Deep learning-based pathological image batch effect optimization method and system

Abstract

The invention discloses a pathological image batch effect optimization method and system based on deep learning, and the method comprises the steps: obtaining a pathological sample image, and carrying out the image preprocessing, and obtaining to-be-processed multi-center pathological image data; a batch discriminator training module and a batch confrontation module are introduced, and a batch effect optimization neural network model is constructed; and performing pathological image batch effect optimization processing on the to-be-processed multi-center pathological image data based on the batch effect optimization neural network model to obtain optimized multi-center pathological image data. By using the method, the batch effect of the multi-center pathological image data can be weakened, and the adaptability and prediction accuracy of the model to multiple pieces of center data are further improved. The pathological image batch effect optimization method and system based on deep learning can be widely applied to the technical field of image recognition.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a method and system for optimizing batch effects of pathological images based on deep learning. Background Art

[0002] Existing pathological diagnosis mainly relies on pathologists to manually examine H&E-stained tissue sections using a microscope. Although this method is widely used in clinical practice, it is inefficient, and the results are greatly affected by the operator's skills and experience, which is prone to errors. With the development of digital pathology, the use of full-frame digital pathology images (WSI) has become popular, making it possible to use artificial intelligence technologies such as deep learning. These technologies can automate image analysis and improve the efficiency and consistency of diagnosis. However, data from multicenter studies usually produce batch effects due to differences in acquisition protocols, equipment manufacturers, equipment drift, and other factors between different centers. These effects cause the performance of the model established during training to deteriorate when applied to data from new centers, limiting the generalization ability of the model. At present, although some methods such as gradient distortion correction, bias field correction, and intensity normalization are used in the data preprocessing stage, these technologies only alleviate the batch effect to a certain extent, but fail to fundamentally solve the problem. In addition, existing processing methods usually rely on statistical tests or machine learning algorithms, which have a high dependence on the distribution and characteristics of the data and are not effective in all cases. Existing models are usually trained and optimized on specific data sets, but in real-world applications, especially cross-center application scenarios, there are significant changes in data distribution and conditions, and existing models often find it difficult to adapt to such changes. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a pathology image batch effect optimization method and system based on deep learning, which can reduce the batch effect of multi-center pathology image data, thereby improving the adaptability and prediction accuracy of the model to multiple center data.

[0004] The first technical solution adopted by the present invention is: a method for optimizing batch effects of pathological images based on deep learning, comprising the following steps:

[0005] Acquire pathological sample images and perform image preprocessing to obtain multi-center pathological image data to be processed;

[0006] Introduce batch discriminator training module and batch adversarial module to build a batch effect optimization neural network model;

[0007] The batch effect optimization neural network model is used to optimize the batch effect of the multi-center pathology image data to be processed, and the optimized multi-center pathology image data is obtained.

[0008] Furthermore, the step of acquiring the pathological sample image and performing image preprocessing to obtain the multi-center pathological image data to be processed specifically includes:

[0009] Perform image data extraction and processing based on the pathological sample data to obtain a pathological sample image;

[0010] Performing format conversion processing on the pathological sample image by using a digital pathological image scanner to obtain a pathological sample image in a digital format;

[0011] The pathological sample images in digital format are sequentially subjected to image enhancement, threshold segmentation and target region cutting processes to obtain multi-center pathological image data to be processed.

[0012] Furthermore, the batch effect optimization neural network model includes a data processing module, a batch discriminator training module and a batch adversarial module, the first output end of the data processing module is connected to the first input end of the batch discriminator training module, the second output end of the data processing module is connected to the first input end of the batch adversarial module, the output end of the batch discriminator training module is connected to the second input end of the batch adversarial module, and the output end of the batch adversarial module is connected to the second input end of the batch discriminator training module.

