Detection of retinal pigment epithelium detachment from OCT images using multiscale Gaussian filtering

Abstract

Macular diseases, including neovascular age-related macular degeneration (nvAMD), are leading causes of irreversible blindness and visual impairment. One prominent feature of nvAMD is the detachment of the retinal pigment epithelium. The aim of this study is to implement an automated method for the segmentation of the pigment epithelial detachment (PED) using optical coherence tomography (OCT). OCT datasets from 8 patients with nvAMD were acquired during multiple sessions. At each session, 17 images with a resolution of 1020 $\times$ 640 pixels were obtained. The images were segmented using Gaussian filtering and template matching for the detection of the upper and lower border of the PED, respectively. The results of the method were compared with the ones obtained from the manual segmentation of the images by an expert. Four well-known metrics were used to evaluate the performance of the method with respect to the manual segmentation, resulting in high scores of consistency. Furthermore, the proposed method was also compared with four other well-known methods providing similar or superior performance.

Keywords

Optical coherence tomography (OCT)retinal pigment epithelium detachment segmentation Gaussian filtering

1. Introduction

The neovascular age-related macular degeneration (nvAMD) is a progressive disorder leading to severe visual impairment or even vision loss. Loss of vision in patients diagnosed with age-related macular degeneration (AMD) occurs due to the development of geographic atrophy (GA) and choroidal neovascularization (CNV) [1, 2, 3]. In those cases, it has been proven that the admission of anti-vascular endothelial growth factor (anti-VEGF) agents has dramatically improved the clinical condition of nvAMD patients [4]. As a result, those anti-VEGF therapeutic schemes have become the standard treatment for nvAMD. Despite this fact, there are limitations associated with anti-VEGF based treatments. The primary reason for this is that anti-VEGF therapeutic schemes require frequent clinical assessment in order to monitor the response of the patient to treatment. Moreover, in some cases, monthly intravitreal injections are required for amending the selected treatment plan [5, 6].

In general, nvAMD can be diagnosed using spectral domain optical coherence tomography (SD-OCT). OCT is a powerful imaging modality used to image various aspects of biological tissues, such as structural information, blood flow, elastic parameters, change of polarization state, and molecular content. SD-OCT provides the same high-resolution images as spectral-domain-based devices, which measure the intensity of the backscattered light. In order to provide additional information, polarization-sensitive OCT is often employed for identifying the polarization state of the examined retina. Studies have shown that most cellular parts of the retina preserve the polarization state of the incoming and reflected light, with the exception of melanin [7]. Melanin is found mostly in retinal pigment epithelium (RPE) cell organelles, but was also observed in other structures, including cells affected by pathologic conditions such as pigmentary changes, as well as inside pigment epithelium detachments (PED) [8].

Usually, the definition of the required regions of interest (ROI) is performed manually, through visual inspection by medical experts. However, this is a cumbersome and highly subjective process, which may lead to inconsistencies and reduced accuracy, especially in cases where poor quality OCT images are involved. Alternatively, diagnosis or treatment evaluation may be greatly improved through the application of semi- or fully automatic computer algorithms for detecting the required ROI. In digital imaging, this process is often referred to as image segmentation [8]. Numerous image segmentation algorithms have been proposed over the years, some of which are specially tailored for particular medical imaging data, such as CT, MRI, fundus or dental radiographs [9].

In general, segmentation techniques can be grouped in three wide categories: (a) intensity-based, (b) region-based and (c) model-based. Intensity-based techniques range from conceptually simple algorithms, such as Watershed and various thresholding approaches, to more complex classification schemes, such as Bayesian or k-Means classifiers [10, 11, 12]. In any case, those techniques examine the intensity level of individual image pixels, in order to classify them in two or more groups (background/foreground or several independent ROIs). On the other hand, region-based methods, detect and isolate regions as groups of pixels with common properties, based on specific correlations (such as intensity differences) between neighbouring pixels. This category includes more complex techniques, such as Fuzzy Connectedness, Region Growing or filters based on Markov Random Fields [13, 14, 15]. Model-based segmentation methods rely on predefined geometrical models for detecting depicted objects. Those techniques usually employ a repetitive process where the selected model is constantly updated (morphed), until it successfully detects the shape of the desired ROI. Deformable models, such as Active Contours (snakes) or Level-set based algorithms fall into this category of techniques, which can achieve accurate ROI detection, given an appropriate initial model estimate, but due to their iterative nature, they usually come with an increased computational cost [16, 17].

