Ultrasound breast tumor image segmentation method based on hierarchical cross-correlation learning

Abstract

The present invention proposes a method for ultrasound breast tumor image segmentation based on hierarchical cross-correlation learning, which includes: performing multi-stage feature extraction on ultrasound images; aligning and fusing the multi-stage representation features of ultrasound images, and learning tumor boundary features in combination with a residual structure; aligning and fusing the multi-stage representation features of ultrasound images, and enhancing and learning them using a residual structure; performing association modeling on the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features using a hierarchical cross-attention mechanism, and decoding them in combination with the enhanced tumor segmentation features to obtain a tumor segmentation prediction map. The present invention utilizes a boundary perception fusion module to explicitly enhance the perception of tumor boundary areas during the feature learning process, significantly alleviating the problems of blurred boundaries and inaccurate positioning in traditional methods, and improving the precision and consistency of tumor contour extraction.

Description

Technical Field

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

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

[0003] However, ultrasound images often suffer from low contrast, strong noise, artifacts, and blurred boundaries, making the automatic segmentation of breast tumors a significant challenge. Traditional image segmentation methods rely on handcrafted features or simple image processing techniques, which struggle to effectively address the complex and changing characteristics of ultrasound imaging and are less adaptable to images collected from different patients and devices.

[0004] In recent years, the rapid development of deep learning technology has driven advancements in medical image segmentation. Convolutional neural networks (CNNs), in particular, have demonstrated exceptional performance in feature extraction and spatial modeling, enabling significant progress in deep learning-based breast tumor segmentation methods. However, CNN architectures primarily excel at capturing local feature relationships and lack the ability to model global contextual information, which can easily lead to problems such as inaccurate boundaries and internal voids in segmentation results. Furthermore, tumor regions often exhibit complex morphological variations, making it difficult to fully characterize the structural characteristics of tumor regions by relying solely on feature extraction and fusion at a single scale or branch.

[0005] To address these issues, some studies have introduced multi-scale feature fusion, boundary-aware mechanisms, and Transformer architectures to enhance the model's global dependency modeling and fine-grained segmentation capabilities. In particular, cross-feature hierarchical association and interactive feature fusion strategies effectively capture information at different scales and levels while helping to achieve more accurate boundary localization and regional segmentation. However, existing methods still have limitations in cross-level feature interaction and hierarchical feature association learning, making it difficult to fully exploit the complex structural characteristics of tumor regions and their fine-grained differences from 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 proposes a method for ultrasonic breast tumor image segmentation based on hierarchical cross-correlation learning, which comprises the following steps:

[0008] Step 1: Perform multi-stage feature extraction on the ultrasound image to obtain multi-stage representation features of the ultrasound image;

[0009] Step 2: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features. Combine the multi-scale tumor boundary features with the residual structure to perform tumor boundary feature learning and obtain a tumor boundary prediction map.

[0010] Step 3: Align and fuse the multi-stage representation features of the ultrasound image to obtain a fused multi-scale tumor segmentation feature, and enhance and learn the fused multi-scale tumor segmentation feature using a residual structure to obtain an enhanced tumor segmentation feature;

[0011] Step 4: Use the hierarchical cross-attention mechanism to perform association modeling on the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features to obtain the tumor refined features with complementary enhancement;

[0012] Step 5: Convolution and upsampling operations are performed on the complementary enhanced tumor refinement features and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map;

[0013] A tumor boundary loss is constructed based on the tumor boundary prediction map, and a tumor segmentation loss is constructed based on the tumor segmentation prediction map. The image segmentation model is optimized using the tumor boundary loss and the tumor segmentation loss to obtain an optimized image segmentation model. The ultrasound image is input into the optimized image segmentation model to obtain the final segmentation result.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. By introducing a cross-attention interactive fusion module, this paper effectively realizes the deep correlation and complementary fusion between features of different scales and semantic levels, fully exploits the fine-grained structural characteristics and contextual information of the tumor area, thereby improving the accuracy and completeness of breast tumor segmentation;

[0016] 2. This invention utilizes a boundary perception fusion module to explicitly enhance the ability to perceive tumor boundary areas during feature learning, significantly alleviating the problems of blurred boundaries and inaccurate positioning in traditional methods, and improving the precision and consistency of tumor contour extraction.

