Impact of Multiple En Face Image Averaging on Quantitative Assessment from Optical Coherence Tomography Angiography Images Akihito Uji, MD, PhD,1,2 Siva Balasubramanian, MD, PhD,1,2 Jianqin Lei, MD,1,2 Elmira Baghdasaryan, MD,1,2 Mayss Al-Sheikh, MD,1,2 SriniVas R. Sadda, MD1,2 Purpose: To investigate the impact of multiple en face image averaging on quantitative measurements of the retinal microvasculature using optical coherence tomography angiography (OCTA). Design: Prospective, observational, cross-sectional case series. Participants: Twenty-one healthy individuals with normal eyes. Methods: Macular OCTA images were acquired from all participants using the Zeiss Cirrus 5000 with Angioplex OCTA software (Carl Zeiss Meditec, Dublin, CA). Nine OCTA cube scans per eye were obtained and 9 superficial retinal layer (SRL) and deep retinal layer (DRL) en face OCTA image slabs were averaged individually after registration. Quantitative parameters from the retinal microvasculature were measured on binarized and skeletonized OCTA images and compared with single OCTA images without averaging. Main Outcome Measures: Vessel density (VD), vessel length density (VLD), vessel diameter index (VDI), and fractal dimension (FD). Results: Participants with artifact or poor image quality were excluded, leaving 18 eyes for the analysis. After averaging, qualitatively there was apparent reduction in background noise, and fragmented vessels in the images before averaging became continuous with smoother walls and showed sharper contrast in both the SRL and DRL. Binarized and skeletonized derivates of these averaged images also showed fewer line fragments and dots in nonvascular areas and more continuous vessel images than those of images without averaging. In both SRL and DRL, VD (P ¼ 0.0010 and P ¼ 0.0003, respectively), VLD (P < 0.0001 for both), and FD (P < 0.0001 for both) significantly decreased and VDI significantly increased after averaging (P < 0.0001 for both). Conclusions: Averaging of multiple en face OCTA images improves image quality and also significantly impacts quantitative measurements. Reducing noise that could be misinterpreted as flow and annealing discontinuous vessel segments seem to be major mechanisms by which averaging may be of benefit. Ophthalmology 2017;-:1e9 ª 2017 by the American Academy of Ophthalmology
Optical coherence tomography angiography (OCTA) is a promising technology that can be used to image the retinal microvasculature noninvasively without the use of contrast agent dye.1e4 By segmenting the cubic OCTA data into specific layers or slabs, OCTA can provide en face images of the deep retinal plexus separately from the superficial plexus that cannot be evaluated adequately by fluorescein angiography.2 With its high-contrast imaging for individual retinal layers, OCTA offers a significant advantage for quantitative characterization of the microvascular morphologic features. Foveal avascular zone area, vessel density (VD), and vessel length per unit area have been shown to be useful parameters in normal eyes and diseased eyes, including those with diabetic retinopathy and retinal vascular occlusions.5e8 Moreover, fractal dimension (FD), which can describe the complexity of the vasculature, has been gaining interest as another quantitative parameter for evaluating microvascular pathologic features in OCTA imaging.5,9,10 ª 2017 by the American Academy of Ophthalmology Published by Elsevier Inc.
Multiple B-scan averaging is a commonly used technique and one of the most effective strategies to reduce the level of speckle noise and enhance the image quality of optical coherence tomography (OCT) B-scan images.11,12 Today, this technique is implemented in most of the commercially available OCT devices. Given that OCTA en face images also contain noise, application of this averaging technique on multiple en face images generated from different volume data sets also may be expected to increase the contrast of OCTA images. Moreover, the technique may make discontinuous vessel segments or fragments more continuous. However, it may be hypothesized that improvement in image quality could impact the quantitative analyses from OCTA images. For example, FD calculation on images with fragmented vessel segments with background noise would be expected to be different from images with continuous vessels and less noise in the intervascular spaces. Thus in the present study, we assessed the impact of multiple en face http://dx.doi.org/10.1016/j.ophtha.2017.02.006 ISSN 0161-6420/17
1
Ophthalmology Volume -, Number -, Month 2017 image averaging on quantitative parameters from the retinal microvasculature using OCTA.
