A recognition and classification method and system for breast cancer pathology images

Abstract

The present invention discloses a recognition and classification method and system for breast cancer pathology images, including: collecting breast cancer pathology images and preprocessing them, extracting multi-scale texture features based on convolutional neural networks, capturing local and global texture information of breast tissue by designing convolution kernels with different receptive field sizes, and obtaining texture features; adaptively weighting texture features by attention mechanism, dynamically adjusting feature weights according to the importance of different regions, and obtaining key texture features; using a sequence feature selection algorithm to screen the optimal texture feature subset based on the key texture features, and obtaining the most discriminative feature combination through iterative search and cross-validation; constructing a recognition and classification model based on ensemble learning according to the feature combination, inputting the pathology image to be predicted into the recognition and classification model for processing, and obtaining recognition and classification results. The present invention effectively improves the accuracy, reliability and interpretability of breast cancer pathology image classification.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition and classification, and in particular relates to a recognition and classification method and system for breast cancer pathology images. Background Art

[0002] Intelligent recognition of breast cancer pathology images faces the technical challenge of texture feature extraction and selection. The structure of breast tissue is complex and changeable, and there are a large number of subtle texture differences. Traditional texture analysis methods such as gray-level co-occurrence matrix and wavelet transform are difficult to fully characterize the spatial arrangement and structural characteristics of breast cancer tissue. At the same time, different texture features contribute to breast cancer classification to different degrees. How to select the most discriminative feature combination from numerous texture features and build an efficient breast cancer classification model is a key problem that needs to be solved urgently. In addition, the quality of breast cancer pathology images varies, and there are interference factors such as uneven staining, background noise, and tissue folding, which have a serious impact on the extraction and selection of texture features. How to achieve standardized acquisition and preprocessing of breast cancer pathology images in a complex clinical environment and improve the reliability and stability of texture features is also a technical problem that needs to be overcome urgently. Only by solving the core problem of texture feature extraction and selection can a solid foundation be laid for the development and application of intelligent breast cancer classification systems. Summary of the invention

[0003] To solve the above technical problems, the present invention provides a method and system for identifying and classifying breast cancer pathology images. A method for identifying and classifying breast cancer pathology images includes:

[0004] Collecting and preprocessing breast cancer pathology images to obtain preprocessed breast cancer pathology images;

[0005] Extracting multi-scale texture features of the preprocessed breast cancer pathology image based on a convolutional neural network, capturing local and global texture information of breast tissue by designing convolution kernels with different receptive field sizes, and obtaining texture features;

[0006] Adaptively weighting the texture features through an attention mechanism, dynamically adjusting feature weights according to the importance of different regions, and obtaining key texture features;

[0007] Based on the key texture features, a sequence feature selection algorithm is used to select the optimal texture feature subset, and the most discriminative feature combination is obtained through iterative search and cross-validation;

[0008] A recognition and classification model of breast cancer pathology images based on ensemble learning is constructed according to the feature combination, and the pathology image to be predicted is input into the recognition and classification model for processing to obtain a recognition and classification result.

[0009] Preferably, the process of collecting breast cancer pathology images and performing preprocessing includes:

[0010] A Gaussian filter is used to smooth the image and remove high-frequency noise, and then a histogram equalization algorithm is used to enhance the breast cancer pathology image to enhance the contrast of the image and obtain an enhanced image with more uniform staining distribution and clearer background;

[0011] The enhanced image is subjected to wavelet transformation, and the Daubechies4 wavelet basis function is selected to perform three-layer decomposition on the enhanced image to remove high-frequency noise components in the enhanced image, thereby obtaining a denoised image with higher quality and more obvious texture features.

[0012] Preferably, the process of extracting multi-scale texture features of the preprocessed breast cancer pathology image based on a convolutional neural network includes:

[0013] According to the pre-established texture feature model, the texture features of the denoised image are extracted using the gray-level co-occurrence matrix to obtain the feature vector reflecting the morphological structure of the lesion tissue;

[0014] The feature vector is input into a pre-trained convolutional neural network model, and the deep semantic features of the denoised image are extracted through multi-layer convolution and pooling operations.

