WO2012149687A1

WO2012149687A1 - Method for retinal vessel extraction

Info

Publication number: WO2012149687A1
Application number: PCT/CN2011/073694
Authority: WO
Inventors: 田捷; 向德辉; 杨鑫; 徐敏; 秦承虎
Original assignee: 中国科学院自动化研究所
Priority date: 2011-05-05
Filing date: 2011-05-05
Publication date: 2012-11-08

Abstract

A method for retinal vessel extraction, comprises steps: performing multi-scale filtration on a retinal image; acquiring a centerline of the retinal vessels, and initializing vessel cut by marks; constructing a graph according to the filtration image and the initially cut image, and performing calculation on the graph to acquire the vessels in the retinal image. The method in this invention effectively overcomes the influence of image noise and low contrast of vessels, and accurately extracts the retinal vessels, thereby assisting doctors to diagnose ocular diseases, diabetes, hypertension etc., and being of important application value in the field of vessel cut in cerebrovascular and cardiovascular images.

Description

Retinal blood vessel extraction method

Technical field

The invention relates to image processing technology, in particular to a method for extracting retinal blood vessels. Background technique

Retinal blood vessel extraction (segmentation) is the accurate extraction of blood vessels from the retina (the fundus) image. It is one of the important research contents in image processing, and it is also a typical research field in which image processing systems are applied clinically. Retinal blood vessel extraction results can be used to register images, and can also be used to diagnose ophthalmic diseases and other diseases (such as diabetes, high blood pressure, etc.). Therefore, retinal blood vessel extraction has attracted a lot of efforts from domestic and foreign researchers.

In 1989, Chaudhuri et al. (S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum, "Detection of blood vessels in retinal images using two-dimensional matched filters," IEEE Transactions on Medical Imaging , vol. 8, no. 3, pp. 263-269, 1989) proposed a method for detecting retinal blood vessels based on matched filters. The method constructs 12 two-dimensional discrete matched filter kernels to detect possible directions of blood vessels to enhance blood vessel and segmentation. Staal et al. (J. Staal, M. Abr^amoff, M. Niemeijer, M. Viergever, and B. van Ginneken, "Rail-based vessel segmentation in color images of the retina," IEEE Transactions on Medical Imaging, vol. 23, no. 4, pp. 501-509, 2004) A ridge-based automatic retinal blood vessel extraction method is proposed. The method first assumes that the ridges of the blood vessels are consistent with the centerline, and constructs the line element to divide the image into several segments. Then, a feature vector is constructed for each pixel based on the properties of these line elements and slices. Finally, the fully automated blood vessel extraction is achieved using the kNN method. Scares et al. (J. Soares, J. Leandro, R. Cesar, H. Jelinek, and M. Cree, "Retinal vessel segmentation using the 2-D Gabor wavelet and supervised classification," IEEE Transactions On Medical Imaging, vol. 25, no. 9, pp. 1214-1222, 2006) Multi-scale Gabor filtering is used to filter the retinal image to obtain multi-scale response and construct a high-level eigenvector. Finally, the Gaussian mixture model is used to classify The device completes the extraction of blood vessels. Mendot^a et al. (A. Mendon^a and A. Campilho, "Segmentation of retinal blood vessels by bringing the detection of centerlines and morphological reconstruction," IEEE Transactions on Medical Imaging, vol. 25, no. 9, pp. 1200 -1213, 2006) The linear detection filter first enhances the blood vessels of small blood vessels and low contrast, and then uses the gradient information to detect the center line of the blood vessels, and uses morphological methods and regional growth methods to obtain blood vessels. Ricci et al. (E. Ricci and R. Perfetti, "Retinal blood vessel segmentation using line operators and support vector classification," IEEE Transactions on Medical Imaging, vol. 26, no. 10, pp. 1357-1365, 2007) A directional linear detector is then constructed using two mutually perpendicular linear detectors, and finally the supervised classification of blood vessels is performed using the support vector machine method. Wang et al. (L. Wang, A. Bhalerao and R. Wilson, Analysis of retinal vasculature using a multiresolution hermite model, IEEE Transactions on Medical Imaging, vol. 26, no. 2, pp. 137-152, 2007) Resolution Hermite model to model and label blood vessels. Chen et al. (J. Chen, J. Tian, Z. Tang, J. Xue, Y. Dai, and J. Zheng, "Retinal vessel enhancement and extraction based on directional field," Journal of XRay Science and Technology, vol. 16, no. 3, pp. 189-201, 2008 ) First, the directional field and Gabor filtering methods are used to enhance the retinal image, and finally the vascular structure is reconstructed using morphological methods. Lam et al. (B. Lam and H. Yan, "A novel vessel segmentation algorithm for pathological retina images based on the divergence of vector fields," IEEE Transactions on Medical Imaging, vol. 27, no. 2, pp. 237-246 , 2008) Using the gradient vector field to detect the vascular structure and extract the centerline, using the centerline pruning technique to further eliminate the noise to obtain the segmentation result. In 2010, Lam et al. (B. Lam, Y. Gao, and A. Liew, "General Retinal Vessel Segmentation Using Regularization-Based Multiconcavity Modeling," IEEE Transactions on Medical Imaging, vol. 29, no. 7, pp. 1369-1381, 2010) again proposes a retinal vessel extraction method based on multi-concavity modeling . The method utilizes linear concavity and local normalized concavity methods to filter out noise in the retina and extract blood vessels based on statistical distribution.

