Digital pathology image standardization system based on task constraint enhanced generative adversarial network
By constructing a large-scale source dataset and a multi-layered nested U-Net network, combined with a task-constrained adversarial generative network, the problems of insufficient template adaptation and resolution in digital pathological image standardization are solved, generating high-resolution standardized images and improving the performance of computer-aided pathological diagnosis systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2024-01-05
- Publication Date
- 2026-06-23
AI Technical Summary
Existing digital pathology image standardization technologies suffer from insufficient template adaptability and low image resolution, leading to a decline in the performance of computer-aided pathology diagnostic systems.
We employ a task-constrained enhanced adversarial generative network (PGN) to convert RGB images into grayscale images by constructing a large-scale source dataset. We then use a multi-layered nested U-Net generator and Markov discriminator, combined with adversarial loss, generated image difference loss, and task loss, to optimize the generator and discriminator and generate high-resolution normalized images.
It improves the adaptability and resolution of digital pathology images, ensuring that the generated images are adapted to computer-aided pathology diagnosis systems for specific tasks, thereby enhancing system performance.
Smart Images
Figure CN117877688B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical and health information technology, and in particular relates to a digital pathological image standardization system based on task-constrained enhanced adversarial generative networks. Background Technology
[0002] Histopathological image analysis is a key standard for the diagnosis of various cancers in clinical practice. Pathologists use biopsy methods to obtain diseased tissue from living patients for pathological examination, prepare pathological slides, and observe them under a microscope to make timely and accurate diagnoses and assess treatment efficacy. In the context of the digital revolution, digital slide scanners convert pathological slides into whole-slide images (WSI), facilitating subsequent applications in preservation, teaching, and remote slide reading. With the rapid development of digital pathology and computational pathology aided by scanning equipment, computing servers, and deep learning algorithms, research on digital pathological images of various diseases has achieved significant results in cancer diagnosis, subtype classification, and prognostic stratification, potentially helping clinicians provide better medical services. Meanwhile, deep learning algorithms are efficient and highly reproducible, saving pathologists considerable time and effort in histopathological image analysis. Therefore, many computer-aided pathology diagnostic systems are also applied in clinical practice based on deep learning models trained for different tasks. However, the data used in clinical practice and the data used in training models come from different medical institutions. Different medical institutions have different procedures, staining agents and scanners used in the preparation of digital pathology images, which leads to many differences in digital pathology images. These differences reduce the performance of computer-aided pathology diagnosis systems.
[0003] Currently, to address this issue, many studies have attempted to standardize digital pathology images prepared by different centers for clinical use and align them with the training data to improve the performance of computer-aided diagnostic systems. Among these, the technical solution most similar to that claimed in this patent is:
[0004] ① A staining separation method based on staining density maps. This technique, based on Beer-Lambert's law, separates the absorption density maps and color bases of different stains from digital pathological images. Then, it aligns the color base of the digital pathological image to be standardized with the color base of a carefully selected template image, thus completing the standardization process. Vahadane et al., based on the assumption that hematoxylin and eosin stains are linearly separable in the optical density color space, developed a pathological image standardization scheme based on sparse non-negative matrix factorization. This scheme can decompose different staining density maps and color bases while preserving the structural information of the original pathological image. However, firstly, different patients have different physical conditions, disease progression, and the location of tissue sections within the tumor. This results in differences in the tissue structure of the selected template digital pathological image and the tissue components contained in the digital pathological image to be standardized, making it impossible for the color base to be fully adapted. This leads to some information loss and changes in the standardized digital pathological image. Secondly, the separation process of staining absorption density maps and color bases only considers the color information of the image, while simultaneously breaking down the tissue structure, which is considered as a whole, into individual pixels for staining separation.