The isolation of anatomical regions of interest can also be achieved through the definition of the region’s boundaries, through the employment of intensity, region or model-based segmentation algorithms. The identification of anatomical structures (such as the PED) in OCT images is a common application for such techniques. To that end, several studies have presented numerous approaches for detecting or segmenting retinal layer boundaries with varying levels of success. Garvin et al. developed a specific automated approach for three-dimensional (3D) macular segmentation on OCT scans for the division of the retina into five layers [18]. This method was applied to 24 images acquired from 12 patients. The particular technique was assessed by comparing the detected ROIs to the ones defined by experts (golden standard), in terms of the perceived border distance between the two cases, which was found to be 6.1 $\pm$ 2.9 $\mu$ m. The limitation of this method is that it does not fare well with images where the desired ROI is not clearly visible or when noise is present in the ROI surface. Chiu et al. proposed a fully automatic approach for segmenting 8 retina layer structures in ocular images, using graph theory and dynamic programming which significantly reduced processing time [19]. The method was applied on 100 B-scan images and its results were compared to the ones obtained from manual segmentation (golden standard, as this was defined by experts), resulting in average error (deviation from the golden standard) ranging from 0.16% for the retinal pigment epithelium, to 3.88% for the nerve fiber layer. Kafieh et al. proposed a segmentation method capable of detecting 12 retinal boundaries by means of diffusion mapping on 2D and 3D OCT datasets [20]. The particular method was tested on 23 images from two subject groups (glaucoma and normal). The mean border position error, compared to the actual ROI borders was found to be 8.52 $\pm$ 3.13 $\mu$ m and 7.56 $\pm$ 2.95 $\mu$ m for the 2D and 3D datasets, respectively. Penha et al. evaluated an automated quantitative algorithm for the volumetric measurement and detection of retina PED from 58 patients diagnosed with AMD, using SD-OCT images [21]. However, the particular approach may provide inaccurate results if the PED area is only partially depicted. Zheng et al. used an interactive segmentation method to quantify the intraretinal and subretinal fluids in retinal SD-OCT volume scans from patients with exudative AMD [22]. This method was applied to 37 SD-OCT images from patients diagnosed with AMD. In this case, the detected ROIs closely matched the actual ones, as correlation coefficient measurements ranging from 0.897 to 0.979 and dice coefficient ranging from 0.721 to 0.785 were obtained. Finally, Sun et al. proposed an automated framework for PED segmentation consisting of four steps: (a) segmentation of abnormal retinal layers using multi-scale graph search, (b) segmentation refinement by means of AdaBoost, (c) serous PED segmentation using shape-constrained graph cut and (d) false positive removal using morphological operators [23]. The particular algorithm was tested on 25 patients with serous PED. In this case, quite accurate results were obtained, as the average true positive fraction was found to be 90.08%, the false positive volume fraction was 0.22%, the dice similarity coefficient was 91.20% and the positive predictive value was 92.62%.

In this context, the present study proposes a novel image processing technique for detecting PED regions and estimating their volume in SD-OCT images. The proposed technique is defined as a three-step procedure for (a) detecting the upper PED border, (b) detecting the lower PED border and (c) estimate the volume of the area included within the detected PED borders. The detection of the borders, which is the core of the proposed methodology, is based on multiscale Gaussian filtering. The proposed technique was applied to a total of 19 SD-OCT datasets from eight patients and achieved high levels of agreement with the manual segmentation from an expert.

