A method for generating images of lesions in different positions based on mammographic images

By adding classifier branches and ECA-ResNet50 network to the CycleGAN network, combined with the PatchGAN network, the problem of image generation and classification of mammography images is solved, improving image quality and detail authenticity, and helping to improve diagnostic accuracy.

CN118762238BActive Publication Date: 2024-11-08HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411240986.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-05
Publication Date
2024-11-08
Estimated Expiration
2044-09-05

AI Technical Summary

Technical Problem

The prior art has failed to effectively use generative adversarial networks to study the relationship between different positions in mammogram images, resulting in missing data, increasing the difficulty of diagnosis for doctors, and the generated image details and textures are not realistic enough.

Method used

Add classifier branches to the generator structure of the CycleGAN network, and add the ECA-ResNet50 network as a perceived loss extractor. Combined with the PatchGAN network structure, it realizes the generation and classification of lesion images of different positions, and improves the authenticity of image details and textures.

Benefits of technology

Through the improved network model, the generated mammography images are of better quality, helping doctors and artificial intelligence to make more accurate judgments, the classification performance is better than general models, and the image details are more realistic.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118762238B_ABST
    Figure CN118762238B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for generating images of lesions in different positions based on mammographic images. The method first obtains medical image data of breast cancer, performs preprocessing, and obtains image data of lesion areas in mammographic images. Secondly, the image data of lesion areas in mammographic images are divided into a training set and a test set, and data enhancement and sample balancing are performed on the sample data of the training set. Then, a lesion image generation model is constructed, and the processed training set data is input to obtain the lesion image generation result. Finally, the test set is used for testing to evaluate the quality of the generated image. The present invention realizes the generation and classification of lesion images in different positions of mammographic images, improves the detail information of the generated image, and makes the generated image quality better.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of image processing and relates to a method for generating images of lesions in different body positions based on mammographic X-ray images. Background Art

[0002] With the development of medical imaging, medical imaging plays an increasingly important role in clinical practice and is of great help to the diagnosis and treatment of patients. Medical imaging can reflect the patient's pathology and tumor location and other information. Professional physicians can diagnose the disease of the patient's tissues and organs by analyzing the relevant information of medical imaging.

[0003] Mammographic images usually contain two body position images, namely the cranio-caudal (CC) image and the mediolateral oblique (MLO) image. There is complementary image information between images in different body positions. During the diagnosis process, the doctor obtains the final diagnosis result by making a comprehensive diagnosis of the image information of the two body positions. When collecting experimental data, it was found that the image data of some patients only had a single body position image. After investigation, it was found that the reason for this phenomenon may be that the data was lost or damaged during the collection and transfer process. At the same time, some hospitals only take single body position images in order to reduce the amount of radiation that patients are exposed to due to image shooting. The lack of data greatly increases the difficulty of diagnosis for doctors.

[0004] At present, generative adversarial networks are widely used in the field of medical image generation. They can solve problems such as data scarcity and category imbalance, but no one has applied them to study the relationship between images in different body positions and improve the classification performance of the model based on this relationship. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems and provide a method for generating images of lesions in different positions based on mammographic images. A classifier branch is added to the generator structure of the CycleGAN network, and an ECA-ResNet50 network structure is added to the overall model as a perceptual loss extractor. The classifier branch can classify the input image into benign and malignant while generating the image, and the perceptual loss makes the detailed texture part of the generated image more realistic. The discriminator adopts the PatchGAN network structure. By adding a classifier and a perceptual loss to the CycleGAN network, it is possible to generate images of lesions in different positions while classifying the input image, and its classification performance is better than that of the general model.

[0006] To achieve the above object, the technical solution of the present invention is:

[0007] A method for generating images of lesions in different positions based on mammographic images comprises the following steps:

[0008] Step 1: Obtain medical imaging data of breast cancer, perform format conversion on the acquired mammographic images, filter the converted data, and remove images with poor image quality. Extract the ROI area images of the remaining images according to the annotations. Finally, obtain the image data of the lesion area of ​​the patient's mammographic image.