[0005] ② Style Transfer-Based Methods. This approach treats image standardization as a style transfer problem within the deep learning domain. It uses deep learning algorithms to learn the style of digital pathology images in the training dataset and then transfers this learned style to actual digital pathology images generated by medical institutions in other clinical practices. Salehi et al., using a pix2pix network architecture, minimized the generative adversarial loss and the difference loss between the generated and original images to learn the style of digital pathology images. They converted digital pathology images from different medical institutions in clinical practice into grayscale images and input them into the trained model to generate standardized digital pathology images. However, the first approach, using Generative Adversarial Networks (GANs), constrains the generated images by optimizing the discriminative and L1 losses, which cannot guarantee that the standardized images can be adapted to computer-aided pathology diagnostic systems for specific tasks. Secondly, the output image resolution of pix2pix and other GAN frameworks is not high. This is because higher resolution not only leads to an exponential increase in training time but also results in inaccuracies in the details of the conversion, while some computer-aided pathology diagnostic systems for specific tasks require high-resolution input images. Summary of the Invention
[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a digital pathology image standardization system based on task-constrained enhanced adversarial generative networks.
[0007] The objective of this invention is achieved through the following technical solution: a digital pathology image standardization system based on task-constrained enhanced adversarial generative networks, comprising:
[0008] The dataset building module is used to build source datasets based on the training set of digital pathological images of the deep learning model on which the computer-aided pathology diagnosis system for a specific task is based.
[0009] The data preprocessing module is used to clean the digital pathology images in the source dataset and convert the RGB images into grayscale images to obtain a grayscale dataset.
[0010] A generator module is used to input grayscale images of digital pathology images into a generator to generate standardized digital pathology images. The generator consists of multiple nested U-Net layers.
[0011] The discriminator module is used to determine whether a digital pathology image was generated by the generator module.
[0012] The task constraint module is used to input standardized digital pathology images and original digital pathology images into a deep learning model trained according to a specific task, and obtain the model calculation results.
[0013] The loss calculation module is used to calculate the loss based on the data obtained from the generator module, discriminator module, and task constraint module, and to optimize the generator module and discriminator module.
[0014] The loss includes adversarial loss, generated image difference loss, and task loss. The adversarial loss is used to measure the result of the discriminator's judgment. The generated image difference loss is the average of the absolute values of the differences between each pixel between the generated image and the original image. The task loss is the average of the squared differences between the generated image and the original image obtained through the deep learning model, calculated sequentially by position.
[0015] Furthermore, the input to the task-specific deep learning model f is a fixed-size digital pathology image patch, where each patch is an RGB image of size λ×λ, where λ is a multiple of 2, denoted as λ=2. m , m≥7.
[0016] Furthermore, in the data preprocessing module, the source dataset X = {x1, x2, ..., x...} is processed. n}∈R λ×λ×3×n Each digital pathology image block x i Converting an RGB image to a grayscale image results in the following grayscale value: L = 0.299 × R + 0.587 × G + 0.144 × B. The resulting grayscale dataset is denoted as X′ = {x′1, x′2, ..., x′}. n}∈R λ×λ×n , where n is the number of images in the dataset.
[0017] Furthermore, in the generator module, generator G will process the input digital pathological image grayscale image x′. i Converted into standardized digital pathology images Right now Obtain a standardized image dataset
[0018] Furthermore, the generator G consists of k nested U-Net layers, G = {G1, G2, ..., G...} k}, where G k The input is x′ i,k =x′ i ∈R λ×λ G k-1 The input is x′ i The image after half-sampling is Until the output of G1 is
[0019] In actual operation, first x′ i,1 Input G1 to get the output The obtained output is used as additional information to supplement G2 to obtain the output. Until the final output is obtained
[0020] Furthermore, in the discriminator module, discriminator D determines whether the input image pair is real or fake. For image pair (x′) i ,x i ) should be classified as real image pairs, for image pairs They should be classified as forged image pairs;
[0021] To achieve parallel computing, As input, where The concatenation operation involves concatenating two matrices along their third dimension to obtain the output result. The output dataset D of the discriminator module is obtained by calculating all datasets. real ={d 1,real ,d 2,real ,…,d n,real},D g ={d 1,g ,d 2,g ,…,d n,g}
[0022] Furthermore, the discriminator D uses a Markov discriminator as the discrimination model. The Markov discriminator uses a convolutional neural network to predict whether a local region of the input image is a real image. It maps an image block of size λ×λ to a judgment matrix of size p×p. Each block in the matrix is a real or fake label generated for a local region in the image. Finally, a fully connected layer is used to map the features to a value between 0 and 1 as the real or fake label of the entire image.
