Breast cancer her2 gene mutation prediction method and device based on ABVS

Abstract

The application discloses a breast cancer HER2 gene mutation prediction method and device based on ABVS, and comprises the following steps: S1, acquiring an ABVS image; S2, performing preliminary segmentation on the ABVS image to obtain a breast tissue part; S3, performing secondary segmentation on the basis of the segmented breast tissue part to obtain a lesion ROI region; S4, obtaining an imageomics feature according to the lesion ROI region, wherein the imageomics feature comprises first-order gray statistical features, shape features, texture features and transformation-based features; and S5, combining segmentation features generated in the segmentation process with the imageomics feature, and using a deep feature classifier to predict a HER2 gene mutation state. The application predicts the HER2 mutation state in breast cancer by using a non-invasive method, is accurate in prediction, and provides help for clinical treatment of breast cancer.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a breast cancer HER2 gene mutation prediction method and device based on ABVS. BACKGROUND

[0002] A receptor type tyrosine kinase-human epidermal growth factor receptor 2 (HER2) has gene amplification and protein overexpression in 20% to 30% of primary breast ductal papillomas. HER2-positive breast tumors have strong invasion and short disease-free survival, and have poor prognosis.

[0003] Among breast cancer risk level evaluation factors, although the lymph node status is still the most important factor, the HER2 status directly affects the risk level. When the lymph node is negative or only 1 to 3 lymph nodes have metastasis, if HER2 is overexpressed or gene amplified, the risk level is increased from low to medium and from medium to high, respectively. Existing inventions show that the molecular biological behavior of a tumor determines the pathological morphological characteristics of the lesion, which makes it possible to explore the correlation between molecular biology and imaging.

[0004] Currently, HER2-positive mutations can only be detected after HER2-positive mutations, and generally, after the patient appears HER2-positive, immunohistochemistry (IHC) is used to detect HER2 receptor protein overexpression, and fluorescence in situ hybridization (FISH) and chromogenic in situ hybridization (CISH) are used to detect the level of HER2 gene amplification.

[0005] However, various factors such as fixation and preservation of the specimen, selection of the antibody and probe, interpretation of the results, and experience of the operator can cause deviation of the detection results.

[0006] Existing CT image-based methods can predict HER2-positive mutations, but have low resolution and are prone to misjudgment. SUMMARY

[0007] To solve the above technical problems, the application provides a breast cancer HER2 gene mutation prediction method and device based on ABVS.

[0008] To achieve the above purpose, the technical scheme of the application is as follows:

[0009] On the one hand, the application discloses a breast cancer HER2 gene mutation prediction method based on ABVS, comprising the following steps:

[0010] Step S1: acquiring an ABVS image;

[0011] Step S2: preliminary segmentation is performed on the ABVS image to obtain a breast tissue part;

[0012] Step S3: secondary segmentation is performed on the basis of the segmented breast tissue part to obtain a lesion ROI region;

[0013] Step S4: according to the lesion ROI region, an imageomic feature is obtained, wherein the imageomic feature includes: first-order gray statistical feature, shape feature, texture feature and feature based on transformation;

[0014] Step S5: the segmentation features generated in the segmentation process and the imageomic features are combined, and a deep feature classifier is used to predict the HER2 gene mutation state.

[0015] On the basis of the above technical solutions, the following improvements can be made:

[0016] As a preferred scheme, step S1 further includes: pre-processing the obtained ABVS image, and the pre-processing includes one or more of the following operations: isotropic interpolation of each layer of ABVS, adjusting image gain, normalization processing.

[0017] As a preferred scheme, step S2 specifically includes:

[0018] Step S2.1: using a first segmentation model to perform preliminary segmentation on the ABVS image to obtain a preliminary segmentation region;

[0019] Step S2.2: performing morphological closing operation on the preliminary segmentation region to obtain the breast tissue part.

[0020] As a preferred scheme, step S3 specifically includes:

[0021] Step S3.1: generating a feature map according to the breast tissue part;

[0022] Step S3.2: extracting channel attention according to the feature map, and dividing the feature map into three parts of axial plane, coronal plane and sagittal plane according to the multi-view of the ABVS image, using learnable parameters to give corresponding weights, and then extracting spatial attention;

[0023] Step S3.4: fusing the channel attention, spatial attention and feature map according to the weight coefficient;

[0024] Step S3.5: on the basis of the fused feature map, a second segmentation model is used for secondary segmentation to obtain the lesion ROI region.

