An Image and Segmentation Label Generation Model for Tree-Structured Data and Its Application

Abstract

The present invention provides an image and segmentation label generation model for tree-structured data and its application. The model includes a tree-structured image emulator model designed based on a small amount of expert knowledge and a generation network model based on a morphological loss function. The present invention can generate labeled simulated tree-structured images that are extremely similar in style to real target images without any manually labeled information. This technology can generate images and segmentation labels for tree-shaped targets in various medical images (such as: brain neurons, retinal blood vessels, pulmonary trachea, etc.), and the data quality can reach the level of manually labeled data. The present invention is the first model with the ability to automatically generate segmentation-level data, and has the advantages of simple and flexible model configuration, high-quality generated data, and wide application range. It can be used to replace the high-cost and time-consuming process of manually labeled data, and is expected to completely solve the bottleneck of the lack of labeled data in the processing of medical tree-shaped image segmentation by deep learning.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical engineering and relates to an image and segmentation label generation model for tree-structured data and its applications. Without any manually labeled information, it can generate labeled simulation images that are extremely similar to real tree-structured images, which is a highly potential technology to solve the problem of neural networks relying on high-quality labeled data. This technology can be applied to various scenarios such as brain neurons, retinal blood vessels, and pulmonary airways, and has extremely wide practical value. Background Art

[0002] The basic form of a tree structure can be represented as a hierarchical nested structure with multi-level branches and three-dimensional extension morphological characteristics. Starting from a root node, branches grow along the trunk, and after a split at each bifurcation node in each layer, growth continues, thus forming a multi-level tree structure. In biomedical image processing, the segmentation of tree-structured images has a wide range of application scenarios, such as brain neuron reconstruction, retinal blood vessel segmentation, pulmonary airway segmentation, and many other scenarios.

[0003] A data-driven deep neural network is a model that can perform high-precision segmentation on tree-structured images. However, implementing a strongly robust deep network requires a large amount of high-quality expert-labeled data as training data. However, the annotation of such tree-structured images, especially 3D images, is extremely time-consuming and laborious. Different biologists may miss some subtle structures when analyzing branch images, and it is very difficult for annotators to distinguish some areas where the foreground branch signals are submerged in background noise due to the limitations of microscope imaging level.

[0004] In order to reduce the difficulty of obtaining labeled data and thus help such tree-structured images to use the powerful model of deep learning, we expect to use computer simulation to generate labeled data with rich features to replace the cumbersome manual annotation work.

[0005] Computer simulation-based methods include two major categories: model-based methods and simulation-based methods.

[0006] Model-based methods model the entire imaging process. In order to make the simulated images more realistic, many details must be carefully considered, including the length of the branches, the branch radius, the image pixel intensity (or color), the characteristics of imaging, noise, etc., even if some parameters need to be detailedly understood or statistically analyzed from real images. Simulating images in this way is a time-consuming and laborious task, and whenever the conditions of the images change, the parameters must be adjusted by experts.

[0007] Learning-based methods utilize emerging deep learning techniques to generate visually realistic images. Such methods learn the features in real images through generative adversarial networks. The generative adversarial model mainly consists of two deep neural networks, a generator and a discriminator. The input of the generator G is random data z, and its goal is to learn relevant knowledge in real data as much as possible, with the output being G(z). The input of the discriminator D is of two types, the generated image G(z) by the generator and the real image x, and its role is to distinguish the real image from the generated image as much as possible. In each iteration of the model, the generator tries to generate an image similar to the real image as much as possible to deceive the discriminator. And the discriminator tries to distinguish between the real image and the synthetic image as much as possible. By adjusting the image style of the simulated image, the simulated image is made realistic enough. However, the simulated images generated by such methods do not provide pixel or voxel-level labels, resulting in limited application in segmentation tasks. Summary of the Invention

[0008] Aiming at the problem of the lack of high-quality training data for the above-mentioned tree-like image segmentation task, the purpose of the present invention is to provide an image and segmentation label generation model for tree-like structure data, which is an image and segmentation label generation model for tree-like structure. The model includes two modules. The first module is a tree-like structure image emulator model designed based on a small amount of expert knowledge, which is used to generate coarse-grained tree-like structure images. The second module is a generative network model based on a morphological loss function, which is used to learn the style of real tree-like structure images and adjust the details of the simulated images. Through the tree-like structure image generation model of the two modules, a large number of high-quality labeled segmentation-level training data can be automatically generated, completely solving the bottleneck of the lack of training data in deep learning.