[0013] Furthermore, the step of performing batch effect optimization processing on the multi-center pathology image data to be processed based on the batch effect optimization neural network model to obtain optimized multi-center pathology image data specifically includes:

[0014] Determine batch training labels based on the multi-center pathology image data to be processed;

[0015] Inputting the multi-center pathology image data to be processed and the batch training labels into the batch effect optimization neural network model;

[0016] The data processing module based on the batch effect optimization neural network model performs slicing and feature extraction on the multi-center pathology image data to be processed, and obtains the multi-center pathology image feature vector data;

[0017] Based on the batch effect optimization neural network model, the batch discriminator training module performs batch prediction on the multi-center pathology image feature vector data and batch training labels to obtain the multi-center pathology image batch prediction probability value;

[0018] Based on the batch effect optimization neural network model's batch adversarial module, the batch prediction probability values ​​and batch training labels of multi-center pathology images are back-propagated and updated to obtain the optimized multi-center pathology image data.

[0019] Furthermore, the batch discriminator training module based on the batch effect optimization neural network model performs batch prediction on the multi-center pathology image feature vector data and batch training labels to obtain the multi-center pathology image batch prediction probability value, which specifically includes:

[0020] Inputting the multi-center pathology image feature vector data and batch training labels into a batch discriminator training module based on a batch effect optimization neural network model, wherein the batch discriminator training module includes a first linear transformation layer, a first nonlinear transformation layer, a second linear transformation layer, and a second nonlinear transformation layer;

[0021] Based on the first linear transformation layer of the batch discriminator training module, the feature dimension of the multi-center pathology image feature vector data is halved to obtain the multi-center pathology image feature vector data after linear transformation;

[0022] Based on the first nonlinear transformation layer of the batch discriminator training module, nonlinearly transform the linearly transformed multi-center pathology image feature vector data to obtain the nonlinearly transformed multi-center pathology image feature vector data;

[0023] Based on the second linear transformation layer of the batch discriminator training module, linear transformation is performed on the multi-center pathology image feature vector data after nonlinear transformation to obtain the multi-center pathology image batch prediction probability;

[0024] Based on the second nonlinear transformation layer of the batch discriminator training module, the batch prediction probability of the multi-center pathology images and the batch training labels are nonlinearly transformed and converted with confidence to obtain the batch prediction probability value of the multi-center pathology images.

[0025] Furthermore, the batch adversarial module based on the batch effect optimization neural network model performs back propagation update on the batch prediction probability values ​​and batch training labels of the multi-center pathology images to obtain the optimized multi-center pathology image data, which specifically includes:

[0026] The batch training labels are inverted to obtain inverted batch training labels;

[0027] The inverted batch training labels and the multi-center pathology image batch prediction probability values ​​are input into the batch adversarial module based on the batch effect optimization neural network model to calculate the batch adversarial loss, and obtain the multi-center pathology image batch loss value;

[0028] updating the batch discriminator training module through back propagation according to the batch loss values ​​of the multi-center pathology images to obtain an updated batch discriminator training module;

[0029] Based on the updated batch discriminator training module, the optimized multi-center pathology image data is output.

[0030] Furthermore, the batch discriminator training module adopts a topkpatch selection mechanism and an early stopping training strategy to improve the data processing efficiency of the batch discriminator training module.

[0031] The second technical solution adopted by the present invention is: a pathological image batch effect optimization system based on deep learning, comprising:

[0032] The first module is used to acquire pathological sample images and perform image preprocessing to obtain multi-center pathological image data to be processed;

[0033] The second module is used to introduce the batch discriminator training module and the batch adversarial module to build a batch effect optimization neural network model;

[0034] The third module is used to optimize the batch effect of the multi-center pathology image data to be processed based on the batch effect optimization neural network model to obtain the optimized multi-center pathology image data.

[0035] The beneficial effects of the method and system of the present invention are as follows: the present invention obtains pathological sample images and performs image preprocessing to obtain multi-center pathological image data to be processed, and then introduces a batch discriminator training module and a batch adversarial module to construct a batch effect optimization neural network model. Finally, based on the batch effect optimization neural network model, the multi-center pathological image data to be processed is subjected to pathological image batch effect optimization processing. By introducing the batch discriminator training module and the batch adversarial module, the extracted features and batch labels are used to enhance the identification effect of the batch discriminator, and the batch adversarial module is used to optimize the training of the entire network to reduce the batch effect and improve the adaptability and prediction accuracy of the model to multiple center data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the steps of a method for optimizing batch effects of pathological images based on deep learning of the present invention;

