CN103854284B - Based on graphics search serous pigmentary epithelial pull-up from retina dividing method - Google Patents

Abstract

The invention discloses a retinal segmentation method for searching for serous pigment epithelial layer detachment based on a three-dimensional image, including: (1) a block-speed B-scan image alignment method based on interface segmentation on the retina; A multi-resolution image search segmentation method with the segmented interface as a constraint; (3) A method of image search algorithm with different constraints to obtain the lower interface of the pigment epithelium with raised areas and the smooth sub-retinal interface; (4) Segmentation method of serous pigment epithelial layer detachment based on the position difference between the lower interface of the pigment epithelium and the interface of the bottom part of the retina, combined with the area size and brightness information; (5) The method of segmenting and correcting the outer retinal layer after flattening the image . The segmentation result of the invention has high accuracy, can replace manual segmentation, and can play an important auxiliary role in the diagnosis and treatment of clinically relevant ophthalmic diseases.

Description

Retinal segmentation method based on three-dimensional image search for serous pigment epithelial detachment

technical field

The invention belongs to a method for segmenting retinal images, in particular to a method for segmenting tissue levels and lesion regions in retinal images of SD-OCT (frequency-domain optical coherence tomography).

Background technique

The retina is an extension of the neural tissue of the brain and has a complex multi-layered organizational structure. SD-OCT technology has become a powerful tool for non-destructive evaluation of retinal diseases. It can provide fast, high-resolution, three-dimensional images showing the internal layers of the retina, and provide clinical ophthalmologists with a great tool for diagnosis and treatment of diseases. help. The segmentation of retinal OCT images is of great significance to clinical practice: the segmentation of lesion areas and the quantitative analysis of their shape, size, and location play a key role in the diagnosis and treatment of diseases; the segmentation of retinal tissue levels and the analysis of various tissue shapes and brightness Quantitative analysis plays an important role in discovering early lesions, observing disease course and studying pathology. However, at present, most ophthalmologists use manual methods to conduct quantitative analysis of retinopathy displayed by OCT, which is highly subjective, cannot guarantee accuracy and consistency, and is difficult to comprehensively analyze the large amount of data brought by 3D scanning.

The current automatic retinal OCT image segmentation algorithm has the following defects: (1) Most of the algorithms are two-dimensional algorithms, that is, they are segmented independently in each slice image (x-z plane image, called B-scan image). The three-dimensional context information is not fully utilized, and it is more susceptible to image noise or artifacts, resulting in segmentation errors. (2) Most of the existing retinal tissue hierarchical segmentation algorithms are designed for the normal retina, and these algorithms will fail when the retinal tissue is greatly deformed due to lesions.

Serous pigment epithelial detachment may be caused by a variety of choroidal/retinal diseases, such as age-related macular degeneration, polypoidal choroidal vasculopathy, central serous chorioretinopathy, uveitis, etc. So far, there is no report on a systematic three-dimensional automatic segmentation method for all resolvable tissue layers and lesion regions in retinal OCT images of serous pigment epithelial detachment.

Contents of the invention

The present invention provides a solution to the above problems. For the first time, it provides a feasible and effective three-dimensional automatic system for all resolvable tissue layers and lesion areas in retinal OCT images of serous pigment epithelial detachment. split method. The organizational hierarchy includes 10 layers: nerve fiber layer (NFL), ganglion cell layer (GCL), inner plexiform layer (IPL), inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL) +Inner segment layer (ISL), connecting cilia (CL), outer segment layer (OSL), Verhoff membrane (VM), pigment epithelium (RPE), a total of 11 interfaces. In addition, due to the detachment of the pigment epithelium and the retinal base, the retinal base forms a separate interface, so the present invention can detect 12 interfaces in total.