[0009] Step 2: First, divide the image data of the mammographic image lesion area into a training set and a test set. Then, perform data enhancement and sample balancing on the sample data of the training set, and expand the training set through rotation and mirroring operations.

[0010] Step 3: Improve the generator structure and loss function of the CycleGAN network, build a lesion image generation model, input the training set data processed in step 2, and obtain the lesion image generation result.

[0011] The generator structure and loss function of the CycleGAN network are improved. A classifier branch is added to the original generator structure. At the same time, a ResNet50 convolutional neural network with an attention mechanism is added to calculate the perceptual loss. At the same time, the calculated perceptual loss and classification loss are combined with the original GAN ​​loss and cycle consistency loss of the network. The total loss is used as the loss function of the new model and is iteratively updated during the model training process.

[0012] Step 4: Test the test set to evaluate the quality of the generated image.

[0013] The images generated by the improved network are compared and analyzed with the original images. The quality of the images generated by this network is evaluated by indicators such as Structural Similarity Index (SSIM), Peak Signal to Noise Ratio (PSNR), and LPIPS (Learned Perceptual Image Patch Similarity).

[0014] Step 5: Classify the good indicators of breast cancer, draw the ROC curve of network classification, calculate the AUC, ACC, specificity, accuracy, sensitivity and other indicators of the model, and evaluate the classification performance of this model. In addition, by comparing the classification effect of the generated image with the real image, the effectiveness and practicality of the present invention are further verified.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The present invention relates to the generation and classification of lesions in different body positions in breast X-ray images based on a generative adversarial network. The generation and classification of lesions in different body positions in breast X-ray images is achieved by changing the original structure of the CycleGAN network model, and the transfer learning strategy is used to add perceptual loss to the generative adversarial network to enhance the detail information of the generated images, thereby improving the quality of the generated images and helping doctors and even artificial intelligence to make more accurate judgments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0018] Figure 1 A flowchart of image data processing provided by an embodiment of the present invention;

[0019] Figure 2 The image data ROI region extraction diagram provided by the embodiment of the present invention;

[0020] Figure 3 A structural diagram of an improved CycleGAN network generator provided in an embodiment of the present invention;

[0021] Figure 4 The improved overall CycleGAN network structure diagram provided by the embodiment of the present invention;

[0022] Figure 5 PSNR index evaluation diagram of the generated image and the real image of the model and other models provided in the embodiment of the present invention;

[0023] Figure 6 SSIM index evaluation diagram of the generated images and real images of the model and other models provided in the embodiment of the present invention;

[0024] Figure 7 LPIPS index evaluation diagram of the generated images and real images of the model and other models provided in the embodiment of the present invention;

[0025] Figure 8 ROC curve diagram of the classification effect of this model and other models on the benign and malignant classification task provided in the embodiment of the present invention.

[0026] Fig. 9 ROC curve diagram of the classification effects of real images and generated images in this model for the benign and malignant classification task provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, a method for generating images of lesions in different positions based on mammographic images comprises the following steps:

[0029] Step 1: Obtain breast imaging data

[0030] The medical images used in the present invention were all taken by the same model of Hologic digital mammography machine and Hologic digital breast tomosynthesis imaging fusion machine. The patient's relevant case information came from the Affiliated Cancer Hospital of Fudan University. Then, the collected raw image data were processed by format conversion, screening, cleaning, ROI area extraction, data pairing and data set division, etc. Figure 2 shown.

[0031] The specific steps are as follows:

[0032] First, the DICOM format image data generated by the initial radiological diagnosis and treatment equipment needs to be converted. By reading the image-related information in the DICOM format, it is losslessly compressed and converted into a JPG image format with high resolution and high versatility. Then the converted patient data is screened to remove data with poor image quality or missing data. The lesion area is segmented according to the location information of the lesion area marked by the doctor. In order to make the information contained in the lesion area images between different body positions as similar as possible, the location information needs to be processed to a certain extent. The same patient uses a mask of the same size and the lesion area is placed in the center of the mask.

