CN105551038A - Method for fully automatically classifying and segmenting retinal branch artery obstruction based on three-dimensional OCT image - Google Patents

Abstract

The invention discloses a method for fully automatic classification and segmentation of retinal branch arterial occlusion based on three-dimensional OCT images. The AdaBoost classifier is used to automatically classify the acute phase and atrophic phase of branch retinal artery occlusion; the segmentation of the acute phase of branch retinal artery occlusion: firstly, the Bayesian posterior probability is used to initialize the segmentation of the blocked area; then based on the graph The search-graph-cut algorithm accurately segmented the blocked area; (4) Segmentation of the atrophic phase of retinal branch artery occlusion: the blocked area in the atrophic phase was automatically segmented by establishing an inner retinal thickness model. The invention can accurately classify and segment retinal branch arterial occlusion areas, and can replace manual classification and segmentation.

Description

A method for fully automatic classification and segmentation of branch retinal artery occlusion based on 3D OCT images

technical field

The present invention relates to the classification of lesions in retinal images of SD-OCT (frequency-domain optical coherence tomography) and the segmentation method of lesion areas, in particular to a method for fully automatic classification and segmentation of retinal branch artery occlusion based on three-dimensional OCT images, belonging to Method for classifying and segmenting retinal images Technical field.

Background technique

Branch retinal artery occlusion is one of the acute diseases in ophthalmology. It has a poor prognosis, rapid onset, and usually painless monocular visual impairment. Retinal artery occlusion interrupts the nutrient supply of the corresponding retinal area, leading to hypoxia and ischemia in the local area of the retina, resulting in edema and rapid death of retinal cells, resulting in visual dysfunction.

So far, most of the work related to branch retinal artery occlusion has focused on the qualitative analysis of branch retinal artery occlusion, such as: H. Chen et al. proposed a framework for analyzing the light intensity of each layer of the retina in OCT images; CKSLeng et al. Measuring macular and peripapillary retinal nerve fiber layer thickness and visual acuity to investigate the relationship between retinal structure and function in patients with branch retinal artery occlusion; B. Asefzadeh and K. Ninyo Analysis of longitudinal fundus of peripapillary nerve fiber layer thickness in branch retinal artery occlusion Change.

These methods are all qualitative analysis of branch retinal artery occlusion and cannot detect and segment the occluded area fully automatically. Therefore, accurate quantitative information such as shape, size and location, etc. of the obstructed area cannot be provided to the clinician. In general, the current methods for branch retinal artery occlusion have the following defects: (1) most algorithms do not classify branch retinal artery occlusion (acute phase and atrophic phase), and the organizational structure of retina in different stages of branch retinal artery occlusion A big difference. (2) Most of the methods are not fully automatic, relying on manual measurement or marking. (3) Most of the algorithms do not perform specific analysis on the occlusion area of branch retinal artery occlusion.

Retinal branch arterial occlusion is mostly caused by embolism or thrombus, and the degree of visual impairment and fundus performance depends on the location and degree of obstruction. If the blocked area can be automatically and accurately segmented, it can help doctors make a diagnosis and formulate a corresponding treatment plan, which is very meaningful for the recovery of the patient's vision. However, the shape, size, and location of the blocked area of retinal branch artery occlusion are arbitrary, and the boundary between the blocked area and surrounding tissues is blurred, and the retinal OCT image itself is noisy. Therefore, fully automated segmentation of branch retinal artery occluded regions is a challenging task.

Contents of the invention

In view of the deficiencies in the prior art, the purpose of the present invention is to provide a method for fully automatic classification and segmentation of retinal branch artery occlusion based on three-dimensional OCT images, which can accurately classify and segment retinal branch artery occlusion areas, and can replace manual classification and segmentation.

