A polar-transunet network-based intravascular ultrasound image boundary extraction method

Abstract

The application relates to a Polar-TransUnet network-based intravascular ultrasound image boundary extraction method, which comprises the following steps: data enhancement; obtaining an IVUS image classification result; and extracting an IVUS image boundary.The application has the beneficial effects that: by customizing data enhancement and building a TransUnet network, the application obtains a pixel-level classification result of an IVUS image, realizes accurate boundary extraction of the IVUS image, has important significance for the diagnosis of clinical coronary atherosclerosis, and also has reference significance for the processing of other annularly distributed medical images.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of medical image engineering, and more particularly to an intravascular ultrasound image boundary extraction method based on a Polar-TransUnet network. BACKGROUND

[0002] The latest China Cardiovascular Health and Disease Report shows that among many diseases, the mortality rate of cardiovascular disease has always been the highest, and coronary heart disease accounts for the second highest proportion, only next to cerebral apoplexy, so the research demand for coronary heart disease diagnosis and treatment is very urgent. Intravascular ultrasound (IVUS) image is the first imaging method to observe the internal structure of blood vessels, which can directly diagnose intravascular lesions and assist PCI operation process, and plays an important role in coronary heart disease diagnosis and treatment. A typical IVUS image contains a "bright-dark-bright" three-layer membrane structure, the lumen boerder depicts the boundary of the blood flow area, and the boundary position is the main parameter for evaluating the degree of stenosis of the blood vessel; the media-adventita border depicts the shape of the blood vessel, and the boundary shape is an important criterion for the health of the blood vessel. Therefore, accurate extraction of the lumen and media-adventita borders of the IVUS image plays an important role in the process of coronary heart disease diagnosis.

[0003] The low resolution characteristics of the ultrasound image and the complexity of the intravascular structure make it difficult to extract the boundary of the IVUS image. In the clinic, most of the IVUS images present a variety of physiological tissues (such as: tube wall, plaque, bifurcation), image performance (such as catheter area, shadow area, artifact), and external inserted guide wire, stent, etc. (hereinafter collectively referred to as "structure"). The sound reflection of some structures is similar, and the sound interface is weak, which further increases the difficulty of boundary extraction.

[0004] Early IVUS image boundary extraction is mostly directly based on boundary or region characteristics, and representative algorithms include Sobel form gradient operator, dynamic programming algorithm, heuristic graph search technology, active contour model, etc. Such algorithms are difficult to handle the complex relationship between structures, and cannot accurately extract the two boundary lines in complex IVUS images. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies in the prior art, and provide an intravascular ultrasound image boundary extraction method based on a Polar-TransUnet network, the technical solution is as follows:

[0006] Step 1, obtaining an IVUS original image I in the test set, performing data enhancement on the IVUS original image I to obtain an IVUS enhanced image I * , so as to obtain the enhanced data set corresponding to the test set;

[0007] Step 2, training the TransUnet network based on the enhanced data set, and then classifying the IVUS original image I in the test set to obtain an IVUS image classification result;

[0008] Step 3, extracting an IVUS image boundary based on the IVUS image classification result of step 2.

[0009] As preferred, the step 1 specifically comprises the following steps:

[0010] Step 1.1, modeling side vessels, vessel bifurcations, guide wire artifacts and shadows, the modeling results including side vessels I side , vessel bifurcations I bf , guide wire artifacts I artifact and shadows I sh , which can be rotated by 360 degrees;

[0011] Step 1.2, obtaining an IVUS original image I, rotating each modeling result in step 1.1 by an arbitrary angle, and then superimposing the modeling results onto the IVUS original image I with a random probability to obtain an IVUS enhanced image I * ; the IVUS enhanced image I * can be represented as:

[0012] I * = I * ε(I sh ) * ε(I bf ) * ε(I side ) * ε(I artifact )

[0013] Wherein, ε is a switching function.