[0033] Step 2: Preprocess the segmented image data

[0034] Since the amount of benign sample data in the breast X-ray images used is relatively small and there is sample imbalance in the data, we need to perform data enhancement and data balancing on the segmented breast image data. First, the data set is randomly divided into a training set and a test set in a ratio of 7:3. When expanding the data set, only the training set is expanded, and the test set data remains unchanged. Secondly, for the sample imbalance problem in the training set, the data enhancement method is chosen, that is, the image data after sample balancing is rotated and mirrored, and the purpose of sample balance is achieved by expanding the small number of samples. Finally, for the benign and malignant classification task of breast cancer, a total of 492 patient data are used. There are 344 cases in the training set, including 71 benign samples and 273 malignant samples. After data balancing, there are 284 benign samples and 273 malignant samples; there are 148 cases in the test set, including 31 benign samples and 117 malignant samples.

[0035] Step 3: Design of CCycleGAN network

[0036] like Figure 4 As shown in the figure, the lesion image generation model CCycleGAN consists of three parts: a generator, a discriminator, and a perceptual loss extractor. The input of the generator is the real image, and the output is the benign and malignant prediction results of the generated image and the input image; the input of the discriminator is the real image and the generated image, and the output is the probability value of the image being true. The input of the perceptual loss extractor is the real image and the generated image, and the output is the similarity between the real image and the generated image in the deep feature space. The lesion image generation model adopts the idea of ​​cyclic generation of the CycleGAN model as a whole.

[0037] The generator structure of this model is divided into four parts: encoder 1, encoder 2, classifier, and decoder. Encoder 1 inputs the real image and outputs the deep features of the image. Encoder 2 inputs the real image and outputs the classification features of the image. The classifier inputs the deep features of the image and the classification features are concatenated after the attention mechanism, and the output is the prediction result of the image. The decoder inputs the deep features of the image and outputs the generated image. Both encoders use the residual module with an added attention mechanism as the basic unit. The residual module adds residual connections compared to the ordinary convolution module, which can train a deeper network. The channel attention mechanism can make the model focus on the area of ​​interest. Therefore, adding a channel attention mechanism to the residual module can make the model focus more on the area of ​​interest.

[0038] The discriminator of this model uses the PatchGAN network without any other modifications. The perceptual loss extractor of this network uses the ResNet50 network with the attention mechanism introduced (ECA-ResNet50 network), which is obtained by adding the channel attention mechanism ECA to the residual module of the original ResNet50 network. The weight parameters of the perceptual loss extractor are obtained by pre-training the ECA-ResNet50 network.

[0039] The encoder 1 is composed of a convolution module and a plurality of convolution units cascaded together, and the convolution module is composed of a convolution layer, a normalization layer and an activation function layer cascaded together.

[0040] The convolution unit is composed of a plurality of residual modules 1 and a final cascaded residual module 2.

[0041] The residual module 1 is composed of two convolution modules and an ECA cascade, and the input features of the residual module 1 are concatenated with the output features of the ECA through a residual connection to obtain the output features of the residual module 1. The residual module 2 is composed of two convolution modules and an ECA cascade, and the input features of the residual module 1 are concatenated with the output features of the ECA through a convolution module through a residual connection to obtain the output features of the residual module 1.

[0042] The encoder 2 is composed of a plurality of cascaded encoding units, and the encoding unit is composed of a cascaded convolution module, an ECA and a pooling layer.

[0043] The classifier is composed of a classifier module and a fully connected layer cascaded, and the structure of the classifier module is the same as that of encoder 2.

[0044] The loss function of the model is composed of GAN loss ( ), cycle-consistent loss( ), Perceptual Loss( ) and classification loss ( ) consists of four parts. Representation Generator , Represents the discriminator, Indicates the input image, represents the true label image, Indicates the generated image, represents the probability of class 1.

[0045] GAN loss ) is the classic loss function in the generative adversarial network, which is obtained by minimizing the difference between the generator and the discriminator. Its formula is shown as follows:

[0046] In the formula represents the input image, Represents a label image, represents the distribution of the sample, Represents the discriminator to sample The probability of judging it to be true, Represents the discriminator to generate The probability that the generated image is classified as true.

