A microvascular tree generation method based on U-Net conditional generative adversarial network

Abstract

The present invention provides a method for generating a microvascular tree based on a U‑Net conditional generative adversarial network. First, microvascular bifurcation pair data is extracted from a vascular network to construct a training set. The microvascular bifurcation pair data is a set of maternal and daughter vascular bifurcation data. Then, a conditional generative adversarial network and a pair are input into a discriminator to ensure mapping. A generator for generating lower-level daughter vascular bifurcations based on upper-level maternal vascular bifurcations is obtained through comparative training. Finally, the entry bifurcation of the microvascular tree to be generated is obtained. Sub-line vascular bifurcations are generated step by step according to the trained generator, and the microvascular bifurcation tree is grown at the same time. The present invention establishes a microvascular tree regeneration model based on a U‑Net conditional generative adversarial network. The model can learn the morphology and topological pattern of the microvascular tree from the real vascular network structure, and then generate a simulated microvascular tree similar to the real microvascular tree in morphology and topological structure.

Description

Technical Field

[0001] The invention relates to an artificial generation method of a microvascular tree structure, belonging to the field of biomedicine. Background Art

[0002] Vascular network regeneration is a physiological process involving multiple time scales. It is difficult to simultaneously observe physiological processes in the short and long time courses by controlling experimental conditions in animal experiments, and the process of obtaining a true microvascular network is difficult and complicated. Traditional methods are mostly based on statistics and fractal theory to generate microvascular trees. The generated microvascular trees have high homogeneity and cannot reflect the real microvascular tree structure with high heterogeneity. Therefore, researchers try to simulate the generation of microvascular tree structures. How to generate microvascular trees that conform to the actual physiological structure and topological characteristics has important research value. Summary of the invention

[0003] The conversion between images such as grayscale images, color images, gradient images, etc. is called image translation. These tasks can be solved using the same framework, which is called the Pix2Pix algorithm. The algorithm is implemented based on a conditional generative adversarial network, and the input is an image rather than a random vector. The present invention refers to and draws on the basic idea of ​​the Pix2Pix algorithm, using a conditional generative adversarial network and pair inputs to the discriminator to ensure mapping, and extract and compare high-dimensional array features. In view of the shortcomings of traditional methods, the present invention establishes a microvascular tree regeneration model based on a conditional generative adversarial network, which can learn the morphology and topological pattern of the microvascular tree from the real vascular network structure, and then generate a simulated microvascular tree that is similar to the real microvascular tree in morphology and topological structure.

[0004] The technical solution adopted by the present invention to solve its technical problem is:

[0005] A method for generating a microvascular tree based on a U-Net conditional generative adversarial network comprises the following steps:

[0006] a. Extracting microvascular bifurcation pair data from the vascular network, the microvascular bifurcation pair data is a set of maternal and daughter vascular bifurcation data, wherein the maternal vascular bifurcation data includes the maternal main branch node coordinates and diameter information, the maternal left branch node coordinates and diameter information, and the maternal right branch node coordinates and diameter information of the maternal vascular bifurcation; the daughter vascular bifurcation data is the vascular bifurcation node coordinates and diameter information of a branch in the maternal vascular bifurcation data as the main branch, including the daughter main branch node coordinates and diameter information, the daughter left branch node coordinates and diameter information, and the daughter right branch node coordinates and diameter information.

[0007] b. Based on a deep convolutional neural network, a U-Net conditional generative adversarial network consisting of a generator and a discriminator is constructed; wherein the generator adopts a U-Net structure, the input is the mother vessel bifurcation data of the bifurcation pair, and the output is the generated daughter vessel bifurcation data; the input of the discriminator is the real microvascular bifurcation pair obtained in step a and the bifurcation pair consisting of the fake daughter vessel bifurcation generated by the generator and the real mother vessel bifurcation, which is used to distinguish whether the input data comes from a real sample or a fake sample generated by the generator.

[0008] c. Train the generator and discriminator based on the bifurcation of step a, and obtain the trained generator through the game of loss function;

[0009] d. Generate sub-vascular bifurcations step by step based on the trained generator and grow the microvascular bifurcation tree at the same time.

