A breast ultrasound image segmentation method and system based on decoupled learning

Abstract

The application provides a breast ultrasound image segmentation method and system based on decoupled learning, relates to the technical field of breast ultrasound image processing and analysis, and comprises the following steps: acquiring a breast ultrasound image to be segmented; inputting the breast ultrasound image into a decoupled learning segmentation model to generate a prediction probability map, and mapping the prediction probability map into a final lesion segmentation result; the decoupled learning adopts a two-stage training method to train an encoder and a decoder of the segmentation model respectively; in the first-stage training process, the ambiguity activation realized based on region reorganization is used to dynamically replace a high-confidence region with a fuzzy block, the gradient flow is enhanced by disturbing the stable context, structure conflicts and semantic variability are generated in the high-confidence region, and the encoder is forced to learn semantic invariance under the context change. The application addresses the challenges of the current breast ultrasound image segmentation method in the optimization deficiency in processing fuzzy regions, feature representation fragmentation and boundary quality decline, so as to improve the segmentation precision and robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of breast ultrasound image processing and analysis, and particularly relates to a breast ultrasound image segmentation method and system based on decoupled learning. BACKGROUND

[0002] In the field of breast ultrasound image processing, segmentation of breast ultrasound images (BUS) is crucial for early diagnosis and treatment planning of breast cancer. Accurate lesion segmentation can help doctors quickly identify lesion areas and their boundaries, improving diagnostic efficiency. However, due to the presence of fuzzy regions in breast ultrasound images, these regions are often caused by blurred boundaries, degraded texture, or high similarity in pixel intensity distribution between lesions and surrounding tissues, traditional segmentation methods face significant challenges in handling such regions. These fuzzy regions often produce an uncertainty probability close to 0.5 in prediction, resulting in weak gradient signals and becoming the main bottleneck for segmentation accuracy.

[0003] The goal of lesion segmentation is to accurately separate the lesion region from the input breast ultrasound image. Given the input image , traditional methods usually predict the lesion mask through pixel-level classification or region segmentation. The optimization problem can be expressed as: where is the loss function, often using cross-entropy loss or Dice loss, by minimizing the difference between the predicted mask and the true mask to optimize the model parameters. However, traditional methods such as threshold-based segmentation or manual feature extraction techniques are difficult to effectively deal with the high uncertainty of fuzzy regions and complex background interference. These methods usually assume that image regions have clear boundaries, ignoring the weak gradient problem of fuzzy regions, resulting in fragmented representation or decreased boundary quality near the lesion boundary in the segmentation results.

[0004] With the development of deep learning technology, segmentation methods based on convolutional neural networks (CNN) have significantly improved the performance of breast ultrasound image processing. U-Net and its variants (such as U-Net++, Attention U-Net) enhance multi-scale feature extraction capability through encoder-decoder structure and skip connection, and have better adaptability to fuzzy boundaries than traditional methods. Recent studies attempt to handle fuzzy areas through loss reweighting, dynamic sampling or contrastive learning. For example, Focal Loss improves model attention to fuzzy areas by weighting difficult samples, SegNet assists upsampling by recording pooling indexes, and DeepLab uses dilated convolution to preserve spatial details. However, these methods still have limitations: they mainly improve the visibility of fuzzy areas through passive methods such as weighting or sampling, but fail to truly activate these areas, resulting in insufficient learning of fuzzy areas during optimization, which manifests as semantic gaps and inaccurate boundaries.

[0005] To solve the above problems, some studies have proposed improvement schemes. For example, some methods model the uncertainty of fuzzy areas through dedicated branches or contrastive 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, the weak gradient signal of fuzzy areas leads to insufficient optimization, and the model tends to focus on high-confidence areas, ignoring the fine-grained structure of fuzzy areas; on the other hand, the semantic misalignment of deep-shallow features and inconsistent multi-scale feature transmission are not fully addressed. Therefore, existing technologies cannot actively activate fuzzy areas and optimize their semantic consistency and boundary integrity, resulting in insufficient accuracy and robustness of breast ultrasound image segmentation. SUMMARY

[0006] To solve the above problems, the present application proposes a breast ultrasound image segmentation method and system based on decoupled learning, which addresses the challenges of current breast ultrasound image segmentation methods in handling fuzzy areas, optimizing insufficient features, and declining boundary quality, to improve segmentation accuracy and robustness.