[0023] Furthermore, in the task constraint module, the source dataset X is input into the deep learning model f to obtain the result dataset, denoted as Y = f(X); the digital pathological images generated by the generator G are input into the deep learning model f to obtain the task result. Merge all generated results into the resulting dataset
[0024] Furthermore, in the loss calculation module, the calculation formulas for each loss are as follows:
[0025] L GAN =E[log(D) g )]+E[log(1-D real )]
[0026]
[0027]
[0028] Where L GAN To combat the losses, L image To generate image difference loss, L task Let E(·) represent the expected value of the task, and ∑|·| represent the summation of the absolute values of the differences between the pixel values of the two images. 2 This function represents the sum of the squared differences between the matrix generated by the generator and the matrix generated by the original image and the matrix generated by the deep learning model f, arranged sequentially by position;
[0029] Total loss L = L GAN +αL image +βL task , where α and β are hyperparameters controlling the weights;
[0030] The final generator is obtained through training iterations based on actual tasks. This means that the generator G minimizes the loss L, while the discriminator D maximizes the loss L.
[0031] The present invention also provides a digital pathology image standardization device based on task constraints and adversarial generative networks, comprising a memory and a processor, wherein the memory is used to store program data and the processor is used to execute the program data to implement the above-described digital pathology image standardization system based on task constraints and enhanced adversarial generative networks.
[0032] The beneficial effects of this invention are:
[0033] 1. Existing staining separation techniques based on staining density maps rely on specially selected digital pathology image templates. The limited number of these templates cannot accommodate all digital pathology images generated in clinical practice. Furthermore, the pixel-by-pixel breakdown of tissue structures in digital pathology images destroys the overall structural information. This invention, by constructing a large-scale source dataset, is more adaptable to the complex and ever-changing clinical practice compared to the limited number of digital pathology image templates. Simultaneously, converting digital pathology images into grayscale images as input to the generator preserves the structural information of the original digital pathology images.
[0034] 2. Existing style transfer-based solutions employ a pix2pix generative adversarial network framework, resulting in low-resolution output images that may not learn the image features required for specific tasks. This invention utilizes a multi-layered nested U-Net network architecture in the generator module, enabling the generation of higher-resolution digital pathology images. Furthermore, task constraints are introduced to enhance the generative adversarial network, ensuring that the generator learns the necessary image features for specific tasks, thus exhibiting strong universality. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 A structural diagram of a digital pathology image standardization system is shown as an exemplary embodiment.
[0037] Figure 2 A flowchart illustrating image normalization and loss calculation as an exemplary embodiment;
[0038] Figure 3 A structural diagram of a digital pathology image standardization device is shown as an exemplary embodiment. Detailed Implementation
[0039] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0040] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.
[0041] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.
[0042] This application provides a digital pathology image standardization system based on task-constrained enhanced adversarial generative networks. It is a multi-center digital pathology image standardization scheme designed for computer-aided pathology diagnosis systems targeting specific tasks, such as... Figure 1 As shown, the image standardization system mainly includes a dataset construction module, a data preprocessing module, a generator module, a discriminator module, a task constraint module, and a loss calculation module.
[0043] The dataset construction module collects the source datasets needed to optimize each module; the data preprocessing module cleans and transforms the digital pathology images in the source dataset into grayscale datasets that can be used as input for the generator module and for model training; the generator module generates standardized digital pathology images from the input grayscale digital pathology images; the discriminator module determines whether the output digital pathology images were generated by the generator module; the task constraint module, for a deep learning model trained according to a specific task, inputs standardized and original digital pathology images into the model to obtain the model's calculation results; and the loss calculation module optimizes the generator and discriminator modules by calculating the loss.
[0044] The following description further provides some embodiments of the implementation of each module of the digital pathology image standardization system based on task-constrained enhanced adversarial generative networks, which conforms to the requirements of this application.