The present invention provides a retinal segmentation method for searching for serous pigment epithelial detachment based on a three-dimensional image. The method mainly includes five steps:

Step S01, image preprocessing: mainly perform OCT denoising and alignment between B-scan images;

Step S02, segmentation of each layer of the inner retina: using a multi-resolution image search algorithm, sequentially segmenting according to the order of interface contrast from high to low, to obtain the nerve fiber layer (NFL), ganglion cell layer (GCL), inner plexus Inner layer (IPL), inner core layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL) + inner segmental layer (ISL) interface;

Step S03, segmentation of the pigment epithelium and estimation of the retinal base: in the outer retinal area, use different constraints to perform a graph search algorithm to obtain the lower interface of the pigment epithelium with raised areas and the smooth interface of the retinal base;

Step S04, segmentation of the pigment epithelium detachment area: the area between the lower interface of the pigment epithelium and the smooth sub-retinal interface is the detachment area of the pigment epithelium, and the false detection area is removed according to the size of the area or the brightness information;

Step S05, segmentation of each layer of the outer retina: flatten the image according to the lower interface of the pigment epithelium, and then use the image search algorithm to detect each layer of the outer retina, and obtain connecting cilia (CL), outer segment layer (OSL), Verhoff The interface between membrane (VM) and pigment epithelium (RPE).

The above five steps are described in detail as follows:

(1) Image preprocessing

Image preprocessing mainly includes the following two steps: denoising and B-scan image alignment.

(a) OCT image denoising

The 3D images obtained by the OCT eye imager contain more speckle noise. In order to ensure the effect of subsequent segmentation, the edge information in the image must be preserved as much as possible while effectively removing noise. The present invention uses a fast bilateral filter to denoise each B-scan image. The result of bilateral filtering is:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; f</span>}}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; G</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; ||</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; ||</span><span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>$

in

Here p is the currently processed pixel, q is the pixel in the neighborhood S of p, I _p and I _q are the gray values of p and q respectively, is the gray value of the filtering result, is the normalization coefficient, and It is a Gaussian function whose standard deviation is σ _s and σ _r respectively, and the two parameters of σ _s and σ _r are valued according to the image size and edge contrast respectively.

(b) B-scan image alignment

The movement of the eye during the imaging process can cause the position of the retina to fluctuate up and down in the continuous B-scan image, that is, the image is discontinuous in the slice direction (y-direction). This can cause difficulties in 3D segmentation. The present invention is based on the B-scan image alignment of the segmentation results of the supraretinal interface, because this interface has the highest contrast among all layered interfaces, and can be correctly segmented even on misaligned images. The process of segmentation using the multi-resolution image search algorithm is as follows: the denoised 3D image is down-sampled in the vertical direction (z-direction) so that the number of pixels in this direction is reduced to half, and the process is repeated once to obtain three Images with different resolutions are expressed as scales 1, 2, and 3 from low to high resolution. Segmentation is first performed on scale 1 of the lowest resolution, and based on the obtained results, further accurate segmentation is performed in the vicinity of scale 2, and so on, and finally the segmentation result on the original image is obtained. The segmentation process is the process of finding the segmentation plane with the least cost, which is completed by the graph search algorithm. The cost function of the interface on the nerve fiber layer is calculated by the Sobel operator, and the cost function is smaller at the edge position from dark to bright. In order to distinguish it from the interface between the inner and outer layers, that is, the upper interface connecting the cilia, on scale 1, another component is added to the cost function, which is the sum of the brightness of several pixels above each image point. In this way, since the upper interface of the nerve fiber layer is a darker background area, its corresponding cost is less than the cost of connecting the position of the interface on the cilium, and can be correctly detected. After the upper interface segmentation of the nerve fiber layer is completed, its average height, that is, the average z value, is calculated on each B-scan image. Points in the middle of the image are excluded from the calculation because these points have a large displacement due to the fovea or lesion. Move the image up or down according to the average height of the upper interface of the nerve fiber layer obtained in each B-scan, so that the average height of the upper interface of the nerve fiber layer in the result is a constant, which plays a role in aligning the images