[0007] According to some embodiments, the present application adopts the following technical solutions:

[0008] A breast ultrasound image segmentation method based on decoupled learning, comprising:

[0009] Obtaining a breast ultrasound image to be segmented;

[0010] Inputting the breast ultrasound image into a decoupled learning segmentation model to generate a predicted probability map, and mapping the predicted probability map to a final lesion segmentation result;

[0011] The segmentation model is constructed based on U-Net, and the decoupled learning is a two-stage training method, and the encoder and the decoder of the segmentation model are trained respectively.

[0012] According to some embodiments, the present application adopts the technical scheme as follows:

[0013] A breast ultrasound image segmentation system based on decoupled learning comprises:

[0014] An image acquisition module is configured to acquire a breast ultrasound image to be segmented.

[0015] A lesion segmentation module is configured to input the breast ultrasound image into a decoupled learning segmentation model, generate a predicted probability map, and map the predicted probability map into a final lesion segmentation result.

[0016] The segmentation model is constructed based on U-Net, and the decoupled learning is a two-stage training method, and the encoder and the decoder of the segmentation model are trained respectively.

[0017] According to some embodiments, the present application adopts the technical scheme as follows:

[0018] A computer program product comprises a computer program, and the computer program is executed by a processor to implement the breast ultrasound image segmentation method based on decoupled learning.

[0019] According to some embodiments, the present application adopts the technical scheme as follows:

[0020] A non-transitory computer-readable storage medium is used to store computer instructions, and the computer instructions are executed by a processor to implement the breast ultrasound image segmentation method based on decoupled learning.

[0021] According to some embodiments, the present application adopts the technical scheme as follows:

[0022] An electronic device comprises a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the breast ultrasound image segmentation method based on decoupled learning.

[0023] Compared with the prior art, the beneficial effects of the present application are:

[0024] To further improve the segmentation accuracy, the present application adopts the method of decoupled learning to perform two-stage training on the segmentation model. In the first stage, AAR (Ambiguity Activation Module) dynamically replaces the originally high-predicted confidence area with ambiguous image blocks during the training process, and enhances the gradient flow by disturbing the stable context. In the second stage, the encoder is frozen to retain the ambiguity-sensitive features, while the decoder performs structure recovery through two modules. Among them, the semantic alignment of perceptual ambiguity (ASA) module ensures the consistency of deep and shallow features, while the boundary recovery of focused ambiguity (ABR) module enhances the boundary details through multi-scale filtering. Experimental results show that the segmentation method proposed in the present application has good performance and shows significant improvement in the segmentation of ambiguous areas. BRIEF DESCRIPTION OF DRAWINGS

[0025] The drawings accompanying the specification of the present application form a part thereof and serve to provide further understanding of the present application, the illustrative embodiments of the present application and its description serve to explain the present application, and do not constitute an improper limitation of the present application.

[0026] Figure 1 The method flowchart of example 1.

[0027] Figure 2 The structure schematic diagram of the segmentation model in example 1.

[0028] Figure 3 The ASA structure schematic diagram in example 1.

[0029] Figure 4 The ABR structure schematic diagram in example 1. DETAILED DESCRIPTION

[0030] The present application will be further described below in conjunction with the drawings and examples.

[0031] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0032] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0033] Embodiment 1

[0034] In an embodiment of the present application, a breast ultrasound image segmentation method based on decoupled learning is provided, comprising:

[0035] Step S1: obtaining a breast ultrasound image to be segmented;

[0036] Step S2: inputting the breast ultrasound image into a decoupled learning segmentation model to generate a prediction probability map, and mapping the prediction probability map into a final lesion segmentation result;

[0037] The segmentation model is constructed based on U-Net, and the decoupled learning is implemented by using a two-stage training method to train the encoder and the decoder of the segmentation model respectively. In the first stage of training the encoder, the ambiguity activation is realized based on region reorganization, the high-confidence region is dynamically replaced with a fuzzy block, the gradient flow is enhanced by disturbing the stable context, the structural conflict and semantic variability are generated in the high-confidence region, and the encoder is forced to learn semantic invariance under the change of context.

