Method and system for reconstructing three-dimensional model of blood vessels in CT images

Abstract

The present invention discloses a method and system for reconstructing a three-dimensional model of blood vessels in CT images. The three-dimensional model reconstruction method comprises: establishing an image segmentation model based on a neural network; the image segmentation model is used to segment the blood vessels in the CT image and output a two-dimensional segmentation map of the blood vessels; the CT image to be modeled is input into the image segmentation model to obtain multiple two-dimensional segmentation maps; and the two-dimensional segmentation map is three-dimensionally reconstructed. The image segmentation model constructed by the present invention can quickly and accurately segment the blood vessels in the CT image, and output a two-dimensional segmentation map of the blood vessels, thereby accurately constructing a three-dimensional model of the blood vessels.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a method and system for reconstructing a three-dimensional model of blood vessels from a CT (computer tomography) image. Background Art

[0002] Current CT technology can already visualize the lesion area in the human body in 2D (plane graphics), but it is difficult to imagine the 3D structure of the lesion tissue from the 2D CT image. The construction of a 3D CAD (computer-aided design) model can help doctors make a more accurate diagnosis of the disease.

[0003] In the prior art, the three-dimensional reconstruction of CT images often adopts the method of manual marking: relying on personnel with certain professional experience to manually mark the CT images, and further segmenting and processing them using related software. The obvious disadvantages of this method are the additional manual marking cost, manual marking time and the errors that may be caused by manual work. The latter two points may cause the delay of medical diagnosis and affect the accuracy of medical diagnosis, thereby delaying the patient's further diagnosis and treatment, resulting in more serious consequences.

[0004] In recent years, the technology of automatic image segmentation based on pixel contrast difference has developed rapidly, but it is only suitable for images where the segmented object has obvious color and contrast differences with other surrounding objects. For blurred CT images, the segmented object (such as blood vessels) may be very blurred, so the image segmentation effect based on pixel contrast difference will be very poor, and even multiple unnecessary objects will be segmented.

[0005] It can be seen that the accuracy of segmenting CT images with blurred segmentation objects using the existing image automatic segmentation technology is very low and cannot meet the quality requirements of subsequent three-dimensional model reconstruction. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the defect that the existing automatic image segmentation technology segments CT images with blurred segmentation objects with very low accuracy, and to provide a method and system for reconstructing the three-dimensional model of blood vessels in CT images.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] A method for reconstructing a three-dimensional model of a blood vessel in a CT image, the method comprising:

[0009] An image segmentation model is established based on a neural network; the image segmentation model is used to segment blood vessels in a CT image and output a two-dimensional segmentation map of the blood vessels;

[0010] Inputting the CT image to be modeled into the image segmentation model to obtain a plurality of two-dimensional segmentation images;

[0011] The two-dimensional segmentation map is reconstructed in three dimensions.

[0012] Preferably, the step of establishing an image segmentation model based on a neural network specifically includes:

[0013] Obtain the marked CT images as training samples;

[0014] The training samples are input into a neural network model, and the neural network model is trained according to a loss function of the neural network model to obtain the image segmentation model.

[0015] Preferably, the loss function is:

[0016]

[0017] SI=∑ i,j M predict (i,j)×M true (i,j);

[0018] Sum(M predict )=∑ i,j |M predict (i,j)|;

[0019] Sum(M true )=∑ i,j |M true (i,j)|;

[0020] Wherein, L represents the loss function; M true (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map of the training sample; M predict (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map output by the neural network model; η represents the smoothing parameter.

[0021] Preferably, the step of performing three-dimensional reconstruction on the two-dimensional segmentation map specifically includes:

[0022] Converting the two-dimensional segmentation map into a three-dimensional image in a three-dimensional binary Analyze format;

[0023] Reconstructing the three-dimensional image of the Analyze format based on a advancing cube algorithm;

[0024] The reconstructed 3D image is fitted based on NURBS.

[0025] Preferably, before the step of fitting the reconstructed three-dimensional image, the method further comprises:

[0026] The reconstructed three-dimensional image is smoothed.