(2) Segmentation of each layer of the inner retina

The inner retina is less affected by the lesion, so it is segmented first. The segmentation method is a multi-resolution image search method similar to that in step (1). First, the upper interface connecting the cilia, that is, the interface between the inner and outer layers is detected as a constraint. Segmentation is carried out in the sub-image below the upper interface of the nerve fiber layer. Because the lesion causes fluid in the retina, some tissues of the outer layer of the retina are not visualized in the OCT image, so the detected interface between the inner and outer layers is actually connecting the upper interface of cilia and The lower surface of the pigment epithelium was merged, which was defined as the merged interface connecting the ciliated pigment epithelium. Then, according to the order of the edge contrast of each interface, using the segmented interface as a constraint, segment the ganglion cell layer (GCL), inner plexiform layer (IPL), inner inner layer (INL), outer plexiform layer (OPL), The interface of the outer nuclear layer (ONL) + inner segmental layer (ISL). Interfaces with low border contrast may not be segmentable at low scales, so start segmentation at a higher resolution. In order to eliminate the influence of noise, mean filtering is performed on the obtained results in the x direction to obtain a smooth segmentation interface.

(3) Pigment epithelium segmentation and retinal base estimation

In retinal OCT images of serous pigment epithelial detachment, the pigment epithelium appears as a smooth bulge in the detachment area, with large changes in the anterior and posterior B-scan images. Below is a darker area of effusion, the bottom interface of which may not be visualized. The present invention uses the same cost function and different interface smoothness constraint conditions when detecting the two interfaces based on the graph search algorithm. When the smoothness constraint parameter is large, generally when the smoothness constraint parameter is 5-10, the segmentation result is the lower interface of the pigment epithelium with local bulges. When the value of the parameter is small and the smoothness constraint parameter is 1-4, the segmentation result is a smooth retinal bottom part interface.

(4) Breakaway region segmentation

The area between the lower interface of the pigment epithelium and the interface at the bottom of the retina obtained from the segmentation in step (3) is the detachment area of the pigment epithelium. However, due to the influence of noise, local errors in interface segmentation may occur, and false detection may occur. First, all the pixel points between the lower interface of the pigment epithelium and the lower interface of the retina form several three-dimensional connected areas, and the volume and average brightness of these areas are calculated respectively. When the volume is smaller than a certain predetermined value or the average brightness is greater than a certain predetermined value, this region is considered to be a falsely detected detachment region and removed.

While obtaining the three-dimensional detachment area, a two-dimensional distribution diagram of the detachment area in the x-y direction can also be obtained. This will be used for the correction of the outer retinal segmentation results in the next step.

(5) Segmentation of each layer of the outer retina

The tissues of the outer retina have a relatively flat shape in the normal retina. When there is detachment of the pigment epithelium, these tissues also rise with the uplift of the pigment epithelium, and may not be developed above the raised area, so in the OCT image, these tissues are discontinuous, so certain constraints must be added to Ensure the correctness of segmentation. In the present invention, based on the segmentation result of the lower interface of the pigment epithelium, the retinal image is flattened, that is, each column in the image is moved up and down, so that the lower interface of the pigment epithelium becomes a plane. This restores the raised part of the pigment epithelium to a flat shape, which approximates the normal flat shape of the outer retina. On the image after flattening, firstly, based on the merging interface of the connecting cilium pigment epithelium obtained in step (2), according to the result of step (4), the second-order polynomial is used for interpolation correction in the detachment area of the pigment epithelium to obtain the connecting cilium upper interface. Then use the graph search algorithm to segment the interface of the outer segment layer (OSL), Verhoff's membrane (VM), and pigment epithelium (RPE), and the interface between the outer segment layer (OSL) and Verhoff's membrane (VM). Regions of epithelial detachment were also corrected for by interpolation using a second-order polynomial. Remap the interface connecting cilia (CL), outer segment layer (OSL), Verhoff membrane (VM), and pigment epithelium (RPE) obtained in this step back to the original image to obtain the final segmentation result.

