Automated segmentation of peripapillary retinal boundaries in OCT combining a convolutional neural network and a multi-weights graph search.

Quantitative analysis of the peripapillary retinal layers and capillary plexuses from optical coherence tomography (OCT) and OCT angiography images depend on two segmentation tasks - delineating the boundary of the optic disc and delineating the boundaries between retinal layers. Here, we present a method combining a neural network and graph search to perform these two tasks. A comparison of this novel method's segmentation of the disc boundary showed good agreement with the ground truth, achieving an overall Dice similarity coefficient of 0.91 ± 0.04 in healthy and glaucomatous eyes. The absolute error of retinal layer boundaries segmentation in the same cases was 4.10 ± 1.25 µm.


Introduction
Optical coherence tomography (OCT) provides noninvasive, structural images of eye fundus tissue based on interferometric analysis of low-coherence light [1]. By considering blood flow induced temporal variation in the signal garnered from OCT, vasculature can be distinguished from static tissue. There are many versions of this technique; collectively they are termed OCT angiography (OCTA) [2][3][4][5][6][7][8]. Measurement of retinal layer thickness from structural OCT and analysis of capillary plexuses from OCTA can both help clinical diagnosis and early detection of glaucoma, which is the leading cause of irreversible blindness globally [9][10][11][12][13]. But the clinical utility of such measurements requires accuracy and precision, both of which depend critically on the segmentation of both the optic disc boundary and peripapillary retinal boundaries. Segmentation of these anatomical regions is, then, a critically important task.
Since manual segmentation is time-consuming, several methods to segment the optic disc and peripapillary retinal boundaries have been proposed [14][15][16][17][18][19][20][21][22]. For peripapillary retinal boundaries segmentation, graph search algorithms based on intensity differences between anatomical slabs from structural OCT have been used frequently and show good results. Antony et al. proposed a 3D graph search method for the segmentation of both the optic disc boundary and the peripapillary retinal boundaries [16]. Zang et al. proposed a method which detected the optic disc boundary and segmented peripapillary retinal boundaries separately using a dynamic-programming based graph search algorithm [20]. Gao et al. proposed a method which combined the active appearance model and graph search to segment the peripapillary retinal boundaries [21]. Yu et al. proposed a shared-hole graph search method which first segments the optic disc boundary and then segments the peripapillary retinal boundaries [22]. However, speckle noise and vessel shadows both seriously detrimentally impact segmentation accuracy based just on graph search.
Nowadays, deep learning plays an important role in medical image processing and several learning-based methods exist for segmentation of OCT data [23][24][25][26][27][28]. Devalla et al. proposed a dilated-residual U-Net to segment optic nerve head tissues such as the lamina cribrosa, choroid, sclera and so on [25]. But the peripapillary retinal boundaries were not segmented in this study. Kugelman et al. proposed a retinal boundary segmentation method for macular OCT based on a combination of recurrent neural networks and graph search [26]. However, the anatomical disruption caused by the optic disc makes peripapillary retinal boundaries segmentation much more challenging than the macular region. Networks trained on macular OCT scans therefore may not generalize well to the peripapillary region.
In this study, we propose an automated segmentation method for optic disc boundary detection and peripapillary retina layer segmentation. We designed two separate neural networks and trained one each to segment the optic disc boundary and peripapillary retinal layers. The final peripapillary retinal boundaries were calculated based on the prediction and gradient maps using a multi-weights graph search algorithm.

Patient recruitment and data acquisition
In this study, 46 healthy and 63 participants with glaucoma were recruited and tested at the Casey Eye Institute, Oregon Health & Science University. The diagnoses of all the participants were made by an expert clinical examination. The participants were enrolled after informed consent in accordance with an Institutional Review Board approved protocol. The study was conducted in compliance with the Declaration of Helsinki.
The peripapillary retinal area was scanned using a commercial 70-kHz spectral-domain OCT system (Avanti RTVue-XR, Optovue Inc) with 840-nm central wavelength. The scan regions were 4.5 × 4.5 mm and 1.6 mm in depth (304 × 304 × 640 pixels) centered on the optic disc. Two repeated B-frames were captured at each line-scan location. The blood flow of each line-scan location was detected using the split-spectrum amplitude-decorrelation angiography (SSADA) algorithm based on the speckle variation between two repeated B-frames [2,29]. The OCT structural images were obtained by averaging two repeated B-frames. For each data set, two volumetric raster scans (one x-fast scan and one y-fast scan) were registered and merged through an orthogonal registration algorithm to reduce motion artifacts [30].