2. Proposed framework

2.1 Data acquisition

The case samples consist of 2 women and 6 men, $n=$ 19 with right (OD) and left (OS) eyes, with neovascular age-related macular degeneration (nvAMD). All eyes had a pigment epithelial detachment (PED) as part of their lesion. The PED was either serous (sPED), vascularized (vPED), or both (mixed PED). All eyes were treated with non-selective VEGF (vascular endothelial growth factor) blockade using either bevacizumab (Avastin, Roche – 1.25 mg/0.1 cc) or ranibizumab (Lucentis, Novartis – 0.05 mg/0.1 cc) and a “treat-and-extend” therapeutic strategy. Secondary to persistent exudation a therapeutic switch to aflibercept (Eylea, Bayer – 2 mg/0.1 cc) was performed in a non-randomized fashion. The image raw data collect from Avanti RTVue XR OCT system. OCT images are acquired by scanning a patient’s eye(s) according to a predefined scan protocol. The data is in raster format with raw data size (17 $\times$ 1020 $\times$ 640). Raster scan consist of 17 horizontal line scans in (8 mm $\times$ 5 mm) area with scan depth 1.974 mm by default. The grayscale image size is 1020 columns and 640 lines.

Figure 1.

High level diagram of the segmentation system for OCT image.

Figure 2.

The retinal pigment epithelium forms the upper border of PED.

2.2 The proposed methodology

The proposed methodology consists of three steps, as shown in Fig. 1:

•
Detection of the upper border of the PED.
•
Detection of the lower border of the PED.
•
Estimation of the volume of the segmented PED.

2.3 Detection of the upper border

The upper border of the PED is the retinal pigment epithelium (RPE) and can be detected by taking into account that RPE is a bright structure surrounded by darker pixels, as shown in Fig. 2.

Figure 3.

(a) Zoomed version of Fig. 2. (b) Pixel values along the black line (perpendicular to RPE).

Figure 4.

The eigenvectors (white lines) corresponding to the most negative eigenvalue are perpendicular to a bright structure (RPE in our case).

Specifically, the pixels lying at the centerline of the RPE can be detected as the local maxima of the image in the direction that is perpendicular to the RPE’s local orientation (Fig. 3). As is known, a function of one variable will present local maximum at a point, if the first derivative vanishes at that point and the second derivative is negative. For a function of two variables, such as an image, the corresponding conditions in order to identify whether a pixel belongs to the centerline of the RPE are: the Hessian matrix should possess a negative eigenvalue and the first derivative of image along the corresponding eigenvector should be zero [24].

The detection of the centerline pixels of the RPE is performed using differential geometry as follows [24]. Let $I(x,y)$ denotes the gray level of a pixel at position $(x,y)$ . Initially, the Hessian matrix, $\mathbf{H}(x,y)$ , is calculated at each pixel position $(x,y)$ . The Hessian matrix is a 2D symmetric matrix that contains the values of the second order partial derivatives of the image at $(x,y)$ :

$\displaystyle\mathbf{H}(x,y)=\begin{bmatrix}I_{xx}(x,y)&I_{xy}(x,y)\\ I_{xy}(x,y)&I_{yy}(x,y)\\ \end{bmatrix}$ (1)

where $I_{xx}(x,y)=\frac{\partial^{2}I(x,y)}{\partial x^{2}}$ , $I_{yy}(x,y)=\frac{\partial^{2}I(x,y)}{\partial y^{2}}$ and $I_{xy}(x,y)=\frac{\partial^{2}I(x,y)}{\partial x\partial y}$ .

The partial derivatives are estimated by convolving the image $I(x,y)$ with the second order derivatives of the 2D Gaussian function with parameter $\sigma$ , $g_{\sigma}(x,y)=\frac{1}{2\pi\sigma^{2}}e^{-\frac{x^{2}+y^{2}}{2\sigma^{2}}}$ . The convolution of the image with the derivatives of the Gaussian is equivalent with taking the derivatives of the convolution of the image with the Gaussian itself. The convolution with the Gaussian ensures that the noise that is inherent in the image will not be enhanced during differentiation. The parameter $\sigma$ of the Gaussian determines the scale of analysis. Small values of $\sigma$ (for example around 1) can be used to detect thin line structures, but also can produce false structures due to noise. On the other hand, larger value of $\sigma$ produce more robust results with respect to noise, but can miss very thin structures.