[0017] 3. Based on a multi-scale feature extraction strategy and combined with a hierarchical interaction mechanism, this invention achieves efficient integration and optimization of features at different resolutions. This enables the system to adapt to the diverse variations in morphology, size, and texture complexity of breast tumors, and enhances the robustness and generalization ability of the image segmentation model in complex ultrasound image scenarios.

[0018] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flowchart of the steps of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in the present invention.

[0020] Figure 2 This is a structural diagram of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in this invention.

[0021] Figure 3 This is a schematic diagram of the multi-scale context fusion module of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in the present invention.

[0022] Figure 4 This is a comparison chart of the results of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in this invention and the existing methods.

[0023] Figure 5 This is the visualization result of the predicted segmentation map and boundary map of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in this invention.

[0024] Figure 6 This paper presents the quantitative comparative experimental results of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in this invention and the existing methods on the datasets BUSI, UDIAT, and BUS320.

[0025] Figure 7 This is the quantitative comparative experimental results of the ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning proposed in this invention and the existing methods on the data sets SRBUI and STU. DETAILED DESCRIPTION

[0026] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0027] These and other aspects of the embodiments of the present invention will become clear with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to provide 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 thereto.

[0028] See also Figure 1 The embodiment of the present invention proposes a method for ultrasonic breast tumor image segmentation based on hierarchical cross-correlation learning, which includes the following steps:

[0029] Step 1: Perform multi-stage feature extraction on the ultrasound image to obtain multi-stage representation features of the ultrasound image.

[0030] See also Figure 2 In step 1, multi-stage feature extraction is performed on the ultrasound image to obtain the multi-stage representation features of the ultrasound image. The corresponding relationship between the process is as follows:

[0031] ;

[0032] in, Representation features of multiple phases of ultrasound images, The encoder backbone network representing the RGB image stream, represents the input ultrasound image, Indicates that the feature dimension is The space of real numbers, Indicates the Channel index number of the stage tumor feature map, Indicates the The height of the stage tumor feature map, Indicates the The width of the stage tumor feature map.

[0033] It should be noted that a multi-level feature extraction network is constructed based on the encoder-decoder structure. Figure 2 In the figure, 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; cross-attention interaction fusion also includes linear layers.

[0034] Step 2: Align and fuse the multi-stage representation features of the ultrasound image to obtain the fused multi-scale tumor boundary features. Combine the fused multi-scale tumor boundary features with the residual structure to perform tumor boundary feature learning to obtain a tumor boundary prediction map.

[0035] See also Figure 3 In step 2, the multi-stage representation features of the ultrasound image are aligned and fused to obtain the fused multi-scale tumor boundary features. The fused multi-scale tumor boundary features are combined with the residual structure to perform tumor boundary feature learning to obtain a tumor boundary prediction map. The specific steps include:

[0036] Align and fuse the multi-stage representation features of ultrasound images to obtain fused multi-scale tumor boundary features;

[0037] The fused multi-scale tumor boundary features are enhanced to obtain enhanced tumor boundary feature expression;

[0038] The feature expression of the enhanced tumor boundary is enhanced using the residual structure to obtain the enhanced tumor boundary feature;

[0039] The enhanced tumor boundary features are sequentially subjected to convolution and upsampling operations to obtain a tumor boundary prediction map.

[0040] The multi-stage representation features of the ultrasound image are aligned and fused to obtain the fused multi-scale tumor boundary features. The corresponding relationship is as follows:

[0041] ;

[0042] in, 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 of kernel convolution 1x1, 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, The convolution kernel is The convolution operation, Represents feature concatenation operation;

[0043] In the step of enhancing the fused multi-scale tumor boundary features to obtain the enhanced tumor boundary feature expression, the corresponding process has the following relationship:

[0044] ;

[0045] in, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the first receptive field, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the second receptive field, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the third receptive field, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the fourth receptive field, Indicates a convolution operation with a convolution kernel of 3x1. Indicates a convolution operation with a convolution kernel of 1x3. Indicates a convolution operation with a convolution kernel of 5x1. Indicates a convolution operation with a convolution kernel of 1x5. Indicates a convolution operation with a convolution kernel of 7x1. Indicates a convolution operation with a convolution kernel of 1x7;

[0046] In the step of enhancing the feature expression of the enhanced tumor boundary using the residual structure to obtain the enhanced tumor boundary feature, the corresponding process has the following relationship:

[0047] ;

[0048] in, represents the enhanced tumor boundary features, Indicates the enhanced tumor boundary feature expression after splicing, Represents element-by-element addition operation;

[0049] 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 corresponding process has the following relationship:

[0050] ;

[0051] in, represents the upsampled tumor boundary features, represents the tumor boundary prediction map, Represents the Sigmoid activation function.