Neural network designing
The neural network used in this study was designed based on the architecture of the classic U-Net [31,32] (Fig. 1). Three max-pooling and (de)convolution layers were separately used in the down-sampling and up-sampling towers. Because each peripapillary retinal layer cannot be identified based just on the upper and lower boundaries, the global position in the whole retina is also an important feature. In order to capture both the relative and absolute location of each peripapillary retinal layer, a 3 × 3 normal and atrous-convolution layer [33,34] were cascaded together in each layer of the down-sampling and up-sampling towers. In addition, a global block was also designed to capture the local and global information before the final classification layer. The batch normalization [35] and exponential linear unit (ELU) function [36] were used after each convolution layer (except the output layer) to improve the stability of the final classification.
The Dice similarity coefficient (DSC) for each channel of the output map was used in the loss function: where N c is th Out n and Lab segmented by (momentum = minimum valu function kept rate of 0.1). T and peripapill The desig processing wa (R) Core (TM

Optic dis
The major cha of the optic di between eyes diametral B-f images used t

Optic d
The optic disc disc is not alw the image cen the optic disc determine its c in OCT imag volumes we demarcates th he number of fi b n are the n th y a certified = 0.9) was us ue of the loss f increasing ove he same netwo lary retinal bou gned neural ne as performed in M) i7-8700K CP sc boundary s allenge of the p isc, which is to . Because the frames were ge to train the netw disc center de c center is need ways aligned at nter (Fig. 2(A)) that leverages center. The inte es after proper designed a c he three stronge nal classes, eps channels of t grader. Stoch sed to optimiz function [37]. T er three consec ork architecture undaries. etwork was tr n MATLAB 20 PU @ 3.70GHz .
where Conv(• 1 10 − in the fir was then gen threshhold (Fi was detected generated, we ( Fig. 2(H)). T binarization (i en face image accumulation as: The optic dis obtained. Tho shadows, the c symmetry   oundaries segm ndaries between gate errors due as the combina ta set contained es (Fig. 5(A)). segmented retin f the correspon l correspondin n retinal layers ea constituted c binary image rk through a c med a multi-an irst, the bumpy e convex hull o Fig. 4).
a volumetric OCT d optic disc center metral B-frames are°. The green line wo blue lines is the (C) The generated y segmented EZ + prediction maps of A). (C) The bump sc region after the mentation netw n retinal layers e to layer disto ation of several d channels with Each input lab nal layers of th nding image. Th ng to the area ou ( Fig. 5(C)).

Results
In this study, 7 scanned. Amo glaucoma pati segmentation batch size wa glaucoma trai data. The trai used to test th between the c

Qualitati
In Fig. 7 [13,39]. An e within the sam scanned from critically imp peripapillary visualized on The superotem a green line in  Fig. 9(B)). optic disc boundary ace average projec m-to-top 3D view metric OCT scann OCT of (C). illary retinal fr es outside and ascular complex all structures b generated by p n, the segment wn in Fig. 9. The diagnosis of otably, the glau . 9(B)) and NF ss could also cl shown in rly shown er 80% of NL border orrelation OCT data , which is the radial ect can be )) [9][10][11][12][13]

Quantita
We tested 219 performance standard devi labels was 0. boundaries an 0.91 ± 0.05 in For the pe absolute error our method an healthy and gl and diagnosis compared wit deviation valu µm in glaucom  As another test of performance for the algorithm presented here, we also compared our results to those obtained with our previous method, which was based exclusively on the graph search algorithm [20]. The comparisons of the segmentation accuracy of peripapillary retinal boundaries is shown in Table 2. Through Table 2, it is clear that the segmentation accuracy and stability were both improved after combining the neural network with the classic graph search.

Neural network analysis
Inside the neural network, the addition of the atrous-convolution layer in each atrous-block and the global block greatly improved the performance of the neural networks. In order to further analyze the neural network design, we compared the validation accuracy (based on DSC) of the peripapillary retinal layers segmentation between the four architectures below: original U-Net, U-Net + global block, U-Net + cascaded atrous-block, and U-Net + global block + cascaded atrous-block (proposed) ( Table 3). Clearly, adding the cascaded atrous-convolution layers in the down and up sampling towers and global block at the end of the network critically improved the convergence of the neural network. In addition, the validation accuracies of the healthy and glaucoma data based on the inputs using only one channel (the middle one) instead of the 5 used in our algorithm were 84.11% and 83.53% respectively. These accuracies were about 2% lower than the accuracies shown in the last column of Table 3 which proved the five channels input design was effective.  Figure 10 shows example feature maps learned by the network in the normal convolution layers of the global block. It is clear that in each map the network is learning different retinal layers, as each map highlights specific layers or combinations thereof. The result of each map then yields a complete segmentation.

Discussio
The structure constitute thre optic disc nee to its unique reorientation geometry. Th which greatly addition, the guaranteed th stage, the rea same directio segmentation adjacent layer For the ne network could input data an boundaries se Though th limitations we breadth of the scan patterns to be expande learn either ne nearly the sam training data s

Conclusio
We combined optic disc and The addition o The quantifie plexuses, hav detection of g