In order to achieve the above object, the present invention is achieved through the following technical solutions:

A method for fully automatic classification and segmentation of retinal branch artery occlusion based on three-dimensional OCT images of the present invention, including

Include the following steps:

(1) Preprocessing: the retina is layered through the graph search algorithm (existing algorithm), and then each layer of the retina is flattened according to the pigment epithelium;

(2) Use AdaBoost (an algorithm that produces a final strong classifier by iterating a weak classifier) classifier to automatically classify the acute phase and atrophic phase of branch retinal artery occlusion;

(3) Segmentation of the acute stage of branch retinal artery occlusion: first, the Bayesian posterior probability is used to initialize the segmentation of the blocked area; then, the blocked area is accurately segmented based on the graph search-graph cut algorithm;

(4) Segmentation of retinal branch arterial occlusion in atrophic phase: the blocked area in atrophic phase is automatically segmented by establishing an inner retinal thickness model.

In step (1), the cost function of the graph search algorithm is defined as:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; +</span>\underset{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

where v is a voxel, S is the required surface, c _v is an edge-based cost that is inversely related to the likelihood that S contains voxel v, (p,q) is a pair of adjacent voxels, N is the set of voxels in the image, p and q are both in the image N, h _p,q are the cost of changing the shape of the surface S on p, q, S(p) is the position of the voxel p on the surface S.

In step (2), the method for automatically classifying the acute phase and atrophic phase of branch retinal artery occlusion using the AdaBoost classifier is as follows:

(a) extracting texture features, shape features and position features of the retina;

(b) Use the AdaBoost classifier to train and select the features in step (a), and classify a total of 23 cases of 3D OCT images of branch retinal artery occlusion, including 12 cases of acute stage images of branch retinal artery occlusion, and 11 cases of branch retinal artery occlusion images. Image of blocked atrophy phase. And use the one-out method, that is, select a three-dimensional image of a patient for testing each time, and use the remaining images for training, so that each image can be divided into acute phase of branch retinal artery occlusion or atrophy phase.

In step (3), the method of initializing and segmenting the blocked area using Bayesian posterior probability is as follows:

(3-1) Use the Bayesian posterior probability to estimate the probability that each voxel belongs to the blocked area. The calculation formula of this probability probability is:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; P</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">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; P</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">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; P</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">\; no</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</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>}<span\; class="notranslate">\; )</span>$

Among them, I _p represents the brightness value of the voxel, occ represents blocking, and non represents not blocking; P(occ) and P(non) represent the probability of blocking areas and non-blocking areas in the entire image; P(I _p |occ) and P(I _p |non) represents the probability that a voxel p with brightness I _p belongs to the blocked area or not; in the training phase, each voxel in the inner retina is marked as blocked or not blocked, and this label is based on clinical Manually labeled under the guidance of a doctor; during the test phase, each voxel of the inner retina is given a probability between 0 and 1 to estimate the likelihood that it is an occluded region, resulting in a probability map of the entire image, where the The probability map is used as a constraint in the cost function of graph search-graph cut during segmentation;

(3-2) After obtaining the probability map, first, use the mean filter to smooth the image; then, use the threshold to separate the points with high probability, and these points are regarded as blocked areas; finally, through the morphological The opening and closing operations obtain initial segmentation results.

In step (3), based on the graph search-graph cut algorithm (existing algorithm), the maximum flow-min cut algorithm is used to automatically segment the image according to the minimum cost function. The cost function of the graph search-graph cut algorithm is defined as:

E(f)=E(Surfaces)+E(Regions)+E(Interactions)

(2)

Among them, E(Surfaces) represents all costs related to surface segmentation, E(Regions) represents the cost related to segmented regions, and E(Interactions) represents the cost of constraints between surfaces and regions;

Based on the initialization segmentation result, use the morphological erosion operation (prior art) to obtain the foreground seed points required by the graph search-graph cut algorithm, and then use the dilation operation (prior art) to obtain the foreground seed points required by the graph search-graph cut algorithm. The desired background seed point.

