CN108629769B - Method and system for optic disc location in fundus images based on best sibling similarity - Google Patents

Abstract

The invention discloses a fundus image optic disc positioning method and system based on the best sibling similarity, comprising the following steps: Step (1): selecting a healthy fundus image, and extracting the minimum circumscribed rectangle according to the minimum circumscribed rectangle including the optic disc region of the healthy fundus image The image in the rectangle is used as the template image; Step (2): Preprocess both the template image and the image to be matched; Step (3): Calculate the best sibling similarity between the template image and the image to be matched; Step (4) : Search in the image to be matched, and the area corresponding to the maximum sibling similarity is the final positioning result. A new concept was put forward for optic disc localization in fundus images and good results were obtained.

Description

Method and system for optic disc location in fundus images based on best sibling similarity

technical field

The present invention relates to a fundus image optic disc location method and system based on Best-BuddiesSimilarity.

Background technique

In recent years, the location and identification of physiological structures based on fundus images have been widely studied and applied, which is of great significance and role in the timely prevention and diagnosis of cataract, glaucoma and other blinding ophthalmic diseases. However, due to the increasing number of patients suffering from ophthalmic diseases, and the number of ophthalmologists cannot meet the needs of a large number of patients, more and more experts in the field of computer vision and medicine pay attention to and study the automatic computer recognition and diagnosis system based on medical images. In fundus images, the optic disc is one of the most important physiological structures, so efficient and accurate positioning of the optic disc is a very important step in the automatic analysis and processing of fundus images.

Researchers at home and abroad have proposed a variety of optic disc localization detection methods according to the characteristics of the fundus optic disc. These methods can be roughly divided into three categories: (1) optic disc localization methods based on blood vessel characteristics; (2) optic disc localization based on optic disc appearance characteristics Methods; (3) An optic disc localization method that comprehensively utilizes the blood vessel information of the fundus image and the characteristics of the optic disc.

The algorithm based on blood vessel characteristics utilizes the special structure and orientation information of blood vessels in the optic disc area of the fundus. When the appearance characteristics of the optic disc are not obvious due to interference factors such as lesions or noise in the image, it can still effectively and accurately locate the optic disc. Most of them are based on strict geometric templates, which leads to high algorithm complexity of such algorithms. The construction of the geometric template needs to be based on the accurate segmentation of the blood vessel structure. Regardless of whether there is a diseased area in the image, the complete and accurate blood vessel segmentation itself is a relatively difficult and complex work, which leads to the characteristics of blood vessels. The optic disc localization method usually requires a long processing time and cannot locate and detect the fundus optic disc in real time, but also makes the automatic optic disc localization work more complicated.

The algorithm based on the appearance features of the optic disc focuses on the brightness, color, shape, size and texture characteristics of the optic disc itself. However, when the lesion area in the fundus image is similar to the brightness of the optic disc, it is very likely that the lesion area will be detected by mistake only by the brightness, shape and other characteristics of the optic disc, resulting in failure to locate the optic disc. As a result, the brightness and integrity of the optic disc are disturbed and destroyed, which also seriously affects the accuracy of optic disc positioning.

Therefore, it is an urgent problem for those skilled in the art to reduce the interference of the lesion area in the fundus image to the optic disc positioning work, avoid the accurate segmentation of the blood vessel structure in the fundus image, reduce the complexity of the algorithm, and reduce the workload.

SUMMARY OF THE INVENTION

Based on the problems existing in the traditional fundus image optic disc location method, the present invention proposes a fundus image optic disc location method and system based on the best sibling similarity. The optimal sibling similarity does not only consider the appearance of the fundus optic disc, but optimizes and improves the optimal sibling similarity measurement method in the algorithm through a new template matching idea, combined with the gradient change characteristics of the optic disc of the fundus image. For the first time, this algorithm is applied to the field of optic disc positioning in fundus images, and the RGB appearance difference and spatial position difference between the template image and the image to be matched are calculated. The localization of the optic disc in the image has better robustness, while avoiding the segmentation of the blood vessel structure in the fundus image, and has a higher localization success rate and lower algorithm complexity.

Based on the idea of template matching, the present invention selects a healthy optic disc image as the template image, takes the fundus image to be located as the image to be matched, calculates the best sibling similarity between the template image and the image to be matched, and searches for the maximum value of the similarity The area is the optic disc area. The invention provides a new idea for optic disc location in fundus images.