[0009] To achieve the above purpose, an image and segmentation label generation model for tree-like structure data designed by the present invention includes the following two modules:

[0010] (1) A tree-like structure image emulator model M designed based on a small amount of expert knowledge γ : The emulator is constructed by inducing the prior knowledge of real images to generate tree-like structure simulated images and their corresponding segmentation labels. The prior knowledge of tree-like images includes features such as the morphology of the tree (branch bifurcation, branch thickness, and branch length), the histogram of pixel or voxel intensity, background noise, and the blurring effect of branches.

[0011] (2) A generative network model R based on a morphological loss function: While ensuring that the underlying segmentation-level labels of the image remain unchanged, it learns the style and morphological features of real images and adjusts the detailed texture and overall pixel or voxel distribution style of the simulated image.

[0012] Preferably, the specific content of constructing the image emulator model M in the step (1) γ is as follows:

[0013] (1-1) Establish a tree structure image simulation model M by summarizing the internal and external features of the tree structure image γ . The internal features refer to the geometric features of the branches in the image, including the tree depth T, branch bifurcation degree B, branch radius R, branch length L, and branch direction θ, these 5 branch geometric features. The external features refer to the texture characteristics of the image, branch strength distribution, and noise distribution.

[0014] (1-2) The tree structure image simulation model M γ The internal feature is used to simulate the shape of the tree. Here, a series of nodes with different radii are used to simulate the typical tree-like data structure:

[0015]

[0016] where, n i is a single node, and its position is p i =(x i , y i , z i ), and the radius is r i . The node n j is the parent node of the node n i , and the node n 0 is the root node. Based on this data structure, for each node n i , it is necessary to specify its node radius R, branch length L, branch angle θ, branch bifurcation degree B, and tree depth T.

[0017] The tree depth T refers to the number of tree bifurcations starting from the root node, and is obtained according to the uniform distribution U(T min , T max ), and its probability density is:

[0018]

[0019] where, T max , T min are the maximum and minimum tree depths respectively.

[0020] The branch bifurcation degree B refers to the number of branch sub-trees at the bifurcation, and is obtained according to the exponential distribution here:

[0021]

[0022] where, α takes 1.

[0023] The node radius R on each branch is obtained according to the uniform distribution U(R min , R max ), and the branch radius decreases at a rate of δ r at each tree depth:

[0024]

[0025]

[0026] Among them, R max , R min are the maximum and minimum branch radii respectively. The branch radius decay rate is δ r . Particularly, the proportion of thinner branches can be appropriately increased to enhance the proportion of branches with weaker signals in the generated data distribution.

[0027] The branch length L is obtained according to the uniform distribution U(L / 2, 3L / 2), and the branch length decreases at the rate of δ l at each tree depth.

[0028]

[0029]

[0030] Among them, L is the average length of the branches obtained according to statistics, and the branch length decay rate is δ l .

[0031] The direction of the sub-branch is determined according to the direction of its parent branch. By rotating successively along the Z-axis and Y-axis, the rotation angles are θ z and θ y . The angle θ y takes a smaller angle and is obtained according to the uniform distribution U(π / 4, π / 2) to make the direction of the sub-branch approximately the same as that of the parent branch:

[0032]

[0033] The angle θ z is obtained according to an arithmetic progression:

[0034]

[0035] Among them, B is the branch bifurcation degree, so that the sub-branches are evenly distributed around the central axis.

[0036] By representing the above tree branch characteristics, the topological morphology of the tree image can be modeled.

[0037] (1 - 3) The external features of the tree structure image simulation model M γ are used to model the texture characteristics, branch strength distribution, and noise distribution of the image. In the tree structure image, the foreground is set as the branch signal, and the background is the noise except the branches. The features considered in the external features include the foreground and background intensity distributions, the blurring effect during the imaging process, and the noise pattern.

[0038] In the image, the intensity M of the pixels or voxels of the branches of the tree structure f and the background part M b respectively follow normal distributions M f ~N(μ f , σ f ) and M b ~N(μ b , σ b ) for simulation:

[0039]

[0040]

[0041] Among them, μ f , σ f are the mean and variance of the foreground, and μ b , σ b are the mean and variance of the background.

[0042] In addition, for the branches of the foreground of the image, Gaussian kernels are used for blurring to simulate the effect of microscope imaging. At the same time, circular spots with random sizes are added at random positions in the background to simulate the common speckle-like noise signals in the image.