[0037] Figure 2 It is a structural block diagram of a pathological image batch effect optimization system based on deep learning of the present invention;

[0038] Figure 3 It is a schematic diagram of a batch effect optimization neural network model provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0040] Reference Figure 1 The present invention provides a pathological image batch effect optimization method based on deep learning, the method comprising the following steps:

[0041] S100, acquiring a pathological sample image and performing image preprocessing to obtain multi-center pathological image data to be processed;

[0042] Specifically, image data extraction and processing are performed based on the pathological sample data to obtain pathological sample images; format conversion processing is performed on the pathological sample images through a digital pathological image scanner to obtain pathological sample images in a digital format; image enhancement, threshold segmentation and target area cutting processing are performed on the pathological sample images in a digital format in sequence to obtain multi-center pathological image data to be processed.

[0043] In this embodiment, after the image is collected from the pathological sample, it is converted into a digital format by a digital pathology image scanner and transmitted to the computing platform, and then the software running on the computing platform performs image preprocessing, including image enhancement, threshold segmentation and target area cutting.

[0044] S200, introduce batch discriminator training module and batch adversarial module to build batch effect optimization neural network model;

[0045] Specifically, Figure 3 As shown, the batch effect optimization neural network model includes a data processing module, a batch discriminator training module and a batch adversarial module. The first output end of the data processing module is connected to the first input end of the batch discriminator training module, the second output end of the data processing module is connected to the first input end of the batch adversarial module, the output end of the batch discriminator training module is connected to the second input end of the batch adversarial module, and the output end of the batch adversarial module is connected to the second input end of the batch discriminator training module.

[0046] S300, performing pathology image batch effect optimization processing on the multi-center pathology image data to be processed based on the batch effect optimization neural network model to obtain optimized multi-center pathology image data.

[0047] Specifically, batch training labels are determined according to the multicenter pathology image data to be processed; the multicenter pathology image data to be processed and the batch training labels are input into the batch effect optimization neural network model; based on the data processing module of the batch effect optimization neural network model, the multicenter pathology image data to be processed are sliced ​​and feature extracted to obtain multicenter pathology image feature vector data; based on the batch discriminator training module of the batch effect optimization neural network model, batch predictions are performed on the multicenter pathology image feature vector data and the batch training labels to obtain multicenter pathology image batch prediction probability values; based on the batch adversarial module of the batch effect optimization neural network model, back-propagation updates are performed on the multicenter pathology image batch prediction probability values ​​and the batch training labels to obtain optimized multicenter pathology image data.

[0048] In this embodiment, the batch effect optimization neural network model includes a data processing module, a batch discriminator training module, and a batch adversarial module, wherein the input of the data processing module is data from multiple centers, and the output is data after preprocessing such as slicing. The output of the data processing module is used as the input data of the main task model. The main task model extracts feature vectors of all data, which are used as inputs of the batch discriminator training module and the batch adversarial module. The output of the batch discriminator training module is connected to the batch discriminator in the batch adversarial module, and the output of the batch adversarial module is used to assist the main task training to improve model performance.

[0049] That is, the preprocessed multi-center pathology image data is input into the batch effect optimization neural network model; the feature vector is extracted through the main task model; the batch discriminator is trained by the feature vector extracted by the main task model and the corresponding batch label to enhance the effect of the batch discriminator module; the batch adversarial module is used to encourage the model to generate feature vectors with reduced batch effects to enhance the consistency of data from different centers; finally, the performance of the main task model is optimized and improved through the mutual promotion effect of the batch discriminator module and the batch adversarial module.

[0050] Among them, it should be noted that the batch discriminator training module learns the deep features of multi-center pathology image data and identifies the batch features of each data center to enhance the identification effect of the batch discriminator. An adversarial effect is formed between the batch adversarial module and the batch discriminator training module, so that the two are mutually enhanced. The batch adversarial module guides the model to generate feature vectors that are not easily affected by batch effects through batch adversarial loss, so as to reduce the visible differences between different batches of data.

[0051] In addition, the batch discriminator training module will select the top k patches with the highest scores as training data, and during the batch discriminator verification process, if the three verification losses are less than the currently recorded minimum verification loss, the batch discriminator training will be stopped. This allows the discriminator to be sufficiently trained while also ensuring a certain training efficiency.