Methods This was a prospective, observational, cross-sectional case series. This study was approved by the Institutional Review Board of the University of California, Los Angeles, and was conducted in accordance with the tenets of the Declaration of Helsinki and in compliance with the regulations set forth by the Health Insurance Portability and Accountability Act. Written informed consent was obtained from all examined healthy individuals before they participated in the study.
Participants Twenty-one healthy volunteers with no history of ocular or systemic disease were recruited into the study. All imaging was performed without pupil dilation.
Optical Coherence Tomography Imaging The OCTA images were obtained using the Zeiss Cirrus 5000 with Food and Drug Administrationecleared Zeiss Angioplex software (Carl Zeiss Meditec, Dublin, CA), which acquires 4 sequential OCT scans in the same location and generates en face OCTA images using the optical microangiography algorithm. One eye from each participant was selected randomly, and the 33-mm scans centered on the fovea were obtained repeatedly until 9 OCTA cubes with sufficient image quality could be obtained. On the scan quality check screen, the quality of the scans was assessed according to acceptance criteria. Acceptable images should have clear and sharp focus, few to no artifacts (e.g., motion lines), minimal saccades (identified by horizontal misalignment of vessel segments on en face images), and signal strength of 7 or more. Moreover, images should be centered on the fovea and illuminated uniformly without dark corners. If acceptable scans meeting these criteria (for all 9 acquisitions) could not be acquired, the eye was excluded from the analysis. En face images of the superficial retinal capillary layer (SRL) and deep retinal capillary layer (DRL) were obtained using the commercial default automated segmentation boundaries and exported at a size of 10241024 pixels for further analyses.
Multiple En Face Image Averaging The SRL and DRL en face images were analyzed separately with 9 en face images from each layer separately stacked to generate a 9-frame video. Because of the eye motion between image acquisitions, videos demonstrated obvious evidence of misalignment between frames, which included not only translation and rotational differences, but even elastic or nonlinear differences from frame to frame.13,14 Thus, simple averaging produced blurry images because of poor overlapping of vessels (Fig 1GeI). To overcome this issue, both linear and nonlinear (elastic) image registration were performed using ImageJ (developed by Wayne Rasband, National Institutes of Health, Bethesda, MD; available at http:// rsb.info.nih.gov/ij/index.html), which has been reported as an effective method to improve image quality in adaptive optics imaging.15 Briefly, after importing the 9-frame video of SRL as 8-bit grayscale images into ImageJ, we cropped a central rectangular area of 800800 pixels to exclude the Zeiss logo in the image, which would be expected to compromise the accuracy of the registration (Fig 2). First, linear image registration was performed to align position gaps using the plug-in Stackreg (available at
2
http://bigwww.epfl.ch/thevenaz/stackreg/).16 Subsequently, elastic image registration was performed to correct for intraframe distortions that were uncorrectable by linear registration using the plug-in bUnwarpJ (available at http://imagej.net/BUnwarpJ).17,18 During the registration, corresponding points of interest in the 2 images were extracted using the plug-in Feature Extraction (available at http://imagej.net/Feature_Extraction), which is an algorithm for Multi-Scale Oriented Patches; these corresponding points were used as weighted landmarks.19 Of the 9 frames, the best-quality frame was selected as the reference or base frame and the remaining 8 frames were warped using this process to match the reference frame. Transformation information obtained from the SRL registration then was applied to the DRL, because the SRL and DRL from the same case presumably are perfectly aligned. As soon as the elastic registration process was complete, the 9 registered frames for the SRL and DRL were averaged separately for subsequent quantitative analyses. The reference or base frame used in registration was defined as the original image and was compared with the averaged image. A macro that automates a series of ImageJ commands was created, and all of the digital image processing described above was executed automatically.