[0047] Cycle-consistent loss ( ) is measured by the L1 norm and The difference between and The difference between them can be obtained, which can prompt the generator to learn a bidirectional consistent mapping relationship. The formula is as follows:

[0048]

[0049] In the formula represents the input image, Represents a label image, Represents the distribution of the sample.

[0050] Classification loss( ) is calculated by the predicted results and the actual results. The formula is as follows:

[0051]

[0052] in, represents a generator, represents the classifier, is the input image, is the category label.

[0053] Perceived loss It is calculated by the ECA-ResNet50 network. The specific calculation process is as follows: first, the feature maps corresponding to the real image and each feature layer of the generated image can be obtained through each convolution layer of ECA-ResNet50. Secondly, the loss of the feature maps corresponding to the selected feature layers is calculated respectively. Finally, the losses of the feature maps corresponding to all the selected feature layers are added together, which is the perceptual loss.

[0054] ECA-ResNet50 is divided into 5 stages, among which Stage0 has a relatively simple structure, which only preprocesses the input data. The last 4 stages are composed of Bottleneck and have similar structures. Stage1 contains 3 Bottlenecks, and the other 3 stages contain 4, 6, and 3 Bottlenecks respectively. This paper chooses to use the last 3 Bottlenecks of Stage4 and the 3 Bottlenecks of Stage5, a total of 6 feature maps to calculate the perceptual loss. The loss function calculation formula for perceptual loss is as follows:

[0055]

[0056] In the formula Perceptual loss extractor In the The feature map extracted from the layer, Expressed as The calculated Gram matrix, Represents the size of the j-th layer feature map, and Represent the generated image and the real image respectively.

[0057] By adding the perceptual loss and classification loss to the loss function of the CycleGAN network, we get the improved loss function as shown below:

[0058]

[0059] in, are the weight parameters of the three loss functions, with values ​​of 10, 5, and 5 respectively.

[0060] The network structure of the model is as follows Figure 3 In the parameter iteration process of model training, the generator G is first fixed and the discriminator D is trained so that the discriminator can accurately distinguish true and false data, that is, the output probability of the real data is 1 and the output probability of the generated data is 0, thereby maximizing , Represents the distance between the real data and the generated data. By maximizing log(D(x)), we make the probability of the classification model output as close to 1 as possible when given the real data, and as close to 0 as possible when given the generated data. This can effectively train the discriminator to better distinguish between real data and generated data.

[0061] Then, the discriminator D is fixed and the generator G is trained so that the probability of the data generated by the generator after being judged by the discriminator is as close to 1 as possible, thereby minimizing , the purpose of minimizing log(1 - D(G(z))) is that when the output ( G(z) ) of the generator ( G ) is judged as real data by the discriminator D, D(G(z)) is close to 1, so the value of log(1 - D(G(z))) is close to 0. This means that the greater the probability that the data G(z) generated by the generator is misjudged as real data by the discriminator, the smaller the value of this objective function, indicating that the performance of the generator is better.

[0062] This model uses the Adam optimizer for gradient update. During the model training process, the value of Adam's hyperparameter β_1 is set to 0.5, and the value of β_2 is set to 0.99. In the embodiment, the Batch Size is set to 2, that is, two images are run each time. In view of the fact that the GAN model and the classifier have inconsistent convergence during the training process, setting a fixed learning rate may cause deviation in the generation effect or poor classification effect. Therefore, this paper dynamically updates the learning rate, and adjusts it by 95% every 50 batches of training. The initial learning rate is set to 1e-3.

[0063] Step 4: Model Results Evaluation

[0064] The generated images of this model are compared with the real images, and the SSIM, PSNR, and LPIPS indicators are used to evaluate the effects of the enhanced image sequences generated by the generative adversarial network and interpolation.

[0065] (1) Peak Signal Ratio (PSNR)

[0066] PSNR is the most widely used evaluation index, which measures the pixel difference between the reconstructed image and the original image. The unit of PSNR is dB, and its formula is as follows:

[0067]

[0068] Where Peak is the maximum value of the pixel. For different pictures, if the sampling interval is 8, then Peak is 255. If the sampling interval is n, then Peak is From the relationship between PSNR and MSE, we can see that the smaller the MSE value, the higher the PSNR value. Therefore, the higher the PSNR value, the better the image reconstruction effect.