[0045] I. Dataset Construction Module
[0046] Obtain the training set of a deep learning model for a specific task and organize it into a source dataset X, where X = {x1, x2, ..., x...} n} represents the digital pathology images of a deep learning model for a specific task, and there are a total of n images.
[0047] A deep learning model for a specific task can be represented as f, whose function is to obtain the target result Y = f(X), where Y = {y1, y2, ..., y}. n} represents the output result for each input.
[0048] Generally, the input to a deep learning model for a specific task is a fixed-size digital pathology image patch, then X∈R λ×λ×3×n Each digital pathology image patch is an RGB image of size λ×λ. To facilitate model processing, λ is usually a multiple of 2, denoted as λ=2. m (m≥7).
[0049] II. Data Preprocessing Module: The data preprocessing module performs the following steps:
[0050] Each digital pathology image patch x in the source dataset X i Converting an RGB image to a grayscale image, the resulting grayscale value is L = 0.299 × R + 0.587 × G + 0.144 × B. The converted dataset is denoted as X′ = {x′1, x′2, ..., x′}. n}∈R λ×λ×n .
[0051] Combine X, X′, and Y to form a training dataset Z = {(x1, x′1, y1), (x2, x′2, y2), ..., (x n ,x′ n ,y n )}.
[0052] III. Generator Module: The generator module performs the following steps:
[0053] The generator G will take the input digital pathology image patch x′ i Converted into standardized digital pathology images Right now
[0054] Specifically, the generator consists of k nested U-Nets (U-shaped Networks), where G = {G1, G2, ..., G...} k}. Among them, G k The input is x′ i,k =x′ i ∈R λ×λ G k-1 The input is x′ i The image after half-sampling is Until the output of G1 is Therefore, we can calculate k = log₂λ - 7 + 1 = m - 6. In actual operation, we first set x′ i,1 Input G1 to get the output The obtained output is used as additional information to supplement G2 to obtain the output. The final output result is obtained by repeating this process. Generate a standardized image dataset using a generator. The generated standardized image dataset is used in the discriminator module, task constraint module, and loss calculation module. The generator architecture adopts a multi-layer U-Net, with each U-Net consisting of an encoder and a decoder. The encoder encodes the input image into a low-dimensional feature vector, and the decoder decodes this feature vector into a high-resolution output image. The encoding process involves progressively downsampling the input image, saving the resulting feature map after each downsampling. The decoding process involves progressively upsampling the corresponding encoding process to restore it to its original size. Each upsampling operation acquires the corresponding downsampled feature map. At this point, the output images of other low-level U-Nets are used as additional feature maps and stitched into the corresponding decoding process to supplement the image details.
[0055] IV. Discriminator Module: The discriminator module performs the following steps:
[0056] Discriminator D determines whether the input image pair is real or fake. For image pair (x... i ′ ,x i ) should be classified as real image pairs, for image pairs These should be classified as forged image pairs. In practice, to achieve parallel processing, [the following will be implemented]... As input, where This is a concatenation operation, where two matrices are concatenated along their third-dimensional axis. The output result can then be obtained. Calculating all datasets yields the output dataset D of the discriminator module. real ={d 1,real ,d 2,real ,…,d n,real},D g ={d 1,g ,d 2,g ,…,d n,g}, where D real D is the set of discrimination results for real images. g The result set for identifying forged images is used in the loss calculation module.
[0057] Specifically, the discriminator uses a Markov discriminator (Patch-based Discriminator Generative Adversarial Networks, PatchGAN) as the discrimination model, but is not limited to this. PatchGAN uses a convolutional neural network to predict whether local regions of the input image are real images. It maps λ×λ image patches to p×p judgment matrices. Each patch in the matrix is a real or fake label generated for a local region in the image. Finally, a fully connected layer is used to map the features to a value between 0 and 1 as the real or fake label for the entire image. Using PatchGAN can flexibly adapt to images of different sizes in different tasks, while better capturing the local structure and detail information of the image.
