A multi-stage generative adversarial network supporting generation of pairs of medical segmentation data

Abstract

The application provides a multi-stage generative adversarial network supporting generation of medical segmentation data pairs, and belongs to the field of medical image generation.The method comprises the following steps: decoupling medical images in a segmentation data set; dividing the generation of the medical images into three stages; obtaining a medical image without a lesion by using human body structure symmetry; and generating diverse realistic data pairs by using the multi-stage generative adversarial network.The image decoupling refers to dividing the medical images into lesion and non-lesion regions by using a segmentation mask; the three stages of image generation refer to non-lesion region generation, lesion region shape generation, and non-lesion and lesion region texture enhancement.In order to alleviate the problem of a lack of medical image segmentation data sets, the medical image generation process is decoupled, a multi-stage generative adversarial network capable of simultaneously generating medical images and corresponding segmentation masks is constructed, and by generating different parts of a complex medical image in stages, the quality of the generated image is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image generation, and particularly relates to a multi-stage generative adversarial network supporting generation of medical segmentation data pairs. BACKGROUND

[0002] Medical images play an important role in disease diagnosis and treatment, especially magnetic resonance images (MRI) and computed tomography images (CT) have shown their effectiveness in many tumor diagnosis and treatment, especially tumor segmentation. For example, MRI is often used in brain tumor diagnosis because it can display high-resolution images, while CT is often used in imaging of some organs with simple structures such as lungs and kidneys because of its lower equipment loss. However, in the case of uneven distribution of medical resources and continuous growth of medical image data, traditional manual labeling gradually exposes the problems of long labeling period and uneven labeling accuracy, which also makes automatic labeling a focus of attention. Because of the single style and fuzzy boundary of medical images, traditional semantic segmentation methods do not work well on medical image segmentation and need human assistance, and cannot achieve fully automated labeling. In order to reduce the burden of doctors and improve the accuracy of segmentation, computer-aided diagnosis and treatment technology based on deep learning has been widely used in recent years. However, the progress of these methods depends heavily on a large number of pixel-by-pixel labeled data sets, which are usually expensive and labor-intensive. To solve this problem, two methods are mainly used. One is the weakly supervised method, which uses weak annotations such as bounding boxes to train CNN-based models, which is relatively low-cost and fast in training. The other is the generative method based on generative adversarial network (GAN), which generates a large number of fake data pairs and uses the generated data to build a large data set and then applies it to supervised segmentation methods. Because these data pairs can build a huge segmentation data set, training supervised segmentation networks using the newly built large data set usually achieves better performance than weakly supervised methods. Due to the special adversarial training method of GAN, it has excellent performance in data fitting, and GAN has therefore been widely used in image generation tasks. Because of the high time cost of GAN to generate 3D medical images, most GAN-based medical generation methods focus on 2D medical images. SUMMARY

[0003] The present application aims to provide a multi-stage generative adversarial network supporting generation of medical segmentation data pairs, characterized by comprising the following steps:

[0004] S1: decoupling the medical images in the segmentation data set;

[0005] S2: The decoupled medical image is divided into three stages, which are non-lesion region generation, lesion region shape generation, and non-lesion region and lesion region texture enhancement;

[0006] S3: Obtain a medical image without a lesion by using the symmetry of the human structure;

[0007] S4: Generate a variety of realistic data pairs using a multi-stage generative adversarial network.

[0008] Further, in S1, decoupling is: using a separation mask to divide the medical image into a lesion region and a non-lesion region, the lesion region is obtained by multiplying the medical image and the corresponding separation mask; the non-lesion region is obtained by inversely multiplying the medical image and the corresponding separation mask.

[0009] Further, S3 is specifically: by the symmetry of the human structure, it is obtained that the tissue organs in the medical image have symmetry, based on the symmetry of the tissue organs, the lesion region pixels are replaced with non-lesion pixels of symmetric coordinates, and the adjacent pixels of the lesion region boundary pixels are replaced, to obtain a medical image without a lesion.

