Breast segmentation model training method and device for breast ultrasound image lesion area identification

Abstract

The invention discloses a mammary gland segmentation model training method and device for mammary gland ultrasound image focus area identification. The mammary gland segmentation model training method for mammary gland ultrasonic image lesion area identification comprises the steps of obtaining a mammary gland patient ultrasonic image data set, wherein the mammary gland patient ultrasonic image data set comprises at least one mammary gland patient ultrasonic image; a breast segmentation model is obtained, the breast segmentation model comprises an MGSA module, and the MGSA module is used for obtaining a guide feature map with a weight according to the original image; and training the mammary gland segmentation model through the mammary gland patient ultrasonic image data set to obtain a trained mammary gland segmentation model. According to the method, more reliable support can be provided for auxiliary diagnosis of the breast ultrasonic image. For the problems of fuzzy lesion segmentation boundary, insufficient context information capture and the like of the breast ultrasonic image, higher segmentation precision and stronger robustness can be ensured.

Description

Technical Field

[0001] The present application relates to the technical field of image recognition, and particularly to a method for training a breast segmentation model for identifying lesion regions in breast ultrasound images and a device for training a breast segmentation model for identifying lesion regions in breast ultrasound images. Background Art

[0002] Due to its high incidence and mortality rates, breast cancer has become a major threat in the field of women's health globally. Despite the continuous progress of modern medical diagnosis and treatment technologies, the early diagnosis and treatment of breast cancer still face huge challenges. Breast ultrasound imaging, as a non-invasive, convenient, and economical diagnostic tool, has become an important means for breast cancer screening and early diagnosis. With the development of medical technology, especially the continuous innovation of medical imaging technology, the quality of breast ultrasound imaging has been significantly improved and widely applied in clinical practice. However, the low contrast, noise interference, and blurred boundaries between lesions and the background in breast ultrasound images pose significant challenges to automated image segmentation. These challenges limit the performance of computer-aided diagnosis systems (CAD), thereby affecting the accuracy and efficiency of tumor detection.

[0003] The tumor segmentation of traditional breast ultrasound images relies on doctors to manually annotate the lesion regions. Although this method is accurate, it not only requires doctors to have rich experience and professional knowledge but also has extremely low efficiency when dealing with a large number of images, unable to meet the growing clinical needs. To improve the diagnostic efficiency and reduce the workload of doctors, automated tumor segmentation technology has become a research hotspot. In recent years, the rapid development of deep learning technology, especially the successful application of convolutional neural networks (CNN) in medical image analysis, has provided new solutions for the automatic segmentation of breast ultrasound images.

[0004] Therefore, there is a need for a technical solution to solve or at least alleviate the above deficiencies of the existing technology. Summary of the Invention

[0005] The object of the present invention is to provide a method for training a breast segmentation model for identifying lesion regions in breast ultrasound images to at least solve one of the above technical problems.

[0006] The present invention provides the following solutions:

[0007] According to one aspect of the present invention, there is provided a method for training a breast segmentation model for identifying lesion regions in breast ultrasound images, the method for training a breast segmentation model for identifying lesion regions in breast ultrasound images comprising:

[0008] Obtain a breast patient ultrasound image dataset, the breast patient ultrasound image dataset including at least one breast patient ultrasound image;

[0009] Obtain a breast segmentation model, where the breast segmentation model includes an MGSA module, and the MGSA module is used to obtain a guided feature map with weights based on the original image;

[0010] Train the breast segmentation model through the breast patient ultrasound image dataset to obtain a trained breast segmentation model.

[0011] Optionally, the breast segmentation model further includes a channel shuffle attention module, and the channel shuffle attention module is used to perform dimension permutation, capture channel interdependencies, perform channel shuffle operations, and enhance spatial attention on the guided feature map with weights to obtain a final output feature map.

[0012] Optionally, the breast segmentation model further includes a DySample upsampling module, and the DySample upsampling module is used for pixel-level spatial rearrangement of the final output feature map to determine the positions of upsampling points using dynamically generated offsets, which reduces the neglect of positional relationships and the disordered shifting of offsets.