[0025] As a preferred scheme, step S5 specifically includes:

[0026] Step S5.1: combine the segmentation features generated by the segmentation process and the radiomics features to form joint features;

[0027] Step S5.2: perform dimension reduction processing on the joint features using a dimension reduction algorithm;

[0028] Step S5.3: use a deep feature classifier to predict the HER2 gene mutation state.

[0029] In another aspect, the present application also discloses an ABVS-based breast cancer HER2 gene mutation prediction device, comprising:

[0030] An acquisition module is configured to acquire an ABVS image.

[0031] A preliminary segmentation module is configured to preliminarily segment the ABVS image to obtain a breast tissue part.

[0032] A secondary segmentation module is configured to perform secondary segmentation based on the segmented breast tissue part to obtain a lesion ROI region.

[0033] A radiomics feature acquisition module is configured to acquire radiomics features from the lesion ROI region, wherein the radiomics features include first-order gray statistical features, shape features, texture features, and transformation-based features.

[0034] A prediction module is configured to combine the segmentation features generated by the segmentation process and the radiomics features, and use a deep feature classifier to predict the HER2 gene mutation state.

[0035] As a preferred scheme, the acquisition module comprises:

[0036] An acquisition unit is configured to acquire an ABVS image.

[0037] A preprocessing unit is configured to preprocess the acquired ABVS image, and the preprocessing includes isotropic interpolation of each layer of the ABVS, adjustment of image gain, and normalization processing.

[0038] As a preferred scheme, the preliminary segmentation module specifically comprises:

[0039] A preliminary segmentation unit is configured to preliminarily segment the ABVS image using a first segmentation model to obtain a preliminary segmentation region.

[0040] A morphological closing operation unit is configured to perform morphological closing operation on the preliminary segmentation region to obtain the breast tissue part.

[0041] As a preferred scheme, the secondary segmentation module specifically comprises:

[0042] A feature map generation unit is configured to generate a feature map according to the breast tissue part.

[0043] a multi-attention extraction unit configured to extract channel attention according to the feature map, and divide the feature map into three parts of an axial plane, a coronal plane and a sagittal plane according to multiple perspectives of the ABVS image, and use learnable parameters to assign corresponding weights and then extract spatial attention thereof;

[0044] a fusion unit configured to fuse the channel attention, the spatial attention and the feature map according to the weight coefficients;

[0045] a secondary segmentation unit configured to perform secondary segmentation on the fused feature map by using a second segmentation model to obtain a lesion ROI region.

[0046] As a preferred scheme, the prediction module specifically comprises:

[0047] a joint feature forming unit configured to combine the segmentation features generated in the segmentation process and the radiomics features to form joint features;

[0048] a dimension reduction unit configured to perform dimension reduction processing on the joint features by using a dimension reduction algorithm;

[0049] a prediction unit configured to predict the HER2 gene mutation state by using a deep feature classifier.

[0050] The application discloses a breast cancer HER2 gene mutation prediction method and device based on ABVS, which performs secondary segmentation on the basis of an ABVS image, performs first segmentation on a breast tissue part, and performs second segmentation on a lesion ROI region.

[0051] The application predicts the HER2 mutation state in breast cancer by using a non-invasive method, is accurate in prediction, and provides help for clinical treatment of breast cancer. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0053] Figure 1 The flowchart of the breast cancer HER2 gene mutation prediction method provided by the embodiments of the application.

[0054] Figure 2A step block diagram of the breast cancer HER2 gene mutation prediction method provided by the embodiment of the present application is provided.

[0055] Figure 3 A pretreatment and breast tissue partial segmentation image provided by the embodiment of the present application is provided.

[0056] Figure 4 A fusion attention block diagram provided by the embodiment of the present application is provided.

[0057] Figure 5 A composition diagram of an imageomics feature group provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0058] The preferred embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0059] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0060] The ordinal numbers “first”, “second”, “third” and the like are used to describe common objects only to indicate different instances of the similar objects, and are not intended to imply that the objects thus described must have a given order in time, space, sequence or in any other way.