[0010] Further, in S4, while generating a variety of realistic data pairs using a multi-stage generative adversarial network, a medical image and a corresponding segmentation mask are also generated, which specifically includes the following steps:

[0011] S41: The obtained medical image without a lesion, the corresponding segmentation mask, and the medical image form a set of training data pairs;

[0012] S42: Construct a multi-stage generative adversarial network;

[0013] S43: Train the multi-stage generative adversarial network;

[0014] S44: Generate a medical image and a corresponding segmentation mask.

[0015] Further, in S42, constructing a multi-stage generative adversarial network includes the following steps:

[0016] S421: According to the different generation targets, the generative adversarial network structure is divided into a background stage, a mask stage, and an enhancement stage;

[0017] S422: construct a three-stage generative adversarial network, the three-stage adversarial network includes three VAEGAN architectures and a BiDiscriminator discriminator, the VAEGAN architecture includes an encoder, a decoder and a discriminator, the encoder maps the input image to the latent vector space, the decoder reconstructs the latent vector into an image, and the decoder and the discriminator jointly constrain the generation process of the VAEGAN architecture; the BiDiscriminator discriminator improves the correlation between the image space and the latent vector space by discriminating the array including the latent vector and the real image and the array including the real vector and the generated image;

[0018] S423: generate a lesion-free image through a background stage, generate a segmentation mask through a mask stage, and generate enhanced textures of the lesion area and the non-lesion area through an enhancement stage, all of which are constrained by the encoder and the decoder and the discriminator in the multi-stage generative adversarial network;

[0019] S424: verify the necessity of three-stage generation by designing an ablation experiment.

[0020] Further, in S43, the training of the stage-wise generative adversarial network specifically includes the following steps:

[0021] S431: synthesize a lesion-free image and corresponding segmentation mask and medical image as input data and generation target, i.e., input sample;

[0022] S432: train the stage-wise generative adversarial network using the input sample, and set the training times, which satisfy that the model tends to be stable after the stage-wise generative adversarial network is trained for the training times;

[0023] S433: save the final model.

[0024] Further, S44 specifically includes the following steps:

[0025] S441: randomly generate three vectors and a noise vector in the latent vector space;

[0026] S442: load three generators of the trained stage-wise generative adversarial network;

[0027] S443: connect any two vectors in the latent vector space with the noise vector to obtain two synthetic vectors, the two synthetic vectors being a first synthetic vector and a second synthetic vector; and the remaining vector in the latent vector space is taken as a third synthetic vector;

[0028] S444: input the first synthetic vector into the generator in the background stage to obtain a lesion-free image; input the second synthetic vector into the generator in the mask stage to obtain a mask image, and multiply the mask image by the lesion-free image to obtain an image containing only non-lesion areas; input the third synthetic vector into the generator in the enhancement stage to obtain an enhanced image, and superimpose the enhanced image on the lesion-free image to obtain a realistic medical image.

[0029] Further, in S424, designing the ablation experiment includes the following steps:

[0030] S4241: design a two-stage generation model, including the following steps:

[0031] S42411: extract the first two stages of the three-stage generative adversarial network;

[0032] S42412: take the medical image as the input data and generation target of the first stage, and take the segmentation mask as the input data and generation target of the second stage;

[0033] S41423: take the generated data of the first stage as the medical image, and take the generated data of the second stage as the corresponding segmentation mask to form a medical image segmentation data pair;

[0034] S4242: design a four-stage generation model, including the following steps:

[0035] S42421: add a structure identical to the third stage after the three-stage generative adversarial network as the fourth stage;

[0036] S42422: take the synthetic lesion-free image as the input data and generation target of the first stage, take the segmentation mask as the input data of the second stage, and take the medical image as the input data of the third and fourth stages;

[0037] S42423: take the generated data of the fourth stage as the medical image, and take the generated data of the second stage as the corresponding segmentation mask to form a medical image segmentation data pair;

[0038] S4243: train the two-stage generation model, the three-stage generation model and the four-stage generation model to obtain generation results, respectively, and complete the verification of the necessity of the three-stage generation by comparing the three generation results.

