Open Access
Add to BibSonomy
Abstract
Fovea serves to be one of the crucial landmarks of the retina. The automatic detection of the foveal center in optical coherence tomography (OCT) images helps in diagnosing retinal diseases. However, challenges arise due to retinal structure damage and the demand for high time performance. In this study, we propose a fast and robust fovea detection framework for OCT and OCT angiography (OCTA) images. We focus on detecting the foveal center based on the foveal avascular zone (FAZ) segmentation. Firstly, the proposed framework uses a lightweight neural network to quickly segment the FAZ. Further, the geometric center of the FAZ is identified as the position of the foveal center. We validate the framework’s performance using two datasets. Dataset A contains two modalities of images from 316 subjects. Dataset B contains OCT data of 700 subjects with healthy eyes, choroidal neovascularization, geographic atrophy, and diabetic retinopathy. The Dice score of the FAZ segmentation is 84.68%, which is higher than that of the existing algorithms. The success rate (< 750 µm) and distance error of fovea detection in OCTA images are 100% and 92.3 ± 90.9 µm, respectively, which are better than that in OCT. For different disease situations, our framework is more robust than the existing algorithms and requires an average time of 0.02 s per eye. This framework has the potential to become an efficient and robust clinical tool for fovea detection in OCT images.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Optical coherence tomography (OCT) is a non-invasive imaging technology for acquiring high-resolution, three-dimensional (3D) cross-sectional images of the retina, and it is one of the most important ancillary tools for the diagnosis and management of macular diseases [1]. OCT allows a detailed in-vivo analysis of the interior of the retina, especially the fovea [shown in Fig. 1(a)], which supports the highest visual acuity. The foveal center is also a key reference landmark on the retina. For example, the early treatment diabetic retinopathy study (ETDRS) grid generated from the foveal center is an important tool for diagnosing retinal diseases [2]. Therefore, the accurate detection of the foveal center is of great significance for disease evaluation and diagnosis.
Fig. 1. The anatomy of the fovea and structure of OCT (a), a normal case of the fovea (b) and the challenges of fovea detection: vanished layer boundaries (c), abnormal retinal thickness (d), irregular foveal shape in B-scan (e).
The foveal center in healthy eyes was identified as the deepest point of the foveal pit, as shown in Fig. 1(b). However, technically implementing the automatic detection of the foveal center is extremely challenging for OCT images, due to the following complexities and variabilities[Figs. 1(c)–1(e)]: (1) vanished layer boundary, (2) abnormal retinal thickness due to edema, and (3) irregular foveal shape in B-scan.
Each of these characteristics poses the following particular challenges for fovea detection methods: (1) The disappearance of layer boundaries makes it difficult to locate the foveal center using the thinning and confluence of retinal layers near the fovea [3–6]. (2) The method of locating fovea using the regions of thinning retinal thickness is only suitable for healthy eyes and lacks universal applicability for cases with abnormal retinal thickness due to edema [7–9]. (3) Due to the diversity of retinal diseases, it is difficult to classify the fovea based on their different shapes [10], as shown in Fig. 1(e). (4) Liefers et al. [11] used a fully convolutional neural network (CNN) to classify the fovea; however, this framework locates few error regions that are similar to the foveal shape without regional restrictions. More importantly, it is inefficient in predicting the fovea pixel by pixel in 3D volume data of OCT. As fovea detection is the premise of the retinal indicator quantification process [2,12,13], high time performance requirements are put forward for automatic fovea detection. In short, there is still an urgent need for a fast and robust fovea detection framework in clinical practice.
In OCT projection maps, we can notice that there are few vessels around the fovea, which are called the foveal avascular zone (FAZ) [14] (marked as yellow circles in Fig. 1). In particular, the appearance of optical coherence tomography angiography (OCTA) makes FAZ easier to be observed [15–17]. We can describe the boundary of FAZ more accurately through OCTA projection maps (in Fig. 1). We also observe that the location relationship between the FAZ region and foveal center is stable in most of the cases. Based on these observations, we can conclude that the FAZ can be used to locate the foveal center, and a robust fovea detection framework can be designed.