[0061] In addition, the expression “including” the element is an “open” expression, which only means that the corresponding component exists, and should not be interpreted as excluding additional components.

[0062] In order to achieve the purpose of the present application, in some embodiments of the breast cancer HER2 gene mutation prediction method and device based on ABVS, the ABVS image is used for prediction. The resolution of the ABVS image for soft tissue is higher than that of the CT image for soft tissue, and the ABVS can obtain breast tissue structure information range far greater than the tissue structure information obtained by CT, and the ABVS can collect real-time tumor blood flow signal and blood vessel spectrum image information.

[0063] In the embodiments, as shown in Figures 1-2 The breast cancer HER2 gene mutation prediction method based on ABVS includes the following steps:

[0064] Step S1: acquiring an ABVS image;

[0065] Step S2: performing preliminary segmentation on the ABVS image to obtain a breast tissue part;

[0066] Step S3: based on the segmented breast tissue part, secondary segmentation is performed to obtain a lesion ROI region;

[0067] Step S4: according to the lesion ROI region, an imageomic feature is obtained, wherein the imageomic feature includes: first-order gray statistical feature, shape feature, texture feature, and feature based on transformation;

[0068] Step S5: the segmentation features generated by the segmentation process and the imageomic features are combined, and a deep feature classifier is used to predict the HER2 gene mutation state.

[0069] Each step will be described in detail below.

[0070] Step S1 further includes: pre-processing the obtained ABVS image, the pre-processing includes: isotropic interpolation of each layer of ABVS, adjusting image gain, and normalization processing.

[0071] Step S2 specifically includes:

[0072] Step S2.1: using a first segmentation model to perform preliminary segmentation on the ABVS image to obtain a preliminary segmentation region;

[0073] Step S2.2: performing morphological closing operation on the preliminary segmentation region to obtain a breast tissue part.

[0074] Specifically, in step S2, for breast cancer lesion segmentation, the breast tissue part is segmented first, and then the lesion ROI (Region Of Interest) region is segmented based on the breast tissue part.

[0075] In this embodiment, the first segmentation model (such as a traditional U-Net model) can be used to segment the breast tissue part. Due to the influence of the surrounding glandular tissue image, the segmented breast tissue part is partially missing, so morphological closing operation is further performed on the segmentation region, thereby a complete breast tissue region can be obtained, providing a more accurate search space for subsequent breast lesion tissue segmentation.

[0076] In step S2, the first segmentation model used is trained and evaluated on a public dataset containing 422 available ABVS images and segmentation results of patients. At the same time, the dataset is used as a test set for the above segmentation task and for the downstream classification task, which includes 211 available ABVS images.

[0077] After the same pre-processing, the voxel intensity is clipped to the range [-1200, 300], resampled to an anisotropic resolution of 1, 1, 1.5 mm3, and the breast is cut using U-Net (R231) pre-training weights in the two datasets.

[0078] Due to the surrounding tissue factors, there are small gaps in the segmentation results. To solve this problem, morphological closing operation is used to obtain the complete public, and the breast image is resized to 256x256x256. The b-spline and nearest interpolation of the image are used for tumor segmentation. The tumor segmentation can be used for 144 patients, of which 117 patients are known HER2 mutation status. The use of these data sets provides a stable basis for the experiment, ensuring the reliability and effectiveness of the experiment.

[0079] The results of ABVS preprocessing and breast tissue segmentation are shown in Figure 3 The figure shows the influence of surrounding tissue factors, the problem of missing breast tissue segmentation, and the process of correction by morphological closing operation.

[0080] As shown in Figure 4 S3 specifically comprises:

[0081] Step S3.1: generating a feature map according to the breast tissue part;

[0082] Step S3.2: extracting channel attention according to the feature map, and dividing the feature map into three parts of axial plane, coronal plane and sagittal plane according to the multi-view of the ABVS image, using learnable parameters to give corresponding weights, and then extracting spatial attention;

[0083] Step S3.4: fusing the channel attention, spatial attention and feature map according to the weight coefficient;

[0084] Step S3.5: based on the fused feature map, using a second segmentation model for secondary segmentation to obtain the lesion ROI region.

