Semantic Segmentation of the Choroid in Swept Source Optical Coherence Tomography Images for Volumetrics

The choroid is a complex vascular tissue that is covered with the retinal pigment epithelium. Ultra high speed swept source optical coherence tomography (SS-OCT) provides us with high-resolution cube scan images of the choroid. Robust segmentation techniques are required to reconstruct choroidal volume using SS-OCT images. For automated segmentation, the delineation of the choroidal-scleral (C-S) boundary is key to accurate segmentation. Low contrast of the boundary, scleral canals formed by the vessel and the nerve, and the posterior stromal layer, may cause segmentation errors. Semantic segmentation is one of the applications of deep learning used to classify the parts of images related to the meanings of the subjects. We applied semantic segmentation to choroidal segmentation and measured the volume of the choroid. The measurement results were validated through comparison with those of other segmentation methods. As a result, semantic segmentation was able to segment the C-S boundary and choroidal volume adequately.

image acquisition. SS-OCT (Plex Elite 9000-TM, Carl Zeiss Meditec, Inc., Dublin, California, USA) 6 × 6-mm (a 500 × 500 sampling in one volume scan) OCT images of the healthy eyes were acquired using OCT angiography mode between November 1, 2018 and March 24, 2019, after application of mydriatics (5 mg/ml tropicamide and 5 mg/ml phenylephrine sodium chloride, Santen Pharmaceutical Co. Ltd., Osaka, Japan). Two ophthalmic imaging experts took SS-OCT images. The SS-OCT instrument has a central wavelength of 1,060 nm, a bandwidth of 100 nm, A-scan depth of 3.0 mm in tissue, full-width (at half-maximal axial resolution) of about 5 µm in tissue, and lateral resolution at the retinal surface of about 20 µm. The FastTrac motion-correction software built into the SS-OCT was used during image acquisition. We used the structural OCT data only for the analysis.
Semantic segmentation. Three thousand B-scan images from six cube scans of six eyes were used as the ground truth data for training. The labeling of the choroidal boundary on the SS-OCT images was carried out by one retina specialist (Y.S.) manually (Fig. 1). The C-S boundary was defined along the extended low contrast line between the choroid and the sclera, not the outer boundary of the large choroidal vessel. The area of the choroid was defined as the area between the line beneath the retinal pigment epithelium and the C-S boundary on a B-scan image. We approached choroidal segmentation using the SegNet, which was developed by members of the Computer Vision and Robotics Group at the University of Cambridge, UK, based on the VGG19 network, www.nature.com/scientificreports www.nature.com/scientificreports/ primarily designed for large-scale natural image classification and transfer learning. This network consisted of 16 convolutional layers, with pooling as the encoder and three upsampling transposed convolutional layers as the decoder 23 . This decoder was trained to convert the features from the encoder to class labels, accompanying skip layers being used to process higher resolution features from the lower layers.
In the VGG19 network, the final layer uses a softmax function to produce a probability output. The area ratio of the choroidal area to the outside of the choroidal area on the images was about 1:9 in an SS-OCT image in the present study, so we balanced the classes using class weighting. For adapting this network to various SS-OCT images, we augmented the training images with translation and rotation. Stochastic gradient descent with momentum was used for training this network, with a momentum of 0.9, a learning rate of 0.001, and mini-batch size to use for a prediction of 1. We found that this network converged within 100 epochs. c-S boundary segmentation. A cubic smoothing spline was applied to determine the C-S boundary. The cubic smoothing spline f minimizes where n represents the number of entries of x, and the integral is over the smallest interval containing the whole entries of x. In addition, x j and y j refer to the jth entries of x and y, respectively, and D 2 f represents the second order derivative of the function f. In the present study, the smoothing parameter p was 0.01. The number was selected for smoothing the edges of the C-S boundary correctly and keeping its details. Graph cut segmentation was performed as described previously by Chiu, S. et al. 24 .

Validation of similarities. The Sørensen-Dice coefficient (DSC) among the segmentations performed with
the SegNet and other methods (the graph cut and manual segmentation) was calculated to evaluate similarities in 290 SS-OCT B scan images randomly selected from 37 eyes of 31 individuals. Three retina specialists (A.O., A.K., T.S.) segmented the choroid independently using the same method as that used for ground truth labeling; they enclosed the area of the choroid in a polygonal shape using a personal computer. The following equation was used to calculate the DSC.
A and B are the choroidal region segmented by either method respectively.

