A method for predicting the metastasis status of axillary lymph nodes in breast cancer

Abstract

The present invention provides a method for predicting the axillary lymph node metastasis status of breast cancer, comprising the following steps: S1: obtaining a B-mode ultrasound image and a shear wave elasticity image of the lymph node, and intercepting regions of interest from each of the B-mode ultrasound image and the shear wave elasticity image; S2: processing the two intercepted regions of interest to generate a square B-mode image and a square SWE image; S3: extracting features from the square B-mode image and the square SWE image using a multimodal adaptive fusion network and performing adaptive feature fusion to obtain an adaptive fusion result; S4: preliminarily predicting the class probability of the adaptive fusion result, comparing the class probability with a pre-acquired optimal cutoff value, and obtaining a prediction result. The present invention provides a method for predicting the axillary lymph node metastasis status of breast cancer, addressing the current difficulty in achieving high accuracy in ultrasonic lymph node assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to a method for predicting the status of axillary lymph node metastasis of breast cancer. Background Art

[0002] Breast cancer is the most common cancer among women worldwide. Its incidence ranks first among all newly diagnosed female cancers and it is the leading cause of cancer-related death in women, posing a huge challenge to public health worldwide.

[0003] The sentinel lymph node is the first lymph node to receive drainage from the tumor area. In breast cancer diagnosis, the axillary lymph node (ALN) is the most common site of breast cancer metastasis, and metastasis is the leading cause of cancer-related death in breast cancer patients. Axillary lymph node metastasis (ALNM) is one of the most important prognostic factors for breast cancer. Identifying lymph node metastasis is crucial for pathological staging, prognosis, and treatment guidance for breast cancer patients.

[0004] Invasive surgical procedures for assessing lymph node status include sentinel lymph node biopsy (SLNB) and axillary lymph node dissection (ALND). ALND removes more lymph nodes than SLNB but is associated with more lymphatic and neurologic side effects, such as lymphedema and numbness. Therefore, using SLNB to detect metastatic cancer is generally highly effective and less invasive than the older method of ALND. SLNB is associated with fewer complications compared to ALND, but it is not a risk-free procedure and still has some significant limitations, including considerable additional anesthesia time and expense, and complications such as arm numbness or upper limb edema in 3.5–10.9% of patients. Therefore, both ALND and SLNB carry the risk of overtreatment, and these invasive surgical procedures are considered controversial due to their invasive nature, potential complications, and potential for overtreatment.

[0005] Noninvasive methods for detecting axillary lymph node metastasis involve assessing lymph node status through ultrasound. There are generally two methods: B-mode ultrasound, which primarily detects malignant lymph nodes based on morphological characteristics; and shear wave elastography (SWE), which uses SWE to measure the stiffness of lymph node tissue, based on the significant difference in stiffness between healthy and metastatic lymph nodes, to quantitatively assess lymph node metastasis. However, the interpretation of ultrasound images relies on the subjective assessment of radiologists, which leads to interobserver variability and makes it difficult to achieve a high degree of assessment accuracy. Summary of the Invention

[0006] The present invention aims to overcome the technical defect that it is difficult to achieve a high assessment accuracy rate by ultrasonic evaluation of lymph node status, and provides a method for predicting the axillary lymph node metastasis status of breast cancer.

[0007] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0008] A method for predicting the status of axillary lymph node metastasis in breast cancer comprises the following steps:

[0009] S1: Obtain B-mode ultrasound (B-mode) images and shear wave elasticity (SWE) images of the lymph node, and intercept the regions of interest of the B-mode ultrasound image and SWE image respectively;

[0010] S2: Process the two intercepted regions of interest separately to generate a square B-mode image and a square SWE image;

[0011] S3: The features of the square B-mode image and the square SWE image are extracted and adaptively fused through the multimodal adaptive fusion network to obtain the adaptive fusion result;

[0012] The multimodal adaptive fusion network consists of 2L self-attention encoders, of which 2N-2 self-attention encoders are used for feature extraction and 2L-2N+2 self-attention encoders are used for feature adaptive fusion;

[0013] S4: Preliminary prediction of the class probability of the adaptive fusion result and comparison of the class probability with the pre-acquired optimal cutoff value;

[0014] If the class probability is greater than the optimal cutoff value, the prediction result is transfer;

[0015] If the class probability is not greater than the optimal cutoff value, the prediction result is non-transition.