[0038] As an embodiment, the breast ultrasound image segmentation method based on decoupled learning of the present application addresses the challenges of current breast ultrasound image segmentation methods in handling the optimization deficiency of fuzzy regions, feature representation fragmentation, and boundary quality decline, so as to improve the segmentation accuracy and robustness, as shown in Figure 1 The specific implementation process is as follows:

[0039] Step 1: obtaining and preprocessing the breast ultrasound image dataset, specifically:

[0040] Step 1.1: obtaining the public breast ultrasound image datasets BUSI-E9, Dataset B and STU as the training dataset, and preliminarily processing the training dataset, reading the data in the order of image acquisition, and dividing the original training dataset and the original test dataset .

[0041] Step 1.2: preprocessing the training dataset:

[0042] The original training dataset and the original test dataset The standardization is performed to obtain a pre-processed training data set and a pre-processed test data set .

[0043] Step 2: constructing an encoder of the segmentation model and performing first-stage training, specifically:

[0044] Step 2.1: constructing an encoder as shown in Figure 2 , using a pre-trained RexNet as the encoder to generate multi-scale feature maps;

[0045] Step 2.1.1: selecting RexNet as the encoder base network, which obtains initial parameters by pre-training on the ImageNet dataset, effectively solving the cold start problem of the network model;

[0046] Step 2.1.2: inputting the pre-processed training data set to the RexNet to extract features through different stages of the network;

[0047] Step 2.1.3: setting five down-sampling stages in the encoder, i.e. Figure 2 Step1-5 in , , , and , which generate multi-scale feature maps ,

[0048] , ,

[0049] and , respectively, with resolutions decreasing to 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image, respectively, and channel numbers being 32, 38, 61, 128 and 1280, respectively;

[0050] Step 2.2: in order to perform first-stage training, a segmentation head is constructed, and the structure of the segmentation head is as simple as possible to avoid providing additional structures with feature extraction capability outside the encoder module, so as to optimize the performance of the encoder as much as possible, therefore the segmentation head only contains a convolution layer;

[0051] The fuzzy regions usually produce uncertain predictions close to 0.5, resulting in weak gradients that hinder the model learning. To actively activate these regions, an uncertainty-guided region reorganization mechanism is introduced during the training process, which enhances the activation of fuzzy regions by explicitly perturbing the spatial context. The goal is to generate structural conflicts and semantic variability in high-confidence regions, forcing the model to learn semantic invariance under context variations.

[0052] Step 2.4.1: denotes the input images of the training batch, denotes the corresponding ground truth segmentation masks, denotes the predicted probability maps, which are the prediction results obtained by inputting the images of the batch into the above steps 2.1-2.3.

[0053] Step 2.4.2: For each predicted probability map , its uncertainty map is calculated as the basis for dynamically locating the fuzzy regions in the training batch:

[0054]

[0055] where, quantifies the uncertainty of the prediction, reaching a maximum value when the prediction is most uncertain.

[0056] Step 2.4.3: To extract clear foreground regions from the source image , a threshold is used to threshold the mask :

[0057]

[0058] This determines the spatial extent of the object of interest (foreground) .

[0059] Step 2.4.4: Next, the minimum bounding box of the foreground region is calculated:

[0060]

[0061] defines the foreground patches for reorganization .

[0062] Before cropping, a series of transformations are applied to to introduce context variations:

[0063]

[0064]

[0065]

[0066] where, is the enhanced foreground block, which is used as the blur block to be pasted, , and is sampled from a predefined range to ensure the diversity of enhancement.