[0015] Preferably, the convolutional neural network includes 3 convolutional layers and a fully connected layer, wherein the first layer uses 3×3, 5×5 and 7×7 convolution kernels to extract local texture features respectively, the second layer uses a 2×2 maximum pooling operation to downsample the local texture features, and the third layer uses an 11×11 convolution kernel to extract global texture features;

[0016] Finally, the local texture features and the global texture features are concatenated through the fully connected layer to obtain a multi-scale texture feature representation of the image.

[0017] Preferably, by designing convolution kernels with different receptive field sizes to capture local and global texture information of breast tissue, the process of obtaining texture features includes:

[0018] Performing a convolution operation on the standardized input image using a plurality of convolution kernels with different receptive field sizes to extract local texture features of the standardized input image;

[0019] Performing a pooling operation on the local texture features to obtain downsampled local texture features;

[0020] Using a convolution kernel with a larger receptive field to perform a convolution operation on the standardized input image to extract global texture features of the standardized input image;

[0021] The downsampled local texture features are fused with the global texture features to obtain multi-scale texture features of the standardized input image.

[0022] Preferably, the texture features are adaptively weighted by an attention mechanism, and the feature weights are dynamically adjusted according to the importance of different regions. The process of obtaining key texture features includes:

[0023] Acquire a pre-established breast cancer classification knowledge base, wherein the breast cancer classification knowledge base includes key texture patterns and corresponding importance information; construct a texture feature attention model according to the key texture patterns and the corresponding importance information;

[0024] Inputting the texture features into the texture feature attention model, calculating the importance of different regions through the attention mechanism, and obtaining an adaptive feature weight distribution;

[0025] The texture features are weighted and adjusted according to the feature weight distribution, the features corresponding to the key texture patterns are highlighted, the features of the non-key areas are weakened, and the optimized key texture features are obtained;

[0026] If the key texture feature has insufficient discriminative power in the subsequent classification process, the texture feature attention model is returned, the feature weight is dynamically updated according to the classification feedback, the texture feature is iteratively optimized until the classification requirements are met, the classification result is output, and the iteratively optimized texture feature representation and attention weight are added to the breast cancer classification knowledge base.

[0027] Preferably, the process of weighting and adjusting the texture features according to the feature weight distribution includes:

[0028] The importance of texture features is modeled using gated recurrent units, and the multi-scale texture features of breast cancer pathology images are extracted using a pre-trained Inception-v3 network. The initial feature vector of 1024 dimensions is obtained through global average pooling.

[0029] Then the initial feature vector is input into the attention model, and the attention weight of each region is calculated by the softmax function to obtain a 1024-dimensional adaptive feature weight vector;

[0030] The texture features are weighted and adjusted according to the adaptive feature weight vector, so as to highlight the dimensions corresponding to the key textures and weaken the non-key dimensions, thereby obtaining optimized key texture features.

[0031] Preferably, the process of selecting the optimal texture feature subset based on the key texture features by using a sequence feature selection algorithm, and obtaining the most discriminative feature combination through iterative search and cross-validation includes:

[0032] Using the sequential feature selection algorithm, a number of features are selected from the candidate feature set to form a feature subset through iterative search;

[0033] For each feature subset, the cross-validation method is used to divide the data set into a training set and a test set;

[0034] Train the breast cancer classification model on the training set and evaluate the classification performance of the model on the test set; obtain the optimal feature combination in the current iteration by comparing the classification performance of the models corresponding to different feature subsets;

[0035] Determine whether the performance of the current feature combination is better than that of the previous iteration. If so, update the optimal feature combination, otherwise, keep it unchanged.

[0036] The above iterative search and cross-validation process is repeated until the preset number of iterations is reached or the performance improvement is less than the threshold, and the final optimal feature subset is determined.

[0037] Preferably, a recognition and classification model of breast cancer pathology images based on ensemble learning is constructed according to the feature combination, and the pathology image to be predicted is input into the recognition and classification model for processing. The process of obtaining the recognition and classification result includes:

[0038] Constructing several convolutional neural networks with different structures as base learners, and independently training each base learner using the feature combination to obtain a trained base learner model;

[0039] Input the pathological image to be predicted into the trained base learner model to obtain the classification prediction result of each base learner model;

[0040] According to the performance of each base learner model on the validation set, the weight coefficient of the prediction result is calculated to obtain the weighted voting result;

[0041] The predicted probability values ​​of the base learner model are weighted averaged according to the weight coefficients to obtain the final classification prediction result of the integrated learning model.

