Ultrasonic breast tumor image segmentation method based on hierarchical cross correlation learning

Abstract

The invention provides an ultrasonic breast tumor image segmentation method based on hierarchical cross correlation learning. The method comprises the following steps: performing multi-stage feature extraction on an ultrasonic image; performing alignment and fusion processing on the multi-stage representation features of the ultrasonic image, and performing tumor boundary feature learning in combination with a residual structure; performing alignment and fusion processing on the multi-stage representation features of the ultrasonic image, and performing enhancement and learning by using a residual structure; and performing association modeling on the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features by using a hierarchical cross attention mechanism, and performing decoding in combination with the enhanced tumor segmentation features to obtain a tumor segmentation prediction map. According to the invention, the boundary sensing fusion module is utilized to explicitly strengthen the sensing ability to the tumor boundary area in the feature learning process, so that the problems of fuzzy boundary and inaccurate positioning in the traditional method are significantly relieved, and the fineness and coherence of tumor contour extraction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and medical image processing, and particularly to an ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning. Background Art

[0002] In the field of medical image processing, ultrasonic breast tumor image segmentation has gradually become an important research direction in recent years. Breast ultrasound images have been widely used in the screening and diagnosis of breast diseases due to their non-invasiveness, low cost, and real-time imaging capabilities. Tumor segmentation aims to accurately extract the shape and boundary information of breast tumors by analyzing ultrasound image data, providing support for high-level medical tasks such as subsequent clinical diagnosis, pathological analysis, and treatment planning.

[0003] However, ultrasonic images usually have problems such as low contrast, strong noise, artifact interference, and blurred boundaries, which pose great challenges to the automatic segmentation task of breast tumors. Traditional image segmentation methods rely on manually designed features or simple image processing techniques, making it difficult to effectively cope with the complex and variable ultrasonic imaging characteristics, and having poor adaptability to images collected from different patients and different devices.

[0004] In recent years, the rapid development of deep learning technology has promoted the progress of medical image segmentation technology. In particular, convolutional neural networks (CNNs) have demonstrated excellent performance in feature extraction and spatial modeling, enabling significant achievements in breast tumor segmentation methods based on deep learning. However, the CNN structure is mainly good at capturing local feature relationships and lacks the ability to model global context information, easily leading to problems such as inaccurate boundaries and internal holes in the segmentation results. In addition, the tumor region usually has complex morphological changes, and it is difficult to comprehensively characterize the structural characteristics of the tumor region by relying solely on single-scale or single-branch feature extraction and fusion.

[0005] To address the above problems, some studies have introduced multi-scale feature fusion, boundary awareness mechanisms, and Transformer architectures to enhance the model's global dependence modeling and fine-grained segmentation capabilities. In particular, cross-feature level correlation and interactive feature fusion strategies, while effectively capturing information at different scales and different levels, contribute to more accurate boundary localization and region segmentation. However, existing methods still have certain limitations in cross-level feature interaction and hierarchical feature correlation learning, and it is still difficult to fully exploit the complex structural features of the tumor region and its fine-grained differences from the background tissue. Summary of the Invention

[0006] In view of the above situation, the main purpose of the present invention is to propose an ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning to solve the above technical problems.

