Self-supervised dyeing standardization method based on de-noising diffusion probability model

Abstract

The invention provides a self-supervised staining standardization method based on a de-noising diffusion probability model, which is mainly used for solving the problems of low robustness and generalization when a computer-aided system analyzes a histopathological image, and information loss, staining errors and the like existing in a current staining standardization algorithm. Compared with the prior art, by means of the denoising diffusion probability model and the self-supervising training method, staining standardization of histopathology images of various staining styles is achieved, the cell morphological structure is reserved, the accuracy of disease diagnosis of a computer-aided system in downstream tasks such as classification and segmentation is improved, and the method is suitable for popularization and application. And clinical practice and medical research promotion are facilitated.

Description

Technical Field

[0001] The invention relates to the technical field of histopathology image staining standardization, and in particular to a self-supervised staining standardization method based on a denoising diffusion probability model. Background Art

[0002] With the development of computer-aided detection / diagnosis, histopathological images have become increasingly important for the diagnosis and prognosis of cancer. However, due to differences in staining techniques, operating skills, and scanner specifications, histopathological images present different staining styles. These diverse staining styles reduce the robustness of computer-aided detection / diagnosis algorithms.

[0003] In recent years, staining standardization methods are limited to traditional staining vector decomposition, learning-based generative adversarial networks, and their variants. Traditional staining standardization methods usually require medical professionals to screen staining templates, which takes a lot of time and effort. Among the learning-based methods, supervised methods require paired images, which are difficult to obtain in clinical applications. Unsupervised methods often have problems with staining errors and changes in cell structure. Self-supervised methods are mostly used to map grayscale images to color images, which have problems with poor generalization ability and information loss. Summary of the invention

[0004] In view of the shortcomings of existing staining standardization technology, the purpose of the present invention is to provide a self-supervised staining standardization method based on a denoising diffusion probability model. The method systematically converts the staining normalization task into a self-supervised pixel-aligned color mapping, avoids the need for paired data, adapts to a variety of staining styles, and can be applied to different staining scenarios without retraining. At the same time, a re-planning sampling algorithm is designed, which has the characteristics of fast speed, high accuracy, and no destruction of cell morphological structure. It provides a robust and effective solution for the staining standardization of digital tissue pathology images, which is beneficial to improve the accuracy of disease diagnosis and subsequent analysis, and promote clinical practice and medical research.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] A self-supervised coloring normalization method based on a denoising diffusion probability model, the method comprising the following steps:

[0007] Obtain different stained tissue pathology images and divide the target staining data set and other staining data sets;

[0008] Preprocess the other acquired staining data sets and integrate the extracted staining vectors into a staining matrix database;

[0009] Design a self-supervised dye standardization training strategy and use the dye vectors in the dye matrix database to perform dye enhancement preprocessing on the target dye dataset;

[0010] Construct a dye normalization model based on the denoising diffusion probability model, and use the preprocessed dataset to perform self-supervised training on the denoising diffusion probability model;

[0011] Determine whether the training has achieved the expected effect. If so, use the renormalization sampling strategy to sample the tissue pathology images on the trained model to obtain the dyeing standardization results; otherwise, readjust the self-supervised dyeing standardization training strategy and retrain the model using the adjusted strategy until a satisfactory standardization effect is achieved.

[0012] According to a self-supervised staining standardization method based on a denoising diffusion probability model provided by the present invention, the acquired tissue pathology image data set is sorted, including:

[0013] Traverse all images in the acquired tissue pathology image data set and divide the tissue pathology images with different staining styles;

[0014] Filter the target stained images of a staining style and store them as data set 0, and store the other stained images as data set 1.