[0043] Preferably, the specific content of the generator based on the adversarial network in step (2) is as follows:

[0044] (2-1) The generation network model of the second module based on the morphological loss function uses a generative adversarial network to learn relevant features from complex real tree image data, and adjusts the style of the simulated image generated by the first module model so that it has the characteristics of a real image while maintaining the original underlying segmentation-level labels. In our generation model, the synthetic image z from the simulator M γ is input into the refiner R to generate a refined image R(z). Then, the real image x and the refined image R(z) are fed into the discriminator D, which learns to distinguish between real images and refined images. The discriminator D and the refiner R can be optimized by the following formula

[0045]

[0046]

[0047] Among them, x comes from the real image distribution p real , and z comes from the first-stage simulated image distribution p sim ; among them, the objective function of the adjuster R has two terms. One is the adversarial objective function for the adjuster to learn the style of real images, and the other is to control the change of the segmentation labels of the input images

[0048] (2-2) Through the way of adversarial learning, the mapping relationship between the distribution of simulated images and the distribution of real images can be learned. However, in the process of solving this mapping by the neural network, the problem of unstable training is very likely to occur. Adjusting the network R often overemphasizes some image features to deceive the current discriminator network, resulting in problems such as image drift and all-black images. In addition, when adjusting through the adjuster R, the underlying segmentation labels often undergo large deformations. Specifically, for the branches of a tree, its radius may change, resulting in the branches of the adjusted image becoming thinner or thicker, thus mismatching with its original segmentation label. Therefore, in the training process, external data features of the image are added to control the similarity between the generated image and the overall distribution of the original image:

[0049]

[0050] Among them, z is the simulated image, and R(z) is the image generated using the adjuster R. represents the image similarity loss designed for external image features, is the image shape preservation loss designed for internal image features, and α and β are hyperparameters for balancing each loss function.

[0051] Image similarity loss: It is a commonly used loss function in the current unsupervised generative learning model. Its purpose is to control the stability of the tuning model during training and prevent the model from falling into a single mapping method. We write the data similarity loss term as:

[0052]

[0053] ||·|| 1 is the L1 regularization term.

[0054] The shape preservation loss can keep the basic underlying segmentation label unchanged during the process of the adjuster R performing style transfer on the model, and is defined as:

[0055]

[0056] Among them, z F represents the foreground of the simulated image, and R(z) F represents the foreground of the generated image, that is, the branches in the image. measures the voxel value difference between the foregrounds of the simulated image and the generated image.

[0057] We use the following two formulas to calculate |z F ∪R(z) F | and |z F ∩R(z) F |:

[0058]

[0059]

[0060] Among them,

[0061]

[0062] Among them, K is a relatively large value used to clearly distinguish the foreground and background of the image. μ is the threshold of the foreground and background, which can be automatically statistically obtained according to the pre-defined segmentation labels of the image:

[0063]

[0064] Among them, p b (x) is the image background distribution. When the value of K is large, when and only when z F and R(z) F are both foregrounds, |z F ∩R(z) F | tends to 1, otherwise it tends to 0. Similarly, when z F or R(z) F is the foreground, |z F ∪R(z) F | tends to 1, otherwise it tends to 0. Using this differentiable function, the neural network can be computed during backpropagation. Through this term, we can control the distribution of the foreground and background of the image after passing through the adjuster.

[0065] So far, the construction of the image and segmentation label generation model for tree-structured data is completed.

[0066] The second object of the present invention is to provide the application of the system architecture of the model in various medical image segmentations including tree structures, such as 2D or 3D medical image segmentation tasks such as neuron segmentation, trachea segmentation, and retinal blood vessel segmentation. Through the generation model of the present invention, tree-structured simulation images and their corresponding segmentation-level labels that are extremely similar to the above-mentioned various medical tree-structured images can be automatically generated, which are used to replace the high-cost and time-consuming manual annotation data process. These labeled generated data can be directly used to train the segmentation network of the corresponding image to solve various image segmentation problems.

[0067] For the application of the tree - structured image emulator model of the first module, it is necessary to statistically analyze the required prior knowledge according to the characteristics of different images. Specifically, it includes internal image features (depth of the tree, degree of branch bifurcation, branch radius, branch length, and branch direction) and external image features (intensity distribution of foreground and background, noise characteristics, and blurring effect). The statistical information based on the above - mentioned features serves as the expert knowledge required for constructing the simulation model.