In order to achieve the above object, the present invention adopts the following technical solutions:

As a first aspect of the present invention, a fundus image optic disc location method based on the best sibling similarity is provided;

The optic disc localization method in fundus images based on the best sibling similarity includes the following steps:

Step (1): select a healthy fundus image, and extract the image in the minimum circumscribed rectangle as a template image according to the minimum circumscribed rectangle of the healthy fundus image that includes the optic disc region;

Step (2): preprocess both the template image and the image to be matched;

Step (3): calculate the best sibling similarity between the template image and the image to be matched;

Step (4): Search in the image to be matched, and the area corresponding to the maximum value of the best sibling similarity is the final positioning result.

In the step (1), the image to be matched is the fundus image where the optic disc to be located is located.

In the step (2), both the template image and the image to be matched are preprocessed, and the size of the template image and the image to be matched is adjusted according to the size of the image sub-block, that is, the length and width of the image are in pixels, The size of the template image and the image to be matched is an integer multiple of the size of the image sub-block; then the template image and the template image are divided into several non-overlapping image sub-blocks of the same size, and then each image sub-block is divided into It is regarded as a point, thereby forming a point set composed of image sub-blocks, which are respectively the point set corresponding to the template image and the point set corresponding to the image to be matched;

In the step (3), based on the point set corresponding to the template image formed in step (2) and the point set corresponding to the image to be matched, calculate the point set corresponding to the template image and the point set corresponding to the image to be matched. Find the best sibling pair (Best-BuddiesPairs) of the two point sets, and calculate the best sibling similarity according to the number of best sibling pairs;

和

其中

表示模板图像的特征点集，

表示待匹配图像中候选区域的特征点集，U,V分别代表点集中特征点的个数，其中r_i，s_j∈R^d。存在这样一个点对{r_i∈R，s_j∈S}，当r_i到集合S的最近邻点为s_j并且s_j到集合R的最近邻点也为r_i时，那么称这样的点对为最佳兄弟点对(Best-BuddiesPairs,BBP)。The definition of the best sibling point pair is as follows: first define two point sets,

and

in

the set of feature points representing the template image,

Represents the feature point set of the candidate region in the image to be matched, U and V respectively represent the number of feature points in the point set, where r _i , s _j ∈ R ^d . There is such a point pair {ri _i ∈ R, s _j ∈ S}, when the nearest neighbor of _ri to set S is s _j and the nearest neighbor of s _j to set R is also _ri , then we call such a The pair is the best sibling pair (Best-BuddiesPairs, BBP).

The mathematical expression for judging the best sibling point pair Seg(r _i , s _j , R, S) is:

Among them, NN(r _i , S)=argmin d(r _i , s), s∈S represents the method for discriminating the nearest neighbors from _ri to point set S, and d( _ri , s) represents an arbitrary distance metric , ∧ represents AND operation, NN(r _i , S)=s _j represents the nearest neighbor point of _ri in point set S is s _j , if NN(r _i , S)=s _j ∧NN(s _j , R )=r _i indicates that _ri and s _j are the nearest neighbors to each other, that is, the best sibling pair, then Seg(r _i , s _j , R, S)=1, if it is not the best sibling pair, then Seg (r _i , s _j , R, S)=0.

By counting the total number of best sibling point pairs between point sets R and S, and then normalizing the total number of point pairs obtained by processing, the result is the result value of the best sibling similarity, and the best sibling similarity measure The value BBS(R, S) is expressed as the following formula (2):

Represent each candidate area in the template image and the image to be matched as a point in the xyRGB space, and calculate the value of the best sibling similarity measure between the two point sets. First, you need to get the distance value between each point pair , divides the distance metric between two points into three parts, one is the difference in color gray value between the two points, the other is the spatial position difference between the two points, and the third is the image gradient difference.

The distance value d(r _i , s _j ) between the points in the point set corresponding to the template image and the points in the point set corresponding to the image to be matched, is as follows formula (3):

表示模板图像对应的点集中的点的颜色像素值；

表示待匹配图像对应的点集中的点的颜色像素值；

表示模板图像对应的点集中的点的空间位置；

表示待匹配图像对应的点集中的点的空间位置；

表示模板图像对应的点集中的点的梯度值；

表示待匹配图像对应的点集中的点的梯度值；β表示为权重系数。in,

Represents the color pixel value of the point in the point set corresponding to the template image;

Represents the color pixel value of the point in the point set corresponding to the image to be matched;

Represents the spatial position of the point in the point set corresponding to the template image;

Represents the spatial position of the point in the point set corresponding to the image to be matched;

Represents the gradient value of the point in the point set corresponding to the template image;

represents the gradient value of the point in the point set corresponding to the image to be matched; β represents the weight coefficient.

