Pathological image classification integration method based on feature information guidance

Abstract

The invention relates to a pathological image classification integration method based on feature information guidance, and belongs to the field of artificial intelligence and medical image processing. Comprising the following steps: acquiring an unlabeled pathological image patch data set; utilizing a plurality of pre-trained visual language models to extract respective visual embedding features and category probability distribution; integrating prediction results of multiple models in an average fusion mode to obtain robust preliminary semantic prediction; multi-model visual features are spliced to construct joint representation, and a Gaussian mixture model is used for modeling class condition structure prior; introducing KL divergence to construct a dual alignment objective function, and performing joint optimization on semantic prediction and structure prior; label distribution and structure prior parameters are alternately optimized through a block coordinate descent method, and mutual guidance and enhancement of the label distribution and the structure prior parameters are achieved. According to the method, multi-model knowledge and structure information are fully fused, the classification performance and generalization ability of the pathological image are effectively improved, and a more accurate and efficient solution is provided for intelligent pathological auxiliary diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to a pathological image classification ensemble method based on feature information guidance, belonging to the technical field of artificial intelligence and medical image processing BACKGROUND

[0002] With the development of digital pathology, whole slide images (WSIs) of tissue sections have gradually become an important data source for assisting cancer diagnosis and treatment planning. Due to the extremely high resolution of WSI images and the complex tissue structure, traditional computer vision methods often need to rely on dense annotation and complex training process to achieve effective classification. However, these methods generally rely on a large number of expert annotated samples, which restricts their application in real clinical scenarios.

[0003] In recent years, vision-language models (VLMs) have been gradually introduced into pathological image analysis tasks due to their strong cross-modal alignment ability and zero-shot generalization performance. Such models can achieve semantic matching between images and natural language descriptions through large-scale image-text data pre-training, thereby promoting the development of zero-shot classification tasks. Existing VLMs models customized for pathological images (such as PLIP, CONCH, and MUSK) have shown significant performance on multiple public datasets. However, these models exhibit significant differences in performance across different datasets, which is attributed to differences in their pre-training data, training objectives, and architecture design, resulting in certain limitations in their generalization ability and robustness, making it difficult to meet the stability requirements of real clinical deployment.

[0004] To address the above problems, ensemble learning methods, as a strategy to improve model robustness without additional training, have gradually attracted attention. Although traditional ensemble methods such as average fusion and maximum voting can improve performance to some extent, they only perform score-level fusion at the output layer, failing to fully utilize the rich visual feature information within the vision-language model, thus limiting the improvement effect. In addition, although hybrid expert architectures can theoretically coordinate the capabilities of multiple sub-models, their training process highly depends on the joint optimization of routing mechanisms, making it difficult to be directly applied to zero-shot deployment scenarios. SUMMARY

[0005] The present application aims at the shortcomings of the above-mentioned existing methods, and proposes a pathological image classification integration method based on feature information guidance, which is used to fully integrate the intermediate visual features of multiple visual language models without retraining, to integrate the classification results of different visual language models, and to improve the overall prediction performance. Unlike the traditional method of only fusing the output results, the present method constructs a unified visual representation space, excavates the complementarity between the features of different models, models the semantic structural information in the feature distribution by means of the Gaussian mixture model, so as to realize the prior guidance classification of structure perception.

[0006] The technical scheme of the present application is: a pathological image classification integration method based on feature information guidance, and the specific steps of the feature information integration zero-shot pathological image classification method are as follows:

[0007] Step 1, obtaining an input image patch data set, which is an unannotated pathological image patch set;

[0008] Step 2, using multiple pre-trained visual language models to process each image patch respectively, extracting the visual embedding features and class probability prediction corresponding to the model, and forming the representation output of multiple models;

[0009] Step 3, integrating the class prediction results of multiple models, and adopting an average fusion method to obtain a stable preliminary semantic prediction distribution;

[0010] Step 4, splicing the visual embedding features of each model to construct a multi-model multi-view visual joint feature, and modeling the class conditional structure prior by means of the Gaussian mixture model, to obtain the class distribution of the sample in the structure space;

[0011] Step 5, constructing a joint optimization objective function, introducing the structure prior distribution and the semantic prediction result, and performing double alignment through KL divergence to optimize the final label distribution;

[0012] Step 6, adopting a block coordinate descent algorithm to alternately optimize the label distribution and the Gaussian mixture model GMM parameters, and realizing the mutual guidance and enhancement of the structure prior and the semantic prediction through closed-form update iteration convergence.