Image Quality Comparison between Averaged Image and Original Nonaveraged Images Two certified readers (S.B. and E.B.) at the Doheny Image Reading Center masked to image information (i.e., averaged vs. nonaveraged) performed independent expert comparisons of pairs of averaged image and original nonaveraged images. Ten cases were selected and all 9 single images used for averaging were compared with corresponding averaged images, and 90 pairs of images were graded in total. Forty-five pairs were graded using DRL images and the remaining 45 pairs were graded using SRL images. Images were arranged in 2 panels (left and right) to facilitate comparison, with random assignment of averaged versus nonaveraged images to the left and right panels. The grader assigned a score for comparative image quality for 3 parameters: (1) vessel quality (contrast and continuity), (2) nonvascular area quality (background noise level), and (3) an overall image quality score (overall clarity) to each pair of images. A comparative image quality score was assigned to each image pair as follows: 2 ¼ the left image is definitely better; 1 ¼ the left image is slightly better; 0 ¼ the two images are equal; 1 ¼ the right image is slightly better; and 2 ¼ the right image is definitely better. Cases with disagreement between graders were resolved by open adjudication between graders to yield a single determination for each image pair.
Quantitative Measurements Vessel density, vessel length density (VLD), vessel diameter index (VDI), and FD were measured for comparison between the original image and averaged image in both the SRL and DRL. For measurements, a 680680-pixel rectangular box centered on the fovea was cropped and binarized using a modified version of the previously reported method.5 Briefly, after processing with a top-hat filter, the image was duplicated and a different binarization method was performed. One image was processed first by a Hessian filter, followed by global thresholding using Huang’s fuzzy thresholding method. The other (duplicate) image was binarized using median local thresholding. Finally, the 2 different binarized images were combined to generate the final binarized image in which only pixels that existed on both binarized images were included. Vessel density was assessed on the final binarized image and was defined as the ratio of the area occupied by vessels divided by
Uji et al
OCTA with Multiple En Face Image Averaging
Figure 1. Optical coherence tomography angiography images showing the difference in quality between images obtained before and after multiple en face image averaging. AeC, Optical coherence tomography angiography images before multiple en face image averaging. Images contain background noise and vessels are depicted as discontinuous lines. DeF, Images after multiple en face image averaging. After averaging, noise scattered about on the images is reduced and fragmented vessel images are annealed together, resulting in continuous, smooth, and high-contrast vessel images. GeI, Images after multiple en face image averaging without image registration. The averaging produced blurry images resulting from poor overlapping of vessels. A, D, G, Optical coherence tomography angiography images of the superficial retinal capillary layer (SRL). B, E, H, Optical coherence tomography angiography images of the deep retinal capillary layer (DRL). C, F, I, Composite images: SRL is green and DRL is red.
the total area. After skeletonization of the binarized image, VLD, which represents the vessel length per unit area, was evaluated as described previously.5,10 Vessel diameter index, which represents the average vessel caliber, was calculated by dividing the total vessel area in the binarized image by the total vessel length in the skeletonized image.5 Finally, FD was calculated on the
skeletonized image using Fractalyse (TheMA, Besanç on Cedex, France).9 The box-counting method was used for calculation. The FD can range from 0 to 2, and images with a more complex vessel branching pattern will have a higher FD.20 To characterize further the impact of averaging and to assess the minimum level of averaging required, averaged images also were constructed
3
Ophthalmology Volume -, Number -, Month 2017
Figure 2. Multiple en face image averaging of optical coherence tomography angiography images. A, The 33-mm optical coherence tomography angiography (OCTA) image of the superficial retinal capillary layer (SRL) centered on the fovea obtained using a Zeiss Angioplex spectral-domain OCTA device (Carl Zeiss Meditec, Dublin, CA). B, The 800800-pixel rectangular area was cropped from (A) (outlined in yellow). C, Reference (first) frame with deformation grids and landmarks automatically extracted by the software. D, Second frame with deformation grids and landmarks. The image is deformed elastically to look as similar as possible to the reference image. E, Diagram showing the registration strategy. Nine en face images (each) of the SRL and deep retinal capillary layer (DRL) were stacked separately to generate a 9-frame video. The best-quality frame was selected as the reference (base) frame and the remaining 8 frames were warped to match the reference frame. Transformation information used in the SRL registration was applied to the DRL registration. After registration, the 9 frames were averaged for SRL and DRL.