[0067] To dynamically determine the target image during the training process, a suitable pasting area is selected based on the uncertainty map A sliding window search is performed on the uncertainty map to select the area with the lowest average uncertainty (i.e., high confidence area), where the window size matches the size of the enhanced foreground block , which can be expressed as:

[0068]

[0069] where, is the starting position of the area.

[0070] The blur block is pasted to the indicated high confidence area introduces controllable semantic contrast, which helps to distinguish structures, but the overlap with the target foreground may cause structural interference. To avoid this situation, the overlapping pixels are removed:

[0071]

[0072] An intra-sample region replacement strategy is adopted to insert the blur block into the high confidence area of the same image to simulate semantic conflict, which can be expressed as:

[0073]

[0074] where, is the smallest rectangular frame containing the foreground pixels; is a binary mask indicating which pixels belong to the non-overlapping foreground area to be pasted; is the starting coordinate of the best pasting position, determined by the sliding window search, aiming to find the area with the lowest average uncertainty in the target image. ; here, + is not arithmetic addition, but indicates that the mask is placed at the position with as the top-left corner; is the scaling factor, which determines the size variation ratio of the foreground block; is the rotation angle; is the contrast adjustment coefficient; is the foreground threshold used to distinguish foreground and background. It is an uncertainty graph. The larger the value of a pixel, the more uncertain the model is about classifying that pixel.

[0075] 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.

[0076] Step 3: Construct the decoder for the segmentation model and perform the second stage of training;

[0077] 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.

[0078] 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.

[0079] Step 3.3: Construct the decoder, introducing the semantic alignment module for perceiving ambiguity (ASA) and the boundary recovery module for focusing ambiguity (ABR).

[0080] 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.

[0081] 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. .

[0082] Step 3.3.3: At each upsampling stage, the perceptual ambiguity-aware semantic alignment module (ASA) and the ambiguity-aware boundary restoration module (ABR) are cascaded to gradually restore spatial details and output the segmentation mask;

[0083] Step 3.4: The perceptual ambiguity-aware semantic alignment module (ASA) 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), such fusion can easily lead to semantic misalignment, i.e., high-level semantics cannot accurately guide the restoration of low-level details, and low-level noise interferes with the expression of high-level semantics. To this end, as shown in FIG. 4, the specific steps of ASA are as follows: Figure 3

[0084] The first-layer ASA module receives the encoder feature and the semantic feature from the previous ABR module. To align their resolutions, the is first upsampled to . Subsequently, and are projected to a unified embedding space through 3x3 convolution and batch normalization (BN) to eliminate channel mismatch problems:

[0085]

[0086]

[0087] Then, the two transformed features are fused by element-wise addition:

[0088]

[0089] The resulting feature map combines local structural cues and global semantic information. However, its limited receptive field limits the ability to model context, which is particularly important for ambiguous regions. To this end, a 2x2 average pooling followed by a 3x3 convolution is applied. The processed feature is then upsampled back to the original resolution:

[0090]

[0091]

[0092] To further optimize semantic alignment, a residual attention mechanism is used to adaptively reweight the semantic ambiguity of each spatial region:

[0093] ​​

[0094] where denotes a sigmoid function. The generated attention map highlights regions that are semantically inconsistent or under-representative, this perceptionally ambiguous attention map is used to modulate the shallow encoder features:

[0095]

[0096] By suppressing overconfident background responses while simultaneously boosting responses in ambiguous regions, the optimized features f i improve the representational fidelity of uncertain regions. This output is subsequently passed to subsequent modules for further structural recovery.

[0097] Step 3.5: The Ambiguity-Focused Boundary Recovery module (ABR), which progressively enhances boundary completeness and recovers fine-grained structural cues, especially for the boundaries of uncertain regions, such as Figure 4 as shown, specifically:

[0098] semantic alignment features from the ASA and upsampled results from the previous decoder output . Both are fused through element-wise addition:

[0099]

[0100] Subsequently, a 3x3 convolution is applied to obtain a richer representation:

[0101]

[0102] To capture boundary structures over multiple receptive field ranges, two parallel convolution branches are employed:

[0103]

[0104]

[0105] where, focus on local contour details, while capture more extensive contextual features. These two outputs are aggregated to form a unified representation:

[0106]

[0107] To adaptively balance the contributions of different receptive fields, we compute a soft fusion map:

[0108]

[0109]

[0110] Blur areas usually exhibit lower edge contrast. To this end, we separate the high-frequency gradient information by subtracting the local average version of the feature map:

[0111]

[0112]

[0113] The resulting residual term highlights the structural gradients lost in the early fusion process.