[0013] Further, in Step 1, an unannotated image patch set is extracted from a high-resolution pathological section As the input data of zero-shot classification, N is the total number of patches.

[0014] Further, in Step 2, M multi-modal pre-trained visual language models are selected, and each model m∈{1,…,M} processes the input image patch x i Output a set of visual embedding features wherein The image encoder for model m, meanwhile, computes the probability distribution of the image belonging to each class: where the probability of each class The similarity of the image to the text description of the class is computed by a temperature-scaled softmax function where denotes the similarity score of the image to the k-th class text, τ is the temperature parameter, used to control the confidence smoothing degree of the class probability distribution, K is the number of classes, Δ K denotes the K-dimensional probability simplex, used to ensure that each soft label z i constitutes a valid probability distribution for each class.

[0015] Further, the Step3 includes the following:

[0016] Given the set of class probabilities from M multimodal pre-trained visual language models First, the element-wise average ensemble method is used to obtain a smoothed preliminary prediction as follows:

[0017]

[0018] where Ensemble(·) represents the common ensemble learning method to integrate the output, which is based on the element-wise average operation on the input multiple prediction vectors.

[0019] Further, the Step4 includes the following steps:

[0020] First, a multi-model multi-view visual joint feature is constructed from the visual embedding features of all models forming a multi-model multi-view visual joint feature that fuses the knowledge of different models, where denotes the visual embedding feature of model m for sample x i extracted from the visual embedding feature;

[0021] Then, a Gaussian mixture model is used to model the distribution of these joint features under the condition of the class, assuming that each sample comes from a class-specific Gaussian distribution mixture component; for each class k∈{1,…,K}, let its corresponding Gaussian distribution have mean μ k , covariance matrix ∑ k , and mixing coefficient π k ; according to the Gaussian mixture model, the probability of sample x i belonging to class k is as follows:

[0022]

[0023] where, denotes the multivariate Gaussian distribution density function; and finally obtains the structure-aware class probability distribution vector p i = [p i,1 ,…,p i,K ] T as a sample x i The structural soft assignment results of each class in the joint feature space.

[0024] Further, in Step 5, a joint optimization objective function based on KL divergence is constructed to jointly model and constrain the sample label distribution; this step introduces a double alignment mechanism to integrate the smoothed preliminary prediction obtained by the element average integration in Step 3 and the structure-aware class probability distribution vector p i in Step 4, to guide the final label distribution to converge to semantic and structural consistency;

[0025] The joint optimization objective function based on KL divergence is:

[0026]

[0027] where the first term is used to measure the difference between the final label distribution and the structural prior, thereby strengthening the model's classification consistency in the feature space structure; the second term is the weighted KL divergence:

[0028]

[0029] where the hyperparameter λ>0 controls the relative weight of the semantic prior in the final label decision, used to adjust the fusion ratio between structure information and semantic information, z i,k denotes the model's prediction probability that the sample x i belongs to class k; denotes the prediction value of the sample x i in class k obtained by multi-model average integration, Δ K denotes the K-dimensional probability simplex, ensuring that each soft label z i constitutes an effective probability distribution for each class, and N is the total number of patches.

[0030] Further, Step 6 includes:

[0031] To optimize the objective function in Step 5, the block coordinate descent (BCD) strategy is adopted, which divides the optimization variables into two parts: the soft label distribution z and the Gaussian mixture model parameters (μ,∑); in each iteration, one variable is updated alternately while the other is fixed until convergence; the specific steps are as follows:

[0032] Step 6.1, initialization:

[0033] First, initialize the pseudo label distribution z i and Gaussian mixture model parameters; specifically, the pseudo-label vector is initialized to the average of all visual language model predictions: in Represents the mean of the prediction results of different models, that is, the element average integration method is used to obtain a smooth preliminary prediction; at the same time, the fused visual features Modeled as a balanced multivariate Gaussian mixture distribution; where each class K corresponds to a specific mean vector μ k , while the covariance matrix ∑ is shared among all categories and is a diagonal matrix;

[0034] Step 6.2, Update the soft label distribution z:

[0035] When the Gaussian mixture model parameters (μ, ∑) are fixed, the label distribution z of each sample is updated i ; The update strategy aims to make z i Simultaneously align the structural prior p i and semantic priors That is, a smooth preliminary prediction is obtained by element-wise averaging integration; the closed-form solution is obtained by minimizing the following objective function:

[0036]

[0037] where α is the Lagrange multiplier, which is used to force z i Satisfy the normalization constraint; About z i,k Taking the derivative and setting it to zero, we get log z i,k The linear combination form of , further derives the following closed-form update formula:

[0038]

[0039] Among them, ⊙ represents element-by-element multiplication, and the exponential operation is also performed one by one on the element dimension;

[0040] Step 6.3, Update GMM parameters:

[0041] When z is fixed, the Gaussian mixture model parameters (μ, ∑) are updated analytically; when updating the mean μ k When , the objective function becomes a convex function under the condition of fixed z and ∑, and the closed-form solution is obtained by weighted summation:

[0042]

[0043] Similarly, when other variables are fixed, the update of the covariance also has a closed-form solution, as follows:

[0044]

[0045] Based on the updated mu k and ∑ calculate the structure prior probability p of each sample belonging to each category i , that is, the normalized likelihood generated by different Gaussian distributions, p i , that is, the structure-aware class probability distribution vector.

[0046] The application also provides a pathological image classification integrated system based on feature information guidance, which comprises a module for executing the pathological image classification integrated method based on feature information guidance.

[0047] The application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the pathological image classification integrated method based on feature information guidance.

[0048] The application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the pathological image classification integrated method based on feature information guidance.

[0049] The application has the following beneficial effects:

[0050] (1) The application provides a visual language model integration method based on saliency guidance, feature fusion and information gain optimization strategy, which explicitly models the complementarity and uncertainty between different visual language models, and breaks through the bottleneck of poor generalization and insufficient robustness of traditional integration methods in multi-modal medical image tasks.

[0051] (2) The application adopts an average integration strategy, combines the prior knowledge of multi-source visual language models, optimizes the integrated features through feature alignment and information guidance mechanism, effectively enhances the perception ability of the model to the key semantic regions of pathological images, significantly improves the pathological image classification performance without downstream annotation, and provides a more accurate and efficient solution for intelligent pathological auxiliary diagnosis.

[0052] (3) The application has good scalability and deployment flexibility, and can be compatible with visual language models of different architectures and training targets. Experimental results show that the application is superior to single models and traditional integration methods on multiple real pathological image datasets, and has high clinical application potential. DETAILED DESCRIPTION

[0053] Embodiment 1: Pathological image classification method based on feature information guidance, the present application solves the problems of prediction instability and insufficient utilization of structural information in the existing visual language model in the pathological image classification task; the specific steps of the feature information integrated pathological image classification method are as follows:

[0054] Step 1, obtaining an input image patch data set, which is an unannotated pathological image patch set;

[0055] Further, in Step 1, an unannotated image patch set is extracted from a high-resolution pathological section As the input data of zero-shot classification, N is the total number of patches.

[0056] Step 2, using multiple pre-trained visual language models to process each image patch respectively, extracting the visual embedding features and class probability prediction corresponding to the model, forming the representation output of multiple models;

[0057] Further, in Step 2, M multimodal pre-trained visual language models are selected, each model m∈{1,…,M} processes the input image patch x i Output a set of visual embedding features Wherein is the image encoder of model m, at the same time, the probability distribution of the image belonging to each class is calculated: Where the probability of each class is The similarity between the image and the class text description is calculated by the temperature scaled softmax function Where Indicates the similarity score between the image and the kth class text, τ is the temperature parameter, used to control the confidence smoothing degree of the class probability distribution, K is the number of classes, Δ K Indicates a K-dimensional probability simplex, used to ensure that each soft label z i All constitute an effective probability distribution for each class.

[0058] Step 3, integrating the class prediction results of multiple models, and adopting an average fusion method to obtain a stable preliminary semantic prediction distribution;

[0059] Further, Step 3 includes the following:

[0060] Given the class probability set from M multimodal pre-trained visual language models First, the element average integration method is used to obtain a smooth preliminary prediction as follows:

[0061]

[0062] Among them, Ensemble(·) represents the common ensemble learning method to integrate the output, which is based on the element-by-element averaging operation of multiple prediction vectors of the input.