using a smaller number of frames (2e9), and resultant quantitative metrics were compared with the original image.
Statistical Analysis All values were expressed as the mean standard deviation. The intraclass correlation coefficient was used to evaluate interobserver agreement. Differences in VD, VLD, VDI, and FD between the original image and averaged image were assessed using the paired t test. Comparisons of VDL and FD of the 9 different averaging images (original image and averaged image with 2 to 9 frames) were carried out using repeated measures analyses of variance, and differences between the 2 groups were analyzed using the paired
4
t test followed by Bonferroni correction. A P value of less than 0.05 was considered statistically significant. All analyses were performed using StatView software version 5.0 (SAS Institute, Cary, NC).
Results Two eyes with off-center scans and 1 eye with nonuniform illumination in some scans did not meet the prespecified quality criteria and were excluded from the analyses. Thus, 18 eyes from 18 participants were included in the final analysis. These participants had an average age of 34.95.9 years (range, 24e49 years). Eight participants were men and 10 were women.
Uji et al
OCTA with Multiple En Face Image Averaging
Image quality assessment by masked reading center graders (S.B. and E.B.) revealed that in all cases, the averaged image was rated to have less background noise and more continuous vessels for both the SRL and DRL compared with the original nonaveraged image (Fig 1). Eighty-four image pairs (93.3%) for vessel quality, 79 image pairs (87.8%) for the nonvascular area, and 90 image pairs (100%) for overall image quality were graded to have definitely better image quality with the averaged image compared with the original nonaveraged image. Only 6 image pairs (6.7%) for vessel quality and 11 image pairs (12.2%) for the nonvascular area were graded to have slightly better image quality in the averaged image compared with the original image. The intraclass correlation coefficient was 0.996 for the vessel, 0.970 for the nonvascular area, and 0.999 for the overall image quality assessment, indicating good interobserver agreement. Figure 3 shows the images after binarization and skeletonization for use in quantitative measurements. Binarized averaged images of the SRL and DRL showed fewer line fragments and dots in nonvascular area and more continuous vessels than binarized original images. A similar observation was made for skeletonized images. Moreover, vessels in binarized original images showed a spiny, branching-like morphologic feature, whereas the vessels in the averaged images appeared smooth with a more uniform caliber. These spiny, branching-like features in the binarized original images were converted to many discontinuous branches of vessels in the skeletonized images. Table 1 summarizes the measurement results from the binarized and skeletonized images. In both the SRL and DRL, VD (P ¼ 0.0010 and P ¼ 0.0003, respectively) and VLD (P < 0.0001 for both) showed significantly lower values in averaged images than in the original image, whereas VDI showed a significantly higher value in averaged images than in the original image (P < 0.0001 for both). Fractal dimension in averaged images was significantly lower than in the original images (P < 0.0001). For the SRL, the magnitude of the difference (9 average compared with original) was 4.344.52%, 9.051.99%, 5.133.21%, and 1.250.24% for VD, VLD, VDI, and FD, respectively. For the DRL, the magnitude of the difference (9 average compared with original) was 6.636.16%, 13.494.91%, 7.891.89%, and 2.090.77% for VD, VLD, VDI, and FD, respectively. Figure 4 shows the relationships between quantitative parameters (VLD and FD) and the number of frames (2e9) used for averaging. Significant differences were detected among the different levels of averaging in VLD (P < 0.0001 for both SRL and DRL) and FD (P < 0.0001 for both SRL and DRL), and VLD and FD decreased as the number of frames used for averaging increased. For the SRL, there was a statistically significant difference in VLD and FD between original images and the images derived by averaging 2 to 9 frames. Of note, the biggest improvement or reduction in VD and FD came with the first level of averaging (2 frames), with a diminished magnitude of benefit after 5 frames of averaging. For the DRL, the images derived by averaging more than 3 frames showed significant differences from the original image. Significant differences were not found among images derived by averaging more than 6 frames in the DRL.