[0058] V. Task Constraint Module: The task constraint module completes the following steps:
[0059] The task constraint module is used to input the generated image into a deep learning model for a specific task to obtain the task result. Merge all generated results into a generated results dataset The generated dataset is used in the loss calculation module.
[0060] VI. Loss calculation module, such as Figure 2 As shown, the loss calculation module completes the following steps:
[0061] The loss calculation module is used to calculate the loss L based on the data obtained from the generator module, discriminator module, and task constraint module. The loss L is obtained from three parts, the first part being the adversarial loss L. GAN The second part is the generated image difference loss L. image The third part is the task loss L. task .
[0062] Specifically, the first part addresses the loss L. GAN The formula used to measure the result of the discriminator's judgment is as follows:
[0063] L GAN (G,D)=E[log(D g )]+E[log(1-D real )]
[0064] Where E(·) represents the mathematical expectation.
[0065] Part Two: Image Difference Loss L image The formula for averaging the absolute values of the differences between each pixel in the generated image and the original image is as follows:
[0066]
[0067] Where ∑|·| represents a function that sums the absolute values of the differences between the pixel values of two images.
[0068] Part Three: Mission Loss L task The formula is as follows:
[0069]
[0070] Where ∑(·) 2 This function represents the sum of the squared differences between the result matrix generated by the task constraint module and the result matrix of the original image after passing through the deep learning model f, according to their positions.
[0071] After calculating the losses from the three parts, they can be combined to obtain the final loss, as shown in the following formula:
[0072] L = L GAN +αL image +βL task
[0073] Where α and β are hyperparameters that control the weights, and are adjusted during training according to the actual task.
[0074] Finally, the generator is obtained through training iterations based on actual tasks. This means that the generator G minimizes the loss L, while the discriminator D maximizes the loss L.
[0075] This invention constructs a source dataset based on the training set of a deep learning model used in a task-specific computer-aided pathology diagnosis system. Data preprocessing methods from this system can be incorporated into the dataset construction module. The invention converts RGB images to grayscale as input to the generator module, preserving the tissue structure features of the original digital pathology images. By introducing a pix2pixHD adversarial generative network framework, this invention learns the style of digital pathology images from the dataset used to train the deep learning model, achieving end-to-end digital pathology image standardization without requiring the selection of a specific template image. Based on a multi-layered nested U-Net network framework, this invention can generate high-resolution standardized digital pathology images.
[0076] This invention introduces task constraints, using the trained deep learning model of the computer-aided pathology diagnosis system for a specific task as an additional discrimination module. Additional task loss is introduced into the loss function for task constraints, so that the standardized digital pathology image retains as many key features as possible in the original image required for the specific task. This ensures that the generated standardized image can be adapted to the computer-aided pathology diagnosis system for the specific task, and the trained generator can also be closely integrated with the computer-aided pathology diagnosis system.
[0077] Corresponding to the aforementioned embodiments of the digital pathology image standardization system based on task-constrained enhanced adversarial generative networks, the present invention also provides embodiments of a digital pathology image standardization device based on task constraints and adversarial generative networks.
[0078] See Figure 3 The digital pathology image standardization device based on task constraints and adversarial generative networks provided in this embodiment includes one or more processors for implementing the digital pathology image standardization system based on task constraints and enhanced adversarial generative networks in the above embodiment.
[0079] The embodiments of the digital pathology image standardization device based on task constraints and adversarial generative networks of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 3 The diagram shown is a hardware structure diagram of any device with data processing capabilities, including the digital pathology image standardization device based on task constraints and adversarial generative networks according to the present invention. (Except for...) Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.
[0080] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.
[0081] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0082] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the digital pathology image standardization system based on task-constrained enhanced adversarial generative networks described in the above embodiments.
[0083] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.
[0084] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.