Since the Hessian matrix is symmetric, it possesses real eigenvalues. The pixel at $(x,y)$ is considered a centerline pixel of a bright structure if it has at least one negative eigenvalue. Let $\lambda_{m}$ be the most negative eigenvalues and $\mathbf{u}_{m}=(u_{mx},u_{my})$ be the corresponding eigenvector. The eigenvector $\mathbf{u}_{m}$ will be perpendicular to the structure, as is shown in Fig. 4.

Having identified the pixels $(x,y)$ that have a negative eigenvalue of the Hessian matrix, the final decision whether these pixels do belong to the centerline of a bright structure can be made by examining the first derivative of the image along the direction of the corresponding eigenvector. The method described in [24] can be used to accomplish this task. In short, the Taylor expansion around a candidate pixel (with negative eigenvalue) using terms up to second order is considered. Then, taking the first derivative of the Taylor expansion along the direction of the eigenvector $\mathbf{u}_{m}=(u_{mx},u_{my})$ , parameterized by a parameter $t$ , the following relation is obtained:

$\displaystyle t=-\frac{u_{m,x}I_{x}(x,y)+u_{m,y}I_{y}(x,y)}{u_{m,x}^{2}I_{xx}(% x,y)+2u_{m,x}u_{m,y}I_{xy}(x,y)+u_{m,y}^{2}I_{yy}(x,y)}$ (2)

where $I_{x}$ and $I_{y}$ denote the first derivatives of the image. If $tu_{m,x},tu_{m,y}\in\left[-\frac{1}{2},\frac{1}{2}\right]$ then the pixel at $(x,y)$ is considered a centerline pixel.

Figure 5.

Centrelines detection after Gaussian filtering for various values of the standard deviation ( $\sigma$ ). (a) $\sigma=$ 8. (b) $\sigma=$ 9.(c) $\sigma=$ 10. (d) $\sigma=$ 11.

As mentioned above, the parameter $\sigma$ determines the scale of analysis. For better results, multiple values of the parameter were used, as has been proposed in the literature [25]. After experimentation four values were selected, namely 8, 9, 10 and 11. For each value of $\sigma$ , the pixels belonging to bright centrelines were detected (Fig. 5). For each detected pixel, the maximal absolute eigenvalue, as well as the orientation of the corresponding eigenvector were stored, in two 3D matrices; matrix $\lambda(x,y,\sigma)$ for the eigenvalues and $\theta(x,y,\sigma)$ for the orientation. For pixels that did not belong to centrelines, the corresponding entries of the two matrices were zero.

For each pixel $(x,y)$ at a centerline, the value of $\sigma_{\max}(x,y)$ corresponding to the maximum of the four eigenvalues was calculated:

$\displaystyle\sigma_{\max}(x,y)=\arg\operatorname{\max_{\sigma}}\lambda(x,y,\sigma)$ (3)

Using the values of $\sigma_{\max}(x,y)$ , two 2D matrices with the corresponding eigenvalues and orientations were created:

$\displaystyle\lambda_{\max}(x,y)=\lambda(x,y,\sigma_{\max}(x,y))$ (4) $\displaystyle\theta(x,y)=\theta(x,y,\sigma_{\max}(x,y))$ (5)

Figure 6.

Image with the maximal eigenvalues.

Figure 7.

Binary image after thresholding.

Figure 8.

Example of labelling the binary image.

Figure 6 shows the image with the maximal eigenvalues for the aforementioned example.

The next step is to create a binary image, $B(x,y)$ , of the centrelines, using thresholding with hysteresis. In particular, a high threshold ( $\textit{THR}_{H}$ ) and a low threshold ( $\textit{THR}_{L}$ ) on the values of $\lambda_{\max}(x,y)$ are defined. The thresholds are defined with respect to the maximum value of $\lambda_{\max}(x,y)$ . In this study, it was selected $\textit{THR}_{H}=0.5\max(\lambda_{\max}(x,y))$ and $\textit{THR}_{L}=0.2\max(\lambda_{\max}(x,y))$ . Starting from a seed pixel $(x_{s},y_{s})$ with $\lambda_{\max}(x_{s},y_{s})\geqslant THR_{H}$ , neighboring pixels $(x,y)$ are added as long as $\lambda_{\max}(x,y)\geqslant THR_{L}$ . The obtained image $B(x,y)$ contains connected components, where each component has at least one “strong” pixel, with $\lambda_{\max}$ value above $\textit{THR}_{H}$ . The above procedure produces a binary image containing the RPE and possibly other unwanted structures. Furthermore, in general, the RPE is not fully connected. The image $B(x,y)$ is thinned to create structures with one pixel width. Furthermore, gaps are filled using suitable morphological operations producing a binary image $C(x,y)$ (Fig. 7).

