A two-stage generative adversarial network guided cervical cancer pathology image analysis method

Abstract

The present invention relates to the field of image processing technology, and discloses a cervical cancer pathology image analysis method guided by a two-stage generative adversarial network, comprising the following steps: collecting real pathology images of cervical cancer and marking them with real labels; constructing a cervical cancer pathology data generation network based on the two-stage generative adversarial network, including a pathology image generation network and a pathology image label generation network; the cervical cancer pathology data generation network receives real pathology images of cervical cancer and marks them with real labels, outputs generated pathology images and corresponding generated labels, and is used to jointly train a downstream segmentation network; and uses the trained downstream segmentation network to segment the cervical cancer pathology image to identify cancerous tissue, solid tissue and background parts. The present invention improves the efficiency and accuracy of cervical cancer pathology image analysis by generating pathology image data.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a cervical cancer pathology image analysis method guided by a two-stage generative adversarial network. Background Art

[0002] In clinical practice, the pathological diagnosis of cervical cancer mainly relies on histopathological examination of full-field digital slides. Pathologists perform staging by repeatedly examining pathological slides under a microscope and estimating the proportion of solid areas in cancerous areas. However, current conventional pathological diagnosis has certain limitations. Manual diagnosis is very time-consuming, especially when dealing with large pathological datasets. This heavy workload may compromise the timeliness and accuracy of diagnosis. To address these limitations, computer-assisted pathological staging diagnosis of cervical cancer can be achieved by performing multi-class segmentation in pathological images. In recent years, the rise of deep learning has significantly promoted the progress of computer-assisted diagnosis and improved the accuracy of medical image segmentation. However, training accurate multi-class deep learning segmentation models requires a large number of annotated pathological image datasets, which are challenging to obtain. Due to concerns about patient privacy, pathological image acquisition is subject to strict ethical and legal restrictions. In addition, imperfect data sharing mechanisms hinder access to large-scale and diverse datasets, limiting the progress of related research. Semantic segmentation labeling is labor-intensive and requires specialized domain expertise because every pixel must be labeled. This process may lead to fatigue-induced errors such as omissions and mislabeling. Therefore, the scarcity of labeled training data, i.e., ultra-low data mode, poses a significant challenge to existing deep learning-based segmentation methods, which may lead to overfitting or poor model generalization. The literature (Park S, Moon J, Hwang E. Data generation scheme for photovoltaic power forecasting using Wasserstein GAN with gradient penalty combined with autoencoder and regression models. Expert Systems with Applications (2024)) proposed a new data generation scheme that combines Wasserstein GAN with gradient penalty, autoencoder and regression model to generate tabular data for photovoltaic power prediction. The literature (Aeschbacher D, Meisner J, Miletic M, et al. Use and Evaluation of GANs for Synthetic Data Generation in Pharmacogenetics. Studies in Health Technology and Informatics November (2024)) Dominic et al. studied the performance of two GAN models (CTGAN and CTAB-GAN+) in generating synthetic pharmacogenomics (PGx) data.The literature (Subramaniam P, Kossen T, Ritter K, et al. Generating 3D TOF-MRA volumes and segmentation labels using generative adversarial networks. Medical Image Analysis (2022)) introduced 3D GANs to generate 3D medical images and their corresponding labels, while applying mixed precision to reduce computational costs. However, current generation methods mainly focus on the visual quality of other medical images, ignoring the semantic information in pathological images and the accurate transmission of pathological features. The generated images may be difficult to provide sufficient clinical reference value in actual clinical applications. Summary of the invention

[0003] The purpose of the present invention is to solve the problems in the prior art.

[0004] The technical solution adopted by the present invention to solve the technical problem is: to provide a cervical cancer pathology image analysis method guided by a two-stage generative adversarial network, comprising the following steps:

[0005] Collect real pathological images of cervical cancer and label them with real labels;

[0006] A cervical cancer pathology data generation network is constructed based on a two-stage generative adversarial network, including a pathology image generation network and a pathology image label generation network;

[0007] The cervical cancer pathology data generation network receives real pathology images of cervical cancer and adds real labels, and outputs generated pathology images and corresponding generated labels for joint training of downstream segmentation networks;

[0008] Use the trained downstream segmentation network to segment cervical cancer pathology images and identify cancerous tissue, solid tissue, and background.

