A method for predicting HER2 status changes based on radiomics attention network

Abstract

The present invention discloses a HER2 state change prediction method based on an ultrasound imaging genomics feature attention network, comprising the steps of outlining the tumor area in the ultrasound image of breast cancer; extracting the imaging genomics features of the ultrasound image using imaging genomics technology to obtain an imaging genomics feature matrix; selecting features of the imaging genomics feature matrix to form a key feature matrix; inputting it into a network module with an attention mechanism, outputting a key feature matrix with attention weights related to the HER2 state change; finally, training the key feature matrix with attention weights using a fully connected neural network, thereby obtaining a prediction of whether the HER2 state of breast cancer has changed. The present invention realizes the selection and learning of features related to HER2 state change by combining imaging genomics technology with a deep learning model, and predicts whether the HER2 state has changed after neoadjuvant therapy by using a fully connected neural network.

Description

Technical Field

[0001] The present invention belongs to the field of medical information processing, and specifically relates to a method for predicting changes in HER2 status of breast cancer based on ultrasound imaging genomics attention network. Background Art

[0002] Breast cancer is the most common disease in women worldwide. It is estimated that by 2040, the incidence of breast cancer in the world will increase to more than 3 million new cases and 1 million deaths per year due to population growth and aging alone. HER2 is an established treatment target for most breast cancer patients, and targeted therapy is effective for breast cancer patients with HER2 overexpression and / or amplification. According to the testing practice guidelines issued by the American Society of Clinical Oncology and the American College of Pathologists, HER2 positivity, i.e. overexpression, is defined as: immunohistochemistry score 3+ (IHC3+) or IHC2+ in situ hybridization positive, and all states other than positive are simply classified as negative. HER2 expression status can be divided into three types, namely HER2 positive, HER2 zero expression, and HER2 low expression. There are great differences in the treatment effects of breast cancer patients with different HER2 status using the same anti-HER2 treatment. Therefore, by studying HER2 status, the effect of therapeutic drugs can be better exerted, the blindness of treatment can be reduced, and new treatment opportunities will be brought to a certain proportion of patients.

[0003] Early studies on HER2 status focused on identifying its status through magnetic resonance images. Some researchers used radiomics knowledge to evaluate the HER2 expression level in patients with breast cancer who were overexpressing HER2 and received neoadjuvant chemotherapy using magnetic resonance images, and found that intra-tumor radiomic heterogeneity was not equal to pathological intra-tumor heterogeneity. Other researchers used three-dimensional visualization software to extract radiomics features, and used minimum absolute shrinkage and regression models for feature selection, realizing clinical radiomics feature extraction, showing good classification function, and identifying the HER2 expression status of breast cancer patients.

[0004] Recently, researchers have reviewed the value of HER2 status as a prognostic indicator and clinical guide for breast cancer treatment. It is believed that the most practical application of detecting HER2 expression in breast cancer is that it can predict the tumor's response to treatment and thus guide treatment, which strongly supports that this examination should be used as a routine medical examination. However, current studies only focus on the determination of HER2 status and cannot predict changes in patients after receiving neoadjuvant therapy.

[0005] In recent years, deep learning has developed rapidly as a research hotspot in the field of artificial intelligence, especially in image data processing. At the same time, some deep learning-based methods have been proposed for diagnosing HER2 status. However, these methods use deep learning methods to learn the feature representation of other imaging data, and then identify HER2 status based on the learned features. They do not use ultrasound images and imaging omics technology to predict whether HER2 status changes after neoadjuvant therapy. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a method for predicting the change of human epidermal growth factor receptor-2 (HER2) status in breast cancer based on ultrasound imaging feature attention network, which can obtain a large number of imaging features of ultrasound image data, and realize the selection and learning of features related to HER2 status change by combining imaging technology with deep learning model, and classify breast cancer ultrasound images through fully connected neural network, thereby realizing the prediction of whether HER2 status changes after neoadjuvant therapy. .

