An automatic carotid artery localization and segmentation method based on adaptive feature optimization

Abstract

The present invention discloses a method for automatic positioning and segmentation of the carotid artery based on adaptive feature optimization. First, a standard matching template is designed based on the grayscale characteristics of the shape of the lumen in the ultrasonic carotid artery cross-section image, and a multi-scale template group is generated based on possible scale adjustments. Second, the multi-scale template group is matched with the image to obtain the possible position of the carotid artery. The most likely area is further selected by combining the adaptive circular grayscale information and relative position information, and the initial contour of the carotid artery lumen is generated accordingly. Finally, the initial contour is evolved by the DRLSE algorithm that optimizes the external energy term to achieve contour segmentation with high accuracy and high robustness. This method achieves robust position detection by designing a reasonable multi-scale matching template group, and improves segmentation accuracy and robustness by optimizing the energy term that drives the evolution. The present invention can accurately and effectively achieve automatic positioning and segmentation of the carotid artery, providing guidance for the processing and analysis of medical images.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and relates to a method for identifying the lumen of an ultrasonic carotid artery cross section, and in particular to an adaptive carotid artery automatic positioning and segmentation algorithm based on improved grayscale features of the carotid artery lumen shape. Background Art

[0002] The carotid artery, one of the body's major blood vessels, is closely linked to the development of cardiovascular and cerebrovascular diseases. Vascular ultrasound can detect early signs of disease, which is crucial for the prevention and treatment of cardiovascular and cerebrovascular diseases. The structure of the carotid artery lumen directly affects blood supply and blood pressure, making it a key indicator for assessing cardiovascular and cerebrovascular health. Therefore, accurate segmentation of the carotid artery lumen helps identify and analyze the likelihood and causes of disease, significantly improving the accuracy and efficiency of disease diagnosis.

[0003] Traditional carotid artery diagnosis relies on ultrasound scanning and subjective analysis by doctors, which is time-consuming and easily influenced by personal experience and skills. With the development of computer-aided diagnosis technology, automated segmentation methods have emerged, including those based on thresholds, regions, and models. However, these methods often require the user to provide prior information, such as seed points. To this end, some automatic auxiliary diagnosis systems propose the idea of first finding the target area and then further implementing segmentation. Template matching is a commonly used target detection algorithm, but its performance is highly dependent on the design and selection of the template. If the template design is unreasonable or the scale is inappropriate, it may lead to poor matching results.

[0004] Image segmentation technology has made significant progress in recent years. The distance regularized level set evolution method (DRLSE) has performed well in target contour extraction. This method introduces a distance regularization term to correct the deviation between the level set function and the signed distance function, avoiding the need for periodic reinitialization. However, in the field of medical image processing, the DRLSE method still faces challenges, such as complex backgrounds and multiple types of noise. In addition, existing DRLSE methods are still sensitive to the choice of initial contours, and their performance needs to be improved when processing images of target objects with weak boundaries. Summary of the Invention

[0005] To overcome the limitations of existing carotid artery segmentation systems, the present invention provides an automatic carotid artery location and segmentation method based on adaptive feature optimization. This method achieves robust location detection by designing a rational multi-scale matching template set and improves segmentation accuracy and robustness by optimizing the energy term that drives the evolution. This method automatically adjusts key parameters during the detection process using image features. It can handle carotid artery images of varying sizes and shapes, providing highly accurate segmentation results. This method effectively overcomes the limitations of existing technologies and offers a novel approach for computer-assisted ultrasound diagnosis.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization includes the following steps:

[0008] Step 1: By analyzing the grayscale features of the lumen in the ultrasonic carotid artery cross-section image, a circular ring is designed as the standard matching template for the carotid artery. The possible size of the carotid artery lumen is calculated by the image features of the statistical data set, and the carotid artery standard template is scaled to obtain a multi-scale template group.

[0009] Step 2: For the carotid artery ultrasound image to be detected, the templates of each scale in the multi-scale template group are traversed. For each scale template, the area with the highest similarity is recorded as the best matching area at that scale. A simulated circle is generated and its size is adaptively adjusted. The pixel grayscale difference and relative position information of the simulated circle are further analyzed to select the most likely lumen area and generate a simulated carotid artery lumen contour.

[0010] Step 3. The external energy term of the DRLSE algorithm is optimized in combination with the shape and grayscale features of the carotid artery lumen. Two new energy terms, the shape constraint term and the shape region energy term, are added to constrain the shape and grayscale distribution. An improved level set function evolution equation is established. The level set function evolution equation drives the contour evolution, and a two-step evolution stage is designed to quickly approach the target contour first and then perform fine segmentation to better meet the different characteristics of the contour during the evolution process, and finally achieve contour segmentation with high accuracy and high robustness.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] (1) Traditional template matching methods are highly dependent on the template itself, and are less effective when the target object is deformed or changes size, or when the image conditions are not ideal. The present invention provides a carotid artery segmentation method based on the grayscale feature design of the carotid artery shape, an adaptive multi-scale template matching positioning method, and an optimized level set algorithm. This method independently designs a template by deeply analyzing the grayscale features of the carotid artery shape, and combines multi-scale transformation with an adaptive adjustment mechanism, so that the algorithm has good robustness and enhances its adaptability in changing scenarios. In addition, by further analyzing the regional grayscale features and relative position information, the complete dependence on the template is effectively reduced, thereby maintaining a stable segmentation effect in a variety of situations.