The invention integrates steps such as bilateral filter denoising, B-scan alignment, three-dimensional graph cutting technology, connected region segmentation, image flattening, and segmentation result correction. The segmentation result has high accuracy and can replace manual segmentation. Disease diagnosis and treatment can play an important auxiliary role.

Description of drawings

Fig. 1 is a structural representation of the present invention;

Figure 2 is an image of the retinal tissue hierarchy, Figure 2(a) is the original B-scan image, and Figure 2(b) is the 10-layer tissue and 11 interfaces of the normal retina;

Figure 3 is the denoising result of the bilateral filter, Figure 3(a) is the original image, and Figure 3(b) is the image after denoising;

Figure 4 is the x-y plane view of the B-scan alignment result, Figure 4(a) the original image, and Figure 4(b) the aligned image;

Figure 5 shows the interface segmentation results of connecting cilia (CL), outer segment layer (OSL), Verhoff's membrane (VM), and pigment epithelium (RPE). Figure 5(a) shows the lower interface of the pigment epithelium and the bottom part of the retina The results of interface segmentation, Figure 5(b) is the image after flattening the interface under the pigment epithelium, Figure 5(c) is the segmentation result of interface 7-10 in Figure 5(b), and Figure 5(d) is After mapping back to the original image, the results of the interface connecting cilia (CL), outer segment layer (OSL), Verhoff’s membrane (VM), and pigment epithelium (RPE) (wherein Verhoff’s membrane (VM), pigment epithelium ( RPE) interfaces substantially overlap in this image);

Figure 6 is the OCT image segmentation results of serous pigment epithelium detachment, Figure 6(a) is a two-dimensional display of the layered results, with 12 interfaces from top to bottom, and Figure 6(b) is the upper interface and connections of the nerve fiber layer The three-dimensional display of the segmentation results of the upper interface of the cilia, the lower interface of the pigment epithelium, and the interface at the bottom of the retina. Figure 6(c) is a two-dimensional display of the detachment area segmentation, and Figure 6(d) is a three-dimensional display of the detachment area segmentation.

In Fig. 2, the reference signs are as follows, 1 nerve fiber layer (NFL), 2 ganglion cell layer (GCL), 3 inner plexiform layer (IPL), 4 inner inner plexiform layer (INL), 5 outer plexiform layer (OPL) , 6 outer nuclear layer (ONL) + inner segment layer (ISL), 7 connecting cilia (CL), 8 outer segment layer (OSL), 9 Verhoff membrane (VM), 10 pigment epithelium (RPE).

detailed description

The present invention will be further described below in combination with specific embodiments.

As shown in Figure 1, this method mainly includes five steps: image preprocessing, segmentation of each layer of the inner retina, segmentation of the pigment epithelium and estimation of the retinal base, segmentation of the detachment region, and segmentation of each layer of the outer retina.

As shown in Figure 2, the retinal tissue hierarchy includes 10 layers: nerve fiber layer (NFL) 1, ganglion cell layer (GCL) 2, inner plexiform layer (IPL) 3, inner inner layer (INL) 4, outer plexiform layer (OPL) 5, outer nuclear layer (ONL) + inner segment layer (ISL) 6, connecting cilia (CL) 7, outer segment layer (OSL) 8, Verhoff membrane (VM) 9, pigment epithelium (RPE) 10. There are 11 interfaces in total.

The present invention is a retinal segmentation method based on a three-dimensional map search for serous pigment epithelial layer detachment, specifically described as follows,

(1) Image preprocessing

Image preprocessing mainly includes the following two steps: denoising and B-scan image alignment.