[0027] A three-dimensional model reconstruction system for blood vessels in CT images, the three-dimensional model reconstruction system comprising:

[0028] A modeling module, used to establish an image segmentation model based on a neural network; the image segmentation model is used to segment blood vessels in a CT image and output a two-dimensional segmentation map of the blood vessels;

[0029] The three-dimensional reconstruction module is used to input the CT image to be modeled into the image segmentation model to obtain multiple two-dimensional segmentation maps, and perform three-dimensional reconstruction on the two-dimensional segmentation maps.

[0030] Preferably, the modeling module specifically includes:

[0031] A data acquisition unit, used for acquiring the marked CT images as training samples;

[0032] The model training unit is used to input the training samples into the neural network model, train the neural network model according to the loss function of the neural network model, and obtain the image segmentation model.

[0033] Preferably, the loss function is:

[0034]

[0035] SI=∑ i,j M predict (i,j)×M true (i,j);

[0036] Sum(M predict )=∑ i,j |M predict (i,j)|;

[0037] Sum(M true )=∑ i,j |M true (i,j)|;

[0038] Wherein, L represents the loss function; M true (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map of the training sample; M predict (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map output by the neural network model; η represents the smoothing parameter.

[0039] Preferably, the three-dimensional reconstruction module specifically includes:

[0040] A format conversion unit, used for converting the two-dimensional segmentation map into a three-dimensional image in a three-dimensional binary Analyze format;

[0041] A three-dimensional reconstruction unit, used for reconstructing the three-dimensional image in the Analyze format based on a advancing cube algorithm;

[0042] The fitting unit is used to fit the reconstructed three-dimensional image based on NURBS.

[0043] Preferably, the three-dimensional reconstruction module further includes:

[0044] A smoothing unit is used to call the fitting unit after performing smoothing on the reconstructed three-dimensional image.

[0045] The positive and progressive effect of the present invention is that the image segmentation model constructed by the present invention can quickly and accurately segment the blood vessels in the CT image, and output the two-dimensional segmentation map of the blood vessels, thereby accurately constructing a three-dimensional model of the blood vessels. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a first flow chart of a method for reconstructing a three-dimensional model of blood vessels in a CT image according to Embodiment 1 of the present invention.

[0047] Figure 2 This is a second flow chart of the method for reconstructing a three-dimensional model of blood vessels in a CT image according to Embodiment 1 of the present invention.

[0048] Figure 3 This is a module schematic diagram of a system for reconstructing a three-dimensional model of blood vessels in a CT image according to Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below by way of examples, but the present invention is not limited to the scope of the examples.

[0050] Example 1

[0051] This embodiment provides a method for reconstructing a three-dimensional model of blood vessels in a CT image. Figure 1 As shown, the three-dimensional model reconstruction method includes the following steps:

[0052] Step 101: Establish an image segmentation model based on a neural network.

[0053] The image segmentation model is used to segment the CT image and output a two-dimensional segmentation map of the blood vessels in the CT image. In this embodiment, a fully convolutional neural network is used as the basic structure for automatic blood vessel segmentation for CT images.

[0054] See also Figure 2, the following describes the process of establishing the image segmentation model:

[0055] Step 101 - 1 : Obtain a marked CT image as a training sample.

[0056] In step 101-1, a large number of relevant high-precision manually labeled CT images are obtained, part of which is used as training samples of the neural network model for training the neural network model; and part of which is used as test samples for testing the accuracy of the neural network model after each iterative training.

[0057] Step 101 - 2: input the training samples into the neural network model, train the neural network model according to the loss function of the neural network model, and obtain an image segmentation model.

[0058] In this embodiment, a dedicated convolutional neural network loss function is designed for the specific situation of blood vessel segmentation. The loss function L can be, but is not limited to:

[0059]

[0060] SI=∑ i,j M predict (i,j)×M true (i,j);

[0061] Sum(M predict )=∑ i,j |M predict (i,j)|;

[0062] Sum(M true )=∑ i,j |M true (i,j)|;

[0063] Among them, M true represents the real segmentation map (two-dimensional segmentation map) data of the training sample CT image. The real segmentation map is obtained by the trained labeler labeling the original CT image with the corresponding labeling software; M predict Represents the predicted segmentation map (two-dimensional segmentation map) data output by the neural network model during model training; M true (i, j) represents the pixel value at position (i, j) in the true segmentation map; M predict (i, j) represents the pixel value at position (i, j) in the predicted segmentation map; SI represents the segmentation intersection between the true segmentation map and the predicted segmentation map; η represents the smoothing parameter.