[0013] Optionally, the MGSA module includes:

[0014] A feature weight conversion module, which is used to perform Otsu binarization on the original image, segment the foreground and background, and splice the edge feature map obtained by using the Scharr operator and the reverse feature map obtained by reverse attention with the decoded feature map output by the decoder. Next, the spliced feature map is converted into feature weights through Sigmoid;

[0015] A product module, which is used to multiply the encoded feature map obtained from the encoded features by the feature weights to obtain a guided feature map with weights.

[0016] Optionally, the feature weight conversion module obtains feature weights through the following formula:

[0017] Guide 1i = σ(Scharr(Pred i ))

[0018] Guide 2i = σ(Otsu(O ri ))

[0019] Guide 3i = 1 - σ(Pred i );

[0020] Where Guide 1i is the edge-guided feature weight, Guide 2iis the background-guided feature weight, Guide 3i is the reverse attention-guided feature weight, Pred i is the decoded prediction feature map, O ri is the original image.

[0021] Optionally, the product module obtains the guided feature map with weights through the following formula:

[0022] Guide i = Cat(Guide 1i ×F ei )(Guide 2i ×F ei )(Guide 3i ×F ei ); where,

[0023] where, Guide i is the guided feature map with weights, Guide 1i is the edge-guided feature weight, Guide 2i is the background-guided feature weight, Guide 3i is the reverse attention-guided feature weight, Pred i is the decoded prediction feature map, F ei is the encoded feature map.

[0024] Optionally, the dimension permutation and inter-channel dependency capture of the guided feature map with weights include:

[0025] The guided feature map with weights contains multiple channels, and the spatial size of each channel is H×W. The guided feature map with weights first undergoes dimension permutation, changing from C×H×W to W×H×C, and then the inter-channel dependencies are captured through a two-layer multi-layer perceptron MLP; where, the first layer of MLP reduces the number of channels to \(1 / 4\) of the original, and then introduces non-linearity through the ReLU activation function. Then, the number of channels is restored to the original dimension through the second layer of MLP. Finally, an inverse permutation is performed to restore to C×H×W, and a channel attention map is generated through the Sigmoid activation function. The input feature map and the channel attention map are multiplied element-wise to obtain the enhanced feature map.

[0026] Optionally, the channel shuffle operation is specifically as follows:

[0027] The enhanced feature map is divided into (4) groups, each group containing (C / 4) channels. A transpose operation is performed on the grouped feature map to shuffle the channel order within each group. Subsequently, the shuffled feature map is restored to the original shape (C×H×W), thereby obtaining the shuffled feature map.

[0028] Optionally, the spatial attention enhancement includes:

[0029] The shuffled feature map passes through a 7x7 convolutional layer, and the number of channels is reduced to (1 / 4) of the original. Then, through batch normalization and the ReLU activation function for non-linear transformation. Next, the number of channels is restored to the original dimension C through the second 7x7 convolutional layer, followed by a batch normalization layer. Finally, a spatial attention map is generated through the Sigmoid activation function;

[0030] The guided feature map with weights and the channel attention map are multiplied element-wise to obtain the final output feature map.

[0031] The present application also provides a breast segmentation model training device for breast ultrasound image lesion area recognition, characterized in that the breast segmentation model training device for breast ultrasound image lesion area recognition includes:

[0032] A training set acquisition module, which is used to acquire a breast patient ultrasound image data set, and the breast patient ultrasound image data set includes at least one breast patient ultrasound image;

[0033] A breast segmentation model acquisition module, which is used to acquire a breast segmentation model. Among them, the breast segmentation model includes an MGSA module, and the MGSA module is used to obtain a guided feature map with weights according to the original image;

[0034] A training module, which is used to train the breast segmentation model through the breast patient ultrasound image data set to obtain a trained breast segmentation model.