[0014] As preferred, in step 2, the IVUS image classification result includes foreground and background; the foreground is the plaque part between the intima and the media-adventitia in the IVUS image, and the background is the remaining part in the IVUS image.

[0015] As preferred, the step 2 specifically comprises the following steps:

[0016] Step 2.1, building a TransUnet network framework, the TransUnet network framework having a U-shaped structure and a skip connection feature;

[0017] Step 2.2, cutting the IVUS image to obtain an image block x p , and then linearly projecting the image block x p , the formula of linear projection being:

[0018]

[0019] wherein E is a linear mapping matrix, E pos is a position encoding, z0 is a mapped vector;

[0020] Step 2.3, adding Polar-bias encoding on the image block, the encoding formula is:

[0021] E pos = Norm(R) + cosθ

[0022] wherein Norm represents normalization, R represents the Euclidean distance between the image block center point and the image center point, and θ represents the angle between the image block center point and the positive direction of the horizontal coordinate;

[0023] Step 2.4, taking the image block processed in step 2.3 as the input of a Transformer module, wherein the Transformer module comprises a multi-head self-attention unit, a first normalization unit, a multi-layer perception unit and a second normalization unit;

[0024] Step 2.5, connecting l Transformer modules in series to obtain the output of the lth Transformer module, wherein the output of the lth Transformer module is:

[0025] z l = MLP(LN(z l ))+z′ l

[0026] wherein MLP is the multi-layer perception unit of the lth Transformer module, LN is the layer normalization operation, z l ′ represents the output of the (l-1)th Transformer module, and the expression is:

[0027] z l ′ = MSA(LN(z l-1 ))+z l-1

[0028] wherein MSA is the multi-head self-attention mechanism unit of the (l-1)th Transformer module.

[0029] As a preferred, in step 2.4, the multi-head self-attention unit is used to combine multiple self-attentions, and the calculation formula of a single self-attention is:

[0030]

[0031] wherein Q represents the current image block, K represents the correlation between the current image block and the queried image block, and V represents the queried image block, B represents the network threshold value.

[0032] As preferred, step 3 specifically comprises the following steps:

[0033] Step 3.1, explicitly define a curve using snake model, which is expressed as:

[0034] x(s) = (x(s), y(s))

[0035] wherein x(s) on the left side of the equation is a vector representing the curve position, and x(s) and y(s) on the right side of the equation represent the curve projections on the two coordinate axes, respectively;

[0036] Step 3.2, construct energy function The energy function is expressed as:

[0037]

[0038] wherein E snake is the curve energy, E int is the internal energy, and E image is the external energy;

[0039] E int is expressed as:

[0040]

[0041] wherein x' represents the first derivative of curve x, x" represents the second derivative of curve x, and a and β correspond to the weight coefficients of the two derivatives, respectively;

[0042] E image is expressed as:

[0043] E image (v(s)) = ω line E line + ω edge E edge + ω term E term

[0044] wherein ω line E line represents the line energy, ω edge E edge represents the edge energy, and ω term E term represents the corner energy, and E image is composed of the line energy, the edge energy, and the corner energy weightedly;

[0045] Step 3.3, define GVF force field and energy functional ε, wherein the GVF force field is expressed as:

[0046] v(x,y) = [u(x,y),v(x,y)]

[0047] wherein v(x,y) is a GVF force field, u(x,y) and v(x,y) are two coordinate components of the GVF force field;

[0048] The energy functional ε is expressed as:

[0049]

[0050] wherein f represents the boundary information response of the image, μ is a regularization parameter and μ>0, and controls the balance of the two force fields, when is small, the current position (x,y) is in a uniform area, the first term plays a major role, and controls the curve to slowly diffuse around; when is large, the current position (x,y) is in an edge area, only can make the energy functional minimum, so the curve will evolve rapidly towards the target boundary;

[0051] Step 3.4, obtaining the IVUS image classification result in step 3, according to the above steps 3.1 to 3.3, improving the energy function of curve evolution and improving the external force field.