[0007] The present invention provides a method for predicting changes in HER2 status of breast cancer based on ultrasound imaging feature attention network, specifically:

[0008] Step 1: delineate the tumor area in the breast cancer ultrasound image, and then segment the tumor area and convert it into a tumor area image in the format of the Neuro-Imaging Informatics Technology Initiative;

[0009] Step 2, extracting the imaging features of the tumor region image to obtain an imaging feature matrix X;

[0010] Step 3: Perform feature selection on the imaging genomics feature matrix X to obtain a key feature matrix

[0011] Step 4: Transform the key feature matrix Input a network module with an attention mechanism, weight the attention of different key features, and output a key feature matrix with attention weights related to HER2 status changes;

[0012] Step 5: Use a fully connected neural network to train the key feature matrix with attention weights to obtain a classification result, thereby predicting whether the HER2 status of breast cancer has changed.

[0013] Furthermore, the specific content of step 1 is: outlining the tumor region on N breast cancer ultrasound images, and then segmenting the tumor region and converting it into a tumor region image in the format of the Neuro-Imaging Informatics Technology Initiative.

[0014] Furthermore, in step 2, specifically, the Pyradiomics toolkit is used to extract the imaging omics features of the tumor region image, and the imaging omics features include ultrasound image shape features, first-order features, texture and high-order features.

[0015] Furthermore, the shape features include surface area, perimeter and maximum two-dimensional diameter; the surface area refers to the surface area of ​​the tumor region; the perimeter refers to the length of the edge of the tumor region; the maximum two-dimensional diameter is the maximum pairwise Euclidean distance between tumor surfaces.

[0016] Furthermore, the first-order features include first-order energy features, entropy features and root mean square features; the first-order energy features are a measure of the size of voxel values ​​in the tumor area image; the entropy features represent the randomness of the image values; and the root mean square features are the square root of the square mean of all intensity values.

[0017] Furthermore, the texture and high-order features include grayscale co-occurrence matrix features, grayscale size area matrix features, grayscale run length matrix features, adjacent grayscale difference matrix features, and grayscale correlation matrix features; the grayscale co-occurrence matrix features are used to measure the fineness and coarseness of the texture; the grayscale size area matrix features are used to measure the index of the distribution of small-sized areas, and the larger the value, the more small areas there are and the finer the texture; the grayscale run length matrix features are indicators for measuring the distribution of run lengths, and are used to describe the coarseness of the texture; the adjacent grayscale difference matrix features are used to measure the average difference between the central voxel and its neighboring voxels, and are an indication of the spatial rate of change; the grayscale correlation matrix features are used to indicate the magnitude of the dependence intensity.

[0018] Furthermore, in step 3, the LassoNet module is used to perform feature selection on the imaging feature matrix X, retaining the key features related to the change of HER2 status to form a key feature matrix Specifically:

[0019] The LassoNet module has a total of L layers, and its learning method is implemented by the following sub-steps:

[0020] Assume that the weight matrix of the LassoNet module is W, and the network output of the lth layer is The weight matrix is ​​W (l) ; The final network output label f ​​of the LassoNet module θ,W (X) is obtained by learning the radiomics feature matrix X, and its objective function is:

[0021]

[0022] in, is the weight of the jth feature in the first hidden layer network, θ is the sparse parameter vector, θj is the sparse parameter of the jth feature, Y is the real data label, l(f θ,W (X), Y) represents the loss calculation of the network prediction label and the true result label, and l is the cross entropy loss function; It means that LassoNet adjusts To adjust the sparsity and nonlinear representation of network features, where M is the sparsity and nonlinear control parameter, for Infinity norm; the value range of M is [0, +∞): When M = 0, all other hidden layer nodes are inactive, only At this time, the network is an ordinary linear Lasso regularization; when M tends to positive infinity, that is, M→+∞, the network becomes a standard nonlinear feedforward neural network; therefore, the LassoNet model selects appropriate θ and M values ​​to adjust the sparsity and nonlinear representation of network features respectively;

[0023] After LassoNet training, the sparse parameters of features that are irrelevant or weakly correlated with HER2 status changes are constrained to 0, and then these features are removed from the matrix X, retaining the key features related to HER2 status changes, that is, the key feature matrix is ​​obtained Among them, d is the dimension of the data after feature selection, and there are d <p。

[0024] Furthermore, in step 4, the key feature matrix Input a network module with an attention mechanism, weight the attention of different key features, and output a key feature matrix with attention weights related to HER2 status changes, which is achieved by the following sub-steps:

[0025] Step 4.1: transform the key feature matrix Input to the network module with attention mechanism, after the parallel process of global average pooling operation and global maximum pooling operation, the intermediate feature matrix is ​​obtained.

