CN106846338A - Retina OCT image based on mixed model regards nipple Structural Techniques - Google Patents

Abstract

Nipple Structural Techniques are regarded the invention discloses a kind of retina OCT image based on mixed model, is comprised the following steps：1）Retinal images are filtered, even up treatment by image preprocessing；2）According to the mark point of hand labeled, active appearance models are set up, and the coarse segmentation of retinal images structure is carried out with active appearance models；3）With step 2）In result be constraints, the Accurate Segmentation of retinal images structure is carried out with graph search method.The present invention provide first it is a kind of with feasibility and validity to the SD OCT centered on optic disk（Domain optical coherence fault imaging）The method that retinal images are split, makes microscope imaging equipment that the imaging of noncontact, high-resolution, high-res can be in this way carried out with profit.

Description

Segmentation method of optic papilla structure in retinal OCT images based on mixture model

technical field

The invention belongs to an image processing method, in particular to a method for segmenting an SD-OCT (frequency-domain optical coherence tomography) retinal image centered on an optic disc.

Background technique

Optical coherence tomography (OCT) is an imaging technique that has developed rapidly in recent years and has been widely used to acquire high-resolution retinal images. The retinal structure of the human eye is very complex, especially in the area centered on the optic nerve head (ONH), its structural changes are more complex and the changes are more significant, so it is necessary to develop a reliable and automatic retinal optic nerve head area Structural segmentation and measurement methods are very important.

So far, some researches that have been published are mainly to segment the retinal OCT images of the macular area. For the segmentation of the retinal OCT images with the optic disc as the center area, although some studies have focused on the structural segmentation of this area, However, there are some methods based on two-dimensional color fundus photos to segment the optic disc and optic cup; some methods are based on three-dimensional OCT images, but because of the complexity of the ONH structure and the structural changes caused by the lesion, As a result, the accuracy of segmentation is not very high.

Contents of the invention

The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a feasible and effective segmentation method for SD-OCT (frequency-domain optical coherence tomography) retinal images centered on the optic disc.

In order to solve the above-mentioned technical problems, the present invention provides a method for segmenting the optic disc structure in retinal OCT images based on a hybrid model.

The present invention provides a kind of automatic segmentation method based on AAM (active appearance model) and Graph-search (graph search) technology, and this method mainly comprises 3 steps: image preprocessing, AAM modeling and coarse segmentation, Graph-search ( graph search) method for fine segmentation.

1) Image preprocessing, filtering and flattening the retinal image;

2) Establish an AAM (Active Appearance Model) based on the manually marked points, and use the AAM (Active Appearance Model) to complete the coarse segmentation of the retinal image structure;

3) With the result in step 2) as a constraint condition, a Graph-Search (graph search) method is used to realize the fine segmentation of the retinal image structure.

In step 1), the image filtering process adopts the algorithm of gradient anisotropic diffusion to remove noise.

In step 1), the image leveling process uses the external membrane as a standard for leveling, and first specifies the detected position of the external membrane as a fixed value, and then uses this fixed value as a reference to align the complete image.

In step 2), the steps of establishing an active appearance model are:

First, manually mark the contour line of the segmentation target on each frame of image preprocessed in step 1), then represent the 3D segmentation target as a stack of 2D contour lines, and mark multiple markers on the marked contour line point;

After all the labeled points of the training data are marked, a standard active appearance model is used to build a model of the retinal image structure.

Active appearance model includes two parts: shape model and texture model:

in, is the average shape model, is the average texture model corresponding to the average shape, and Q _s and Q _g are transformation matrices formed by the shape calculated by principal component analysis and the characteristic components of texture principal components. s is the shape parameter to control the shape change; t is the texture parameter to control the texture change; x is the shape model, g is the texture model; use the active appearance model to locate and segment the region of interest on the test data, and obtain the layer 1-layer 7 initial results.