[0063] Step 4: Concatenate the visual embedding features of each model to construct a multi-model and multi-view visual joint feature, and use the Gaussian mixture model to model the class conditional structure prior to obtain the category distribution of samples in the structure space;

[0064] Furthermore, Step 4 includes the following steps:

[0065] First, construct a multi-model multi-view visual joint feature from the visual embedding features of all models Form a multi-model multi-view visual joint feature that integrates knowledge from different models, where Represents model m for sample x i Extracted visual embedding features;

[0066] Then, the distribution of these joint features under category conditions is modeled using a Gaussian mixture model, assuming that each sample comes from a category-specific Gaussian distribution mixture component; for each category k∈{1,…,K}, let its corresponding Gaussian distribution have mean μ k , covariance matrix Σ k , the mixing coefficient is π k ; Get sample x according to Gaussian mixture model i The probability of belonging to class k is as follows:

[0067]

[0068] in, Represents the multivariate Gaussian distribution density function; finally, the structure-aware category probability distribution vector p is obtained i =[p i,1 ,…,p i,K ] T As a sample x i Structural soft assignment results for each class in the joint feature space.

[0069] Step 5: Construct a joint optimization objective function, introduce the structural prior distribution and semantic prediction results, perform double alignment through KL divergence, and optimize the final label distribution;

[0070] Further, in Step 5, to realize effective fusion of semantic prediction results and structural prior information, thereby obtaining more stable and accurate final label prediction, the application constructs a joint optimization objective function based on Kullback-Leibler (KL) divergence to jointly model and constrain the sample label distribution; this step introduces a double alignment mechanism to integrate the smoothed preliminary prediction obtained by using the element average integration method in Step 3 and the structure-aware class probability distribution vector p i in Step 4, to guide the final label distribution to converge to semantic and structural consistency, thereby improving the generalization ability of zero-shot classification;

[0071] The joint optimization objective function based on KL divergence is as follows:

[0072]

[0073] wherein the first term is used to measure the difference between the final label distribution and the structural prior, thereby strengthening the classification consistency of the model in the feature space structure; the second term is a weighted KL divergence:

[0074]

[0075] wherein the hyperparameter λ>0 controls the relative weight of the semantic prior in the final label decision, to adjust the fusion ratio between the structural information and the semantic information, z i,k represents the prediction probability of the model for the sample x i belonging to the class k; represents the prediction value of the sample x i in class k obtained by multi-model average integration, and ΔK represents a K-dimensional probability simplex, to ensure that each soft label z i constitutes an effective probability distribution for each class, and N is the total number of patches.

[0076] Step 6, the block coordinate descent algorithm is used to alternately optimize the label distribution and the Gaussian mixture model (GMM) parameters, and the iterative convergence is realized through closed-form updating, to realize mutual guidance and enhancement of the structural prior and the semantic prediction. Further, Step 6 includes:

[0077] To optimize the objective function in Step 5, the block coordinate descent (BCD) strategy is used to divide the optimization variables into two parts: the soft label distribution z and the Gaussian mixture model parameters (μ,∑); in each iteration, one variable is updated alternately while the other variable is fixed until convergence; the specific steps are as follows:

[0078] Step 6.1, initialization:

[0079] First, initialize the pseudo label distribution z i and Gaussian mixture model parameters; specifically, the pseudo-label vector is initialized to the average of all visual language model predictions: in Represents the mean of the prediction results of different models, that is, the element average integration method is used to obtain a smooth preliminary prediction; at the same time, the fused visual features Modeled as a balanced multivariate Gaussian mixture distribution; where each class K corresponds to a specific mean vector μ k , while the covariance matrix ∑ is shared among all categories and is a diagonal matrix;

[0080] Step 6.2, Update the soft label distribution z:

[0081] When the Gaussian mixture model parameters (μ, ∑) are fixed, the label distribution z of each sample is updated i ; The update strategy aims to make z i Simultaneously align the structural prior p i and semantic priors That is, a smooth preliminary prediction is obtained by element-wise averaging integration; the closed-form solution is obtained by minimizing the following objective function:

[0082]

[0083] where α is the Lagrange multiplier, which is used to force z i Satisfy the normalization constraint; About z i,k Taking the derivative and setting it to zero, we get log z i,k The linear combination form of , further derives the following closed-form update formula:

[0084]

[0085] Among them, ⊙ represents element-by-element multiplication, and the exponential operation is also performed one by one on the element dimension;

[0086] Step 6.3, Update GMM parameters:

[0087] When z is fixed, the Gaussian mixture model parameters (μ, ∑) are updated analytically; when updating the mean μ k When , the objective function becomes a convex function under the condition of fixed z and ∑, and the closed-form solution is obtained by weighted summation:

[0088]

[0089] Similarly, when other variables are fixed, the update of the covariance also has a closed-form solution, as follows:

[0090]

[0091] Based on the updated mu k and Sigma calculate the structure prior probability p of each sample belonging to each category i , that is, the normalized likelihood of its features generated by different Gaussian distributions, p i , that is, the structure-aware class probability distribution vector.