[0068] For the application of the generative network model based on the morphological loss function of the second module, it is necessary to construct an end - to - end generative adversarial network. This network can achieve style transfer of the simulation image by learning the pixel or voxel distribution characteristics of the target in real tree - structured images. In particular, the morphological loss function is used during the network training process to keep the underlying segmentation labels of the image unchanged.

[0069] The present invention is the first model capable of automatically generating tree - structured images and segmentation - level labels, with the advantages of simple and flexible model configuration, high - quality generated data, and wide application scope. It can be used to replace the high - cost and time - consuming process of manually annotating data, and is expected to completely solve the bottleneck of the lack of annotated data in the deep - learning - based medical tree - structured image segmentation problem, showing great potential for practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Schematic diagram of the two - stage tree - structured image generation model architecture.

[0071] Figure 2 Technical flow chart of the two - stage tree - structured image generation model.

[0072] Figure 3 Neuron images generated by the emulator in the first stage and corresponding labels.

[0073] Figure 4 Network architecture of the generative model in the second stage.

[0074] Figure 5 Ablation experiment of the neuron tree - structured image generation model and comparison with real neuron images. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0075] To better understand the embodiments of the present invention, the following will be further described in conjunction with the drawings and specific embodiments.

[0076] Example 1

[0077] In this example, mouse brain neuron images are selected as an example of the application scenario. Neuron morphology reconstruction is considered one of the core steps in understanding the working principle of the biological brain and revealing the mechanism of life operation. An important subtask of the neuron reconstruction algorithm is to segment the three-dimensional neuron image, separating the foreground of the neuron with high intensity values from the low-intensity noise background in the image.

[0078] However, the annotation of such three-dimensional neuron images is extremely time-consuming and laborious. Different neurologists may miss some subtle branch structures when analyzing the images; for some areas where the neuron signals in the foreground are submerged in the background noise due to the limitations of optical microscope imaging level, it is very difficult for annotators to distinguish them.

[0079] To reduce the difficulty of obtaining annotated data, thereby helping the neuron reconstruction task to use the powerful model of deep learning, an image simulation method based on a two-stage generation model proposed by the present invention is used to automatically generate image data with high-quality segmentation-level labels, solving the bottleneck of the lack of training data in the neuron reconstruction task by deep learning. The model architecture diagram is as Figure 1 shown.

[0080] The flow of the dendritic structure data generation method provided by the present invention is as follows, as Figure 2 shown:

[0081] Step A: Construct a neuron dendritic image simulator. Based on the prior knowledge of neuron morphology, such as the radius of the neuron branches, the length of the branches, the angle of the branches, the degree of bifurcation, and the depth of the neuron tree, an initial neuron image simulator is designed. The user needs to count and input the above prior knowledge according to the real neuron images. When constructing the dendritic structure image simulator model, the specific parameter settings are as follows: the tree depth T min and T max are set to 5 and 8; the node radii R min and R max are set to 2 and 4; the branch radius decay rate δ r is set to 0.9; the branch length L is set to 50; the branch length decay rate δ l is set to 0.9; when generating an 8-bit image, the mean μ f and variance σ f of the foreground are set to 200 and 10, and the mean μ b and variance σ b of the background are set to 20 and 5; when generating a 16-bit image, the mean μ f and variance σ f of the foreground are set to 1000 and 50, and the mean μ b and variance σ bSet them to 20 and 5. In addition, the neuron image simulator supports multi-process parallel computing and can quickly generate a large amount of simulation data, including simulated neuron images (8-bit or 16-bit) and their corresponding segmentation-level labels (0-1), such as Figure 3 as shown.

[0082] Step B: Generate the training data segmentation of the network. Since the size of 3D neuron images is large, directly inputting them into the deep network will result in a huge and difficult-to-train model. Therefore, this model adopts the form of cutting neurons into small pieces, with each small piece sized 64*64*32. The specific implementation method is to use the form of a sliding window to slide and select positions on the neuron image. When the area with a large branch intensity is included in the window, the image block is cut into a training image.

[0083] Step C: Construct a generative adversarial network based on the morphological loss function. The second stage adopts a model based on the generative adversarial network, including an adjuster R and a discriminator D. Input the image z in the first stage into the adjuster R containing n residual modules and output the adjusted image R(z). Then input the generated image R(z) and the real image into the discriminator network D. The discriminator will try its best to distinguish the generated image from the real image and backpropagate the gradient to the adjuster R, so that in the next generation, R will generate an image as similar to the real image as possible, making it impossible for the discriminator to distinguish. The network architecture of the adjuster R is as Figure 4 shown. Through the form of confrontation between the two networks, the styles of the generated image and the real image become close. After the model loss value tends to be stable, the training can be stopped. Use the trained adjuster R, input the simulation image in the first stage, and output the final generated image.