[0015] According to a self-supervised dyeing standardization method based on a denoising diffusion probability model provided by the present invention, the preprocessing of the other dyeing data sets obtained includes:

[0016] For each tissue pathology image in dataset 1, the RGB color space of the image is mapped to the LAB color space, and the image is thresholded on the brightness channel to obtain the tissue area pixel mask;

[0017] The RGB color space of the image is mapped to the optical density space, and the optical density image is multiplied by the tissue area pixel mask to remove the area with almost no staining to highlight the staining range of interest;

[0018] The truncated image is flattened into a vector, and then singular value decomposition is used to obtain two vertical coloring matrices. The coloring matrix of each image in dataset 1 constitutes the coloring matrix database.

[0019] According to a self-supervised coloring standardization method based on a denoising diffusion probability model provided by the present invention, the self-supervised coloring standardization training strategy is designed, comprising:

[0020] Randomly select a staining matrix from the staining matrix database;

[0021] For a randomly selected staining matrix, a perturbation is added;

[0022] For the perturbed coloring matrix, the K nearest neighbor algorithm is used in the coloring matrix database to search for samples within the Euclidean distance of 0.1, and the number of neighbors searched is ensured to be greater than 5, otherwise the coloring matrix is ​​returned to the coloring matrix database and reselected for enhancement;

[0023] For the input image, calculate its staining matrix;

[0024] Correcting the coloring matrix of the input image to the perturbed coloring matrix to obtain a coloring enhanced image;

[0025] Gaussian blur is added to the dye-enhanced image to obtain the final dye-enhanced preprocessed image.

[0026] According to a self-supervised color normalization method based on a denoising diffusion probability model provided by the present invention, a disturbance is added to a randomly selected color matrix, which is expressed as the following formula:

[0027] M′=M×a+b

[0028] Where M is a randomly selected coloring matrix, M′ is the coloring matrix after adding perturbations, a is a random number ranging from 0.3 to 1.7, and b is a random number ranging from -0.7 to 0.7.

[0029] According to a self-supervised coloring normalization method based on a denoising diffusion probability model provided by the present invention, the construction of the denoising diffusion probability model includes:

[0030] Construct a forward noise adding process, and gradually add noise to the image through a parameterized Markov chain until its distribution is close to the standard Gaussian distribution;

[0031] Construct a UNet network for predicting noise, which receives a noisy image x t , the input of the image to be standardized x and the current time t, and the output of the predicted noise under the current state

[0032] Construct a backward denoising process by predicting random noise and the noisy image x t , and through reverse reasoning, we get the noisy image x at the next moment t-1 , until the noise is completely removed and the stained and standardized image is obtained.

[0033] According to a self-supervised color normalization method based on a denoising diffusion probability model provided by the present invention, noise is gradually added to the image through a parameterized Markov chain until its distribution is close to a standard Gaussian distribution, which is expressed as the following formula:

[0034]

[0035] Among them, x0 represents the clean image without noise, x t is the noisy image at time t, ∈ is the random noise sampled from a normal distribution, is the hyperparameter of the denoising diffusion probability model at time t.

[0036] According to a self-supervised coloring normalization method based on a denoising diffusion probability model provided by the present invention, the step of judging whether the training meets the expected effect includes:

[0037] When comparing the performance of the dyed standard image generated by the training model with the actual required dyed standardized image in terms of structural similarity SSIM, peak signal-to-noise ratio PSNR, and learning-perceptual image block similarity LPIPS, it is expressed as the following formula:

[0038]

[0039]

[0040]

[0041]

[0042] Where μ is the mean, σ is the standard deviation, and σ xy is the covariance of image x and image y, MSE is the mean square error of the two images, and MaxValue is the maximum value that the image pixel can take. and is the i-th layer feature extracted by the pre-trained neural network, w l is the corresponding weight vector.

[0043] The dyeing normalization results are applied to related downstream tasks such as classification and segmentation. The corresponding indicators of the computer-aided detection / diagnosis system for the dyeing normalized images are better than those for the non-dying normalized images.