[0010] Furthermore, in step a, extracting microvascular bifurcation pair data from the vascular network specifically includes the following sub-steps:

[0011] a1 separates the micro-arteriolar tree from the original microvascular network, and stores the extracted coordinates and diameter of each microvascular bifurcation node in a two-dimensional array; each line of the two-dimensional array stores the x, y coordinates and diameter information of the main branch, left branch, and right branch nodes of each microvascular bifurcation.

[0012] a2 normalizes the node coordinate values ​​and pipe diameter values ​​respectively so that their values ​​are between 0 and 1. The normalization formula is:

[0013] normValue=(value-min) / (max-min)

[0014] Among them, value represents the pipe diameter value or coordinate node, min represents the minimum value of the pipe diameter or node coordinate in the data set, and max represents the maximum value of the pipe diameter or node coordinate in the data set;

[0015] a3 fills 0 after each row of data that is less than N in the two-dimensional array array, forming a 9×N matrix, where N is the maximum number of child nodes contained in the microvascular branch; in the diameter information of the main branch, left branch, and right branch, the index of the zero-filled position of the branch is set to 1 as a label. To facilitate the input of data into the neural network for training, add a row to the array and fill it with 0, so that the matrix size becomes 10×N.

[0016] a4 selects paired mother and daughter vascular bifurcation data from the microvascular bifurcations in the two-dimensional array array to form 20×N microvascular bifurcation pair data as bifurcation pair data.

[0017] Furthermore, the step a further comprises a data amplification sub-step, specifically:

[0018] In the range of 0° to 360°, the microvascular bifurcation tree stored in the array array is rotated by increasing α degrees each time to amplify the data. The rotation formula is as follows:

[0019] x1=x0 cosα-y0 sinα

[0020] y1=x0 sinα+y0 cosα

[0021] Among them, x0 and y0 are the node coordinates of the microvascular bifurcation, and x1 and y1 are the node coordinates of the microvascular bifurcation after rotating α degrees clockwise around the origin;

[0022] Furthermore, in step b, the U-Net structure deep convolutional neural network used by the generator has 10 layers, of which the first 5 layers are convolutional layers and the last 5 layers are deconvolutional layers. The number of output nodes is 10×N.

[0023] Furthermore, the discriminator network is also constructed using a deep convolutional neural network model. Its input is in the form of a microvascular bifurcation pair matrix obtained in step a, and the input matrix size is 10×N×2. The number of layers of the deep convolutional neural network model is set to 5, of which the first 4 layers are convolutional layers, and the last layer is an output layer, and the number of nodes in the output layer is 1.

[0024] Furthermore, in step c, the loss function is:

[0025]

[0026] in:

[0027] L GAN (G,D)=E k [logD(k)]+E z [log(1-D(G(z)′))]

[0028] L L1 (G) = E z,k [||kG(z)′||1]

[0029] L GAN (G, D) is the loss function of the basic generative adversarial network, L L1The (G) function is used to stimulate less ambiguity and increase the similarity of true and false bifurcation pairs in the discriminator. z represents the parent vascular bifurcation in the original microvascular bifurcation data k, E represents the expectation; G(z) represents the false lower-level sub-vascular bifurcation generated by the generator after the input vector z; D(k) represents the output of the discriminator D after the input k; G(z)′ represents the combination of G(z) and z; D(G(z)′) represents the output of G(z)′ after the discriminator D, ||*||1 is a norm processing, and λ is a hyperparameter.

[0030] Furthermore, in step d, the sub-branch vascular bifurcations are generated step by step according to the trained generator, and the microvascular bifurcation tree is grown, including the following sub-steps:

[0031] d1 selects a microvascular bifurcation with the largest average branch diameter from the microvascular bifurcation library as the entrance bifurcation of the vascular tree to be generated, that is, the 0th level bifurcation of the microvascular tree, and takes the entrance bifurcation as the mother vascular bifurcation and inputs it into the trained generator to generate the next level of daughter vascular bifurcation.

[0032] d2 translates the endpoint (endX, endY) of the generated daughter microvascular bifurcation main branch to the position of the parent vascular bifurcation node (nodeX, nodeY) of the previous level, and then rotates the next level of daughter vascular bifurcation with the node position as the origin, so that the angle between its main branch and the positive direction of the x-axis is the same as the angle between the growth branch of the previous bifurcation and the positive direction of the x-axis.