In our fovea detection framework, we generate two-dimensional (2D) projection maps from 3D volume data of OCT and OCTA. Further, a lightweight U-Net with multi-scale dilated convolution is designed for FAZ segmentation in OCT and OCTA projection maps. Finally, the probability map of the FAZ region is obtained as the output from the proposed network, and the geometric center of the FAZ region is calculated as the foveal center.
We validate the performance of our framework using the following two aspects. (1) The first being the types of modality. A dataset containing 316 OCT volumes and their corresponding OCTA volumes are used to explore the applicability of our framework in two image modalities, OCT and OCTA. (2) The second being the types of retinopathy. An OCT dataset containing 700 subjects with four definite disease labels was used to test the robustness of our framework under different disease situations.
This research was approved by the institutional Human Subjects Committee and followed the tenets of the Declaration of Helsinki. Our contributions in this paper can be highlighted as follows: (1) We firstly use the FAZ to locate the foveal center. (2) We propose a novel lightweight CNN to achieve highly-accurate FAZ segmentation. (3) The proposed framework can be applied to both OCT and OCTA images. (4) The proposed framework is robust to several disease cases. (5) The proposed framework is efficient and more accurate than other reported algorithms. (6) The proposed framework does not depend on retinal layer segmentation.
2. Methods
2.1 Overview
As mentioned above, the proposed framework is based on the FAZ segmentation, and it calculates the geometric center of the FAZ to locate the foveal center. The overview of the proposed framework is schematically described in Fig. 2.
In this framework, the FAZ segmentation in OCT and OCTA is different as it is difficult to get the accurate FAZ boundary from OCT images alone. Therefore, a dual training process is proposed for the FAZ segmentation in OCT images. This process will be described in Section 2.3.
2.2 Image pre-processing
The data collected are 3D OCT and OCTA volumes. Image pre-processing mainly generates projection maps from 3D volume data for the FAZ segmentation. For OCT, the projection map generated is the average projection (Fig. 1). For OCTA data, the projection map generated is the maximum projection between the internal limiting membrane layer (ILM) and outer plexiform layer (OPL) (Fig. 1). These are the mainstream projection methods.
The size of the OCT volumes in the two datasets and the size of projection maps are different. For Dataset A, the size of the generated projection maps is 400 px ×400 px, and it corresponds to the actual size of 6 mm × 6 mm in the retina. For Dataset B, the size of the projection maps is 512 px ×128 px, and it corresponds to the actual size of 6 mm × 6 mm in the retina. To keep the resolution of each direction of the image consistent, we use a bilinear interpolation to stretch the projection maps in Dataset B to a size of 512 px ×512 px.
2.3 Lightweight U-Net for FAZ segmentation
A large number of network structures have been applied to medical image segmentation, such as FCN [18] and U-Net [19]. However, the fovea detection algorithm is the premise of some retinal indicator quantification algorithms, and it has a high demand for time performance. To improve the time performance of the current framework, a lightweight U-Net structure with fewer parameters was proposed to achieve a fast segmentation of the FAZ.
2.3.1 Network architecture
The proposed network architecture is shown in Fig. 3. Compared with U-Net [19], this network has fewer convolution layers and channels. This setting is related to the scale of the data and complexity of the problem. Further, fewer network parameters are more conducive to network training. At the bottom of the network, we increase the pool kernel size to reduce the size of the feature maps, and use multi-scale dilated convolution at the bottom layers to increase the receptive field of the network. These settings allow the bottom of the network to have a keener insight into the location of the FAZ, while the top pays more attention to the boundary of the FAZ. In addition, the image is not cropped and resized in the entire architecture, thus the size consistency of the upsampling and downsampling portion is maintained. Further, maintaining the consistency of image position information is conducive to the positioning task. Moreover, the network can allow different input sizes in different modalities. Specifically, the input and output sizes of dataset A are 400 px × 400 px, while the input and output sizes of dataset B is 512 px × 512 px.
2.3.2 Training pattern
In OCTA projection maps, the FAZ boundaries can be outlined and the ground truth for training the FAZ segmentation network can be obtained (Fig. 4 - top), and the labeling method is introduced in Section 2.5. In OCT projection maps, describing the exact boundary of the FAZ is difficult; thus, making it difficult to obtain its ground truth. To ensure the accuracy of the FAZ region labeling in OCT projection maps, we design a dual training process to segment the obvious FAZ target and obvious background areas (as shown in Fig. 4 - bottom).