[0052] 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 constructed respectively. The boundary multi-scale context fusion module and the boundary decoder constitute the boundary detection branch, and the multi-stage representation features of the ultrasound image are used to learn the tumor boundary features in the boundary detection branch.

[0053] 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.

[0054] In step 3, the multi-stage representation features of the ultrasound image are aligned and fused to obtain a fused multi-scale tumor segmentation feature. The fused multi-scale tumor segmentation feature is enhanced and learned using a residual structure to obtain an enhanced tumor segmentation feature. The specific steps include the following:

[0055] Align and fuse the multi-stage representation features of ultrasound images to obtain fused multi-scale tumor segmentation features;

[0056] The fused multi-scale tumor segmentation features are enhanced to obtain enhanced ultrasound image tumor segmentation feature expression;

[0057] The feature expression of enhanced ultrasound image tumor segmentation is combined with the residual structure to perform feature enhancement to obtain enhanced tumor segmentation features.

[0058] The multi-stage representation features of the ultrasound image are aligned and fused to obtain the fused multi-scale tumor segmentation features. The corresponding relationship is as follows:

[0059] ;

[0060] in, 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 fused multi-scale tumor segmentation features;

[0061] In the step of enhancing the fused multi-scale tumor segmentation features to obtain the enhanced ultrasound image tumor segmentation feature expression, the corresponding process has the following relationship:

[0062] ;

[0063] in, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the first receptive field, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the second receptive field, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the third receptive field, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the fourth receptive field;

[0064] In the step of combining the feature expression of enhanced ultrasound image tumor segmentation with the residual structure to enhance the features and obtain the enhanced tumor segmentation features, the corresponding process has the following relationship:

[0065] ;

[0066] in, represents the enhanced tumor segmentation feature expression after splicing, Represents the enhanced tumor segmentation features.

[0067] 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 ultrasound image are enhanced and learned in the region segmentation branch.

[0068] Step 4: Use the hierarchical cross-attention mechanism to perform association modeling on the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features to obtain tumor refined features with complementary enhancement.

[0069] In step 4, the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features are associated with each other using a hierarchical cross-attention mechanism to obtain tumor refined features with complementary enhancement. The specific steps are as follows:

[0070] A learnable linear layer of a key vector and a learnable linear layer of a value vector are used for the tumor boundary feature and the tumor segmentation feature to obtain a key vector of the tumor boundary feature and a value vector of the tumor boundary feature respectively;

[0071] Using a learnable linear layer on the query vector, we obtain the query vector of tumor segmentation features.

[0072] The key vector of the tumor boundary feature, the query vector of the tumor segmentation feature, and the value vector of the tumor boundary feature are subjected to attention weighting processing using the cross-attention mechanism to obtain the attention-weighted tumor features; the attention-weighted tumor features and the fused multi-scale tumor segmentation features are added element-by-element using the residual structure to obtain the tumor refined features with complementary and enhanced features.

[0073] The tumor boundary features and tumor segmentation features are respectively obtained by using a learnable linear layer of key vectors and a learnable linear layer of value vectors to obtain the key vector of the tumor boundary features and the value vector of the tumor boundary features. The corresponding relationship between the processes is as follows:

[0074] ;

[0075] in, The key vector representing the tumor boundary features, represents the key vector for the tumor boundary to learn the linear layer parameters, Indicates the Tier-level tumor boundary features, A vector of values representing the tumor boundary features, The vector of values representing the tumor boundary can be used to learn the linear layer parameters;

[0076] In the step of obtaining the query vector of tumor segmentation features using the learnable linear layer of the query vector, the corresponding process has the following relationship:

[0077] ;

[0078] in, The query vector representing the tumor segmentation features, represents the query vector for tumor segmentation to learn the linear layer parameters, Indicates the Hierarchical tumor segmentation features;