[0039] A terminal device for running a multi-stage generative adversarial network supporting generation of medical segmentation data pairs, characterized by comprising a storage medium, a memory and a processor, the storage medium being a computer readable storage medium for storing at least one program, the memory being used to load the storage medium and support various parameters and instructions of the multi-stage generative adversarial network supporting generation of medical segmentation data pairs, and the processor loading the various parameters and executing the instructions.

[0040] Compared with the prior art, the beneficial effects of the present application mainly include:

[0041] 1、The present application expands the medical image data set by constructing a multi-stage generative adversarial network supporting generation of medical segmentation data pairs, alleviates the problem of lack of labeled data for medical image segmentation, obtains optimal medical image generation quality as a data generation model, and achieves effective data enhancement capability in medical image segmentation as a data expansion method.

[0042] 2、Based on the present application, doctors can make more accurate diagnosis of tumor location and characteristics based on the segmentation data, effectively solving the problems of long labeling period and uneven labeling accuracy of artificial labeling. The present application embodiment can generate medical segmentation data pairs with optimal quality on a computer terminal, effectively alleviating the problem of lack of medical image labeling. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a medical image decoupling method of the present application embodiment;

[0044] Figure 2 is a non-lesion medical image synthesis method of the present application embodiment;

[0045] Figure 3 is a two-dimensional brain magnetic resonance image schematic diagram of a certain acoustic neuroma patient;

[0046] Figure 4 is Figure 3 a corresponding real annotation map;

[0047] Figure 5 is a multi-stage generative adversarial network structure diagram of the present application embodiment;

[0048] Figure 6 is an explanatory diagram of using the multi-stage generative adversarial network to generate medical image segmentation data pairs;

[0049] Figure 7 is an ablation experiment comparison diagram of the necessity of using a private acoustic neuroma data set to prove the necessity of a three-stage generative network of the present application embodiment;

[0050] Figure 8An ablation experiment is performed to demonstrate the necessity of the three-stage generative network using a public kidney dataset Kits19 for the exemplary use of the present application. DETAILED DESCRIPTION

[0051] The multi-stage generative adversarial network for generating pairs of medical segmentation data will be described in more detail below with reference to the accompanying drawings, in which a preferred embodiment of the present application is shown, it being understood that a person skilled in the art can modify the present application described herein while still achieving the advantageous effects of the present application, therefore, the following description should be understood as a broad knowledge for a person skilled in the art, and not as a limitation of the present application.

[0052] A multi-stage generative adversarial network for generating pairs of medical segmentation data comprises the following steps:

[0053] Step 1: decoupling medical images in the segmentation dataset

[0054] A medical image segmentation dataset is usually composed of Figure 3 medical images and Figure 4 segmentation masks.

[0055] Decoupling of images is to divide the image into lesion and non-lesion regions using the segmentation mask.

[0056] Since the textures of the foreground (lesion region) and the background (non-lesion region) of the medical image are too similar, a single GAN cannot generate a medical image with good quality, therefore, the medical image needs to be decoupled. The decoupling method of the medical image is as shown in Figure 1 , multiplying the medical image with the corresponding segmentation mask to obtain the lesion region, and multiplying the medical image with the inverse of the corresponding segmentation mask to obtain the non-lesion region.

[0057] Step 2: dividing the medical image into three stages

[0058] According to the decoupling method of the medical image, the generation process of the medical image can be divided into three stages, i.e. non-lesion region generation, lesion region shape generation, and non-lesion and lesion region texture enhancement.

[0059] Step 3: obtaining a medical image without lesions using the symmetry of human body structure

[0060] As shown in Figure 2 , according to the symmetry of the human body structure, most of the tissues and organs in the medical image have symmetry, the lesion region pixels are replaced with non-lesion pixels at the symmetric coordinates, and the adjacent 8 pixels of the lesion region boundary pixels are replaced, which are within a 3 by 3 range centered on the pixel.