[0009] Among them, both the pathological image generation network and the pathological image label generation network adopt generative adversarial networks; the pathological image generation network includes a fully connected layer, an image generator, and an image discriminator connected in sequence; the pathological image label generation network includes a pathological prior feature guided loss block, a label generator, and a label discriminator connected in sequence.

[0010] Preferably, the pathological image generation network receives a real pathological image and outputs a generated pathological image, and the process includes the following steps:

[0011] After obtaining the random hidden variable z from the real pathological image and regularizing it, it is input into the fully connected layer to obtain the decoupled intermediate hidden vector s, which is expressed as:

[0012] ;

[0013] Among them, W is the weight coefficient and b is the bias vector;

[0014] The intermediate hidden vector s is input into the image generator together with the real pathological image to generate features of different scales, which are input into the image discriminator for discrimination; the image generator uses the pathological style feature information fusion normalization module to add random noise y in the upsampling stage z , enrich the detailed information of generated pathological images;

[0015] The image discriminator outputs a pathological image.

[0016] Preferably, the pathological style feature information fusion normalization module is based on the self-guided feature map y w and a random vector y z As input, the output is the pathological style feature information fusion normalized perception map y n , expressed as:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] Among them, conv5 is a 5×5 convolution operation; AdaFLN is an adaptive instance feature layer normalization operator, α and ρ are learnable parameters used to constrain the feature vectors extracted by different normalization modules; AdaIN is an adaptive instance normalization operator; AdaLN is an adaptive layer normalization operator, l w represents the normalized input feature map of the adaptive layer, l z represents the input feature vector of the adaptive layer normalization; AdaFN is the adaptive feature normalization operator, f w represents the input feature map of adaptive feature normalization, f z represents the input feature vector of adaptive feature normalization; μ and σ represent the mean and variance operations, respectively.

[0022] Preferably, the pathological image generation network adopts a loss function based on CycleGan, which is expressed as:

[0023] ;

[0024] Among them, D ITIrepresents the loss function of the image discriminator, G ITI represents the loss function of the image generator, X represents the predicted label of the pathological image generation process, Y represents the true label of the pathological image generation process, F represents the true source domain of the pathological image, and L ITI is the loss function of the pathological image generation network, L GAN To combat the loss, L Cycle is the cycle consistency loss, L Identity is the identity loss, λ a , b and λ c denote the weights of adversarial loss, cycle consistency loss, and identity loss, respectively.

[0025] Preferably, in the pathological image generation network, the fully connected layer includes eight 512×512 depth-separable convolutional layers; the image generator adopts 8-layer ResNet50; and the image discriminator adopts 3-layer ResNet50.

[0026] Preferably, the pathological image label generation network receives a real pathological image with a real label and a generated pathological image output by the pathological image generation network, and outputs a generated label, and the process includes the following steps:

[0027] The pathology prior feature guided loss block receives the real pathology image with the real label and the generated pathology image output by the pathology image generation network, generates the pathology prior feature guided loss, and inputs it into the label generator;

[0028] The adversarial relationship between the label generator and the label discriminator is used to generate labels corresponding to the generated pathological images.

[0029] Preferably, the pathological prior feature guides the loss L ITL-A It is expressed as:

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] Among them, λ d , e and λ f are the weights of cross entropy loss, boundary loss and distance feature penalized cross entropy loss respectively; represents the cross entropy loss, represents the i-th element of the true label, Represents the probability that the model predicts that it belongs to the i-th category; represents the boundary loss, is the distance feature calculated by the distance transform of the true segmentation label, is the pixel-level probability value predicted by the model, α represents the index of each pixel in the image, and N GB is the total number of pixels in the image used for normalization; represents the distance feature penalty cross entropy loss, N DFP is the total number of pixels, c is the category of pixels, is the probability value of pixel β belonging to class c, is the true value of pixel β belonging to class c, is the distance penalty term for class c.

[0035] Preferably, the total loss function L of the pathological image label generation network is ITL It is expressed as:

[0036] ;

[0037] in, is the loss function of the pathological image generation network, g represents the weight of the loss function of the pathological image generation network, and h represents the weight value of the loss guided by the pathological prior feature.

[0038] Preferably, in the pathological image label generation network, the label generator adopts 7-layer ResNet50; the label discriminator adopts 3-layer ResNet50.