[0114] We then introduce a spatial attention mechanism to guide the feature responses to focus on the boundary regions with information quantity:

[0115]

[0116]

[0117] To further optimize the representation, we compute the channel attention map by spatial average pooling and max pooling, and concatenate a shared transformation function:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] By combining the adaptive scale fusion with the attention-guided refinement process, we obtain the output of the ABR module:

[0124]

[0125] To enhance the discriminative ability of the multi-level features in the decoder, we introduce an auxiliary supervision branch after the output of each ABR module. Specifically, we first compress the output of each ABR module by a 1x1 convolution and then upsample it to the original image size to obtain an auxiliary prediction map which is used to calculate the loss with the ground truth, denoted as:

[0126]

[0127] Each auxiliary prediction are calculated with the real mask M:

[0128]

[0129] wherein, , are Binary Cross-Entropy Loss and Dice Loss respectively, which are used to measure the difference between the predicted mask and the real mask, and the specific definitions are as follows:

[0130]

[0131]

[0132] wherein, N represents the total number of pixels, and represent the values of the th auxiliary prediction and the real mask at the th pixel position, is a smoothing term, which is used to avoid the denominator being zero.

[0133] Finally, the total loss of Stage II gives different weights to different levels, emphasizing deep semantic output while retaining shallow detail guidance:

[0134]

[0135] Step 4: Use the trained model to make lesion segmentation prediction;

[0136] Step 4.1: Prepare input training data , input sequence , is the batch size, indicating the number of samples input at a 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 point.

[0137] Step 4.2: Use the encoder part of the trained model to extract multi-scale features from the input data. Input the input tensor X into the model trained in step 3, first extract features through the encoder; the encoder generates multi-scale feature maps ( ) through layer-by-layer downsampling, wherein , , the number of channels are 32, 38, 61, 128 and 1280 respectively;

[0138] Step 4.3: Process the feature maps and generate the prediction probability map by semantic alignment of perceptual ambiguity (ASA) and boundary recovery of empty focus ambiguity (ABR) in the decoder; process the feature maps output by the encoder passed to the decoder through the skip connection, and the decoder is upsampled layer by layer through transposed convolution;

[0139] In the ASA module, the encoder features and the upsampled features from the previous layer are processed to perform cross-level feature alignment, context enhancement and ambiguity perception modulation to generate semantic alignment features .

[0140] In the ABR module, the features from the ASA and the upsampled output of the previous layer decoder are processed to perform multi-scale feature extraction, gradient enhancement and attention-guided refinement to generate refined features .

[0141] Step 4.4: Map the prediction probability map to the final lesion segmentation result to generate the segmentation mask ; post-process the prediction probability map generated in step 4.3 to generate a binary segmentation mask using a thresholding operation , the process is represented as: , if , then . (lesion area), otherwise (background area); output the final segmentation mask , representing the lesion area of each input image, for subsequent evaluation or clinical application.

[0142] Table 1 and Table 2 are the results of comparing different representative segmentation algorithms on BUSI-E9 and Dataset B datasets, which fully demonstrate the superiority of the image segmentation task of the method of the embodiment.