[0061] Step 4: Use a multi-stage generative adversarial network to generate diverse and realistic data pairs.

[0062] Specifically: using, as Figure 5 The multi-stage generative adversarial network shown generates medical images and corresponding segmentation masks simultaneously.

[0063] The process of generating medical images and corresponding segmentation masks simultaneously using a multi-stage generative adversarial network includes the following steps:

[0064] Step 4.1: Construct the required training data for multi-stage generation using the segmented dataset.

[0065] First, lesion-free images are synthesized using the above method by taking advantage of the symmetry of human body structure. Then, the corresponding lesion-free images, segmentation masks, and medical images are combined to form a set of training data pairs, which are used as inputs for the three generation stages.

[0066] Step 4.2: Construct a multi-stage generative adversarial network

[0067] The construction of a multi-stage adversarial generative network specifically includes the following steps:

[0068] Step 4.2.1: Based on the different generation targets, the structure of the generative adversarial network is divided into a background stage, a masking stage, and an enhancement stage to construct a three-stage generative adversarial network;

[0069] Specifically, the three-stage adversarial network includes three VAEGAN architectures and a BiDiscriminator. The VAEGAN architecture includes an encoder, a decoder, and a discriminator. The encoder maps the input image to the latent vector space, the decoder reconstructs the image from the latent vectors, and the decoder and discriminator jointly constrain the generation process of the VAEGAN architecture. The BiDiscriminator improves the correlation between the image space and the latent vector space by distinguishing between an array containing latent vectors and real images and an array containing real vectors and generated images.

[0070] Step 4.2.2: The background stage generates a lesion-free image, the mask stage generates a segmentation mask, and the enhancement region generates enhanced textures for both lesion and non-lesion regions. All generation processes are constrained by the encoder, decoder, and discriminator in the multi-stage generative adversarial network.

[0071] Specifically, considering that medical images and their corresponding segmentation masks can be clustered into various latent low-dimensional vectors, VAEGAN is used to generate images at each stage. An additional discriminator is added to the three VAEGAN stages to constrain the relationship between the images and the latent low-dimensional vectors. Since the VAEGAN encoder can map high-dimensional inputs to low-dimensional vectors, the encoders in the three stages (background stage, mask stage, and enhancement stage) extract background, segmentation mask, and texture information from the image and construct latent one-hot vectors. This forces the generators in the three stages to generate images similar to the encoder input. Simultaneously, under the supervision of the discriminator in each VAEGAN, the generator tends to generate more realistic images, effectively improving the generation quality. Furthermore, to increase the diversity of generated data, an additional discriminator is added to constrain the correlation between the images and the code. In this way, this multi-stage generative adversarial network can generate various structurally sound medical segmentation data pairs.