[0026]

[0027] Among them, F Gp () represents the global average pooling and maximum pooling operations mentioned above;

[0028] Step 4.2: The intermediate feature matrix X' is weighted through two fully connected layers to obtain the network output f(X',W)

[0029]

[0030] in, represents the ReLU activation function, W1 and W 2 Represent the weight matrices of the two fully connected layers respectively; f(X',W) is activated by sigmoid to obtain the weight vector W with attention A , expressed as follows:

[0031] W A =F w (X',W)=g(f(X',W)) (4)

[0032] Among them, g() represents the sigmoid activation function, F w Represents the entire attention weighting process;

[0033] Step 4.3: The attention weight vector W A With key feature matrix After a dot multiplication between the vectors, we get the feature matrix X with attention weights. A , the steps are as follows:

[0034]

[0035] Among them, ⊙ represents the dot product operation between vectors;

[0036] Finally, we get the key feature matrix with attention weights

[0037] Furthermore, the cross entropy loss function of the breast cancer HER2 status change prediction method is:

[0038]

[0039] Among them, y p (x n ) is the probability of the HER2 status of the nth patient sample changing, y q (x n ) is the probability of HER2 status change of the nth patient sample predicted by the radiomics attention network.

[0040] Beneficial effects: The present invention combines radiomics technology with a deep learning model to perform radiomics feature extraction, key feature selection, key feature weighting and classification tasks on breast cancer ultrasound images, thereby realizing the prediction of whether HER2 status changes after neoadjuvant therapy, which has the following advantages:

[0041] (1) Using radiomics technology to extract a large number of radiomics features such as breast cancer ultrasound image shape, first-order features, and texture features, and using the radiomics features for subsequent deep learning model training to integrate the representation information of ultrasound image data into the deep learning model;

[0042] (2) LassoNet was used to complete the radiomics feature selection. By applying sparse constraints to each feature, irrelevant and redundant features were removed, and key features related to HER2 status changes were retained;

[0043] (3) The key feature matrix is ​​input into the attention weighted network module to achieve adaptive weighting of different key features to measure the contribution of different key features to the HER2 status change prediction task;

[0044] (4) A fully connected neural network is used to learn key feature data with attention weights to obtain the predicted probability of a change in the HER2 status of patient samples, thereby achieving the HER2 status change prediction task. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to illustrate the embodiments of the invention. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.

[0046] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0047] 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.

[0048] The embodiment of the present invention provides a method for predicting the state change of human epidermal growth factor receptor-2 (HER2) in breast cancer based on ultrasound imaging feature attention network, and the specific steps are as follows:

[0049] Step 1: Given N breast cancer ultrasound images, experienced senior doctors outline the tumor area on the breast cancer ultrasound images, and then segment the tumor area and convert it into a tumor area image in the format of the Neuro-Imaging Informatics Technology Initiative.

[0050] Step 2: Use radiomics technology to calculate and extract a large number of radiomics features such as ultrasound image shape, first-order features, and texture features to obtain the radiomics feature matrix:

[0051] The Pyradiomics toolkit was used to extract radiomic features from medical images in the Neuroimaging Informatics Technology Initiative format. There were a total of p = 1333 extracted radiomic features, including 267 first-order features (accounting for 20.0%), 16 shape features (accounting for 1.2%), and 1050 texture and high-order features (accounting for 78.8%). The texture and high-order features included 335 grayscale co-occurrence matrix features (accounting for 31.9%), 224 grayscale size area matrix features (accounting for 21.3%), 225 grayscale run length matrix features (accounting for 21.4%), 71 neighboring grayscale difference matrix features (accounting for 6.8%), and 195 grayscale correlation matrix features (accounting for 18.6%).