Fig. 4. Network training patterns of the FAZ segmentation in OCTA images (top) and OCT images (bottom).
For target segmentation, we draw a circle with radius r = 30 px, which is less than the average radius of the FAZ, as shown in Fig. 4(c). The inside of the circle is labeled as 1, while the outside of the circle is labeled as 0. We use these labels to train the first network and get the target’s probability map from the output of the network as ${\mathcal T}$. For background segmentation, we draw a circle with radius R = 70 px, which is more than the average radius of the FAZ, as shown in Fig. 4(e). The inside of the circle is labeled as 0, while the outside of the circle is labeled as 1. We use these labels to train the second network and get the background’s probability map from the output of the network as ${\mathcal B}$. The target probability map ${\mathcal T}$ and the background probability map ${\mathcal B}$ have complementary characteristics. ${\mathcal T}$ usually contains few misidentified areas [Fig. 4(d)], but its small radius is significantly helpful in determining the foveal center. The recognition of the background using ${\mathcal B}$ is highly accurate, which can correct the wrong segmentation in ${\mathcal T}$. However, the non-background region has a larger radius [Fig. 4(f)], which is not conducive for the determination of the foveal center. To combine the advantages of ${\mathcal T}$ and ${\mathcal B}$, the FAZ region ${\mathcal F}$ is calculated by using the following formula:
where $\vert\vert{\ast}\vert\vert _\alpha $ represents threshold binarization and ${\alpha \; }$( = 0.5) is the selected threshold.2.4 Fovea detection
After the FAZ segmentation, we calculate the geometric center of the FAZ and regard it as the final position of the foveal center ${\mathcal P},$ described as follows:
Fig. 5. An example of fovea detection. (a) FAZ segmentation result. (b) OCT projection map. (c) B-scan with fovea.
2.5 Dataset and ground truth
We validated the performance of our framework using two datasets: (1) Dataset A (including OCT and OCTA) is mainly used to verify the performance of our framework in different modalities. (2) Dataset B (including OCT only) is mainly used to verify the performance of our framework in different disease situations. Table 1 lists the pathology distribution in the two datasets.
Table 1. The number of volumes with different pathologies in Dataset A and Dataset B.
Dataset A includes 316 OCTA volumes and their corresponding OCT volumes. The data were collected using a commercial 70 kHz spectral-domain OCT system with a center wavelength of 840 nm (RTVue XR, Optovue, CA). Each volume has a size of 640 px × 400 px × 400 px corresponding to a 2 mm × 6 mm × 6 mm volume centered at the retinal macular region. The diseases in Dataset A mainly include age-related macular degeneration (AMD), diabetic retinopathy (DR), central serous chorioretinopathy (CSC), choroidal neovascularization (CNV), and retinal vein occlusion (RVO). Other diseases, which are not listed, include retinal detachment, optic atrophy, retinitis pigmentosa, retinoschisis, etc.
Dataset B includes 700 OCT volumes with CNV, DR, geographic atrophy (GA), and normal retina. The OCT cubes are 1024 px × 512 px × 128 px in size corresponding to a 2 mm × 6 mm × 6 mm volume centered at the retinal macular region generated by a Cirrus HD-OCT device (Carl Zeiss Meditec, Inc.).
The ground truth includes the following two parts:
- 1) The FAZ label. The FAZ in the OCT images is determined by drawing circles using an automatic algorithm as described in Section 2.3.2. In the OCTA images, the FAZ has a relatively clear boundary and can be manually labeled. An expert marked the FAZ in the OCTA projection image. The criteria for labeling are finding an area without blood flow signals around the foveal center and drawing its maximum closed loop as shown in Fig. 6.
