Breast ultrasonic image segmentation method and system based on decoupling learning

Abstract

The invention provides a breast ultrasound image segmentation method and system based on decoupling learning, and relates to the technical field of breast ultrasound image processing and analysis, and the method comprises the steps: obtaining a to-be-segmented breast ultrasound image; inputting the breast ultrasound image into the decoupling learning segmentation model to generate a prediction probability graph, and mapping the prediction probability graph into a final focus segmentation result; according to decoupling learning, a two-stage training method is adopted, an encoder and a decoder of a segmentation model are trained respectively, in the first-stage training process, fuzzy activation is achieved based on region recombination, a high-confidence region is dynamically replaced with a fuzzy block, gradient flowing is enhanced through a context with stable disturbance, and the robustness of the segmentation model is improved. Structural conflicts and semantic variability are generated in a high confidence region, forcing an encoder to learn semantic invariance under context changes. According to the method, the challenges that a current breast ultrasound image segmentation method is insufficient in fuzzy region processing optimization, fragmentized in feature representation and reduced in boundary quality are solved, and the segmentation precision and robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of breast ultrasound image processing and analysis technology, specifically to a breast ultrasound image segmentation method and system based on decoupling learning. Background Technology

[0002] In the field of breast ultrasound image processing, segmentation of breast ultrasound images (BUS) is crucial for the early diagnosis and treatment planning of breast cancer. Accurate lesion segmentation helps doctors quickly identify lesion regions and their boundaries, improving diagnostic efficiency. However, because breast ultrasound images often contain blurred regions, typically caused by blurred boundaries, texture degradation, or high similarity in pixel intensity distribution between the lesion and surrounding tissue, traditional segmentation methods face significant challenges in handling such regions. These blurred regions often generate an uncertainty probability close to 0.5 in prediction, resulting in weak gradient signals, thus becoming a major bottleneck for segmentation accuracy.

[0003] The goal of lesion segmentation is to accurately isolate the lesion region from the input breast ultrasound image. Given an input image... Traditional methods typically predict lesion masks through pixel-level classification or region segmentation. The optimization problem can be expressed as: ,in The loss function is often cross-entropy loss or Dice loss, which is achieved by minimizing the prediction mask. With the real mask The differences between them are used to optimize model parameters. However, traditional methods, such as threshold-based segmentation or manual feature extraction techniques, struggle to effectively handle the high uncertainty and complex background interference in blurred regions. These methods typically assume that image regions have clear boundaries, ignoring the weak gradient problem in blurred regions, leading to fragmented representations or degraded boundary quality in the segmentation results near lesion boundaries.

[0004] With the development of deep learning technology, segmentation methods based on convolutional neural networks (CNNs) have significantly improved the performance of breast ultrasound image processing. U-Net and its variants (such as U-Net++ and Attention U-Net) enhance multi-scale feature extraction capabilities through encoder-decoder structures and skip connections, exhibiting better adaptability to blurred boundaries than traditional methods. Recent studies have attempted to handle blurred regions through loss reweighting, dynamic sampling, or contrastive learning. For example, Focal Loss improves the model's attention to blurred regions by weighting hard samples, SegNet assists upsampling through record pooling indexes, and DeepLab uses dilated convolutions to preserve spatial details. However, these methods still have limitations: they mainly improve the visibility of blurred regions through passive means (such as weighting or sampling) but fail to truly activate these regions, resulting in blurred regions not being fully learned during the optimization process, manifesting as semantic gaps and inaccurate boundaries.

[0005] To address these issues, some studies have proposed improvement schemes. For example, some methods model the uncertainty of blurred regions through dedicated branches or contrast constraints, or enhance boundary accuracy through multi-scale fusion and edge-aware loss. However, these methods still face challenges in breast ultrasound image segmentation: on the one hand, weak gradient signals in blurred regions lead to insufficient optimization, with models tending to focus on high-confidence regions and neglecting the fine-grained structure of blurred regions; on the other hand, the semantic misalignment between deep and shallow features and the inconsistency in multi-scale feature transfer have not been adequately addressed. Therefore, existing techniques cannot actively activate blurred regions and optimize their semantic consistency and boundary integrity, resulting in insufficient accuracy and robustness in breast ultrasound image segmentation. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a breast ultrasound image segmentation method and system based on decoupled learning. This method addresses the challenges of current breast ultrasound image segmentation methods, such as insufficient optimization in handling ambiguous regions, fragmented feature representations, and degraded boundary quality, thereby improving segmentation accuracy and robustness.