[0013] (2) Compared with the traditional DRLSE algorithm, the present invention adopts an initial contour that is close to the position and shape of the carotid artery as the starting point, which reduces the number of iterations required for the evolution process, thereby reducing the consumption of computing resources and time. The external energy function is optimized in combination with the unique shape grayscale features of the carotid artery lumen ultrasound image, and two new energy terms are introduced to consider and restrict the shape and regional grayscale features respectively, so that the accuracy and reliability of the lumen contour segmentation can be maintained even under conditions of poor image quality such as contour defects and blurred edges. The analysis and introduction of the unique characteristics of the carotid artery make this algorithm show stronger applicability and anti-interference ability in the carotid artery segmentation task than the general algorithm.

[0014] (3) The present invention adopts a two-step structure of first positioning and then segmentation to achieve automatic segmentation of the carotid artery, reducing dependence on physician guidance and thus improving the system's operating efficiency. Based on the first step of positioning, the second step of segmentation uses the positioning results as an initial reference, enhancing the overall scalability and optimizability of the system. This structured approach provides an efficient and automated solution for carotid artery segmentation and has important clinical application value.

[0015] (4) The method proposed in the present invention can accurately and effectively realize the automatic positioning and segmentation of the carotid artery, providing guidance for the processing and analysis of medical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flowchart of the method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization;

[0017] Figure 2 This is the carotid artery standard matching template image;

[0018] Figure 3 Examples of simulated circles obtained at different scales;

[0019] Figure 4 This is an example of the initial outline of the carotid artery lumen;

[0020] Figure 5 An example of the final contour obtained for carotid artery lumen segmentation. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0022] The present invention discloses a method for automatic positioning and segmentation of the carotid artery based on adaptive feature optimization. First, a standard matching template is designed based on the grayscale characteristics of the lumen shape in the ultrasonic carotid artery cross-section image, and a multi-scale template group is generated based on possible scale adjustments. Second, the multi-scale template group is matched with the image to obtain the possible position of the carotid artery. The most likely area is further selected by combining the adaptive circular grayscale information and relative position information, and the initial contour of the carotid artery lumen is generated accordingly. Finally, the initial contour is evolved by the DRLSE algorithm that optimizes the external energy term to achieve contour segmentation with high accuracy and high robustness. Figure 1 As shown, it is divided into three steps, the specific steps are as follows:

[0023] Step 1: By analyzing the grayscale features of the lumen in the ultrasound carotid artery cross-section image, a circular ring of a certain scale is designed as a standard matching template. The possible size of the carotid artery lumen is calculated by the image features of the statistical data set, and the standard template is scaled to obtain a multi-scale template group. The specific steps are as follows:

[0024] Step 1-1: Design the carotid artery standard matching template:

[0025] Step 1-1-1: Based on the characteristics of the carotid artery lumen being approximately circular, with low grayscale values inside, high grayscale values in the vascular membrane, and low grayscale values in the surrounding tissue, a circular ring is designed as a template, where the inner circle simulates the lumen, the annular area simulates the vascular membrane, and the outer area of the ring simulates the surrounding tissue.

[0026] Step 1-1-2: Use F carotid artery ultrasound images as template references, and take the ratio of the distance from the center point of the lumen to the intima and adventitia respectively as the approximate ratio of the inner and outer circle radii of the standard matching template.

[0027] Step 1-1-3: For the fth template reference image, manually select the region of interest (ROI) that just contains the carotid adventitia area, and use an approximate ring to simulate the carotid artery lumen. The outer circle of the simulated ring is denoted as C. ot,f , where ot represents the outer circle, then the outer circle diameter D ot,f Defined as:

[0028]

[0029] Among them, W x,f is the pixel size in the x direction of the region, W y,f is the pixel size of the region in the y direction.