[0016] In the above scheme, by processing the B-mode and SWE dual-modal images of the lymph nodes into square B-mode images and square SWE images respectively, the original tissue morphology information of the lymph nodes is prevented from being changed during image processing. At the same time, redundant tissue information in the B-ultrasound image can be removed, so that the multimodal adaptive fusion network can better focus on the analysis of lymph node tissue; the global feature information of the image is extracted and adaptively fused through the multimodal adaptive fusion network with adjustable parameters L and N, which greatly improves the prediction accuracy of the axillary lymph node status of breast cancer; and the prediction accuracy is further improved by introducing the optimal truncation value as the soft threshold.

[0017] Preferably, in step S1 , the region of interest of the shear wave elasticity image is first intercepted, and then the region of interest of the B-mode ultrasound image is intercepted using the region of interest of the shear wave elasticity image as a reference.

[0018] Preferably, the region of interest includes an entire lymph node region.

[0019] Preferably, step S2 specifically includes: placing the two intercepted regions of interest in different square black frames, respectively, to obtain a square SWE image and a square B-mode image accordingly.

[0020] Preferably, the side length of the square black frame is equal to the longest side of the region of interest.

[0021] In the above scheme, the square SWE images and square B-mode images obtained through processing are one-to-one corresponding, without changing the original tissue morphology information of the lymph node and removing redundant tissue information from the B-ultrasound image; thus, in deep learning tasks, the multimodal adaptive fusion network can better focus on the analysis of lymph node tissue.

[0022] Preferably, the output Attention_out of any self-attention encoder is expressed as follows:

[0023] Attention_out=MLP(MutiHead(pathces))

[0024] Among them, MLP represents the multi-layer perceptron in the self-attention encoder; pathces represents multiple slices of the image after being equally divided by the Embeded Patches layer in the self-attention encoder;

[0025] MutiHead represents the multi-head attention module in the self-attention encoder, which is obtained by splicing multiple single-head attention modules, as shown below:

[0026] MutiHead=Concat(head1,...,head i ,...,head h )W O

[0027] The i-th single-head attention module head i The definition is as follows:

[0028]

[0029]

[0030] Q represents the query matrix; K represents the content to be focused on; V represents the actual feature information; head1 represents the first single-head attention module; headh represents the hth single-head attention module; Concat represents the concatenation operation; W O represents the parameter matrix; represents the parameter matrix of Q; represents the parameter matrix of K; represents the parameter matrix of V; Attention represents the single-head attention module; softmax represents the normalization function in the single-head attention module; d k Represents the scaling factor.

[0031] In the above scheme, the image features are extracted through the self-attention encoder, which increases the representation ability of unimodal image features and lays the foundation for the adaptive fusion of attention features.

[0032] Preferably, feature adaptive fusion is expressed as follows:

[0033]

[0034]

[0035] Where, j∈(N+1,…,L-1,L); Represents the self-attention encoder A j The characteristic information of the obtained square B-mode image; Represents the self-attention encoder A j-1 The characteristic information of the obtained square B-mode image; Represents the self-attention encoder B j Obtain feature information of the obtained square SWE image; Represents the self-attention encoder B j-1 Get the feature information of the square SWE image; W j Represents the hyperparameter of the multimodal adaptive fusion network, and its value is between 0 and 1.

[0036] Preferably, the adaptive fusion result C is:

[0037]

[0038] in, Represents the self-attention encoder A L The characteristic information of the obtained square B-mode image; Represents the self-attention encoder B L The feature information of the obtained square B-mode image.

[0039] Preferably, the optimal cutoff value is obtained by the following steps:

[0040] S4.1: Obtain several groups of sample images with transfer state labels and input them into a multimodal adaptive fusion network, thereby obtaining several corresponding sample adaptive fusion results; each group of sample images includes a B-mode ultrasound sample image and a shear wave elasticity sample image;

[0041] S4.2: Obtain the category probability of each sample adaptive fusion result through classification;

[0042] S4.3: Use the roc_curve function to obtain several sets of true positive rates, false positive rates, and cutoff values based on the class probability of each sample's adaptive fusion result and the corresponding transition state label;

[0043] S4.4: Calculate the Youden index for each group based on the true positive rate and false positive rate;

[0044] S4.5: Select the cutoff value in the group with the largest Youden index as the optimal cutoff value.

[0045] Preferably, the formula for calculating the Youdens' Index is as follows:

[0046] Youdens′Index=tpr+fpr-1

[0047]

[0048]

[0049] Among them, tpr represents the true positive rate; fpr represents the false positive rate; TP represents the number of correctly classified positive samples; FN represents the number of incorrectly classified positive samples; FP represents the number of incorrectly classified negative samples; TN represents the number of correctly classified negative samples.