In the step (4), based on the step (3), the region with the highest similarity value is searched column by column in the image to be matched, which is the final positioning result.

In the step (4), the similarity image is constructed: the level of the best sibling similarity is displayed in the form of color gradient, the color of the part with high similarity is dark, and the color of the part with low similarity is light. Find the area with the darkest color, which is the area with the highest degree of matching with the template image.

As a second aspect of the present invention, a fundus image optic disc positioning system based on the best sibling similarity is provided;

A fundus image optic disc localization system based on the best sibling similarity, comprising: a memory, a processor, and computer instructions stored in the memory and executed on the processor, the computer instructions are executed by the processor to complete any of the above methods. the steps described.

As a third aspect of the present invention, a computer-readable storage medium is provided;

A computer-readable storage medium on which computer instructions run, and when the computer instructions are executed by a processor, completes the steps described in any of the above methods.

The beneficial effects of the present invention are:

(1) The present invention utilizes the template matching method based on the best sibling similarity, which is a similarity measurement algorithm with strong robustness and high operating efficiency, which can effectively overcome background clutter and occlusion , this feature is applied to the optic disc location of the fundus image. Combined with the characteristics of the optic disc area, the algorithm is optimized by introducing the image gradient based on the first derivative, which can not only overcome the changes in the lesion area and the illumination change in the fundus image. It can also avoid the high computational complexity caused by the accurate segmentation of the blood vessel structure in the fundus image, and the defect that real-time optic disc positioning cannot be realized. In the step (2), image sub-blocks are used to form a point set, instead of a point set composed of pixels in the traditional algorithm.

(2) After optimizing the best sibling similarity measurement method, the present invention applies it to the work of locating the optic disc in the fundus image. The fundus image optic disc location method based on the best sibling similarity proposed by the present invention has achieved good results. The present invention compares the experimental results with other methods for locating the optic disc. The present invention's localization success rate and localization error show very outstanding performance, and its localization success rate is on the two public data sets of DRIVE and DIARETDB1. are 100% and 97.7%, respectively, and the localization average errors are 10.4 and 12.9, respectively, in pixels.

Description of drawings

The accompanying drawings that form a part of the present application are used to provide further understanding of the present application, and the schematic embodiments and descriptions of the present application are used to explain the present application and do not constitute improper limitations on the present application.

FIG. 1 is a flow chart of a video disc positioning method of the present invention.

Figure 2(a) and Figure 2(b) are the extraction of template images.

Fig. 3(a)-Fig. 3(h) are the steps of positioning the optic disc according to the present invention.

Fig. 4(a)-Fig. 4(l) are some examples of optic disc localization in the DRIVE dataset of the present invention.

Figures 5(a)-5(t) are some examples of optic disc localization in the DIARETDB1 dataset of the present invention.

Detailed ways

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

As shown in Figure 1, in a fundus image optic disc location method based on Best-Buddies Similarity:

(1) Construction of the template image: In the healthy fundus image, according to the marking of the optic disc area by the doctor, the minimum circumscribed rectangle containing the optic disc area is taken as the template image, as shown in Figure 2(a) and Figure 2(b), the rectangle is the template image.

(2) Image preprocessing: After step (2), in order to ensure that the points in the point set are sufficient for matching without affecting the accuracy of the experimental results, and at the same time to ensure that the amount of calculation is small, the size of the divided sub-blocks should be based on the Select the size of the template to make appropriate adjustments. When the template image is large, the size of the sub-block can be 3 or even larger; and when the template image is small, a single pixel can be directly used as a point. The size of this value can be determined according to specific application scenarios. Then, according to the size of the sub-block, re-adjust the size of the original template image and the image to be matched, so that the size of the original image is an integer multiple of the size of the sub-block.

Assuming that the size of the image sub-block is 3*3, according to this size, the length and width of the template image and the image to be matched are adjusted to an integer multiple of the length and width of the image sub-block. For example: the template image is 523*478, resized to 522*477.