[0044] According to a self-supervised staining normalization method based on a denoising diffusion probability model provided by the present invention, the renormalization sampling strategy is used to sample the tissue pathology image, comprising:

[0045] Add a certain degree of noise to the image to be standardized and use it as the initial image in the backward denoising process;

[0046] Through the non-Markov chain method, some steps in the backward denoising process are skipped, so as to quickly and accurately obtain the stained normalized image of the denoised diffusion probability model.

[0047] According to a self-supervised color normalization method based on a denoising diffusion probability model provided by the present invention, some steps in the process of jumping backward denoising are expressed as the following formula:

[0048]

[0049] Among them, x prev prev indicates that there can be multiple steps between t.

[0050] It can be seen that, therefore, compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. Automation: Traditional histopathology image staining standardization usually requires professionals to manually select staining templates based on the staining effect. The present invention makes full use of the denoising diffusion probability model to achieve staining standardization of differently stained histopathology images, thereby significantly reducing the burden on human resources and accelerating the entire process of histopathology image analysis by introducing automated processing.

[0052] 2. Accuracy: Through sufficient training and learning of the self-supervised staining standardization training strategy, the present invention has a higher level of ability to standardize the staining of histopathology images and preserve the morphological structure. Compared with the existing methods, the present invention more accurately identifies the key information in histopathology images, thereby providing more reliable staining standardization.

[0053] 3. High efficiency: The present invention adopts an efficient re-planning sampling algorithm to solve the shortcoming of the slow speed of the denoising diffusion probability model. The speed is increased by 120 times compared with the initial sampling process of the denoising diffusion probability model, reaching a speed comparable to that of the traditional method. This real-time analysis process provides strong support for rapid diagnostic decisions.

[0054] 4. Decision support: The present invention standardizes the staining of histopathological images with different staining styles, thereby solving the problem of poor generalization ability of computer-aided detection / diagnosis systems for histopathological images with different staining styles. By combining the analysis results of the computer-aided detection / diagnosis system with the experience and judgment of the doctor, the accuracy of disease diagnosis and prognosis analysis and the effectiveness of treatment can be improved, which helps to make more accurate diagnosis and treatment decisions.

[0055] 5. Reduced reliance on paired data: Self-supervised learning can use unpaired data for training, thereby reducing reliance on a large number of paired, histopathological images with different staining styles. Through self-supervised learning, the model can learn the inherent laws and features in the data, so that it can show better generalization ability when faced with new, unseen data.

[0056] 6. Powerful generation capability: The denoising diffusion probability model (DDM) performs well in the field of image generation and can generate high-quality, realistic images. This generation capability provides a solid foundation for staining standardization, enabling the model to better simulate and restore the target staining style. DDM trains and learns the data by gradually adding and removing noise, so that the model can better adapt to pathological images with different staining styles and improve the accuracy and robustness of staining standardization.

[0057] 7. The present invention designs a self-supervised staining standardization training strategy and dynamically adjusts it according to the training effect, so that the model can adapt to different data distributions and staining styles to achieve better standardization effects. The renormalization sampling strategy is used on the trained model to sample the tissue pathology images, and high-precision staining standardization results can be obtained. This sampling strategy can ensure that the standardized images have a consistent staining style while maintaining the original tissue structure.

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of an embodiment of a self-supervised coloring normalization method based on a denoising diffusion probability model of the present invention.

[0060] Figure 2 It is a flow chart of a self-supervised coloring standardization training strategy for an algorithm model in an embodiment of a self-supervised coloring standardization method based on a denoising diffusion probability model of the present invention.

[0061] Figure 3 It is a schematic diagram of the overall framework of the algorithm model in an embodiment of a self-supervised coloring standardization method based on a denoising diffusion probability model of the present invention.

[0062] Figure 4 It is a schematic diagram of the effect of data set 1 in an embodiment of a self-supervised coloring normalization method based on a denoising diffusion probability model of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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.