- 2) The location of the foveal center. The location of the foveal center is labeled by two experts according to the following criteria: (i) In a healthy retina, the foveal center can be defined as the deepest point of retinal depression, as shown in Fig. 7(a). (ii) In the case of retinal diseases, by examining the convergence of the retinal layer structure, the thinnest position of the inner retinal layer is found and identified as the foveal center, as shown in Fig. 7(b). (iii) If the retinal layer structure is severely damaged by edema, as shown in Fig. 7(c), it is difficult to determine the position of the fovea from the OCT B-scan images. In this case, we first find the area of the FAZ according to the blood vessel distribution in the OCT projection map, and then, determine the foveal center based on the center of the FAZ, as shown in Fig. 7(d).
Fig. 6. The ground truth labeling of the FAZ. (a) 3D visualization of OCTA volume. (b) OCTA maximum projection between ILM and OPL. (c) The ground truth of the FAZ.
Fig. 7. The ground truth labeling of the foveal center. (a) B-scan image of a healthy retina. (b)-(c) B-scan images of the retina with different diseases. (d) OCT projection map between OPL and BM layer.
2.6 Evaluating indicator
We use the following five indexes to evaluate the FAZ segmentation results quantitatively:
where TP is true positives, TN is true negatives, FP is false positives, and FN is false negatives.We use two indexes to evaluate the fovea detection results quantitatively. (1) Accuracy (Acc.) is defined as the percentage of cubes whose detection results are 750 µm (radius of fovea) less than the ground truth. (2) Distance error (Dis.) gives the Euclidean distance between the automatic positioning result and the expert’s labeling.
2.7 Baselines
To further evaluate the performance of our framework, several baselines were considered in this study. We implemented three baselines for the FAZ segmentation. The first baseline, proposed in [16], is a method based on an active shape model, which is generalized gradient vector flow (GGVF). This method has several parameters that need to be adjusted according to the different images, and we adjusted them based on our datasets. The second baseline [20] is based on edge detection and morphological processing. This method has good applicability in the OCTA projection maps of different devices, but it has poor adaptability to lesions areas, such as non-perfusion areas. We also considered a deep learning method (FCN [18]) as the baseline. In this implementation, we adopted the same parameters as in the case of the proposed lightweight U-Net, as described in the next section, which only differed in terms of network structure. The codes for these methods are publicly available.
Two baselines are considered to compare the accuracy of foveal detection. A thickness-based method [8] uses layer segmentation to generate a retinal thickness map, and then, uses a saliency detection method for fovea detection. This method is generally applicable to healthy retinal images. We obtained the original code and tested it on our dataset. A deep learning method [11] introduced a fully CNN structure to classify each pixel block of the OCT B-scan into two classes, fovea and background. Further, after being processed by a Gaussian filter, the maximum response is considered as the foveal center. This framework is implemented in our datasets. The parameter settings are the same as in [11]. This framework has high accuracy but is not quite fast. The performance differences with our framework are analyzed in Section 3.3.
2.8 Implementation details
We implemented the proposed framework with 3-fold cross-validation on two datasets. Each dataset was randomly divided into three parts, two parts as the training set and one part as the test set. Each group of experiments was conducted three times to obtain the test results of all the data. In the training stage, we used the Adam stochastic optimization of the TensorFlow framework. We ran 3000 training iterations on 2 NVIDIA GeForce GTX 1080 Ti GPUs. We used cross-entropy loss function with a batch size of 8 and initial learning 10−4. A standard normal initialization method is used to initialize the network with a variance of 0.02. The network does not use dropout and batch normalization methods. The training process takes approximately 0.5 h, and the testing speed is approximately 0.01 s/eye.
3. Experiments and analysis
3.1 Experiment I: performance in different modality types
To explore the foveal center detection performance of the proposed framework in different modalities, we implemented 3-fold cross validation in Dataset A, which contained OCTA volumes and corresponding OCT volumes from 316 eyes. The quantitative results are listed in Table 2. It shows that the accuracy of using the OCTA projection for fovea detection can reach 100%, which is much higher than that achieved when using the OCT image. Figure 8. shows four results of the FAZ segmentation and foveal center localization. As the FAZ region has clear boundaries in OCTA projection maps, the segmentation and foveal center localization results are more accurate than that of OCT images. For OCT projection maps, the FAZ region has a blurred boundary, but the result of FAZ segmentation is quite acceptable, and the result of foveal center localization is still accurate.