[0035] In view of the challenges such as noise interference and blurred boundaries encountered in the segmentation process of breast ultrasound images (BUS), the present application proposes a breast segmentation model training method for breast ultrasound image lesion area recognition. The method of the present application can be widely applied in a complex clinical environment, providing more reliable support for the auxiliary diagnosis of breast ultrasound images. For problems such as blurred boundaries of breast ultrasound image lesion segmentation and insufficient capture of context information, it can ensure higher segmentation accuracy and stronger robustness. Description of the Drawings

[0036] Figure 1 is a schematic flowchart of a breast segmentation model training method for breast ultrasound image lesion area recognition in an embodiment of the present application;

[0037] Figure 2 is an architecture diagram of a breast segmentation model in an embodiment of the present application.

[0038] Figure 3 is an architecture diagram of the MGSA module.

[0039] Figure 4 It is a visualization result graph for comparison with existing methods.

[0040] Figure 5 It is a visualization result graph for ablation experiments. Specific implementation manners

[0041] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] As Figure 1 shown, the method for training a breast segmentation model for identifying lesion regions in breast ultrasound images includes:

[0043] Step 1: Obtain a breast patient ultrasound image dataset, where the breast patient ultrasound image dataset includes at least one breast patient ultrasound image;

[0044] Step 2: Obtain a breast segmentation model, where the breast segmentation model includes an MGSA module, and the MGSA module is used to obtain a weighted guidance feature map according to the original image;

[0045] Step 3: Train the breast segmentation model with the breast patient ultrasound image dataset to obtain a trained breast segmentation model.

[0046] This application proposes a method for training a breast segmentation model for identifying lesion regions in breast ultrasound images in view of challenges such as noise interference and blurred boundaries encountered in the segmentation process of breast ultrasound images (BUS). The method of this application can be widely applied in complex clinical environments, providing more reliable support for the auxiliary diagnosis of breast ultrasound images. For problems such as blurred boundaries of breast ultrasound image lesion segmentation and insufficient capture of context information, it can ensure higher segmentation accuracy and stronger robustness.

[0047] In this embodiment, Step 1: Obtain a breast patient ultrasound image dataset, where the breast patient ultrasound image dataset includes at least one breast patient ultrasound image can be achieved by the following method:

[0048] Prepare the ultrasound image dataset of breast patients. The dataset we used consists of the ultrasound image data of breast cancer patients in Peking Union Medical College Hospital from January 2021 to July 2023, which is divided into two categories: low HER2 expression and overexpression of HER2. First, it is necessary to annotate the lesion areas of these two types of datasets, and convert the annotated datasets into txt files in COCO format by labelme2coco after annotation for the next training.

[0049] 2) Divide the ultrasound image dataset of breast patients into a training set, a test set, and a validation set at a ratio of 8:1:1; the training set is used to train the neural network, the test set is used to observe the training effect to determine the training termination time, and the validation set is used to verify the change of training indicators after a certain period.

[0050] See Figure 2 and Figure 3 , in this embodiment, the breast segmentation model uses U 2 -Net as the baseline of our model building. The network architecture of U 2 -Net is a U-Net nested U-Net architecture. Its main body consists of 6 encoders and 5 decoders. Next, the MGSA module is incorporated between the encoder and the decoder to guide the learning of background and edge features to improve the segmentation of breast BUS images. The Otsu algorithm and the Scharr operator are applied in the MGSA module.

[0051] The MGSA module performs Otsu binarization on the original image to segment the foreground and background, and respectively performs Sigmoid processing on the edge feature map obtained by using the Scharr operator and the reverse feature map obtained by reverse attention with the decoded prediction feature map output by the decoder to convert them into corresponding guiding feature weights:

[0052] Guide 1i = σ(Scharr(Pred i ));

[0053] Guide 2i = σ(Otsu(O ri ));

[0054] Guide 3i = 1 - σ(Pred i );

[0055] Among them, Guide 1i is the edge guiding feature weight, Guide 2i is the background guiding feature weight, Guide 3i is the reverse attention guiding feature weight, Pred i is the decoded prediction feature map, Ori is the original image.

[0056] Multiply the above guiding feature weights by the encoded feature map weights obtained by the encoder to obtain the guiding feature map Guide of the weights i :

[0057] Guide i = Cat(Guide 1i × F ei )(Guid 2i × F ei )(Guid 3i × F ei )

[0058] where F ei is the encoded feature map.