(a) OCT image denoising

The 3D images obtained by the OCT eye imager contain more speckle noise. In order to ensure the effect of subsequent segmentation, the edge information in the image must be preserved as much as possible while effectively removing noise. The present invention uses a fast bilateral filter to denoise each B-scan image. The result of bilateral filtering is:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; f</span>}}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; G</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; ||</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; ||</span><span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>$

in

Here p is the currently processed pixel, and q is the pixel in the neighborhood S of p. I _p and I _q are the gray values of p and q, respectively. is the gray value of the filtering result, is the normalization coefficient. and are Gaussian functions with standard deviations σ _s and σ _r respectively. The two parameters σ _s and σ _r take values according to the image size and edge contrast respectively. The denoising results are shown in Figure 3.

(b) B-scan image alignment

The movement of the eye during the imaging process can cause the position of the retina to fluctuate up and down in the continuous B-scan image, that is, the image is discontinuous in the slice direction (y-direction). This can cause difficulties in 3D segmentation. The present invention is based on B-scan image alignment of the segmentation results of the upper interface of the nerve fiber layer 1, that is, the upper interface of the retina, because this interface has the highest contrast among all layer interfaces, and can be correctly segmented even on misplaced images. The process of segmenting the upper interface of the nerve fiber layer 1 with the multi-resolution image search algorithm is as follows: the denoised three-dimensional image is down-sampled in the vertical direction (z-direction) so that the number of pixels in this direction becomes half, repeat Once this process is performed, three images with different resolutions are obtained, which are represented as scale 1, scale 2, and scale 3 in order of resolution from low to high. Segmentation is first performed on scale 1 of the lowest resolution, and based on the obtained results, further accurate segmentation is performed in the vicinity of scale 2, and so on, and finally the segmentation result on the original image is obtained.

The segmentation process is the process of finding the segmentation plane with the least cost, which is completed by the graph search algorithm. The cost function of the upper interface of the nerve fiber layer 1 is calculated by the Sobel operator, and the cost function is smaller at the edge from dark to bright. In order to distinguish it from the interface between the inner and outer layers, that is, the upper interface connecting cilia 7, on scale 1, another component is added to the cost function, which is the sum of the brightness of several pixel points above each image point. In this way, since there is a darker background area above the upper interface of the nerve fiber layer 1, its corresponding cost is less than the cost of connecting the position of the upper interface of the cilium 7, and can be correctly detected. After the segmentation of the upper interface of the nerve fiber layer 1 is completed, its average height, that is, the average z value, is calculated on each B-scan image. Points in the middle of the image are excluded from the calculation because these points have a large displacement due to the fovea or lesion. According to the average height of the upper interface of the nerve fiber layer 1 obtained in each B-scan, the image is moved up or down, so that the average height of the upper interface of the nerve fiber layer 1 in the result is a constant, which plays a role in aligning the images. The image effect after alignment can be seen from the x-y direction, as shown in Figure 4.

(2) Segmentation of each layer of the inner retina

The inner retina is less affected by the lesion, so it is segmented first. The segmentation method is a multi-resolution image search method similar to that in step (1).

First, the upper interface connecting the cilia 7, that is, the interface between the inner and outer layers is detected as a constraint condition. Segmentation is carried out in the sub-image below the upper interface of the nerve fiber layer 1. Because the lesion causes fluid accumulation in the retina, some tissues of the outer layer of the retina are not visualized in the OCT image, so the interface detected between the inner and outer layers is actually connected to the cilia 7 The interface formed by merging with the interface on the pigment epithelium 10 is defined as the merging interface connecting the ciliated pigment epithelium.

Then, according to the order of the edge contrast of each interface, with the segmented interface as a constraint condition, the ganglion cell layer (GCL) 2, inner plexiform layer (IPL) 3, inner inner plexiform layer (INL) 4, outer plexiform layer ( OPL)5, the interface of outer nuclear layer (ONL) + inner segment layer (ISL)6. Interfaces with low border contrast may not be segmentable at low scales, so start segmentation at a higher resolution. The specific segmentation sequence, upper and lower constraint surfaces, corresponding boundary changes, and the starting scale of multi-resolution segmentation are shown in Table 1. In order to eliminate the influence of noise, mean filtering is performed on the obtained results in the x direction to obtain a smooth segmentation interface.