In step (4), the method of automatically segmenting the blocked area in the atrophy period by establishing the inner retinal thickness model is as follows:

(4-1) Image alignment: Since the retinal image is divided into left eye and right eye, the retinal structure is a mirror image; in order to obtain a unified inner retinal thickness model, the left eye image is flipped to make it consistent with the right eye;

(4-2) Establishing the thickness model of the inner retina: in the retinal OCT image, with the macula as the center, the thickness of the inner retina at different positions is recorded to form a 2D thickness model;

(4-3) Segmentation of blocked regions in the atrophic stage: select retinal OCT images of several normal eyes, use the average of their inner retinal thickness models as the standard inner retinal thickness model of normal human eyes, and compare with the inner retinal thickness of patients in the atrophic stage Compare to segment out the blocking area.

The present invention uses the AdaBoost classification algorithm to automatically distinguish the acute stage and the atrophic stage of branch retinal artery occlusion, and uses different segmentation methods to automatically segment it according to its tissue structure and texture features; combined with Bayesian posterior probability and graph search-graph Cutting algorithm to accurately segment the blocked area of retinal branch artery occlusion in the acute stage; and use the inner retinal thickness model to efficiently and quickly segment the blocked area in the atrophic stage; both classification and segmentation have high accuracy, so this method can replace Manual classification and segmentation can play an important auxiliary role in the diagnosis and treatment of clinically relevant ophthalmic diseases.

Description of drawings

Figure 1(a) is the effect of retinal delamination and flattening (delamination in the atrophic period);

Figure 1(b) shows the effect of retinal delamination and flattening (delamination in the acute phase);

Fig. 2 is the relationship diagram of posterior probability and voxel brightness;

Figure 3(a) is the original image of Bayesian initialization;

Figure 3(b) is a probability map of Bayesian initialization;

Figure 3(c) is the segmentation result diagram of Bayesian initialization;

Figure 4(a) is the inner retinal thickness model of the normal retina;

Figure 4(b) is the inner retinal thickness model of the first branch retinal artery occlusion patient in the atrophic stage;

Figure 4(c) is the inner retinal thickness model of the first branch retinal artery occlusion atrophic patient;

Figure 5(a) is the segmentation result of branch retinal artery occlusion in the acute stage (the first column is the original image, the second column is the reference standard, and the third column is the segmentation result);

Figure 5(b) shows the segmentation results of branch retinal artery occlusion and atrophy phase (the first column is the original image, the second column is the reference standard, and the third column is the segmentation result).

detailed description

In order to make the technical means, creative features, goals and effects achieved by the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.

The invention comprises the following four steps: image preprocessing, classification based on AdaBosst, segmentation of the acute stage of branch retinal artery occlusion and segmentation of the atrophy stage of branch retinal artery occlusion.

(1) Image preprocessing

In order to obtain the information of each layer of the retina, a graph search method is used to realize the layering of the retina. Its cost function is defined as:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; +</span>\underset{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

where S is the required surface and _cv is an edge-based cost that is inversely related to the likelihood that S contains voxel v. (p,q) is a pair of adjacent voxels. h _p,q is the cost of changing the shape of surface S over p,q.

The layers of the retina are then leveled against the pigment epithelium. The effect of retinal delamination and flattening is shown in Figure 1(a), Figure 1(b). Only the four layers used in the present invention are marked in the figure, which are the upper boundary of the retinal nerve fibers, the lower boundary of the outer reticular layer, the lower boundary of the outer nuclear layer and the lower boundary of the pigment epithelium from top to bottom. These 4 layers divide the retina into the inner retina between the first and second lines (top to bottom) and the outer retina between the third and fourth lines (top to bottom) Below).

(2) Classification based on AdaBoost

Because branch retinal artery occlusion has an acute phase and an atrophic phase, and the retinal structure and texture are different in both, it is necessary to classify branch retinal artery occlusion first. The present invention uses AdaBoost classifier to classify the acute phase and atrophy phase of branch retinal artery occlusion, which includes two parts, specifically described as follows:

(a) Features

The present invention extracts 50 features for each voxel, including inner retinal thickness, voxel brightness, gray level co-occurrence matrix energy, entropy, inertia, correlation and Gabor filter transformation, as shown in Table 1. These features describe the texture, structure and location information of each voxel.