[0042] The present invention also provides a recognition and classification system for breast cancer pathology images, comprising:

[0043] An image preprocessing module is used to collect and preprocess breast cancer pathology images to obtain preprocessed breast cancer pathology images;

[0044] A feature extraction module, for extracting multi-scale texture features of the preprocessed breast cancer pathology image based on a convolutional neural network, capturing local and global texture information of breast tissue by designing convolution kernels with different receptive field sizes, and obtaining texture features;

[0045] An attention mechanism module is used to adaptively weight the texture features through an attention mechanism, dynamically adjust feature weights according to the importance of different regions, and obtain key texture features;

[0046] A feature selection module, for selecting an optimal texture feature subset based on the key texture features by using a sequence feature selection algorithm, and obtaining the most discriminative feature combination through iterative search and cross-validation;

[0047] A classification model building module, used to build a recognition and classification model of breast cancer pathology images based on ensemble learning according to the feature combination;

[0048] The model optimization and analysis module is used to input the pathological image to be predicted into the recognition and classification model for processing to obtain the recognition and classification results.

[0049] Compared with the prior art, the present invention has the following advantages and technical effects:

[0050] The present invention discloses a method for classifying breast cancer pathology images based on deep learning. First, the pathology images are preprocessed and enhanced to eliminate interference factors. Then, a multi-scale convolutional neural network is used to extract comprehensive texture features, and an attention mechanism is introduced to highlight key texture features. The optimal feature subset is selected through a sequence feature selection algorithm to construct an integrated learning classification model. During the training process, adversarial sample generation technology is introduced to improve the robustness of the model. Finally, the model is analyzed for interpretability to reveal the basis for classification. The present invention effectively improves the accuracy, reliability and interpretability of breast cancer pathology image classification through the integrated application of multiple deep learning technologies, and provides a powerful intelligent auxiliary tool for clinical classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0052] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;

[0053] Figure 2 The figure is a schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0056] Embodiment 1

[0057] like Figure 1 As shown, this embodiment provides a method for identifying and classifying breast cancer pathology images, including:

[0058] Collecting and preprocessing breast cancer pathology images to obtain preprocessed breast cancer pathology images;

[0059] The multi-scale texture features of the preprocessed breast cancer pathology images are extracted based on convolutional neural networks. The local and global texture information of breast tissue is captured by designing convolution kernels with different receptive field sizes to obtain texture features.

[0060] The texture features are adaptively weighted through the attention mechanism, and the feature weights are dynamically adjusted according to the importance of different regions to obtain key texture features;

[0061] Based on key texture features, a sequential feature selection algorithm is used to select the optimal texture feature subset, and the most discriminative feature combination is obtained through iterative search and cross-validation;

[0062] According to the feature combination, a recognition and classification model of breast cancer pathology images based on ensemble learning is constructed, and the pathology images to be predicted are input into the recognition and classification model for processing to obtain the recognition and classification results.

[0063] Furthermore, the process of collecting and preprocessing breast cancer pathology images includes:

[0064] A breast cancer pathology image dataset is collected, wherein the breast cancer pathology image dataset includes multiple breast cancer pathology images; specifically, this embodiment uses a dataset including 500 breast cancer pathology images, including 300 benign tumor images and 200 malignant tumor images.

[0065] Aiming at the uneven staining and background noise interference factors in breast cancer pathology images, a Gaussian filter is used to smooth the image and remove high-frequency noise. Then, the histogram equalization algorithm is used to enhance the breast cancer pathology image to enhance the image contrast and obtain an enhanced image with more uniform staining distribution and clearer background. By calculating the image grayscale histogram, the grayscale value of the original image is mapped to a uniformly distributed grayscale range, thereby stretching the image contrast and highlighting the detail information.

[0066] The enhanced image is subjected to wavelet transform, and the Daubechies4 wavelet basis function is used to perform three-layer decomposition on the enhanced image to remove the high-frequency noise components in the enhanced image, thus obtaining a denoised image with higher quality and more obvious texture features.