[0079] The key vector of the tumor boundary feature, the query vector of the tumor segmentation feature, and the value vector of the tumor boundary feature are subjected to attention weighting using the cross-attention mechanism to obtain the attention-weighted tumor features. The attention-weighted tumor features and the fused multi-scale tumor segmentation features are then element-wise added using the residual structure to obtain the feature-complementary enhanced tumor refined features. The corresponding process has the following relationship:

[0080] ;

[0081] in, represents the cross attention weight map, express activation function, represents the transpose of the matrix, represents the characteristic dimension of the key vector, represents the tumor features after attention weighting, represents an element-wise multiplication operation, Represents tumor refinement with enhanced feature complementarity.

[0082] It should be noted that a cross-attention interaction fusion module is constructed based on the Transformer architecture, and 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 association modeling processing.

[0083] Step 5: Convolution and upsampling operations are performed on the complementary enhanced tumor refinement features and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map;

[0084] A tumor boundary loss is constructed based on the tumor boundary prediction map, and a tumor segmentation loss is constructed based on the tumor segmentation prediction map. The segmentation model is optimized using the tumor boundary loss and the tumor segmentation loss to obtain an optimized segmentation model. The ultrasound image is input into the optimized segmentation model to obtain the final segmentation result.

[0085] In step 5, convolution and upsampling operations are performed on the complementary enhanced tumor refinement features and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map. The relationship between the corresponding processes is as follows:

[0086] ;

[0087] in, represents the upsampled tumor segmentation features, Represents the tumor segmentation prediction map.

[0088] The tumor boundary loss is constructed based on the tumor boundary prediction map. The relationship between the corresponding process is as follows:

[0089] ;

[0090] in, represents the tumor boundary loss, represents the binary cross entropy loss function, represents the ground truth map of the tumor boundary;

[0091] It should be noted that the tumor boundary loss is used to focus on optimizing the tumor boundary area, guiding the network to focus on edge details, improving the model's ability to recognize boundary contours, and further improving the accuracy and robustness of the overall segmentation.

[0092] In the step of constructing the tumor segmentation loss based on the tumor segmentation prediction map, the corresponding process has the following relationship:

[0093] ;

[0094] in, represents the tumor segmentation loss, represents the ground truth map of the tumor boundary, represents the intersection-over-union loss function.

[0095] Furthermore, there is also the use of tumor boundary loss and tumor segmentation loss to obtain the total loss. The relationship between the corresponding process is as follows:

[0096] ;

[0097] in, represents the total loss, represents a constant used to balance the tumor boundary loss and segmentation loss.

[0098] See also Figure 4 , Figure 4 The following figure shows the comparison 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 tumor segmentation truth map, TBANet (Ours) represents the result of the method proposed in the present invention, and MDFNet, ESKNet, NUNet, CF2Net, MSUNet, MDANet, ExpdUnet, Stan, SKUNet, and RDAUNet represent the results of other existing ultrasound tumor segmentation methods. Figure 4 It can be seen from the figure that the method proposed in the embodiment of the present invention can obtain better segmentation results.

[0099] See also Figure 5 , Figure 5 The visualization results of the predicted segmentation map and boundary map of the method proposed in the embodiment of the present invention are shown.

[0100] See also Figure 6 , Figure 6 The quantitative comparative experimental results of the method proposed in the embodiment of the present invention and the existing methods on the datasets BUSI, UDIAT, and BUS320 are shown.

[0101] See also Figure 7 , Figure 7 The quantitative comparative experimental results of the method proposed in the embodiment of the present invention and the existing methods on the SRBUI and STU datasets are shown.

[0102] Furthermore, the present invention adopts a dual-branch decoder structure to decode the multi-scale features output by the encoder respectively: the boundary detection branch is used for tumor boundary area detection, and the region segmentation branch is used for tumor region segmentation. The multi-scale context fusion and feature enhancement module improves the expression ability of tumor region and boundary information, and realizes the joint modeling of region and boundary.