[0039] Preferably, the downstream segmentation network is jointly trained with the cervical cancer pathology data generation network, and the adopted loss function is expressed as:

[0040]

[0041] in, represents the predicted label of the downstream segmentation network, w d represents the trainable parameters in the downstream segmentation network, which is used to constrain the accuracy of the segmentation network prediction results. z represents the true label of the pathological image, y represents the predicted label of the pathological image generated after the true label constraint, and w e represents the trainable parameters in the generative network, which is used to constrain the accuracy of the generated results; θ represents the normalization of the segmentation network, Ф represents the regularization of the generative network, and L d is the regularization term of the downstream segmentation network, L e is the regularization term of the pathological data generation network, is a balanced hyperparameter.

[0042] The present invention has the following beneficial effects:

[0043] (1) The pathological style feature information fusion normalization module of the present invention addresses the problem of diverse features and complex edge features of pathological images in the image generation stage. It proposes adaptive instance feature layer normalization to normalize the input features, and then adjusts the features of the target image through statistics of the style image to achieve independent processing of different style features.

[0044] (2) Aiming at the problem that the features of tumor and solid tumor regions in cervical cancer pathology images are similar and difficult to distinguish during the label generation stage, the present invention proposes a pathology prior feature guided loss block, which includes a generative cross entropy loss, a generative boundary loss, and a distance feature penalized cross entropy loss. This loss block improves the feature extraction capability of small sample solid tumors, enhances the accuracy of boundary marking, and increases the model's sensitivity to the target region boundary. The proposed method utilizes prior pathology knowledge to enhance the feature extraction capability of pathology images and improve the accuracy of label generation.

[0045] (3) The present invention proposes a two-layer learning scheme to optimize the pathology data generation task, generate pathology images and corresponding labels of pathology images, and cooperate with downstream tasks to help generate a pathology dataset that is close to the real image and segmentation standard and is beneficial to the downstream segmentation task.

[0046] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A method step diagram of an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of a flow chart of an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of a pathological image generation network structure according to an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of a network structure diagram for generating pathological image labels according to an embodiment of the present invention;

[0051] Figure 5 is a real pathological image according to an embodiment of the present invention;

[0052] Figure 6 Generating a pathological image for an embodiment of the present invention;

[0053] Figure 7 A schematic diagram of generating labels according to an embodiment of the present invention;

[0054] Figure 8 A schematic diagram of the downstream segmentation results of an embodiment of the present statement. DETAILED DESCRIPTION

[0055] See also Figure 1 and Figure 2 The method step diagram and flow chart of an embodiment of the present invention are shown, which include the following steps:

[0056] S101, collects real pathological images of cervical cancer and adds real labels;

[0057] S102, constructing a cervical cancer pathology data generation network based on a two-stage generative adversarial network, including a pathology image generation network and a pathology image label generation network;

[0058] S103, the cervical cancer pathology data generation network receives the real pathology image of cervical cancer and adds the real label, and outputs the generated pathology image and the corresponding generated label for joint training of the downstream segmentation network;

[0059] S104, using the trained downstream segmentation network to segment the cervical cancer pathology image to identify cancerous tissue, solid tissue and background parts;

[0060] Among them, both the pathological image generation network and the pathological image label generation network adopt generative adversarial networks; the pathological image generation network includes a fully connected layer, an image generator, and an image discriminator connected in sequence; the pathological image label generation network includes a pathological prior feature guided loss block, a label generator, and a label discriminator connected in sequence.

[0061] Specifically, due to the privacy and complexity of pathological images, it is difficult to obtain pathological image data. Insufficient training data will lead to overfitting or poor model generalization ability. Therefore, a two-stage generative adversarial network is used to generate pathological images and labels corresponding to the images. In order to ensure that different elements of the control vector in the original image can independently control various visual features, enrich the detailed information while ensuring the unity of global features, and construct a pathological image generation network such as Figure 3 As shown in the figure, it mainly includes a fully connected layer, a generator and a discriminator, and incorporates a pathological style feature information fusion normalization module into the upsampling process of the generator. Its internal execution process includes the following steps:

[0062] Obtain a random hidden variable z from the original pathological image, regularize it, and input it into 8 fully connected layers to obtain a decoupled intermediate hidden vector s to control the style of the generated image. This process can be expressed as:

[0063] ;

[0064] Among them, s is the intermediate vector obtained after decoupling, W is the weight coefficient, z is the random latent variable obtained in the original pathological image, and b is the bias vector.