[0007] According to some embodiments, the present invention adopts the following technical solution: A breast ultrasound image segmentation method based on decoupling learning, comprising: Obtain the breast ultrasound image to be segmented; Breast ultrasound images are input into a decoupled learning segmentation model to generate a prediction probability map, which is then mapped to the final lesion segmentation result. The segmentation model is built based on U-Net, and the decoupled learning adopts a two-stage training method to train the encoder and decoder of the segmentation model separately. In the first stage of encoder training, fuzzy activation based on region recombination is used to dynamically replace high-confidence regions with fuzzy blocks. Gradient flow is enhanced by perturbing the stable context, generating structural conflicts and semantic variability in high-confidence regions, forcing the encoder to learn semantic invariance under context changes.

[0008] According to some embodiments, the present invention adopts the following technical solution: A breast ultrasound image segmentation system based on decoupling learning, comprising: The image acquisition module is configured to acquire breast ultrasound images to be segmented; The lesion segmentation module is configured to: input breast ultrasound images into a decoupled learning segmentation model, generate a prediction probability map, and map the prediction probability map to the final lesion segmentation result; The segmentation model is built based on U-Net, and the decoupled learning adopts a two-stage training method to train the encoder and decoder of the segmentation model separately. In the first stage of encoder training, fuzzy activation based on region recombination is used to dynamically replace high-confidence regions with fuzzy blocks. Gradient flow is enhanced by perturbing the stable context, generating structural conflicts and semantic variability in high-confidence regions, forcing the encoder to learn semantic invariance under context changes.

[0009] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned breast ultrasound image segmentation method based on decoupling learning.

[0010] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned breast ultrasound image segmentation method based on decoupling learning.

[0011] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the breast ultrasound image segmentation method based on decoupling learning.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: To further improve segmentation accuracy, this invention employs a decoupled learning approach for two-stage training of the segmentation model. In the first stage, the AAR (Fuzzy Activation Module) dynamically replaces regions with high prediction confidence with blurred image patches during training, enhancing gradient flow by perturbing the stable context. In the second stage, the encoder is frozen to preserve blur-sensitive features, while the decoder performs structure recovery through two modules. Specifically, the ASA (Awareness of Ambiguity) module ensures consistency between deep and shallow features, while the ABR (Aspect-Based Memory Recovery) module enhances boundary details through multi-scale filtering. Experimental results demonstrate that the proposed segmentation method achieves excellent performance, showing a significant improvement in the segmentation of blurred regions. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is a flowchart of the method in Example 1. Figure 2 This is a schematic diagram of the segmentation model in Example 1.

[0015] Figure 3 This is a schematic diagram of the ASA structure in Example 1. Figure 4 This is a schematic diagram of the ABR structure in Example 1. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0019] Example 1 One embodiment of the present invention provides a breast ultrasound image segmentation method based on decoupling learning, comprising: Step S1: Obtain the breast ultrasound image to be segmented; Step S2: Input the breast ultrasound image into the decoupled learning segmentation model to generate a prediction probability map, and map the prediction probability map to the final lesion segmentation result; The segmentation model is built based on U-Net, and the decoupled learning adopts a two-stage training method to train the encoder and decoder of the segmentation model separately. In the first stage of encoder training, fuzzy activation based on region recombination is used to dynamically replace high-confidence regions with fuzzy blocks. Gradient flow is enhanced by perturbing the stable context, generating structural conflicts and semantic variability in high-confidence regions, forcing the encoder to learn semantic invariance under context changes.

[0020] As one embodiment, this invention provides a breast ultrasound image segmentation method based on decoupled learning, addressing the challenges of current breast ultrasound image segmentation methods such as insufficient optimization in handling blurred regions, fragmented feature representations, and degraded boundary quality, thereby improving segmentation accuracy and robustness. Figure 1 As shown, the specific implementation process is as follows: Step 1: Acquire and preprocess the breast ultrasound image dataset, specifically: Step 1.1: Obtain the publicly available breast ultrasound image datasets BUSI-E9, Dataset B, and STU as training datasets. Simultaneously, perform preliminary processing on the training datasets by reading the data according to the image acquisition order and dividing the original training dataset. and the original test dataset .

[0021] Step 1.2: Preprocess the training dataset: For the original training dataset obtained in step 1.1 and the original test dataset The training dataset is obtained by standardization. and the preprocessed test dataset .