(3) Pigment epithelium segmentation and retinal base estimation

In retinal OCT images of serous pigment epithelial detachment, the pigment epithelium appears as a smooth bulge in the detachment area, with large changes in the anterior and posterior B-scan images. Below is a darker area of effusion, the bottom interface of which may not be visualized. The present invention uses the same cost function and different interface smoothness constraint conditions when detecting the two interfaces based on the graph search algorithm. When the smoothness constraint parameter is 5-10, the segmentation result is the lower interface of the pigment epithelium with local uplift. When the smoothness constraint parameter is 1 to 4, the segmentation result is a smooth retinal bottom part interface.

(4) Breakaway area segmentation

The area between the lower interface of the pigment epithelium layer 10 and the interface at the bottom of the retina obtained by the segmentation in step (3) is the area of pigment epithelial layer detachment. However, due to the influence of noise, local errors in interface segmentation may occur, and false detection may occur. First, all the pixels between the lower interface of the pigment epithelium 10 and the interface at the bottom of the retina form several three-dimensional connected areas, and the volume and average brightness of these areas are calculated respectively. When the volume is smaller than a certain predetermined value or the average brightness is greater than a certain predetermined value, this region is considered to be a falsely detected detachment region and removed.

While obtaining the three-dimensional detachment area, a two-dimensional distribution diagram of the detachment area in the x-y direction can also be obtained. This will be used for the correction of the outer retinal segmentation results in the next step.

(5) Segmentation of each layer of the outer retina

The tissues of the outer retina have a relatively flat shape in the normal retina. When there is detachment of the pigment epithelium, these tissues also rise with the uplift of the pigment epithelium, and may not be developed above the raised area, so in the OCT image, these tissues are discontinuous, so certain constraints must be added to Ensure the correctness of segmentation.

In the present invention, based on the segmentation result of the lower interface of the pigment epithelium 10, the retinal image is flattened, that is, each column in the image is moved up and down, so that the lower interface of the pigment epithelium 10 becomes a plane. This restores the raised part of the pigment epithelium to a flat shape, which approximates the normal flat shape of the outer retina. On the image after flattening, firstly, based on the merging interface of connecting ciliated pigment epithelium obtained in step (2),

According to the result of step (4), the second-order polynomial is used for interpolation correction in the detachment area of the pigment epithelium to obtain the upper interface of the connecting cilium 7 . Then, the graph search algorithm was used to segment the interfaces of the outer segment layer (OSL)8, Verhoff's membrane (VM)9, and pigment epithelium (RPE)10. The segmentation sequence and constraints are also shown in Table 1. For the interface that separates the outer segment layer (OSL)8 and Verhoff's membrane (VM)9, the second-order polynomial should also be used for interpolation correction in the detachment area of the pigment epithelium. Remap the interface of connecting cilia (CL)7, outer segment layer (OSL)8, Verhoff membrane (VM)9, and pigment epithelium (RPE)10 obtained in this step back to the original image to obtain the final segmentation result.

The result of this step is shown in Figure 5.

Table 1 Segmentation method of each layer of retina

(6) Experimental results

Some experimental results are shown in Figure 6.

For hierarchical segmentation, the average of the results of two independent manual segmentations by two experts was used as the gold standard. The absolute value of the difference between the automatic segmentation result and the z value of the gold standard segmentation interface is the segmentation error. The absolute value of the difference between the z-scores of the two expert segmentation results represents the inter-observer difference. The experimental results on 20 samples show that the average segmentation error of the present invention is 2.25 ± 0.96 pixels (7.87 ± 3.36 microns), compared with the inter-observer difference of 2.23 ± 0.73 pixels (7.81 ± 2.56 microns), there is no significant statistical difference, can be considered basically the same. Therefore, this method can replace the manual segmentation method.