[0052] As preferred, in step 3.4, the improved external force field specifically includes the following steps:

[0053] Step 3.4.1, according to the IVUS image classification result, giving the overall force field trend of the image;

[0054] Step 3.4.2, according to the IVUS image classification result, making the force field of the surrounding pixels point to the boundary position;

[0055] Step 3.4.3, using a boundary filter in the IVUS original image I, fine-tuning the force field at the position where the IVUS image classification result is inaccurate.

[0056] The beneficial effects of the present application are: the present application obtains the pixel-level classification result of the IVUS image by self-defining data enhancement and building the TransUnet network, realizes accurate boundary extraction of the IVUS image, has important significance for the diagnosis of clinical coronary atherosclerosis, and also has reference significance for processing other annularly distributed medical images. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The modeling result schematic diagram provided for the present application;

[0058] Figure 2An IVUS image enhancement schematic diagram provided for the present application;

[0059] Figure 3 A structure schematic diagram of a TransUnet network provided for the present application;

[0060] Figure 4 A flowchart of a Transformer module provided for the present application;

[0061] Figure 5 A Polar-bias position encoding schematic diagram provided for the present application;

[0062] Figure 6 An intima and media-adventitia boundary extraction result schematic diagram of an IVUS image provided for the present application. DETAILED DESCRIPTION

[0063] The present application will be further described below in conjunction with embodiments. The following description of the embodiments is only to help understand the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, the present application can be modified in several ways, and these modifications and improvements also fall within the scope of protection of the claims of the present application.

[0064] Embodiment 1:

[0065] In order to realize accurate boundary extraction of IVUS images, the present application proposes a blood vessel intraluminal ultrasound image boundary extraction method based on a Polar-TransUnet network, which includes the following steps:

[0066] Step 1, obtaining an IVUS original image I in the test set, performing data enhancement on the IVUS original image I to obtain an IVUS enhanced image I * , thereby obtaining an enhanced data set corresponding to the test set;

[0067] Step 2, training a TransUnet network based on the enhanced data set, and then classifying the IVUS original image I in the test set to obtain an IVUS image classification result;

[0068] Step 3, extracting the IVUS image boundary based on the IVUS image classification result of step 2.

[0069] Step 1 specifically includes the following steps:

[0070] Step 1.1, modeling the side blood vessel, blood vessel bifurcation, guide wire artifact and shadow, as shown in Figure 1 , the modeling results in Figure 1 from left to right are side blood vessel I side , blood vessel bifurcation I bf , guide wire artifact I artifact and shadow Ish , the modeling results can be rotated by 360 degrees;

[0071] It should be noted that the deep learning algorithm needs to rely on a large amount of data, and the number of medical images that can be provided by the clinic is far less than that of natural images, therefore, data enhancement needs to be performed on the IVUS image, but the traditional data enhancement method includes rotation, cropping, scaling, folding and the like, which cannot highlight the characteristics of the IVUS image. Since the side blood vessel, vessel bifurcation, guide wire artifact and shadow are all factors hindering boundary extraction, the side blood vessel, vessel bifurcation, guide wire artifact and shadow are modeled in the present application, and the IVUS image can be subjected to targeted data enhancement based on the modeling results.

[0072] Step 1.2, obtaining an IVUS original image I, rotating each modeling result in step 1.1 by an arbitrary angle, and then superimposing it on the IVUS original image I with a random probability to obtain an IVUS enhanced image I * ; the IVUS enhanced image I * can be represented as:

[0073] I * =I*ε(I sh )*ε(I bf )*ε(I side )*ε(I artifact )

[0074] Wherein, ε is a switching function.