[0022] Step 2: Construct the encoder for the segmentation model and perform the first stage of training, specifically: Step 2.1: Construct as follows Figure 2 The encoder shown uses a pre-trained RexNet as the encoder to generate multi-scale feature maps; Step 2.1.1: RexNet is selected as the base network for the encoder. This network obtains its initial parameters through pre-training on the ImageNet dataset, which effectively solves the cold start problem of the network model. Step 2.1.2: Input the preprocessed training dataset RexNet extracts features through different stages of the network; Step 2.1.3: Set up a five-level downsampling stage in the encoder, i.e. Figure 2 Steps 1-5 generate multi-scale feature maps respectively. , , , and The resolutions of these images are reduced sequentially to 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input image, with the number of channels being 32, 38, 61, 128, and 1280, respectively. Step 2.2: In order to carry out the first stage of training, construct the segmentation head. Keep the structure of the segmentation head as simple as possible to avoid providing an additional structure with feature extraction capabilities outside the encoder module, thereby optimizing the performance of the encoder as much as possible. Therefore, the segmentation head contains only one convolutional layer. Step 2.3: Subsample the features obtained in Step 5 using the fifth-level downsampling method. The image is enlarged to the original input size and then passed through a segmentation head to obtain the predicted result p, i.e., the predicted probability map. After passing through a sigmoid function, the loss is calculated. At the same time, the predicted probability map p is input into the fuzzy activation module (AAR module) implemented by region recombination for activation.

[0023] Step 2.4: Fuzzy Activation Module (AAR Module) for Region Reorganization; Ambiguous regions typically produce uncertain predictions close to 0.5, resulting in weak gradients and hindering model learning. To actively activate these regions, an uncertainty-guided region reorganization mechanism is introduced during training. This mechanism enhances the activation of ambiguous regions by explicitly perturbing the spatial context. Its goal is to generate structural conflicts and semantic variability in high-confidence regions, forcing the model to learn semantic invariance under contextual changes.

[0024] Step 2.4.1: This represents the input image for the training batch. This represents the corresponding real segmentation mask. The predicted probability map is the prediction result obtained by inputting the images of this batch into steps 2.1-2.3 above.

[0025] Step 2.4.2: For each predicted probability map Calculate its uncertainty graph This serves as the basis for fuzzy regions in dynamic localization training batches:

[0026] in, It quantifies the uncertainty of the forecast, reaching its maximum value when the forecast is most uncertain.

[0027] Step 2.4.3: In order to obtain from the source image Extract the clear foreground region and use a threshold. For mask Perform threshold processing:

[0028] This determines the spatial extent of the object of interest (foreground). .

[0029] Step 2.4.4: Next, calculate the foreground region. Minimum bounding box:

[0030] Define the foreground block for reorganization. .

[0031] Before cutting, Apply a series of transformations to introduce contextual changes:

[0032]

[0033]

[0034] in, The enhanced foreground block is used as the blurred block to be pasted. , and Sampling is performed from a predefined range to ensure enhanced diversity.

[0035] In order to dynamically determine the target image during training Appropriate paste area, in uncertainty graph Perform a sliding window search to select the region with the lowest average uncertainty (i.e., the high-confidence region), where the window size is related to the enhanced foreground block. The dimensional matching can be expressed by the formula:

[0036] in, This is the starting position of the region.

[0037] Paste the blur block to The high-confidence regions indicated introduce controlled semantic contrast, which helps with structural differentiation. However, overlap with the target foreground can lead to structural interference. To avoid this, overlapping pixels are removed.

[0038] An in-sample region replacement strategy is employed to insert blurred blocks into high-confidence regions of the same image to simulate semantic conflicts, expressed by the formula:

[0039] in, It is the smallest rectangle containing the foreground pixels; It is a binary mask that indicates which pixels belong to the non-overlapping foreground area to be pasted; These are the starting coordinates of the optimal paste position, determined by a sliding window search, designed to find the region in the target image with the lowest average uncertainty. The + here is not arithmetic addition, but rather indicates that the mask will be added. Placed in The position is the top left corner; It is the scaling factor, which determines the proportion of change in the size of the foreground block; The rotation angle; This is the contrast adjustment factor; It is the foreground threshold, used to distinguish between the foreground and the background. It is an uncertainty graph. The larger the value of a pixel, the more uncertain the model is about classifying that pixel.