Fig. 8. Examples of the FAZ segmentation and foveal center detection in OCTA projection maps (a) and OCT projection maps (b). Green line represents the FAZ segmentation result of our network. Red line represents the ground truth of the FAZ region. Green dot represents our foveal center detection result. Red dot represents the experts’ labeling.
Table 2. The accuracy and distance error of our framework for foveal center detection.
It is assumed that the success of the proposed framework in both the modalities relies on the following facts: (1) The position relationship between the FAZ region and surrounding large vessels is relatively fixed. Our network trains the projection maps and can accurately grasp the global semantic information. (2) The location relationship between the FAZ center and foveal center is consistent. It is a reliable method to determine the foveal center based on the center of the FAZ region. Therefore, our framework can recognize FAZ accurately and lead to a more robust fovea localization. In the next experiment, we will explore the performance of our algorithm in different disease cases.
3.2 Experiment II: performance in different retinopathy types
To explore the performance of the proposed framework in different retinopathy types, we perform the experiments in Dataset A and Dataset B, and we list the quantitative results of different retinopathy types in Table 2. From the accuracy perspective, DR has a low accuracy in both the datasets while using OCT images, which is related to the effect of the disease on the OCT projection map. Further, an example with DR having a relatively poor image quality caused due to the DR disease is shown in Fig. 9(a). When using the OCTA image, the accuracy of the eye has significantly improved, as shown in Fig. 9(b). It also indicates that DR is often accompanied by the appearance of the non-perfusion areas, which is similar in characteristics to the FAZ. Although the presence of the non-perfusion areas reduces the accuracy of the FAZ segmentation, the final positioning result is still within an acceptable range. This is because the non-perfusion areas at the edge of the image have not been misclassified into the FAZ. From the point of distance error, the different disease conditions correspond to the different error ranges. In particular, the distance error of fovea detection in GA disease is the largest in Dataset B. Figure 9(c). shows an example of the proposed algorithm failing in GA. The GA directly affects the gray distribution of the projection map. It further affects the FAZ segmentation, which leads to the failure of foveal center detection.
Fig. 9. Two relatively poor results of the fovea detection in DR (a) (b) and GA (c). (a) and (c) are the OCT projection maps. (b) is the OCTA projection map. The yellow areas represent the FAZ segmentation results. The green circles represent the fovea detection results and the red circles represent the ground truths.
We also observed the performance of our framework when the lesion destroyed the shape of foveal depression. In this study, we classify the OCT images into five categories according to the morphological differences of the fovea. Figure 10 shows the fovea detection results of the five types of images. Our framework, which is based on the FAZ segmentation, does not rely on the foveal shape. Thus, it can detect the foveal center robustly even in the cases of retinal edema and choroidal thickening. The above facts have proved that our framework is suitable for different disease cases.
Fig. 10. The fovea detection results in five types of fovea shapes. The green line indicates the FAZ segmentation results. The green and red circles indicate the foveal center detection results and the ground truth, respectively. The method performs well in case of A: normal; B: fibrosis; C: GA; D: absent or minor foveal depression; E: large edema.
3.3 Experiment III: performance comparison with others
To further illustrate the effectiveness of our framework, it was compared against several baselines. The proposed framework is based on the FAZ segmentation, and the segmentation results of the FAZ regions play a key role in determining the final location of the fovea. Therefore, we first explore the performance of the FAZ segmentation in our network in comparison with two published FAZ segmentation methods [16,20]. The quantitative results are listed in Table 3. The Dice coefficient of the FAZ segmentation in the OCTA projection map is 0.84, which is higher than that of the other two reported methods. We also compared our method with FCN [18] (Table 3). Our network performance was found to be better than that of the FCN network architecture, which is closely related to the structure of multi-scale dilated convolution and the adjustment of pooling layer to increase the network receptive field.
Table 3. Quantitative comparison of different FAZ segmentation methods.