[0085] Due to the small volume of breast lesion ROI region, limited feature availability, unbalanced sample distribution and other factors, it is challenging. As a standard model for medical image segmentation, CNN has been widely invented, especially the U-Net network. The network consists of a symmetric encoder and decoder network with skip connections, and has become a common choice in the field of medical image analysis. In addition, some improved versions based on the U-Net network have appeared, such as Res-UNet and HDenseUNet. In addition, some inventions use AutoML to search for UNet architecture or a collection of 2D and 3D features, such as C2FNAS uses a two-stage NAS to search for a 3D architecture, and uses a meta-learner to learn a collection of 2D and 3D features.

[0086] Although these architectures have made significant progress in various two-dimensional and three-dimensional medical image segmentation tasks, they lack the ability to learn global context and long-term spatial dependencies, and their performance decreases on challenging small lesion segmentation tasks.

[0087] Therefore, in step S3 of the present application, an innovative three-dimensional architecture VA_UNet (Variational Attention Unet) is proposed for the breast lesion ROI region segmentation part. First, the three-dimensional attention of the medical image is extracted and projected onto the two-dimensional section, and is fused with the feature map to map and fuse the 2D and 3D attention in the encoder and decoder block to obtain a better context representation. In this way, the joint context information can effectively learn the global context and long-term spatial dependency, improve the performance of the small lesion segmentation task, and better obtain the lesion ROI region.

[0088] Table 1 shows the primary breast invasive ductal carcinoma lesion segmentation results, and it can be seen that after introducing the attention mechanism, the lesion segmentation results are better than the NNUNet.

[0089] Table 1 VA_UNet model and NNUNet model result comparison

[0090] Network MSD RECALL RRECION nnUNet 2D full 58.71.％ 69.57％ 62.22％ nnUNet 3D full 60.71％ 67.76％ 58.09％ VA_UNet 65.02％ 77.76％ 67.76％

[0091] The model of the present application is better than the NNUnet. Based on the model, it is found through comparison that segmenting the breast tissue can indeed effectively reduce the lesion search space and more accurately segment the breast lesion region. The ablation experiment results are shown in Table 2.

[0092] Table 2 Segmentation of breast tissue step ablation experiment

[0093] Network MSD RECALL RRECION TTTNet(no lung seg) 60.02％ 77.76％ 67.76％ TTTNet(lung seg) 65.02％ 77.76％ 67.76％

[0094] In recent years, more and more inventions have shown the feasibility of predicting the gene mutation state of breast cancer based on imaging signs, and have invented the correlation between tumor appearance and gene mutation. By combining gene invention with medical imaging, the imaging invention of breast cancer can be expanded to the molecular field, and the biological characteristics of breast cancer can be observed, which can help to guide the pathological classification, treatment and efficacy monitoring of tumors in clinic.

[0095] At present, the detection of HER2 gene mutation state is mainly through surgical or puncture methods to obtain tissue specimens, which is expensive and complex. If the gene mutation state can be judged through ABVS signs, not only can the burden of patients be reduced, but also more accurate guidance can be provided for the diagnosis, treatment and prognosis of breast cancer.

[0096] Step S4 and step S5 disclose a new breast cancer patient gene mutation state evaluation method, which uses the ABVS imaging performance of the breast cancer patient, extracts HER2 gene mutation related features such as texture, shape, intensity and the like through image recognition technology and image feature analysis algorithm. The method does not need to obtain breast cancer tissue specimens, is convenient to operate, has strong non-invasiveness, can more comprehensively and accurately evaluate the gene mutation state of the breast cancer patient.

[0097] The core steps of the method include: collecting a large number of ABVS images of breast cancer patients with known HER2 gene mutation states, establishing a prediction model and using machine learning or deep learning algorithm for prediction. The method has wide application prospects, provides important support for the treatment and efficacy monitoring of breast cancer, and has the following advantages: 1) convenience, high efficiency and high accuracy, which plays an important role in the clinical treatment of breast cancer; 2) without obtaining breast cancer tissue specimens, the operation is convenient, has strong non-invasiveness, can more comprehensively and accurately evaluate the gene mutation state of the breast cancer patient, and provides strong support for the treatment and efficacy monitoring of breast cancer.