[0007] The present invention provides an ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning, and the method includes the following steps: Step 1: Perform multi - stage feature extraction on the ultrasonic image to obtain the multi - stage representation features of the ultrasonic image; Step 2: Perform alignment and fusion processing on the multi - stage representation features of the ultrasonic image to obtain the fused multi - scale tumor boundary features. Perform tumor boundary feature learning on the multi - scale tumor boundary features combined with the residual structure to obtain a tumor boundary prediction map; Step 3: Perform alignment and fusion processing on the multi - stage representation features of the ultrasonic image to obtain the fused multi - scale tumor segmentation features. Use the residual structure to enhance and learn the fused multi - scale tumor segmentation features to obtain enhanced tumor segmentation features; Step 4: Perform correlation modeling on the fused multi - scale tumor boundary features and the fused multi - scale tumor segmentation features using a hierarchical cross - attention mechanism to obtain refined tumor features with complementary and enhanced features; Step 5: Perform convolution operations and up - sampling operations on the refined tumor features with complementary and enhanced features and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map; Construct a tumor boundary loss based on the tumor boundary prediction map, construct a tumor segmentation loss based on the tumor segmentation prediction map, optimize the image segmentation model using the tumor boundary loss and the tumor segmentation loss to obtain an optimized image segmentation model, and input the ultrasonic image into the optimized image segmentation model to obtain the final segmentation result.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing a cross - attention interaction and fusion module, the present invention effectively realizes the deep association and complementary fusion between features of different scales and different semantic levels, fully excavates the fine - grained structural characteristics and context information of the tumor region, thereby improving the accuracy and integrity of breast tumor segmentation; 2. The present invention uses a boundary - aware fusion module to explicitly enhance the perception ability of the tumor boundary region during the feature learning process, significantly alleviates the problems of blurred boundaries and inaccurate positioning in traditional methods, and improves the fineness and coherence of tumor contour extraction; 3. Based on the multi - scale feature extraction strategy and combined with a hierarchical interaction mechanism, the present invention realizes the efficient integration and optimization of features with different resolutions, enables the system to adapt to the diverse changes in the morphology, size, and texture complexity of breast tumors, and enhances the robustness and generalization ability of the image segmentation model in complex ultrasonic image scenarios; The additional aspects and advantages of the present invention will be partially given in the following description, partially become obvious from the following description, or be understood through the embodiments of the present invention. Description of the Drawings

[0009] Figure 1 This is the flowchart of the steps of the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention.

[0010] Figure 2 This is the structural diagram of the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention.

[0011] Figure 3 This is the schematic diagram of the multi - scale context fusion module of the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention.

[0012] Figure 4 This is the result comparison chart between the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention and the existing method.

[0013] Figure 5 This is the visualization result of the predicted segmentation map and the boundary map of the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention.

[0014] Figure 6 This is the quantitative comparison experiment result between the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention and the existing method on the datasets BUSI, UDIAT, and BUS320.

[0015] Figure 7 This is the quantitative comparison experiment result between the ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning proposed by the present invention and the existing method on the datasets SRBUI and STU. Detailed implementation manners

[0016] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] Referring to the following description and drawings, these and other aspects of the embodiments of the present invention will be clear. In these descriptions and drawings, some specific implementation manners in the embodiments of the present invention are specifically disclosed as some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited by this.

[0018] Please refer to Figure 1 , an embodiment of the present invention proposes an ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning. The method includes the following steps: Step 1: Perform multi-stage feature extraction on the ultrasound image to obtain the multi-stage representation features of the ultrasound image.

[0019] Please refer to Figure 2 , in Step 1, when performing multi-stage feature extraction on the ultrasound image to obtain the multi-stage representation features of the ultrasound image, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the multi-stage representation features of the ultrasound image, represents the encoder backbone network of the RGB image stream, represents the input ultrasound image, represents the real number set space with a feature dimension of , represents the channel index number of the tumor feature map in the th stage, represents the th stage height of the tumor feature map, represents the th stage width of the tumor feature map.

[0020] It should be noted that a multi-level feature extraction network is constructed based on the encoder-decoder structure. In Figure 2 , MSCF represents multi-scale context fusion, TCFE represents tumor context feature enhancement, CBR represents convolution, batch normalization, and ReLU activation function, CAIF represents cross-attention interaction fusion, U represents upsampling, and S represents Softmax activation function; the encoder Res2Net in the multi-level feature extraction network adopts a multi-scale residual structure to obtain rich semantic information, providing basic support for subsequent tumor region and boundary feature modeling; the cross-attention interaction fusion also includes a linear layer.

[0021] Step 2: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features, and perform tumor boundary feature learning on the fused multi-scale tumor boundary features combined with the residual structure to obtain the tumor boundary prediction map.

[0022] Please refer to Figure 3 , in Step 2, when aligning and fusing the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features, and performing tumor boundary feature learning on the fused multi-scale tumor boundary features combined with the residual structure to obtain the tumor boundary prediction map, it specifically includes the following steps: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features; Enhance the fused multi-scale tumor boundary features to obtain an enhanced feature expression of the tumor boundary; Feature enhancement of the enhanced tumor boundary is performed using a residual structure to obtain enhanced tumor boundary features; Convolution operations and upsampling operations are sequentially performed on the enhanced tumor boundary features to obtain a tumor boundary prediction map.