[0052] Furthermore, it should be noted that the data processing process for the batch discriminator training module is:

[0053] The multi-center pathology image feature vector data and the batch training labels are input into a batch discriminator training module based on a batch effect optimization neural network model, wherein the batch discriminator training module includes a first linear transformation layer, a first nonlinear transformation layer, a second linear transformation layer and a second nonlinear transformation layer; based on the first linear transformation layer of the batch discriminator training module, the feature dimension of the multi-center pathology image feature vector data is halved to obtain the multi-center pathology image feature vector data after linear transformation; based on the first nonlinear transformation layer of the batch discriminator training module, the multi-center pathology image feature vector data after linear transformation is nonlinearly transformed to obtain the multi-center pathology image feature vector data after nonlinear transformation; based on the second linear transformation layer of the batch discriminator training module, the multi-center pathology image feature vector data after nonlinear transformation is linearly transformed to obtain the batch prediction probability of the multi-center pathology image; based on the second nonlinear transformation layer of the batch discriminator training module, the batch prediction probability of the multi-center pathology image and the batch training labels are nonlinearly transformed and converted with confidence to obtain the batch prediction probability value of the multi-center pathology image.

[0054] Specifically, the batch adversarial loss obtained for each discriminator in the batch adversarial module requires special weights to be added to the main task loss to process different data batches and improve the generalization ability of the model. The specific operations of the batch discriminator training module include combining the batch label with the predicted label obtained by the discriminator inference feature vector, and optimizing the network using the loss function to make the network output as close to the true label as possible. Specific reasoning process of the discriminator: These feature vectors first undergo a linear transformation to halve the feature dimension, and then undergo a nonlinear transformation through an activation function to enhance the expressive power of the model. Subsequently, after a second linear layer, the feature dimension is further converted into an output to represent the predicted probability of each category. Finally, the activation function is applied to convert the output into a probability value to generate the predicted batch probability.

[0055] In summary, the data processing for the batch discriminator training module can be summarized as follows:

[0056] 1) Feature dimension compression;

[0057] First, the feature dimension is halved through a linear transformation (Fully Connected Layer) to reduce computational complexity while retaining key information.

[0058] Next, activation functions (such as ReLU, Leaky ReLU, etc.) are applied for nonlinear transformation to improve the expressive power of the model and enable the network to learn more complex patterns.

[0059] 2) Output layer;

[0060] After the second linear layer, the feature vector is further processed to transform its dimension into an output with the same dimension as the number of categories, representing the predicted probability of each category.

[0061] Finally, an activation function (such as Softmax) is applied to convert the outputs into probability values ​​that represent the model’s confidence in each class, forming the predicted batch probabilities.

[0062] Furthermore, it should be noted that the data processing process for the batch adversarial module is:

[0063] The batch training labels are negated to obtain the negated batch training labels; the negated batch training labels and the batch prediction probability values ​​of the multi-center pathology images are input into the batch adversarial module based on the batch effect optimization neural network model to calculate the batch adversarial loss to obtain the batch loss value of the multi-center pathology images; the batch discriminator training module is updated through back propagation according to the batch loss value of the multi-center pathology images to obtain the updated batch discriminator training module; based on the updated batch discriminator training module, the optimized multi-center pathology image data is output.

[0064] Specifically, the specific operations of the batch adversarial module include combining the negated batch labels with the predicted labels obtained by the discriminator inference feature vector, and optimizing the network using the loss function so that the features generated by the model and their corresponding predicted batches deviate as much as possible from the true labels.

[0065] In summary, the data processing of the batch adversarial module can be summarized as follows:

[0066] 1) Training process;

[0067] During the training process, a batch adversarial strategy is used to generate high-quality features. Specifically, the inverted batch labels (i.e., each label in the true label is replaced by another label) are combined with the predicted labels generated by the feature vector inferred by the discriminator.

[0068] Use a loss function (such as cross entropy loss) to evaluate the gap between the output generated by the model and the true label. The loss will be used to optimize the network weights, with the goal of making the output as close to the true label as possible.