[0065] Then the intermediate vector s is input into the generator composed of 8 layers of ResNet50 together with the original image to generate features of different scales, which are then discriminated by the discriminator composed of 3 layers of ResNet50. In order to integrate the style information of the pathological image in the generator and further enrich the detail information of the generated image, the pathological style feature information fusion normalization module is introduced, and random noise y is added in the upsampling stage. z , enrich the detailed information of generated pathological images.

[0066] Specifically, the pathological style feature information fusion normalization module is a residual link structure, which includes adaptive instance feature layer normalization, activation function and depth-separable convolution layer. The activation function in the pathological style feature information fusion normalization module introduces randomness, allowing the leakage slope to be randomly set during training, which helps to reduce the risk of overfitting. The pathological style feature information fusion normalization module uses the self-guided feature map y w and vector y z As input, the output is the pathological style feature information fusion normalized perception map y n , and the decoded feature map y w Combined for further processing. Overall, by combining the feature vector i in the original input image with the intermediate vector s derived from the mapping and normalization of the latent vector z, the pathological style feature information fusion normalization module retains the independent style information of the latent vector while utilizing the feature clues extracted from the original image, ensuring that the fusion result has visual fidelity and style preservation. This promotes the generation of images that are closely aligned with the true standard. In order to expand the receptive field, a 5×5 convolution layer is applied on the adaptive instance feature layer normalization during the convolution operation. The final feature map y n The formula is as follows:

[0067] ;

[0068] where y w is the feature map, y z is a random vector, y n is the pathological style feature information fused normalized perception map, conv5 is a 5×5 convolution operation, and AdaFLN is an adaptive instance feature layer normalization operator.

[0069] Specifically, the adaptive instance feature layer normalization adjusts the weights of adaptive layer normalization, adaptive instance normalization, and adaptive feature normalization according to different tasks and input data while retaining the content structure and feature style information, achieving global normalization and making feature normalization more flexible. Adaptive instance feature layer normalization provides the generator with feature information containing the structure and style of pathological images, improving the performance of the model in pathological image generation. Adaptive instance feature layer normalization is defined as follows:

[0070] ;

[0071] Where α and ρ are learnable parameters that constrain the feature vectors extracted by different normalization modules. AdaIN is an adaptive instance normalization operator, AdaLN is an adaptive layer normalization operator, and AdaFN is an adaptive feature normalization.

[0072] Adaptive instance normalization, adaptive layer normalization, and adaptive feature normalization have the same mathematical formula structure. Adaptive instance normalization is expressed as:

[0073] ;

[0074] where μ and σ represent the mean and variance operations respectively. is the normalization of the previous layer input, σ(y w ) and μ(y z ) are its scaling and bias terms.

[0075] The implementation process of adaptive feature normalization is similar to that of adaptive instance normalization. The intermediate feature vector is expanded into a contraction factor and a deviation factor through a learnable affine transformation. Then these two operators are weighted summed with the normalized convolution output to complete the process of the intermediate vector affecting the original output. The formula is as follows:

[0076] ;

[0077] in, is the normalization of the previous layer features, σ(f w ) and μ(f z ) are its scaling and bias terms.

[0078] AdaLN is an adaptive layer normalization operator, expressed as:

[0079] ;

[0080] Among them, σ(l w ) and μ(l z ) represent the scaling and bias terms of AdaIN respectively.

[0081] Specifically, the pathological image generation network adopts the CycleGan loss function. The CycleGAN loss function is used to constrain the image generation process, which consists of three main components: adversarial loss, cycle consistency loss, and identity loss. These three loss functions work together to achieve image-to-image conversion, ensuring that the generated images are both realistic and consistent. The total loss function is expressed as follows:

[0082] ;

[0083] Among them, D ITI represents the loss function of the image discriminator, G ITI represents the loss function of the image generator, X represents the predicted label of the pathological image generation process, Y represents the true label of the pathological image generation process, F represents the true source domain of the pathological image, and L ITI is the loss function of the pathological image generation network, L GAN To combat the loss, L Cycle is the cycle consistency loss, L Identity is the identity loss, λ a , b and λ c denote the weights of adversarial loss, cycle consistency loss, and identity loss, respectively.