[0023] Alignment and fusion processing are performed on the multi-stage representation features of the ultrasound image to obtain fused multi-scale tumor boundary features. The relationship in the corresponding process is as follows: ; Among them, represents the small-scale tumor boundary features after feature alignment, represents the large-scale tumor boundary features after feature alignment, represents the feature upsampling layer, represents the ReLU activation function, represents the convolution operation with a 1x1 convolution kernel, represents the small-scale tumor boundary features to be fused, represents the large-scale tumor boundary features to be fused, represents the fused multi-scale tumor boundary features, represents a convolution operation with a convolution kernel of ; represents the feature concatenation operation; In the step of enhancing the fused multi-scale tumor boundary features to obtain the feature expression of the enhanced tumor boundary, the relationship in the corresponding process is as follows: ; Among them, represents the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation using the first receptive field, represents the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation using the second receptive field, represents the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation using the third receptive field, represents the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation using the fourth receptive field, represents the convolution operation with a 3x1 convolution kernel, represents the convolution operation with a 1x3 convolution kernel, represents the convolution operation with a 5x1 convolution kernel, represents the convolution operation with a 1x5 convolution kernel, represents the convolution operation with a 7x1 convolution kernel, represents the convolution operation with a 1x7 convolution kernel; In the step of enhancing the feature expression of the enhanced tumor boundary using the residual structure to obtain the enhanced tumor boundary features, the relational expressions in the corresponding process are as follows: ; Among them, represents the enhanced tumor boundary features, represents the concatenated enhanced tumor boundary feature expression, represents the element-wise addition operation; In the step of sequentially performing a convolution operation and an upsampling operation on the enhanced tumor boundary features to obtain a tumor boundary prediction map, the relational expressions in the corresponding process are as follows: ; Among them, represents the upsampled tumor boundary features, represents the tumor boundary prediction map, represents the Sigmoid activation function.

[0024] It should be noted that based on the feature alignment and fusion mechanism, a boundary multi-scale context fusion module and a segmentation multi-scale context fusion module are respectively constructed. The boundary multi-scale context fusion module and the boundary decoder constitute a boundary detection branch, and the tumor boundary features are learned from the multi-stage representation features of the ultrasound image in the boundary detection branch.

[0025] Step 3: Perform alignment and fusion processing on the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor segmentation features, and use the residual structure to enhance and learn the fused multi-scale tumor segmentation features to obtain the enhanced tumor segmentation features.

[0026] In Step 3, performing alignment and fusion processing on the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor segmentation features, and using the residual structure to enhance and learn the fused multi-scale tumor segmentation features to obtain the enhanced tumor segmentation features specifically includes the following steps: Perform alignment and fusion processing on the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor segmentation features; Perform enhancement processing on the fused multi-scale tumor segmentation features to obtain the feature expression of the enhanced ultrasound image tumor segmentation; Perform feature enhancement on the feature expression of the enhanced ultrasound image tumor segmentation in combination with the residual structure to obtain the enhanced tumor segmentation features.

[0027] Perform alignment and fusion processing on the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor segmentation features, and the relational expressions in the corresponding process are as follows: ; Among them, represents the small-scale tumor segmentation feature after feature alignment, represents the large-scale tumor segmentation feature after feature alignment, represents the small-scale tumor segmentation feature to be fused, represents the large-scale tumor segmentation feature to be fused, represents the multi-scale tumor segmentation feature after fusion; In the step of enhancing the multi-scale tumor segmentation feature after fusion to obtain the feature expression of ultrasonic image tumor segmentation, the relational expressions in the corresponding process are as follows: ; Among them, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the first receptive field, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the second receptive field, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the third receptive field, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the fourth receptive field; In the step of enhancing the feature expression of ultrasonic image tumor segmentation by combining the residual structure to obtain the enhanced tumor segmentation feature, the relational expressions in the corresponding process are as follows: ; Among them, represents the enhanced tumor segmentation feature expression after splicing, represents the enhanced tumor segmentation feature.