[0143] Table 1 BUSI-E9 dataset comparison results

[0144]

[0145] Table 2 Dataset B dataset comparison results

[0146]

[0147] Embodiment 2

[0148] In an embodiment of the present application, a breast ultrasound image segmentation system based on decoupled learning is provided, comprising:

[0149] an image acquisition module configured to acquire a breast ultrasound image to be segmented;

[0150] a lesion segmentation module configured to input the breast ultrasound image into a decoupled learning segmentation model, generate a predicted probability map, and map the predicted probability map into a final lesion segmentation result;

[0151] The segmentation model is constructed based on a U-Net, and the decoupled learning is implemented by using a two-stage training method to respectively train an encoder and a decoder of the segmentation model. In a first-stage training process of the encoder, ambiguity activation is realized based on region recombination, a high-confidence region is dynamically replaced with a fuzzy block, gradient flow is enhanced by disturbing a stable context, structural conflicts and semantic variability are generated in the high-confidence region, and the encoder is forced to learn semantic invariance under context changes.

[0152] Embodiment 3

[0153] In an embodiment of the present application, a computer program product is provided, which comprises a computer program. When the computer program is executed by a processor, the computer program implements the breast ultrasound image segmentation method based on decoupled learning.

[0154] Embodiment 4

[0155] In an embodiment of the present application, a non-transitory computer-readable storage medium is provided, which is used to store computer instructions. When the computer instructions are executed by a processor, the computer instructions implement the breast ultrasound image segmentation method based on decoupled learning.

[0156] Embodiment 5

[0157] In an embodiment of the present application, an electronic device is provided, which comprises a processor, a memory, and a computer program. The processor is connected with the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device implements the breast ultrasound image segmentation method based on decoupled learning.

[0158] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the computer or other programmable data processing device produce a device that implements the functions described in the flowcharts and / or block diagrams. Figure 1apparatuses that implement functionalities described in one or more flows and / or blocks. Figure 1

[0159] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable data processing devices provide a process for implementing the flows Figure 1 apparatuses that implement functionalities described in one or more flows and / or blocks. Figure 1 steps of apparatuses that implement functionalities described in one or more flows and / or blocks.

[0160] The above description is made in connection with the preferred embodiments of the application. It is not intended to limit the scope of the application. Various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the application.​

Claims

1. A method for breast ultrasound image segmentation based on decoupled learning, characterized in that, The method comprises the following steps: acquiring a breast ultrasound image to be segmented; inputting the breast ultrasound image into a decoupled learning segmentation model to generate a predicted probability map, and mapping the predicted probability map to a final lesion segmentation result; wherein the segmentation model is constructed based on a U-Net, and the decoupled learning is a two-stage training method, and the encoder and the decoder of the segmentation model are trained respectively; in the first stage of training the encoder, a blur activation realized by region reorganization is used to dynamically replace a high-confidence region with a blur block, to enhance gradient flow by disturbing a stable context, to generate structural conflicts and semantic variability in the high-confidence region, and to force the encoder to learn semantic invariance under context changes; the blur activation realized by region reorganization is used to construct new first-stage training samples, specifically: acquiring original training samples, including a breast ultrasound image, a corresponding real segmentation mask and a predicted probability map; for each predicted probability map, calculating an uncertainty map; segmenting a foreground region from the breast ultrasound image to construct a blur block; performing a sliding window search on the uncertainty map to select a region with the lowest average uncertainty as a high-confidence region; and inserting the blur block into the high-confidence region of the same breast ultrasound image to obtain new training samples; the decoder comprises a semantic alignment module for perception ambiguity and a boundary recovery module for focusing ambiguity; the semantic alignment module is used to perform cross-level feature alignment, context enhancement and ambiguity perception modulation on multi-scale features of the encoder and up-sampling features from an upper-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 processing on the semantic alignment features from the semantic alignment module and the up-sampling features from the upper-level decoder to generate fine features; the generation of the semantic alignment features specifically comprises the following steps: projecting the multi-scale features of the encoder and the up-sampling features from the upper-level decoder into a unified embedding space through convolution and batch normalization; fusing the two projected features through element-wise addition; performing average pooling, convolution and up-sampling on the fused features to obtain new features; generating an attention map by adaptively re-weighting the new features according to the semantic ambiguity of each spatial region by using a residual attention mechanism; and modulating the multi-scale features of the encoder with the attention map to obtain the semantic alignment features; the generation of the fine features specifically comprises the following steps: fusing the semantic alignment features and the up-sampling features from the upper-level decoder through element-wise addition; performing convolution on the fused features by using two parallel convolution branches, and aggregating the convolution results; separating high-frequency gradient information by subtracting the local average version of the feature map; introducing a spatial attention mechanism to guide the feature response to focus on the boundary regions with information quantity; calculating a channel attention map through spatial average pooling and maximum pooling, and connecting a shared transformation function; and obtaining the fine features by combining the adaptive scale fusion and the attention-guided fine processing.