Our network has achieved high performance for FAZ segmentation. Further, we compared the performance of our fovea location method with that of other such methods. A thickness-based method [8] and a deep learning method [11] were applied on Dataset B. The quantitative results listed in Table 4 indicate that our method performs better than the thickness-based method [8] in all the four categories and better than the deep learning method [11] in case of different disease situations. The method [11] classifies all the pixels in OCT volume, which makes it more accurate when locating the fovea in normal cases. However, due to technical limitations, the method [11] cannot utilize an entire OCT volume to train the neural network. The lack of global information and complexity of the foveal structure in the case of diseases reduces the detection accuracy. Our method uses a projection image, which contains ample global information about the retina, for improving the accuracy of fovea detection in disease cases.
Table 4. The accuracy and distance error of three methods in different disease cases for foveal center detection.
Our framework also has a distance error similar to manual markings by human experts. We calculate the foveal center location results between our automatic framework and the manual markings by two experienced physicians on Dataset B, as listed in Table 5. The result of the proposed algorithm shows a higher consistency when compared to that of the experienced physicians, and the stability of our algorithm is better than that of the physicians.
Table 5. The distance between our results and manual marking (two experts) for 700 cubes.
Most importantly, our framework is quite fast. As the foveal center location is the base of OCT data processing and analysis, the time performance of the algorithm is strictly required. Our algorithm only takes 0.02 s for each OCT volume, which makes it faster than the other algorithms, as listed in Table 6. The method described in [8] needs retinal thickness maps, generation of which needs retinal layer segmentation. However, layer segmentation is often time-consuming. In this study, we have used layer segmentation software (OCTExplorer 3.8.0) to segment the ILM and BM layers, which cost more than 2 min for each OCT volume. Although the method [8] takes little longer to locate the foveal center on the thickness map, it is still a time-consuming method if the time of layer segmentation is taken into account. The method described in [11] classifies the 3D volume data at the pixel level, which consumes ample space and time. Our framework does not need the retinal layer segmentation and uses a lightweight network to detect the fovea in projection maps; thus, achieving a faster speed.
Table 6. The run time of different methods for foveal center detection.
4. Conclusion
In this paper, we presented a fast and robust automatic foveal center detection framework, which is based on the FAZ segmentation and applies to both OCT and OCTA images. The experiments conducted on two datasets showed that the proposed framework could achieve better performance than the existing methods in several retinal disease cases, and had a higher consistency compared to the experienced physicians. The proposed framework has the potential of becoming an efficient and accurate clinical tool in fovea detection.
Funding
National Natural Science Foundation of China (61671242, 61701222); Key Research and Development Program of Jiangxi Province (BE2018131); Suzhou Industrial Innovation Project (SS201759).
Disclosures
The authors declare no conflicts of interest.
References
1. W. Geitzenauer, C. K. Hitzenberger, and U. M. Schmidt-Erfurth, “Retinal optical coherence tomography: Past, present and future perspectives,” Br. J. Ophthalmol. 95(2), 171–177 (2011). [CrossRef]
2. P. Massin, A. Erginay, B. Haouchine, A. B. Mehidi, M. Paques, and A. Gaudric, “Retinal thickness in healthy and diabetic subjects measured using optical coherence tomography mapping software,” Eur. J. Ophthalmol. 12(2), 102–108 (2002). [CrossRef]
3. A. Lang, A. Carass, M. Hauser, E. S. Sotirchos, P. A. Calabresi, H. S. Ying, and J. L. Prince, “Retinal layer segmentation of macular OCT images using boundary classification,” Biomed. Opt. Express 4(7), 1133 (2013). [CrossRef]
4. R. Kafieh, H. Rabbani, F. Hajizadeh, M. D. Abramoff, and M. Sonka, “Thickness mapping of eleven retinal layers segmented using the diffusion maps method in normal eyes,” J. Ophthalmol.2015, (2015).