[0064] It should be noted that η can be set according to actual needs. In this embodiment, for numerical calculation considerations, the "smoothing parameter" is set to 1. For the neural network model, the architecture of the training model can be selected according to actual needs (such as different case CT images, accuracy, etc.), and the corresponding model parameters are also different.

[0065] Take a training sample (2D CT image, size W×L, W represents the width of the CT image, T represents the length of the CT image) as an example. The pixels in the segmented area (that is, the area where the blood vessels are located) in the "true segmentation map" are white, and the pixels outside the segmented area are black. The "true segmentation map" is represented by the matrix M true ∈R W×L , then M true (i, j) = 1 represents the pixel point of the segmentation area in the real segmentation map, that is, the white pixel point; M true (i, j) = 0 represents the pixels outside the segmentation area in the true segmentation map, that is, the black pixels. Similarly, the "predicted segmentation map" output by the neural network model is represented as M predict ∈R W×L , then M predict (i, j) = 1 indicates the pixel point of the segmentation area in the predicted segmentation map, that is, the white pixel point; M true (i, j) = 0 indicates the pixel points outside the segmentation area in the predicted segmentation map, that is, the black pixel points.

[0066] During the model training process of step 101-2, the structure of the convolutional neural network is adjusted with reference to the accuracy observed during the entire training and testing process (the accuracy of the model is calculated based on the identified test samples). When the loss function converges to a local optimal point and satisfactory performance is observed on the test data, the iteration is stopped, thereby improving the training accuracy.

[0067] Step 102: Input the CT image to be modeled into the image segmentation model to obtain a plurality of two-dimensional segmentation images of blood vessels.

[0068] Step 103: reconstruct the two-dimensional segmentation map into three dimensions.

[0069] In this embodiment, step 103 specifically includes:

[0070] Step 103 - 1 : Convert each two-dimensional segmentation map into a three-dimensional image in a three-dimensional binary Analyze format (a medical imaging format).

[0071] Specifically, ITK (an image analysis extension software tool) is used to convert each two-dimensional segmentation map into a three-dimensional image in a three-dimensional binary Analyze format.

[0072] The generated three-dimensional image contains multiple segmented areas that are not connected to each other. ITK is used to calculate the connected area with the largest volume and remove other pixels that are not connected to the largest segmented area. In other words, isolated pixels in the image are removed, and only the part of the image that needs to be reconstructed in three dimensions is retained, which is the three-dimensional image.

[0073] Step 103 - 2 : Reconstruct the 3D image in the Analyze format based on the Marching Cubes algorithm.

[0074] In this embodiment, the advancing cube algorithm is mainly divided into two parts:

[0075] The first part is to use the divide-and-conquer method to find the surface position of the segmented object in the image. Each cube determines how the object surface intersects with it and moves to the next cube position. The vertices of each cube are assigned 1 and 0 according to whether they are inside or outside the segmented object. A cube has 6 vertices, and different combinations of the 6 vertices represent different intersecting states of the object surface and the cube. Excluding symmetry, there are 14 different intersecting modes. For each intersecting mode, the algorithm can determine which edge intersects with the surface by interpolating the pixel values ​​of the vertices.

[0076] The second part of the advancing cube algorithm is to calculate the unit normal vector of the triangle formed by the object and the cube. The normal vector can be used to determine the inside and outside of the surface. To determine the normal vector of the triangle, the algorithm takes the approach of interpolating the gradient vectors of the cube's vertices. The gradient vector for each vertex is obtained by the central difference method:

[0077]

[0078]

[0079]

[0080] Where D(i,j,k) represents the pixel value at position (i,j,k) in the 3D image. Δx, Δy, and Δz represent the side lengths of the cube. x (i,j,k),G y (i,j,k) and G z (i, j, k) represent the gradient values ​​in the x, y, and z directions at the vertex, respectively. The algorithm obtains the gradient vector of the intersection point by interpolating the two vertices of the intersecting edge, and finally advances the cube across the entire image to obtain the surface of the entire segmented object and obtain the node coordinate information of the three-dimensional grid.