和

其中

表示模板图像的特征点集，

表示待匹配图像中候选区域的特征点集，U,V分别代表点集中特征点的个数，其中r_i，s_j∈R^d。存在这样一个点对{r_i∈R，s_j∈S}，当r_i到集合S的最近邻点为s_j并且s_j到集合R的最近邻点也为r_i时，那么称这样的点对为最佳兄弟点对(Best-BuddiesPairs,BBP)。判断最佳兄弟点对的数学表达式为：(3) Calculate the best sibling similarity between point pairs: The best sibling similarity depends on the number of best sibling pairs between two point sets, and the best sibling pair is defined as follows: First define two set of points,

and

in

the set of feature points representing the template image,

Represents the feature point set of the candidate region in the image to be matched, U and V respectively represent the number of feature points in the point set, where r _i , s _j ∈ R ^d . There is such a point pair {ri _i ∈ R, s _j ∈ S}, when the nearest neighbor of _ri to set S is s _j and the nearest neighbor of s _j to set R is also _ri , then we call such a The pair is the best sibling pair (Best-BuddiesPairs, BBP). The mathematical expression for judging the best sibling pair is:

Among them, NN(r _i , S)=argmin d(r _i , s), s∈S represents the method for discriminating the nearest neighbor points from _ri to point set S, where d( _ri , s) represents any one Distance metric, ∧ represents AND operation, NN(r _i , S)=s _j represents the nearest neighbor point of _ri in point set S is s _j , NN(r _i , S)=s _j ∧NN(s _j , R)=ri means that _ri and _{s j are} the nearest neighbors to each other, that is, the best sibling pair, and vice versa, then Seg(i, j)=1, if it is not the best sibling pair, then Seg( i,j)=0. By counting the total number of best sibling point pairs between point sets R and S, and then normalizing the total number of point pairs obtained by processing, the result is the result value of the best sibling similarity, and the best sibling similarity measure The value expression is as follows in formula (2):

We represent each candidate region in the template and the image to be matched as a point in the xyRGB space, and calculate the value of the best sibling similarity measure between the two point sets. First, we need to get the distance value between each point pair. , combined with the optic disc feature of the fundus image, the image gradient based on the first derivative is introduced, and the distance measure between two points is divided into three parts, one is the difference in color gray value between them, and the other is the difference between them The spatial position difference, and the third is the image gradient difference. The specific distance measurement between point pairs is as follows:

The superscript A represents the color pixel value difference between the feature points, the superscript L represents the spatial position distance difference between the feature points, normalized to [0,1], β represents the weight coefficient, and the superscript G represents the gradient values of ri and _{s j of} the sub-block.

(4) According to the value of the best sibling similarity, a similarity image is constructed, and the region with the largest similarity value is searched for the final positioning result. The positioning is successful if the distance between the center of the located area and the center of the optic disc is less than 60 pixels.

Fig. 3(a), Fig. 3(e) are the images to be matched, Fig. 3(b), the rectangular area in Fig. 3(f) is the template image,

Figures 3(c) and 3(g) are the similarity images, and the crosses in Figures 3(d) and 3(h) are the final positioning results.

Fig. 4(a), Fig. 4(c), Fig. 4(e), Fig. 4(g), Fig. 4(i), Fig. 4(k) are the correct positions of the optic disc,

Figure 4(b), Figure 4(d), Figure 4(f), Figure 4(h), Figure 4(j), Figure 4(l) The crosses in Figure 4(l) are the experimental results of the present invention.

Figure 5(a), Figure 5(c), Figure 5(e), Figure 5(g), Figure 5(i), Figure 5(k), Figure 5(m), Figure 5(o), Figure 5 (q), the cross in Fig. 5(s) is the correct position of the optic disc,

Figure 5(b), Figure 5(d), Figure 5(f), Figure 5(h), Figure 5(j), Figure 5(l), Figure 5(n), Figure 5(p), Figure 5 (r), the cross in Fig. 5(t) is the experimental result of the present invention.

The above descriptions are only preferred embodiments of the present application, and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included within the protection scope of this application.