[0069] (2) Structural Similarity (SSIM)

[0070] SSIM compares the similarity between the reconstructed image and the original image by integrating the brightness, contrast and structure of the image. The SSIM value is between 0 and 1. The closer it is to 1, the higher the similarity of the images. The SSIM calculation formula is as follows:

[0071]

[0072] in, and Represent the pixel averages of the reconstructed image and the original image, respectively. and Represent the pixel variance of the reconstructed image and the original image respectively, represents the covariance, and are all constants.

[0073] (3) LPIPS

[0074] LPIPS is a measure of the perceived image block similarity between two images. LPIPS uses deep learning methods to learn the perceived differences between images, thereby more accurately evaluating the similarity between images. Its calculation formula is as follows:

[0075]

[0076] The evaluation results of PSNR, SSIM and LPIPS evaluation indicators of generated images and real images are as follows Figure 5 , Figure 6 and Figure 7 As shown. The inclusion of ECA in the model name indicates that the model structure introduces the attention mechanism, and the inclusion of Pcp in the model name indicates that the model structure adds perceptual loss. The model used in the present invention is named CCycleGAN, which includes an attention mechanism, perceptual loss, and a classifier. The other models are Res_cycle, ECARes_cycle, and ECARes_cycle_Pcp models.

[0077] The comparison of the effects of various indicators in the figure shows that the model of the present invention is superior to other models in terms of image generation quality.

[0078] Step 5: Evaluation of benign and malignant classification task results

[0079] The ultimate goal of the present invention is to improve the benign and malignant prediction performance of the entire model by generating tasks between different body positions. Therefore, the ROC curves of each model are drawn by comparing with other models. Figure 8 and Fig. 9 shown.

[0080] It can be seen from Table 1 that the classification performance of the model of the present invention is better than that of other models, and it can be seen from Table 2 that although there are differences between the generated images and the real images, the differences are not significant. A t-test was performed on the benign and malignant classification results of the generated images and the real images, and the P-value values ​​were all greater than 0.05, indicating that there was no significant difference between the generated images and the real images.

[0081] Table 1 Classification results of each model

[0082]

[0083] Table 2 Comparison of classification results between generated images and real images in different body positions

[0084]