[0081] Step 103 - 3 , smoothing the reconstructed three-dimensional image.

[0082] In this embodiment, in order to make the generated segmented object surface smoother or to reduce the change of the large curvature of the surface, in this embodiment, two consecutive Gaussian smoothing steps are used to implement the smoothing processing algorithm.

[0083] In the Gaussian smoothing process, the position of each vertex of the triangle unit of the 3D mesh is recalculated by the vertices of its neighboring triangle units. The coordinate vector V of a vertex m The vertex coordinate vector V of the adjacent triangle unit n , refers to all and V m The triangle unit vertices that are on the same edge or on the same triangle unit, and do not include vertex V m If V m The vertices of adjacent triangle units are defined as S m = {V n :n=1,2,3,…,p}, where p is the total number of neighboring vertices, an average vector ΔV m is calculated as:

[0084]

[0085] Among them, w mo is a trade-off parameter whose value is set to 1≤o≤p; Vertex V m The coordinates are updated to:

[0086] V m ′=V m +λΔV m ;

[0087] The value of the stretch parameter λ is in the range of 0 to 1.

[0088] In this embodiment, the first step of Gaussian smoothing uses a positive extension parameter λ. The second step of Gaussian smoothing uses a negative extension parameter μ. The two parameters satisfy the relationship 0<λ<-μ<1. Assume that the vertex coordinates obtained by the first Gaussian smoothing are V m ′, the second Gaussian smoothing process is:

[0089] V m ″=V m ′+μΔV m ′;

[0090] Where, ΔV m ′ is the average vector obtained based on the new vertex coordinates after the first smoothing process, V m″ is the new vertex coordinates of the second smoothing process obtained based on the new vertex coordinates after the first smoothing process. By continuing to use the above two-step Gaussian smoothing process on the surface after the smoothing process, the obtained surface can be smooth enough.

[0091] Step 103 - 4 : Fit the smoothed three-dimensional image based on NURBS.

[0092] The following is a further explanation of the image fitting process:

[0093] (1) Analysis of the main directions of point clouds in three-dimensional space

[0094] For the convenience of description, the three-dimensional point cloud set V of the smoothed three-dimensional image is m ″Redefine the point cloud X{X1,X2,…,X r}∈R 3 , and define the first three principal directions obtained by PCA (Principal Component Analysis) decomposition at the beginning of NURBS CAD 3D reconstruction as the point cloud coordinate system E{E1, E2, E3}. The solution process is briefly described as follows:

[0095] Constructing the adjoint matrix in Perform eigenvalue decomposition on the adjoint matrix, that is, solve ∑E=λE. Take the first three eigenvectors of E as the initial coordinate system of the point cloud E{E1,E2,E3}.

[0096] (2) Rigid body rotation of point cloud in three-dimensional space

[0097] Due to the different conditions when shooting the point cloud of vascular reconstruction, the spatial orientation and position are uncertain when starting the NURBS CAD 3D reconstruction. In order to ensure that the surface has the same coordinate system when entering the surface reconstruction stage, the point cloud rigid body must first be rotated to a unified global coordinate system. The method used in this embodiment is SVD (Singular-value decomposition). Assuming the global coordinate system is e{e1, e2, e3}, the similarity adjoint matrix is ​​constructed as follows: in H can be considered as a singular value matrix. Then it can be decomposed into H = U∑V T By decomposing the similar adjoint matrix H, the rotation matrix R can be obtained from UV T The rigid body displacement t can be obtained by So, the point cloud x{x1,x2,…,x r}∈R 3 It is x=RX+t.