Discussion In this study, we evaluated the impact of multiple en face image averaging on quantitative parameters commonly derived from OCTA images. Our findings showed a significant difference in quantitative measurements between the
original image and averaged images, suggesting that the noise and vessel discontinuities apparent in unaveraged images do have a significant impact on these quantitative parameters. Considering that current OCTA measurements are highly dependent on the result of binarization, a higher background noise level could affect the thresholding level for binarization. In addition, accurately assessing the complexity of the vessel branching pattern with fractal dimension requires that discrete vessel segments are actually separate vessels and not a component of an otherwise continuous vessel. Thus, a background noise and annealing of vessel segments into continuous vessels would seem to be an important preprocessing step before binarization. Along with foveal avascular zone area, VD and VLD have been reported to be clinically relevant quantitative metrics from OCTA imaging. Several groups, including our own, have demonstrated the clinical relevance of VD measurement in patients with diabetic retinopathy.5,6 In this study, VD and VLD significantly decreased after averaging. Given that reduced noise decreases VD and VLD and that increase in continuous vessels (increase in area occupied by vessels) increases VD or VLD, the effect of averaging on quantitative measurements seems to be greater in terms of noise reduction than in terms of vessel continuity. However, VDI that reflects the average vessel caliber significantly increased after averaging. As shown in Figure 3, the irregular caliber of the capillaries in the binarized original image became smooth and of uniform caliber after averaging. Although the high contrast and numerous features present in OCTA images lend themselves to highly accurate automated image registration, subtle misregistrations cannot be excluded, and such misalignments also potentially could slightly expand vessel diameters and increase the VDI. However, the overall reduction in VD suggests that this is likely a small effect (if present) compared with the impact of noise reduction as a result of averaging. Fractal dimension, which represents the degree of pattern complexity, can be measured with the box-counting method by assessing the relationship between the size of the box and the number of blocks that the image touches.10,20,21 Therefore, as noted above, the presence of noise and vessel discontinuity was assumed to reduce the precision of FD measurements. The results showed a significant decrease in FD in both SRL and DRL after averaging, suggesting that the averaged image had less complexity than the original image. We suspect that the spiny, branching-like elements contained in the binarized and skeletonized original images may have increased the pattern complexity artificially and contributed to an inflated FD. Vessel length density and FD decreased as the number of frames used for averaging increased. Considering that significant differences were not observed for the DRL among images with more than 6 times averaging, it would seem that at least 6 frames are needed for ideal binarization. In contrast, for the SRL, there seemed to be a continued significant change in measurements as the number of frames for averaging was increased further. However, the biggest impact was observed by averaging even only 2 frames, and after 5 frames, the differences were much smaller. These
5
Ophthalmology Volume -, Number -, Month 2017
Figure 3. Binarization and skeletonization of optical coherence tomography angiography image before and after multiple en face image averaging. AeD, Optical coherence tomography angiography images. EeH, Binarized images. IeL, Skeletonized images. A, E, I, Optical coherence tomography angiography images of superficial retinal capillary layer (SRL) before multiple en face image averaging. E, Binarized image showing noise in the foveal avascular zone (FAZ) and many vessels showing a spiny, branching-like shape (arrows), suggesting the possibility that background noise may be mistakenly included as flow. B, F, J, Optical coherence tomography angiography images of SRL after multiple en face image averaging. B, Optical coherence tomography angiography image showing continuous, smooth, and high-contrast vessel images and noise reduction in the FAZ. F, Binarized image showing smooth and uniform caliber after averaging. Fewer line fragments and dots are detected in the nonvascular areas, including the FAZ, compared with the original image (E). J, Fewer line fragments and dots are detected in the nonvascular area than in the original image (I). C, G, K, Optical coherence tomography angiography images of deep retinal capillary layer (DRL) before multiple en face image averaging. G, Binarized image showing noise in FAZ and many vessels showing a spiny, branching-like feature (arrows). D, H, L, Optical coherence tomography angiography images of DRL after multiple en face image averaging. D, Optical coherence tomography angiography image showing continuous, smooth, and high-contrast vessel images and noise reduction. H, Binarized image showing smooth and uniform caliber after averaging. Fewer line fragments and dots are detected in nonvascular areas, including the FAZ, compared with the original image (G). L, Fewer line fragments and dots are detected in nonvascular areas than the original image (K).