[0072] like Figure 5As shown, the network comprises three VAEGANs and one BiDiscriminator. The specific architecture of the VAEGANs includes an encoder, a decoder, and a discriminator. The encoder maps the input image to a latent vector space, the decoder reconstructs the latent vectors into an image, and the discriminator and encoder / decoder jointly constrain the VAEGAN generation process. The BiDiscriminator, acting as a discriminator, improves the correlation between the image space and the latent vector space by discriminating between an array containing latent vectors and real images and an array containing real vectors and generated images. Four encoders map the input images to latent vectors, and three generators reconstruct them into realistic images before outputting them. The performance of both the generators and encoders is constrained by three ordinary supervised discriminators and an additional bidiscriminator, which not only improves the generation quality but also enhances the correlation between the image and the latent vectors. Specifically, the input images for the three stages are a synthesized lesion-free image (Untumor), a segmentation mask from the dataset (Mask), and a medical image from the dataset (Origin). e_x represents the latent vector encoded by the encoder E_x. b, m, a represent randomly generated one-hot vectors, and vector z is random noise following a normal distribution. Since one-hot vectors can indicate categories, the dimension of the latent vector is a hyperparameter dependent on the diversity of the dataset. Specifically, the dimensions of e_b / b and e_a / a should be similar to the number of training data pairs, while the dimension of e_m / m is one-quarter of theirs due to the consistency of lesion shape and location. Furthermore, the output images G_untumor, G_mask, and A of the three generators are the generated lesion-free image, the generated mask, and the enhanced image, respectively. The generation result G_unmask in the masking stage is the product of the inverses of G_untumor and G_mask. Although multiple VAEGANs should ideally generate tumor-free, segmented, masked, and enhanced images, the complex texture of medical images means that the constraints of a single ordinary discriminator cannot extract the required information into the one-hot vector space and reconstruct it correctly. A discriminator in each stage cannot drive the encoder and generator to converge. To address this issue, a Bi-discriminator is introduced. This discriminator improves the correlation between the image and the latent vector by judging the authenticity of different datasets. Specifically, the discriminator considers the set consisting of the generated image and the real vector to be fake, and the set consisting of the input image and the generated vector to be real.

[0073] Step 4.2.3: Design ablation experiments to verify the necessity of the three-stage generation.

[0074] The design of an ablation experiment includes the following steps:

[0075] Step 4.2.3.1: Design a two-stage generative model

[0076] Specifically, the process involves extracting the first two stages of a three-stage generative adversarial network (GAN); using the medical image as the input data and generation target for the first stage, and the segmentation mask as the input data and generation target for the second stage; using the generated data from the first stage as the medical image, and the generated data from the second stage as the corresponding segmentation mask, forming a medical image segmentation data pair. Two VAEGANs are used to generate the original medical image and the mask image. The input and generation target for the first stage is the original medical image. The mask is set as the input and training target for the second stage. To ensure the generated medical image corresponds to the mask, a hidden code f is set in the first stage and mapped to the hidden code s in the second stage. The network structure is the same as the background and mask stages of DPGAN.

[0077] Step 4.2.3.2: Design a four-stage generative model

[0078] Specifically, a fourth stage is added after the three-stage generative adversarial network (GAN), identical to the third stage. The synthesized lesion-free image is used as the input and generation target of the first stage, the segmentation mask as the input of the second stage, and the medical image as the input of both the third and fourth stages. The generated data from the fourth stage is used as the medical image, and the generated data from the second stage is used as the corresponding segmentation mask, forming a medical image segmentation data pair. Four VAEGANs are used to generate the medical data pairs. VAEGANs are then appended to the three-stage GAN as the fourth stage, and the original medical image is set as the input and generation target for this stage. The latent vector for this stage should be the same as 'a' in the third stage.

[0079] Step 4.2.3.3: Compare the two-stage generative model, the four-stage generative model, and the proposed three-stage generative adversarial model, including: training the two-stage, three-stage, and four-stage generative models; and comparing the generation results.

[0080] Specifically, data pairs of the same scale are generated using the same data. Samples are then taken from the generated data and analyzed. For example... Figure 7 , Figure 8As shown, when trained on the private acoustic neuroma dataset and the public Kits19 dataset, the two-stage generative model struggles to generate complex medical image textures in a single stage (the first stage). Furthermore, the two-stage generative model may fail to generate masks correctly when the foreground and background are too similar. Since the proposed three-stage generative adversarial model can generate not only lesion regions but also complex non-lesion regions, it is effective in medical image generation. While the four-stage generative model outperforms the three-stage model on quantitative metrics, its model complexity is unacceptable (in our experiments, training the four-stage generative model on the Kits19 dataset required at least 10GB of GPU memory).

[0081] Step 4.3: Train a multi-stage generative adversarial network, including the following steps:

[0082] Step 4.3.1: Use the synthesized lesion-free image, the corresponding segmentation mask, and the medical image as input data and generation target to form the training samples of DPGAN.