The next step is to isolate the RPE from the other structures: each column of the image $C(x,y)$ is scanned in a bottop-up direction until a white pixel is found. This pixel gets the label “1”. All the white pixels that are above that pixel in the same column get the label “2”. Let $D(x,y)$ denotes the image with the labelled pixels Fig. 8.

Then the connected components of the image $D(x,y)$ , without taking into account the label number, are extracted. If the majority of the pixels of a connected component have label “1” then it will be accepted, otherwise it will be rejected. Since the REP may not be fully connected, the Bresenham’s line algorithm is applied [26]. Furthermore, components with less than 40 pixels are removed (Fig. 9).

Figure 9.

Examples of RPE border detection.

Figure 10.

(a) Image without RPE border. (b) Corresponding binary image.

2.4 Detection of the lower border

In this stage, the upper border of the PED is available for every OCT image. Since the lower border cannot be easily isolated, due to the inherent noise of the OCT images, a semi-automated approach is applied, based on the observation that usually the first images of an OCT scan do not contain RPE. Thus, one such image is selected by the user, serving as a reference image (Fig. 10).

Figure 11.

The two characteristic points on the RPE border, where a change of the sign of the orientation takes place.

Figure 12.

Purple line is from upper border and green line is from reference image. (a) After vertical displacement of the green line. (b) After rotation of the green line with respect to P1.

Figure 13.

(a) Superposition of the detected on the original image. (b) Segmented image.

For each subsequent image, the RPE is scanned from left to right and the first pixel $(x_{1},y_{1})$ that corresponds to a negative-to-positive change of orientation $\theta(x,y)$ is marked as point P1. Similarly, the RPE is scanned from right to left to find the point first point P2 $(x_{1},y_{1})$ with a negative-to-positive change (Fig. 11).

Using the points P1 and P2, the border of the reference image is displaced and rotated: initially the border is displaced vertically by $dy=y^{\prime}_{1}-y_{1}$ , where $y^{\prime}_{1}$ is the y-coordinate of the white pixel in the reference image with x-coordinate equal to $x_{1}$ . Then, a rotation of the reference border around point P1 takes place until the reference border meets the point P2 (Fig. 12).

Figure 13 displays an example of superimposing the detected PED on the original image, as well as the corresponding segmented image.

Figure 14.

Example of images without a PED.

3. Results and discussions

The proposed method was applied on the datasets described in the previous Section. A total number of 323 images were available. Since 141 images did not include a PED (see Fig. 14), the method was applied on the remaining 182 images.

The performance of the method was initially evaluated by means of visual inspection. Figure 15 presents several examples of successful detection of PED.

Furthermore, the performance of the method was evaluated quantitatively using the manual segmentation of the PED from an expert (A. Charonis) as ground truth. Specifically, let $S_{g}$ and $S_{t}$ be the binary images that are obtained from the manual and the proposed segmentation, respectively. In both images, the pixels from the segmented structures are set to 1. All other pixels are set to 0. The quantitative evaluation was based on the following metrics [27]:

•
DICE coefficient: $\textit{DICE}=\frac{2|S_{g}\cap S_{t}|}{|S_{g}|+|S_{t}|}$
•
True positive rate: $\textit{TPR}=\frac{|S_{g}\cap S_{t}|}{|S_{g}|}$
•
Positive Predictive Value: $\textit{PPV}=\frac{|S_{g}\cap S_{t}|}{|S_{t}|}$
•
False positive rate: $\textit{FPR}=\frac{|S_{t}|-|S_{g}\cap S_{t}|}{|S_{r}-S_{g}|}$

Figure 15.
Examples of successful segmentation using the proposed method.