[0084] Specifically, the image pathology image label generation network is as follows Figure 4 As shown in the figure, it mainly includes a pathological prior feature guided loss block, a label generator and a label discriminator. The pathological prior feature guided loss block includes generating cross entropy loss, generating boundary loss and distance feature penalty cross entropy loss. Its internal execution process is as follows: the original image, the corresponding label and the generated image are input into the pathological prior feature guided loss block, and then a one-to-one label is generated for the pathological image through the adversarial process between the label generator and the label discriminator.

[0085] Due to the limitation of resolution and the complexity of multi-class pathological image features, it may be difficult to extract complete edge structures and detailed feature information when generating labels for pathological images, making it difficult to generate stable results. Therefore, three loss functions are proposed in the pathological prior feature guided loss block to address the problems of multi-class parameters, complex edge features, and medical image noise.

[0086] First, in order to solve the problem of multi-class pathological images in the dataset, a generative cross entropy loss is proposed to measure the gap between the model's predicted results and the actual results. When the probability distribution of the model-generated labels completely matches the true label distribution, the cross entropy loss is 0, indicating that the labels generated by the model are completely accurate. When there is a difference between the labels generated by the model and the true labels, the cross entropy loss increases, guiding the model to reduce this difference through optimization during training, so that the generated labels are closer to the true labels. In this study, since there are three categories of pathological images - background, tumor, and solid tumor - it is necessary to classify the input data x into these three categories. For each input data x 0 , define a three-dimensional vector y ^ , where y ^ i Represents x 0 The probability of belonging to the i-th category. The training is optimized so that y^ As close as possible to the true label distribution y. Assuming that the true label y is a three-dimensional vector, only one element is 1 and the rest are 0, indicating that x belongs to the mth category, the formula for generating the cross entropy loss is:

[0087] ;

[0088] where x i Represents the i-th element of the true label, y i It indicates the probability that the model predicts that x belongs to the i-th category. When the two probability distributions are closer, the cross entropy loss is smaller, indicating that the model prediction result is more accurate.

[0089] Secondly, we address the challenge of generating complex image boundary features by introducing a generative boundary loss function, using a distance metric on contour space instead of region space. The generative boundary loss computes the integral over the boundaries between regions, alleviating the problems associated with region loss in highly imbalanced segmentation tasks. An integral method is used to compute the change in the boundary, avoiding local differentiation of contour points. The final generative boundary loss is a linear function of the sum of the SoftMax probabilities of the output region, allowing it to be used with the existing region loss. The generative boundary s loss formula is as follows:

[0090] ;

[0091] in, is the distance feature calculated by the distance transform of the true segmentation label, is the pixel-level probability value predicted by the model, α represents the index of each pixel in the image, and N GB is the total number of pixels in the image used for normalization.

[0092] Finally, the distance feature penalized cross entropy loss uses the weight term of the distance feature in the true label and optimizes the network learning based on pixel-level error prediction to improve the network's feature extraction. By fusing multi-scale information, large areas and detailed areas can be processed efficiently. The distance feature penalized cross entropy loss is calculated, weighted, and summed at different scales to balance detailed information and global information, thereby enhancing the network's ability to learn feature information at different scales. It is defined as follows:

[0093] ;

[0094] Among them, N DFP is the total number of pixels, c is the category of pixels, is the probability value of pixel β belonging to class c, is the true value of pixel β belonging to class c, is the distance penalty term for class c. The distance feature penalty cross entropy loss gives greater weight to pixels on the boundary of c.

[0095] Therefore, the network is guided to focus on the boundary areas that are difficult to segment. The formula is as follows:

[0096] ;

[0097] Among them, L ITL-A is the loss function of the pathological prior feature guided loss block; d To generate the weight of cross entropy loss, the value of this embodiment of the present invention is 0.35, λ e To generate the weight of the boundary loss, the value of λ in the embodiment of the present invention is 0.35. f The cross entropy loss is penalized for the distance feature, and the value of this embodiment of the present invention is 0.3.

[0098] In the embodiment of the present invention, the loss functions of the label generator and the label discriminator in the pathological image label generation network are the same as those of the pathological image generation network, so the total loss function formula of the image generation label is as follows:

[0099] ;

[0100] Among them, L ITL The loss function for generating labels for images, λ g is the weight of the loss function of the generated adversarial network, λ h The pathology prior features guide the loss function weights of the loss block.