[0075] In step 1.2, each enhancement is first rotated by an arbitrary angle, and then superimposed on the original image with a random probability. For example, as shown in Figure 2 , from left to right are respectively the data enhancement results of the randomly selected original image, the vessel bifurcation and the side blood vessel, the shadow area and the guide wire artifact. According to the above method, 10 enhanced images different from the original image are generated for each image, forming a new training set.

[0076] In addition, in step 2, the IVUS image classification result includes foreground and background; the foreground is the plaque part between the intima and the media-adventitia in the IVUS image, and the background is the remaining part in the IVUS image.

[0077] Step 2 specifically includes the following steps:

[0078] Step 2.1, building a TransUnet network framework, as shown in Figure 3 , the TransUnet network framework has a U-shaped structure and a skip connection feature;

[0079] Step 2.2, cutting the IVUS image to obtain an image block x p , the IVUS image is a small size feature map, and then the image block xp Linear projection is performed, and the formula of linear projection is:

[0080]

[0081] wherein, E is a linear mapping matrix, E pos is a position code, and z0 is a mapped vector;

[0082] Step 2.3, Polar-bias coding is added to the image block, and the coding formula is:

[0083] E pos = Norm (R) + cos θ

[0084] wherein, Norm represents normalization, R represents the Euclidean distance between the image block center point and the image center point, and θ represents the angle between the image block center point and the positive direction of the horizontal coordinate;

[0085] Step 2.4, the image block processed in step 2.3 is taken as the input of the Transformer module, as shown in Figure 4 , the Transformer module includes a multi-head self-attention unit, a first normalization unit, a multi-layer perception unit and a second normalization unit; the Transformer module aims to explore the connection and mapping relationship between different image blocks;

[0086] Step 2.5, a plurality of Transformer modules are connected in series, and the output of the lth Transformer module is obtained, wherein the output of the lth Transformer module is:

[0087] z l = MLP (LN (z l ′) ) + z l

[0088] wherein, MLP is the multi-layer perception unit of the lth Transformer module, LN is the layer normalization operation, z l ′ represents the output of the (l-1) th Transformer module, and the expression is:

[0089] z l ′ = MSA (LN (z l-1 )) + z l-1

[0090] wherein, MSA is the multi-head self-attention mechanism unit of the (l-1) th Transformer module.

[0091] In step 2.4, the multi-head self-attention mechanism combines multiple self-attention mechanisms and reduces the computational loss by reducing the dimension. The core variables of self-attention include Query (Q), Key (K) and Value (V), and the calculation formula of a single self-attention is:

[0092]

[0093] wherein Q represents the current image block, K represents the correlation between the current image block and the queried image block, and V represents the queried image block, , represents the normalization of the current score, and B represents the network threshold.

[0094] The essential difference between images and texts lies in the spatial correlation of images. The original Transformer adopts one-dimensional linear coding, which has achieved excellent performance in image classification problems, but it is difficult to handle complex pixel-level classification problems. IVUS images have very unique annular structure distribution characteristics, which can be summarized as: images within a similar radius show similar performance, and images within a small angle range show similar performance. Traditional CNN methods can better learn the above spatial position information by using convolution kernel translation and pooling, and the transformer series algorithm is difficult to surpass CNN on small sample datasets due to the lack of such inductive bias. However, in step 2.3, as shown in the figure, Figure 5 , a Polar-bias coding is added to the image block, so that each image block is bound to its polar coordinate position through this coding, forming an inductive bias that conforms to the distribution rule of IVUS images, which can reduce the network's demand for data volume and improve the classification effect.