[0052] The first-order features describe the distribution of voxel values ​​and intensity values ​​in the tumor area through basic metrics, such as first-order energy features, entropy features, and root mean square, etc. Among them, the first-order energy feature is a measure of the size of the voxel value in the image; the entropy feature represents the randomness of the image value; and the root mean square feature is the square root of the mean of the squares of all intensity values.

[0053] The shape features specifically include surface area, perimeter, and maximum two-dimensional diameter, etc. The surface area refers to the surface area of ​​the tumor region in the ultrasound image; the perimeter refers to the length of the edge of the tumor region; and the maximum two-dimensional diameter is the maximum pairwise Euclidean distance between the tumor surfaces.

[0054] Texture and high-order features include grayscale co-occurrence matrix features, grayscale size area matrix features, grayscale run length matrix features, adjacent grayscale difference matrix features and grayscale correlation matrix features.

[0055] The main feature of the grayscale co-occurrence matrix feature is the autocorrelation feature, which is used to measure the fineness and coarseness of the texture; the main feature of the grayscale size area matrix feature is the small area emphasis feature, which is an indicator for measuring the distribution of small-sized areas. The larger the value, the more small areas there are and the finer the texture; the main feature of the grayscale run length matrix feature is the long and short-term run emphasis, which is an indicator for measuring the distribution of run length and can describe the coarseness of the texture; the main feature of the neighboring grayscale difference matrix feature is the coarseness feature, which is used to measure the average difference between the central voxel and its neighboring voxels and is an indication of the spatial rate of change; the main feature of the grayscale correlation matrix feature is the dependence emphasis feature, which is used to indicate the magnitude of the dependence intensity.

[0056] Finally, all the extracted p features are used as the imaging omics feature matrix

[0057] Step 3, the imaging genomics feature matrix X generated in the previous step is used to complete feature selection using the LassoNet module. By calculating the sparse coefficient of each feature, irrelevant and redundant features are removed, and key features related to HER2 status changes are retained to form a key feature matrix;

[0058] The LassoNet module has a total of L layers, and its learning method is implemented by the following sub-steps:

[0059] Assume that the weight matrix of the LassoNet module is W, and the network output of the lth layer is The weight matrix is ​​W (l) The final network output label f ​​of the LassoNet module is θ,W (X) is obtained by learning the radiomics feature matrix X, and its objective function is:

[0060]

[0061] in, is the weight of the jth feature in the first hidden layer network, θ is the sparse parameter vector, θ j is the sparse parameter of the jth feature, Y is the real data label, l(f θ,W (X), Y) represents the loss calculation of the network prediction label and the true result label, and l is the cross entropy loss function. It means that LassoNet adjusts To adjust the sparsity and nonlinear representation of network features, where M is the sparsity and nonlinear control parameter, for The infinite norm of M. The value range of M is [0, +∞): When M = 0, all other hidden layer nodes are inactive, only At this time, the network is a normal linear Lasso regularization; when M tends to positive infinity, that is, M→+∞, the network becomes a standard nonlinear feedforward neural network. Therefore, LassoNet can adjust the sparsity and nonlinear representation of network features by selecting a suitable M value.

[0062] The goal of LassoNet is not only to make the network sparse, but also to select relevant input features of the entire network in a structured way. Since the network is feedforward, only the weights of the first hidden layer need to be constrained, and there is no need to penalize the weights of the remaining hidden layers.

[0063] After LassoNet training, the sparse parameters of features that are irrelevant or weakly correlated with HER2 status changes are constrained to 0, and then these features are removed from the data X, retaining the key features related to HER2 status changes, achieving the effect of key feature selection. The final output of the LassoNet module is the key feature matrix related to HER2 status changes Among them, d is the dimension of the data after feature selection, and there are d <p。

[0064] Step 4: Through the key feature matrix Input the network module with attention mechanism to realize the weighted attention of different key features, and output the key feature matrix with attention weights related to the change of HER2 status, which is realized by the following sub-steps:

[0065] Step 4.1, for the output of the LassoNet module It is input into the network module with attention mechanism, and after the parallel process of global average pooling operation and global maximum pooling operation, the intermediate feature matrix is ​​obtained.