Table 1. Differences in Vessel Density and Fractal Dimension between Single Image and Averaged Image in Optical Coherence Tomography Angiography Imaging Superficial Retinal Capillary Layer
Deep Retinal Capillary Layer
Parameter
Original Image
Averaged Image
P Value
Original Image
Averaged Image
P Value
Vessel density Vessel length density Vessel diameter index Fractal dimension
0.3650.022 0.0740.004 4.8180.106 1.5430.010
0.3410.019 0.0670.003 5.0620.100 1.5240.010
0.0010 <0.0001 <0.0001 <0.0001
0.3320.025 0.0720.004 4.6310.120 1.5380.012
0.3090.026 0.0620.008 4.9950.118 1.5060.015
0.0003 <0.0001 <0.0001 <0.0001
Data are mean standard deviation unless otherwise indicated.
6
Uji et al
OCTA with Multiple En Face Image Averaging
Figure 4. The relationships between quantitative measurements and number of frames used for multiple en face image averaging in optical coherence tomography angiography imaging. Vessel length density (VLD) and fractal dimension (FD) decreased as the number of frames used for averaging increased in both the superficial retinal capillary layer (SRL) and deep retinal capillary layer (DRL). A, C, In SRL, there was a statistically significant difference in VLD and FD between original images and the images derived by averaging 2 to 9 frames. B, D, In DRL, the images generated by averaging more than 3 frames showed significant differences from the original image. Meanwhile, significant differences were not found among images derived by averaging more than 6 frames. *P < 0.05, paired t test followed by Bonferroni correction. Numbers show pairs to compare.
smaller differences (0.21% in FD between 5 and 9 frames), although statistically significant, may not be clinically relevant. Establishing the minimum level of clinically relevant averaging is of critical importance because obtaining 9 acquisitions may not be practical clinically. However, by exploring the impact of averaging on measurements and measurement precision, the right level of averaging could be tailored for the specific clinical application. Regardless, the implementation of the averaging approach described in this report should be relatively easy for OCTA instrument manufacturers to implement. The concept of compounding multiple OCTA images into 1 en face image has been reported previously by Mo et al.22 However, their study focused on the optic disc and used only the superficial OCTA slab. Given the many published reports describing the use of OCTA to evaluate macular perfusion and the availability of analytic software for automated density measurements in commercial OCTA devices, our study, which focused on the macular circulation, would seem to be of clinical relevance. Our study has several additional limitations that should be considered when assessing our findings. First, our sample size is small. Thus, although differences in DRL metrics
after 6 averaging were no longer statistically significant, our study was underpowered to detect very small differences. However, such small differences are less likely to be clinically relevant. Second, we had to crop the OCTA images to obtain the final averaged images because of the manufactured logo in the lower right corner. As such, we could not make comparisons between our cohort and previously reported measurements from these 33-mm images. Third, we used the default automated segmentation boundaries to obtain the 2 OCTA en face images from different depths. Although we could not identify segmentation errors in this study of young, healthy participants, variability in segmentation from scan to scan potentially could compromise the accuracy of registration of en face images as well as the resultant quantitative metrics. In the setting of disease, one would expect that such automated segmentation differences from scan to scan may occur. However, the SRL that was used for the primary registration in our technique may be the least prone to segmentation error, because detection of the internal limiting membrane surface by commercial algorithms tends to be robust and accurate. Regardless, follow-up studies in diseased eyes are required to address this issue.