[0028] It should be noted that the segmentation multi-scale context fusion module and the segmentation decoder constitute the region segmentation branch, and the multi-stage representation features of the ultrasonic image are enhanced and learned in the region segmentation branch.

[0029] Step 4: Use the hierarchical cross-attention mechanism to perform correlation modeling on the fused multi-scale tumor boundary feature and the fused multi-scale tumor segmentation feature to obtain the refined tumor feature with complementary enhanced features.

[0030] In Step 4, using the hierarchical cross-attention mechanism to perform correlation modeling on the fused multi-scale tumor boundary feature and the fused multi-scale tumor segmentation feature to obtain the refined tumor feature with complementary enhanced features specifically includes the following steps: For the tumor boundary features and tumor segmentation features, use the learnable linear layer of the key vector and the learnable linear layer of the value vector to obtain the key vector of the tumor boundary features and the value vector of the tumor boundary features respectively; Use the learnable linear layer of the query vector to obtain the query vector of the tumor segmentation features; For the key vector of the tumor boundary features, the query vector of the tumor segmentation features, and the value vector of the tumor boundary features, use the cross-attention mechanism for attention weighting processing to obtain the tumor features after attention weighting; perform element-wise addition calculation on the tumor features after attention weighting and the fused multi-scale tumor segmentation features using the residual structure to obtain the refined tumor features with complementary features enhanced.

[0031] For the tumor boundary features and tumor segmentation features, use the learnable linear layer of the key vector and the learnable linear layer of the value vector to obtain the key vector of the tumor boundary features and the value vector of the tumor boundary features respectively. The relational expressions in the corresponding process are as follows: ; Among them, represents the key vector of the tumor boundary features, represents the parameters of the learnable linear layer of the key vector for the tumor boundary, represents the tumor boundary features at the th level, represents the value vector of the tumor boundary features, represents the parameters of the learnable linear layer of the value vector for the tumor boundary; In the step of using the learnable linear layer of the query vector to obtain the query vector of the tumor segmentation features, the relational expressions in the corresponding process are as follows: ; Among them, represents the query vector of the tumor segmentation features, represents the parameters of the learnable linear layer of the query vector for the tumor segmentation, represents the tumor segmentation features at the th level; In the step of using the cross-attention mechanism for attention weighting processing on the key vector of the tumor boundary features, the query vector of the tumor segmentation features, and the value vector of the tumor boundary features to obtain the tumor features after attention weighting; performing element-wise addition calculation on the tumor features after attention weighting and the fused multi-scale tumor segmentation features using the residual structure to obtain the refined tumor features with complementary features enhanced, the relational expressions in the corresponding process are as follows: ; Among them, represents the cross-attention weight mapping graph, denote activation function denote the transpose of a matrix denote the characteristic dimension of the key vector denote the tumor features after attention weighting denote the element-wise multiplication operation denote the refined tumor features with enhanced feature complementarity

[0032] It should be noted that based on the Transformer architecture, a cross-attention interaction fusion module is constructed. The fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features are input into the cross-attention interaction fusion module, and the cross-attention mechanism is used for correlation modeling processing

[0033] Step 5: Perform convolution operations and upsampling operations on the refined tumor features with enhanced feature complementarity and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map Construct a tumor boundary loss based on the tumor boundary prediction map, construct a tumor segmentation loss based on the tumor segmentation prediction map, use the tumor boundary loss and the tumor segmentation loss to optimize the segmentation model, obtain an optimized segmentation model, and input the ultrasound image into the optimized segmentation model to obtain the final segmentation result

[0034] In Step 5, convolution operations and upsampling operations are performed on the refined tumor features with enhanced feature complementarity and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map. The corresponding relationship is as follows ; where denote the upsampled tumor segmentation features denote the tumor segmentation prediction map

[0035] Construct a tumor boundary loss based on the tumor boundary prediction map. The corresponding relationship is as follows ; where denote the tumor boundary loss denote the binary cross-entropy loss function denote the tumor boundary ground truth map It should be noted that the tumor boundary loss is used to focus on optimizing the tumor boundary region, guiding the network to pay attention to edge details, improving the model's recognition ability of the boundary contour, and further enhancing the accuracy and robustness of the overall segmentation

[0036] In the step of constructing the tumor segmentation loss based on the tumor segmentation prediction map, the corresponding relationship is as follows ; Among them, represents the tumor segmentation loss, represents the ground truth map of the tumor boundary, represents the intersection over union loss function.