In step 3), the specific steps adopted are as follows:

First, the retinal image is defined as a three-dimensional matrix I(x, y, z), whose size is X×Y×Z, where x, y, and z are spatial coordinates, and X, Y, and Z are volumes in three directions, respectively. prime number;

The surface to be detected is defined as a function S(x, y), x∈{0,...,X-1}, y∈{0,...,Y-1}, and S(x,y)∈ {0,...,Z-1};

The parameter Δx defines the smooth constraint condition in the x direction, and the parameter Δy defines the smooth constraint condition in the y direction; the maximum distance between two adjacent surfaces in the x direction satisfies |S(x+1, y)-S(x, y )|≤Δx, the maximum distance in the y direction satisfies |S(x, y+1)-S(x, y)|≤Δy;

Then, a vertex-weight directed graph G=(V, E) is established based on image voxels, which contains a set V of vertices v and a set E of edges e; in this directed graph, any vertex v∈ V corresponds to a voxel point in the image I(x, y, z), and any arc <v _i , v _j >∈E connects two vertices v _i , v _j ; each vertex v∈V , the cost value c(x,y,z) of {V(x,y,z)|(z>0)} is calculated according to the gradient magnitude of the OCT image, indicating the possibility that a pixel does not belong to the target surface , the weight of each vertex v∈V is calculated according to the cost value, as shown in the following formula (2):

The problem of finding the optimal surface is transformed into retrieving the minimum cost closed set in the directed graph G to segment the region of interest in the image.

Also includes the step of segmenting the optic disc image:

Detect the boundary of the optic disc image, and hide the layer information in the optic disc area; obtain the projection image in the z direction between layer 6 and layer 7 according to the layering results, and then use the shape prior model algorithm to segment the optic disc area image .

The beneficial effect that the present invention reaches:

The present invention provides a feasible and effective method for segmenting retinal images of SD-OCT (frequency-domain optical coherence tomography) centered on the optic disc for the first time, so that microscope imaging equipment can use this method for non-contact , high-resolution, high-resolution imaging.

Description of drawings

Figure 1 is a flow chart of the ONH structure segmentation algorithm for retinal OCT images based on the hybrid model.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solution of the present invention more clearly, but not to limit the protection scope of the present invention.

The basic block diagram of the method of the present invention is as shown in accompanying drawing 1, mainly comprises 3 steps: image preprocessing, do initial segmentation with AAM (active appearance model algorithm), do fine segmentation with Graph-Search (graph search algorithm). The specific description is as follows.

(1) Image preprocessing

Image preprocessing mainly includes the following steps: filtering and leveling of retinal images.

(a) Image filtering

The algorithm of gradient anisotropic diffusion is used to filter the image to remove noise as much as possible and retain the detailed information of the image.

(b) Flattening of the image

In the process of collecting retinal OCT images, due to the eyeball rotation of the subject, there will be large jumps between different frames of the collected images, and the smoothness of the images in some positions is not high enough. Affected by this, resulting in Stratification results worse. The flattening operation of the image is to use the external membrane, which is relatively easy to detect, as a flattening reference, first specify the detected external membrane position as a fixed value, and then use this as a reference to align the entire image.

(2) Establish AAM model (active appearance model)

Firstly, the contour line of the segmentation target is manually marked on each frame of the collected image, and then the 3D segmentation target is represented as a stack of 2D contour lines. A plurality of marker points are marked on the marked contour line. In the present embodiment, 13 contour lines are marked altogether. Except that the first ILM (inner limiting membrane) has marked 16 mark points, all the other 12 contours have only marked 6 points. marker points.

After all the labeled points of the training data are marked, the standard AAM (Active Appearance Model) is used to build a model of the retinal image structure.