[0085] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A method for generating images of lesions in different positions based on mammographic images, It is characterized in that The following steps are involved: Step 1: Obtain medical imaging data of breast cancer, perform preprocessing, and obtain imaging data of the lesion area of ​​the mammographic image; Step 2: Divide the image data of the lesion area of ​​the mammographic X-ray image into a training set and a test set, and perform data enhancement and sample balancing on the sample data of the training set; Step 3: Improve the generator structure and loss function of the CycleGAN network, build a lesion image generation model, input the training set data processed in step 2, and obtain the lesion image generation result; The lesion image generation model CCycleGAN consists of three parts: a generator, a discriminator, and a perceptual loss extractor. The input of the generator is a real image, and the output is the prediction result of the generated image and the input image; the input of the discriminator is the real image and the generated image, and the output is the probability value of the image being true; the input of the perceptual loss extractor is the real image and the generated image, and the output is the similarity between the real image and the generated image in the deep feature space; the lesion image generation model CCycleGAN adopts the idea of ​​cyclic generation of the CycleGAN model; The generator structure is divided into four parts: encoder 1, encoder 2, classifier, and decoder. The encoder 1 inputs the real image and outputs the deep features of the image. The encoder 2 inputs the real image and outputs the classification features of the image. The classifier inputs the deep features of the image and the classification features of the image, which are often concatenated after the self-attention mechanism, and the output is the prediction result of the image. The decoder inputs the deep features of the image and outputs the generated image. The discriminator adopts the PatchGAN network; The perceptual loss extractor adopts a ResNet50 network with an attention mechanism, which is obtained by adding a channel attention mechanism ECA to the residual module of the original ResNet50 network; the weight parameters of the perceptual loss extractor are obtained by pre-training the ECA-ResNet50 network; The encoder 1 is composed of a convolution module and a plurality of convolution units cascaded, and the convolution module is composed of a convolution layer, a normalization layer and an activation function layer cascaded; The convolution unit is composed of a plurality of residual modules 1 and a final cascaded residual module 2; The residual module 1 is composed of two convolution modules and an ECA cascade, and the input features of the residual module 1 are spliced ​​with the output features of the ECA through residual connections to obtain the output features of the residual module 1; The residual module 2 is composed of two convolution modules and an ECA cascade, and the input features of the residual module 1 are concatenated with the output features of the ECA through a convolution module through a residual connection to obtain the output features of the residual module 1; The encoder 2 is composed of a cascade of several encoding units, and the encoding unit is composed of a cascade of a convolution module, an ECA and a pooling layer; The classifier is composed of a classifier module and a fully connected layer cascaded, and the structure of the classifier module is the same as that of encoder 2; The loss function of the lesion image generation model CCycleGAN is composed of GAN loss Cycle-Consistency Loss Perceptual loss L pcp (G) and classification loss L clf (G,C 1 ) consists of four parts, where G represents the generator G, D Y represents the discriminator, X represents the input image, Y represents the real label image, F represents the generated image, C 1 represents the probability of category 1; GAN Loss It is obtained by minimizing the difference between the generator and the discriminator, and its formula is as follows: Where x represents the input image, y represents the label image, and p data represents the distribution of samples, D Y (y) represents the probability that the discriminator classifies sample y as true, D Y (G(x)) represents the probability that the discriminator classifies the image generated by the generator G as real; Cycle-Consistency Loss The difference between F(G(x)) and x is measured by the L1 norm and The difference between G(F(y)) and y is obtained, and its formula is as follows: Classification loss L clf (G,C 1 ) is calculated by the predicted results and the actual results. The formula is as follows: Among them, G represents the generator, C 1 represents the classifier, x is the input image, and l is the category label; Perceptual loss L pcp (G) is calculated by the ECA-ResNet50 network. The specific calculation process is as follows: first, the real image and the feature maps corresponding to each feature layer of the generated image are obtained through each convolutional layer of ECA-ResNet50; secondly, the loss of the feature maps corresponding to the selected feature layers is calculated respectively; finally, the losses of the feature maps corresponding to all the selected feature layers are added together, which is the perceptual loss. The calculation formula of the perceptual loss function is as follows: Where φ j is the feature map extracted by the perceptual loss extractor φ in the jth layer, K j Represented as φ j The calculated Gram matrix, C j H j W j Represents the size of the j-th layer feature map, and y represent the generated image and the real image respectively; By adding the perceptual loss and classification loss to the loss function of the CycleGAN network, we get the improved loss function as shown below: Among them, λ, μ, ξ are the weight parameters of the three loss functions; Step 4: Test the test set to evaluate the quality of the generated image.

2. The method for generating images of lesions in different positions based on mammographic images according to claim 1, It is characterized in that The preprocessing includes: acquiring medical imaging data of breast cancer patients, performing format conversion operations on the acquired breast X-ray images, performing image quality screening on the converted data, extracting ROI area images of the remaining images according to the annotations, and obtaining the patient's breast X-ray image lesion area image data.

3. The method for generating images of lesions in different body positions based on mammographic images according to claim 1, It is characterized in that The data enhancement and sample balancing processing are specifically as follows: the training set is expanded through rotation and mirroring operations.

4. The method for generating images of lesions in different positions based on mammographic images according to claim 3, It is characterized in that In the parameter iteration process of the lesion image generation model training, the generator G is first fixed, and the discriminator D is trained so that the discriminator can accurately distinguish true and false data, that is, the output probability of the real data is 1, and the output probability of the generated data is 0, and log(D(x)) is maximized, where log(D(x)) represents the distance between the real data and the generated data. The discriminator is trained by maximizing log(D(x)); Then, fix the discriminator D and train the generator G so that the probability of the data generated by the generator after being judged by the discriminator is as close to 1 as possible, thereby minimizing log(1-D(G(z))).