[0037] Furthermore, there is also a total loss obtained by using the tumor boundary loss and the tumor segmentation loss. The corresponding relationship in the process is as follows: ; Among them, represents the total loss, represents a constant used to balance the tumor boundary loss and the segmentation loss.

[0038] Please refer to Figure 4 , Figure 4 which shows a comparison chart of the prediction results of the method proposed in the embodiment of the present invention and the existing methods. Image represents the ultrasound tumor image, GT represents the ground truth map of tumor segmentation, TBANet (Ours) represents the result of the method proposed in the present invention, and MDFNet, ESKNet, NUNet, CF2Net, MSUNet, MDANet, ExpdUnet, Stan, SKUNet, RDAUNet represent the results of other existing ultrasound tumor segmentation methods. It can be seen from Figure 4 that the method proposed in the embodiment of the present invention can obtain better segmentation results.

[0039] Please refer to Figure 5 , Figure 5 which shows the visualization results of the predicted segmentation map and the boundary map of the method proposed in the embodiment of the present invention.

[0040] Please refer to Figure 6 , Figure 6 which shows the quantitative comparison experiment results of the method proposed in the embodiment of the present invention and the existing methods on the datasets BUSI, UDIAT, and BUS320.

[0041] Please refer to Figure 7 , Figure 7 which shows the quantitative comparison experiment results of the method proposed in the embodiment of the present invention and the existing methods on the datasets SRBUI and STU.

[0042] Furthermore, the present invention adopts a double-branch decoder structure to decode the multi-scale features output by the encoder respectively: the boundary detection branch is used for detecting the tumor boundary region, and the region segmentation branch is used for segmenting the tumor region. The expression ability of the tumor region and boundary information is enhanced through the multi-scale context fusion and feature enhancement module, and the joint modeling of the region and the boundary is realized.

[0043] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one of the following techniques known in the art or a combination thereof can be used: discrete logic circuits having logic gate circuits for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0044] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0045] The above-described embodiments merely represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed, but should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

Claims

1. An ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning, characterized in that, The method includes the following steps: Step 1: Perform multi-stage feature extraction on the ultrasound image to obtain the multi-stage representation features of the ultrasound image; Step 2: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features, and perform tumor boundary feature learning on the fused multi-scale tumor boundary features combined with the residual structure to obtain the tumor boundary prediction map; Step 3: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor segmentation features, and enhance and learn the fused multi-scale tumor segmentation features using the residual structure to obtain the enhanced tumor segmentation features; Step 4: Use the hierarchical cross-attention mechanism to perform correlation modeling on the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features to obtain the refined tumor features with complementary enhanced features; Step 5: Perform convolution operations and upsampling operations on the refined tumor features with complementary enhanced features and the enhanced tumor segmentation features in sequence to obtain the tumor segmentation prediction map; Construct a tumor boundary loss based on the tumor boundary prediction map, construct a tumor segmentation loss based on the tumor segmentation prediction map, use the tumor boundary loss and the tumor segmentation loss to optimize the image segmentation model, obtain the optimized image segmentation model, and input the ultrasound image into the optimized image segmentation model to obtain the final segmentation result.

2. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 1, characterized in that In the above Step 1, when performing multi-stage feature extraction on the ultrasound image to obtain the multi-stage representation features of the ultrasound image, the relational expressions in the corresponding process are as follows: ； Among them, represents the representation features of multiple stages of ultrasound images, represents the encoder backbone network of the RGB image stream, represents the input ultrasound image, represents that the feature dimension is the real number set space of, represents the channel index number of the tumor feature map in the th stage, represents the th stage height of the tumor feature map, represents the th stage width of the tumor feature map.