[0083] Step 4.3.2: Use the above samples for training and set the number of iterations. After enough iterations, the model will stabilize and the final model will be saved.

[0084] Step 4.4: Generate the medical image and the corresponding segmentation mask, including the following steps:

[0085] Step 4.4.1: Randomly generate three vectors and one noise vector in the latent vector space.

[0086] Step 4.4.2: Load the three generator synthesis input vectors of the trained multi-stage generative adversarial network. Specifically, the vectors in the two latent vector spaces are concatenated with the noise vector to obtain the input vectors of the generators in the background stage and the masking stage. The vector in the third latent vector space is used as the input vector of the generator in the enhancement stage.

[0087] Step 4.4.3: Input the three synthesized vectors into three generators respectively to obtain the generated medical image and the corresponding mask. This includes the following steps:

[0088] Step 4.4.3.1: Input the first synthesized vector into the generator in the background stage to obtain a lesion-free image.

[0089] Step 4.4.3.2: Input the second synthesized vector into the generator in the masking stage to obtain the mask image. Multiply the mask image with the lesion-free image to obtain an image containing only non-lesion regions.

[0090] Step 4.4.3.3: Input the third latent spatial vector into the generator of the enhancement stage to obtain the enhanced image. Superimpose the enhanced image with the non-lesion area image synthesized in the background stage to obtain a realistic medical image.

[0091] like Figure 6 As shown, after training, the three-stage generator will be able to convert vectors into real original medical images G_origin and corresponding masks G_mask.

[0092] The data pair generation process is as follows:

[0093] (1) Randomly generate one-hot vectors b, m, a and normally distributed noise z.

[0094] (2) Set the cascaded b and z as the input of G_b, set the cascaded m and z as the input of G_m, and set a as the input of G_a.

[0095] (3) Execute the generation and save the generated original image G_origin and mask G_mask as segmentation data pairs.

[0096] On the other hand, the technical solution adopted in the embodiments of the present invention is: a multi-stage generative adversarial network that supports the generation of medical segmentation data pairs, the use of which includes the following implementation conditions:

[0097] A computer-readable storage medium for storing at least one program.

[0098] A terminal device includes at least one processor and a computer-readable memory, the memory loading a computer-readable storage medium and supporting parameters and instructions for a multi-stage generative adversarial network that generates medical segmentation data pairs, the processor loading and executing these instructions.

[0099] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A multi-stage generative adversarial network supporting the generation of medical segmentation data pairs, characterized in that, Includes the following steps: S1: Decouple medical images within the segmented dataset; S2: The decoupled medical image is divided into three stages: non-lesion region generation, lesion region shape generation, and enhancement of texture between non-lesion and lesion regions. S3: Utilizing the symmetry of human body structure to obtain lesion-free medical images; S4: Use a multi-stage generative adversarial network to generate diverse and realistic data pairs; In step S4, a multi-stage generative adversarial network is used to generate diverse and realistic data pairs, including medical images and corresponding segmentation masks, specifically including the following steps: S41: The obtained lesion-free medical images are combined with the corresponding segmentation masks and medical images to form a set of training data pairs; S42: Constructing a multi-stage generative adversarial network; S43: Train the staged generative adversarial network; S44: Generate medical images and corresponding segmentation masks; In step S42, constructing the multi-stage generative adversarial network includes the following steps: S421: Based on the different generation targets, the structure of generative adversarial networks is divided into background stage, masking stage and enhancement stage; S422: Construct a three-stage generative adversarial network, which includes three VAEGAN architectures and a BiDiscriminator. The VAEGAN architecture includes an encoder, a decoder, and a discriminator. The encoder maps the input image to the latent vector space, and the decoder reconstructs the latent vectors into an image. The decoder and discriminator jointly constrain the generation process of the VAEGAN architecture. The BiDiscriminator improves the correlation between the image space and the latent vector space by distinguishing between an array containing latent vectors and real images and an array containing real vectors and generated images. S423: Generate a lesion-free image through the background stage, generate a segmentation mask through the masking stage, and generate enhanced textures for lesion and non-lesion regions through the enhancement stage. All generation processes are constrained by the encoder, decoder and discriminator in the multi-stage generative adversarial network. S424: Verify the necessity of the three-stage generation by designing ablation experiments; In S43, training a phased generative adversarial network specifically includes the following steps: S431: The synthesized lesion-free image, the corresponding segmentation mask, and the medical image are used as input data and generation targets, i.e., input samples; S432: Use the input samples to train a staged generative adversarial network, and set the number of training iterations, such that the model tends to stabilize after the staged generative adversarial network has been trained that many times. S433: Save the final model; S44 specifically includes the following steps: S441: Randomly generate three vectors and one noise vector in the latent vector space; S442: Load the three generators of the staged generative adversarial network that have been trained; S443: Connect any two vectors in the latent vector space with the noise vector to obtain two composite vectors, which are the first composite vector and the second composite vector, respectively; use the remaining vectors in the latent vector space as the third composite vector. S444: Input the first composite vector into the generator in the background stage to obtain a lesion-free image; input the second composite vector into the generator in the mask stage to obtain a mask image, and multiply the mask image with the lesion-free image to obtain an image containing only non-lesion areas; input the third composite vector into the generator in the enhancement stage to obtain an enhanced image, and superimpose the enhanced image with the lesion-free image to obtain a realistic medical image.