[0085] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. A digital pathology image standardization system based on task-constrained enhanced adversarial generative networks, characterized in that, include: The dataset building module is used to build source datasets based on the training set of digital pathological images of the deep learning model on which the computer-aided pathology diagnosis system for a specific task is based. The input to the deep learning model is a fixed-size digital pathology image patch, where each digital pathology image patch is of size [size missing]. RGB image, Take multiples of 2 and denote them. , ; The data preprocessing module is used to clean the digital pathology images in the source dataset and convert the RGB images into grayscale images to obtain a grayscale dataset. The generator module is used to input grayscale images of digital pathology images into the generator to generate standardized digital pathology images. The generator G consists of... Composed of nested U-Net layers, ,in The input is , The input is The image after half-sampling is ,until The output is ;in, Let i be the grayscale image of the i-th digital pathology image; In actual operation, first enter Get output The obtained output is used as additional information to supplement the input. Get output until the final output is obtained. ; The discriminator module is used to determine whether a digital pathology image was generated by the generator module. The task constraint module is used to input standardized digital pathology images and original digital pathology images into a deep learning model trained according to a specific task, and obtain the model calculation results. The loss calculation module is used to calculate the loss based on the data obtained from the generator module, discriminator module, and task constraint module, and to optimize the generator module and discriminator module. The loss includes adversarial loss, generated image difference loss, and task loss. The adversarial loss is used to measure the result of the discriminator's judgment. The generated image difference loss is the average of the absolute values of the differences between each pixel between the generated image and the original image. The task loss is the average of the squared differences between the generated image and the original image obtained through the deep learning model, calculated sequentially by position.
2. The digital pathology image standardization system based on task-constrained enhanced adversarial generative networks according to claim 1, characterized in that, In the data preprocessing module, the source dataset is... Digital pathology image blocks Converting an RGB image to a grayscale image, resulting in grayscale values. The converted grayscale image dataset is denoted as ,in This represents the number of images in the dataset.
3. The digital pathology image standardization system based on task-constrained enhanced adversarial generative networks according to claim 2, characterized in that, In the generator module, generator G will process the input digital pathology image grayscale image. Converted into standardized digital pathology images ,Right now To obtain a standardized image dataset .
4. The digital pathology image standardization system based on task-constrained enhanced adversarial generative networks according to claim 1, characterized in that, In the discriminator module, discriminator D determines whether the input image pair is real or fake. They should be classified as real image pairs. They should be classified as forged image pairs; To achieve parallel computing, As input, where The concatenation operation involves concatenating two matrices along their third dimension to obtain the output result. The output dataset of the discriminator module is obtained by calculating all datasets. .
5. The digital pathology image standardization system based on task-constrained enhanced adversarial generative networks according to claim 1, characterized in that, The discriminator D employs a Markov discriminator as its discrimination model. The Markov discriminator uses a convolutional neural network to predict whether a local region of the input image is a real image. Image patches of size are mapped as The judgment matrix is used, where each block in the matrix is used to judge the true or false labels of local regions in the image. Finally, a fully connected layer is used to map the features to a value between 0 and 1, which serves as the true or false label for the entire image.
6. The digital pathological image standardization system based on task-constrained enhanced adversarial generative networks according to claim 4, characterized in that, In the task constraint module, the source dataset The result dataset obtained by inputting a deep learning model f is denoted as f. The digital pathological images generated by generator G are input into the deep learning model f to obtain the task results. Merge all generated results into the final dataset. .
7. The digital pathology image standardization system based on task-constrained enhanced adversarial generative networks according to claim 6, characterized in that, In the loss calculation module, the calculation formulas for each loss are as follows: in To combat the losses, To generate image difference loss, For mission losses, Represents the mathematical expectation. This function represents the summation of the absolute values of the differences between the pixel values of two images. This indicates that the image generated by the generator is processed by a deep learning model. The resulting matrix and the original image are processed by a deep learning model The result matrix is a function that sums the squares of the differences in order of position; Total loss ,in Hyperparameters for controlling weights; The final generator is obtained through training iterations based on actual tasks. This means that the generator G minimizes the loss L, while the discriminator D maximizes the loss L.
8. A digital pathology image standardization device based on task constraints and adversarial generative networks, comprising a memory and a processor, characterized in that, The memory is used to store program data, and the processor is used to execute the program data to implement the digital pathology image standardization system based on task-constrained enhanced adversarial generative networks as described in any one of claims 1-7.