[0040] This process actively introduces fuzzy patterns into high-confidence regions, enriching structural variability and enhancing gradient flow. It effectively forces the encoder to focus on underoptimized regions and learn invariant representations under contextual perturbations. The encoder is input again to generate new predictions, the loss is calculated, and the encoder parameters are optimized.

[0041] Step 3: Construct the decoder for the segmentation model and perform the second stage of training; Step 3.1: Prepare input training data Input image , The batch size represents the number of samples input at one time. H and W represent the height and width of the image, respectively. C is the number of channels, representing the feature dimension of each pixel.

[0042] Step 3.2: Initialize encoder parameters, load the encoder weights obtained in the first stage of training in Step 2, and generate multi-scale feature maps from the input X through the encoder. , , , and The resolution of each feature map is reduced to 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image, and the number of channels is 32, 38, 61, 128 and 1280, respectively. The multi-scale feature maps are passed to the decoder through skip connections. Step 3.3: Construct the decoder, introducing the semantic alignment module for perceiving ambiguity (ASA) and the boundary recovery module for focusing ambiguity (ABR). Step 3.3.1: Design the decoder structure, which includes a five-level upsampling stage corresponding to the encoder. Each upsampling stage gradually enlarges the feature map resolution to the original input size through transpose convolution operation. Step 3.3.2: As Figure 2 As shown, in each upsampling stage, skip connection feature maps are received from the encoder. and the output of the previous decoder Upsampled feature map The common inputs are fed into the semantic alignment module to obtain the output. Then Compared with upsampling The connection is established, and then the input is fed into the Ambiguity Boundary Recovery (ABR) module to obtain the output. .

[0043] Step 3.3.3: In each upsampling stage, embed a semantic alignment module (ASA) for perceptual ambiguity and an ambiguity-focusing boundary recovery module (ABR) to progressively recover spatial details and output a segmentation mask through a cascaded approach; Step 3.4: The ASA (Awareness of Ambiguity) module is mainly used in architectures such as U-Net, where the decoder needs to fuse deep features containing rich semantic information but low resolution with shallow features containing rich details but weak semantics. However, in ambiguous regions (such as lesion boundaries), this fusion can easily lead to semantic misalignment, i.e., high-level semantics cannot accurately guide the recovery of low-level details, and low-level noise interferes with the expression of high-level semantics. Therefore, such as... Figure 3 As shown, the specific steps of ASA are as follows: No. The ASA module of the layer receives encoder features. and semantic features from the previous ABR module To align their resolutions, first... Upsampling Subsequently, and All are projected into a unified embedding space through 3×3 convolution and batch normalization (BN) to eliminate the channel mismatch problem:

[0044]

[0045] Then, these two transformed features are fused by adding them element by element:

[0046] The obtained feature map It integrates local structural cues and global semantic information. However, its limited receptive field restricts its context modeling capabilities, which is particularly important for ambiguous regions. Therefore, 2×2 average pooling is applied, followed by a 3×3 convolution. The processed features are then upsampled back to the original resolution.

[0047]

[0048] To further optimize semantic alignment, a residual attention mechanism is adopted to adaptively reweight based on the semantic ambiguity of each spatial region:

[0049] in Represents the sigmoid function. The generated attention map. This attention map, which highlights semantically inconsistent or underrepresented regions, is used to modulate shallow encoder features:

[0050] By suppressing overconfident background responses while enhancing responses in blurred regions, the optimized feature f i This improves the fidelity of the representation of uncertain regions. The output is then passed to subsequent modules for further structure recovery.

[0051] Step 3.5: Focus on Ambiguous Boundary Recovery Module (ABR). This module progressively enhances boundary integrity and recovers fine-grained structural cues, especially for boundary recovery in uncertain regions, such as... Figure 4 As shown, specifically: Semantic alignment features from ASA and the upsampled result from the output of the previous layer decoder. The two are fused by adding elements one by one:

[0052] Subsequently, a 3×3 convolution is applied to obtain a richer representation:

[0053] To capture boundary structures across multiple receptive fields, two parallel convolutional branches were employed:

[0054]

[0055] in, Focusing on local contour details, This captures a broader range of contextual features. The two outputs are then aggregated to form a unified representation:

[0056] To adaptively balance the contributions of different receptive fields, we compute a soft fusion map:

[0057]

[0058] Blurred regions typically exhibit low edge contrast. Therefore, we separate high-frequency gradient information by subtracting the locally averaged version of the feature map:

[0059]

[0060] The resulting residuals highlight the structural gradients lost during the early fusion process.