[0092] The application also provides a pathological image classification integration system based on feature information guidance, which comprises a module for executing the pathological image classification integration method based on feature information guidance.

[0093] The application also provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the pathological image classification integration method based on feature information guidance.

[0094] The application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the pathological image classification integration method based on feature information guidance.

[0095] In order to verify the effectiveness of the method in the pathological image classification task, comparative experiments were conducted on five public pathological image datasets (SICAP-MIL, NCT-CRC, LC-LUNG, SKINCANCER and WSSS4LUAD), and the accuracy was used as the evaluation index. The experimental results are shown in Table 1:

[0096] Table 1 shows the performance comparison of different methods on different pathological data sets

[0097]

[0098] The FIRE framework proposed in the application is compared with three representative visual language models (CONCH, PLIP and MUSK) and their test-time adaptive versions based on TransCLIP. The results show that MUSK performs best overall under single model conditions (average accuracy 67.04%), TransCLIP can effectively improve the accuracy of each baseline model by 5-11%, but the method achieves optimal performance on all data sets, with an average accuracy of 85.59%, which is 8.89% higher than the strongest baseline (MUSK+TransCLIP). This significant improvement verifies that the feature-level integration strategy proposed by us has stronger robustness and generalization ability without additional training, and shows its application potential in the zero-shot classification task of pathological images.

[0099] The specific embodiments of the present application have been described above in detail, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A pathological image classification integration method based on feature information guidance, characterized in that: The method comprises: Step 1: Obtain an input image patch dataset, which is a collection of unlabeled pathological image patches. Step 2: Use multiple pre-trained visual language models to process each image patch separately, extract the corresponding visual embedding features and category probability predictions of the model, and form a multi-model representation output; Step 3: Integrate the category prediction results of multiple models and use the average fusion method to obtain a stable preliminary semantic prediction distribution; Step 4: Concatenate the visual embedding features of each model to construct a multi-model and multi-view visual joint feature, and use the Gaussian mixture model to model the class conditional structure prior to obtain the category distribution of samples in the structure space; Step 5: Construct a joint optimization objective function, introduce the structural prior distribution and semantic prediction results, perform double alignment through KL divergence, and optimize the final label distribution; Step 6: Use the block coordinate descent algorithm to alternately optimize the label distribution and the Gaussian mixture model (GMM) parameters, and achieve mutual guidance and enhancement of structural prior and semantic prediction through closed-form update iterative convergence.

2. The pathological image classification integration method based on feature information guidance according to claim 1, characterized in that: In Step 1, a set of unlabeled image patches is extracted from high-resolution pathological sections. As the input data for zero-shot classification, N is the total number of patches.

3. The pathological image classification integration method based on feature information guidance according to claim 1, characterized in that: In Step 2, M multimodal pre-trained visual language models are selected, each model m∈{1,…,M} is trained on the input image patch x i Output a set of visual embedding features in For the image encoder of model m, at the same time, calculate the probability distribution of the image belonging to each category: The probability of each category The similarity between the image and the category text description is calculated by the temperature-scaled softmax function in represents the similarity score between the image and the k-th category text, τ is the temperature parameter used to control the confidence smoothness of the category probability distribution, K is the number of categories, Δ K represents a K-dimensional probability simplex, which is used to ensure that each soft label z i They all constitute a valid probability distribution for each category.

4. The pathological image classification integration method based on feature information guidance according to claim 1, characterized in that: Step 3 includes the following: Given a set of category probabilities from M multimodal pre-trained visual language models First, we use the element-wise average integration method to get a smooth preliminary prediction as follows: Among them, Ensemble(·) represents the common ensemble learning method to integrate the output, which is based on the element-by-element averaging operation of multiple prediction vectors of the input.