Validation of feasibility and reproducibility.
To evaluate the feasibility and reproducibility of the SegNet model, we measured the intraclass correlation coefficient (ICC) between the repeatedly measured choroidal volumes of twenty-five eyes. The volumetric data were calculated from the SS-OCT cube images. We registerd the center of the fovea in the two images taken in row using ImageJ (version 1.52n) 25 along the x-, y-, and z-axes. We cropped the original 3D choroidal volume at sizes of 250 × 500 × 500 pixels (Height × Width x Depth) to 170 × 450 × 400 pixels to remove the outside of the overlapping area.
Visualization of choroidal thickness. The choroidal thickness was plotted as a heat map to visualize the regional choroidal thickening 8,26 , and was converted using the following normalization method:

Results
Similarities in B-scan images. We compared the results of the semantic segmentation (SegNet) to manual segmentaion and that made by graph cut method (Fig. 2). The SegNet model showed a similar segmentation to the manual segmentation, whereas the graph cut method partially failed to delineate the C-S boundary. Table 1 shows the areas measured using each method. We analyzed the DSC between the segmentation by three manual graders (A, B, and C) individually, the SegNet model, and the graph cut method ( Table 2). The median of intergrader similarity between the three manual segmentations was 0.9184-0.9358. All of the DSC medians between the SegNet and manual segmentation by each grader were above 0.9000, meaning that the automated segmentation of the choroid using SegNet was not inferior to manual segmentation. The DSC median between the graph cut method and manual segmentation was low in each manual segmentation compared to that using SegNet.

Reproducibility.
We were able to measure fifty choroidal volumes of all twenty-five pairs of cube scans using SegNet. The ICC between the choroidal volumes in the repeatedly measured two images was 0.985 (95% (2020) 10:1088 | https://doi.org/10.1038/s41598-020-57788-z www.nature.com/scientificreports www.nature.com/scientificreports/ confidence interval, 0.967 to 0.993) (Fig. 3). The mean choroidal volume in the posterior pole (5.4 × 4.8 mm) was 7.4671 ± 0.8168 (standard deviation) mm 3 in twenty-five healthy eyes. We analyzed the DSC of the twenty-five pairs of choroidal volumes calculated using SegNet. The DSC median was 0.9090 (standard deviation, 0.0554; range 0.7650-0.9647). Figure 4 shows 2D choroidal thickness maps before (a) and after normalization (b). The regional changes are enhanced in the normalized thickness map.

Discussion
The median DSC among the SegNet and manual segmentation by each grader were 0.9494 (grader A), 0.9447 (grader B), and 0.9275 (grader C). The DSC medians among the manual segmentations were 0.9184 to 0.9358, and the SegNet results were almost equivalent to those of the manual segmentation ( Table 2). The high ICC value (0.985) in the choroidal volumetrics suggested good reproducibility of SS-OCT measurement and 3D segmentation.   www.nature.com/scientificreports www.nature.com/scientificreports/ SS-OCT, which has a high penetration laser light source and provides depth-enhanced images, is suitable for imaging of the choroid. 3D image analysis and choroidal 2D maps using SS-OCT images are potentially useful to explore the unknown pathogenesis of the choroid. To accomplish this purpose, the choroidal segmentation has several problems that cannot be easily overcome; low contrast of the C-S boundary, a deficit of the inner surface of the sclera, namely scleral canals for vessels or nerves, and the presence of a posterior stromal layer and lamina fusca 27 .
Enhancement of the C-S boundary in the preprocessing of OCT images has been reported to improve the success rate of boundary detection 17,18,28 . Even if the C-S boundary is completely delineated, the sclera has a pathway through which blood vessels and nerves pass 28 . The pathway in the sclera should be corrected to make the outer surface of the choroid. Interpolation in the processing of the outer surface of the choroid 17 can correct the deletion of the boundary. Since statistical image segmentation (maximally stable extremal regions) can be robust against scleral defects caused by the scleral canal, shadowing, and other elements of the choroid in the OCT images, software using these techniques has made it possible to successfully delineate the C-S boundary. However, in previous studies, the C-S boundary was delineated along the outer border of the choroidal vessels, detected using automated segmentation 17,[29][30][31] . Recently, the choroidal posterior stromal layer has led to misinterpretations of C-S boundaries that have been delineated automatically 32 . In the current study, we trained the SegNet model to identify the outer edge of the posterior stromal layer, not the C-S boundary at the posterior limit of the choroidal vessel. Therefore, the SegNet achieved choroidal segmentation with high similarity to manual segmentation under the same rule, which was superior to the graph cut segmentation.
An additional advantage of the SegNet model is that it was able to successfully delineate the boundary (Fig. 2) without denoising or edge enhancement, although the previous segmentation model required edge enhancement or inclination adjustment 17,31 .
Limitations. The present study has several limitations. First, in this series, we did not examine eyes with retinal and choroidal diseases, such as diabetic maculopathy or AMD, so we do not know whether our method can segment the choroids of such eyes effectively. If not, additional training using images of eyes with diseases can improve the segmentation results in such eyes, since the DCNN model is versatile. Second, the parameters to make 3D spline can influence the volume of the choroid, especially in eyes with an irregular C-S boundary, such as in CSC 10,33 or polypoidal choroidal vasculopathy 34 . Third, histopathological correlation to SS-OCT images  Table 2. The median similarity indices (DSC) (upper) and similarity index ranges (lower) among segmentation methods.