7
Ophthalmology Volume -, Number -, Month 2017 Despite these limitations, our study did demonstrate the significant changes in quantitative measurements after averaging. Multiple en face image averaging has the potential to yield quantitative measurements that are more insulated from the influence of noise or imperfect vessel delineation. In summary, multiple en face image averaging reduced VD, VLD, and FD and increased VDI. Reducing background noise and improving the continuity of retinal vessels seem to be important benefits of averaging. Future studies are needed to explore the clinical relevance of multiple en face image averaging in diseased eyes.
References 1. Nagiel A, Sadda SR, Sarraf D. A promising future for optical coherence tomography angiography. JAMA Ophthalmol. 2015;133(6):629-630. 2. Spaide RF, Klancnik Jr JM, Cooney MJ. Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography. JAMA Ophthalmol. 2015;133(1):45-50. 3. Jia Y, Tan O, Tokayer J, et al. Split-spectrum amplitudedecorrelation angiography with optical coherence tomography. Opt Express. 2012;20(4):4710-4725. 4. Savastano MC, Lumbroso B, Rispoli M. In vivo characterization of retinal vascularization morphology using optical coherence tomography angiography. Retina. 2015;35(11): 2196-2203. 5. Kim AY, Chu Z, Shahidzadeh A, et al. Quantifying microvascular density and morphology in diabetic retinopathy using spectral-domain optical coherence tomography angiography. Invest Ophthalmol Vis Sci. 2016;57(9):OCT362-OCT370. 6. Al-Sheikh M, Akil H, Pfau M, Sadda SR. Swept-source OCT angiography imaging of the foveal avascular zone and macular capillary network density in diabetic retinopathy. Invest Ophthalmol Vis Sci. 2016;57(8):3907-3913. 7. Suzuki N, Hirano Y, Tomiyasu T, et al. Retinal hemodynamics seen on optical coherence tomography angiography before and after treatment of retinal vein occlusion. Invest Ophthalmol Vis Sci. 2016;57(13):5681-5687. 8. Balaratnasingam C, Inoue M, Ahn S, et al. Visual acuity is correlated with the area of the foveal avascular zone in diabetic retinopathy and retinal vein occlusion. Ophthalmology. 2016;123(11):2352-2367. 9. Zahid S, Dolz-Marco R, Freund KB, et al. Fractal dimensional analysis of optical coherence tomography angiography in eyes with diabetic retinopathy. Invest Ophthalmol Vis Sci. 2016;57(11):4940-4947.