[0033] d3 re-inputs the daughter vascular bifurcation as the training sample of the next round of generator. At this time, the daughter vascular bifurcation becomes the mother vascular bifurcation, generating the corresponding next-level daughter vascular bifurcation, that is, the second-level bifurcation.

[0034] D4 matches and splices the second-level bifurcation with the first-level bifurcation tree completed by D2 to form a second-level bifurcation tree.

[0035] d5 repeats steps d2-d4 to complete the generation of microvascular bifurcations and the matching connection of microvascular bifurcation trees step by step, until the diameter value is less than a certain threshold or the number of recursions is greater than a specified value, and the generation of a microvascular tree is completed.

[0036] The beneficial effects of the present invention are mainly manifested in that the generative adversarial network can generate false samples of the next-level bifurcation that conform to the distribution of the original data based on the given previous-level bifurcation, and process the microvascular bifurcation data in the form of a two-dimensional array, which can effectively learn the branch structure with different numbers of nodes, and then generate microvascular bifurcations while performing an iterative processing method of microvascular tree splicing, and finally generate a vascular tree with extremely small errors in morphology and topological structure from the real microvascular bifurcation tree. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1This is a flow chart of a method for generating a vascular tree based on a generative adversarial network according to the present invention.

[0038] Figure 2 A schematic diagram of a fork pair

[0039] Figure 3 Schematic diagram of the overall structure of the generative adversarial network.

[0040] Figure 4 This is the generator network structure diagram.

[0041] Figure 5 This is the structure diagram of the discriminator network.

[0042] Figure 6 Flowchart for generating microvascular trees. DETAILED DESCRIPTION

[0043] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below in conjunction with the accompanying drawings through specific embodiments. In the present example, a mouse mesenteric vascular network dataset is selected. However, the present invention is not limited to this dataset and can be applied to different vascular networks.

[0044] For example, in this example, the mouse mesenteric vascular network is composed of micro-arterial trees, micro-venular trees and capillaries, and the microvascular tree is formed by many microvascular bifurcations. Each microvascular bifurcation has three branches ( Figure 2 ), and each branch is composed of a different number of node coordinates and pipe diameter information.

[0045] Reference Figure 1-Figure 6 , a method for artificially generating a microvascular tree structure, comprising the following steps:

[0046] a. extracting microvascular bifurcation pair data from the vascular network and preprocessing the data; wherein the microvascular bifurcation pair data is a set of maternal and daughter vascular bifurcation data, wherein the maternal vascular bifurcation data includes the maternal main branch node coordinates and diameter information, the maternal left branch node coordinates and diameter information, and the maternal right branch node coordinates and diameter information of the maternal vascular bifurcation; and the daughter vascular bifurcation data is the vascular bifurcation node coordinates and diameter information of a branch in the maternal vascular bifurcation data as the main branch, including the daughter main branch node coordinates and diameter information, the daughter left branch node coordinates and diameter information, and the daughter right branch node coordinates and diameter information.

[0047] b. Construct a generator and a discriminator based on U-Net and a deep convolutional neural network; wherein the generator adopts a U-Net structure, the input is the mother vessel bifurcation data of the bifurcation pair, and the output is the generated daughter vessel bifurcation data; the input of the discriminator is the real microvascular bifurcation pair obtained in step a and the bifurcation pair consisting of the fake daughter vessel bifurcation generated by the generator and the real mother vessel bifurcation, which is used to distinguish whether the input data comes from a real sample or a fake sample generated by the generator.

[0048] c. Train the generator and discriminator based on the conditional generative adversarial network;

[0049] d. Obtain the entry bifurcation of the microvascular tree to be generated, generate microvascular bifurcations according to the trained generator, and grow the microvascular bifurcation tree at the same time.