[0098] Specifically, in step S4, imageomic features are obtained according to the lesion ROI region, wherein the imageomic features include: first-order gray statistical features, shape features, texture features and features based on transformation.

[0099] Step S4 uses imageomic features to analyze the tumor image, and the imageomic features include four parts, as shown in the following table. Figure 5 The imageomic features in this embodiment are manual features, but are not limited to manual features.

[0100] The first part is first-order statistical features, which include 18 features.

[0101] The second part is shape features based on three-dimensional reconstruction, which includes 14 features.

[0102] The third part is statistical texture features, which include 22 features calculated by a gray level co-occurrence matrix (GLCM), 16 features calculated by a gray level size zone matrix (GLSZM), 16 features calculated by a gray level run length matrix (GLRLM), 5 features calculated by a neighborhood gray tone difference matrix (NGTDM) and 14 features calculated by a gray level dependence matrix (GLDM).

[0103] The fourth part is feature data based on LoG and wavelet transformation, which includes 5 Gaussian filter images with different parameters (θ = 1, 2, 3, 4, 5) and 8 wavelet filter images with different directions.

[0104] Step S5 specifically includes:

[0105] Step S5.1: combine the segmentation features and the radiomics features generated by the segmentation process to form joint features;

[0106] Step S5.2: perform dimension reduction on the joint features using a dimension reduction algorithm;

[0107] Step S5.3: use a deep feature classifier to predict the HER2 gene mutation status.

[0108] The segmentation features are generated based on the intermediate output of the lesion ROI segmentation model in the previous step. By introducing an attention mechanism, the network pays more attention to the feature information of the lesion area, which is more accurate and has more potential applications than manually extracted features.

[0109] Specifically, in step S5, the deep segmentation features in the segmentation process are extracted and combined with the radiomics features of the lesion ROI region for dimension reduction, and then passed to the deep feature classifier to predict the epidermal growth factor receptor (HER2) mutation status. Finally, key visual cues and segmentation measurements are provided, and genomic information is predicted to help clinicians make reliable and non-invasive assessments and reduce the risk of biopsy operations.

[0110] The dimension reduction algorithm can use PCA, LDA, etc. The deep feature classifier can use random forest (RF), support vector machine classification (RFC), etc.

[0111] PCA (Principal Component Analysis) and LDA (Linear Discriminant Analysis) are classic dimension reduction techniques that can be used for feature extraction and data visualization of high-dimensional data.

[0112] The main purpose of PCA is to map the original data from high-dimensional space to low-dimensional space while preserving as much original data information as possible. In contrast, LDA aims to extract features related to class information through linear transformation to achieve better classification results. LDA can mine the differences between different classes in the original data space and construct a new low-dimensional feature vector to classify data in this space. LDA preserves the differences between data points while removing intra-class differences, making it more reflective of the essential features of the data and more conducive to subsequent classification or other tasks.

[0113] In summary, although PCA and LDA are both dimension reduction techniques, their application scenarios are slightly different: PCA is suitable for scenarios where the main feature components of the data set are preserved as much as possible and data visualization is improved, while LDA is more suitable for scenarios where information related to classification needs to be mined from the data set.

[0114] Random forest is an ensemble learning method that builds multiple decision trees for classification or regression. When building each tree, random forest randomly selects a subset of data samples and a number of features from the training set for tree construction, and selects the optimal split feature using indicators such as Gini index or entropy. The final prediction result is the average of the prediction results of all decision trees.

[0115] The random forest model can be expressed as:

[0116]

[0117] where T represents the number of decision trees, f i (x) represents the jth decision tree, and x represents the input feature vector.

[0118] In the random forest constructed based on the CART algorithm (classification and regression tree), each tree is defined as:

[0119]

[0120] where: is the split rule of the jth decision tree, R jm is the partition of the data space X by the mth node, I is an indicator function, and x R jm represents the position of the input sample x in the jth tree.