Table 1 Features used for classification

Feature number feature description 1 inner retinal thickness 2 voxel brightness 3-34 Gray level co-occurrence matrix energy, entropy, inertia, mean and standard deviation in 8 directions 35-50 Gabor (Gabor) filter output, two center frequencies, 8 directions

(b) AdaBoost classification

Use the AdaBoost classifier to train and select the features, and use the one-out method to select one patient's three-dimensional data for testing each time, and use the remaining data for training, and divide each data into acute stage of branch retinal artery occlusion or Atrophy period.

(3) Segmentation in the acute phase of branch retinal artery occlusion

In the acute stage of branch retinal artery occlusion, the OCT image shows that the local area of the inner retina has increased brightness, and this brightened area is the blocked area. The invention combines the Bayesian posterior probability and the graph search-graph cut algorithm to automatically segment the blocking area, including the following two steps: initialization and precise segmentation.

(a) Initial segmentation of blocked regions with Bayesian posterior probability

The present invention uses Bayesian posterior probabilities to estimate the likelihood that each voxel belongs to an occluded region. The formula for calculating this probability is:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; P</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">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; P</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">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; P</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">\; no</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</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>}<span\; class="notranslate">\; )</span>$

Among them, I _p represents the brightness value of the voxel, occ represents blocking, and non represents not blocking; P(occ) and P(non) represent the probability of blocking areas and non-blocking areas in the entire image; P(I _p |occ) and P(I _p |non) represents the probability that a voxel p with brightness I _p belongs to an occluded area and is not an occluded area. Figure 2 is a plot of the posterior probability versus voxel brightness.

Referring to Fig. 3(a), during the training phase, each voxel of the inner retina was marked as occluded or not occluded, which was manually labeled according to the clinician's guidance. In the testing phase, each voxel of the inner retina is given a probability between 0 and 1 to estimate its probability of being an occluded region, resulting in a probability map for the entire image, as shown in Fig. 3(b). This probability map will be used as a constraint in the cost function of the segmentation.

After obtaining the probability map, the initial segmentation results are obtained through some post-processing operations. First, the image is smoothed using a mean filter. Then, a threshold of 0.5 is used to separate the points with high probability, and these points are regarded as blocked areas. Finally, the initialization segmentation results are obtained through morphological opening and closing operations. The small scattered points outside the blocked area are removed by the open operation, and the blank points inside the blocked area are filled by the closed operation. The initialization result is shown in Figure 3(c).

(b) Automatic segmentation based on graph search-graph cut algorithm

This method uses the graph search-graph cut algorithm to accurately segment the blocked area, and the cost function is designed as:

E(f)=E(Surfaces)+E(Regions)+E(Interactions)

(2)

Among them, E(Surfaces) represents all costs related to surface segmentation, E(Regions) represents the cost related to segmented regions, and E(Interactions) represents the cost of constraints between surfaces and regions.

The foreground and background seed points required for graph search-graph cut are automatically obtained by the morphological algorithm. Based on the initial segmentation results, the foreground seed points are obtained by the morphological erosion operation, and then the background seed points are obtained by the dilation operation.

(4) Segmentation of branch retinal artery occlusion atrophy period

The atrophic phase of branch retinal artery occlusion shows a decrease in inner retinal thickness on OCT images, so it is reasonable to segment the obstructed area in the atrophic phase based on changes in thickness. However, since the thickness of the retina itself is different in different positions (thicker near the macula, thinner farther away, as shown in Figure 4(a)), it is easy to affect the judgment. The present invention proposes a method of using a thickness model to segment the obstructed area in the atrophic phase, including the following three steps: (1) Image alignment. Since the retinal image is divided into left and right eyes, its retinal structure is a mirror image. In order to obtain a unified inner retinal thickness model, in the present invention, the image of the left eye is flipped to make it consistent with the right eye. (2) Establish a model of inner retinal thickness. In the retinal OCT image, record the thickness of the inner retina at different positions centered on the macula to form a 2D thickness model. (3) Segmentation of the blocked area in the atrophic period. Select the retinal images of 20 normal eyes, use the average of their inner retinal thickness models as the standard inner retinal thickness model of normal human eyes, and use the inner retinal thickness model of atrophic patients (as shown in Figure 4(b) and (c)) The thickness is compared to segment the blocked area.