[0050] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0051] The present invention provides a method for predicting the axillary lymph node metastasis status of breast cancer. By processing the B-mode and SWE dual-modality images of the lymph nodes into square B-mode images and square SWE images, respectively, the original tissue morphology information of the lymph nodes is prevented from being changed during image processing. At the same time, redundant tissue information of the B-ultrasound image can be removed, so that the multimodal adaptive fusion network can better focus on the analysis of lymph node tissue. The multimodal adaptive fusion network with adjustable parameters L and N extracts global feature information of the image and performs adaptive fusion, which greatly improves the prediction accuracy of the axillary lymph node status of breast cancer. The optimal cutoff value is introduced as the soft threshold to further improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of the steps for implementing the technical solution of the present invention;

[0053] Figure 2 It is a schematic diagram of the overall framework of the present invention;

[0054] Figure 3 Schematic diagram of the process of generating a square B-mode image and a square SWE image in the present invention;

[0055] Figure 4 Schematic diagram of the structure of the self-attention encoder in the present invention;

[0056] Figure 5 Schematic diagram of the prediction results of using B-mode images to predict the axillary lymph node metastasis status of breast cancer in the present invention;

[0057] Figure 6 Schematic diagram of the prediction results of using SWE images to predict the axillary lymph node metastasis status of breast cancer in the present invention;

[0058] Figure 7 Schematic diagram of the prediction results of axillary lymph node metastasis status using a dual-modal convolutional neural network in the present invention;

[0059] Figure 8 Schematic diagram of the prediction results of the method for predicting axillary lymph node metastasis status of breast cancer used in the present invention. DETAILED DESCRIPTION

[0060] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0061] In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size;

[0062] It is understandable to those skilled in the art that some well-known structures and descriptions thereof may be omitted in the drawings.

[0063] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0064] Example 1

[0065] like Figure 1-2 As shown, a method for predicting the status of axillary lymph node metastasis of breast cancer includes the following steps:

[0066] S1: Obtain B-mode ultrasound (B-mode) images and shear wave elasticity (SWE) images of the lymph node, and intercept the regions of interest of the B-mode ultrasound image and SWE image respectively;

[0067] S2: Process the two intercepted regions of interest separately to generate a square B-mode image and a square SWE image;

[0068] S3: The features of the square B-mode image and the square SWE image are extracted and adaptively fused through the multimodal adaptive fusion network to obtain the adaptive fusion result;

[0069] The multimodal adaptive fusion network consists of 2L self-attention encoders, numbered A1, A2, ..., A j 、... L-1 、A L and B1, B2, ..., B j 、... L-1 、B L , where 2N-2 self-attention encoders are used for feature extraction, and 2L-2N+2 self-attention encoders are used for feature adaptive fusion;

[0070] In actual implementation, L>N. By training a multimodal adaptive fusion network to adjust the parameters L and N, the prediction accuracy of axillary lymph node status in breast cancer and the generalization ability of the network were greatly improved. In deep learning, the network training method is existing technology and will not be repeated here.

[0071] S4: Preliminary prediction of the class probability of the adaptive fusion result and comparison of the class probability with the pre-acquired optimal cutoff value;

[0072] If the class probability is greater than the optimal cutoff value, the prediction result is transfer;

[0073] If the class probability is not greater than the optimal cutoff value, the prediction result is non-transition.

[0074] During the specific implementation process, the B-mode and SWE dual-modal images of the lymph nodes are processed into square B-mode images and square SWE images respectively, which prevents the original tissue morphology information of the lymph nodes from being changed during image processing. At the same time, the redundant tissue information of the B-ultrasound image can be removed, so that the multimodal adaptive fusion network can better focus on the analysis of lymph node tissue; the global feature information of the image is extracted and adaptively fused through the multimodal adaptive fusion network with adjustable parameters L and N, which greatly improves the prediction accuracy of the axillary lymph node status of breast cancer; and the prediction accuracy is further improved by introducing the optimal truncation value as the soft threshold.

[0075] Example 2

[0076] A method for predicting the status of axillary lymph node metastasis in breast cancer comprises the following steps:

[0077] S1: Obtain B-mode ultrasound (B-mode) images and shear wave elasticity (SWE) images of the lymph node, and intercept the regions of interest of the B-mode ultrasound image and SWE image respectively;

[0078] More specifically, Figure 3 As shown, in step S1, the region of interest (ROI) of the shear wave elasticity image is first intercepted, and then the region of interest (ROI) of the B-mode ultrasound image is intercepted using the region of interest of the shear wave elasticity image as a reference.