Table 1
Comparison of results from the proposed method and other methods from the literature

Method TPR (%) FPR (%) Dice (%) PPV (%)

Sun et al. 90.1 $\pm$ 5.3 0.22 $\pm$ 0.38 91.2 $\pm$ 3.77 92.62 $\pm$ 4.7

Shi et al. 87.1 $\pm$ 21.7 0.37 $\pm$ 0.54 – 81.2 $\pm$ 20.2

Xu et al. 85.7 $\pm$ 8.7 0.38 $\pm$ 0.18 85.7 $\pm$ 8.08 86.0 $\pm$ 8.99

Wu et al. 93.2 $\pm$ 9.0 0.15 $\pm$ 0.13 84.6 $\pm$ 10.6 80.0 $\pm$ 10.6

Proposed 90.9 $\pm$ 18.6 0.18 $\pm$ 0.15 89.6 $\pm$ 20.0 94.4 $\pm$ 14.4

Figure 16.
(a) Regression analysis. (b) Bland Altman plot. The dashed lines denote the limits of agreements (LoA) and the vertical bars the 95% confidence intervals for the LoA.

Figure 17.
Different images with low value of $\sigma(3,4,5,6)$ .

Figure 18.
Examples of binary images with different hysteresis thresholds with $\textit{THR}_{L}(0.1,0.2,0.2,0.5)$ and $\textit{THR}_{H}$ (0.2, 0.3, 0.5, 0.7).

Figure 19.
Images with unclear borders and PED.

In the above formulae, $S_{r}$ is the total retina region, $||$ denotes the number of pixels with value 1, $S_{g}\cap S_{t}$ is the intersection of the two binary images (i.e. a pixel is 1 if and only if the corresponding pixels in the two images are both 1).

The proposed method was compared with the methods by [28, 29, 30, 31] with respect to the aforementioned metrics. Table 1 lists the average value and the standard deviation of the metrics for all methods.

The relationship and the agreement of the measurements between the proposed method and the manual ground truth were quantified by means of regression analysis and the Bland-Altman analysis, respectively. Figure 16a presents the regression analysis of the segmentation volume (mm ${}^{3}$ ) of the proposed method against the manual ground truth method, whereas Fig. 16b presents the Bland-Altman plot. The Pearson correlation coefficient ( $r$ ) is 0.9956 ( $P<$ 0.0001) and the 95% confidence interval for $r$ is 0.9883 to 0.9984.

Since the upper order of the PED is a curvilinear structure, like the ones met in retinal or satellite images, MGF seems like a good candidate for its detection. MGF relies on the definition of the values of the standard deviation ( $\sigma$ ) of the Gaussian kernel. Due to the inherent noise that is present in OCT images, large values of $\sigma$ should be used. In this paper, we used the values 8, 9, 10 and 11 (see Fig. 5). Lower values of $\sigma$ were also tested, giving inferior results, as shown in Fig. 17.

Furthermore, it was also studied the effect of the thresholding process on the results. Figure 18, shows the binary images that were obtained for the following combinations of the threshold pairs $(\textit{THR}_{L},\textit{THR}_{H})$ . As can be seen the best results are obtained for $\textit{THR}_{L}=$ 0.2 and $\textit{THR}_{H}=$ 0.5.

Figure 19 provides cases where the method did not detect the PED successfully.
4. Conclusion

Method	TPR (%)	FPR (%)	Dice (%)	PPV (%)
Sun et al.	90.1 $\pm$ 5.3	0.22 $\pm$ 0.38	91.2 $\pm$ 3.77	92.62 $\pm$ 4.7
Shi et al.	87.1 $\pm$ 21.7	0.37 $\pm$ 0.54	–	81.2 $\pm$ 20.2
Xu et al.	85.7 $\pm$ 8.7	0.38 $\pm$ 0.18	85.7 $\pm$ 8.08	86.0 $\pm$ 8.99
Wu et al.	93.2 $\pm$ 9.0	0.15 $\pm$ 0.13	84.6 $\pm$ 10.6	80.0 $\pm$ 10.6
Proposed	90.9 $\pm$ 18.6	0.18 $\pm$ 0.15	89.6 $\pm$ 20.0	94.4 $\pm$ 14.4

In this paper, we proposed a segmentation method for detecting the PED from OCT images. The method comprises two stages: detection of the upper border of the PED based on multiscale Gaussian filtering (MGF) and detection of the lower border by comparing the current image with a properly selected reference one. MGF is a well-known technique that has been widely used for detection of edges, vessel segmentation in colour retinal images [32], extraction of streets from satellites images [24].