The main idea of the image segmentation method based on graph theory is to map the image to the weighted directed graph (graph), the node of the pixel corresponding map in the image, the adjacent relationship between the pixels corresponds to the edge of the graph, the difference between the pixel features or The similarity corresponds to the weight on the edge, and then the various nodes are used to divide the nodes in the graph on the established graph to achieve the purpose of image segmentation. Image segmentation based on graph theory combines graph theory to easily deal with the advantages of local features and objective functions to easily process global information. At the same time, local features and global features are considered to achieve optimal results. The significance of graph theory network and objective function is The interactions are linked together. Both image local processing and image overall processing can be taken into consideration, and there are advantages not found in other methods. Boyov and V. Kolmogorov, "An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 26, no. 9, pp 1124-1137, 2004) In 2004, an interactive image segmentation method based on graph cut was proposed. This method considers the user's interaction on the one hand, allowing the user to make different points for the target object to be segmented and its background. The user input is used as a constraint; on the other hand, the edge information and the area information inside the image are considered. The graph cut algorithm uses the maximum stream/minimum cut algorithm to find the corresponding optimal segmentation effect. Due to the particularity and complexity of the vascular structure in the retinal image, this method is limited in the application of retinal blood vessels.

It is difficult to accurately extract the retinal vascular network in the prior art because of various noises in the retinal image and low contrast of some blood vessels, which causes the blood vessel extraction algorithm to fail in these regions. Summary of the invention

It is an object of the present invention to provide a retinal blood vessel extraction method which makes it possible to accurately extract blood vessels in a retinal image.

In order to achieve the above object, a retinal blood vessel extraction method comprises the steps of:

Multi-scale filtering of retinal images;

A centerline of the retinal blood vessel is obtained, and a segmentation of the blood vessel is initialized using a marker; a map is created based on the filtered image and the initially segmented image, and the map is calculated to obtain a blood vessel in the retinal image.

The invention effectively overcomes the influence of image noise and low contrast of blood vessels, accurately extracts retinal blood vessels, and can assist doctors in diagnosing ophthalmic diseases and diabetes, hypertension and the like. It has important application value in the fields of cerebrovascular images and blood vessel segmentation of cardiovascular images. DRAWINGS

1 is a flow chart of a method for extracting a retinal blood vessel of the present invention;

2 is an example of blood vessel extraction in a normal retinal image according to the present invention, FIG. 2(a) is a retinal image, FIG. 2(b) is a segmentation result obtained by using a threshold method, and FIG. 2(c) is a center line of a blood vessel, FIG. 2(d) is the segmentation result obtained by using the present invention, FIG. 2(e) is the result of manual segmentation by expert A, and FIG. 2(f) is the result of manual segmentation by expert B;

3 is an example of blood vessel extraction of a retinal image with pathological changes according to the present invention.

3(a) is the retinal image, Fig. 3(b) is the segmentation result obtained using the threshold method, Fig. 3(c) is the center line of the blood vessel; Fig. 3(d) is the segmentation result obtained by using the present invention; e) Manually segment the results for Expert A, Figure 3(f) for Expert B manually segmentation results. detailed description

In order to make the objects, the technical solutions and the advantages of the present invention more comprehensible, the present invention will be further described in detail below with reference to the accompanying drawings.

The core idea of the present invention is to propose a retinal blood vessel extraction method based on iterative graph cutting. Firstly, the retinal image is filtered by using a multi-scale method, and the center line of the blood vessel is extracted by using a threshold method and a refinement method, under the guidance of the center line. The image is iteratively labeled using a graph cut algorithm to obtain the blood vessel extraction results.