[0098] (3) NURBS curve fitting of 2D point cloud

[0099] This patent adopts the fitting of layered NURBS surface. Each layer is fitted as a ring NURBS curve. The curve equation is:

[0100]

[0101] Among them, P i is the control point coordinate vector, ω i and ω j is its corresponding weight value. The basis function N is obtained by the following recursive function:

[0102]

[0103]

[0104] Among them, u={u0,…,u m} is a node vector, which is a set of non-negative, uniform and monotonically increasing node vectors selected arbitrarily; p is the order of the curve, and the value of p can be set according to actual needs. For example, in this embodiment, p = 2. The fitting curve of each layer uses the same set of node vectors. By constructing the least squares equation set N t NP=N t X, we can solve for the control point P i .

[0105] (4) Constructing NURBS surfaces by tensor product

[0106] Since each curve in each layer uses the same node vector, for the same node, a new set of node directions v = {v0,…,v m}. The NURBS surface is calculated by tensor product of NURBS curves in two directions. Its surface equation is:

[0107]

[0108] in, P i,j is the value of the i-th control point in the newly constructed v direction, the j-th layer; n and m are the curve orders in the u and v directions respectively. The values ​​of n and m can be set according to actual needs, for example, in this embodiment, n=2, m=2.

[0109] In this embodiment, the constructed image segmentation model can quickly and accurately segment the blood vessels in the CT image and output a two-dimensional segmentation map of the blood vessels, thereby accurately constructing a three-dimensional model of the blood vessels.

[0110] Example 2

[0111] like Figure 3 As shown, the three-dimensional model reconstruction system of blood vessels in CT images of this embodiment includes: a modeling module 1 and a three-dimensional reconstruction module 2.

[0112] The modeling module 1 is used to establish an image segmentation model based on a neural network. The image segmentation model is used to segment blood vessels in CT images and output a two-dimensional segmentation map of blood vessels in the CT images. In this embodiment, a fully convolutional neural network is used as the basic structure for automated blood vessel segmentation for CT images.

[0113] In this embodiment, the modeling module 1 specifically includes: a data acquisition unit 11 and a model training unit 12 .

[0114] The data acquisition unit 11 is used to acquire the marked CT images as training samples. Specifically, the data acquisition unit 11 acquires a large number of relevant high-precision manually labeled CT images, a part of which is used as training samples for the neural network model for training the neural network model; and a part of which is used as test samples for testing the accuracy of the neural network model after each iterative training.

[0115] The model training unit 12 is used to input the training samples into the neural network model, train the neural network model according to the loss function of the neural network model, and obtain the image segmentation model.

[0116] During the model training process, the model training unit 12 adjusts the structure in the convolutional neural network with reference to the accuracy observed during the entire training and testing process (the accuracy of the model is calculated based on the identified test samples). When the loss function converges to a local optimal point and satisfactory performance is observed on the test data, the iteration is stopped, thereby improving the training accuracy.

[0117] In this embodiment, a dedicated convolutional neural network loss function is designed for the specific situation of blood vessel segmentation. The loss function L can be, but is not limited to:

[0118]

[0119] SI=∑ i,j M predict (i,j)×M true (i,j);

[0120] Sum(M predict )=∑ i,j |M predict (i,j)|;

[0121] Sum(M true )=∑ i,j |M true (i,j)|;

[0122] Among them, M true represents the real segmentation map (two-dimensional segmentation map) data of the training sample CT image. The real segmentation map is obtained by the trained labeler labeling the original CT image with the corresponding labeling software; M predict Represents the predicted segmentation map (two-dimensional segmentation map) data output by the neural network model during model training; M true (i, j) represents the pixel value at position (i, j) in the true segmentation map; M predict (i, j) represents the pixel value at position (i, j) in the predicted segmentation map; SI represents the segmentation intersection between the true segmentation map and the predicted segmentation map; η represents the smoothing parameter.

[0123] It should be noted that η can be set according to actual needs. In this embodiment, for numerical calculation considerations, the "smoothing parameter" is set to 1. For the neural network model, the architecture of the training model can be selected according to actual needs (such as different case CT images, accuracy, etc.), and the corresponding model parameters are also different.