[0030] Step 1-1-4: Take the lumen center point of the f-th image as the seed point, perform binarization on the image by region growing method, extract the edge of the carotid artery lumen, and take the maximum and minimum values of the edge in the y direction as y max,f ,y min,f , the maximum and minimum values in the x direction are x max,f ,x min,f Based on this, the vertical diameter and horizontal diameter of the lumen are defined as the difference between the maximum and minimum longitudinal and transverse values of the edge, respectively. The inner circle of the simulated ring of the f-th image is recorded as C in,f , where in represents the inner circle and the inner circle diameter D in,f It is defined as the average of the vertical diameter and the horizontal diameter and is calculated as:

[0031]

[0032] Step 1-1-5: Count the ratios of the inner and outer diameters of the simulated rings of all F ROI images, and take the average value β as the ratio of the inner and outer diameters of the simulated ring of the standard matching template. The calculation is as follows:

[0033]

[0034] Step 1-1-6: The standard matching template of the carotid artery lumen is constructed based on this ratio to fully conform to the grayscale distribution characteristics of the carotid artery lumen and maintain robustness to possible interference in the ultrasound image. A square with a side length of ts is taken as the initial template area, and the center coordinates are (x c ,y c ), where c represents the center of the template and the outer circle of the template ring C to The radius is r to =ts / 2, where to represents the outer circle of the template ring; the inner circle C ti The radius is r ti =β·r to , where ti represents the inner circle of the template ring. Based on this design, a ring is simulated, with low grayscale values inside the inner circle, high grayscale values in the annular area, and low grayscale values outside the outer circle. This results in a standard matching template T. The grayscale value Gr(x,y) of each pixel (x,y) is specifically designed as follows:

[0035]

[0036] Step 1-2: Construct a multi-scale matching template group:

[0037] Step 1-2-1: Count the maximum and minimum values of the number of pixels occupied by the carotid artery lumen diameter in the existing carotid artery image, and use them as the possible maximum value of the template size S max , minimum value S min , set the scale factor at equal intervals.

[0038] Step 1-2-2: When there are SN scales in the multi-scale ratio group, the snth template scale coefficient si is:

[0039]

[0040] In this way, we get the multi-scale ratio group S SN =[si1,si2,…,si SN ], and the standard matching template T is grouped according to the multi-scale ratio S SN Each scale factor in is scaled SN times to obtain a new multi-scale template group T SN , so that the matching algorithm can maintain high matching accuracy for lumens of different sizes.

[0041] Step 2: For the carotid artery ultrasound image to be detected, the templates at each scale in the multi-scale template group are traversed. For each scale template, the area with the highest similarity is recorded as the best matching area at that scale. A simulated circle is generated and its size is adaptively adjusted. The pixel grayscale difference information and relative position information of the simulated circle are further analyzed to select the most likely lumen area and generate a simulated carotid artery lumen contour. The specific steps are as follows:

[0042] Step 2-1, multi-scale template matching:

[0043] Step 2-1-1: For the carotid artery ultrasound image I to be detected, traverse all possible positions (pi, pj) in the image and calculate the position of the multi-scale template group T. SN Templates of various scales T sn The degree of matching is calculated using the normalized cross-correlation (NCC) as the evaluation score RS(pi,pj) of the position (pi,pj):

[0044]

[0045] Among them, T sn For si sn The template after scaling under the scale factor has a size of M×M. is the average value of the scale template pixels, defined as I(x+pi,y+pj) is the region of the ultrasound image where the similarity is to be calculated; is the average value of the matching area in the image, defined as

[0046] Step 2-1-2, scale coefficient si sn The template is selected, and the area that is most similar to the template is the area with the highest score RG sn , record the corresponding region center coordinates as (x sn ,y sn ), where x sn =pi+ts·si sn / 2,y sn =pj+ts·si sn / 2.

[0047] Step 2-2: Adaptively generate a simulated circle:

[0048] Step 2-2-1, for the obtained si sn The best matching area under the scale coefficient is approximated by a circle to represent the most likely existence form of the carotid artery lumen in the area, and the radius of the approximate circle is r. sn ts·si sn / 2. The radius is adjusted autonomously to obtain the radius r of the simulated circle after adjustment. adj,sn , where adj indicates adjustment to ensure that the circular area does not exceed the image boundary, as follows:

[0049] r adj,sn =min(r sn ,x sn ,y sn ,ts·si sn -y sn ,ts·si sn -x sn )

[0050] Step 2-2-2, take the center coordinate (x sn ,y sn ) is the center of the circle, and the radius of the simulated circle after adjustment is r adj,sn Draw a circle with radius C sn , simulating the carotid artery, the area represented by the circle is

[0051] Step 2-3, calculate the relative position coefficient:

[0052] Step 2-3-1: For templates of different scales T sn The obtained simulated circle C sn There is overlap between them, which means that this area is more likely to be similar to the lumen, so the overlap between each simulated circle area and other areas is judged. sn1 and scale factor si sn2The two circles C obtained sn1 and C sn2 , compare its center (x sn1 ,y sn1 ) and (x sn2 ,y sn2 ) distance and the radius of the two circles and r sn1 +r sn2 When the distance between the centers is less than the sum of the radii, the two circles overlap. sn1 ,C sn2 Overlap coefficient γ sn1,sn2 Calculated as:

[0053]

[0054] Step 2-3-2, set the position weight coefficient to pnt, then the scale coefficient si sn The relative coefficient w of the position of the most likely area obtained by the corresponding template sn The calculation is as follows:

[0055]

[0056] Among them, γ ks,sn For two circles C ks ,C sn The overlap coefficient between .