The AAM model (Active Appearance Model) includes two parts, the shape model s and the texture model g, as shown in formula (1):

in, is the average shape model, is the average texture model corresponding to the average shape, and Q _s and Q _g are transformation matrices formed by the shape calculated by principal component analysis and the characteristic components of texture principal components. s is the shape parameter to control the shape change; t is the texture parameter to control the texture change. x is based on the mean shape model and the shape model obtained by the transformation matrix and parameters, g is based on the average texture model And the texture model obtained by the transformation matrix and parameters. Using this model, the region of interest can be located and segmented on the test data, and the initial results of layer 1-layer 7 can be obtained.

(3) Use Graph-Search (graph search) method to realize accurate segmentation of retinal image structure

The segmentation result obtained with the AAM model (active appearance model) is not accurate enough, so the present invention proposes to use the result of the AAM model (active appearance model) as a constraint condition, and use the Graph-Search (graph search) method to realize the precise segmentation of the retinal image structure Methods.

The single-surface detection 3-D Graph-Search (graph search) method proposed by K.li et al. defines the retinal image as a three-dimensional matrix I(x, y, z), whose size is X×Y×Z, where x, y, z are space coordinates, and X, Y, Z are the number of voxels in three directions respectively. The surface to be detected can be defined as a function S(x,y), x∈{0,...,X-1}, y∈{0,...,y-1}, and S(x,y) ∈{0,...,Z-1}. In addition, the parameter Δx defines the smoothness constraint in the x direction, and the parameter Δy defines the smoothness constraint in the y direction. That is to say, the maximum distance between two adjacent surfaces in the x direction must satisfy |S(x+1, y)-S(x, y)|≤Δx, and the maximum distance in the y direction must satisfy |S(x, y +1)-S(x,y)|≤Δy. The cost value c(x,y,z) of each point indicates the probability that a voxel does not belong to the target surface. Therefore, the optimal surface has the smallest cost value. Then, a vertex-weight directed graph G=(V, E) is established based on image voxels, which includes a set V of vertices v and a set E of edges e. In this directed graph, any vertex v∈V corresponds to a voxel point in the image I(x, y, z), and any arc <v _i , v _j >∈E connects two vertices v _i , v _j . For each vertex v∈V, the cost value c(x,y,z) of {V(x,y,z)|(z>0)} is calculated based on the gradient magnitude of the OCT image, indicating that a pixel The possibility that the point does not belong to the target surface, the weight of each vertex v∈V is calculated according to the cost value, as shown in the following formula (2):

Therefore, the problem of finding the optimal surface is transformed into retrieving the minimum cost closed set in the directed graph G, so that the maximum flow/minimum cut algorithm can be used to calculate the minimum closed set and segment the region of interest in the image.

In this algorithm, the cost value uses the Sobel edge operator to calculate the gradient magnitude in the z direction. Two types of cost equations are adopted respectively, among which, layers 1, 5, and 6 are changes from top to bottom, from black to white; layers 2, 3, 4, and 7 are from top to bottom, from white to black Variety.

(4) Segmentation of optic disc image

Inside the neural tube image, the hierarchical structure of the retinal image organization is not clear, and the automatic layering method cannot effectively layer. Therefore, this algorithm detects the boundary of the optic disc image and hides the layer information in the optic disc area. . Firstly, the projection image in the z direction between layer 6 and layer 7 is obtained according to the layering results. Then, the optic disc area is segmented using the shape prior model algorithm.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the technical principle of the present invention, some improvements and modifications can also be made. It should also be regarded as the protection scope of the present invention.

Claims (7)

1. A retina OCT image optic papilla structure segmentation method based on a hybrid model is characterized by comprising the following steps:

1) image preprocessing, namely filtering and flattening the retinal image;

2) establishing an active appearance model according to the manually marked mark points, and performing rough segmentation on the retina image structure by using the active appearance model;

3) and (3) accurately segmenting the retina image structure by using the graph searching method by taking the result in the step 2) as a constraint condition.

2. The retinal OCT image optic papillary structure segmentation method based on the hybrid model as claimed in claim 1, wherein in step 1), the image filtering process adopts gradient anisotropic diffusion algorithm to remove noise.