[0066]

[0067] Among them, F Gp () represents the global average pooling and maximum pooling operations mentioned above;

[0068] Step 4.2: The intermediate feature matrix X' is weighted through two fully connected layers to obtain the network output f(X',W)

[0069]

[0070] in, represents the ReLU activation function, W 1 and W 2 Represent the weight matrices of the two fully connected layers respectively. After sigmoid activation, f(X',W) obtains the weight vector W with attention A , which can be expressed by the following formula:

[0071]

[0072] Among them, g() represents the sigmoid activation function, F w Represents the entire attention weighting process.

[0073] Step 4.3, after learning in the previous step, the attention weight vector W is obtained A , and then with the key feature matrix After a dot multiplication between the vectors, we get the feature matrix X with attention weights.A , the steps are as follows:

[0074]

[0075] Among them, ⊙ represents the dot product operation between vectors.

[0076] Finally, the attention-weighted feature matrix is ​​obtained

[0077] Step 5, using a fully connected neural network to train the key feature matrix with attention weights to obtain a classification result of the model, thereby realizing the prediction of whether the HER2 status of breast cancer has changed;

[0078] The cross entropy loss function of the breast cancer HER2 status change prediction method based on ultrasound imaging omics attention network is:

[0079]

[0080] Among them, y p (x n ) is the probability of the HER2 status of the nth patient sample changing, y q (x n ) is the probability of change in HER2 status of the nth patient sample predicted by the radiomics attention network. The radiomics attention network includes a radiomics feature extraction module, a LassoNet module, a network module with an attention mechanism, and a fully connected neural network; the radiomics feature extraction module is a Pyradiomics toolkit.

[0081] The radiomics attention network model proposed in this application is optimized through multiple iterations to obtain the predicted probability Y of the final HER2 status change of all N patient samples. q , and there is y q (x n )∈Y q In Y q In the above analysis, if the probability of the patient sample is greater than 0.5, it means that the HER2 status of the patient will change after neoadjuvant therapy; if the probability of the patient sample is less than 0.5, it means that the HER2 status of the patient will not change after neoadjuvant therapy.

[0082] The final output of the radiomics attention network model is Y q That is, the predicted result of HER2 status change predicted for the patient sample.

[0083] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The specific steps are: Step 1: delineate the tumor area in the breast cancer ultrasound image, and then segment the tumor area and convert it into a tumor area image in the format of the Neuro-Imaging Informatics Technology Initiative; Step 2, extracting the imaging features of the tumor region image to obtain an imaging feature matrix X; Step 3, using the LassoNet module to perform feature selection on the imaging genomics feature matrix X. After the LassoNet module training is completed, the sparse parameters of the features that are irrelevant or weakly correlated with the change in the HER2 status of breast cancer are constrained to 0, and then these features are removed from the matrix X, and the key features related to the change in the HER2 status of breast cancer are retained, that is, the key feature matrix is ​​obtained. Step 4: Transform the key feature matrix Input a network module with an attention mechanism, weight the attention of different key features, and output a key feature matrix with attention weights related to changes in HER2 status of breast cancer; Step 5: Use a fully connected neural network to train the key feature matrix with attention weights to obtain a classification result, thereby predicting whether the HER2 status of breast cancer has changed.

2. According to claim 1, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The specific content of step 1 is: outlining the tumor area on N breast cancer ultrasound images, and then segmenting the tumor area and converting it into a tumor area image in the format of the Neuro-Imaging Informatics Technology Initiative.

3. According to claim 1, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: Specifically, step 2 includes using the Pyradiomics toolkit to extract the imaging features of the tumor region image, wherein the imaging features include ultrasound image shape features, first-order features, texture, and high-order features.