[0067] Furthermore, the process of extracting multi-scale texture features of the preprocessed breast cancer pathology image based on the convolutional neural network includes:

[0068] According to the pre-established texture feature model, the gray-level co-occurrence matrix is ​​used to extract the texture features of the denoised image, and a feature vector reflecting the morphological structure of the diseased tissue is obtained; among them, the texture features of the image are extracted using the gray-level co-occurrence matrix, including 14 features such as contrast, correlation, energy and entropy, to form a feature vector reflecting the morphological structure of the diseased tissue.

[0069] The feature vector is input into the pre-trained convolutional neural network model, and the deep semantic features of the denoised image are extracted through multi-layer convolution and pooling operations.

[0070] Furthermore, the convolutional neural network includes three convolutional layers and a fully connected layer, where the first layer uses 3×3, 5×5, and 7×7 convolution kernels to extract local texture features, respectively, the second layer uses a 2×2 maximum pooling operation to downsample local texture features, and the third layer uses an 11×11 convolution kernel to extract global texture features;

[0071] Finally, the local texture features and the global texture features are concatenated through a fully connected layer to obtain a multi-scale texture feature representation of the image.

[0072] Furthermore, by designing convolution kernels with different receptive field sizes, the local and global texture information of breast tissue is captured. The process of obtaining texture features includes:

[0073] The standardized input image is convolved with multiple convolution kernels of different receptive field sizes to extract the local texture features of the standardized input image;

[0074] Perform a pooling operation on the local texture features to obtain the downsampled local texture features;

[0075] A convolution kernel with a larger receptive field is used to perform a convolution operation on the standardized input image to extract the global texture features of the standardized input image;

[0076] The downsampled local texture features are fused with the global texture features to obtain the standardized multi-scale texture features of the input image.

[0077] Furthermore, the texture features are adaptively weighted through the attention mechanism, and the feature weights are dynamically adjusted according to the importance of different regions. The process of obtaining key texture features includes:

[0078] Obtaining a pre-established breast cancer classification knowledge base, the breast cancer classification knowledge base including key texture patterns and corresponding importance information; constructing a texture feature attention model based on the key texture patterns and the corresponding importance information;

[0079] The texture features are input into the texture feature attention model, and the importance of different regions is calculated through the attention mechanism to obtain an adaptive feature weight distribution;

[0080] The texture features are weighted and adjusted according to the feature weight distribution, the features corresponding to the key texture patterns are highlighted, the features of the non-key areas are weakened, and the optimized key texture features are obtained;

[0081] If the discriminative power of key texture features is insufficient in the subsequent classification process, the texture feature attention model is returned, the feature weights are dynamically updated according to the classification feedback, and the texture features are iteratively optimized until the classification requirements are met. The classification results are output and the iteratively optimized texture feature representation and attention weights are added to the breast cancer classification knowledge base.

[0082] Furthermore, the process of weighting and adjusting the texture features according to the feature weight distribution includes:

[0083] The importance of texture features is modeled using gated recurrent units, and the multi-scale texture features of breast cancer pathology images are extracted using a pre-trained Inception-v3 network. The initial feature vector of 1024 dimensions is obtained through global average pooling.

[0084] Then the initial feature vector is input into the attention model, and the attention weight of each region is calculated by the softmax function to obtain a 1024-dimensional adaptive feature weight vector;

[0085] The texture features are weighted and adjusted according to the adaptive feature weight vector to highlight the dimensions corresponding to the key textures and weaken the non-key dimensions, thus obtaining the optimized key texture features.

[0086] Furthermore, based on the key texture features, a sequential feature selection algorithm is used to select the optimal texture feature subset. Through iterative search and cross-validation, the process of obtaining the most discriminative feature combination includes:

[0087] Using the sequential feature selection algorithm, a number of features are selected from the candidate feature set to form a feature subset through iterative search;

[0088] For each feature subset, the cross-validation method is used to divide the data set into a training set and a test set;

[0089] Train the breast cancer classification model on the training set and evaluate the classification performance of the model on the test set; obtain the optimal feature combination in the current iteration by comparing the classification performance of the models corresponding to different feature subsets;

[0090] Determine whether the performance of the current feature combination is better than that of the previous iteration. If so, update the optimal feature combination, otherwise, keep it unchanged.

[0091] The above iterative search and cross-validation process is repeated until the preset number of iterations is reached or the performance improvement is less than the threshold, and the final optimal feature subset is determined.