[0057] Step 2-4: Calculate the grayscale value difference between the inner and outer sides of the simulated ring:

[0058] Step 2-4-1: Further evaluate the grayscale characteristics of the area in the form of a simulated ring. sn The outer radius of the optimal simulation ring under the scale coefficient template is r calculated in step 2-2. adj,sn , the inner ring radius is β times the outer ring radius, and the area inside the simulated ring is denoted as Q ir,sn , where ir represents the inner circle; the area between the rings is Q om,sn , where om represents the ring, calculate the average gray value difference between the two, and compare it with the position relative coefficient w obtained in step 2-3 sn The weighted grayscale difference score of the shape area is used to calculate the final score diff of each area sn , the specific calculation is:

[0059]

[0060] Where count represents the number of pixels in the area, and Gr(x,y) is the grayscale value of the pixel (x,y).

[0061] Step 2-4-2: Take the maximum value of the grayscale difference score of the shape area obtained at all scales diff maxThe corresponding area is the most likely lumen area and serves as the target area for segmentation in the next stage.

[0062] Step 2-4-3, take the center point of the most likely lumen area (x b ,y b ) is the center of the circle, where b represents the most likely presence of the carotid artery in the image, and the corresponding adjusted inner radius of the simulated ring r b Draw a circle with radius C b , then C b To simulate the carotid artery lumen contour.

[0063] Step 3: Optimize the external energy term of the DRLSE algorithm based on the shape and grayscale features of the carotid artery lumen, add two new energy terms, the shape constraint term and the shape region energy term, to constrain the shape and grayscale distribution, and establish an improved level set function evolution equation. The level set function evolution equation drives the contour evolution, and designs a two-step evolution stage of first quickly approaching the target contour and then fine segmentation to better meet the different characteristics of the contour during the evolution process, and finally achieve contour segmentation with high accuracy and high robustness. The specific steps are as follows:

[0064] Step 3-1, establish the level set function algorithm equation:

[0065] Step 3-1-1, the simulated carotid artery lumen contour C obtained in step 2 b As the initial contour, and as the initial value of the level set function φ that is evolved in this step. The level set function is defined as:

[0066] φ(x,y,t)=0

[0067] Where x and y represent the coordinates of a point in two-dimensional space, and t represents the number of iterations.

[0068] Step 3-1-2: Drive the evolution of the level set function through the energy function E(φ), which is specifically defined as:

[0069] E(φ)=μR(φ)+E ext (φ)

[0070] Among them, μ is the regularization coefficient, R(φ) is the level set function regularization term, which is specifically defined as:

[0071]

[0072] Where Ω is the neighborhood, and p is the double-well potential function, which is determined by the variable parameter s and regularizes the area near the zero level set. It is specifically defined as follows:

[0073]

[0074] E ext(φ) is the external energy term, ext represents the external driving energy part, which drives the level set function to evolve towards the position with minimum energy. It is specifically defined as:

[0075] E ext (φ)=λ Lt L g (φ)+λ Ar A g (φ)+λ sc ξ sc (φ)+λ sa ξ sa (φ)

[0076] Among them, λ Lt Minimize the energy coefficient for a fixed weighted length, Lt represents the length term; L g (φ) is the weighted length minimization energy term, which calculates the weighted line integral along the zero level set of the level set function. When the zero level set is at the edge of the image contour, the energy of the length term reaches the minimum value to ensure the smoothness of the contour curve. It is specifically defined as:

[0077]

[0078] Among them, g is the boundary detection function, which is specifically defined as:

[0079]

[0080] Among them, G σ is a Gaussian kernel function with a standard deviation of σ, and I is the ultrasound image to be detected. The noise can be removed to a certain extent by convolution of the two.

[0081] λ Ar is a fixed weighted area energy coefficient, Ar represents the area term coefficient, A g (φ) is the weighted area energy term, and the integral region is the region where the level set function is less than zero. It can accelerate the curve evolution process and ensure that when the initial boundary is far away from the image contour boundary, the function can quickly approach the ideal boundary. It is defined as follows:

[0082] A g (φ)=∫ Ω gH(-φ)dx

[0083] Where H is the Heaviside function.

