Automatic midium segmentation method in carotid artery ultrasonic image
Technical Field
The invention relates to an automatic intima-media segmentation method in a carotid artery ultrasonic image, and belongs to the field of medical image processing.
Background
Cardiovascular disease is a leading disease that threatens human health. Atherosclerosis causes the thickening of the arterial vessel wall and the narrowing of the vessel cavity, thus being the main cause of cardiovascular diseases. The vessel wall is divided into an outer membranous layer, a middle membranous layer and an inner membranous layer which are mutually clung from inside to outside. The Intima-Media thickness (IMT) is defined as the distance from the luminal-Intima (LI) boundary to the Media-Adventitia (MA) boundary, i.e. the thickness of the Intima and Media portions is added. An increase in IMT is an early clinical symptom of atherosclerosis. Numerous studies have demonstrated that there is a significant correlation between carotid IMT and cardiovascular and cerebrovascular disease and that it can be a strong predictor of future cardiovascular events. Today, intima-media thickness in the carotid artery has been considered as an important indicator for preliminary discrimination between the degree of carotid atherosclerosis and the condition of cardiovascular pathologies.
The medical ultrasonic imaging technology can carry out clearer imaging on the carotid artery, and is an effective carotid atherosclerosis screening means. By means of the ultrasonic imaging technology, the carotid intima-media membrane can be imaged, and then the boundary of the intima-media membrane can be segmented, so that the measurement of IMT is realized. However, the method of acquiring the intima-media boundary by the manual segmentation by the doctor in clinical use is time-consuming and tedious, and there is a large difference between observers, so that there is an urgent need for an image segmentation method capable of automatically segmenting the intima-media boundary.
At present, a plurality of carotid intima-media segmentation algorithms exist, and more accurate intima-media segmentation can be realized. However, ultrasound images suffer from inherent speckle noise, low signal-to-noise ratio, and insufficient anatomical detail. Serious speckle noise can even cover the inner and middle membrane layers, and the success rate of the inner and middle membrane automatic segmentation method is influenced. Aiming at the problem, the invention provides an inner and middle membrane automatic segmentation method with better robustness based on algorithms such as ant colony optimization and the like.
An Ant Colony Optimization (ACO) is a bionic intelligent optimization algorithm developed based on the inspiration of ant colony crawling characteristics, and is characterized in that the overall optimal solution of the problem is determined jointly through the local optimality of bionic adaptive individuals. The algorithm has a self-learning function and a strong searching capability, has the advantages of parallelization, strong robustness and positive feedback, and is introduced into the field of image segmentation. For the intima-media segmentation, an ant colony algorithm can be adopted to convert the contour extraction problem of the MA and LI boundaries into an optimization problem for solving.
Disclosure of Invention
In order to replace the manual segmentation step, the invention provides an automatic segmentation method for a tunica media in an ultrasonic image. The invention provides a two-leg ant colony optimization algorithm according to a unique double-layer parallel interface structure of an intima-media structure, and realizes automatic segmentation of the intima-media by combining a multi-scale Gaussian kernel multiplication method, an edge detection operator and a Snake model.
The technical scheme of the invention is as follows:
an automatic division method for tunica media in an ultrasonic image comprises the following steps:
1) inhibiting background information except for the intima-media by using a multi-scale Gaussian nuclear phase multiplication method, and enhancing the intima-media boundary to obtain an edge map with a prominent intima-media boundary;
2) acquiring partial contour line segments of the LI and MA boundaries by using an initial contour detection method based on an Otsu threshold method and a Sobel operator according to the edge map obtained in the step 2);
3) designing a two-leg ACO algorithm on the basis of the initial contour obtained in the step 2), and connecting the initial contour line segments to obtain a complete contour;
4) further optimizing the contour obtained in the step 3) by using a contour optimization algorithm based on a Snake model, so that the final contour is smoother, continuous and close to a real edge.
And 1) suppressing background information except for the inner tunica media by using a multi-scale Gaussian nuclear multiplication method, highlighting the LI and MA boundaries of the inner tunica media, and obtaining an edge map. The edge map is defined as the product of the convolution results of the image with two filters having gaussian density kernels of different scales, and only the part with positive values is retained:
wherein the content of the first and second substances,
is a two-dimensional Gaussian function of small scale, sigma
1The value is 1-3. And G
σ2(y) is a one-dimensional Gaussian function of large scale, σ
2The value is 10-20.
Step 2) an initial contour detection method based on an Otsu threshold value method and a Sobel operator is used, and the method mainly comprises the following steps:
(1) based on the edge map obtained in the step 1), obtaining a binary image by using an Otsu threshold method;
(2) extracting the upper edge and the lower edge of the MA interface and the upper edge and the lower edge of the LI interface from the binary image by using a horizontal Sobel operator;
(3) deleting redundant edge lines and defective edge lines;
(4) and for the reserved part, taking the middle point of the upper edge line and the lower edge line as the final single edge line after each pair of edge lines are combined.