[0092] Further optimization scheme, this embodiment adopts sequence forward feature selection algorithm, selects the optimal feature combination from the candidate set by iterative search, and in each iteration, adds a feature in turn from the currently unselected feature to form a new feature subset, and for each feature subset, uses the 5-fold cross validation method to randomly divide the data set into a training set and a test set, wherein the training set accounts for 80% and the test set accounts for 20%. ResNet-50 convolutional neural network is used to train the breast cancer classification model on the training set, and the classification performance of the model is evaluated on the test set, with accuracy, sensitivity and specificity as evaluation indicators. By comparing the classification performance of different feature subsets, the optimal feature combination under the current iteration is obtained. If the average accuracy of the current optimal combination is higher than that of the previous iteration, the optimal feature combination is updated, otherwise it remains unchanged. Repeat the above iterative process until the performance improvement of 5 consecutive iterations is less than 5% and stops. Finally, 20 optimal texture features are selected from 500 candidate features, including gray level co-occurrence matrix, wavelet transform coefficients and Gabor filter response. The classification model trained based on these 20 features achieved an accuracy of 95% on the test set, which is 5 percentage points higher than using all 500 features, indicating that feature selection can remove redundant and irrelevant features and improve the efficiency and performance of breast cancer classification.

[0093] Furthermore, a recognition and classification model of breast cancer pathology images based on ensemble learning is constructed according to the feature combination, and the pathology image to be predicted is input into the recognition and classification model for processing. The process of obtaining the recognition and classification result includes:

[0094] Construct several convolutional neural networks with different structures as base learners, and use feature combinations to train each base learner independently to obtain a trained base learner model;

[0095] Input the pathological image to be predicted into the trained base learner model to obtain the classification prediction result of each base learner model;

[0096] According to the performance of each base learner model on the validation set, the weight coefficient of the prediction result is calculated to obtain the weighted voting result;

[0097] The predicted probability values ​​of the base learner models are weighted averaged according to the weight coefficients to obtain the final classification prediction result of the integrated learning model.

[0098] Further optimization scheme, this embodiment collects breast cancer tissue sample images of 5000 breast cancer patients from multiple hospitals, and can obtain multiple characteristics of normal breast tissue and cancerous tissue by analyzing the tissue structure, cell characteristics, nuclear division images, interstitial changes, vascular changes, tumor margins, necrotic areas, and special staining of breast cancer tissue sample images under a microscope. The breast cancer tissue sample images are preprocessed, uniformly adjusted to 256×256 pixel size, and normalized and data enhanced to expand the data set to 20,000 cases. Then, 5 convolutional neural networks with different structures are constructed as base learners, using classic network architectures such as ResNet-50, Inception-v3, DenseNet-121, EfficientNet-B3 and MobileNetV2. Using the preprocessed medical imaging data, each base learner is trained independently, using the Adam optimizer, the initial learning rate is set to 0.01, the batch size is 32, and 100 epochs are trained to obtain 5 trained base learner models. The performance of each base learner was evaluated on the validation set, and their accuracy, sensitivity and specificity were calculated. The weight coefficient of each model was calculated according to the weighted average formula. When the pathological image to be predicted was input, it was simultaneously input into the five base learners to obtain their respective classification prediction probability values. Then, these probability values ​​were weighted averaged according to the weight coefficients to obtain the final prediction result of the ensemble learning model. The classification performance of the ensemble model was evaluated on the test set. The results showed that the model achieved 95% accuracy, 92% sensitivity and 98% specificity in the breast cancer classification task, which was better than any single base learner, proving the effectiveness of the ensemble learning method in improving the accuracy of breast cancer classification.

[0099] At the same time, during the model training process, adversarial sample generation technology is introduced to slightly perturb the original breast cancer pathology images to simulate images under different quality and interference conditions, thereby enhancing the model's generalization and anti-interference capabilities.