[0079] More specifically, the region of interest includes an entire lymph node region.

[0080] S2: Process the two intercepted regions of interest separately to generate a square B-mode image and a square SWE image;

[0081] More specifically, step S2 specifically includes: placing the two intercepted regions of interest in different square black frames, respectively, to obtain a square SWE image and a square B-mode image accordingly.

[0082] More specifically, the side length of the square black frame is equal to the longest side of the region of interest.

[0083] In this embodiment, the longest side of the region of interest is scaled to 224 pixels suitable for network input, and the short side is scaled at the same ratio. The size of the square black frame is set to 224*224, and the scaled region of interest is centered in the square black frame to obtain a corresponding square image with a size of 224*224.

[0084] During the specific implementation process, the square SWE images and square B-mode images obtained through processing are one-to-one corresponding, without changing the original tissue morphology information of the lymph nodes and removing redundant tissue information from the B-ultrasound image; thus, in deep learning tasks, the multimodal adaptive fusion network can better focus on the analysis of lymph node tissue.

[0085] S3: The features of the square B-mode image and the square SWE image are extracted and adaptively fused through the multimodal adaptive fusion network to obtain the adaptive fusion result;

[0086] The multimodal adaptive fusion network consists of 2L self-attention encoders, numbered A1, A2, ..., A j ,…A L-1 、A L and B1, B2, ..., B j ,…B L-1 、B L , where 2N-2 self-attention encoders are used for feature extraction, and 2L-2N+2 self-attention encoders are used for feature adaptive fusion;

[0087] In actual implementation, L>N, and the parameters L, N and hyperparameter W are adjusted by training the multimodal adaptive fusion network. j, enabling the multimodal adaptive fusion network to automatically learn features that contribute greatly to network classification prediction, suppress features with less contribution, and retain the essence while discarding the dross, greatly improving the prediction accuracy of axillary lymph node status in breast cancer and the generalization ability of the network. In deep learning, the network training method is an existing technology and will not be elaborated here.

[0088] like Figure 4 As shown in Figure 1, a single self-attention encoder consists of an Embeded Patches layer, two layer normalization modules, a multi-head attention module, and a multi-layer perceptron (MLP).

[0089] More specifically, the output Attention_out of any self-attention encoder is expressed as follows:

[0090] Attention_out=MLP(MutiHead(pathces))

[0091] Among them, MLP represents the multi-layer perceptron in the self-attention encoder, which includes two linear layers, two dropout layers, and one activation layer;

[0092] Pathces represent multiple slices of an image after being equally divided by the Embeded Patches layer in the self-attention encoder. That is, the Embeded Patches layer divides the square SWE image and the square B-mode image into multiple slices (patches). Then, the linear mapping layer maps each patch to a one-dimensional space, and finally obtains several token vectors, and adds position information and category information to each token vector.

[0093] MutiHead represents the multi-head attention module in the self-attention encoder, which is obtained by splicing multiple single-head attention modules, as shown below:

[0094] MutiHead=Concat(head1,...,head i ,...,head h )W O

[0095] The i-th single-head attention module head i The definition is as follows:

[0096]

[0097]

[0098] Q represents the query matrix; K represents the content to be focused on; V represents the actual feature information; head1 represents the first single-head attention module; head h represents the hth single-head attention module; Concat represents the concatenation operation; W O represents the parameter matrix; represents the parameter matrix of Q; represents the parameter matrix of K; represents the parameter matrix of V; Attention represents the single-head attention module; softmax represents the normalization function in the single-head attention module; d k Represents the scaling factor.

[0099] In the specific implementation process, the image features extracted through the self-attention encoder increase the representation ability of unimodal image features, laying the foundation for the adaptive fusion of attention features.

[0100] More specifically, feature adaptive fusion is expressed as follows:

[0101]

[0102]

[0103] Where, j∈(N+1,…,L-1,L); Represents the self-attention encoder A j The characteristic information of the obtained square B-mode image; Represents the self-attention encoder A j-1 The characteristic information of the obtained square B-mode image; Represents the self-attention encoder B j Obtain feature information of the obtained square SWE image; Represents the self-attention encoder B j-1 Get the feature information of the square SWE image; W j Represents the hyperparameter of the multimodal adaptive fusion network, and its value is between 0 and 1.