[0069] 2) Optimization process;

[0070] Through the back-propagation algorithm, the calculated loss value is used to update the network parameters to improve the model's predictive ability. Through continuous iteration, the model will gradually learn better feature representations, making it perform better on new data.

[0071] Finally, the batch effect optimization neural network model of the present invention is evaluated by the calculation of classification accuracy (ACC) and the calculation of the receiver operating characteristic curve (ROC) and the area (AUC) under it. First, by comparing the prediction results of the model with the true label, and statistically predicting the correct number of samples, the classification accuracy is obtained by dividing this number by the total number of samples. The specific formula is: accuracy = correct prediction number / total number of samples, and further calculating the true positive rate (TPR) and false positive rate (FPR) under multiple different threshold settings. The true positive rate is defined as the number of true positives divided by the actual positive total number, and the false positive rate is defined as the number of false positives divided by the actual negative total number. Then, draw a ROC curve with false positive rate as the horizontal coordinate and true positive rate as the vertical coordinate. Finally, calculate the area under this curve (AUC), and use this area value as an evaluation index of model classification performance, wherein the closer the AUC value is to 1, the better the model performance is, and finally the baseline performance of the model before using the method of the present invention is compared with the performance after use, focusing on the improvement of performance. Specific evaluation indicators include classification accuracy and the area (AUC) under the receiver operating characteristic curve (ROC). The evaluation criteria set are: after applying the method of the present invention, the AUC value and classification accuracy of the model should show a significant improvement, so as to verify the effectiveness of the method of the present invention.

[0072] In summary, the embodiment of the present invention extracts image features through a feature extraction module, uses the extracted features and batch labels to enhance the identification effect of the batch discriminator, and applies a batch adversarial module to optimize the training of the entire network to reduce the batch effect and improve the adaptability and prediction accuracy of the model to data from multiple centers. By reducing the batch effect, the embodiment of the present invention can significantly improve the accuracy of pathological image analysis. Using the newly designed batch adversarial loss and batch discriminator, the model can more effectively process image data from different centers, thereby reducing errors caused by differences in scanners or operating protocols. The data results of its evaluation are shown in Table 1.

[0073] \ Max-MILACC MAX-MILAUC PMILACC PMILAUC Baseline 0.5813±0.0523 0.7016±0.0790 0.6187±0.0601 0.7187±0.0389 Ours 0.6875±0.0541 0.7727±0.0308 0.6500±0.0559 0.7391±0.0444

[0074] Reference Figure 2 , a deep learning-based pathological image batch effect optimization system, comprising:

[0075] The first module 201 is used to acquire pathological sample images and perform image preprocessing to obtain multi-center pathological image data to be processed;

[0076] The second module 202 is used to introduce a batch discriminator training module and a batch adversarial module to build a batch effect optimization neural network model;

[0077] The third module 203 is used to perform pathology image batch effect optimization processing on the multi-center pathology image data to be processed based on the batch effect optimization neural network model to obtain optimized multi-center pathology image data.

[0078] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0079] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A deep learning-based pathological image batch effect optimization method, characterized in that: The following steps are involved: Acquire pathological sample images and perform image preprocessing to obtain multi-center pathological image data to be processed; Introduce batch discriminator training module and batch adversarial module to build a batch effect optimization neural network model; The batch effect optimization neural network model is used to optimize the batch effect of the multi-center pathology image data to be processed, and the optimized multi-center pathology image data is obtained.

2. According to the deep learning-based pathological image batch effect optimization method of claim 1, it is characterized in that: The step of acquiring the pathological sample image and performing image preprocessing to obtain the multi-center pathological image data to be processed specifically includes: Perform image data extraction and processing based on the pathological sample data to obtain a pathological sample image; Performing format conversion processing on the pathological sample image by using a digital pathological image scanner to obtain a pathological sample image in a digital format; The pathological sample images in digital format are sequentially subjected to image enhancement, threshold segmentation and target region cutting processes to obtain multi-center pathological image data to be processed.