[0084] λ sc is the shape constraint coefficient, a variable parameter, and sc represents the shape constraint. sc (φ) is the shape constraint term, which compares the contour φ during the calculation evolution process with the given prior shape φ pcThe difference is weighted by the boundary detection function g, and the segmentation result is limited to be as close to a circle as possible to weaken the adverse effects of over-segmentation or under-segmentation, so that the segmented lumen shape is consistent with the actual anatomical structure and the smoothness and accuracy of the segmentation contour are maintained, where pc represents the prior shape. The shape constraint term is specifically defined as follows:

[0085] ξ sc (φ)=∫ Ω (gH(-φ)-gH(-φ pc ))dx

[0086] λ sa is the shape area energy coefficient, a variable parameter, and sa represents the shape area term. sa (φ) is the shape area energy term, which is calculated by calculating the shape φ of each evolution and its external expansion shape φ et The difference between the lumen and the membrane is simulated, which further highlights the grayscale distribution characteristics of the carotid artery lumen, where et represents the external extended shape. The shape region energy term is defined as follows:

[0087] ξ sa (φ)=∫ Ω (gH(-φ)-gH(-φ et ))dx

[0088] Among them, φ et is the extended shape corresponding to the carotid artery contour to simulate the peripheral contour of the vascular membrane, which is adaptively generated by the shape φ of each evolution: et =φ⊕K. Where K is the side length of (1-β) / 2·r φ The square structural element, r φ is the average inner diameter of the evolved shape.

[0089] Step 3-2, contour evolution:

[0090] Step 3-2-1: In numerical calculation, use numerical approximation of Dirac function δ ε (x) and the numerical approximation of the Heaviside function H ε (x) respectively approximates the Dirac function δ and the Heaviside function H, where ε is a small fixed parameter, specifically:

[0091]

[0092] After numerical approximate calculation, the energy functional E is obtained ε (φ) is approximately:

[0093]

[0094] The energy functional E is calculated and solved by Euler Lagrange equation and gradient descent method. ε The above energy functional can be minimized by solving the following gradient flow, where d p =p'(s) / s:

[0095]

[0096] Step 3-2-2: Use the inverse finite difference method for evolution iteration. The evolution here is divided into two stages. The first stage quickly approaches the target contour area and iterates m times; the second stage seeks a fine segmentation boundary and iterates n times. The specific evolution process is as follows:

[0097]

[0098] Among them, kt is the number of iteration steps, (i, j) is the coordinate of a point on the contour, and are approximate estimates of the evolution equation in the first and second stages, respectively, and are specifically defined as:

[0099]

[0100] The first stage crosses possible noise and patches at a faster speed to reduce the possibility of falling into local optimality. It is necessary to emphasize shape constraints, λ sc The coefficient takes a larger λ sc1 , where sc1 represents the sc coefficient in the first stage of evolution; there is no need to consider the influence of shape grayscale features too much, λ sa The coefficient takes a smaller λ sa1 , where sa1 represents the sa coefficient of the first stage of evolution. The second stage emphasizes the influence of the shape region energy term to ensure that the segmentation results form a clear grayscale difference between the lumen and the adventitia, and a larger λ is taken sa2 , where sa2 represents the sa coefficient of the second stage of evolution; the shape feature is no longer important, so a smaller λ is used. sc2 , where sc2 represents the sc coefficient of the second stage of evolution.

[0101] Step 3-2-3: After two stages of evolution, the lumen contour φ is obtained f As the final segmentation result, f represents the final result.

[0102] Example:

[0103] This example designs a multi-scale matching template set by analyzing the shape grayscale features of the carotid artery image, and then achieves lumen localization based on a method combining template matching and adaptive feature analysis. Finally, the contour evolution of the lumen is achieved by using the improved DRLSE algorithm based on the grayscale characteristics of the carotid artery shape. The specific steps are as follows:

[0104] First, a large number of two-dimensional ultrasound carotid artery cross-sectional images were collected using an ultrasound imaging system, and the carotid artery lumen contour was manually drawn as the gold standard for algorithm evaluation.

[0105] Execute step 1: Design a multi-scale template set for carotid artery matching. Select F = 100 images of good quality containing a complete and clearly defined carotid artery lumen as the original dataset and manually crop them to obtain the carotid artery ROI. Then calculate the ratio of the inner and outer diameters of the simulated ring β = 0.86. Based on this ratio, establish a standard matching template simulated ring T with a side length of 100, an inner radius of 43, and an outer radius of 50, as shown in the following example: Figure 2 As shown. The size of the lumen in the statistical data set is calculated and the proportional coefficient is S SN =[0.4,0.5,0.6,0.7], and then obtain the multi-scale template group T SN .

[0106] Execute step 2: realize the automatic positioning detection of the carotid artery lumen. Perform the same operation on all 3263 images in the dataset: compare the image to be detected I with the multi-scale template group T SN Perform similarity calculations to obtain the most similar regions at four scales, and make adaptively adjusted simulated circles, as shown in the following example: Figure 3 As shown. Then, for each most similar area, the simulated ring pixel difference and position coefficient are further calculated. The experimental setting of the coincidence position coefficient is pnt = 1.1. Finally, the best similar area is selected and the center point (x b ,y b ) at the adjusted radius of the simulated ring r b Draw a circle with a radius of , as the initial outline of the carotid artery, as shown in the following figure Figure 4 As shown. When the center position of the positioning result (x b ,y b ) is located inside the gold standard outline, the positioning is considered correct, and the result is that 3137 images are correctly positioned, with a success rate of 96.14%.