For the segmentation of pigment epithelial layer detachment, an expert's manual segmentation result was used as the gold standard, and the true positive rate TPVF and false positive rate FPVF were used as the objective indicators of the evaluation method, calculated as follows:

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span>$

Where |·| represents the volume, C _TP represents the true positive point set, C _FP represents the false positive point set, C _GT represents the detachment area point set in the gold standard, and V represents the set of all pixels in the entire retinal region. The experimental results show that the average true positive rate of this method is 87.9%, and the average false positive rate is 0.36%.

So far, an automatic segmentation method for retinal SD-OCT images of serous pigment epithelial detachment has been implemented and validated. The invention integrates steps such as bilateral filter denoising, B-scan alignment, three-dimensional graph cutting technology, connected region segmentation, image flattening, and segmentation result correction. The segmentation result has high accuracy and can replace manual segmentation. Disease diagnosis and treatment can play an important auxiliary role.

The basic principles, main features and advantages of the present invention have been described above. Those skilled in the industry should understand that the present invention is not limited by the above-mentioned embodiments, and what described in the above-mentioned embodiments and the description only illustrates the principles of the present invention, and the present invention will also have other functions without departing from the spirit and scope of the present invention. Variations and improvements are possible, which fall within the scope of the claimed invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims (7)