2. The decoupled learning based breast ultrasound image segmentation method of claim 1, wherein, the segmentation model comprises an encoder and a decoder. The encoder performs multi-level down-sampling on the input breast ultrasound image to generate multi-scale features; The decoder performs multi-level up-sampling on the multi-scale features generated by the encoder to generate a prediction probability map.

3. A decoupled learning based breast ultrasound image segmentation system, characterized by, It comprises: An image acquisition module configured to acquire a breast ultrasound image to be segmented; A lesion segmentation module configured to input the breast ultrasound image into a decoupled learning segmentation model to generate a prediction probability map, and map the prediction probability map to a final lesion segmentation result; The segmentation model is constructed based on U-Net, and the decoupled learning is performed by using a two-stage training method to train the encoder and the decoder of the segmentation model respectively. In the first stage of training the encoder, the ambiguity activation is realized based on region reorganization, the high-confidence region is dynamically replaced with a fuzzy block, the gradient flow is enhanced by disturbing the stable context, the structural conflict and semantic variability are generated in the high-confidence region, and the encoder is forced to learn semantic invariance under the change of context. The region reorganization-based ambiguity activation is used to construct new first-stage training samples, specifically: obtaining original training samples, including breast ultrasound images, corresponding real segmentation masks and prediction probability maps; for each prediction probability map, calculating an uncertainty map; segmenting the foreground region from the breast ultrasound image to construct a fuzzy block; performing sliding window search on the uncertainty map to select the region with the lowest average uncertainty as the high-confidence region; and inserting the fuzzy block into the high-confidence region of the same breast ultrasound image to obtain new training samples. The decoder comprises a semantic alignment module for perception ambiguity and a boundary recovery module for focusing ambiguity; The semantic alignment module is configured to perform cross-level feature alignment, context enhancement and ambiguity perception modulation on the multi-scale features of the encoder and the up-sampled features from the decoder of the previous level to generate semantic alignment features; The boundary recovery module is configured to perform multi-scale feature extraction, gradient enhancement and attention-guided fine processing on the semantic alignment features from the semantic alignment module and the up-sampled features from the decoder of the previous level to generate fine features; The method for generating semantic alignment features comprises the following steps: projecting the multi-scale features of the encoder and the up-sampled features from the decoder of the previous level into a unified embedding space through convolution and batch normalization; fusing the two projected features by element-wise addition; performing average pooling, convolution and up-sampling on the fused features to obtain new features; using a residual attention mechanism to adaptively re-weight the new features according to the semantic ambiguity of each spatial region to generate an attention map; and modulating the multi-scale features of the encoder with the attention map to obtain semantic alignment features. The generating fine features specifically comprises the following steps: fusing the semantic alignment features and the up-sampling features from the upper-level decoder by element-by-element addition; performing convolution on the fused features by using two parallel convolution branches, and aggregating the convolution results; separating high-frequency gradient information by subtracting a local average version of the feature map; introducing a spatial attention mechanism to guide the feature response to focus on the boundary region with information quantity; calculating a channel attention map by spatial average pooling and maximum pooling, and connecting a shared transformation function; and obtaining fine features by combining the adaptive scale fusion and the attention-guided fine process.

4. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the breast ultrasound image segmentation method based on decoupled learning according to any one of claims 1-2.

5. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium is used to store computer instructions, and the computer instructions are executed by the processor to implement the breast ultrasound image segmentation method based on decoupled learning according to any one of claims 1-2.

6. An electronic device, comprising: Comprise: A processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the breast ultrasound image segmentation method based on decoupled learning according to any one of claims 1-2.

Images