4. According to claim 3, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The shape characteristics include surface area, perimeter, and maximum two-dimensional diameter; The surface area refers to the surface area of ​​the tumor region; The perimeter refers to the length around the edge of the tumor area; The maximum two-dimensional diameter is the maximum pairwise Euclidean distance between tumor surfaces.

5. According to claim 3, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The first-order features include first-order energy features, entropy features, and root mean square features; The first-order energy feature is a measure of the size of the voxel value in the tumor region image; The entropy feature represents the randomness of image values; The RMS feature is the square root of the mean of the squares of all intensity values.

6. According to claim 3, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The texture and high-order features include grayscale co-occurrence matrix features, grayscale size region matrix features, grayscale run length matrix features, adjacent grayscale difference matrix features, and grayscale correlation matrix features; The gray level co-occurrence matrix feature is used to measure the fineness and coarseness of the texture; The grayscale size area matrix feature is used to measure the distribution of small-size areas. The larger the value, the more small-size areas there are and the finer the texture. The grayscale run length matrix feature is an indicator for measuring the distribution of run lengths and is used to describe the thickness of textures; the neighboring grayscale difference matrix feature is used to measure the average difference between the central voxel and its neighboring voxels and is an indicator of the spatial change rate; The grayscale correlation matrix feature is used to represent the magnitude of the dependence intensity.

7. According to claim 1, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The LassoNet module described in step 3 has a total of L layers, and its learning method is implemented by the following sub-steps: Assume that the weight matrix of the LassoNet module is W, and the network output of the lth layer is The weight matrix is ​​W (l) ; LassoNet module output label f θ,W (X) is obtained by learning the radiomics feature matrix X, and its objective function is: in, is the weight of the jth feature in the first hidden layer, θ is the sparse parameter vector, θ j is the sparse parameter of the jth feature, Y is the real data label, l(f θ,W (X), Y) represents the loss calculation of the LassoNet module prediction label and the true result label, where l is the cross entropy loss function; It means that the LassoNet module is adjusted To adjust the sparsity and nonlinear representation of LassoNet module features, where M is the sparsity and nonlinear control parameter, for Infinity norm; the value range of M is [0, +∞): When M = 0, all other hidden layer nodes are inactive, only At this time, the LassoNet module is an ordinary linear Lasso regularization; when M tends to positive infinity, that is, M→+∞, the LassoNet module becomes a standard nonlinear feedforward neural network; therefore, the LassoNet module selects appropriate θ and M values ​​to adjust the sparsity and nonlinear representation of the LassoNet module features respectively.

8. According to claim 1, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: Step 4 is implemented by the following sub-steps: Step 4.1: transform the key feature matrix Input to the network module with attention mechanism, after the parallel process of global average pooling operation and global maximum pooling operation, the intermediate feature matrix is ​​obtained. Among them, F Gp () represents the global average pooling and maximum pooling operations mentioned above; Step 4.2: The intermediate feature matrix X' is weighted through two fully connected layers to obtain the output f(X',W) in, represents the ReLU activation function, W1 and W2 represent the weight matrices of the two fully connected layers respectively; f(X',W) is activated by sigmoid to obtain the weight vector W with attention A , expressed as follows: W A =F w (X',W)=g(f(X',W)) (4) Among them, g() represents the sigmoid activation function, F w Represents the entire attention weighting process; Step 4.3: The attention weight vector W A With key feature matrix After a dot multiplication between the vectors, we get the feature matrix X with attention weights. A , the steps are as follows: Among them, ⊙ represents the dot product operation between vectors; Finally, we get the key feature matrix with attention weights 9. According to claim 1, a method for predicting HER2 status changes in breast cancer based on ultrasound imaging feature attention network, characterized in that: The cross entropy loss function of the breast cancer HER2 status change prediction method is: Among them, y p (x n ) is the probability of the nth patient sample's breast cancer HER2 status changing, y q (x n ) is the probability of change in the HER2 status of breast cancer in the nth patient sample predicted by the attention network based on ultrasound imaging features.

Images