Regarding the detection of the lower border of the PED, a registration process was applied. The registration process aligns the border of the reference image with the corresponding parts of the lower border of the current image, using translational and rotational transformations. The proposed approach is quite simple and fast and provides satisfactory results. More sophisticated methods could also be used involving the use of an optimization technique to find the optimum of an evaluation function (for example some kind of distance function) with respect to the parameters of the transform.

The proposed method has similar or better performance with other state of the art methods. According to Table 1, the proposed method ranks second with respect to TPR and Dice coefficient, first with respect to PPV and second with respect to FPR. However, the proposed method will fail to detect the PED, if the dataset does not contain an image that can be used as a reference one or the PED has too unclear borders.

Finally, the average running time is 17 s, using a desktop PC with i7 at 3.6 GHz CPU 16 GB DDR3 RAM under the Matlab programming environment.

Footnotes

Conflict of interest

None to report.

References

Cukras

Wang

Meyerle

Forooghian

Chew

Wong

, et al. Optical coherence tomography-based decision making in exudative age-related macular degeneration: Comparison of time vs spectral-domain devices. Eye 2010; 24(5): 775-83.

Zayit-Soudry

Moroz

Loewenstein

, et al. Retinal pigment epithelial detachment. Surv Ophthalmol 2007; 52(3): 227-43.

Lommatzsch

Heimes

Gutfleisch

Spital

Zeimer

Pauleikhoff

, et al. Serous pigment epithelial detachment in age-related macular degeneration: Comparison of different treatments. Eye 2009; 23(12) 2163-8.

Schmidt-Erfurth

Chong

Loewenstein

Larsen

Souied

Schingemann

Eldem

Monés

Richard

Bandello

, et al. Guidelines for the management of neovascular age-related macular degeneration by the European Society of Retina Specialists (EURETINA). Br J Ophthalmol 2014; 98(9): 1144-67.

Frank

Steffen

Fleckenstein

, et al. Recent developments in the treatment of age-related macular degeneration. JCI 2014; 124(4): 1430-1438.

Makoto

Daisuke

Sawada

Sanae

Takeshi

, et al. Future therapies of wet age-related macular degeneration. Journal of Ophthalmology 2015; id: 138070.

Adrian

, et al. Optical coherence tomography. Journal of Microscopy 2012; 1365-2818.

Sundelin

Nilsson

Brunk

, et al. Lipofuscin-formation in cultured retinal pigment epithelial cells is related to their melanin content. Free Radic Biol Med 2001; 30(1): 74-81.

Pham

Prince

, et al. Current methods in medical image segmentation. Annual Rev Biomed Eng 2000; 2315-337.

10.

Lay

Siau

Weng

, et al. A review of image segmentation methodologies in medical image. Advanced Computer and Communication Engineering Technology 2015; 3151069-1080.

11.

Laurent

Michel

, et al. Watershed of a continuous function. Signal Processing 1994; 38(1): 99-112.

12.

Ong

Foong

KWC

Goh

PYT

Nowinski

, et al. Medical image segmentation using K-means clustering and improved watershed algorithm. IEEE Southwest Symposium on Image Analysis and Interpretation 2006; 61-65.

13.

Pereyra

McLaughlin

, et al. Fast unsupervised Bayesian image segmentation with adaptive spatial regularisation. IEEE Trans Image Process 2017; 26(6): 2577-2587.

14.

Sylvia

Caroline

Aris

Alison

, et al. Feature-based fuzzy connectedness segmentation of ultrasound images with an object completion step. Med Image Anal 2015; 26(1): 30-46.