[0101] Specifically, the downstream segmentation network uses Deeplabv3+ as the basic segmentation network, and further proposes a cooperative training strategy for jointly optimizing network parameters. The pathology data generation network imposes additional constraints on the loss function of the segmentation task, which can pay more attention to boundary information and generate more standard pathology images and more accurate annotation labels. The downstream segmentation network and the pathology data generation network are represented as follows:

[0102] ;

[0103] ;

[0104] Among them, Ф(;w e ) is based on a two-stage generative adversarial network analysis network and the corresponding parameters w e , L d is the regularization term of the downstream segmentation network, L e is the regularization term of the pathological data generation network, λ 3 is a balanced hyperparameter with a value of 0.25. Through the cooperative training strategy and the two-layer optimization network, a more optimized data generation solution can be achieved under the constraints of the downstream segmentation task.

[0105] The effect of the embodiment of the present invention was tested using an Intel CPU-i7-12700KF processor and four NVIDIA GeForce RTX 3090Ti GPUs. The software environment used Python 3.9 and CUDA 11.2 programming languages. The training was performed on 512×512 input images with a batch size of 32 and a training batch size of 160. The initial learning rate was set to 0.005, and the Adam optimizer was used for training. The real pathological images used are as follows: Figure 5 As shown, the pathological image is generated as Figure 6 As shown, the generated labels are as follows Figure 7 As shown in the figure, the segmentation task effect diagram is as follows Figure 8 shown.

[0106] It can be seen that the present invention can improve the efficiency and accuracy of cervical cancer pathological image analysis by generating pathological image data.

[0107] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A two-stage generative adversarial network guided cervical cancer pathology image analysis method, characterized in that: The steps include: Collect real pathological images of cervical cancer and label them with real labels; A cervical cancer pathology data generation network is constructed based on a two-stage generative adversarial network, including a pathology image generation network and a pathology image label generation network; The cervical cancer pathology data generation network receives real pathology images of cervical cancer and adds real labels, and outputs generated pathology images and corresponding generated labels for joint training of downstream segmentation networks; Use the trained downstream segmentation network to segment cervical cancer pathology images and identify cancerous tissue, solid tissue, and background. Among them, both the pathological image generation network and the pathological image label generation network adopt generative adversarial networks; the pathological image generation network includes a fully connected layer, an image generator, and an image discriminator connected in sequence; the pathological image label generation network includes a pathological prior feature guided loss block, a label generator, and a label discriminator connected in sequence; The pathological image generation network receives a real pathological image and outputs a generated pathological image, and the process includes the following steps: After obtaining the random hidden variable z from the real pathological image and regularizing it, it is input into the fully connected layer to obtain the decoupled intermediate hidden vector s, which is expressed as: s = W × z + b; Among them, W is the weight coefficient and b is the bias vector; The intermediate hidden vector s is input into the image generator together with the real pathological image to generate features of different scales, which are input into the image discriminator for discrimination; the image generator uses the pathological style feature information fusion normalization module to add random noise y in the upsampling stage z , enrich the detailed information of generated pathological images; The image discriminator outputs a pathological image; The pathological image label generation network receives a real pathological image with a real label and a generated pathological image output by the pathological image generation network, and outputs a generated label, and the process includes the following steps: The pathology prior feature guided loss block receives the real pathology image with the real label and the generated pathology image output by the pathology image generation network, generates the pathology prior feature guided loss, and inputs it into the label generator; Using the confrontation between the label generator and the label discriminator, the label corresponding to the generated pathological image is generated; The pathological prior features guide the loss L ITL-A It is expressed as: L ITL-A =λ d L GC +λ e L GB +λ f L DFP ; Among them, λ d , e and λ f are the weights of cross entropy loss, boundary loss and distance feature penalty cross entropy loss respectively; L GC represents the cross entropy loss, L GB represents the boundary loss, L DFP represents the distance feature penalized cross entropy loss.