1 is a flow chart of an iterative map-based retinal blood vessel extraction technique provided by the present invention. Such an iterative map-based retinal vessel extraction technique provided in accordance with the present invention is described in detail below in conjunction with specific embodiments.

Step 1: Filter the retinal image using a multi-scale method;

A retinal grayscale image), its multi-scale representation is:

_Wherein,; s denotes variance (scale); * represents the convolution. A Hessian matrix for each pixel in the image is available:

And find the eigenvalue of the matrix ^^A ^, . The eigenvalues can be used to construct a multi-scale filtered response:

V{x, y) max v (x, y, s) where _in ^ _max represent the maximum and minimum scales of the scale space, respectively. Sex: When the radius of the blood vessel matches the scale S, the filtered value is the largest;

c(Ll ( ₅ ^ ₅ ^) ( ,^,Λ ¹ ) ), λ (x,y,s)≤ A, (xys) <0;

1 ο, other.

Where c is a constant. Step 2: First, the center line of the blood vessel is obtained, and the center line of the blood vessel can be obtained using the threshold method and the refinement method; then, a plurality of markers are used to initialize the segmentation of the blood vessel, and in the present invention, four markers are used to initialize the segmentation of the blood vessel.

Use the threshold method for the filtered image (χ, ≥t _A , and then use the refinement algorithm to get the center line of the blood vessel. For the center line with the number of pixels less than ^ from . _te , discard the center line, in order to eliminate the noise with large values. Then use the threshold method four types to mark the pixels corresponding to the image. The four types of markers ^ are: foreground seed, candidate foreground seed. Λ, candidate background seed ^ and background seed. 1) The pixel on the center line is marked as Λ; 2) In addition to the pixels marked as foreground, pixels with (x, y) ≥ t _w are marked as; ³ ) pixels with t _{w >} (x, _y) ≥ t are marked as

4) (^, <^ pixels are marked as .

Step 3: Create a map and calculate the map to obtain blood vessels in the retinal image.

According to the filtering result obtained in the above step 1 (x, and the initial division of the above step 2, the mapping G = <i, f>. With the pixel as the node V of the graph, the adjacent relationship between the pixels is used as the edge of the graph. The K-means clustering algorithm is respectively labeled as candidate foreground seeds and candidate background seed clusters, and calculates the shortest distance c of the candidate foreground seeds and candidate background seeds to the foreground seeds, respectively, and the shortest background seeds. Distance ^, . The junction of the connection graph; ^1 and two virtual sections Point S, the edge of T is defined as Z - / / € £. t—Unk is calculated as

U _s (p) =∞, U _T {P) = O,

= c

; and is a constant, where,

And satisfy 0 _{≤ ί3⁄4} ≤1≤ _ι . The edge of the node corresponding to two adjacent pixels in the connected image is defined as n - link e E o w- //"A: The calculation formula is:

1 where, respectively, represent multi-scale filtered values of nodes and corresponding pixels of g. Use the maximum flow/minimum cut algorithm for graph G to divide the nodes into two categories: source and sink. Change the last tag, the rule is: if it belongs to the source type and the tag is not the foreground seed point, it is marked as the candidate foreground seed point.:, otherwise it is marked as the foreground seed point Λ; if it belongs to the sink type and the mark is not the background seed point, then the mark Click ^ for the candidate background seed, otherwise mark as the background seed point. Calculate the local energy: ep( ^l p) = ^U p( ^l p) ^{+ Ke} N ( ^l _P ) where / represents the current pixel's mark;

Then u _p (l _p ) = D _h (p) is a constant; el _p m , W is a neighborhood of 7; the values are as follows:

Where ς· is a constant with a value of 0 _{≤ ί ≤} 1. Replace the marker according to the local energy model: foreground The seed point and the background seed point do not change the mark; for the candidate foreground seed point /, the two energy values ^ and ; 7^ divide the local energy of the pixel into three segments, and if the energy is less than a given threshold _Λ , ₂ , the mark is changed For the foreground seed point, if the energy is greater than another given threshold ^ then marked as a candidate background seed point, otherwise the mark is unchanged; the new mark can be summarized as follows: f

Similarly, for the candidate background seed point ^, the two energy values divide the local energy of the pixel into three segments. If the energy is less than a given threshold 3⁄4, ₂ , the flag is changed to the background seed point, if the energy is greater than the other given The threshold ^ , , is marked as a candidate foreground seed point, otherwise the flag is unchanged; the new flag can be summarized as follows:

% _e ,l ^{≤ e} p { ^l p = ^b c)> , ^e _P ( ^l p ^{= b} c) ^{< T} lb _c - determine whether the number of iterations is reached, if not, return; if yes, then the candidate foreground seed point Change to foreground seed point, candidate background seed point to background seed point, algorithm terminated.