[0061] Next, a spatial attention mechanism is introduced to guide the feature response to focus on the information-rich boundary region:

[0062]

[0063] To further optimize the representation, we compute the channel attention map using spatial average pooling and max pooling, and then apply a shared transformation function:

[0064]

[0065]

[0066]

[0067]

[0068] By combining adaptive scale fusion with attention-guided refinement processes, the output of the ABR module is obtained:

[0069] To enhance the discriminative power of multi-level features in the decoder, we adjust the output of each ABR module. A supplementary supervision branch is then introduced. Specifically, it first compresses the data using a 1×1 convolution. Then, it is upsampled to the original image size to obtain an auxiliary prediction map. This is used to calculate the loss with the ground truth label, expressed by the formula:

[0070] Each auxiliary prediction The auxiliary loss is calculated using the actual mask M:

[0071] in, , These are the Binary Cross-Entropy Loss and the Dice Loss, used to measure the difference between the predicted mask and the true mask, respectively, and are defined as follows:

[0072]

[0073] Where N represents the total number of pixels, and They represent the first The auxiliary prediction and the true mask are in the first The value at each pixel position, This is a smoothing term used to avoid the denominator being zero.

[0074] Ultimately, the total loss of Stage II assigns differentiated weights to different levels, emphasizing deep semantic output while preserving shallow detail guidance:

[0075] Step 4: Use the trained model to predict lesion segmentation; Step 4.1: Prepare input training data Input sequence , This is the batch size, representing the number of samples input at one time. H and W represent the height and width of the image, respectively, and C is the number of channels, representing the feature dimension of each pixel.

[0076] Step 4.2: Utilize the encoder part of the trained model to perform multi-scale feature extraction on the input data. Input the input tensor X into the model trained in Step 3. First, the encoder extracts features; the encoder generates multi-scale feature maps by downsampling layer by layer. ( ),in , Number of channels The numbers are 32, 38, 61, 128, and 1280 respectively. Step 4.3: Process the feature map and generate a predicted probability map by using semantic alignment for perceptual ambiguity (ASA) and boundary recovery for spatial focus ambiguity (ABR) in the decoder; convert the feature map output by the encoder... The signal is passed to the decoder via skip connections, and the decoder upsamples layer by layer through transposed convolutions. In the ASA module, encoder features and upsampled features from the previous layer Perform cross-level feature alignment, context enhancement, and ambiguity-aware modulation to generate semantically aligned features. .

[0077] In the ABR module, features from ASA are... The upsampled output of the previous decoder is subjected to multi-scale feature extraction, gradient enhancement, and attention-guided refinement processing to generate refined features. . Step 4.4: Map the predicted probability map to the final lesion segmentation result and generate a segmentation mask. ; the predicted probability map generated in step 4.3 Post-processing is performed, and a thresholding operation is used to generate a binary segmentation mask. The process is represented as follows: ,in ,like ,but (lesion area), otherwise (Background area); Output the final segmentation mask This represents the lesion area in each input image, used for subsequent evaluation or clinical applications.

[0078] Tables 1 and 2 compare the results of different representative segmentation algorithms on the BUSI-E9 and Dataset B datasets, and the results fully demonstrate the superiority of the method in this embodiment for image segmentation tasks.

[0079] Table 1 Comparison results of the BUSI-E9 dataset

[0080] Table 2 Comparison results of Dataset B

[0081] Example 2 One embodiment of the present invention provides a breast ultrasound image segmentation system based on decoupling learning, comprising: The image acquisition module is configured to acquire breast ultrasound images to be segmented; The lesion segmentation module is configured to: input breast ultrasound images into a decoupled learning segmentation model, generate a prediction probability map, and map the prediction probability map to the final lesion segmentation result; The segmentation model is built based on U-Net, and the decoupled learning adopts a two-stage training method to train the encoder and decoder of the segmentation model separately. In the first stage of encoder training, fuzzy activation based on region recombination is used to dynamically replace high-confidence regions with fuzzy blocks. Gradient flow is enhanced by perturbing the stable context, generating structural conflicts and semantic variability in high-confidence regions, forcing the encoder to learn semantic invariance under context changes.

[0082] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned breast ultrasound image segmentation method based on decoupling learning.

[0083] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned breast ultrasound image segmentation method based on decoupling learning.

[0084] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the breast ultrasound image segmentation method based on decoupling learning.