3. The method for segmenting retinal OCT image papilla structures according to claim 1, wherein in step 1), the image flattening process uses the outer limiting membrane as a reference for flattening, and the detected position of the outer limiting membrane is first assigned as a fixed value, and then the whole image is aligned with the fixed value as the reference.

4. The retinal OCT image optic papillary structure segmentation method based on the hybrid model as claimed in claim 1, wherein in the step 2), the step of establishing the active appearance model comprises the following steps:

firstly, manually marking the contour line of a segmentation target on each frame of image obtained by preprocessing in the step 1), then representing the 3-dimensional segmentation target into a stack of 2-dimensional contour lines, and marking a plurality of mark points on the marked contour line;

and when all the marking points of the training data are marked, establishing a model of the retina image structure by using a standard active appearance model.

5. The retinal OCT image optic papillary structure segmentation method based on the hybrid model as claimed in claim 4, wherein the active appearance model comprises two parts of a shape model and a texture model:

$x=\stackrel{\&OverBar;}{x}+Q_{s}s$

$g=\stackrel{\&OverBar;}{g}+Q_{g}t$

wherein,is a model of the average shape of the object,is the average texture model, Q, corresponding to the average shape_sAnd Q_gAnd the transformation matrix is formed by the shape and texture principal component characteristic components obtained by principal component analysis and calculation. s is a shape parameter for controlling the shape change; t is a texture parameter for controlling texture change; x is a shape model, g is a texture model; and positioning and segmenting the region of interest on the test data by using the active appearance model to obtain initial results of the layers 1 to 7.

6. The retinal OCT image optic papillary structure segmentation method based on the hybrid model as claimed in claim 1, wherein the specific steps adopted in step 3) are as follows:

first, a retina image is defined as a three-dimensional matrix I (X, Y, Z) with a size of X × Y × Z, where X, Y, and Z are spatial coordinates, and X, Y, Z are the numbers of voxels in three directions, respectively;

the surface to be inspected is defined as a function S (X, Y), X ∈ { 0., X-1}, Y ∈ { 0., Y-1}, and S (X, Y) ∈ { 0., Z-1 };

the parameter Δ x defines a smoothing constraint condition in the x direction, and the parameter Δ y defines a smoothing constraint condition in the y direction; the maximum distance between two adjacent surfaces in the x direction meets the condition that | S (x +1, y) -S (x, y) | is less than or equal to Δ x, and the maximum distance in the y direction meets the condition that | S (x, y +1) -S (x, y) | is less than or equal to Δ y;

then, a vertex-weight directed graph G (V, E) is created based on the image voxels, comprising a set V of vertices V and a set E of edges E, in which directed graph any vertex V ∈ V corresponds to the image I (V, E)x, y, z), and any arc<v_i，v_j>∈ E connects two vertices v_i、v_jThe cost value c (x, y, z) of each vertex V ∈ V, { V (x, y, z) | (z > 0) } is calculated according to the gradient magnitude of the OCT image, which shows the possibility that a pixel point does not belong to the target surface, and the weight value of each vertex V ∈ V is calculated according to the cost value, as shown in the following formula (2):

$w(x,y,z)=\left\{\begin{array}{cc}c(x,y,z),& ifz=0\\ c(x,y,z)-c(x,y,z-1)& otherwise\end{array}\right.---\left(2\right)$

and (3) converting the problem of finding the optimal surface into a step of retrieving a minimum cost closed set in the directed graph G, and segmenting the interested region in the image.

7. The retinal OCT image optic papillary structure segmentation method based on the hybrid model as claimed in claim 5, further comprising the step of segmenting the optic disc image:

detecting the boundary of the video disc image and hiding the layer information in the part of the video disc area; and obtaining a projection image in the z direction between the layers 6 and 7 according to the layering result, and segmenting the optic disc region image by using a shape prior model algorithm.