To verify the validity and utility of the present invention, we conducted experiments on two internationally recognized retinal image databases (STARE and DRIVE). Both databases provide 20 retinal images for algorithm testing and two manually segmented vascular network data sets for reference. Figures 2 and 3 show two specific examples. Figure 2(a) shows a normal retinal image, Figure 2(b) shows the results of threshold segmentation, Figure 2(c) shows the centerline of the blood vessel, and Figure 2(d) shows the blood vessel extracted using the present invention. 2 (e) Expert A manually segmented blood vessels, Figure 2 (f) Expert B manually segmented blood vessels; Figure 3 (a) a pathologically altered retinal image, Figure 3 (b) is the result of using threshold segmentation, Figure 3 (c) for blood The center line of the tube, Fig. 3(d) is the blood vessel extracted using the present invention, Fig. 3(e) the blood vessel manually divided by the expert A, and Fig. 3(f) the blood vessel manually divided by the expert B. As can be seen from the figure, the retinal blood vessel extraction technique based on the iterative graph cut proposed by the present invention can effectively extract blood vessels in the retinal image. Tables 1 and 2 respectively show quantitative comparisons between the retinal blood vessel extraction technique based on the iterative graph cut proposed by the present invention and the prior methods. The evaluation method consists of three indicators: sensitivity, specificity and accuracy. Sensitivity refers to the ratio between the number of pixels in which the algorithm extracts blood vessels (expert B manually divided) and the number of pixels in which expert A manually divides the blood vessels in the retinal image field of view; specificity refers to the retinal image viewport Algorithm extraction (expert B manually splits) the ratio of the number of pixels in the background to the number of pixels manually separated by expert A; the precision refers to the algorithmic extraction in the retinal image viewport (expert B manually splits) the number of pixels in the blood vessel and background And the ratio of the number of pixels in the viewport. It can be seen from Table 1 and Table 2 that the retinal blood vessel extraction technique based on the iterative graph cut proposed by the present invention has improved sensitivity and precision compared with the previous method.

Accuracy comparison of Table 1 segmentation algorithm

STARE sensitivity specific accuracy

Staal et al. 0.6898±0.1558 0.9793 ±0.0133 0.9516

Mendon9a et al. 0.7123 0.9758 0.9479±0.0123

Wang et al. 0.7543 ±0.0596 0.9785±0.0106 ―

Chen et al 0.7737 ±0.0735 0.9738 ±0.0169 0.9490±0.0109 manual division B 0.8951±0.1085 0.9385 ±0.0260 0.9503 ±0.0089 The present invention 0.7208 ±0.0695 0.9759 ±0.0076 0.6898 ±0.1558 Accuracy comparison of the table two segmentation algorithm

DRIVE sensitivity specific accuracy

Staal et al. 0.7194± 0.0694 0.9773 ±0.0087 0.9441 ±0.0065 Mendonpa et al 0.7315 0.9781 0.9463 ±0.0065

Wang et al 0.7810±0.0340 0.9770 ±0.0071 ―

Chen et al 0.7589 ±0.0449 0.9778 ±0.0064 0.9462 ±0.0057 manual split B 0.7760 ±0.0594 0.9725 ±0.0083 0.9473 ±0.0048 0.7732 ±0.0345 0.9685 ±0.0064 0.9445 ±0.0049 of the present invention The experiment shows that the method of the present invention is based on iterative map-cut retinal blood vessel extraction Technique 1 accurately extracts blood vessels from the retinal image.

The above is only a specific embodiment of the present invention. The scope of the present invention is not limited thereto, and any person skilled in the art who is familiar with the technology may make changes or substitutions within the technical scope of the present invention. However, such changes and substitutions are intended to be encompassed within the scope of the invention. Therefore, the scope of the invention should be determined by the scope of the claims.

Claims

Rights request

A method for extracting retinal blood vessels, comprising the steps of:

Multi-scale filtering of retinal images;

Obtaining the centerline of the retinal blood vessels, using markers to initiate segmentation of the blood vessels;

A map is created based on the filtered image and the initially segmented image, and the map is calculated to obtain blood vessels in the retinal image.