15.

William

. Digital Image Processing. 4th Edition. 2007.

16.

Kato

Josian

, et al. Markov random fields in image segmentation. Foundations and Trends in Signal Processing 2011; 5(1-2): 1-155.

17.

Qian

Wang

Guo

, et al. An active contour model for medical image segmentation with application to brain CT image. Med Phys 2013; 40(2): 021911.

18.

Xin

Renjie

Shengdong

, et al. Image segmentation based on level set method. Physics Procedia 2012; 33840-845.

19.

Mona

Micheal

Randy

Stephen

Xiaodong

Milan

, et al. Intraretinal layer segmentation of macular optical coherence tomography images using optimal 3-D graph search. IEEE Trans Med Imaging 2008; 27(10): 1495-1505.

20.

Stephanie

Xiao

Peter

Cynthia

Joseph

Sina

, et al. Automatic segmentation of seven retinal layers in SDOCT images congruent with expert manual segmentation. Optics Express 2010; 18(18): 19413-19428.

21.

Kafieh

Rabbani

Abramoff

Sonka

, et al. Intra-retinal layer segmentation of 3d optical coherence tomography using coarse grained diffusion map. Med Image Anal 2013; 17(8): 907-28.

22.

Penha

Rosenfeld

Gregori

Falcão

Yehoshua

Wang

Feuer

, et al. Quantitative imaging of retinal pigment epithelial detachments using spectral-domain optical coherence tomography. Am J Ophthalmol 2012; 153(3): 515-23.

23.

Zheng

Sahni

Campa

Stangos

Raj

Harding

, et al. Computerized assessment of intraretinal and subretinal fluid regions in spectral-domain optical coherence tomography images of the retina. Am J Ophthalmol 2013; 155(2): 277-286.

24.

Steger

, et al. An Unbiased Detector of Curvilinear Structures. IEEE Trans Pattern Analysis Machine Intelligence 1998; 20(2): 113-125.

25.

David

, et al. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision 2004; 60(2): 91-110.

26.

Lokesh

Kislay

Bharat

, et al. Bresenham’s lines algorithm. IJRST 2014; 4(3): 51-63.

27.

Abdel

Allan

, et al. Metrics for evaluating 3D medical image segmentation: Analysis, selection, and tool. BMC Med Imaging 2015; 15-29.

28.

Zhuli

Haoyu

Fei

Lirong

Weifang

Dehui

Chenglin

Liang

Xinjian

, et al. An automated framework for 3D serous pigment epithelium detachment segmentation in SD-OCT images. Sci Rep 2016; 621739.

29.

Fei

Chen

Zhao

Zhu

Xian

Gao

Sonka

Chen

, et al. Automated 3-D retinal layer segmentation of macular optical coherence tomography images with serous pigment epithelial detachments, IEEE Trans Med Imaging 2015; 34(2): 441-52.

30.

Yan

Kim

Xiuying

Xun

David

, et al. Dual-stage deep learning framework for pigment epithelium detachment segmentation in polypoidal choroidal vasculopathy. Biomed Opt Express 2017; 8(9): 4061-4076.

31.

Fan

Chen

Yuan

Park

, et al. Three-dimensional continuous max flow optimization-based serous retinal detachment segmentation in SD-OCT for central serous chorioretinopathy. Biomed Opt Express 2017; 8(9): 4257-4274.

32.

Staal

AbrÃ moff

Niemeijer

Viergever

Ginneken

, et al. Rigid-Based Vessel Segmentation in Color Images of the retina. IEEE Trans Med Imaging 2004; 23(4): 501-9.

Detection of retinal pigment epithelium detachment from OCT images using multiscale Gaussian filtering

Abstract

Keywords

1. Introduction

2. Proposed framework

2.1 Data acquisition

• Detection of the upper border of the PED. • Detection of the lower border of the PED. • Estimation of the volume of the segmented PED. 2.3 Detection of the upper border

Footnotes

Conflict of interest

References

•
Detection of the upper border of the PED.
•
Detection of the lower border of the PED.
•
Estimation of the volume of the segmented PED.

2.3 Detection of the upper border