[0107] Execute step 3: realize accurate segmentation of carotid artery lumen based on improved DRLSE algorithm. Set the prior shape to be the center of the best similarity region, r b The circle is made of radius. Set ε = 1.5, and the length minimizes the energy coefficient λ Lt =5, area minimization energy coefficient λ Ar = -1.5, the contour extension coefficient is (1-β) / 2 = 0.07 times the average radius of the contour. The number of iterations in the first stage is m = 100 times, and the shape constraint coefficient λ sc1 =0.01, shape area energy coefficient λ sa1= 0.1; the number of iterations in the second stage is n = 10 times, and the shape constraint coefficient λ sc2 =0.002, shape area energy coefficient λ sa2 =1.2. Based on the evolution equation, the contour is iterated according to the evolution steps to obtain the final carotid artery lumen contour, as shown in the following figure. Figure 5 As shown. The Dice coefficient DSC is used as the segmentation accuracy evaluation index, which is defined as:

[0108]

[0109] Among them, Q a is the segmented area obtained by the algorithm proposed in this invention, Q gs is the gold standard segmentation region.

[0110] The average Dice coefficient for carotid artery lumen segmentation across all images was 0.9259. Using the same parameters with the traditional DRLSE algorithm, the Dice coefficient was 0.8712, demonstrating a significant improvement in segmentation performance.

[0111] Through the above embodiments, it can be found that the multi-scale template positioning algorithm that integrates the adaptive features of the carotid artery proposed in the present invention effectively meets the actual needs of accurate positioning of the carotid artery and shows good robustness. Compared with the traditional standard template matching paradigm, it has higher accuracy. Based on the positioning result, the initial contour is generated, and the improved level set segmentation algorithm is further used for evolutionary iteration. In this process, the algorithm fully exploits the inherent characteristics of the carotid artery such as the edge, shape and grayscale, and the obtained segmentation contour shows a very high similarity with the result of manual fine segmentation. This shows that the algorithm has a strong anti-interference ability, can effectively filter out noise and artifact interference, and thus obtain a carotid artery contour close to the real state. Comprehensively considered, the algorithm system described in the present invention can achieve automated and high-precision segmentation of the carotid artery, provide a new solution for clinical diagnosis and treatment decision-making, and has practical potential for clinical promotion and application.

Claims

1. A method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization, characterized in that The method comprises the following steps: Step 1: By analyzing the grayscale features of the lumen in the ultrasonic carotid artery cross-section image, a circular ring is designed as the standard matching template for the carotid artery. The possible size of the carotid artery lumen is calculated by the image features of the statistical data set, and the carotid artery standard template is scaled to obtain a multi-scale template group. Step 2: For the carotid artery ultrasound image to be detected, the templates of each scale in the multi-scale template group are traversed. For each scale template, the area with the highest similarity is recorded as the best matching area at that scale. A simulated circle is generated and its size is adaptively adjusted. The pixel grayscale difference and relative position information of the simulated circle are further analyzed to select the most likely lumen area and generate a simulated carotid artery lumen contour. Step 3. Optimize the external energy term of the DRLSE algorithm based on the shape and grayscale features of the carotid artery lumen. Add two new energy terms, the shape constraint term and the shape region energy term, to constrain the shape and grayscale distribution. Establish an improved level set function evolution equation. The level set function evolution equation drives the contour evolution. A two-step evolution stage is designed, which first quickly approaches the target contour and then performs fine segmentation, to better meet the different characteristics of the contour during the evolution process. Finally, high-accuracy and high-robustness contour segmentation is achieved. The specific steps are as follows: Step 3-1, establish the level set function algorithm equation: Step 3-1-1, the simulated carotid artery lumen contour C obtained in step 2 b As the initial contour, and as the initial value of the level set function φ; Step 3-1-2: Drive the evolution of the level set function through the energy function E(φ), which is specifically defined as: E(φ)=μR(φ)+E ext (f) Among them, μ is the regularization coefficient, R(φ) is the level set function regularization term, E ext (φ) is the external energy term, ext represents the external driving energy part, which drives the level set function to evolve towards the position with minimum energy; Step 3-2, contour evolution: Step 3-2-1: In numerical calculation, use numerical approximation of Dirac function δ ε (x) and the numerical approximation of the Heaviside function H ε (x) respectively approximates the Dirac function δ and the Heaviside function H, where ε is a small fixed parameter, specifically: After numerical approximate calculation, the energy functional E is obtained ε (φ) is approximately: The energy functional E is calculated and solved by Euler Lagrange equation and gradient descent method. ε (φ) is minimized, then the above energy functional is minimized by solving the following gradient flow, where d p =p'(s) / s: Step 3-2-2: Use the inverse finite difference method to perform two-stage evolution iteration. The first stage quickly approaches the target contour area and iterates m times. The second stage seeks a fine segmentation boundary and iterates n times. The specific evolution process is as follows: Among them, kt is the number of iteration steps, (i, j) is the coordinate of a point on the contour, and are approximate estimates of the evolution equation in the first and second stages respectively; Step 3-2-3: After two stages of evolution, the lumen contour φ is obtained f As the final segmentation result, f represents the final result.

2. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 1, characterized in that The specific steps of step 1 are as follows: Step 1-1: Design the carotid artery standard matching template: Step 1-1-1: Based on the characteristics of the carotid artery lumen being approximately circular, with low grayscale values inside, high grayscale values in the vascular membrane, and low grayscale values in the surrounding tissue, a circular ring is used as a template, where the inner circle simulates the lumen, the annular area simulates the vascular membrane, and the outer part of the ring simulates the surrounding tissue; Step 1-1-2: Use F carotid artery ultrasound images as template references, and take the ratio of the distance from the lumen center to the intima and adventitia, respectively, as the approximate ratio of the inner and outer circle radii of the standard matching template; Step 1-1-3: For the fth template reference image, manually select the region of interest that just contains the carotid adventitia area, and use an approximate ring to simulate the carotid artery lumen. The outer circle of the simulated ring is denoted as C. ot，f , where ot represents the outer circle, then the outer circle diameter D ot，f Defined as: Among them, W x，f is the pixel size in the x direction of the region, W y，f is the pixel size of the region in the y direction; Step 1-1-4: Take the lumen center point of the f-th image as the seed point, perform binarization on the image by region growing method, extract the edge of the carotid artery lumen, and take the maximum and minimum values of the edge in the y direction as y max，f ,y min，f , the maximum and minimum values in the x direction are x max，f , x min，f Based on this, the vertical diameter and horizontal diameter of the lumen are defined as the difference between the maximum and minimum longitudinal and transverse values of the edge, respectively. The inner circle of the simulated ring of the f-th image is recorded as C in，f , where in represents the inner circle, the inner circle diameter D in，f It is defined as the average of the vertical diameter and the horizontal diameter and is calculated as: Step 1-1-5: Count the ratios of the inner and outer diameters of the simulated rings of all F ROI images, and take the average value β as the ratio of the inner and outer diameters of the simulated ring of the standard matching template. The calculation is as follows: Step 1-1-6, take a square with a side length of ts as the initial template area, and the center coordinates are (x c ,y c ), where c represents the center of the template and the outer circle of the template ring C to The radius is r to =ts / 2, where to represents the outer circle of the template ring; the inner circle C ti The radius is r ti =β·r to , where ti represents the inner circle of the template ring; based on this design, the ring is simulated, the inner circle has a low grayscale value, the annular area has a high grayscale value, and the outer circle has a low grayscale value, and the standard matching template T is obtained. The grayscale value Gr(x, y) of each pixel (x, y) is specifically designed as follows: Step 1-2: Construct a multi-scale matching template group: Step 1-2-1: Count the maximum and minimum values of the number of pixels occupied by the carotid artery lumen diameter in the existing carotid artery image, and use them as the possible maximum value of the template size S max , minimum value S min , set the proportional coefficient at equal intervals; Step 1-2-2: When there are SN scales in the multi-scale ratio group, the snth template scale coefficient si is: In this way, we get the multi-scale ratio group S SN =[si1,si2,…,si SN ], and the standard matching template T is grouped according to the multi-scale ratio S SN Each scale factor in is scaled SN times to obtain a new multi-scale template group T SN .

3. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 1, characterized in that The specific steps of step 2 are as follows: Step 2-1, multi-scale template matching: Step 2-1-1: For the carotid artery ultrasound image I to be detected, traverse all possible positions (pi, pj) in the image and calculate the position of the multi-scale template group T. SN Templates of various scales T sn The degree of matching is calculated using the normalized correlation coefficient as the evaluation score RS(pi, pj) of the position (pi, pj): Among them, T sn For si sn The template after scaling under the scale factor has a size of M×M. is the average value of the pixels of the scale template; I(x+pi, y+pj) is the area of the ultrasound image where the similarity is to be calculated; is the average value of the matching area in the image; Step 2-1-2, scale coefficient si sn The template is selected, and the area that is most similar to the template is the area with the highest score RG sn , record the corresponding region center coordinates as (x sn ,y sn ), where x sn =pi+ts·si sn / 2,y sn =pj+ts·si sn / 2, ts is the side length of the square; Step 2-2: Adaptively generate a simulated circle: Step 2-2-1, for the obtained si sn The best matching area under the scale coefficient is approximated by a circle to represent the most likely existence form of the carotid artery lumen in the area, and the radius of the approximate circle is r. sn ts·si sn / 2, and the radius is adjusted autonomously to obtain the radius r of the simulated circle after adjustment. adj,sn , where adj indicates the adjustment to ensure that the circular area does not exceed the image boundary, as follows: r adj,sn =min(r sn ,x sn ,and sn ,ts·si sn -and sn ,ts·si sn -x sn ) Step 2-2-2, take the center coordinate (x sn ,y sn ) is the center of the circle, and the radius of the simulated circle after adjustment is r adj,sn Draw a circle with radius C sn , simulating the carotid artery, the area represented by the circle is Step 2-3, calculate the relative position coefficient: Step 2-3-1, for the scale coefficient si sn1 and scale factor si sn2 The two circles C obtained sn1 and C sn2 , compare its center (x sn1 ,y sn1 ) and (x sn2 ,y sn2 ) distance and the radius of the two circles and r sn1 +r sn2 When the distance between the centers is less than the sum of the radii, the two circles overlap, and the two circles C sn1 , C sn2 Overlap coefficient γ sn1，sn2 Calculated as: Step 2-3-2, set the position weight coefficient to pnt, then the scale coefficient si sn The relative coefficient w of the position of the most likely area obtained by the corresponding template sn The calculation is as follows: Among them, γ ks,sn For two circles C ks , C sn The overlap coefficient between Step 2-4: Calculate the grayscale value difference between the inner and outer sides of the simulated ring: Step 2-4-1, si sn The outer radius of the optimal simulation ring under the scale coefficient template is r calculated in step 2-2. adj，sn , the inner ring radius is β times the outer ring radius, and the area inside the simulated ring is denoted as Q ir，sn , where ir represents the inner circle; the area between the rings is Q om，sn , where om represents the ring, calculate the average gray value difference between the two, and compare it with the position relative coefficient w obtained in step 2-3 sn The weighted grayscale difference score of the shape area is used to calculate the final score diff of each area sn , the specific calculation is: Where count represents the number of pixels in the area, and Gr(x, y) is the grayscale value of the pixel (x, y). Step 2-4-2: Take the maximum value of the grayscale difference score of the shape area obtained at all scales diff max The corresponding area is the most likely lumen area and serves as the target area for segmentation in the next stage; Step 2-4-3, take the center point of the most likely lumen area (x b ,y b ) is the center of the circle, where b represents the most likely presence of the carotid artery in the image, and the corresponding adjusted inner radius of the simulated ring r b Draw a circle with radius C b , then C b To simulate the carotid artery lumen contour.

4. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 3, characterized in that described 5. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 1, characterized in that The level set function is defined as: φ(x, y, t) = 0 Where x and y represent the coordinates of a point in two-dimensional space, and t represents the number of iterations.

6. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 1, characterized in that The R(φ) is specifically defined as: Where Ω is the neighborhood, and p is the double-well potential function, which is determined by the variable parameter s and regularizes the area near the zero level set. It is specifically defined as follows:

7. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 1, characterized in that The E ext (φ) is specifically defined as: E ext (φ)=λ Lt L g (φ)+λ Ar A g (φ)+λ sc x sc (φ)+λ sa x sa (f) Among them, λ Lt Minimize the energy coefficient for a fixed weighted length, Lt represents the length term; L g (φ) is the weighted length minimization energy term; λ Ar is a fixed weighted area energy coefficient, Ar represents the area term coefficient, A g (φ) is the weighted area energy term; λ sc is the shape constraint coefficient, sc represents the shape constraint; ξ sc (φ) is the shape constraint; λ sa is the shape area energy coefficient, sa represents the shape area term; ξ sa (φ) is the shape region energy term.

8. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 7, characterized in that The L g (φ) is specifically defined as: Among them, g is the boundary detection function, which is specifically defined as: Among them, G σ is a Gaussian kernel function with a standard deviation of σ, and I is the ultrasound image to be detected; The A g (φ) is defined as follows: A g (φ)=∫ Ω gH(-φ)dx Where H is the Heaviside function; The shape constraint term ξ sc (φ) is defined as follows: x sc (φ)=∫ Ω (gH(-φ)-gH(-φ) pc ))dx Among them, φ pc For a given prior shape, pc represents the prior shape; Shape region energy term ξ sa (φ) is defined as follows: x sa (φ)=∫ Ω (gH(-φ)-gH(-φ) et ))dx Among them, et represents the external extension shape, φ et is the extended shape corresponding to the carotid artery contour, φ et = K is the side length of (1-β) / 2·r φ The square structural element, r φ is the average inner diameter of the evolved shape.

9. The method for automatic positioning and segmentation of carotid arteries based on adaptive feature optimization according to claim 1, characterized in that described and Specifically defined as: Among them, sc1 represents the sc coefficient of the first stage of evolution; sa1 represents the sa coefficient of the first stage of evolution; sa2 represents the sa coefficient of the second stage of evolution; sc2 represents the sc coefficient of the second stage of evolution.

Images