[0104] More specifically, the adaptive fusion result C is:

[0105]

[0106] in, Represents the self-attention encoder A L The characteristic information of the obtained square B-mode image; Represents the self-attention encoder B L The feature information of the obtained square B-mode image.

[0107] In the specific implementation process, firstly, the self-attention encoders A1, A2, ...A N-1 Feature extraction is performed on the square B-mode image through the self-attention encoder B1, B2, ...B N-1 Feature extraction is performed on the square SWE image to obtain a certain amount of information from the two modal images, that is, to obtain a general understanding of the lymph node tissue of a certain modality. N ,…,A L and B N ,…,B L The method of adaptive fusion of dual-modal images is adopted, that is, the characteristic information of tissue morphology and tissue hardness of the same area images of the two modalities is combined, and the characteristic information of tissue morphology and tissue hardness is obtained according to the set hyperparameters W. j The network automatically adjusts the contribution weights of different modal images to the classification prediction results to achieve the effect of adaptive adjustment. j The size of is limited to between 0 and 1, and it will be automatically optimized during the network training process. The reason why the adaptive fusion of bimodal images works is that the network can automatically adjust the importance of a certain modality information before making a decision, suppressing features that are not important for the classification result.

[0108] S4: Preliminary prediction of the class probability of the adaptive fusion result and comparison of the class probability with the pre-acquired optimal cutoff value;

[0109] If the class probability is greater than the optimal cutoff value, the prediction result is transfer;

[0110] If the class probability is not greater than the optimal cutoff value, the prediction result is non-transition.

[0111] In actual implementation, the adaptive fusion results output by the network are passed through a normalization layer to obtain the preliminary predicted category probability.

[0112] Example 3

[0113] This embodiment provides a method for predicting axillary lymph node metastasis status of breast cancer. The difference from Example 2 is that, based on Example 2, this embodiment specifically describes the process of obtaining the optimal cutoff value in step S4.

[0114] More specifically, the optimal cutoff value is obtained by the following steps:

[0115] S4.1: Obtain several groups of sample images with transfer state labels and input them into a multimodal adaptive fusion network, thereby obtaining several corresponding sample adaptive fusion results; each group of sample images includes a B-mode ultrasound sample image and a shear wave elasticity sample image;

[0116] S4.2: Obtain the category probability of each sample adaptive fusion result through classification;

[0117] S4.3: Use the roc_curve function to obtain several sets of true positive rates, false positive rates, and cutoff values based on the class probability of each sample's adaptive fusion result and the corresponding transition state label;

[0118] S4.4: Calculate the Youden index for each group based on the true positive rate and false positive rate;

[0119] More specifically, the formula for calculating the Youdens′Index is as follows:

[0120] Youdens′Index=tpr+fpr-1

[0121]

[0122]

[0123] Among them, tpr represents the true positive rate; fpr represents the false positive rate; RP represents the number of correctly classified positive samples; FN represents the number of incorrectly classified positive samples; FP represents the number of incorrectly classified negative samples; TN represents the number of correctly classified negative samples;

[0124] S4.5: Select the cutoff value in the group with the largest Youden index as the optimal cutoff value.

[0125] Example 4

[0126] In this example, the performance of the method for predicting the axillary lymph node metastasis status of breast cancer is compared with that of a method for predicting the axillary lymph node metastasis status of breast cancer using a single-modality image (either B-mode or SWE) and a method for predicting the axillary lymph node metastasis status using a bimodal convolutional neural network, based on the AUC in the ROC curve.

[0127] The prediction results of using single-modality images (B-mode or SWE) to predict the status of axillary lymph node metastasis in breast cancer are as follows: Figure 5-6 As shown, the highest AUC in the independent test set is 0.863;

[0128] The prediction results of axillary lymph node metastasis status using a bimodal convolutional neural network are as follows: Figure 7 As shown, the highest AUC in the independent test set is 0.872;

[0129] The prediction results of the method for predicting the status of axillary lymph node metastasis of breast cancer are as follows: Figure 8 As shown in the figure, the AUC in the independent test set is 0.938. Therefore, the prediction performance of the method for predicting the axillary lymph node metastasis status of breast cancer is better, far higher than the other methods mentioned above.