[0085] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A breast ultrasound image segmentation method based on decoupling learning, characterized in that, include: Obtain the breast ultrasound image to be segmented; Breast ultrasound images are input into a decoupled learning segmentation model to generate a prediction probability map, which is then mapped to the final lesion segmentation result. The segmentation model is built based on U-Net, and the decoupled learning adopts a two-stage training method to train the encoder and decoder of the segmentation model separately. In the first stage of encoder training, fuzzy activation based on region recombination is used to dynamically replace high-confidence regions with fuzzy blocks. Gradient flow is enhanced by perturbing the stable context, generating structural conflicts and semantic variability in high-confidence regions, forcing the encoder to learn semantic invariance under context changes.

2. The breast ultrasound image segmentation method based on decoupling learning as described in claim 1, characterized in that, The segmentation model includes an encoder and a decoder; The encoder performs multi-level downsampling on the input breast ultrasound image to generate multi-scale features; The decoder performs multi-level upsampling on the multi-scale features generated by the encoder to generate a prediction probability map.

3. The breast ultrasound image segmentation method based on decoupling learning as described in claim 1, characterized in that, The fuzzy activation achieved by the region recombination is used to construct new first-stage training samples, specifically: Obtain the original training samples, including breast ultrasound images and corresponding real segmentation masks and prediction probability maps; For each prediction probability map, calculate the uncertainty map; Segmenting the foreground region from breast ultrasound images and constructing blurred blocks; Perform a sliding window search on the uncertainty map and select the region with the lowest average uncertainty as the high confidence region; By inserting the blurred block into the high-confidence region of the same breast ultrasound image, a new training sample is obtained.

4. The breast ultrasound image segmentation method based on decoupling learning as described in claim 1, characterized in that, The decoder includes a semantic alignment module that senses ambiguity and a boundary recovery module for spatially focused ambiguity; The semantic alignment module is used to perform cross-level feature alignment, context enhancement, and ambiguity-aware modulation on the multi-scale features of the encoder and the upsampled features from the previous level decoder to generate semantic alignment features. The boundary recovery module is used to perform multi-scale feature extraction, gradient enhancement, and attention-guided fine-tuning on the semantic alignment features from the semantic alignment module and the upsampled features from the previous level decoder to generate refined features.

5. The breast ultrasound image segmentation method based on decoupling learning as described in claim 4, characterized in that, The specific steps for generating semantic alignment features are as follows: By using convolution and batch normalization, the multi-scale features of the encoder and the upsampled features from the previous level decoder are projected into a unified embedding space. By adding elements one by one, the two projected features are merged; The fused features are then subjected to average pooling, convolution, and upsampling to obtain new features. A residual attention mechanism is adopted to adaptively reweight new features based on the semantic ambiguity of each spatial region to generate an attention map; The semantic alignment features are obtained by modulating the multi-scale features of the encoder using attention maps.

6. The breast ultrasound image segmentation method based on decoupling learning as described in claim 4, characterized in that, The specific steps for generating refined features are as follows: The semantic alignment features and the upsampled features from the previous level decoder are fused together by adding them element by element. Two parallel convolutional branches are used to convolve the fused features, and the convolutional results are aggregated. High-frequency gradient information is separated by subtracting the local average version of the feature map; A spatial attention mechanism is introduced to guide feature responses to focus on information-rich boundary regions; The channel attention map is computed using spatial average pooling and max pooling, and then connected to a shared transformation function. Refined features are obtained by combining adaptive scale fusion with attention-guided refinement processes.

7. A breast ultrasound image segmentation system based on decoupling learning, characterized in that, include: The image acquisition module is configured to acquire breast ultrasound images to be segmented; The lesion segmentation module is configured to: input breast ultrasound images into a decoupled learning segmentation model, generate a prediction probability map, and map the prediction probability map to the final lesion segmentation result; The segmentation model is built based on U-Net, and the decoupled learning adopts a two-stage training method to train the encoder and decoder of the segmentation model separately. In the first stage of encoder training, fuzzy activation based on region recombination is used to dynamically replace high-confidence regions with fuzzy blocks. Gradient flow is enhanced by perturbing the stable context, generating structural conflicts and semantic variability in high-confidence regions, forcing the encoder to learn semantic invariance under context changes.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the breast ultrasound image segmentation method based on decoupling learning as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a breast ultrasound image segmentation method based on decoupling learning as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a breast ultrasound image segmentation method based on decoupling learning as described in any one of claims 1-6.

Images