[0121] In the experiment, it is found that the LDA dimensionality reduction effect is better than that of PCA dimensionality reduction, and the classification result of the RF classifier is better than that of SVM. The comparison results are shown in the following table, and the commonly used binary classification index F1=2*Precision*Recall / (Precision+Recall) is used to improve the precision and recall rate as much as possible, and the difference between the two is as small as possible. The experimental results are shown in Table 3:

[0122] Table 3 F1 index for predicting HER2 mutation type

[0123] Classifier Dimensionality Reduction PCA LDA RF 52.62％ 62.68％ 72.76％ SVM 47.02％ 56.76％ 77.76％

[0124] In addition, the embodiment of the application also discloses a breast cancer HER2 gene mutation prediction device based on ABVS, comprising:

[0125] An acquisition module is configured to acquire an ABVS image.

[0126] A preliminary segmentation module is configured to preliminarily segment the ABVS image to obtain a breast tissue part.

[0127] A secondary segmentation module is configured to perform secondary segmentation based on the segmented breast tissue part to obtain a lesion ROI region;

[0128] An imageomics feature obtaining module is configured to obtain imageomics features based on the lesion ROI region, wherein the imageomics features include first-order gray scale statistical features, shape features, texture features, and transformation-based features;

[0129] A prediction module is configured to jointly use segmentation features generated in a segmentation process and the imageomics features to predict a HER2 gene mutation state by using a deep feature classifier.

[0130] Further, the obtaining module includes:

[0131] An obtaining unit is configured to obtain an ABVS image.

[0132] A preprocessing unit is configured to pre-process the obtained ABVS image, and the pre-processing includes isotropic interpolation, image gain adjustment, and normalization processing of each layer of the ABVS.

[0133] Further, the preliminary segmentation module specifically includes:

[0134] A preliminary segmentation unit is configured to perform preliminary segmentation on the ABVS image by using a first segmentation model to obtain a preliminary segmentation region.

[0135] A morphological closing operation unit is configured to perform morphological closing operation on the preliminary segmentation region to obtain a breast tissue part.

[0136] Further, the secondary segmentation module specifically includes:

[0137] A feature map generating unit is configured to generate a feature map based on the breast tissue part.

[0138] A multi-attention extracting unit is configured to extract channel attention based on the feature map, divide the feature map into three parts of an axial plane, a coronal plane, and a sagittal plane based on multiple viewing angles of the ABVS image, use learnable parameters to assign corresponding weights, and then extract spatial attention.

[0139] A fusion unit is configured to fuse the channel attention, the spatial attention, and the feature map according to a weight coefficient.

[0140] A secondary segmentation unit is configured to perform secondary segmentation on the fused feature map by using a second segmentation model to obtain a lesion ROI region.

[0141] Further, the prediction module specifically includes:

[0142] A joint feature forming unit is configured to jointly use segmentation features generated in a segmentation process and imageomics features to form joint features.

[0143] a dimension reduction unit configured to perform dimension reduction processing on the joint features by using a dimension reduction algorithm;

[0144] a prediction unit configured to predict the HER2 gene mutation state by using the deep feature classifier.

[0145] The breast cancer HER2 gene mutation prediction device disclosed in the present application is similar to the breast cancer HER2 gene mutation prediction method, and will not be described here.

[0146] The present application discloses a breast cancer HER2 gene mutation prediction method and device based on ABVS, which performs secondary segmentation based on the ABVS image, first segments the breast tissue part, and secondly segments the lesion ROI region.

[0147] The present application predicts the HER2 mutation state in breast cancer by using a non-invasive method, and provides accurate prediction and help for clinical treatment of breast cancer.

[0148] The above examples are only for illustrating the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and implement it, and cannot limit the protection scope of the present application, and any equivalent changes or modifications made according to the spirit and essence of the present application should be covered within the protection scope of the present application.