[0064] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0065] See also Figures 1 to 4 The present invention provides a self-supervised coloring standardization method based on a denoising diffusion probability model, the method comprising the following steps:

[0066] Step S1, obtaining different stained tissue pathology images, and dividing the target staining data set and other staining data sets;

[0067] Step S2, preprocessing the other acquired staining data sets, integrating the extracted staining vectors into a staining matrix database;

[0068] Step S3, designing a self-supervised dyeing standardization training strategy, and using the dyeing vectors in the dyeing matrix database to perform dyeing enhancement preprocessing on the target dyeing dataset;

[0069] Step S4, constructing a dyeing standardization model based on the denoising diffusion probability model, and using the preprocessed data set to perform self-supervised training on the denoising diffusion probability model;

[0070] Step S5, determine whether the training achieves the expected effect, if so, execute step S6.

[0071] Step S6, sampling the tissue pathology image using the renormalization sampling strategy on the trained model to obtain the dyeing standardization result; otherwise, readjust the self-supervised dyeing standardization training strategy, and retrain the model using the adjusted strategy until a satisfactory standardization effect is achieved.

[0072] In the above step S1, the acquired tissue pathology image data set is sorted, including:

[0073] Step S11, traversing all images in the acquired tissue pathology image data set, and dividing the tissue pathology images with different staining styles due to dyes, staining procedures, digital scanners, etc.;

[0074] Step S12, selecting a target dyed image of a dyeing style and storing it as data set 0, and storing other dyed images as data set 1.

[0075] In the above step S2, the other obtained staining data sets are preprocessed, including:

[0076] Step S21, for each tissue pathology image in data set 1, the RGB color space of the image is mapped to the LAB color space, and the image is truncated with a threshold (e.g., 0.8) on the brightness channel to obtain a tissue region pixel mask;

[0077] Step S22, mapping the RGB color space of the image to the optical density space, multiplying the optical density image by the tissue region pixel mask, removing the area with almost no staining to highlight the staining range of interest;

[0078] Optical density space is commonly referred to as OD value. OD value conforms to Lambert-Beer law and is directly related to the quality of the dye. The more the amount of the substance being measured, the higher the OD value, the less light is transmitted, and the darker it is reflected in the photo. Optical density space can better reflect the staining distribution in tissue pathology images than RGB color space. The mathematical expression of Lambert-Beer law and RGB color space conversion to optical density space is:

[0079]

[0080] OD=-log 10 (I) (2)

[0081] Where A is the absorbance, T is the transmittance ratio (transmittance, transmitted light intensity to incident light intensity), K is the molar absorption coefficient, c is the concentration of the absorbing substance, L is the thickness of the absorption layer, and I is the RGB vector of the image.

[0082] Step S23, flattening the truncated image into a vector, and then using singular value decomposition to obtain two vertical coloring matrices, the coloring matrix of each image in the data set 1 constitutes a coloring matrix database.

[0083] In the above step S3, a self-supervised coloring standardization training strategy is designed, including:

[0084] Randomly select a staining matrix from the staining matrix database;

[0085] For a randomly selected staining matrix, a perturbation is added, expressed as the following formula:

[0086] M′=M×a+b (3)

[0087] Where M is a randomly selected coloring matrix, M′ is the coloring matrix after adding perturbations, a is a random number ranging from 0.3 to 1.7, and b is a random number ranging from -0.7 to 0.7.

[0088] For the perturbed coloring matrix, the K nearest neighbor algorithm is used in the coloring matrix database to search for samples within the Euclidean distance of 0.1, and the number of neighbors searched is ensured to be greater than 5, otherwise the coloring matrix is ​​returned to the coloring matrix database and reselected for enhancement;

[0089] For the input image, calculate its staining matrix;

[0090] The coloring matrix of the input image is corrected to the perturbed coloring matrix to obtain a coloring enhanced image. This process can be expressed by the following formula:

[0091]

[0092] Among them, 99 represents the 99th percentile, M x is the coloring matrix of the input image, M′ is the coloring matrix after random selection and adding perturbations, M′ x is the coloring matrix after correction of the input image.