3. The method for segmenting ultrasonic breast tumor images based on hierarchical cross - correlation learning according to claim 2, wherein, In the above Step 2, when aligning and fusing the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features, and performing tumor boundary feature learning on the fused multi-scale tumor boundary features combined with the residual structure to obtain the tumor boundary prediction map, it specifically includes the following steps: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features; Enhance the fused multi-scale tumor boundary features to obtain the feature expression of the enhanced tumor boundary; Perform feature enhancement on the feature expression of the enhanced tumor boundary using the residual structure to obtain the enhanced tumor boundary features; Perform convolution operations and upsampling operations on the enhanced tumor boundary features in sequence to obtain the tumor boundary prediction map.

4. The ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning according to claim 3, wherein, When aligning and fusing the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features, the relational expressions in the corresponding process are as follows: ； Among them, represents the small-scale tumor boundary feature after feature alignment, represents the large-scale tumor boundary feature after feature alignment, represents the feature upsampling layer, represents the ReLU activation function, represents the convolution operation with a 1x1 kernel convolution, represents the small-scale tumor boundary feature to be fused, represents the large-scale tumor boundary feature to be fused, represents the multi-scale tumor boundary feature after fusion, represents that the convolution kernel is of the convolution operation, represents the feature concatenation operation; In the step of enhancing the fused multi-scale tumor boundary features to obtain the feature expression of the enhanced tumor boundary, the relational expressions in the corresponding process are as follows: ； Among them, represents the feature expression of the tumor boundary of the ultrasound image enhanced by the convolution operation using the first receptive field, represents the feature expression of the tumor boundary of the ultrasound image enhanced by the convolution operation using the second receptive field, represents the feature expression of the tumor boundary of the ultrasound image enhanced by the convolution operation using the third receptive field, represents the feature expression of the tumor boundary of the ultrasound image enhanced by the convolution operation using the fourth receptive field, represents the convolution operation with a convolution kernel of 3x1, represents the convolution operation with a convolution kernel of 1x3, represents the convolution operation with a convolution kernel of 5x1, represents the convolution operation with a convolution kernel of 1x5, represents the convolution operation with a convolution kernel of 7x1, represents the convolution operation with a convolution kernel of 1x7; In the step of performing feature enhancement on the feature expression of the enhanced tumor boundary using the residual structure to obtain the enhanced tumor boundary features, the relational expressions in the corresponding process are as follows: ； Among them, represents the enhanced tumor boundary feature, represents the expression of the enhanced tumor boundary feature after splicing, represents the element-wise addition operation; In the step of performing convolution operations and upsampling operations on the enhanced tumor boundary features in sequence to obtain the tumor boundary prediction map, the relational expressions in the corresponding process are as follows: ； Among them, represents the upsampled tumor boundary features, represents the tumor boundary prediction map, represents the Sigmoid activation function.

5. The method for segmenting ultrasonic breast tumor images based on hierarchical cross - correlation learning according to claim 4, wherein, In step 3, the representation features at multiple stages of the ultrasound image are aligned and fused to obtain the fused multi-scale tumor segmentation features, and the fused multi-scale tumor segmentation features are enhanced and learned using a residual structure to obtain the enhanced tumor segmentation features. The specific steps are as follows: Align and fuse the representation features at multiple stages of the ultrasound image to obtain the fused multi-scale tumor segmentation features; Enhance the fused multi-scale tumor segmentation features to obtain the feature expression of the enhanced ultrasound image tumor segmentation; For the feature expression of the enhanced ultrasound image tumor segmentation, combine it with the residual structure for feature enhancement to obtain the enhanced tumor segmentation features.