Experimental results

The present invention shared the data of 23 patients with branch retinal artery occlusion, including 12 cases in the acute stage and 11 cases in the atrophic stage. Diagnosed by experts and manually marked blocked areas as gold standard. The specific experimental results are as follows.

(a) Classification results of branch retinal artery occlusion

The classification results of acute phase and atrophic phase of branch retinal artery occlusion are shown in Table 2, and the overall correct rate of AdaBoost classifier is 87.0%.

Table 2 Classification performance for branch retinal artery occlusion

(b) Segmentation results of branch retinal artery occlusion area

The true positive rate TPVF and the false positive rate FPVF are used as the objective indicators of the evaluation method, which are calculated as follows:

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}{{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}{{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Among them, V _TP , V _FP , V _TN and V _FN are the volumes of true positive, false positive, true negative and false negative, respectively. The experimental results show that for the acute stage of branch retinal artery occlusion, the average true positive rate of this method is 91.1%, and the average false positive rate is 5.5%; for the atrophic stage, the average true positive rate of this method is 90.5%, and the average false positive rate is 91.1%. The rate was 8.7%. Part of the segmentation results are shown in Figure 5(a), (b).

So far, an automatic classification and segmentation method for 3D OCT images of branch retinal artery occlusion has been implemented and validated. The invention combines graph search algorithm, AdaBoost classifier, Bayesian posterior probability, graph search-graph cut algorithm and morphological algorithm to automatically classify and segment retinal branch arterial occlusion, and both classification and segmentation have higher accuracy Therefore, this method can replace manual classification and segmentation, and can play an important auxiliary role in the diagnosis and treatment of clinically relevant ophthalmic diseases.

The basic principles and main features of the present invention and the advantages of the present invention have been shown and described above. Those skilled in the industry should understand that the present invention is not limited by the above-mentioned embodiments. What are described in the above-mentioned embodiments and the description only illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will also have Variations and improvements all 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 (6)

1., based on the method that automated classification and the segmentation Branch Retinal Artery of three-dimensional OCT image block, it is characterized in that, comprise following step:

(1) pre-service: carry out layering to retina by graph search algorithm, then evens up each layer of retina according to pigment epithelial layer;

(2) acute stage using AdaBoost sorter to block Branch Retinal Artery and atrophy phase carry out automatic classification;

(3) Branch Retinal Artery blocks the segmentation of acute stage: first adopt Bayes posterior probability to carry out initialize partition to congested areas; Then cut algorithm based on graph search-Tu and Accurate Segmentation is carried out to congested areas;

(4) Branch Retinal Artery blocks the segmentation of atrophy phase: carry out auto Segmentation by setting up interior retinal thickness model to the congested areas of atrophy phase.

2. the method for the automated classification based on three-dimensional OCT image according to claim 1 and segmentation Branch Retinal Artery obstruction, it is characterized in that, in step (1), the cost function of described graph search algorithm is defined as:

$E\left(S\right)=\underset{v\&Element;S}{\&Sigma;}{c}_v+\underset{(p,q)\&Element;N}{\&Sigma;}{h}_{p,q}\left(S\right(p)-S(q\left)\right)$

Wherein, v is a voxel, and S is the surface of requirement, c _vbe a cost based on edge, comprise the possibility inverse correlation of voxel v with S, (p, q) is a pair adjacent voxel, and N is the set of voxel in image, p and q all in image N, h _p,qbe the cost of shape in the upper change of p, q of surperficial S, s (p) is the position of voxel p on surperficial S.