2. Method according to claim 1, characterized in that the filtered image of the retinal grayscale image / (x, y) is calculated as follows: V x, y) = max v (x, y, s) where s The scale is represented; ^ _in ^ _max represents the maximum and minimum scales of the scale space, respectively, and the above formula has the scale selection characteristic: when the radius of the blood vessel matches the scale s, the filtered value is the largest;

( , c (x, JK, + ( , y, s)^ , ( , y, s) ≤ λ ₂ ( , y, s) <0;

, † 0, other. ,

, , (χ, , are two eigenvalues of the multiscale Hessian matrix of the image; _c is a constant.

3. Method according to claim 1, characterized in that said markers are four.

4. Method according to claim 1, characterized in that the center line of the retinal blood vessels is obtained using a threshold method and a refinement method.

The method according to claim 3, characterized in that the four types of markers / _p are: foreground seed Λ, candidate foreground seed; candidate background seed ^ and background seed.

6. The method of claim 5, comprising:

The pixels on the center line are marked as . In addition to the pixel is marked as foreground, V {x, y)> t pixels _hl is _{labeled; t hl> V (x,} y) ≥ t, the pixel is marked as _{b c; V x, y)} < The pixels of t, are marked as .

The method according to claim 1, characterized in that said map comprises - a node V with a pixel as a graph, and an adjacent relationship between pixels as an edge of the graph.

8. The method according to claim 7, characterized in that the K-means clustering algorithm is used to separately filter the filtering results (labeled as candidate foreground seeds and candidate background seeds, and separately calculate candidate foreground seeds and candidate backgrounds). The shortest distance from the seed to the foreground seed ^(x^) and the shortest distance of the background seed (X, .

9. The method of claim 8 further comprising:

The node pel of the connection graph and the edges of the two virtual nodes are defined as t- //^ _e f ; the edge of the node corresponding to two adjacent pixels in the connected image is defined as .

10. The method according to claim 9, wherein the graph is calculated according to the following rules: if it belongs to a source type and the marker is not a foreground seed point, it is marked as a candidate foreground seed point, otherwise it is marked as a foreground seed point. If it belongs to the sink type and the marker is not a background seed point, it is marked as a candidate background seed point ^, otherwise it is marked as a background seed point.

11. The method of claim 10, further comprising:

Calculate the local energy and replace the mark according to the local energy model;

Determine whether the number of iterations is reached, and if not, return; If so, the candidate foreground seed point is changed to the foreground seed point, and the candidate background seed point is changed to the background seed point.

12. Method according to claim 10, characterized in that said calculation method is a maximum flow, minimum cut algorithm.

13. The method according to claim 9, wherein t- /^ is calculated as follows _:

Us {P) =∞' U _T (p) = 0, Ρ≡ f '

U _T (p) = ∞,

U _S {P)- pb _c ;

Where D _F {p)= ^(P factory; , {ρ) = --ψ^, ^ and are constants, and full df{p) + d _b {p) d _f (p) + d _b (p) Foot 0 ≤ <1

14. The method of claim 9 characterized by a following formula

among them, ? ) represent multi-scale filtered values for nodes and ^ corresponding pixels, respectively.

15. The method of claim 11 wherein said local energy is expressed by:

Where / represents the mark of the current pixel; if / _Ρ =.Λ, then " _ρ (/ _ρ ) = Α ( _; if l _p = b _c , then

" _ρ „ ', Κ is a constant; =∑^ _P , ; N is the neighborhood; the values are as follows:

Where ζ· is a constant and takes a value of 0 _{≤ ≤} 1.

16. The method of claim 1, wherein the method of replacing the marking according to the local energy model is:

The foreground seed point and the background seed point are not replaced with the mark;

For the candidate foreground seed point f _c , the two energy values and the local energy of the pixel are divided into three segments. If the energy is less than a given threshold; ^ ₂ , the flag is changed to the foreground seed point Λ if the energy is greater than the other given The threshold; 7^, is marked as the candidate background seed point δ _£ , otherwise the flag is unchanged.

17. The method of claim 15 wherein the method of replacing the marking according to the local energy model is:

BACKGROUND candidate for the seed point b _c, and two energy values, _two local energy pixel is divided into three sections, if the energy is less than a given threshold value ¾ _2, the seed point is marked to the background, if the energy is larger than the other to The threshold is determined; ^, it is marked as a candidate foreground seed point, otherwise the flag is unchanged.