set to ± 1.96 standard deviations from the mean difference for each morphological feature. To compare the DSC between the clinically active and clinically quiet eyes, Welch’s t-test was selected. The chosen level of statistical significance for all tests was P less than 0.05. Mean and SD are reported as “mean ± SD”.