[0050] As a preferred solution, in step a, the data preprocessing includes the following steps:

[0051] a1 separates the micro-arterial tree from the original microvascular network. And stores the node coordinates and diameter of each extracted microvascular bifurcation into a two-dimensional array array, where array[0,:] stores the x-coordinate of the main branch node, array[1,:] stores the x-coordinate of the left branch node, array[2,:] stores the x-coordinate of the right branch node, array[3,:] stores the y-coordinate of the main branch node, array[4,:] stores the y-coordinate of the left branch node, array[5,:] stores the y-coordinate of the right branch node, array[6,:] stores the main branch diameter information, array[7,:] stores the left branch diameter information, and array[8,:] stores the right branch diameter information. Among them, array[i,:] represents the i-th row and all columns of the array array;

[0052] a2 normalizes the microvascular bifurcation coordinate value and the tube diameter value respectively so that their values ​​are between 0 and 1 to retain the relative size information of the microvascular bifurcation. The normalization formula is:

[0053] normValue=(value-min) / (max-min)

[0054] Among them, value represents the pipe diameter value or coordinate node, min represents the minimum value of the pipe diameter or node coordinate in the data set, and max represents the maximum value of the pipe diameter or node coordinate in the data set;

[0055] In addition, the microvascular bifurcation tree stored in the array array can be rotated within the range of 0° to 360° by increasing α degrees each time to amplify the data. The rotation formula is as follows:

[0056] x1=x0 cosα-y0 sinα

[0057] y1=x0 sinα+y0 cosα

[0058] Among them, x0 and y0 are the node coordinates of the microvascular bifurcation, and x1 and y1 are the node coordinates of the microvascular bifurcation after rotating α degrees clockwise around the origin.

[0059] a3 Since the number of child nodes in each microvascular branch is different, the number of columns in each row of the array array is not uniform. Therefore, 0 is added to the end of each row of coordinates and diameter data in the array to form a 9×N matrix, where N is the maximum number of child nodes in the microvascular branch.

[0060] a4 records the index of the position where zero padding starts in array[0,:]~array[5,:], and sets 0 to 1 at the corresponding index recorded in array[6,:]~array[8,:];

[0061] a5 adds array[9,:] to the array and fills them with 0, making the matrix size 10×N;

[0062] a6 converts the array into a list, selects paired data from the microvascular bifurcations in the list array, takes the mother vessel bifurcation as the matching bifurcation, puts the matching branches uniformly in the left branch position, and matches the branch in the data set. If the matching bifurcation contains the same branch and the bifurcation is different from the matching bifurcation, it is determined to be a daughter vessel bifurcation. The two bifurcations are combined to form a complete bifurcation pair; Figure 2 Shown is a visualization of a bifurcated pair.

[0063] a7 stores the extracted bifurcation node coordinates and diameters into a two-dimensional array array, where array[0,:] stores the x-coordinates of the main branch nodes of the mother system, array[1,:] stores the x-coordinates of the left branch nodes of the mother system, array[2,:] stores the x-coordinates of the right branch nodes of the mother system, array[3,:] stores the y-coordinates of the main branch nodes of the mother system, array[4,:] stores the y-coordinates of the left branch nodes of the mother system, array[5,:] stores the y-coordinates of the right branch nodes of the mother system, array[6,:] stores the diameter information of the main branch of the mother system, array[7,:] stores the diameter information of the left branch of the mother system, array[8,:] stores the diameter information of the right branch of the mother system, array[9,:] is filled with 0, and array[10,:] stores the main branch of the child system Node x coordinate, array[11,:] stores the x coordinate of the left branch node of the subsystem, array[12,:] stores the x coordinate of the right branch node of the subsystem, array[13,:] stores the y coordinate of the main branch node of the subsystem, array[14,:] stores the y coordinate of the left branch node of the subsystem, array[15,:] stores the y coordinate of the right branch node of the subsystem, array[16,:] stores the diameter information of the main branch of the subsystem, array[17,:] stores the diameter information of the left branch of the subsystem, array[18,:] stores the diameter information of the right branch of the subsystem. Among them, array[19,:] is filled with 0, array[i,:] represents the i-th row of the array array, and all columns are unified to form a 20×N matrix, where N is the maximum number of subnodes contained in the microvascular branch.

[0064] Through the above steps, the original microvascular bifurcation data can be processed into a form that can be input into the generative adversarial network.

[0065] As a preferred solution, in step b, a U-Net conditional generative adversarial network consisting of a generator and a discriminator is constructed based on a deep convolutional neural network, specifically comprising the following steps:

[0066] The b1 generator network is constructed using a deep convolutional neural network model with a U-Net structure. Its input is the parent vascular bifurcation of the real bifurcation pair. The number of layers of the deep convolutional neural network model is set to 10, of which the first 5 layers are convolutional layers, the last 5 layers are deconvolutional layers, and the number of output nodes is 10×N.