[0095] Step 3 specifically includes the following steps:

[0096] Step 3.1, a curve is explicitly defined using a snake model, and the curve is represented as:

[0097] x(s)=(x(s),y(s))

[0098] wherein x(s) on the left side of the equation is a vector representing the position of the curve, and x(s) and y(s) on the right side of the equation represent the projections of the curve on the two coordinate axes, respectively; the curve is used to describe the boundary position;

[0099] Step 3.2, an energy function is constructed The energy function is represented as:

[0100]

[0101] wherein E snake is the curve energy, and E intE is internal energy image E is external energy

[0102] E is internal energy int E is external energy

[0103]

[0104] wherein x' represents the first derivative of the curve x, x'' represents the second derivative of the curve x, and a and b respectively correspond to the weight coefficients of the two;

[0105] E is internal energy image E is external energy

[0106] E is internal energy image (v(s)) = ω line E is external energy line + ω edge E is internal energy edge + ω term E is external energy term

[0107] wherein ωli ne Eli ne represents line energy, ω edge E is internal energy edge represents edge energy, ω term E is internal energy term represents corner point energy, E image is composed of line energy, edge energy and corner point energy;

[0108] It should be noted that the core of the algorithm is to construct a suitable energy function, so that when the energy function obtains a minimum value, the curve converges to the target boundary. The present application takes the line integral of the curve energy E snake as the energy function, and the two sub-modules in E snake cooperate with each other to explore the target boundary as a whole. In addition, E int is the weighted result of the first derivative of the curve, which is the elastic force, and the second derivative, which is the bending force. The two forces work together to make the curve smooth; E image represents external energy, which is mainly composed of line energy, edge energy and corner point energy, and is the main driving force for curve evolution.

[0109] Step 3.3, define GVF force field and energy functional ε, the GVF force field is represented as:

[0110] v(x,y) = [u(x,y),v(x,y)]

[0111] wherein v(x,y) is the GVF force field, u(x,y) and v(x,y) are two coordinate components of the GVF force field;

[0112] The energy functional ε is represented as:

[0113]

[0114] where f represents the boundary information response of the image, μ is a regularization parameter and μ > 0, and controls the balance of the two terms of the force field, when μ is small, the current position (x, y) is in the uniform area, the first term plays a major role, and controls the curve to slowly diffuse around; when μ is large, the current position (x, y) is in the edge area, only can make the energy functional minimum, so the curve will evolve rapidly towards the target boundary.

[0115] In step 3.3, GVFs are named gradient vector flow, a new variable f representing the boundary information response of the image is introduced, f can have multiple forms, but under different forms, there are three characteristics: the direction points to the boundary; the amplitude is the strongest at the boundary; the amplitude is approximately 0 in the uniform area.

[0116] The GVFs snake model established through the above steps 3.1 to 3.3 can make the curve slowly evolve in the uniform area.

[0117] Step 3.4, in view of the fact that each structure in the IVUS image has a relatively fixed spatial position relationship with the intima and media-adventitia boundaries, the IVUS image classification result in step 3 is obtained, and the energy function of the curve evolution is improved according to the above steps 3.1 to 3.3 and the external force field is improved.

[0118] In step 3.4, the improvement of the external force field specifically includes the following steps:

[0119] Step 3.4.1, according to the IVUS image classification result, the overall force field trend of the image is given;

[0120] In step 3.4.1, if it is inside the boundary, the force field departs from the center, and vice versa.

[0121] Step 3.4.2, according to the IVUS image classification result, the force field of the surrounding pixels is directed to the boundary position;

[0122] Step 3.4.3, in the IVUS original image I, a boundary filter is used to fine-tune the force field at the position where the IVUS image classification result is inaccurate.

[0123] Through the above steps 3.4.1 to 3.4.3, the force field direction and amplitude of each pixel are calculated from coarse to fine, which jointly guide the boundary evolution process.

[0124] Example 2:

[0125] As Figure 6 ​​As shown, five different frequency, different state of IVUS image segmentation results are shown. The first row is the IVUS original image, the second row is the TransUnet network classification result image, on the basis of the first row of the IVUS original image, the third row of the solid line is the intima and adventitia boundary position automatically extracted by the algorithm, and the fourth row of the dashed line represents the expert hand-drawn gold standard curve with rich clinical experience. As can be seen from the figure, the algorithm proposed in the present study can accurately extract the intima and adventitia boundary position. It is worth noting that the data enhancement method proposed in the present study greatly increases the robustness of the algorithm in the case of shadow, artifact, side vessel and vessel bifurcation, and improves the overall segmentation effect of the algorithm.