[0093] Gaussian blur is added to the dye-enhanced image to obtain the final dye-enhanced preprocessed image.

[0094] In the above step S4, the construction of the denoising diffusion probability model includes:

[0095] Construct a forward denoising process. The purpose of denoising is to enable the denoising diffusion probability model to generate target data samples from noise. The image is gradually added with noise through a parameterized Markov chain until its distribution is close to the standard Gaussian distribution, which is expressed as the following formula:

[0096]

[0097] Among them, x0 represents the clean image without noise, x t is the noisy image at time t, ∈ is the random noise sampled from a normal distribution, is the hyperparameter of the denoising diffusion probability model at time t.

[0098] Construct a UNet network for predicting noise, which receives a noisy image x t , the input of the image to be standardized x and the current time t, and the output of the predicted noise under the current state

[0099] Construct a backward denoising process by predicting random noise and the noisy image x t , and through reverse reasoning, we get the noisy image x at the next moment t-1 , until the noise is completely removed to obtain the stained and standardized image, the formula is:

[0100]

[0101] where z is random noise sampled from a normal distribution, σ t , α t and is the hyperparameter of the denoising diffusion probability model at time t, and gradually obtains the image x at the next moment t-1 Finally, a clean stained standardized image is restored.

[0102] In the above step S5, judging whether the training meets the expected effect includes:

[0103] When comparing the performance of the dyed standard image generated by the training model with the actual required dyed standardized image in terms of structural similarity SSIM, peak signal-to-noise ratio PSNR, and learning-perceptual image block similarity LPIPS, it should be better than the current existing methods and achieve good results from visual subjective evaluation, which can be expressed as the following formula:

[0104]

[0105]

[0106]

[0107]

[0108] Where μ is the mean, σ is the standard deviation, and σ xy is the covariance of image x and image y, MSE is the mean square error of the two images, and MaxValue is the maximum value that the image pixel can take. and is the lth layer feature extracted by the pre-trained neural network, w l is the corresponding weight vector.

[0109] The dyeing normalization results are applied to downstream tasks such as classification and segmentation. The corresponding indicators of the computer-aided detection / diagnosis system for the dyeing normalized images, such as the classification accuracy, which is used to indicate the percentage of the number of correctly predicted samples in all samples, and the segmentation Dice Similarity Coefficient (DSC), which is a metric function used to evaluate the similarity between two samples, are better than those of the unstained normalized images, and are expressed as the following formula:

[0110]

[0111]

[0112] Among them, TP is the number of samples that are actually positive and predicted to be positive; FP is the number of samples that are actually negative but predicted to be positive; TN is the number of samples that are actually negative and predicted to be negative; FN is the number of samples that are actually positive but predicted to be negative. In the classification task, the samples are images, and in the segmentation task, the samples are pixels of the image.

[0113] Furthermore, in the above step S6, the tissue pathology image is sampled using a renormalization sampling strategy for staining standardization, including:

[0114] A certain degree of noise is added to the image to be standardized and used as the initial image in the backward denoising process. The formula is:

[0115]

[0116] Among them, x N It represents the initial image in the backward denoising process, and N represents the initial sampling time step selected by experience, that is, a certain degree of noise.

[0117] By using the non-Markov chain method, some steps in the backward denoising process are skipped, so as to quickly and accurately obtain the stained normalized image of the denoised diffusion probability model, which is expressed as the following formula:

[0118]

[0119] Among them, x prev prev indicates that there can be multiple steps between t.

[0120] The sampling strategy of the original denoising diffusion probability model and the re-planning sampling strategy selected their respective optimal parameters for test set verification. After testing, the relevant data comparison is shown in the following table (1):

[0121]

[0122] Table (1)

[0123] As shown in Table (1), after the re-planning sampling strategy of the present invention, SSIM, PSNR and LPIPS are improved, and the inference speed is increased by 120 times. This shows that the present invention can achieve better color standardization performance and effect with less time complexity.