[0103] It should be understood that various components of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0104] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0105] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for ultrasound breast tumor image segmentation based on hierarchical cross-correlation learning, characterized in that: The method comprises the following steps: Step 1: Perform multi-stage feature extraction on the ultrasound image to obtain multi-stage representation features of the ultrasound image; Step 2: Align and fuse the multi-stage representation features of the ultrasound image to obtain fused multi-scale tumor boundary features. Combine the fused multi-scale tumor boundary features with the residual structure to perform tumor boundary feature learning and obtain a tumor boundary prediction map. Step 3: Align and fuse the multi-stage representation features of the ultrasound image to obtain a fused multi-scale tumor segmentation feature, and enhance and learn the fused multi-scale tumor segmentation feature using a residual structure to obtain an enhanced tumor segmentation feature; Step 4: Use the hierarchical cross-attention mechanism to perform association modeling on the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features to obtain the tumor refined features with complementary enhancement; Step 5: Convolution and upsampling operations are performed on the complementary enhanced tumor refinement features and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map; A tumor boundary loss is constructed based on the tumor boundary prediction map, and a tumor segmentation loss is constructed based on the tumor segmentation prediction map. The image segmentation model is optimized using the tumor boundary loss and the tumor segmentation loss to obtain an optimized image segmentation model. The ultrasound image is input 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 step 1, multi-stage feature extraction is performed on the ultrasound image to obtain the multi-stage representation features of the ultrasound image. The corresponding relationship between the processes is as follows: ； in, Representation features of multiple phases of ultrasound images, The encoder backbone network represents the RGB image stream, represents the input ultrasound image, Indicates that the feature dimension is The space of real numbers, Indicates the Channel index number of the stage tumor feature map, Indicates the The height of the stage tumor feature map, Indicates the The width of the stage tumor feature map.

3. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 2, characterized in that: In step 2, the multi-stage representation features of the ultrasound image are aligned and fused to obtain fused multi-scale tumor boundary features. The fused multi-scale tumor boundary features are combined with the residual structure to perform tumor boundary feature learning to obtain a tumor boundary prediction map, which specifically includes the following steps: Align and fuse the multi-stage representation features of ultrasound images to obtain fused multi-scale tumor boundary features; The fused multi-scale tumor boundary features are enhanced to obtain enhanced tumor boundary feature expression; The feature expression of the enhanced tumor boundary is enhanced using the residual structure to obtain the enhanced tumor boundary feature; The enhanced tumor boundary features are sequentially subjected to convolution and upsampling operations to obtain a tumor boundary prediction map.

4. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 3 is characterized in that: The multi-stage representation features of the ultrasound image are aligned and fused to obtain the fused multi-scale tumor boundary features. The corresponding relationship is as follows: ； in, 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 of kernel convolution 1x1, 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, The convolution kernel is The convolution operation, Represents feature concatenation operation; In the step of enhancing the fused multi-scale tumor boundary features to obtain the enhanced tumor boundary feature expression, the corresponding process has the following relationship: ； in, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the first receptive field, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the second receptive field, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the third receptive field, Indicates the feature expression of the ultrasound image tumor boundary enhanced by the convolution operation of the fourth receptive field, Indicates a convolution operation with a convolution kernel of 3x1. Indicates a convolution operation with a convolution kernel of 1x3. Indicates a convolution operation with a convolution kernel of 5x1. Indicates a convolution operation with a convolution kernel of 1x5. Indicates a convolution operation with a convolution kernel of 7x1. Indicates a convolution operation with a convolution kernel of 1x7; In the step of enhancing the feature expression of the enhanced tumor boundary using the residual structure to obtain the enhanced tumor boundary feature, the corresponding process has the following relationship: ； in, represents the enhanced tumor boundary features, Indicates the enhanced tumor boundary feature expression after splicing, Represents element-by-element 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 corresponding process has the following relationship: ； in, represents the upsampled tumor boundary features, represents the tumor boundary prediction map, Represents the Sigmoid activation function.

5. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 4 is characterized in that: In step 3, the multi-stage representation features of the ultrasound image are aligned and fused to obtain a fused multi-scale tumor segmentation feature. The fused multi-scale tumor segmentation feature is enhanced and learned using a residual structure to obtain an enhanced tumor segmentation feature. Specifically, the steps include: Align and fuse the multi-stage representation features of ultrasound images to obtain fused multi-scale tumor segmentation features; The fused multi-scale tumor segmentation features are enhanced to obtain enhanced ultrasound image tumor segmentation feature expression; The enhanced ultrasound image tumor segmentation feature expression is combined with the residual structure to perform feature enhancement to obtain the enhanced tumor segmentation feature.

6. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 5, characterized in that: The multi-stage representation features of the ultrasound image are aligned and fused to obtain the fused multi-scale tumor segmentation features. The corresponding relationship is as follows: ； in, 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 fused multi-scale tumor segmentation features; In the step of enhancing the fused multi-scale tumor segmentation features to obtain the enhanced ultrasound image tumor segmentation feature expression, the corresponding process has the following relationship: ； in, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the first receptive field, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the second receptive field, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the third receptive field, Indicates the feature expression of enhanced ultrasound image tumor segmentation using the convolution operation of the fourth receptive field; In the step of combining the feature expression of enhanced ultrasound image tumor segmentation with the residual structure to enhance the features and obtain the enhanced tumor segmentation features, the corresponding process has the following relationship: ； in, represents the enhanced tumor segmentation feature expression after splicing, Represents the enhanced tumor segmentation features.

7. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 6, characterized in that: In step 4, the fused multi-scale tumor boundary features and the fused multi-scale tumor segmentation features are associated with each other using a hierarchical cross-attention mechanism to obtain a tumor refined feature with complementary enhancement, which specifically includes the following steps: For tumor boundary features and tumor segmentation features, a learnable linear layer of key vectors and a learnable linear layer of value vectors are used to obtain the key vector of tumor boundary features and the value vector of tumor boundary features respectively; Using a learnable linear layer on the query vector, we obtain the query vector of tumor segmentation features. The key vector of the tumor boundary feature, the query vector of the tumor segmentation feature, and the value vector of the tumor boundary feature are subjected to attention weighting processing using the hierarchical cross-attention mechanism to obtain the attention-weighted tumor features; the attention-weighted tumor features and the fused multi-scale tumor segmentation features are added element-by-element using the residual structure to obtain the tumor refined features with complementary and enhanced features.

8. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 7, characterized in that: The tumor boundary features and tumor segmentation features are respectively obtained by using a learnable linear layer of key vectors and a learnable linear layer of value vectors to obtain the key vector of the tumor boundary features and the value vector of the tumor boundary features. The corresponding relationship between the processes is as follows: ； in, The key vector representing the tumor boundary features, represents the key vector for the tumor boundary to learn the linear layer parameters, Indicates the Tier-level tumor boundary features, A vector of values representing the tumor boundary features, The vector of values representing the tumor boundary can be used to learn the linear layer parameters; In the step of obtaining the query vector of tumor segmentation features using the learnable linear layer of the query vector, the corresponding process has the following relationship: ； in, The query vector representing the tumor segmentation features, represents the query vector for tumor segmentation to learn the linear layer parameters, Indicates the Hierarchical tumor segmentation features; The key vector of the tumor boundary feature, the query vector of the tumor segmentation feature, and the value vector of the tumor boundary feature are subjected to attention weighting using the hierarchical cross-attention mechanism to obtain the attention-weighted tumor features. The attention-weighted tumor features and the fused multi-scale tumor segmentation features are then element-wise added using the residual structure to obtain the feature-complementary enhanced tumor refined features. The corresponding process has the following relationship: ； in, represents the cross attention weight map, express activation function, represents the transpose of the matrix, represents the characteristic dimension of the key vector, represents the tumor features after attention weighting, represents an element-wise multiplication operation, Represents tumor refinement with enhanced feature complementarity.

9. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 8, characterized in that: In step 5, convolution and upsampling operations are performed on the complementary enhanced tumor refinement features and the enhanced tumor segmentation features in sequence to obtain a tumor segmentation prediction map. The corresponding process has the following relationship: ； in, represents the upsampled tumor segmentation features, Represents the tumor segmentation prediction map.

10. The ultrasonic breast tumor image segmentation method based on hierarchical cross-correlation learning according to claim 9, characterized in that: The tumor boundary loss is constructed based on the tumor boundary prediction map. The relationship between the corresponding process is as follows: ； in, 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 the tumor segmentation loss based on the tumor segmentation prediction map, the corresponding process has the following relationship: ； in, represents the tumor segmentation loss, represents the ground truth map of the tumor boundary, represents the intersection-over-union loss function.

Images