[0067] The network of the b2 discriminator is also constructed using a deep convolutional neural network model. Its input is the real microvascular bifurcation pair obtained in step a and the bifurcation pair consisting of the false lower bifurcation generated by the generator and the real upper bifurcation. The input matrix size is 10×N×2. The number of layers of the deep convolutional neural network model is set to 5, of which the first 4 layers are convolutional layers, the last layer is the output layer, and the number of nodes in the output layer is 1. The generator G and discriminator D designed in this embodiment are respectively as follows: Figure 3 , as shown in 4.

[0068] Furthermore, in step c, based on the U-Net conditional generative adversarial network, the generator generates fake samples that conform to the original data distribution according to the upper bifurcation, and the discriminator is used to distinguish whether the input data comes from real samples or fake samples generated by the generator. The loss function is used to conduct a game to reach a balance state. The training process specifically includes the following steps:

[0069] c1 selects a microvascular bifurcation with the largest average branch diameter from the microvascular bifurcation data as the entry bifurcation of the vascular tree to be generated, that is, the upper-level mother vascular bifurcation, and uses this sample as the input of the generator to generate the lower-level daughter vascular bifurcation;

[0070] c2 splices the generated lower-level daughter vascular bifurcation with the upper-level mother vascular bifurcation and inputs it into the discriminator as a false bifurcation pair;

[0071] c3 For the entrance bifurcation, the bifurcation of the blood vessel in the data set is extracted as the bifurcation pair of the parent blood vessel bifurcation, as the real bifurcation pair, and the true and false bifurcation pairs are compared through the discriminator, and the game is carried out through the loss function;

[0072] c4 uses the actual next-level sub-branch vascular bifurcation of the entrance bifurcation as the training sample input of the next round of generator. At this time, the sub-branch vascular bifurcation becomes the upper-level parent vascular bifurcation, and the corresponding sub-branch vascular bifurcation is generated;

[0073] c5 repeats steps c2-c4 to complete the generation of bifurcations step by step until the diameter value is less than a certain threshold or the number of recursions is greater than a specified value, thus completing a round of iterative generation training of generating the next level of microvascular bifurcations from the previous level of microvascular bifurcations.

[0074] Among them, the loss function is:

[0075]

[0076] in:

[0077] L GAN (G,D)=E k [logD(k)]+E z [log(1-D(G(z)′))]

[0078] L L1 (G) = E z,k [||kG(z)′||1]

[0079] L GAN (G, D) is the loss function of the basic generative adversarial network, L L1 The (G) function is used to stimulate less ambiguity and increase the similarity of true and false bifurcation pairs in the discriminator. z represents the parent vascular bifurcation in the original microvascular bifurcation data k, E represents the expectation; G(z) represents the false subordinate vascular bifurcation generated by the generator after the input vector z; D(k) represents the output of the discriminator D after the input k; G(z)′ represents the combination of G(z) and z; D(G(z)′) represents the output of G(z)′ after the discriminator D, ||*||1 is a norm processing, and λ is a hyperparameter. The overall training framework diagram is as follows Figure 5 shown.

[0080] Through the above steps, the discriminator can extract the features of the input data, and the trained generator can generate the next level of microvascular bifurcation that conforms to the distribution of the original data through the previous level of microvascular bifurcation data.

[0081] Further, in step d, a microvascular tree flow chart is generated as follows: Figure 6 As shown, the following steps are included:

[0082] d1 extracts a microvascular bifurcation with the largest average branch diameter from the microvascular tree to be generated as the entrance bifurcation of the vascular tree to be generated, that is, the 0th level bifurcation of the microvascular tree, and takes the entrance bifurcation as the upper-level mother vascular bifurcation, which is input into the trained generator to generate a lower-level daughter vascular bifurcation.