2. According to claim 1, the two-stage generative adversarial network guided cervical cancer pathology image analysis method is characterized in that: The pathological style feature information fusion normalization module is used to self-guide the feature map y w and random noise y z As input, the output is the pathological style feature information fusion normalized perception map y n , expressed as: and n =conv5(AdaFLN(and w ,and z )); AdaFLN=ρAdaIN+αAdaLN+(1-ρ-γ)AdaFN; Among them, conv5 is a 5×5 convolution operation; AdaFLN is an adaptive instance feature layer normalization operator, γ and ρ are learnable parameters used to constrain the feature vectors extracted by different normalization modules; AdaIN is an adaptive instance normalization operator; AdaLN is an adaptive layer normalization operator, l w represents the normalized input feature map of the adaptive layer, l z represents the input feature vector of the adaptive layer normalization; AdaFN is the adaptive feature normalization operator, f w represents the input feature map of adaptive feature normalization, f z represents the input feature vector of adaptive feature normalization; μ and σ represent the mean and variance operations, respectively.

3. The two-stage generative adversarial network guided cervical cancer pathology image analysis method according to claim 1, characterized in that: The pathological image generation network adopts a loss function based on CycleGan, which is expressed as: L ITI =λ a L GAN (G ITI ,D ITI ,X,Y)+λ b L Cycle (G ITI ,F)+λ c L Identity (G ITI ,F); where D ITI represents the loss function of the image discriminator, G ITI represents the loss function of the image generator, X represents the predicted label of the pathological image generation process, Y represents the true label of the pathological image generation process, F represents the true source domain of the pathological image, and L ITI is the loss function of the pathological image generation network, L GAN To combat the loss, L Cycle is the cycle consistency loss, L Identity is the identity loss, λ a , b and λ c denote the weights of adversarial loss, cycle consistency loss, and identity loss, respectively.

4. The two-stage generative adversarial network guided cervical cancer pathology image analysis method according to claim 1, characterized in that: In the pathological image generation network, the fully connected layer includes eight 512×512 depth-separable convolutional layers; the image generator uses an 8-layer ResNet50; and the image discriminator uses a 3-layer ResNet50.

5. The two-stage generative adversarial network guided cervical cancer pathology image analysis method according to claim 1, characterized in that: The cross entropy loss, boundary loss and distance feature penalty cross entropy loss are expressed as: Among them, x i Represents the i-th element of the true label, y i Represents the probability that the model predicts that the class belongs to the i-th category; φ α is the distance feature calculated by the distance transform of the true segmentation label, is the pixel-level probability value predicted by the model, α represents the index of each pixel in the image, and N GB is the total number of pixels in the image used for normalization; N DFP is the total number of pixels, c is the category of pixels, is the probability value of pixel β belonging to class c, is the true value of pixel β belonging to class c, D c is the distance penalty term for class c.

6. The two-stage generative adversarial network guided cervical cancer pathology image analysis method according to claim 5, characterized in that: The total loss function L of the pathological image label generation network ITL It is expressed as: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> ITL <h2 style=";text-align:left;direction:ltr"> =gL<h2 style=";text-align:left;direction:ltr"> ITI <h2 style=";text-align:left;direction:ltr"> +hL<h2 style=";text-align:left;direction:ltr"> ITL-A <h2 style=";text-align:left;direction:ltr"> ; Among them, L ITI is the loss function of the pathological image generation network, g represents the weight of the loss function of the pathological image generation network, and h represents the weight value of the loss guided by the pathological prior feature.

7. The two-stage generative adversarial network guided cervical cancer pathology image analysis method according to claim 1, characterized in that: In the pathological image label generation network, the label generator adopts 7-layer ResNet50; the label discriminator adopts 3-layer ResNet50.

8. The two-stage generative adversarial network guided cervical cancer pathology image analysis method according to claim 1, characterized in that: The downstream segmentation network is jointly trained with the cervical cancer pathology data generation network, and the loss function used is expressed as: Among them, x * ^ represents the predicted label of the downstream segmentation network, w d represents the trainable parameters in the downstream segmentation network, which is used to constrain the accuracy of the segmentation network prediction results. z represents the true label of the pathological image, y represents the predicted label of the pathological image generated after the true label constraint, and w e represents the trainable parameters in the generative network, which is used to constrain the accuracy of the generated results; θ represents the normalization of the segmentation network, Ф represents the regularization of the generative network, and L d is the regularization term of the downstream segmentation network, L e is the regularization term of the pathological data generation network, and λ3 is a balanced hyperparameter.

Images