[0100] Specifically, the preprocessed original breast cancer pathology image dataset and the adversarial sample are merged to construct an enhanced training dataset; a deep learning algorithm is used to construct a breast cancer pathology image classification model, and the breast cancer pathology image classification model is trained using the enhanced training dataset; during the training process, the performance index of the breast cancer pathology image classification model on the validation set is obtained, and the generation strategy and proportion of the adversarial sample are dynamically adjusted according to the performance index to obtain the adjusted adversarial sample; the adjusted adversarial sample is merged into the enhanced training dataset to obtain an updated enhanced training dataset, and the breast cancer pathology image classification model is continued to be trained using the updated enhanced training dataset; after the training process is completed, the breast cancer pathology image classification model is evaluated using a test set containing various quality and interference conditions to obtain an evaluation result; according to the evaluation result, at least one of the network structure, loss function, and regularization strategy of the breast cancer pathology image classification model is adjusted to obtain an optimized breast cancer pathology image classification model.

[0101] An interpretability analysis was performed on the trained breast cancer classification model, and visualization technology was used to reveal the key texture areas and discrimination basis that the model focused on, thereby enhancing the credibility and clinical application value of the classification results.

[0102] Specifically, a breast cancer classification model obtained through training is obtained, and an interpretability analysis is performed on the breast cancer classification model; at least one visualization technology is used to visualize the breast cancer classification model to obtain a heat map of key texture areas that the breast cancer classification model focuses on; based on the highlighted areas in the heat map, a key basis for classification and discrimination by the breast cancer classification model is determined; the classification experience of medical experts stored in advance is obtained, and the key texture areas are compared with the classification experience of medical experts to obtain a comparison result; based on the comparison result, the rationality of the area of ​​concern of the breast cancer classification model is evaluated to obtain an evaluation result; if the evaluation result meets the preset conditions, the feature extraction and discrimination process of the breast cancer classification model is optimized based on the evaluation result to obtain an optimized breast cancer classification model; the optimized breast cancer classification model is applied to clinical classification, and the doctor's trust in the classification result is enhanced through the visualization results of the optimized breast cancer classification model; clinical feedback is continuously collected, and the optimized breast cancer classification model is iteratively optimized based on the clinical feedback to enhance the practical value of the optimized breast cancer classification model in breast cancer classification.

[0103] Furthermore, in order to explain the decision-making process of the breast cancer classification model, this embodiment adopts Grad-CAM visualization technology, and generates a heat map with the same size as the original image by calculating the gradient weighted average of the last convolutional layer feature map of the classifier, in which the brighter area represents the key texture that contributes more to the model decision. By analyzing the heat map of 100 test images, it is found that the model mainly focuses on the characteristics of the nuclear morphology, arrangement and staining depth of breast tissue, which is highly consistent with the classification experience summarized by medical experts. In order to further quantify the rationality of the model's focus area, the present invention invites 3 senior pathologists to mark the highlighted area of ​​the heat map, and calculates its IoU (intersection-over-union ratio) with the model's focus area. The results show that the average IoU of the model's focus area and the doctor's annotated area reaches 85, indicating that the model can accurately capture the key basis for breast cancer classification. Based on the inspiration of the visualization results, this study optimizes the feature extraction and discrimination process of the model. An attention mechanism is added to the backbone network, so that the model can adaptively adjust the weights of different regions; at the same time, a discriminator is introduced before the fully connected layer to determine whether the extracted features are consistent with medical prior knowledge, so as to improve the interpretability of the model. The accuracy of the optimized model on the test set is further improved to 97%, and the visualization results are more consistent with the doctors' classification experience. In clinical applications, the model provides doctors with intuitive visual evidence, which effectively enhances the doctors' trust in the classification results. By continuously collecting clinical feedback, the present invention has iteratively optimized the model three times, continuously improving its practical value in breast cancer classification, and finally achieved an accuracy of 95% in 1,000 real cases, proving the feasibility and effectiveness of the model-assisted classification.

[0104] Embodiment 2

[0105] like Figure 2 As shown, this embodiment also provides a recognition and classification system for breast cancer pathology images, including:

[0106] An image preprocessing module is used to collect and preprocess breast cancer pathology images to obtain preprocessed breast cancer pathology images;

[0107] The feature extraction module is used to extract multi-scale texture features of preprocessed breast cancer pathology images based on convolutional neural networks. By designing convolution kernels with different receptive field sizes, the local and global texture information of breast tissue is captured to obtain texture features.