[0124] Take a training sample (2D CT image, size W×L, W represents the width of the CT image, T represents the length of the CT image) as an example. The pixels in the segmented area (that is, the area where the blood vessels are located) in the "true segmentation map" are white, and the pixels outside the segmented area are black. The "true segmentation map" is represented by the matrix M true ∈R W×L , then M true (i, j) = 1 represents the pixel point of the segmentation area in the real segmentation map, that is, the white pixel point; M true (i, j) = 0 represents the pixels outside the segmentation area in the true segmentation map, that is, the black pixels. Similarly, the "predicted segmentation map" output by the neural network model is represented as M predict ∈R W×L , then M predict (i, j) = 1 indicates the pixel point of the segmentation area in the predicted segmentation map, that is, the white pixel point; M true (i, j) = 0 indicates the pixel points outside the segmentation area in the predicted segmentation map, that is, the black pixel points.

[0125] The three-dimensional reconstruction module 2 is used to input the CT image to be modeled into the image segmentation model, obtain multiple two-dimensional segmentation maps, and perform three-dimensional reconstruction on the two-dimensional segmentation maps.

[0126] In this embodiment, the 3D reconstruction module 2 specifically includes: a format conversion unit 21 , a 3D reconstruction unit 22 , a smoothing processing unit 23 and a fitting unit 24 .

[0127] The format conversion unit 21 is used to convert the two-dimensional segmentation map into a three-dimensional image in the three-dimensional binary Analyze format, and output it to the three-dimensional reconstruction unit 22. Specifically, the format conversion unit 21 uses ITK (an image analysis extension software tool) to convert each two-dimensional segmentation map into a three-dimensional image in the three-dimensional binary Analyze format.

[0128] The generated three-dimensional image contains multiple segmented areas that are not connected to each other. ITK is used to calculate the connected area with the largest volume and remove other pixels that are not connected to the largest segmented area. In other words, isolated pixels in the image are removed, and only the part of the image that needs to be reconstructed in three dimensions is retained, which is the three-dimensional image.

[0129] The three-dimensional reconstruction unit 22 is used to reconstruct the three-dimensional image in the Analyze format based on the marching cube algorithm, and output the reconstructed image to the smoothing processing unit 23 .

[0130] In this embodiment, the advancing cube algorithm is mainly divided into two parts:

[0131] The first part is to use the divide-and-conquer method to find the surface position of the segmented object in the image. Each cube determines how the object surface intersects with it and moves to the next cube position. The vertices of each cube are assigned 1 and 0 according to whether they are inside or outside the segmented object. A cube has 6 vertices, and different combinations of the 6 vertices represent different intersecting states of the object surface and the cube. Excluding symmetry, there are 14 different intersecting modes. For each intersecting mode, the algorithm can determine which edge intersects with the surface by interpolating the pixel values ​​of the vertices.

[0132] The second part of the advancing cube algorithm is to calculate the unit normal vector of the triangle formed by the object and the cube. The normal vector can be used to determine the inside and outside of the surface. To determine the normal vector of the triangle, the algorithm takes the approach of interpolating the gradient vectors of the cube's vertices. The gradient vector for each vertex is obtained by the central difference method:

[0133]

[0134]

[0135]

[0136] Where D(i,j,k) represents the pixel value at position (i,j,k) in the 3D image. Δx, Δy, and Δz represent the side lengths of the cube. x (i,j,k),G y (i,j,k) and Gz (i, j, k) represent the gradient values ​​in the x, y, and z directions at the vertex, respectively. The algorithm obtains the gradient vector of the intersection point by interpolating the two vertices of the intersecting edge, and finally advances the cube across the entire image to obtain the surface of the entire segmented object and obtain the node coordinate information of the three-dimensional grid.

[0137] The smoothing processing unit 23 is used to call the fitting unit 24 after performing smoothing processing on the reconstructed three-dimensional image.

[0138] In this embodiment, in order to make the generated segmented object surface smoother or to reduce the change of the large curvature of the surface, in this embodiment, two consecutive Gaussian smoothing steps are used to implement the smoothing processing algorithm.