[0059] Channel shuffle attention strategy:

[0060] This part designs a global channel-space attention mechanism that integrates channel attention, channel rearrangement, and spatial attention to capture global dependencies. The guiding feature map is first fed into the channel attention sub-module and then into the spatial attention sub-module. In the channel attention sub-module, the guiding feature map first undergoes a dimension permutation, changing from C×H×W to W×H×C. Then, a two-layer multi-layer perceptron MLP is used to capture the dependencies between channels. The first layer of MLP reduces the number of channels to (1 / 4) of the original, then introduces non-linearity through the ReLU activation function, and then restores the number of channels to the original dimension through the second layer of MLP. Finally, an inverse permutation is performed to restore to C×H×W, and a channel attention map is generated through the Sigmoid activation function. The input feature map and the channel attention map are multiplied element-wise to obtain the enhanced feature map:

[0061]

[0062] where F channel is the enhanced feature map, σ is the Sigmoid function, represents element-wise multiplication.

[0063] To further mix and share information, a channel shuffle operation is applied. The enhanced feature map is divided into (4) groups, each group containing (C / 4) channels. A transpose operation is performed on the grouped feature maps to shuffle the channel order within each group. Subsequently, the shuffled feature map is restored to the original shape (C×H×W). This way can better mix the feature information and enhance the feature expression ability.

[0064] F shuffle = ChanelShuffle(F channel );

[0065] Among them, F shuffle is the shuffled feature map.

[0066] In the spatial attention sub-module, the input feature map passes through a 7x7 convolutional layer, and the number of channels is reduced to (1 / 4) of the original. Then, it undergoes non-linear transformation through batch normalization and ReLU activation function. Next, the number of channels is restored to the original dimension C through the second 7x7 convolutional layer, followed by a batch normalization layer. Finally, a spatial attention map is generated through the Sigmoid activation function. The shuffled feature map and the spatial attention map are multiplied element-wise to obtain the final output feature map.

[0067]

[0068] Among them, F spatial is the feature map after spatial attention. The final output feature map contains enhanced features after channel attention, channel shuffling, and spatial attention.

[0069] Dysample Upsampling:

[0070] DySample upsampling is introduced into the network architecture to replace the original upsampling method. The purpose of this improvement is to improve the accuracy and robustness of breast ultrasound image segmentation by utilizing dynamic sampling and local perception functions. DySample upsampling optimizes the upsampling process by adaptively adjusting the sampling position and offset range. In DySample, pixel-level spatial rearrangement of the feature map is performed to determine the position of the upsampling points using dynamically generated offsets, which reduces the neglect of positional relationships and the disordered shift of offsets. This design enables more accurate feature recovery and avoids the artifacts and information loss problems common in traditional upsampling methods. According to the form of the distance factor and the generation method of the offset, this method adopts a PL-style dynamic distance factor. This method first destroys the pixel alignment of the input feature map through a pixel shuffling operation before applying the offset. Then, it generates offsets in each local region, which are subsequently used for upsampling.

[0071] Residual Connection:

[0072] We changed the connection method of normal convolution, aiming to fuse the context information between different channels to improve the segmentation performance of the model. Specifically, we added residual links in the inner structure of the network, including the 3-8 convolutional layers of encoder 1, the 3-7 convolutional layers of encoder 2, the 3-6 convolutional layers of encoder 3, and the 3-5 convolutional layers of encoder 5. The output obtained by convolution, BN layer, and Relu activation is superimposed with the original input, while retaining more original image information and not causing too much interference to the decoding process.

[0073] In this embodiment, the present application further includes the following steps:

[0074] Fine-tune the breast segmentation model, and the specific steps are as follows:

[0075] Our method is implemented in the PyTorch framework, and the environment is configured with Python version 3.11 and PyTorch version 2.3.1. The network is optimized using the Adam optimizer, with an initial learning rate of 10-4 and a weight decay set to 10-7. The size of all input images is adjusted to 288×288, and all experiments are conducted on an NVIDIA GeForce GTX 4060. Each model is trained for 100 epochs with a batch size of 4.