2. The multi-stage generative adversarial network supporting the generation of medical segmentation data pairs according to claim 1, characterized in that, In S1, the decoupling is as follows: the medical image is divided into lesion regions and non-lesion regions using a separation mask. The lesion regions are obtained by multiplying the medical image with the corresponding separation mask; the non-lesion regions are obtained by multiplying the medical image with the corresponding separation mask inversely.

3. The multi-stage generative adversarial network supporting the generation of medical segmentation data pairs according to claim 2, characterized in that, Specifically, S3 involves: using the symmetry of the human body structure to obtain the symmetry of tissues and organs in the medical image; based on the symmetry of the tissues and organs, replacing the pixels in the lesion area with non-lesion pixels with symmetrical coordinates, and replacing the adjacent pixels of the boundary pixels of the lesion area to obtain a lesion-free medical image.

4. The multi-stage generative adversarial network supporting the generation of medical segmentation data pairs according to claim 3, characterized in that, In step S424, designing an ablation experiment includes the following steps: S4241: Design a two-stage generative model, including the following steps: S42411: Extract the first two stages of a three-stage generative adversarial network; S42412: Use medical images as input data and generation targets for the first stage, and use segmentation masks as input data and generation targets for the second stage; S41423: Use the generated data from the first stage as a medical image and the generated data from the second stage as the corresponding segmentation mask to form a medical image segmentation data pair; S4242: Design a four-stage generative model, including the following steps: S42421: Add a fourth stage to the three-stage generative adversarial network, which is completely identical to the third stage. S42422: The synthesized lesion-free image is used as the input data and generation target of the first stage, and the segmentation mask is used as the input data of the second stage. The medical image is used as the input data of the third and fourth stages. S42423: Use the generated data from the fourth stage as a medical image and the generated data from the second stage as the corresponding segmentation mask to form a medical image segmentation data pair; S4243: Train a two-stage generative model, a three-stage generative model, and a four-stage generative model, and obtain the generation results respectively. By comparing the three generation results, the necessity of three-stage generation is verified.

5. A terminal device for running the multi-stage generative adversarial network supporting the generation of medical segmentation data pairs as described in any one of claims 1-4, characterized in that, The system includes a storage medium, a memory, and a processor. The storage medium is used to store at least one program. The memory is used to load the storage medium and support the parameters and instructions of a multi-stage generative adversarial network for generating medical segmentation data pairs. The processor loads the parameters and executes the instructions.

Images