[0139] In the Gaussian smoothing process, the position of each vertex of the triangle unit of the 3D mesh is recalculated by the vertices of its neighboring triangle units. The coordinate vector V of a vertex m The vertex coordinate vector V of the adjacent triangle unit n , refers to all and V m The triangle unit vertices that are on the same edge or on the same triangle unit, and do not include vertex V m If V m The vertices of adjacent triangle units are defined as S m = {V n :n=1,2,3,…,p}, where p is the total number of neighboring vertices, an average vector ΔV m is calculated as:

[0140] ΔV m =∑ Vo∈sm w mo (V o -V m );

[0141] Among them, w mo is a trade-off parameter whose value is set to 1≤o≤p; Vertex V m The coordinates are updated to:

[0142] V m ′=V m +λΔV m ;

[0143] The value of the stretch parameter λ is in the range of 0 to 1.

[0144] In this embodiment, the first step of Gaussian smoothing uses a positive extension parameter λ. The second step of Gaussian smoothing uses a negative extension parameter μ. The two parameters satisfy the relationship 0<λ<-μ<1. Assume that the vertex coordinates obtained by the first Gaussian smoothing are V m ′, the second Gaussian smoothing process is:

[0145] V m ″=V m ′+μΔV m ′;

[0146] Where, ΔV m ′ is the average vector obtained based on the new vertex coordinates after the first smoothing process, V m ″ is the new vertex coordinates of the second smoothing process obtained based on the new vertex coordinates after the first smoothing process. By continuing to use the above two-step Gaussian smoothing process on the surface after the smoothing process, the obtained surface can be smooth enough.

[0147] The fitting unit 24 is used to fit the reconstructed three-dimensional image based on NURBS.

[0148] The working principle of the image fitting by the fitting unit 24 is further described below:

[0149] (1) Analysis of the main directions of point clouds in three-dimensional space

[0150] For the convenience of description, the three-dimensional point cloud set V of the smoothed three-dimensional image is m ″Redefine the point cloud X{X1,X2,…,X r}∈R 3 , and define the first three principal directions obtained by PCA (Principal Component Analysis) decomposition at the beginning of NURBS CAD 3D reconstruction as the point cloud coordinate system E{E1, E2, E3}. The solution process is briefly described as follows:

[0151] Constructing the adjoint matrix in Perform eigenvalue decomposition on the adjoint matrix, that is, solve ∑E=λE. Take the first three eigenvectors of E as the initial coordinate system of the point cloud E{E1,E2,E3}.

[0152] (2) Rigid body rotation of point cloud in three-dimensional space

[0153] Due to the different conditions when shooting the point cloud of vascular reconstruction, the spatial orientation and position are uncertain when starting the NURBS CAD 3D reconstruction. In order to ensure that the surface has the same coordinate system when entering the surface reconstruction stage, the point cloud rigid body must first be rotated to a unified global coordinate system. The method used in this embodiment is SVD (Singular-value decomposition). Assuming the global coordinate system is e{e1, e2, e3}, the similarity adjoint matrix is ​​constructed as follows: in H can be considered as a singular value matrix. Then it can be decomposed into H = U∑V T By decomposing the similar adjoint matrix H, the rotation matrix R can be obtained from UV T The rigid body displacement t can be obtained by So, the point cloud x{x1,x2,…,x r}∈R 3 It is x=RX+t.

[0154] (3) NURBS curve fitting of 2D point cloud

[0155] This patent adopts the fitting of layered NURBS surfaces. Each layer is fitted as a ring NURBS curve. The curve equation is:

[0156]

[0157] Among them, P i is the control point coordinate vector, ω i and ω j is its corresponding weight value. The basis function N is obtained by the following recursive function:

[0158]

[0159]

[0160] Among them, u={u0,…,u m} is a node vector, which is a set of non-negative, uniform and monotonically increasing node vectors selected arbitrarily; p is the order of the curve, and the value of p can be set according to actual needs. For example, in this embodiment, p = 2. The fitting curve of each layer uses the same set of node vectors. By constructing the least squares equation set N t NP=N t X, we can solve for the control point P i .