And 3) adopting a two-leg ACO algorithm to connect the initial contour line segments into a complete contour. It is characterized in that two ants are placed at the starting point of the intima-media membrane and climb from the left end to the right end of the blood vessel. One ant is fixed above the other ant, and then the two ants climb simultaneously, so that the ant can be regarded as two legs of one ant. In addition, the contour line segment obtained by the initial contour detection is defined as a forced path of the ant, when the ant crawls to the part, the ant is forced to crawl along the initial contour line segment, and gaps between the line segments are connected by the ACO algorithm. The method is characterized by comprising the following steps:
(1) equation of transition probability
Ants from manually selected Ps(xs,ys) The dots start, crawl from left to right, moving 1 abscissa to the right each time they are fixed. At time t, the process of moving the l (l ═ 1,2) th leg of the kth ant from pixel (x, i) to the neighboring pixel (x +1, j) is according to the lower transition probability equation:
wherein tau refers to pheromone, η refers to density of the pixel, α and β respectively determine relative influence of the pheromone and the heuristic information, wherein α takes the value of 1-3, and β takes the value of 3-5.
The set of positions where the ith leg representing the kth ant is allowed in the next crawl corresponds to the right neighborhood of pixel (x, i). It has to be noted that the neighborhood pixels cannot be shared by both legs at the same time. Thus, if the right neighborhoods of the two legs coincide, the coincident pixels will be driven from
And (4) excluding.
(2) Global update law
After all ants complete crawling, the pheromone alternates according to the following formula:
τ(x,i)(t+1)=ρτ(x,i)(t)+Δτ(x,i)
wherein, tau(x,i)(t) is the number of pheromones at pixel (x, i) at the t-th iteration, τ(x,i)(t +1) is the number of pheromones at pixel (x, i) at the next iteration. Initial value of pheromone tau0The value is 0.1-1. Rho is an attenuation constant and is used for simulating the volatilization of pheromone, and the value is 0.5-1. Delta tau(x,i)Is the number of pheromones released in this iteration, and the calculation formula is as follows:
where Q is a constant and m is the number of ants. Q is 1, and m is 10-50. C (k) is a consumption function of the kth ant in the path searching process, which is defined as follows:
wherein IN(x,y)(0≤IN(xi,yi) ≦ 1) is the gray value of the pixel (x, i) after normalization. D (k) is antkManually selected end point P before the end point distance of the crawl pathe(xe,ye) The distance of (c). a is a penalty coefficient, the weight of the error distance D (k) in the consumption function is set, and the value is 1-2.
The above step 4) further contours using a contour optimization algorithm based on a Snake model, which contains a smoothing energy term (smoothing energy), an edge energy term (boundary energy) and a uniform energy term (uniform energy), and is implemented by minimizing the following energy functional:
wherein y is1(x) And y2(x) Representing the profile of the interface LI and MA, respectively, the parameter μ controls the weight of the smooth energy term and ν controls the weight of the uniform energy term. The uniform energy term links the two mutually independent profiles of LI and MA, keeping them at a uniform distance. Mu is 0.1-0.3, and v is 1-2.
The invention has the following advantages:
according to the invention, firstly, a multi-scale Gaussian kernel multiplication method is designed according to the characteristics of the carotid artery ultrasonic image, the image is preprocessed, background information except intima-media is inhibited, and the intima-media boundary is enhanced. Then, on the basis of the preprocessed image, a two-leg ant colony optimization algorithm is designed, and is combined with an Otsu threshold method, a Sobel edge detection operator, an ant colony optimization algorithm, a Snake model and other algorithms, so that automatic segmentation of the tunica media in the carotid artery ultrasonic image is achieved. The experimental result based on the clinical image shows that the method obtains an accurate segmentation result, and the error of the segmentation result is smaller than the error between observers of manual segmentation; meanwhile, the method achieves a success rate of 98.7% in clinical data testing, has good robustness, and can deal with images seriously polluted by speckle noise.
Drawings
FIG. 1 is a flow chart of the present invention for segmenting the intima-media of the carotid artery.
Fig. 2 is the result of each step in the initial contour detection process.
Fig. 3 is a typical example of the permitted location set of ant upper leg (upper leg) and ant lower leg (lower leg) in the ant colony optimization algorithm.
FIG. 4 is a diagram illustrating an example of the segmentation result according to the present invention.
Detailed Description
The present invention is further illustrated by the following examples to better understand the technical solutions of the present invention. The method comprises the following steps:
1. an edge map f (x, y) with prominent intima-media boundaries is obtained using a multi-scale gaussian kernel multiplication method. The f (x, y) calculation method is to multiply two filters with gaussian density kernels of different scales with the image convolution result and only retain the positive part of the value:
wherein the content of the first and second substances,
is a two-dimensional Gaussian function with small scale, in this embodiment, sigma
1The value is 1.5. And G
σ2(y) is a one-dimensional Gaussian function of large scale, in this embodiment σ
2The value is 15.