6. The method for segmenting ultrasonic breast tumor images based on hierarchical cross - correlation learning according to claim 5, wherein, Align and fuse the representation features at multiple stages of the ultrasound image to obtain the fused multi-scale tumor segmentation features. The corresponding relationship in the process is as follows: ； Among them, represents the small-scale tumor segmentation features after feature alignment, represents the large-scale tumor segmentation features after feature alignment, represents the small-scale tumor segmentation features to be fused, represents the large-scale tumor segmentation features to be fused, represents the multi-scale tumor segmentation features after fusion; In the step of enhancing the fused multi-scale tumor segmentation features to obtain the feature expression of the enhanced ultrasound image tumor segmentation, the corresponding relationship in the process is as follows: ； Among them, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the first receptive field, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the second receptive field, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the third receptive field, represents the feature expression of ultrasonic image tumor segmentation enhanced by the convolution operation using the fourth receptive field; In the step of combining the feature expression of the enhanced ultrasound image tumor segmentation with the residual structure for feature enhancement to obtain the enhanced tumor segmentation features, the corresponding relationship in the process is as follows: ； Among them, represents the enhanced tumor segmentation feature expression after splicing, represents the enhanced tumor segmentation feature.

7. The ultrasonic breast tumor image segmentation method based on hierarchical cross - correlation learning according to claim 6, wherein In step 4, the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features are associated and modeled using a hierarchical cross-attention mechanism to obtain the refined tumor features with complementary enhanced features. The specific steps are as follows: For the tumor boundary features and the tumor segmentation features, use the learnable linear layer of the key vector and the learnable linear layer of the value vector to obtain the key vector of the tumor boundary features and the value vector of the tumor boundary features respectively; Use the learnable linear layer of the query vector to obtain the query vector of the tumor segmentation features; Perform attention weighting processing on the key vector of the tumor boundary features, the query vector of the tumor segmentation features, and the value vector of the tumor boundary features using the hierarchical cross-attention mechanism to obtain the attention-weighted tumor features; perform element-wise addition calculation on the attention-weighted tumor features and the fused multi-scale tumor segmentation features using the residual structure to obtain the refined tumor features with complementary enhanced features.

8. The method for segmenting ultrasonic breast tumor images based on hierarchical cross-correlation learning according to claim 7, wherein For the tumor boundary features and the tumor segmentation features, use the learnable linear layer of the key vector and the learnable linear layer of the value vector to obtain the key vector of the tumor boundary features and the value vector of the tumor boundary features respectively. The corresponding relationship in the process is as follows: ； Among them, The key vector representing the tumor boundary feature, Represents the learnable linear layer parameters for the key vector of the tumor boundary, Indicates the Tumor boundary feature of the The value vector representing the tumor boundary feature, Represents the learnable linear layer parameters for the value vector of the tumor boundary; In the step of using the learnable linear layer of the query vector to obtain the query vector of the tumor segmentation features, the corresponding relationship in the process is as follows: ； Among them, represents the query vector of tumor segmentation features, represents the learnable linear layer parameters of the query vector for tumor segmentation, represents the tumor segmentation features at the In the step of performing attention weighting processing on the key vector of the tumor boundary features, the query vector of the tumor segmentation features, and the value vector of the tumor boundary features using the hierarchical cross-attention mechanism to obtain the attention-weighted tumor features; perform element-wise addition calculation on the attention-weighted tumor features and the fused multi-scale tumor segmentation features using the residual structure to obtain the refined tumor features with complementary enhanced features. The corresponding relationship in the process is as follows: ； Among them, represents the cross-attention weight mapping graph, represents the activation function, represents the transpose of the matrix, represents the characteristic dimension of the key vector, represents the tumor characteristics after attention weighting, represents the element-wise multiplication operation, represents the refined tumor characteristics with enhanced feature complementarity.

9. The method for segmenting ultrasonic breast tumor images based on hierarchical cross - correlation learning according to claim 8, wherein, In step 5, convolution operations and upsampling operations are sequentially performed on the refined tumor features with complementary enhanced features and the enhanced tumor segmentation features to obtain a tumor segmentation prediction map. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the upsampled tumor segmentation features, represents the tumor segmentation prediction map.

10. The method for segmenting ultrasonic breast tumor images based on hierarchical cross - correlation learning according to claim 9, wherein, Based on the tumor boundary prediction map, a tumor boundary loss is constructed. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the tumor boundary loss, represents the binary cross-entropy loss function, represents the ground truth map of the tumor boundary; In the step of constructing a tumor segmentation loss based on the tumor segmentation prediction map, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the tumor segmentation loss, represents the ground truth map of the tumor boundary, represents the intersection over union loss function.

Images