3. According to the deep learning-based pathological image batch effect optimization method of claim 2, it is characterized in that: The batch effect optimization neural network model includes a data processing module, a batch discriminator training module and a batch adversarial module. The first output end of the data processing module is connected to the first input end of the batch discriminator training module, the second output end of the data processing module is connected to the first input end of the batch adversarial module, the output end of the batch discriminator training module is connected to the second input end of the batch adversarial module, and the output end of the batch adversarial module is connected to the second input end of the batch discriminator training module.

4. According to the deep learning-based pathological image batch effect optimization method of claim 3, it is characterized in that: The step of performing pathology image batch effect optimization processing on the multi-center pathology image data to be processed based on the batch effect optimization neural network model to obtain optimized multi-center pathology image data specifically includes: Determine batch training labels based on the multi-center pathology image data to be processed; Inputting the multi-center pathology image data to be processed and the batch training labels into the batch effect optimization neural network model; The data processing module based on the batch effect optimization neural network model performs slicing and feature extraction on the multi-center pathology image data to be processed to obtain the multi-center pathology image feature vector data; Based on the batch effect optimization neural network model, the batch discriminator training module performs batch prediction on the multi-center pathology image feature vector data and batch training labels to obtain the multi-center pathology image batch prediction probability value; Based on the batch effect optimization neural network model's batch adversarial module, the batch prediction probability values ​​and batch training labels of multi-center pathology images are back-propagated and updated to obtain the optimized multi-center pathology image data.

5. According to the deep learning-based pathological image batch effect optimization method of claim 4, it is characterized in that: The batch discriminator training module based on the batch effect optimization neural network model performs batch prediction on the multi-center pathology image feature vector data and batch training labels to obtain the multi-center pathology image batch prediction probability value. This step specifically includes: Inputting the multi-center pathology image feature vector data and batch training labels into a batch discriminator training module based on a batch effect optimization neural network model, wherein the batch discriminator training module includes a first linear transformation layer, a first nonlinear transformation layer, a second linear transformation layer, and a second nonlinear transformation layer; Based on the first linear transformation layer of the batch discriminator training module, the feature dimension of the multi-center pathology image feature vector data is halved to obtain the multi-center pathology image feature vector data after linear transformation; Based on the first nonlinear transformation layer of the batch discriminator training module, nonlinearly transform the linearly transformed multi-center pathology image feature vector data to obtain the nonlinearly transformed multi-center pathology image feature vector data; Based on the second linear transformation layer of the batch discriminator training module, linear transformation is performed on the multi-center pathology image feature vector data after nonlinear transformation to obtain the multi-center pathology image batch prediction probability; Based on the second nonlinear transformation layer of the batch discriminator training module, the batch prediction probability of the multi-center pathology images and the batch training labels are nonlinearly transformed and converted with confidence to obtain the batch prediction probability value of the multi-center pathology images.

6. According to the deep learning-based pathological image batch effect optimization method of claim 5, it is characterized in that: The batch adversarial module based on the batch effect optimization neural network model performs back propagation update on the multi-center pathology image batch prediction probability value and batch training label to obtain the optimized multi-center pathology image data. This step specifically includes: The batch training labels are inverted to obtain inverted batch training labels; The inverted batch training labels and the multi-center pathology image batch prediction probability values ​​are input into the batch adversarial module based on the batch effect optimization neural network model to calculate the batch adversarial loss, and obtain the multi-center pathology image batch loss value; updating the batch discriminator training module through back propagation according to the batch loss values ​​of the multi-center pathology images to obtain an updated batch discriminator training module; Based on the updated batch discriminator training module, the optimized multi-center pathology image data is output.

7. A method for optimizing batch effects of pathological images based on deep learning according to claim 6, characterized in that: The batch discriminator training module also adopts a topkpatch selection mechanism and an early stopping training strategy to improve the data processing efficiency of the batch discriminator training module.

8. A pathological image batch effect optimization system based on deep learning, characterized in that: Includes the following modules: The first module is used to acquire pathological sample images and perform image preprocessing to obtain multi-center pathological image data to be processed; The second module is used to introduce the batch discriminator training module and the batch adversarial module to build a batch effect optimization neural network model; The third module is used to optimize the batch effect of the multi-center pathology image data to be processed based on the batch effect optimization neural network model to obtain the optimized multi-center pathology image data.

Images