2. And acquiring partial initial contour line segments of the LI and MA boundaries by using an initial contour detection method based on an Otsu threshold method and a Sobel operator. As shown in fig. 2, the method comprises the following steps:
1) acquiring a binary image by using an Otsu threshold method based on the edge map (a) obtained in the previous step, as shown in a map (b);
2) extracting the upper edge and the lower edge of the MA interface and the upper edge and the lower edge of the LI interface from the binarization image (b) by using a horizontal Sobel operator; as shown in FIG. (c), the white line represents the upper edge and the gray line represents the lower edge.
3) After step 2), the ideal result should contain 2 pairs (4) of contours, corresponding to the upper and lower edges of the LI and MA interfaces, respectively. However, the edge lines obtained using the Sobel operator often contain redundant false edge lines or gaps. Therefore, more or less than 2 pairs of edge lines will be deleted as seen in the longitudinal direction, resulting in the result shown in (d).
4) Finally, for the remaining part, the midpoint of the upper edge line and the lower edge line is taken as the final single edge line after each pair of edge lines is merged, as shown in (e).
3. And (3) designing a two-leg ACO algorithm on the basis of the initial contour obtained in the step (2), and connecting the initial contour line segments to obtain a complete contour. The method comprises the following steps:
1) the pheromone matrix is initialized.
2) Placing a straight ant at an initial point P of manual selections(xs,ys) Nearby (upper leg placed 2 pixels above the point and lower leg placed 2 pixels below the point).
3) Ant from Ps(xs,ys) Starting to crawl to the right, moving 1 abscissa to the right every time fixedly, wherein the total steps of crawling are equal to xe-xs。
At time t, the process of moving the l (l ═ 1,2) th leg of the kth ant from pixel (x, i) to the neighboring pixel (x +1, j) is according to the lower transition probability equation:
where τ refers to the pheromone and η to the density of the pixel α and β determine the relative influence of the pheromone and heuristic information, respectively, in this example the values α -1 and β -4.
The set of positions where the ith leg representing the kth ant is allowed in the next crawl corresponds to the right neighborhood of pixel (x, i); however, if the right neighborhoods of the two legs coincide, the coincident pixels will be driven from
And (4) excluding. Therefore, the upper legs of the ants are always above the lower legs, and the final contours can be ensured not to be crossed or overlapped. FIG. 3 illustrates a typical neighborhood during crawling of two-leg ants
Ants are now crawling from x to x +1 position on the abscissa, and the light squares in the figure are the permitted position sets of the upper leg (upper leg) and the lower leg (lower leg) of the ants, and are excluded from the permitted position sets of the upper leg and the lower leg because the dark squares exist in the crawling permitted position sets of the upper leg and the lower leg at the same time.
4) And calculating the pheromone of the pixel point on the ant path.
5) And repeating the steps 2-4 until all the ants finish the crawling task.
6) And updating the pheromone matrix and calculating the volatilization amount of the pheromone.
The pheromone is replaced according to the following formula:
τ(x,i)(t+1)=ρτ(x,i)(t)+Δτ(x,i)
wherein, tau(x,i)(t) is the number of pheromones at pixel (x, i) at the t-th iteration, τ(x,i)(t +1) is the number of pheromones at pixel (x, i) at the next iteration. In this embodiment, the initial value of the pheromone is τ00.5. ρ is an attenuation constant used to simulate the occurrence of pheromones, and ρ is 0.6 in this embodiment. Delta tau(x,i)Is the number of pheromones released in this iteration, and the calculation formula is as follows:
where Q is a constant and m is the number of ants. In the embodiment, Q is 1, and m is 20. C (k) is a consumption function of the kth ant in the path searching process, which is defined as follows:
wherein IN(x,y)(0≤IN(xi,yi) ≦ 1) is the gray value of the pixel (x, i) after normalization. D (k) is antkManually selected end point P before the end point distance of the crawl pathe(xe,ye) The distance of (c). a is a penalty coefficient, and a weight of the error distance d (k) in the consumption function is set, where a is 1 in this embodiment.
7) Repeating steps 2-6 for a number of iterations. The number of iterations is set to 10 in this embodiment.
8) And selecting the ant path with the minimum consumption equation as an optimal path, namely the outline of the interface of the LI and the MA.
4. The contours are further optimized using a contour optimization algorithm based on a Snake model, which is implemented by minimizing the energy functional:
wherein y is1(x) And y2(x) Representing the profile of the interface LI and MA, respectively, the parameter μ controls the weight of the smooth energy term and ν controls the weight of the uniform energy term. In this embodiment, μ ═ 0.1 and ν ═ 1.4 were selected.
Fig. 4 shows the segmentation results of CGACO on three different exemplary graphs. Fig. a is a very noisy image. The inner middle membrane is not clear due to noise pollution. As shown in fig. b, the exact segmentation of the intima-media is not affected by severe speckle noise. Fig. c is an image of a carotid artery with a curve, and fig. e is an image of a carotid artery with a typical significant thickening of the intima-media. The graph d and the graph f correspond to the segmentation results thereof, respectively.