where n and m represent，the sample size and measuring times respectively[12]. Aswe measured four times on each eye, n was calculated to be 28.46. In the agreement analysis, the formula is given as n≥log(1-β)/log(1-α). If the discordance rate (α) was 0.05 and the tolerance probability (β) was 80%, then the sample size (n) was calculated to be more than 32[13].
The en face images of SCP were duplicated into two copies and sent to two independent observers for the FAZ quantification, the sequence of which was randomized to avoid contextual bias. The Triton OCTA images of 3 mm × 3 mm slabs were imported with the original resolution of 320×320 pixels in ImageJ software. Then the FAZ boundary was outlined using Freehand Selection Tool. The FAZ metrics (area, perimeter and circularity) were measured. The manual measurement result was the average of those measured by two observers.
The en face OCTA images in grayscale were imported into an automated customized program named MATLAB that has been introduced by Tang et al.[11]. The non local means (NLM) denoising filter and the phansalkar adaptive local thresholding method were respectively applied in the image denoising and binarization. Then the FAZ was segmented by using the region growing method which started from a seed point. Finally, the FAZ area, perimeter and circularity were measured[18].
After being imported into ImageJ in 8-bit grayscale (Figure 1A), the OCTA images were processed by the Smooth method (Figure 1B), to blur the background noise but still preserve the boundary features. An initial seed (Figure 1C) located inside the FAZ is necessary to start the Level Sets, a modern image segmentation technique by use of the theory of partial differential equations (PDE) (available at https://imagej.net/Level_ Sets). The advanced active contour algorithm is more preferred than the more basic fast marching to make it less sensitive to leaking. The contour will automatically progress like a rubber band (Figure 1D), in which the parameter of curvature provides the strength of the leaking, while the parameter of convergence acts as a criterion for converging. Once it hits the border, the contour will stop and the FAZ is segmented (Figure 1E). Then the measurement results were automatically output (area, perimeter and circularity).
The method of optimizing the parameters has been described in our previous study[19]. We tried different combinations of curvature and convergence in the training dataset and evaluated the performance by analyzing the accuracy and direct visualization. The curvature =1.5 and convergence =0.0015 presented the best performance, with an average accuracy of 0.9960 and Dice coefficient of 0.9443, respectively (Figure 2, Table 1). Different values of grayscale were also tried and presented the various segmentation results especially when the noise exists near the FAZ border, which will be easily mistaken for the vessel signals. It showed that the grayscale of 30 (Dice coefficient, 0.9443) performed better than the grayscale of 10 (Dice coefficient, 0.9310) and 50 (Dice coefficient, 0.8965) (Figure 3). The macro script of the SLSM can be found in Figure 4.
The segmentation results of the first observer were served as the ground truth and compared with those performed by the second observer and the automated methods. The accuracy (ACC), sensitivity (SEN), specificity (SPE)[20], and Dice coefficient[21]were used to evaluate the segmentation performance and calculated based on the following formulas: ACC=(TP+TN)/ (FN+FP+TP+TN), SEN=TP/(TP+FN), SPE=TN/ (TN+FP), Dice=2TP/(2TP+FP+FN), where TP=true positive, TN = true negative, FP=false positive, and FN = false negative. The values of the Dice coefficient among three different methods were compared using the one way ANOVA/Kruskal-Wallis test.
In our study, 35 eyes of 35 healthy subjects were included, with the mean ages of 24.69±2.52 years (range, 20 to 35 years) and mean spherical equivalent of ?2.25±1.93 D (range, ?5.50 to 0.75 D). A total of 140 OCTA images were analyzed in the test dataset, with the mean signal strengths of 72.41±3.11 (range, 60 to 79).
The agreement analyses between automated and manual methods were shown (Table 4, Figures 6-8). For the FAZ area, although there was a statistical difference in the measurement results of two observers (P=0.013), the interobserver agreement was excellent (ICC =0.992). Both MATLAB (ICC =0.968) and SLSM ((ICC =0.973) showed excellent agreement with the manual methods. For the FAZ area, Bland-Altman Plots showed agreement ranging from ?0.021 to 0.033 for the manual method, ranging from ?0.056 to 0.009 for MATLAB, ?0.055 to 0.016 for SLSM. For the FAZ perimeter, the interobserver agreement was excellent (ICC =0.904). The agreement of SLSM with manual methods was good (ICC =0.837) while that of MATLAB was only moderate (ICC =0.554). For the FAZ circularity, all methods showed moderate agreement with manual methods (manual, ICC =0.716; SLSM, ICC =0.520; MATLAB, ICC =0.737).