[0076] Verify the performance of the breast segmentation model and test the detection results. The specific steps are as follows:

[0077] 1) Compare our method with other existing segmentation methods. The experimental results are shown in Table 1.

[0078] Table 1 Comparison with advanced methods

[0079]

[0080] 2) To further demonstrate the superiority of the model, we conducted ablation experiments. We compared the baseline, baseline + residual connection, baseline + Dysample upsampling, baseline + MGSA module, baseline + MGSA module + Dysample with this neural network model respectively. The ablation experiment results of this neural network model are shown in Table 2.

[0081] Table 2 Ablation experiments

[0082]

[0083] In the experiment of comparing with advanced methods, our model compared with the state-of-the-art method in the HER2 low-expression dataset, and the Jaccard, Precision, Sensitivity, and Dice metrics were improved by 2.30, 1.28, 1.36, and 0.83 respectively; in the HER2 over-expression dataset compared with the state-of-the-art method, the Jaccard, Precision, Sensitivity, and Dice metrics were improved by 1.44, 0.90, 2.89, and 1.99 respectively. And from the ablation experiment results, the contributions of different modules or components of the final model can be known.

[0084] 3) The visualized results obtained from the test are as Figure 4 , Figure 5 shown. Among them, Figure 4 is the visualized result of comparing with the advanced method.Figure 5 For the visualization results of ablation experiments.

[0085] The present application also provides a training device for a breast segmentation model for identifying lesion regions in breast ultrasound images. The training device for the breast segmentation model for identifying lesion regions in breast ultrasound images includes a training set acquisition module, a breast segmentation model acquisition module, and a training module. Among them,

[0086] The training set acquisition module is used to acquire a breast patient ultrasound image data set, and the breast patient ultrasound image data set includes at least one breast patient ultrasound image;

[0087] The breast segmentation model acquisition module is used to acquire a breast segmentation model. Among them, the breast segmentation model includes an MGSA module, and the MGSA module is used to obtain a weighted guidance feature map according to the original image;

[0088] The training module is used to train the breast segmentation model through the breast patient ultrasound image data set, so as to obtain a trained breast segmentation model.

[0089] In view of the challenges such as noise interference and blurred boundaries encountered in the segmentation process of breast ultrasound images (BUS), the present invention proposes a method for segmenting lesion regions in breast ultrasound images based on a deep learning neural network. This method can be widely applied in complex clinical environments, providing more reliable support for the auxiliary diagnosis of breast ultrasound images. For problems such as blurred segmentation boundaries and insufficient capture of context information in breast ultrasound image lesion segmentation, it can ensure higher segmentation accuracy and stronger robustness. The present invention is realized through the following technical solutions. Specifically, the present invention is as follows:

[0090] The present invention uses the algorithm of a deep convolutional neural network to design a network model based on U2-Net. This model takes breast ultrasound images (BUS) as the research object, extracts image features by using a deep convolutional neural network, enhances the key information in the image through a multi-guidance attention module, combines the advantages of traditional image processing techniques and deep learning, and realizes high-precision segmentation of the lesion region. This model includes an encoder, a decoder, and a multi-guided channel shuffle attention (MGSA) module, which improves the processing ability of image edges and details. At the same time, the original bilinear upsampling is replaced by DySample upsampling, so that the model can obtain clearer regions and edge details. This method shows excellent performance in breast ultrasound image segmentation, has high accuracy and strong robustness, and is suitable for practical applications in clinical diagnosis.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A breast segmentation model training method for breast ultrasound image lesion area recognition, characterized in that: The breast segmentation model training method for breast ultrasound image lesion area recognition includes: Acquire a breast patient ultrasound image dataset, wherein the breast patient ultrasound image dataset includes at least one breast patient ultrasound image; Acquire a breast segmentation model, wherein the breast segmentation model includes a MGSA module, and the MGSA module is used to acquire a weighted guided feature map according to an original image; The breast segmentation model is trained using the breast patient ultrasound image dataset to obtain a trained breast segmentation model.

2. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 1, characterized in that: The breast segmentation model further includes a channel shuffle attention module, which is used to perform dimension permutation, inter-channel dependency capture, channel shuffle operation, and spatial attention enhancement on the weighted guide feature map, so as to obtain the final output feature map.

3. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 2, characterized in that: The breast segmentation model further includes a DySample upsampling module, which is used for pixel-level spatial rearrangement of the final output feature map to determine the position of the upsampling points using dynamically generated offsets, which reduces the neglect of positional relationships and the disordered shifting of offsets.

4. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 3, characterized in that: The MGSA module includes: A feature weight conversion module is used to perform Otsu binarization on the original image, segment the foreground and background, and splice the decoded feature map output by the decoder using the edge feature map obtained by the Scharr operator and the reverse feature map obtained by reverse attention, and then convert the spliced ​​feature map into feature weights through Sigmoid; A product module, wherein the product module is used to multiply the coded feature map obtained by the coded feature with the feature weight, so as to obtain a guided feature map with a weight.

5. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 4, characterized in that: The feature weight conversion module obtains the feature weight through the following formula: Guide 1i =σ(Scharr(Pred i )) Guide 2i = σ(Otsu(O ri )) Guide 3i =1-σ(Pred i ); Among them, Guide 1i Guide is the edge guide feature weight, 2i Guide is the background guide feature weight. 3i To guide the feature weights in reverse attention, Pred i To decode the predicted feature map, O ri is the original image.

6. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 5, characterized in that: The product module obtains the guided feature map with weights through the following formula: Guide i =Cat(Guide 1i ×F ei )(Guide 2i ×F ei )(Guide 3i ×F ei ); in, Among them, Guide i is a weighted guide feature map, Guide 1i Guide is the edge guide feature weight, 2i Guide is the background guide feature weight. 3i To guide the feature weights in reverse attention, Pred i To decode the predicted feature map, F ei is the encoding feature map.

7. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 6, characterized in that: The step of performing dimension permutation on the weighted guided feature map and capturing the inter-channel dependency relationship includes: The weighted guided feature map contains multiple channels, and the spatial size of each channel is H×W. The weighted guided feature map is first dimensionally permuted from C×H×W to W×H×C, and then the dependency between channels is captured through a two-layer multilayer perceptron MLP. Among them, the first layer of MLP reduces the number of channels to \(1 / 4\) times the original, and then introduces nonlinearity through the ReLU activation function, and then the number of channels is restored to the original dimension through the second layer of MLP. Finally, the inverse permutation is performed to restore it to C×H×W, and the channel attention map is generated through the Sigmoid activation function. The input feature map and the channel attention map are multiplied element by element to obtain an enhanced feature map.

8. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 7, characterized in that: The channel shuffling operation is as follows: The enhanced feature maps are divided into (4) groups, each group contains (C / 4) channels, and the grouped feature maps are transposed to shuffle the order of channels within each group. Subsequently, the shuffled feature maps are restored to their original shape (C×H×W) to obtain the shuffled feature maps.

9. The breast segmentation model training method for breast ultrasound image lesion area recognition according to claim 8, characterized in that: The spatial attention enhancement comprises: The shuffled feature map passes through a 7x7 convolution layer, and the number of channels is reduced to (1 / 4) times the original. Then, it is nonlinearly transformed through batch normalization and ReLU activation function. Then, the number of channels is restored to the original dimension C through the second 7x7 convolution layer, and then passed through the batch normalization layer. Finally, the spatial attention map is generated through the Sigmoid activation function. Multiply the weighted guided feature map and the channel attention map element by element to obtain the final output feature map.

10. A breast segmentation model training device for breast ultrasound image lesion area recognition, characterized in that: The breast segmentation model training device for breast ultrasound image lesion area recognition comprises: A training set acquisition module, wherein the training set acquisition module is used to acquire a breast patient ultrasound image dataset, wherein the breast patient ultrasound image dataset includes at least one breast patient ultrasound image; A breast segmentation model acquisition module, wherein the breast segmentation model acquisition module is used to acquire a breast segmentation model, wherein the breast segmentation model includes an MGSA module, and the MGSA module is used to acquire a weighted guided feature map according to an original image; A training module is used to train the breast segmentation model using the breast patient ultrasound image dataset, thereby obtaining a trained breast segmentation model.

Images