Claims

1. A breast cancer HER2 gene mutation prediction method based on ABVS, characterized in that, The method comprises the following steps: Step S1: acquiring an ABVS image; Step S2: performing preliminary segmentation on the ABVS image to obtain a breast tissue part; Step S3: performing secondary segmentation on the basis of the segmented breast tissue part to obtain a lesion ROI region; Step S4: obtaining an image-based feature according to the lesion ROI region, wherein the image-based feature comprises a first-order gray statistical feature, a shape feature, a texture feature, and a transformation-based feature; Step S5: combining the segmentation features generated in the segmentation process and the image-based features, and using a deep feature classifier to predict the HER2 gene mutation state. The step S3 specifically comprises: Step S3.1: generating a feature map according to the breast tissue part; Step S3.2: extracting channel attention according to the feature map, and dividing the feature map into three parts of an axial plane, a coronal plane, and a sagittal plane according to the multi-view of the ABVS image, using a learnable parameter to assign corresponding weights, and then extracting spatial attention; Step S3.4: fusing the channel attention, the spatial attention, and the feature map according to the weight coefficient; Step S3.5: performing secondary segmentation on the basis of the fused feature map by using a second segmentation model to obtain the lesion ROI region. 2.The breast cancer HER2 gene mutation prediction method according to claim 1, characterized in that, The step S1 further comprises: pre-processing the acquired ABVS image, and the pre-processing comprises one or more of the following operations: isotropic interpolation of each layer of the ABVS, adjusting the image gain, and normalization processing. 3.The breast cancer HER2 gene mutation prediction method of claim 1, characterized by, The step S2 specifically comprises: Step S2.1: using a first segmentation model to perform preliminary segmentation on the ABVS image to obtain a preliminary segmentation region; Step S2.2: performing a morphological closing operation on the preliminary segmentation region to obtain the breast tissue part. 4.The breast cancer HER2 gene mutation prediction method of claim 1, characterized by, The step S5 specifically comprises: Step S5.1: combining the segmentation features generated in the segmentation process and the image-based features to form joint features; Step S5.2: performing dimension reduction processing on the joint features by using a dimension reduction algorithm; Step S5.3: using a deep feature classifier to predict the HER2 gene mutation state.

5. The breast cancer HER2 gene mutation prediction device based on ABVS, characterized in that, The method comprises the following steps: An acquisition module is configured to acquire an ABVS image; A preliminary segmentation module is configured to perform preliminary segmentation on the ABVS image to obtain a breast tissue part; A secondary segmentation module is configured to perform secondary segmentation on the basis of the segmented breast tissue part to obtain a lesion ROI region; An image-based feature obtaining module is configured to obtain an image-based feature according to the lesion ROI region, wherein the image-based feature comprises a first-order gray statistical feature, a shape feature, a texture feature, and a transformation-based feature; A prediction module is configured to combine the segmentation features generated in the segmentation process and the image-based features, and use a deep feature classifier to predict the HER2 gene mutation state. The secondary segmentation module specifically comprises: A feature map generating unit is configured to generate a feature map according to the breast tissue part; A multi-attention extracting unit is configured to extract channel attention according to the feature map, and divide the feature map into three parts of an axial plane, a coronal plane, and a sagittal plane according to the multi-view of the ABVS image, use a learnable parameter to assign corresponding weights, and then extract spatial attention; A fusion unit is configured to fuse the channel attention, the spatial attention and the feature map according to a weight coefficient; A secondary segmentation unit is configured to perform secondary segmentation on the fused feature map by using a second segmentation model to obtain a lesion ROI region. 6.The breast cancer HER2 gene mutation prediction device according to claim 5, characterized by, The acquisition module comprises: An acquisition unit is configured to acquire an ABVS image. A preprocessing unit is configured to pre-process the acquired ABVS image, and the pre-processing comprises: isotropic interpolation, image gain adjustment and normalization processing of each layer of the ABVS.

7. The breast cancer HER2 gene mutation prediction device according to claim 5, characterized by, The preliminary segmentation module specifically comprises: A preliminary segmentation unit is configured to perform preliminary segmentation on the ABVS image by using a first segmentation model to obtain a preliminary segmentation region. A morphological closing operation unit is configured to perform morphological closing operation on the preliminary segmentation region to obtain a breast tissue part.

8. The breast cancer HER2 gene mutation prediction device according to claim 5, characterized by, The prediction module specifically comprises: A joint feature forming unit is configured to combine segmentation features generated in a segmentation process and radiomics features to form joint features. A dimension reduction unit is configured to perform dimension reduction processing on the joint features by using a dimension reduction algorithm. A prediction unit is configured to predict a HER2 gene mutation state by using a deep feature classifier.

Images