3. the method for the automated classification based on three-dimensional OCT image according to claim 1 and segmentation Branch Retinal Artery obstruction, it is characterized in that, in step (2), the method that the acute stage using AdaBoost sorter to block Branch Retinal Artery and atrophy phase carry out automatic classification is as follows:

A () extracts amphiblestroid textural characteristics, shape facility and position feature;

B () adopts AdaBoost sorter train the feature in step (a) and select, and a method is abandoned in use, namely select the 3-D view of a patient to test at every turn, remaining image is trained, thus is divided into by each image Branch Retinal Artery to block acute stage or atrophy phase.

4. the method for the automated classification based on three-dimensional OCT image according to claim 1 and segmentation Branch Retinal Artery obstruction, it is characterized in that, in step (3), the method adopting Bayes posterior probability to carry out initialize partition to congested areas is as follows:

(3 ?1) use Bayes posterior probability to estimate that each voxel belongs to the possibility of congested areas, and the computing formula of this possibility probability is:

$P\left(occ\right|I_p)=\frac{P\left(I_P\right|occ)\&CenterDot;P\left(occ\right)}{P\left(I_P\right|occ)\&CenterDot;P\left(occ\right)+P\left(I_P\right|non)\&CenterDot;P\left(non\right)}---\left(1\right)$

Wherein, I _prepresent the brightness value of voxel, occ represents obstruction, and non represents it is not block; P (occ) and P (non) represents congested areas in whole image and is not the probability of congested areas; P (I _p| occ) and P (I _p| non) represent that brightness is I _pvoxel p belong to congested areas and be not the probability of congested areas; In the training stage, interior amphiblestroid each voxel is marked as and blocks or be not block, and this mark is the hand labeled that instructs according to clinician; At test phase, probability between interior given one 0 to 1 of amphiblestroid each voxel estimates that it is the possibility of congested areas, thus obtain the probability graph of whole image, described probability graph is used in the cost function that graph search-Tu when splitting cuts as constraint;

After (3 ?2) obtain described probability graph, first, mean filter is used to carry out smoothed image; Then, separated by point high for probability by threshold value, these points are regarded as the region of obstruction; Finally, initialize partition result is obtained by morphologic open and close operation.

5. the method for the automated classification based on three-dimensional OCT image according to claim 4 and segmentation Branch Retinal Artery obstruction, it is characterized in that, in step (3), algorithm is cut based on graph search-Tu, with max-flow-minimal cut algorithm, carry out auto Segmentation according to cost function is minimum to image, the cost function that graph search-Tu cuts algorithm is defined as:

E(f)＝E(Surfaces)+E(Regions)+E(Interactions)(2)

Wherein, E (Surfaces) represents all costs relevant to surface segmentation, and E (Regions) represents the cost relevant to cut zone, the cost that E (Interactions) retrains between presentation surface and region;

Based on initialize partition result, with morphological erosion operation obtain described figure Sou Suo ?figure cut foreground seeds point needed for algorithm, then with expansive working obtain described figure Sou Suo ?figure cut background Seed Points needed for algorithm.

6. the method for the automated classification based on three-dimensional OCT image according to claim 1 and segmentation Branch Retinal Artery obstruction, it is characterized in that, in step (4), as follows by setting up the method that interior retinal thickness model carries out auto Segmentation to the congested areas of atrophy phase:

(4-1) image alignment: because retinal images divides left eye and right eye, its retinal structure becomes mirror image; In order to obtain retinal thickness model in unification, left-eye image being overturn, makes it consistent with right eye;

(4-2) set up interior retinal thickness model: in retina OCT image, centered by macula lutea, the interior amphiblestroid thickness at record diverse location place, forms the thickness model of a 2D;

(4-3) atrophy phase congested areas segmentation: the retina OCT image choosing several normal eyes, average as retinal thickness model in the normal eye of standard with retinal thickness model in them, compare with retinal thickness in atrophy phase patient, thus be partitioned into congested areas.