Claims (7)

1. the fundus image optic disc positioning method based on the best sibling similarity, is characterized in that, comprises the following steps: Step (1): select a healthy fundus image, and extract the image in the minimum circumscribed rectangle as a template image according to the minimum circumscribed rectangle of the healthy fundus image that includes the optic disc region; Step (2): preprocess both the template image and the image to be matched; Step (3): calculate the best sibling similarity between the template image and the image to be matched; The best sibling similarity depends on the number of best sibling pairs between two point sets, and the best sibling pair is defined as follows: first define two point sets,

and

in

the set of feature points representing the template image,

Represents the feature point set of the candidate area in the image to be matched, U, V respectively represent the number of feature points in the point set, there is such a point pair {r _i ∈ R, s _j ∈ S}, when ri _i is the closest to the set S When the neighbor point is s _j and the nearest neighbor point from s _j to the set R is also ri _i , then such a point pair is called the best sibling pair (Best-Buddies Pairs, BBP); the mathematics of judging the best sibling pair The expression is:

Among them, NN(r _i , S)=argmin d(r _i , s), s∈S represents the method for discriminating the nearest neighbor points from _ri to point set S, where d( _ri , s) represents any one Distance metric, ∧ means AND operation, NN(r _i , S)=s _j means that the nearest neighbor point of _ri in point set S is s _j , if NN(r _i , S)=s _j ∧NN(s _j , R)=r _i means that _ri and s _j are the nearest neighbors to each other, that is, the best sibling pair, then Seg(r _i , s _j , R, S)=1, if it is not the best sibling pair, Then Seg(r _i , s _j , R, S)=0; by counting the total number of best sibling point pairs between point sets R and S, and then normalizing the total number of point pairs obtained, the result is The resulting value of the best sibling similarity, and the best sibling similarity measure is expressed as formula (2):

Represent each candidate area in the template and the image to be matched as a point in the xyRGB space, and calculate the value of the best sibling similarity measure between the two point sets. First, you need to get the distance value between each point pair, Combined with the optic disc feature of the fundus image, an image gradient based on the first derivative is introduced, and the distance metric between two points is divided into three parts, one is the difference in color gray value between them, and the other is the difference between them. Spatial position difference, the third is the image gradient difference; the specific distance measure between point pairs is as follows:

in,

Represents the color pixel value of the point in the point set corresponding to the template image;

Represents the color pixel value of the point in the point set corresponding to the image to be matched;

Represents the spatial position of the point in the point set corresponding to the template image;

Represents the spatial position of the point in the point set corresponding to the image to be matched;

Represents the gradient value of the point in the point set corresponding to the template image;

represents the gradient value of the point in the point set corresponding to the image to be matched; β represents the weight coefficient; Step (4): Search in the image to be matched, and the area corresponding to the maximum value of the best sibling similarity is the final positioning result. 2. The fundus image optic disc location method based on the best sibling similarity as claimed in claim 1, wherein in the step (1), the image to be matched is the fundus image where the optic disc to be located is located. 3. the fundus image optic disc positioning method based on best brother similarity as claimed in claim 1, is characterized in that, in described step (2), template image and image to be matched are all carried out preprocessing, according to image sub-block The size of the image, that is, the length and width of the image are in pixels, adjust the size of the template image and the image to be matched, so that the size of the template image and the image to be matched are both integer multiples of the size of the image sub-block; then the template image and The template image is divided into several non-overlapping image sub-blocks with the same size and size, and then each image sub-block is regarded as a point, thereby forming a point set composed of image sub-blocks, which are the corresponding point sets of the template image. The point set corresponding to the image to be matched. 4. the fundus image optic disc positioning method based on the best brother similarity as claimed in claim 1, is characterized in that, in described step (4), based on step (3), in the image to be matched column by column search similarity The area with the highest value is the final positioning result. 5. the fundus image optic disc positioning method based on best brother similarity as claimed in claim 1, is characterized in that, in described step (4), construct similarity image: show best brother similarity by color gradient form The color of the part with high similarity is dark, and the color of the part with low similarity is light. Find the area with the darkest color from the image to be matched by visual observation, that is, the area with the highest matching degree with the template image. 6. the fundus image optic disc positioning system based on the best sibling similarity, is characterized in that, comprises: memory, processor and the computer instruction that is stored on memory and runs on processor, when described computer instruction is run by processor, Complete the steps described in any of the methods of claims 1-5 above. 7. A computer-readable storage medium, wherein computer instructions are stored thereon, and when the computer instructions are executed by a processor, the steps described in any one of the methods of claims 1-5 are completed.

Images