[0126] Example 3:

[0127] In order to objectively evaluate the performance of the algorithm proposed in the present application, and also facilitate comparison with other new algorithms, the IVUS image public data set established by Balocco research team in 2014 is adopted, and three evaluation indexes adopted by Balocco et al. in 2014 are used: Jaccard distance (JACC), area difference percentage (PAD) and Hausdorff distance (HD). JACC describes the degree of coincidence between the segmentation result area and the gold standard area. The higher the degree of coincidence, the larger the JACC, and the more accurate the extracted boundary. PAD describes the difference between the area occupied by the segmentation result and the area occupied by the gold standard. The more similar the areas, the smaller the PAD, indicating that the algorithm result is more accurate. HD describes the maximum value of the Euclidean distance between the extracted curve and the gold standard curve. The smaller the value, the more accurate the extracted boundary.

[0128] The 12 statistical evaluation indexes of the two data sets are shown in Tables 1 and 2, and all the evaluation indexes are displayed in the form of mean (variance). The algorithm proposed in the present study is first compared with the standard Unet network and TransUnet network, and then compared with the deep learning improved network DPUnet and IVUSNet widely recognized in the field, and finally compared with the EREL algorithm based on traditional machine learning with the best performance in the past three years. In the longitudinal comparison, some studies only study the data set B, and do not give the PAD index, and the present study does not repeat the experiment. From the comparison of the two data sets, the evaluation indexes of the TransUnet network are generally better than those of the Unet network, and even in some cases can be comparable to other algorithms with fine tuning, which shows the advantage of the network structure. Further, the improved TransUnet network in the present study has achieved the best performance in 11 of the 12 evaluation indexes, fully demonstrating the advantage and effectiveness of its improvement.

[0129] Table 1 Algorithm evaluation results of data set A (center frequency 40 MHz)

[0130]

[0131] Table 2 Algorithm evaluation results for dataset B (center frequency 20 MHz)

[0132]

Claims

1. A method for extracting intravascular ultrasound image boundaries based on Polar-TransUnet network, characterized in that: include: Step 1: Obtain the IVUS original image I in the test set, and perform data enhancement on the IVUS original image I to obtain the IVUS enhanced image I * , thereby obtaining the enhanced data set corresponding to the test set; Step 2: Train the TransUnet network based on the enhanced dataset, and then classify the IVUS original images I in the test set to obtain the IVUS image classification results; In step 2, the IVUS image classification result includes a foreground and a background; the foreground is the plaque portion between the tunica intima and the media-adventitia in the IVUS image, and the background is the rest of the IVUS image; Step 2 specifically includes the following steps: Step 2.1, building a TransUnet network framework, wherein the TransUnet network framework has a U-shaped structure and jump connection characteristics; Step 2.2: Slice the IVUS image to obtain image block x p , and then the image block x p Perform linear projection. The formula for linear projection is: Among them, E is the linear mapping matrix, E pos is the position encoding, z0 is the mapped vector; Step 2.3: Add Polar-bias coding to the image block. The coding formula is: E pos =Norm(R)+cosθ Among them, Norm represents normalization, R represents the Euclidean distance between the center point of the image block and the center point of the image, and θ represents the angle between the center point of the image block and the positive direction of the horizontal coordinate; Step 2.4: Use the image block processed in step 2.3 as the input of the Transformer module, which includes a multi-head self-attention unit, a first normalization unit, a multi-layer perceptron unit, and a second normalization unit; Step 2.5: Connect l Transformer modules in series and obtain the output of the lth Transformer module. The output of the lth Transformer module is: With l =MLP(LN(z′ l ))+z′ l Among them, MLP is the multi-layer perceptron unit of the lth Transformer module, LN is the layer normalization operation, z l ′ represents the output of the l-1th Transformer module, and the expression is: z l ′=MSA(LN(z l-1 ))+z l-1 Among them, MSA is the multi-head self-attention mechanism unit of the l-1th Transformer module Step 3: Based on the IVUS image classification result of step 2, extract the IVUS image boundary.