[0084] To verify the reliability of the generation model of the present invention, analyze the quality of the data generated by the model. Here, show the effect of the ablation experiment of the two-stage model in generating data and verify the effectiveness of each stage of the model.

[0085] For the ablation experiment of the two-stage model, it includes 4 control groups:

[0086] Experiment 1: The simulator of Module 1 (internal features), only generates neuron images with a dendritic topology structure;

[0087] Experiment 2: The simulator of Module 1 (internal features) + the generation network of Module 2. Based on Experiment 1, first generate neuron images with a dendritic topology structure, and then use the generation network to adjust the image details;

[0088] Experiment 3: The simulator of Module 1 (internal features + external features), generates neuron images with a dendritic topology structure and adds external feature rendering of the foreground, background, and noise to the images;

[0089] Experiment 4: The emulator of Module 1 (internal features + external features) + the generation network of Module 2. Based on Experiment 3, generate a simulated image with the internal and external features of neuron images, and then use the generation network to adjust the image details.

[0090] The results of the ablation experiment are as Figure 5 shown. In Experiment 1, a neuron image emulator that only combines the internal features of the image was used to generate an image with the morphological features of neurons, but both the foreground and background of the image were constant values. In Experiment 2, based on Experiment 1, by combining the second-stage generation network model, it can be observed that the branches produced an obvious blurring effect, but its foreground and background did not have the characteristics of real neuron images. For Experiment 3, a neuron image emulator that combines the internal and external features of the image was used, which had the basic morphology of real neurons and the basic pixel or voxel intensity features, but its texture details were still not realistic enough. The image style of the model in Experiment 4 was the most similar to the real image.

[0091] In addition, the dendritic structure image emulator in the first stage can be designed separately for various types of images. The present invention can be conveniently and quickly applied to the generation tasks of labeled data for 2D or 3D medical images, including neuron image reconstruction, trachea segmentation, retinal vessel segmentation, etc. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall fall within the protection scope of the present invention.

Claims

1. A two-stage image and segmentation label generation model for tree-structured data, characterized in that, the model comprises two modules. The first module is a tree-structured image emulator model designed based on a small amount of expert knowledge, which is used to generate a coarse-grained tree-structured image. The second module is a generation network model based on a morphological loss function, which is used to learn the style of real tree-structured images and adjust the details of the simulated images; (1) The tree - structured image simulator model M designed based on a small amount of expert knowledge γ : The simulator is constructed by inducing the prior knowledge of real images, generating tree - structured simulated images and their corresponding segmentation labels. The prior knowledge of the tree - structured images includes the shape of the tree, the histogram of pixel or voxel intensity, background noise, and the blurring effect characteristics of the branches. (2) The generation network model R based on the morphological loss function: Without changing the underlying segmentation-level labels of the image, it learns the style and morphological features of the real image and adjusts the detailed texture and overall pixel or voxel distribution style of the simulated image.

2. The image and segmentation label generation model according to claim 1, characterized in that, In the first module, the tree-structured image emulator model designed based on a small amount of expert knowledge divides the features of the tree-structured image into internal features and external features. The internal features focus on the geometric image features of the branches in the image, including the depth of the tree, the branch bifurcation degree, the branch radius, the branch length, and the branch direction. The external features refer to the texture characteristics of the image, the branch strength distribution, and the noise distribution.