[0124] The best parameters obtained by training for the present invention in breast cancer classification and colorectal gland segmentation were selected for test set verification. After testing, the relevant data comparison is shown in the following table (2) (3):

[0125]

[0126] Table (2)

[0127]

[0128] Table (3)

[0129] It can be seen from the above tables (2) and (3) that the self-supervised staining standardization method based on the denoising diffusion probability model of the present invention has been improved in both breast cancer classification and colorectal gland segmentation, indicating that the present invention has achieved good staining standardization effects and improved the accuracy of disease diagnosis in computer-aided systems in downstream tasks such as classification and segmentation, which is beneficial to clinical practice and the advancement of medical research.

[0130] In summary, compared with the prior art, the present embodiment includes the following steps through a self-supervised staining standardization method based on a denoising diffusion probability model: First, after collecting tissue pathology image data of different staining styles, the image is divided into a target staining style part and other staining style parts. Then, the acquired other style image data sets are preprocessed, including converting the optical density space to intercept the staining area of ​​interest, extracting the staining matrix features, and finally forming a staining matrix database for training. In the preprocessing stage of model training, a staining matrix is ​​randomly selected from the staining matrix database, and disturbances are added. The K nearest neighbor algorithm is used to determine whether the staining matrix is ​​in line with reality, and the staining matrix of the input image is corrected to the perturbed staining matrix, and then Gaussian blur is added to obtain the final preprocessing result. In the model construction stage, a denoising diffusion probability algorithm model is adopted. The overall framework of the model includes a forward denoising process, a UNet network for predicting noise, and a backward denoising process. After training, the performance of the model is evaluated, including structural similarity SSIM, peak signal-to-noise ratio PSNR, learning-perceptual image block similarity LPIPS and inference time, to determine whether the training has achieved the expected results. The results of dyeing standardization can be further applied to related downstream tasks such as classification and segmentation. The stained and standardized images are processed accordingly using the computer-aided detection / diagnosis system to determine whether the recognition ability of the computer system has improved. If the results are in line with expectations, the self-supervised dyeing model based on the denoising diffusion probability model is deployed on the histopathology image processing platform. On the platform, upload histopathology images with different dyeing styles, and then use the trained model to dye and standardize the images, and finally obtain a dyeing standardized image with a unified style. The goal of this overall process is to improve the accuracy of disease detection and diagnosis of computer-aided systems in downstream tasks such as classification and segmentation, and to promote clinical practice and medical research.

[0131] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A self-supervised coloring normalization method based on a denoising diffusion probability model, characterized in that: The method comprises the following steps: Obtain different stained tissue pathology images and divide the target staining data set and other staining data sets; Preprocess the other acquired staining data sets and integrate the extracted staining vectors into a staining matrix database; Design a self-supervised dye standardization training strategy and use the dye vectors in the dye matrix database to perform dye enhancement preprocessing on the target dye dataset; Construct a dye normalization model based on the denoising diffusion probability model, and use the preprocessed dataset to perform self-supervised training on the denoising diffusion probability model; Determine whether the training has achieved the expected effect. If so, use the renormalization sampling strategy to sample the tissue pathology images on the trained model to obtain the dyeing standardization results; otherwise, readjust the self-supervised dyeing standardization training strategy and use the adjusted strategy to retrain the model until a satisfactory standardization effect is achieved.

2. The method according to claim 1, characterized in that Organize the acquired tissue pathology image datasets, including: Traverse all images in the acquired tissue pathology image data set and divide the tissue pathology images with different staining styles; Filter the target stained images of a staining style and store them as data set 0, and store the other stained images as data set 1.