[0108] The attention mechanism module is used to adaptively weight texture features through the attention mechanism, dynamically adjust feature weights according to the importance of different regions, and obtain key texture features;

[0109] A feature selection module is used to select the optimal texture feature subset based on key texture features using a sequence feature selection algorithm, and obtain the most discriminative feature combination through iterative search and cross-validation;

[0110] A classification model building module is used to build a recognition and classification model of breast cancer pathology images based on ensemble learning according to feature combinations;

[0111] The model optimization and analysis module is used to input the pathological image to be predicted into the recognition and classification model for processing to obtain the recognition and classification results.

[0112] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for identifying and classifying breast cancer pathology images, characterized in that: include: Collecting and preprocessing breast cancer pathology images to obtain preprocessed breast cancer pathology images; Extracting multi-scale texture features of the preprocessed breast cancer pathology image based on a convolutional neural network, capturing local and global texture information of breast tissue by designing convolution kernels with different receptive field sizes, and obtaining texture features; Adaptively weighting the texture features through an attention mechanism, dynamically adjusting feature weights according to the importance of different regions, and obtaining key texture features; Based on the key texture features, a sequence feature selection algorithm is used to select the optimal texture feature subset, and the most discriminative feature combination is obtained through iterative search and cross-validation. Specifically, a sequence feature selection algorithm is used to select a number of features from a candidate feature set to form a feature subset through iterative search; For each feature subset, the cross-validation method is used to divide the data set into a training set and a test set; Train the breast cancer classification model on the training set and evaluate the classification performance of the model on the test set; obtain the optimal feature combination in the current iteration by comparing the classification performance of the models corresponding to different feature subsets; Determine whether the performance of the current feature combination is better than that of the previous iteration. If so, update the optimal feature combination, otherwise, keep it unchanged. Repeat the above iterative search and cross-validation process until the preset number of iterations is reached or the performance improvement is less than the threshold, and then determine the final optimal feature subset; A recognition and classification model of breast cancer pathology images based on ensemble learning is constructed according to the feature combination, and the pathology image to be predicted is input into the recognition and classification model for processing to obtain a recognition and classification result.

2. The method for identifying and classifying breast cancer pathology images according to claim 1, characterized in that: The process of collecting and preprocessing breast cancer pathology images includes: A Gaussian filter is used to smooth the image and remove high-frequency noise, and then a histogram equalization algorithm is used to enhance the breast cancer pathology image to enhance the contrast of the image and obtain an enhanced image with more uniform staining distribution and clearer background; The enhanced image is subjected to wavelet transformation, and the Daubechies4 wavelet basis function is selected to perform three-layer decomposition on the enhanced image to remove high-frequency noise components in the enhanced image, thereby obtaining a denoised image with higher quality and more obvious texture features.

3. The method for identifying and classifying breast cancer pathology images according to claim 1, characterized in that: The process of extracting multi-scale texture features of the preprocessed breast cancer pathology image based on a convolutional neural network includes: According to the pre-established texture feature model, the texture features of the denoised image are extracted using the gray-level co-occurrence matrix to obtain the feature vector reflecting the morphological structure of the lesion tissue; The feature vector is input into a pre-trained convolutional neural network model, and the deep semantic features of the denoised image are extracted through multi-layer convolution and pooling operations.

4. The method for identifying and classifying breast cancer pathology images according to claim 3, characterized in that: The convolutional neural network includes three convolutional layers and a fully connected layer, wherein the first layer uses 3×3, 5×5 and 7×7 convolution kernels to extract local texture features respectively, the second layer uses a 2×2 maximum pooling operation to downsample the local texture features, and the third layer uses an 11×11 convolution kernel to extract global texture features; Finally, the local texture features and the global texture features are concatenated through the fully connected layer to obtain a multi-scale texture feature representation of the image.

5. The method for identifying and classifying breast cancer pathology images according to claim 1, characterized in that: By designing convolution kernels with different receptive field sizes, the local and global texture information of breast tissue is captured. The process of obtaining texture features includes: Performing a convolution operation on the standardized input image using a plurality of convolution kernels with different receptive field sizes to extract local texture features of the standardized input image; Performing a pooling operation on the local texture features to obtain downsampled local texture features; Using a convolution kernel with a larger receptive field to perform a convolution operation on the standardized input image to extract global texture features of the standardized input image; The downsampled local texture features are fused with the global texture features to obtain multi-scale texture features of the standardized input image.