[0083] d2 translates the endpoint of the generated daughter microvascular bifurcation main branch to the node position of the parent vascular bifurcation node of the previous level, and then rotates the daughter vascular bifurcation with the parent vascular bifurcation node position of the previous level as the origin, so that the angle between the main branch and the positive direction of the x-axis is the same as the angle between the growth branch of the parent bifurcation and the positive direction of the x-axis. Then splice the upper and lower bifurcations to grow into a first-level bifurcation tree. The coordinate translation calculation method is as follows:

[0084] newArray[0:3,:]=array[0:3,:]-(endX-nodeX)

[0085] newArray[3:5,:]=array[3:5,:]-(endY-nodeY)

[0086] Where array[0:3,:] represents the x-coordinates of the main branch and sub-branch of the daughter vascular bifurcation before translation, and newArray[0:3,:] represents the x-coordinates of the main branch and sub-branch of the daughter vascular bifurcation after translation; array[3:5,:] represents the y-coordinates of the main branch and sub-branch of the daughter vascular bifurcation before translation, and newArray[3:5,:] represents the y-coordinates of the main branch and sub-branch of the daughter vascular bifurcation after translation, (endX,endY) represents the end point of the main branch of the daughter vascular bifurcation, and (nodeX,nodeY) represents the node of the mother vascular bifurcation;

[0087] The matching microvascular bifurcation is rotated clockwise and the calculation formula is as follows:

[0088] x'=(x-nodeX)cosθ+(y-nodeY)sinθ+nodeX

[0089] y'=(y-nodeY)cosθ-(x-nodeX)sinθ+nodeY

[0090] Among them, x and y are the coordinates of the bifurcation of the sub-branch vessels after translation, x' and y' are the coordinates of the bifurcation of the microvessels after x and y rotate clockwise around (nodeX, nodeY), and θ represents the calculated angle between the growth branch of the superior bifurcation and the positive direction of the x-axis;

[0091] d3 uses this generated lower-level daughter vascular bifurcation as the training sample input of the next round of generator. At this time, the lower-level daughter vascular bifurcation becomes the upper-level mother vascular bifurcation, generating the corresponding next-level daughter vascular bifurcation, that is, the second-level bifurcation.

[0092] At the same time, d4 matches and splices the second-level fork with the first-level fork tree completed by d2 to form a second-level fork tree.

[0093] d5 repeats steps d2-d4 to complete the generation of the next level of microvascular bifurcation and the matching connection of the next level of microvascular bifurcation tree, until the diameter value is less than a certain threshold or the number of recursions is greater than a specified value, the growth of a microvascular tree is completed.

[0094] Through the above steps, a microvascular tree similar to the real microvascular tree in morphology and topology can be generated.

[0095] The beneficial effects of the present invention are mainly manifested in that the conditional generative adversarial network can generate false samples of the next-level bifurcation that conform to the distribution of the original data based on the given previous-level bifurcation, and process the microvascular bifurcation data in the form of a two-dimensional array, which can effectively learn the branch structure with different numbers of nodes, and then through the iterative processing method of generating microvascular bifurcations and splicing microvascular trees, finally generate a vascular tree with extremely small errors in morphology and topological structure with the real microvascular bifurcation tree.