3. The method according to claim 1, characterized in that The preprocessing of the other obtained staining data sets includes: For each tissue pathology image in dataset 1, the RGB color space of the image is mapped to the LAB color space, and the image is thresholded on the brightness channel to obtain the tissue area pixel mask; The RGB color space of the image is mapped to the optical density space, and the optical density image is multiplied by the tissue area pixel mask to remove the area with almost no staining to highlight the staining range of interest; The truncated image is flattened into a vector, and then singular value decomposition is used to obtain two vertical coloring matrices. The coloring matrix of each image in dataset 1 constitutes the coloring matrix database.

4. The method according to claim 1, characterized in that: The designed self-supervised dyeing standardization training strategy includes: Randomly select a staining matrix from the staining matrix database; For a randomly selected staining matrix, a perturbation is added; For the perturbed coloring matrix, the K nearest neighbor algorithm is used in the coloring matrix database to search for samples within the Euclidean distance of 0.1, and the number of neighbors searched is ensured to be greater than 5, otherwise the coloring matrix is ​​returned to the coloring matrix database and reselected for enhancement; For the input image, calculate its staining matrix; Correcting the coloring matrix of the input image to the perturbed coloring matrix to obtain a coloring enhanced image; Gaussian blur is added to the dye-enhanced image to obtain the final dye-enhanced preprocessed image.

5. The method according to claim 4, characterized in that For a randomly selected staining matrix, a perturbation is added, expressed as the following formula: M ′ =M×a+b Where M is a randomly selected coloring matrix, M ′ is the coloring matrix after adding disturbance, a is a random number ranging from 0.3 to 1.7, and b is a random number ranging from -0.7 to 0.

7.

6. The method according to claim 1, characterized in that The construction of the denoising diffusion probability model includes: Construct a forward noise adding process, and gradually add noise to the image through a parameterized Markov chain until its distribution is close to the standard Gaussian distribution; Construct a UNet network for predicting noise, which receives a noisy image x t , the input of the image to be standardized x and the current time t, and the output of the predicted noise under the current state Construct a backward denoising process by predicting random noise and the noisy image x t , and through reverse reasoning, we get the noisy image x at the next moment t-1 , until the noise is completely removed and the stained standardized image is obtained.

7. The method according to claim 6, characterized in that The image is gradually added with noise through a parameterized Markov chain until its distribution is close to a standard Gaussian distribution, which can be expressed as the following formula: Among them, x0 represents the clean image without noise, x t is the noisy image at time t, ∈ is the random noise sampled from a normal distribution, is the hyperparameter of the denoising diffusion probability model at time t.

8. The method according to claim 1, characterized in that: The determining whether the training meets the expected effect includes: When comparing the performance of the dyed standard image generated by the training model with the actual required dyed standardized image in terms of structural similarity SSIM, peak signal-to-noise ratio PSNR, and learning-perceptual image block similarity LPIPS, it is expressed as the following formula: Where μ is the mean, σ is the standard deviation, and σ xy is the covariance of image x and image y, MSE is the mean square error of the two images, and MaxValue is the maximum value that the image pixel can take. and is the lth layer feature extracted by the pre-trained neural network, w l is the corresponding weight vector; The staining normalization results are applied to relevant downstream tasks, and the corresponding indicators of the computer-aided detection / diagnosis system for the stained and standardized images are better than those for the unstained and standardized images.

9. The method according to claim 1, characterized in that: The method of sampling the tissue pathology image by adopting a renormalization sampling strategy comprises: Add a certain degree of noise to the image to be standardized and use it as the initial image in the backward denoising process; Through the non-Markov chain method, some steps in the backward denoising process are skipped, so as to quickly and accurately obtain the stained normalized image of the denoised diffusion probability model.

10. The method according to claim 9, characterized in that: Through the non-Markov chain method, some steps in the jump backward denoising process are expressed as the following formula: Among them, x prev prev indicates that there can be multiple steps between t.

Images