6. The method for identifying and classifying breast cancer pathology images according to claim 1, characterized in that: The texture features are adaptively weighted by the attention mechanism, and the feature weights are dynamically adjusted according to the importance of different regions. The process of obtaining key texture features includes: Acquire a pre-established breast cancer classification knowledge base, wherein the breast cancer classification knowledge base includes key texture patterns and corresponding importance information; construct a texture feature attention model according to the key texture patterns and the corresponding importance information; Inputting the texture features into the texture feature attention model, calculating the importance of different regions through the attention mechanism, and obtaining an adaptive feature weight distribution; The texture features are weighted and adjusted according to the feature weight distribution, the features corresponding to the key texture patterns are highlighted, the features of the non-key areas are weakened, and the optimized key texture features are obtained; If the key texture feature has insufficient discriminative power in the subsequent classification process, the texture feature attention model is returned, the feature weight is dynamically updated according to the classification feedback, the texture feature is iteratively optimized until the classification requirements are met, the classification result is output, and the iteratively optimized texture feature representation and attention weight are added to the breast cancer classification knowledge base.

7. The method for identifying and classifying breast cancer pathology images according to claim 6, characterized in that: The process of weighting and adjusting the texture feature according to the feature weight distribution includes: The importance of texture features is modeled using gated recurrent units, and the multi-scale texture features of breast cancer pathology images are extracted using a pre-trained Inception-v3 network. The initial feature vector of 1024 dimensions is obtained through global average pooling. Then the initial feature vector is input into the attention model, and the attention weight of each region is calculated by the softmax function to obtain a 1024-dimensional adaptive feature weight vector; The texture features are weighted and adjusted according to the adaptive feature weight vector, so as to highlight the dimensions corresponding to the key textures and weaken the non-key dimensions, thereby obtaining optimized key texture features.

8. The method for identifying and classifying breast cancer pathology images according to claim 1, characterized in that: The process of constructing a recognition and classification model of breast cancer pathology images based on ensemble learning according to the feature combination, inputting the pathology image to be predicted into the recognition and classification model for processing, and obtaining the recognition and classification results includes: Constructing several convolutional neural networks with different structures as base learners, and independently training each base learner using the feature combination to obtain a trained base learner model; Input the pathological image to be predicted into the trained base learner model to obtain the classification prediction result of each base learner model; According to the performance of each base learner model on the validation set, the weight coefficient of the prediction result is calculated to obtain the weighted voting result; The predicted probability values ​​of the base learner model are weighted averaged according to the weight coefficients to obtain the final classification prediction result of the integrated learning model.

9. A recognition and classification system for breast cancer pathology images, characterized in that: include: An image preprocessing module is used to collect and preprocess breast cancer pathology images to obtain preprocessed breast cancer pathology images; A feature extraction module, for extracting multi-scale texture features of the preprocessed breast cancer pathology image based on a convolutional neural network, capturing local and global texture information of breast tissue by designing convolution kernels with different receptive field sizes, and obtaining texture features; An attention mechanism module is used to adaptively weight the texture features through an attention mechanism, dynamically adjust feature weights according to the importance of different regions, and obtain key texture features; A feature selection module is used to select an optimal texture feature subset based on the key texture features by using a sequence feature selection algorithm, and obtain the most discriminative feature combination through iterative search and cross-validation. Specifically, a sequence feature selection algorithm is used to select a number of features from a candidate feature set through iterative search to form a feature subset; For each feature subset, the cross-validation method is used to divide the data set into a training set and a test set; Train the breast cancer classification model on the training set and evaluate the classification performance of the model on the test set; obtain the optimal feature combination in the current iteration by comparing the classification performance of the models corresponding to different feature subsets; Determine whether the performance of the current feature combination is better than that of the previous iteration. If so, update the optimal feature combination, otherwise, keep it unchanged. Repeat the above iterative search and cross-validation process until the preset number of iterations is reached or the performance improvement is less than the threshold, and then determine the final optimal feature subset; A classification model building module, used to build a recognition and classification model of breast cancer pathology images based on ensemble learning according to the feature combination; The model optimization and analysis module is used to input the pathological image to be predicted into the recognition and classification model for processing to obtain the recognition and classification results.

Images