1. The retinal segmentation method based on three-dimensional image search serous pigment epithelial layer detachment, is characterized in that, comprises the following steps, Step S01, image preprocessing: performing OCT image denoising and B-scan image alignment; Step S02, segmentation of each layer of the inner retina: using a multi-resolution image search algorithm, sequentially segmenting according to the order of interface contrast from high to low, to obtain the nerve fiber layer, ganglion cell layer, inner plexiform layer, inner core layer, The interface between outer plexiform layer, outer nuclear layer and inner segmental layer; Step S03, segmenting the pigment epithelium and estimating the bottom of the retina: in the outer retina area, use different smoothness constraint parameters to perform a graph search algorithm to obtain the lower interface of the pigment epithelium with raised areas and the smooth interface of the bottom of the retina; Step S04, segmenting the pigment epithelium detachment area: the area between the lower interface of the pigment epithelium and the bottom part of the retina is the detachment area of the pigment epithelium, and the false detection area is removed according to the size of the area or the brightness information; Step S05, segmentation of each layer of the outer retina: after flattening the image according to the lower interface of the pigment epithelium, use the image search algorithm to detect each layer of the outer retina, and obtain the connecting cilia, outer segment layer, Verhoff's membrane, and pigment epithelium. Interface; The step S04, the segmentation of the pigment epithelial layer detachment region specifically includes, All the pixels between the lower interface of the pigment epithelium and the interface at the bottom of the retina form several three-dimensional connected regions, and the volume and average brightness of the three-dimensional connected regions are calculated respectively. When the volume of the region is smaller than a predetermined value or the average brightness is greater than When the value is a certain predetermined value, the area is a falsely detected detachment area. 2. the retinal segmentation method based on three-dimensional image search serous pigment epithelial layer detachment according to claim 1, is characterized in that, OCT image denoising specifically comprises the following steps in the step S01, adopts a kind of fast bilateral filter to each After the B-scan image is denoised, the denoised three-dimensional image is obtained, and the result of bilateral filtering is shown in formula (1a): in p is the currently processed pixel, q is the pixel in the neighborhood S of p, I _p and I _q are the gray values of p and q respectively, is the gray value of the filtering result, is the normalization coefficient, and It is a Gaussian function whose standard deviation is σ _s and σ _r respectively, and the two parameters of σ _s and σ _r are valued according to the image size and edge contrast respectively. 3. the retinal segmentation method based on three-dimensional image search serous pigment epithelial layer detachment according to claim 1, is characterized in that, the alignment between described B scan image is to carry out B scan based on the segmentation result to the upper interface of nerve fiber layer Image alignment, specifically includes the following steps, 1-1) Use the multi-resolution image search algorithm to segment the upper interface of the nerve fiber layer: down-sample the denoised 3D image in the vertical direction so that the number of pixels in this direction becomes half, and repeat this process once to obtain Three images with different resolutions are represented as scale 1, scale 2, and scale 3 according to the resolution from low to high; 1-2) Carry out segmentation on the lowest resolution scale 1, based on the segmentation results of scale 1, perform further accurate segmentation in the nearby area on scale 2, and then continue to segment the nearby area on scale 3, and finally get Segmentation results on the original image; 1-3) After the upper interface segmentation of the nerve fiber layer is completed, calculate its average height on each B-scan image, that is, the average z-value, and exclude the points in the middle of the image during the calculation process, according to the nerve fibers obtained in each B-scan The average height of the upper interface of the layer is moved up or down to make the average height of the upper interface of the nerve fiber layer in the result constant, which plays a role in aligning the images. 4. the retinal segmentation method based on three-dimensional image search serous pigment epithelial layer detachment according to claim 3, is characterized in that: the segmentation described in step 1-2) is the process of finding the segmentation plane with the least cost, searched by image The algorithm is completed, and the cost function of the upper interface of the nerve fiber layer is calculated by the Sobel operator. The cost function is smaller at the edge position from dark to bright. , the cost function adds a component, which is the sum of the brightness of several pixels above each image point. 5. the retinal segmentation method based on three-dimensional graph search serous pigment epithelium detachment according to claim 1, is characterized in that, the segmentation of each level of described inner layer retina specifically comprises the following steps, Firstly, the upper interface of the connecting cilia, that is, the interface between the inner and outer layers is detected as a constraint condition, and the segmentation is performed in the submap below the upper interface of the nerve fiber layer. Due to the lesions causing fluid in the retina, some tissues of the outer layer of the retina are not visible in the OCT image. Therefore, the detected interface between the inner and outer layers is actually formed by merging the upper interface connecting the cilium and the upper interface of the pigment epithelium, which is defined as the merged interface connecting the cilium pigment epithelium; Then, according to the order of the edge contrast of the above-mentioned interfaces, and using the segmented interfaces as constraints, segment the ganglion cell layer, inner plexiform layer, inner core layer, outer plexiform layer, and outer nuclear layer+inner ganglion layer interface. 6. The retinal segmentation method for serous pigment epithelial layer detachment based on three-dimensional image search according to claim 1, characterized in that, said step S03 pigmentary epithelial layer segmentation and retinal bottom estimation specifically includes the following steps, based on image search algorithm detection When the lower interface of the pigment epithelium and the interface of the bottom of the retina, the same cost function and different interface smoothness constraints are used. When the smoothness constraint parameter is set to 5-10, the segmentation result is the lower interface of the pigment epithelium with local bulges. When the smoothness constraint parameter is 1-4, the segmentation result is a smooth retinal subsurface. 7. The retinal segmentation method for searching for serous pigment epithelial layer detachment based on a three-dimensional image according to claim 1, wherein the segmentation of each layer of the outer retina in the step S05 specifically comprises the following steps, Based on the segmentation result of the lower interface of the pigment epithelium, the retinal image is flattened, that is, each column in the image is moved up and down, so that the lower interface of the pigment epithelium becomes a plane, so that the raised part of the pigment epithelium is restored to a flat shape. The normal flat shape of the outer retina; On the flattened retinal image, based on the merging interface of the connecting cilium pigment epithelium obtained by segmentation in step S02, according to the result of step S04, interpolation correction is performed with a second-order polynomial in the detachment area of the pigment epithelium to obtain the upper interface of connecting cilia, Use the image search algorithm to segment the interface of the outer segment layer, Verhoff membrane, and pigment epithelium, and the interface between the outer segment layer and Verhoff membrane, and use a second-order polynomial to perform interpolation correction in the detachment area of the pigment epithelium, and connect the connecting cilia obtained above , outer segment layer, Verhoff's membrane, and pigment epithelium interface are remapped back to the original image to obtain the final segmentation result.