10. Reif R, Qin J, An L, et al. Quantifying optical microangiography images obtained from a spectral domain optical coherence tomography system. Int J Biomed Imaging. 2012;2012:509783. 11. Sander B, Larsen M, Thrane L, et al. Enhanced optical coherence tomography imaging by multiple scan averaging. Br J Ophthalmol. 2005;89(2):207-212. 12. Sakamoto A, Hangai M, Yoshimura N. Spectral-domain optical coherence tomography with multiple B-scan averaging for enhanced imaging of retinal diseases. Ophthalmology. 2008;115(6):1071-1078.e7. 13. Spaide RF, Fujimoto JG, Waheed NK. Image artifacts in optical coherence tomography angiography. Retina. 2015;35(11):2163-2180. 14. Ghasemi Falavarjani K, Al-Sheikh M, Akil H, Sadda SR. Image artefacts in swept-source optical coherence tomography angiography. Br J Ophthalmol. 2016 Jul 20. pii: bjophthalmol2016-309104. doi:10.1136/bjophthalmol-2016-309104. [Epub ahead of print]. 15. Uji A, Ooto S, Hangai M, et al. Image quality improvement in adaptive optics scanning laser ophthalmoscopy assisted capillary visualization using B-spline-based elastic image registration. PLoS One. 2013;8(11):e80106. 16. Thevenaz P, Ruttimann UE, Unser M. A pyramid approach to subpixel registration based on intensity. IEEE Trans Image Process. 1998;7(1):27-41. 17. Arganda-Carreras I, Sorzano COS, Marabini R, et al. Consistent and elastic registration of histological sections using vector-spline regularization. In: Beichel RR, Sonka M, eds. Computer Vision Approaches to Medical Image Analysis: Second International ECCV Workshop, CVAMIA 2006, Graz, Austria, May 12, 2006, Revised Papers. Berlin, Heidelberg: Springer Berlin Heidelberg; 2006:85-95. 18. Sorzano COS, Thevenaz P, Unser M. Elastic registration of biological images using vector-spline regularization. IEEE Trans Biomed Eng. 2005;52(4):652-663. 19. Brown M, Szeliski R, Winder S. Multi-image matching using multi-scale oriented patches. Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05). Volume 1. Washington, DC: IEEE Computer Society; 2005:510-517. 20. Masters BR. Fractal analysis of the vascular tree in the human retina. Annu Rev Biomed Eng. 2004;6:427-452. 21. Landini G, Murray PI, Misson GP. Local connected fractal dimensions and lacunarity analyses of 60 degrees fluorescein angiograms. Invest Ophthalmol Vis Sci. 1995;36(13):27492755. 22. Mo S, Phillips E, Krawitz BD, et al. Visualization of radial peripapillary capillaries using optical coherence tomography angiography: the effect of image averaging. PLoS One. 2017;12(1):e0169385.
Footnotes and Financial Disclosures
1
Financial Disclosure(s): The author(s) have made the following disclosure(s): S.R.S.: Consultant e Carl Zeiss Meditec, Optos, Allergan, Genentech, Alcon, Novartis, Roche, Regeneron, Bayer, Thrombogenics, Stemm Cells, Inc, Avalanche; Financial support e Carl Zeiss Meditec, Optos, Allergan, Genentech
2
Supported by the Astellas Foundation for Research on Metabolic Disorders (A.U.).
Originally received: December 10, 2016. Final revision: February 3, 2017. Accepted: February 7, 2017. Available online: ---.
Manuscript no. 2016-1009.
Doheny Image Reading Center, Doheny Eye Institute, Los Angeles, California. Department of Ophthalmology, David Geffen School of Medicine at the University of CaliforniadLos Angeles, Los Angeles, California.
Author Contributions: Conception and design: Uji, Sadda
8
Uji et al
OCTA with Multiple En Face Image Averaging
Analysis and interpretation: Uji, Balasubramanian, Lei, Baghdasaryan, Al-Sheikh, Sadda
angiography; SRL ¼ superficial retinal layer; VD ¼ vessel density; VDI ¼ vessel diameter index; VLD ¼ vessel length density.
Data collection: Uji, Balasubramanian, Lei, Baghdasaryan, Al-Sheikh Obtained funding: none
Correspondence: SriniVas R. Sadda, MD, Doheny Image Reading Center, Doheny Eye Institute, 1355 San Pablo Street, Suite 211, Los Angeles, CA 90033. E-mail:
[email protected].
Overall responsibility: Uji, Sadda Abbreviations and Acronyms: DRL ¼ deep retinal layer; FD ¼ fractal dimension; OCT ¼ optical coherence tomography; OCTA ¼ optical coherence tomography
9