2. The method for extracting intravascular ultrasound image boundaries based on the Polar-TransUnet network according to claim 1, characterized in that: The step 1 specifically includes the following steps: Step 1.1: Model the side vessels, vascular bifurcations, guidewire artifacts, and shadows. The modeling results include side vessel I side , vascular bifurcation I bf , Guidewire Artifact I artifact and Shadow I sh , the modeling results can be rotated 360 degrees; Step 1.2: Obtain the IVUS original image I, rotate each modeling result in step 1.1 by any angle, and then superimpose it on the IVUS original image I with random probability to obtain the IVUS enhanced image I * ;IVUS enhanced image I * It can be expressed as: I * =I*ε(I sh )*ε(I bf )*ε(I side )*ε(I artifact ) Where ε is the switching function.

3. The method for extracting intravascular ultrasound image boundaries based on the Polar-TransUnet network according to claim 2, characterized in that: In step 2.4, the multi-head self-attention unit is used to combine multiple self-attentions. The calculation formula for a single self-attention is: Among them, Q represents the current image block, K represents the correlation between the current image block and the queried image block, and V represents the queried image block. represents the normalization of the current score, and B represents the network threshold.

4. The method for extracting intravascular ultrasound image boundaries based on the Polar-TransUnet network according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: Use the snake model to explicitly define a curve, which is represented by: x(s)=(x(s),y(s)) Where x(s) on the left side of the equation is a vector representing the position of the curve, and x(s) and y(s) on the right side of the equation represent the projections of the curve on the two coordinate axes respectively; Step 3.2: Construct energy function The energy function Expressed as: Among them, E snake is the curve energy, E int is the internal energy, E image For external energy; E int Expressed as: Where x' represents the first-order derivative of the curve x, x" represents the second-order derivative of the curve x, and α and β correspond to the weight coefficients of the two respectively; E image Expressed as: E image (v(s))=ω line E line +oh edge E edge +oh term E term Among them, ω line E line represents the linear energy, ω edge E edge represents the edge energy, ω term E term represents the corner energy, E image It is composed of line energy, edge energy and corner energy weighted; Step 3.3, define the GVF force field and energy functional ε, the GVF force field is expressed as: v(x,y)=[u(x,y),v(x,y)] Among them, v(x,y) is the GVF force field, u(x,y) and v(x,y) are the two coordinate components of the GVF force field; The energy functional ε is expressed as: Among them, f represents the boundary information response of the image, μ is the regularization parameter and μ>0, which controls the balance of the two force fields. When it is small, the current position (x, y) is in a uniform area, and the first term plays a major role, controlling the curve to slowly spread to the surrounding area; when When it is very large, the current position (x, y) is in the edge area, only In order to minimize the energy functional, the curve will evolve rapidly toward the target boundary; Step 3.4: Obtain the IVUS image classification results in step 3, and improve the energy function of the curve evolution according to the above steps 3.1 to 3.

3. and improve the external force field.

5. The method for extracting intravascular ultrasound image boundaries based on the Polar-TransUnet network according to claim 4, characterized in that: In step 3.4, the improved external force field specifically includes the following steps: Step 3.4.1, providing the overall force field trend of the image based on the IVUS image classification result; Step 3.4.2, directing the force field of the surrounding pixels toward the boundary position according to the IVUS image classification result; Step 3.4.3: Use a boundary filter in the IVUS original image I to perform force field fine-tuning on locations where the IVUS image classification results are inaccurate.

Images