In our study, we investigated the feasibility of the SLSM, a free and open-source plugin used for the automated FAZ metrics on Triton OCTA images in healthy subjects. The SLSM showed excellent repeatability and agreement with the manual methods and performed better with a higher Dice coefficient than than MATLAB did.
Although manual methods have demonstrated excellent repeatability and reproducibility in various OCTA devices, they will waste a lot of time and labor. In this case, automated FAZ metrics will fit our need especially when a large number of images need to be analyzed, but the validation of reliability is required before being applied into the clinical practice. Linderman et al.[27] have segmented the FAZ on Optovue OCTA using the AngioVue semiautomatic nonflow measurement tool in healthy eyes. Their study showed that the reliability of all area measurements was excellent (ICC =0.994 manual, 0.969 semiautomatic), while manual segmentation had better repeatability (0.020 mm2 ) than semiautomatic did (0.043 mm2 ). Lim et al.[28] have evaluated the inbuilt algorithm in the Zeiss Cirrus 5000 (AngioPlex? OCTA software) and showed good repeatability with the value of ICC more than 0.75 for automated FAZ metrics. But the agreement with manual measurements has not been given in this study. Our prior study also assessed the reliability of this embedded algorithm in Cirrus 5000 OCTA. Using a systematic way, we found that the Cirrus inbuilt algorithm outlined the border of FAZ wrongly in 22.9% of cases, and the agreement with manual measurements was poor for all FAZ metrics[9]. Besides, some customized algorithms used for the automated FAZ metrics have been reported. Ishii et al.[29] have introduced a macro-based method named the KannoSaitama macro (KSM) for the FAZ area measurement in the Zeiss PLEX Elite 9000, proving that it was feasible and yielded results comparable to those obtained by manual measurement. Díaz et al.[30] have investigated a fully automated system used in Triton OCTA images, which provided accurate results both for healthy and diabetic eyes. But they only reported the agreement with the manual measurements, but not for the repeatability.
Tang et al.[11] have proposed a customized automated program named MATLAB for the superficial capillary network quantification on Triton OCTA images in diabetic eyes. They evaluated the repeatability of MATLAB and reported a lower ICC value of FAZ area (ICC =0.976) than our study’s (ICC =0.983), while that of FAZ circularity (ICC =0.751) was higher than ours (ICC =0.670). Fang et al.[31] have investigated MATLAB in the healthy eyes and found that the repeatability of MATLAB for the FAZ area and perimeter measurements both in the left and right eyes was excellent, while for the FAZ circularity measurement was good (ICC, ranged from 0.969 to 0.996). But they have not reported the agreement between the automated and manual methods. MATLAB was also used for the FAZ quantification in glaucoma patients, though the reliability analyses were not performed in these studies[32-33]. In our study, the SLSM showed better repeatability than MATLAB for all FAZ metrics with higher ICC values. The agreement of the SLSM with manual methods for the FAZ area and perimeter was also better than MATLAB. Though MATLAB was proved to be feasible in Triton OCTA images of both healthy and pathologic eyes, it is not a free and open-source program, thus making it difficult for us to obtain.
Our previous study[19] has introduced the Level Sets macro (LSM) in the FAZ quantification on the Zeiss Cirrus HD-OCT 5000 system, which provided results comparable to those for manual measurement. Different from Cirrus 5000 OCTA images, the background noise in Triton OCTA images was apparent and will affect the detection of the FAZ boundary[34]. MATLAB utilized a non-local means (NLM) denoising filter on the grayscale images to reduce the background noise and improve the signal-to-noise ratio[11]. In our study, the OCTA images were only processed by the Smooth method, which will blur the background noise but the boundary features are still preserved. It can also be written in the macro language and automatically run in the SLSM program, which is more convenient and efficient than MATLAB.
There exists some limitations in our study. First, the reliability of the SLSM has not been evaluated in those eyes with ocular diseases. Secondly, we only investigated the 3 mm × 3 mm macular scanning mode, and other scanning modes have not been accessed yet. Thirdly, the feasibility of this program in other OCTA systems needs further investigation.
In conclusion, the SLSM exhibits better accuracy than MATLAB did and shows excellent repeatability and agreement with manual measurement. This free and open-source program may be a feasible and accessible option for automated FAZ quantification of Triton OCTA images.