5. S. J. Chiu, X. T. Li, P. Nicholas, C. A. Toth, J. A. Izatt, and S. Farsiu, “Automatic segmentation of seven retinal layers in SDOCT images congruent with expert manual segmentation,” Opt. Express 18(18), 19413 (2010). [CrossRef]
6. S. J. Chiu, M. J. Allingham, P. S. Mettu, S. W. Cousins, J. A. Izatt, and S. Farsiu, “Kernel regression based segmentation of optical coherence tomography images with diabetic macular edema,” Biomed. Opt. Express 6(4), 1172 (2015). [CrossRef]
7. A. Montuoro, S. M. Waldstein, B. S. Gerendas, and U. S. Erfurth, “Joint retinal layer and fluid segmentation in OCT scans of eyes with severe macular edema using unsupervised representation and auto-context,” Biomed. Opt. Express 8(3), 1874–1888 (2017). [CrossRef]
8. S. Niu, Q. Chen, L. de Sisternes, T. Leng, and D. L. Rubin, “Automated detection of foveal center in SD-OCT images using the saliency of retinal thickness maps,” Med. Phys. 44(12), 6390–6403 (2017). [CrossRef]
9. F. Wang, G. Gregori, P. J. Rosenfeld, B. J. Lujan, M. K. Durbin, and H. Bagherinia, “Automated detection of the foveal center improves SD-OCT measurements of central retinal thickness,” Ophthalmic Surg. Lasers Imaging 43(6), S32–S37 (2012). [CrossRef]
10. J. Wu, S. M. Waldstein, A. Montuoro, B. S. Gerendas, G. Langs, and U. Schmidt-Erfurth, “Automated Fovea Detection in Spectral Domain Optical Coherence Tomography Scans of Exudative Macular Disease,” Int. J. Biomed. Imaging 2016, 1–9 (2016). [CrossRef]
11. B. Liefers, F. G. Venhuizen, V. Schreur, B. van Ginneken, C. Hoyng, S. Fauser, T. Theelen, and C. I. Sánchez, “Automatic detection of the foveal center in optical coherence tomography,” Biomed. Opt. Express 8(11), 5160 (2017). [CrossRef]
12. W. D. Vogl, S. M. Waldstein, B. S. Gerendas, U. Schmidt-Erfurth, and G. Langs, “Predicting macular edema recurrence from spatio-temporal signatures in optical coherence tomography images,” IEEE Trans. Med. Imaging 36(9), 1773–1783 (2017). [CrossRef]
13. P. K. Yu, Z. Mammo, C. Balaratnasingam, and D. Yu, “Quantitative Study of the Macular Microvasculature in Human Donor Eyes,” Invest. Ophthalmol. Visual Sci. 59(1), 108–116 (2018). [CrossRef]
14. J. Conrath, R. Giorgi, D. Raccah, and B. Ridings, “Foveal avascular zone in diabetic retinopathy: Quantitative vs qualitative assessment,” Eye 19(3), 322–326 (2005). [CrossRef]
15. K. G. Falavarjani, H. Shenazandi, D. Naseri, P. Anvari, K. Pegah, A. Farzaneh, and A. Feras, “Original Article Foveal Avascular Zone and Vessel Density in Healthy Subjects: An Optical Coherence Tomography Angiography Study,” J. Ophthalmic Vision Res. 13(3), 260–265 (2018). [CrossRef]
16. Y. Lu, J. M. Simonett, J. Wang, M. Zhang, T. Hwang, M. Ahmed, D. Huang, D. Li, and Y. Jia, “Evaluation of Automatically Quantified Foveal Avascular Zone Metrics for Diagnosis of Diabetic Retinopathy Using Optical Coherence Tomography Angiography,” Invest. Ophthalmol. Visual Sci. 59(6), 2212 (2018). [CrossRef]
17. L. Giselle, J. S. A. Romo, R. E. Linderman, B. D. Krawitz, S. Mo, A. Zakik, and J. Carroll, “Within-subject assessment of foveal avascular zone enlargement in different stages of diabetic retinopathy using en face OCT reflectance and OCT angiography,” Biomed. Opt. Express 9(12), 5982–5996 (2018). [CrossRef]
18. J. Long, E. Shelhamer, and T. Darrell, “Fully Convolutional Networks for Semantic Segmentation,” IEEE Trans. Pattern Anal. Mach. Intell. 39, 3431–3440 (2014). [CrossRef]
19. O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” Int. Conf. Med. Image Comput. Comput. Interv. (2015).
20. M. Díaz, J. Novo, P. Cutrín, F. Gómez-Ulla, M. G. Penedo, and M. Ortega, “Automatic segmentation of the Foveal Avascular Zone in ophthalmological OCT-A images,” PLoS One (2018).