[0130] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting axillary lymph node metastasis status of breast cancer, characterized in that: The following steps are involved: S1: Obtain a B-mode ultrasound image and a shear wave elasticity image of the lymph node, and intercept the region of interest of the B-mode ultrasound image and the shear wave elasticity image respectively; S2: Process the two intercepted regions of interest separately to generate a square B-mode image and a square SWE image; S3: The features of the square B-mode image and the square SWE image are extracted and adaptively fused through the multimodal adaptive fusion network to obtain the adaptive fusion result; The multimodal adaptive fusion network consists of 2L self-attention encoders, of which 2N-2 self-attention encoders are used for feature extraction and 2L-2N+2 self-attention encoders are used for feature adaptive fusion; Feature adaptive fusion is expressed as follows: where j∈(N+1,...,L-1,L); Represents the self-attention encoder A j The characteristic information of the obtained square B-mode image; Represents the self-attention encoder A j-1 The characteristic information of the obtained square B-mode image; Represents the self-attention encoder B j Obtain feature information of the obtained square SWE image; Represents the self-attention encoder B j-1 Get the feature information of the square SWE image; W j Represents the hyperparameter of the multimodal adaptive fusion network, and its value is between 0 and 1; The adaptive fusion result C is: in, Represents the self-attention encoder A L The characteristic information of the obtained square B-mode image; Represents the self-attention encoder B L The characteristic information of the obtained square B-mode image; S4: Preliminary prediction of the class probability of the adaptive fusion result and comparison of the class probability with the pre-acquired optimal cutoff value; If the class probability is greater than the optimal cutoff value, the prediction result is transfer; If the class probability is not greater than the optimal cutoff value, the prediction result is non-transition.

2. The method for predicting axillary lymph node metastasis of breast cancer according to claim 1, wherein: In step S1 , the region of interest of the shear wave elasticity image is firstly captured, and then the region of interest of the B-mode ultrasound image is captured using the region of interest of the shear wave elasticity image as a reference.

3. A method for predicting axillary lymph node metastasis status of breast cancer according to claim 1 or 2, characterized in that: The region of interest includes the entire lymph node region.

4. The method for predicting axillary lymph node metastasis of breast cancer according to claim 1, wherein: Step S2 specifically includes: placing the two intercepted regions of interest in different square black frames, respectively, to obtain a square SWE image and a square B-mode image.

5. The method for predicting axillary lymph node metastasis of breast cancer according to claim 4, wherein: The side length of the square black frame is equal to the longest side of the region of interest.

6. The method for predicting axillary lymph node metastasis status of breast cancer according to claim 1, wherein: The output Attention_out of any self-attention encoder is expressed as follows: Attention_out=MLP(MutiHead(pathces)) Among them, MLP represents the multi-layer perceptron in the self-attention encoder; pathces represents multiple slices of the image after being equally divided by the Embeded Patches layer in the self-attention encoder; MutiHead represents the multi-head attention module in the self-attention encoder, which is obtained by splicing multiple single-head attention modules, as shown below: MutiHead=Concat(head1,...,head i ,...,head h )W O The i-th single-head attention module head i The definition is as follows: Q represents the query matrix; K represents the content to be focused on; V represents the actual feature information; head1 represents the first single-head attention module; head h represents the hth single-head attention module; Concat represents the concatenation operation; W O represents the parameter matrix; Wi i Q represents the parameter matrix of Q; represents the parameter matrix of K; represents the parameter matrix of V; Attention represents the single-head attention module; softmax represents the normalization function in the single-head attention module; d k Represents the scaling factor.

7. The method for predicting axillary lymph node metastasis of breast cancer according to claim 1, wherein: The optimal cutoff value is obtained by the following steps: S4.1: Obtain several groups of sample images with transfer state labels and input them into a multimodal adaptive fusion network, thereby obtaining several corresponding sample adaptive fusion results; each group of sample images includes a B-mode ultrasound sample image and a shear wave elasticity sample image; S4.2: Obtain the category probability of each sample adaptive fusion result through classification; S4.3: Use the roc_curve function to obtain several sets of true positive rates, false positive rates, and cutoff values based on the class probability of each sample's adaptive fusion result and the corresponding transition state label; S4.4: Calculate the Youden index for each group based on the true positive rate and false positive rate; S4.5: Select the cutoff value in the group with the largest Youden index as the optimal cutoff value.

8. The method for predicting axillary lymph node metastasis of breast cancer according to claim 7, wherein: The formula for calculating Youdens'Index is as follows: Youdens′Index=tpr+fpr-1 Among them, tpr represents the true positive rate; fpr represents the false positive rate; TP represents the number of correctly classified positive samples; FN represents the number of incorrectly classified positive samples; FP represents the number of incorrectly classified negative samples; TN represents the number of correctly classified negative samples.

Images