[0161] (4) Constructing NURBS surfaces by tensor multiplication

[0162] Since each curve in each layer uses the same node vector, for the same node, a new set of node directions v = {v0,…,v m}. The NURBS surface is calculated by tensor product of NURBS curves in two directions. Its surface equation is:

[0163]

[0164] in, P i,j is the value of the i-th control point in the newly constructed v direction, the j-th layer; n and m are the curve orders in the u and v directions respectively. The values ​​of n and m can be set according to actual needs, for example, in this embodiment, n=2, m=2.

[0165] In this embodiment, the constructed image segmentation model can quickly and accurately segment the blood vessels in the CT image and output a two-dimensional segmentation map of the blood vessels, thereby accurately constructing a three-dimensional model of the blood vessels.

[0166] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that this is only for illustration and the protection scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A method for reconstructing a three-dimensional model of blood vessels in a CT image, characterized in that: The three-dimensional model reconstruction method comprises: An image segmentation model is established based on a neural network; the image segmentation model is used to segment blood vessels in a CT image and output a two-dimensional segmentation map of the blood vessels; Inputting the CT image to be modeled into the image segmentation model to obtain a plurality of two-dimensional segmentation images; Performing three-dimensional reconstruction on the two-dimensional segmentation map; The steps of establishing an image segmentation model based on a neural network include: Obtain the marked CT images as training samples; Inputting the training samples into a neural network model, training the neural network model according to a loss function of the neural network model, and obtaining the image segmentation model; The loss function is: SI=∑ i,j M predict (i,j)×M true (i,j); Sum(M predict )=∑ i,j |M predict (i,j)|; Sum(M true )=∑ i,j |M true (i,j)|; Wherein, L represents the loss function; M true (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map of the training sample; M predict (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map output by the neural network model; η represents the smoothing parameter; The step of performing three-dimensional reconstruction on the two-dimensional segmentation map specifically includes: Converting the two-dimensional segmentation map into a three-dimensional image in a three-dimensional binary Analyze format; Reconstructing the three-dimensional image of the Analyze format based on a advancing cube algorithm; The reconstructed 3D image is fitted based on NURBS.

2. The method for reconstructing a three-dimensional model of a blood vessel in a CT image according to claim 1, wherein: Before the step of fitting the reconstructed three-dimensional image, the following steps are also included: The reconstructed three-dimensional image is smoothed.

3. A three-dimensional model reconstruction system for blood vessels in CT images, characterized in that: The three-dimensional model reconstruction system comprises: A modeling module, used to establish an image segmentation model based on a neural network; the image segmentation model is used to segment blood vessels in a CT image and output a two-dimensional segmentation map of the blood vessels; A three-dimensional reconstruction module, used for inputting the CT image to be modeled into the image segmentation model, obtaining a plurality of two-dimensional segmentation maps, and performing three-dimensional reconstruction on the two-dimensional segmentation maps; The modeling module specifically includes: A data acquisition unit, used for acquiring the marked CT images as training samples; A model training unit, used for inputting the training samples into a neural network model, training the neural network model according to a loss function of the neural network model, and obtaining the image segmentation model; The loss function is: SI=∑ i,j M predict (i,j)×M true (i,j); Sum(M predict )=∑ i,j |M predict (i,j)|; Sum(M true )=∑ i,j |M true (i,j)|; Wherein, L represents the loss function; M true (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map of the training sample; M predict (i, j) represents the pixel value at position (i, j) in the two-dimensional segmentation map output by the neural network model; η represents the smoothing parameter; The three-dimensional reconstruction module specifically includes: A format conversion unit, used for converting the two-dimensional segmentation map into a three-dimensional image in a three-dimensional binary Analyze format; A three-dimensional reconstruction unit, used for reconstructing the three-dimensional image in the Analyze format based on a advancing cube algorithm; The fitting unit is used to fit the reconstructed three-dimensional image based on NURBS.

4. The three-dimensional model reconstruction system of blood vessels in CT images according to claim 3, characterized in that: The three-dimensional reconstruction module also includes: A smoothing unit is used to call the fitting unit after performing smoothing on the reconstructed three-dimensional image.

Images