3. The image and segmentation label generation model according to claim 2, characterized in that, The internal features of the tree-structured image are used to simulate the shape of the branches. The basic morphology of the neurons is represented by a tree structure, and its basic morphology is represented as a hierarchical nested structure with multi-level branches and three-dimensional extension morphological features. Starting from a root node, branches are formed along the growth of the branches, and a split is performed at each bifurcation node of each layer and then continues to grow, thus forming a multi-level tree structure. The beginning of the structure is called the root node, the nodes other than the root node are called child nodes, and the nodes that are not connected to other child nodes are called leaf nodes. A series of nodes with different radii are used to simulate a typical tree structure: where n i is a single node, whose position is p i =(x i , y i , z i ), with a radius of r i , node n j is the parent node of node n i , node n 0 is the root node. On this basis, the simulation model M γ also requires the following parameters to generate the simulation tree structure image, including the node radius R, branch length L, branch angle, branch bifurcation degree B, and tree depth T, which are set according to the results of the relevant work counted according to the form of the branches in the image; The depth T of the tree refers to the number of bifurcations of the tree starting from the root node, obtained according to the uniform distribution U(T min ,T max ), and its probability density is: Among them, T max , T min are the maximum and minimum tree depths respectively; The branch bifurcation degree B refers to the bifurcated sub-branch trees at the bifurcation, and is obtained according to the exponential distribution here: where α takes 1; The node radius R on each branch is obtained according to a uniform distribution U(R min ,R max ), and the branch radius decreases at a rate of δ r at each depth of the tree layer: wherein, R max , R min are the maximum and minimum branch radii respectively, and the branch radius decay rate is δ r ; The branch length L is obtained according to a uniform distribution U(L / 2, 3L / 2), and the trunk length decreases at a rate of δ l at each depth of the tree layer: Among them, L is the average length of the branches and trunks obtained according to statistics, and the attenuation rate of the branch and trunk length is δ l ; The direction of the sub-branch is determined according to the direction of its parent branch, by rotating successively along the Z-axis and the Y-axis, with the rotation angles being θ z and θ y , where the angle θ y takes a smaller angle, obtained according to the uniform distribution U(π / 4, π / 2), to make the direction of the sub-branch roughly the same as that of the parent branch: Angle θ z Obtained according to an arithmetic progression: where B is the branch bifurcation degree, so that the sub-branches are evenly distributed around the central axis.

4. The image and segmentation label generation model according to claim 2, characterized in that, For the external features of the tree-structured image, in the tree-structured image, the foreground is set as the branch signal, and the background is the noise except the branches. The external features include the foreground and background intensity distributions, the blurring effect during the imaging process, and the noise pattern; In the image, the intensity M of the pixels or voxels of the branches of the tree-like structure f and the background part M b are respectively simulated using normal distributions M f ~N(μ f ,σ f ) and M b ~N(μ b ,σ b ): where μ f , σ f are the mean and variance of the foreground, and μ b , σ b are the mean and variance of the background; For the branches in the foreground of the image, Gaussian kernels are used for blurring to simulate the effect of microscope imaging. At the same time, circular spots with random sizes are added at random positions in the background to simulate the common speckle-like noise signals in the image.

5. The image and segmentation label generation model according to claim 1, characterized in that, In the second module, while ensuring that the underlying labels of the images remain unchanged, the underlying style of the real tree-like images is learned to adjust the simulated images. The synthetic image z from the simulator M γ is input into the refiner R to generate the refined image R(z). Then, the real image x and the refined image R(z) are fed into the discriminator D, which learns to distinguish between the real image and the refined image. The discriminator D and the refiner R can be optimized by the following formulas: where x is from the real image distribution p real , z is from the first-stage simulated image distribution p sim , where the objective function of the regulator R has two terms. One is the adversarial objective function for the regulator to learn the style of real images, and the other is to control the change of the segmentation label of the input image 6. The image and segmentation label generation model according to claim 5, characterized in that, Objective function for controlling the change of segmentation labels of the input image in the adjuster R Control the similarity between the generated image and the overall distribution of the original image: where z is the simulated image and R(z) is the image generated using the adjuster R, represents the image similarity loss designed for the external features of the image, is the image shape preservation loss designed for the internal features of the image, and α and β are hyperparameters for balancing each loss function.

7. The image and segmentation label generation model according to claim 6, characterized in that, The image similarity loss is a relatively commonly used loss function in unsupervised generative learning models. Its purpose is to control the stability of the tuning model during training and prevent the model from falling into a single mapping method. The data similarity loss term is written as: ‖·‖ 1 is the L1 regularization term; The shape preservation loss keeps the underlying segmentation labels unchanged during the style transfer of the model by the optimizer R, and is defined as: Among them, z F represents the foreground of the simulation image, and R(z) F represents the foreground of the generated image. To measure the difference in pixel or voxel values between the foregrounds of the simulation image and the generated image, the following two expressions are used to calculate |z F ∪R(z) F | and |z F ∩R(z) F |: Among them, Among them, K is a relatively large number used to clearly distinguish the foreground and background of the image. μ is the threshold of the foreground and background, which is automatically obtained by statistically counting according to the pre-defined segmentation labels of the image: where p b (x) is the image background distribution. When the value of K is large, and when both z F and R(z) F are foregrounds, |z F ∩R(z) F | tends to 1, otherwise it tends to 0. Similarly, when either z F or R(z) F is a foreground, |z F ∪R(z) F | tends to 1, otherwise it tends to 0.

Images