5. The pathological image classification integration method based on feature information guidance according to claim 1, characterized in that: Step 4 includes the following steps: First, construct a multi-model multi-view visual joint feature from the visual embedding features of all models Form a multi-model multi-view visual joint feature that integrates knowledge from different models, where Represents model m for sample x i Extracted visual embedding features; Then, the distribution of these joint features under category conditions is modeled using a Gaussian mixture model, assuming that each sample comes from a category-specific Gaussian distribution mixture component; for each category k∈{1,…,K}, let its corresponding Gaussian distribution have mean μ k , covariance matrix ∑ k , the mixing coefficient is π k ; Get sample x according to Gaussian mixture model i The probability of belonging to class k is as follows: in, Represents the multivariate Gaussian distribution density function; ultimately, the structure-aware category probability distribution vector is obtained As a sample x i Structural soft assignment results for each class in the joint feature space.

6. The pathological image classification integration method based on feature information guidance according to claim 1, characterized in that: In Step 5, a joint optimization objective function based on KL divergence is constructed to jointly model and constrain the sample label distribution; this step introduces a dual alignment mechanism to integrate the element-wise average integration method in Step 3 to obtain a smooth preliminary prediction. and the structure-aware category probability distribution vector p in Step 4 i , guiding the final label distribution to converge towards the dual consistency of semantics and structure; The joint optimization objective function based on KL divergence is: Among them, the first It is used to measure the difference between the final label distribution and the structural prior, thereby strengthening the classification consistency of the model in the feature space structure; the second item is the weighted KL divergence: Among them, the hyperparameter λ>0 controls the relative weight of the semantic prior in the final label decision, which is used to adjust the fusion ratio between structural information and semantic information. i,k Represents the model for sample x i The predicted probability of belonging to category k; Represents the sample x obtained by multi-model average integration i The predicted value on category k, Δ K Represents a K-dimensional probability simplex, ensuring that each soft label z i They all constitute the effective probability distribution for each category, and N is the total number of patches.

7. The pathological image classification integration method based on feature information guidance according to claim 1, characterized in that: Step 6 includes: To optimize the objective function in Step 5, a block coordinate descent (BCD) strategy is used to divide the optimization variables into two parts: the soft label distribution z and the Gaussian mixture model parameters (μ, ∑). In each iteration, one of the variables is updated alternately while the other variables are fixed until convergence. The specific steps are as follows: Step 6.1, Initialization: First, initialize the pseudo label distribution z i and Gaussian mixture model parameters; specifically, the pseudo-label vector is initialized to the average of all visual language model predictions: in Represents the mean of the prediction results of different models, that is, the element average integration method is used to obtain a smooth preliminary prediction; at the same time, the fused visual features Modeled as a balanced multivariate Gaussian mixture distribution; where each class K corresponds to a specific mean vector μ k , while the covariance matrix ∑ is shared among all categories and is a diagonal matrix; Step 6.2, Update the soft label distribution z: When the Gaussian mixture model parameters (μ, ∑) are fixed, the label distribution z of each sample is updated i ; The update strategy aims to make z i Simultaneously align the structural prior p i and semantic priors That is, a smooth preliminary prediction is obtained by element-wise averaging integration; the closed-form solution is obtained by minimizing the following objective function: where α is the Lagrange multiplier, which is used to force z i Satisfy the normalization constraint; About z i,k Taking the derivative and setting it to zero, we get logz i,k The linear combination form of , further derives the following closed-form update formula: Among them, ⊙ represents element-by-element multiplication, and the exponential operation is also performed one by one on the element dimension; Step 6.3, Update GMM parameters: When z is fixed, the Gaussian mixture model parameters (μ, ∑) are updated analytically; when updating the mean μ k When , the objective function becomes a convex function under the condition of fixed z and ∑, and the closed-form solution is obtained by weighted summation: Similarly, when other variables are fixed, the update of the covariance also has a closed-form solution, as follows: Based on the updated μ k And ∑ calculate the structural prior probability p of each sample belonging to each category i , that is, its features are regarded as the normalized likelihood generated by different Gaussian distributions, p i That is the structure-aware category probability distribution vector.

8. A pathological image classification integration system guided by feature information, characterized in that: The system includes: a module for executing the pathological image classification integration method based on feature information guidance as described in any one of claims 1 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the pathological image classification integration method based on feature information guidance as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the pathological image classification integration method based on feature information guidance as described in any one of claims 1 to 7 is implemented.

Images