[0096] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for generating microvascular trees based on U-Net conditional generative adversarial network, characterized in that: The following steps are involved: a. extracting microvascular bifurcation pair data from the vascular network to construct a training set, wherein the microvascular bifurcation pair data is a set of maternal and daughter vascular bifurcation data, wherein the maternal vascular bifurcation data includes the maternal main branch node coordinates and diameter information, the maternal left branch node coordinates and diameter information, and the maternal right branch node coordinates and diameter information of the maternal vascular bifurcation, and the daughter vascular bifurcation data is the vascular bifurcation node coordinates and diameter information of a branch in the maternal vascular bifurcation data as the main branch, including the daughter main branch node coordinates and diameter information, the daughter left branch node coordinates and diameter information, and the daughter right branch node coordinates and diameter information; b. Based on a deep convolutional neural network, a U-Net conditional generative adversarial network consisting of a generator and a discriminator is constructed; wherein the generator adopts a U-Net structure, the input is the mother vessel bifurcation data of the bifurcation pair, and the output is the generated daughter vessel bifurcation data; the input of the discriminator is the real microvascular bifurcation pair obtained in step a and the bifurcation pair consisting of the fake daughter vessel bifurcation generated by the generator and the real mother vessel bifurcation, which is used to distinguish whether the input data comes from a real sample or a fake sample generated by the generator; c. Train the generator and discriminator according to the bifurcation of step a, and obtain the trained generator through the game of loss function; d. Obtaining the entry bifurcation of the microvascular tree to be generated, and generating sub-vascular bifurcations step by step according to the trained generator while growing the microvascular bifurcation tree; including the following sub-steps: d1 extracts a microvascular bifurcation with the largest average branch diameter from the microvascular tree to be generated as the entrance bifurcation of the vascular tree to be generated, that is, the 0th level bifurcation of the microvascular tree, and takes the entrance bifurcation as the mother vascular bifurcation and inputs it into the trained generator to generate the next level of daughter vascular bifurcation; d2 translates the end point (endX, endY) of the generated daughter microvascular bifurcation main branch to the position of the parent vascular bifurcation node (nodeX, nodeY) of the previous level, and then rotates the next level of daughter vascular bifurcation with the node position as the origin, so that the angle between the main branch and the positive direction of the x-axis is the same as the angle between the growth branch of the previous bifurcation and the positive direction of the x-axis; d3: The daughter vascular bifurcation is used as the training sample input of the next round of generator. At this time, the daughter vascular bifurcation becomes the mother vascular bifurcation, and the corresponding next-level daughter vascular bifurcation, i.e., the second-level bifurcation, is generated. d4 matches and splices the second-level fork with the first-level fork tree completed by d2 to form a second-level fork tree; d5 repeats steps d2-d4 to complete the generation of microvascular bifurcations and the matching connection of microvascular bifurcation trees step by step, until the diameter value is less than a certain threshold or the number of recursions is greater than a specified value, and the generation of a microvascular tree is completed.

2. The method for generating a microvascular tree according to claim 1, characterized in that: In step a, extracting microvascular bifurcation pair data from the vascular network specifically includes the following sub-steps: a1 separates the arteriolar tree from the original microvascular network, and stores the extracted node coordinates and diameter of each microvascular bifurcation into a two-dimensional array; each line of the two-dimensional array stores the x, y coordinates and diameter information of the main branch, left branch, and right branch nodes of each microvascular bifurcation; a2 normalizes the node coordinate values ​​and pipe diameter values ​​respectively so that their values ​​are between 0 and 1; a3 fills 0 after each row of data that is less than N in the two-dimensional array array, and uniformly forms a 9×N matrix, where N is the maximum number of child nodes contained in the microvascular branch; in the diameter information of the main branch, the left branch, and the right branch, the index of the zero-filling position of the branch is set to 1 as a label; add a row to the array and fill it with 0 at the same time, so that the matrix size becomes 10×N; a4 selects paired mother and daughter vascular bifurcation data from the microvascular bifurcations in the two-dimensional array array to form 20×N microvascular bifurcation pair data as bifurcation pair data.

3. The method for generating a microvascular tree according to claim 1, characterized in that: The step a also includes a data amplification sub-step, specifically: In the range of 0° to 360°, the microvascular bifurcation tree stored in the array array is rotated by increasing α degrees each time to amplify the data. The rotation formula is as follows: x1=x0 cosα-y0 sinα y1=x0 sinα+y0 cosα Among them, x0 and y0 are the node coordinates of the microvascular bifurcation, and x1 and y1 are the node coordinates of the microvascular bifurcation after rotating α degrees clockwise around the origin.

4. The method for generating a microvascular tree according to claim 1, characterized in that: In the step b, the deep convolutional neural network with a U-Net structure adopted by the generator has 10 layers, of which the first 5 layers are convolutional layers and the last 5 layers are deconvolutional layers.

5. The method for generating a microvascular tree according to claim 1, characterized in that: In step c, the loss function is: in: L GAN (G,D)=E k [logD(k)]+E z [log(1-D(G(z)′))] L L1 (G)=E z,k [||k-G(z)′||1] L GAN (G, D) is the loss function of the basic generative adversarial network, L L1 The (G) function is used to stimulate less ambiguity and increase the similarity of true and false bifurcation pairs in the discriminator; z represents the parent vascular bifurcation in the original microvascular bifurcation data k, and E represents expectation; G(z) represents the false subordinate vascular bifurcation generated by the generator after the input vector z; D(k) represents the output of the discriminator D after the input k; G(z)′ represents the merger of G(z) and z; D(G(z)′) represents the output of G(z)′ after the discriminator D, ||*||1 is a norm processing, and λ is a hyperparameter.

Images