A breast cancer lesion segmentation system based on asymmetric U-shaped network

Abstract

The present invention provides a breast cancer lesion segmentation system based on an asymmetric U-shaped network, comprising: an image acquisition module to be processed, for acquiring an image to be processed; a processing module, for inputting the image to be processed into the asymmetric U-shaped network to obtain a segmentation result; the asymmetric U-shaped network comprises a multi-branch residual encoder, an external attention module and a boundary detection module. In the breast cancer lesion segmentation system of the present invention, an asymmetric U-shaped network design is adopted to improve the accuracy and robustness of lesion area segmentation in breast ultrasound images; and a multi-branch residual encoder based on an attention mechanism is introduced to enhance the representation ability of the network; in addition, a deep supervision boundary detection module is proposed to improve the segmentation accuracy of the lesion boundary, which can more effectively process breast ultrasound images, achieve accurate segmentation of breast cancer lesions, provide doctors with more reliable auxiliary information, and help improve the early diagnosis and treatment of breast cancer.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a breast cancer lesion segmentation system based on an asymmetric U-shaped network. Background Art

[0002] In recent years, convolutional neural networks (CNNs), with their powerful ability to learn image features, have achieved incredible breakthroughs in computer vision. In medical image analysis, CNNs have also seen improved performance across various image analysis tasks. Mainstream breast ultrasound lesion segmentation methods focus on three key aspects: multi-scale feature extraction, attention mechanisms, and deep supervision. Multi-scale dilated convolutions with varying kernel sizes are a common strategy for extracting multi-scale features, such as STAN and RRCNet. These variants refine breast tumor feature information using convolutional kernels at different scales, thereby improving breast tumor segmentation accuracy to varying degrees. With the emergence of attention mechanisms, various networks integrating codecs with attention mechanisms have been developed, such as MSGRAP, AE U-net, and GG-net. While the introduction of different attention modules has improved segmentation performance, the complex computational complexity and susceptibility to interference from various factors inherent in the attention mechanism remain research challenges. The introduction of deep supervision mechanisms can guide the network to learn to optimize segmentation results from coarse to fine scales, mitigating the interference from external factors to a certain extent.

[0003] In the clinical diagnosis process, the requirements for the segmentation accuracy and ease of use of breast ultrasound images are increasing. Therefore, there is an urgent need for an automatic segmentation method for breast cancer ultrasound lesions to assist in the early diagnosis of breast cancer. Summary of the Invention

[0004] One of the objectives of the present invention is to provide a breast cancer lesion segmentation system based on an asymmetric U-shaped network to solve the above-mentioned problem.

[0005] An embodiment of the present invention provides a breast cancer lesion segmentation system based on an asymmetric U-shaped network, comprising:

[0006] The image to be processed acquisition module is used to acquire the image to be processed;

[0007] A processing module is used to input the image to be processed into the asymmetric U-shaped network to obtain the segmentation result;

[0008] Among them, the asymmetric U-shaped network includes: a multi-branch residual encoder, an external attention module and a boundary detection module.

[0009] Preferably, the multi-branch residual encoder is used to extract features from the image to be processed using a preset number of branches, and then fuse each set of features to form a final representation feature.

[0010] Preferably, the multi-branch residual encoder uses a preset number of branches to extract features from the image to be processed, and then fuses each set of features to form a final representation feature, including:

[0011] Construct the feature map corresponding to the image to be processed;

[0012] Extract details from the feature map through multiple branches with shared weights to obtain the output of the model of each branch;

[0013] Based on the feature map and the output of the model of each branch, the representation feature is determined. The calculation formula is as follows:

[0014]

[0015] In the formula, f is the characterization feature, I0 is the feature map, is the output of the model of the g-th branch, and G is the total number of branches.

[0016] Preferably, detail extraction is performed on the feature map through multiple branches with shared weights to obtain the output of the model of each branch, including:

[0017] Each branch performs three convolutions and normalization on the feature map.

[0018] Preferably, the external attention module includes: a key matrix and a value matrix;

[0019] The representation feature is multiplied by the transposed matrix of the key matrix and normalized, and then multiplied by the value matrix to obtain the output feature.

[0020] Preferably, the boundary detection module includes: multi-level boundary detection for outputting lesion boundary prediction maps of different resolutions at different levels of the decoder; the prediction maps at each level serve as supervisory signals to focus on complex boundary contours;

[0021] The multi-level boundary detection module outputs lesion boundary prediction maps of different resolutions at different levels of the decoder and performs the following operations:

[0022] Upsample the preset high-resolution image to obtain a feature map with the same size as the label image;

[0023] Performing a convolution operation to generate a first parameter set containing boundary information of a potential lesion area;

[0024] Pass the first parameter set through the maximum pooling layer to obtain the second parameter set;

[0025] The absolute difference set between the first parameter set and the second parameter set is calculated as a boundary prediction map.

[0026] Preferably, the total loss function used by the asymmetric U-shaped network is as follows:

[0027]

[0028] Among them, N is the total number of CNN layers with side output, Used to measure the similarity between the prediction result of the i-th lateral layer and the true label, and They represent the boundary loss and segmentation loss of the i-th layer respectively, ω1 is the preset first weight, ω2 is the preset second weight, and ω3 is the preset third weight.

[0029] Preferably, before the processing module inputs the image to be processed into the asymmetric U-shaped network to obtain the segmentation result, the processing module further performs the following operations:

[0030] Based on the preset initial feature extraction template, feature extraction is performed on the image to be processed and an image feature set is constructed based on the extracted feature values;

[0031] Matching the image feature set with each standard feature set in the set of standard feature sets corresponding to each asymmetric U-shaped network model in the preset model repository;

[0032] Based on the matching results, the fitness of each asymmetric U-shaped network model is determined;

[0033] Extract the asymmetric U-shaped network with the highest fitness.

[0034] Preferably, the standard feature set associated with the asymmetric U-shaped network model is constructed by the following steps:

[0035] Obtain training data for the asymmetric U-shaped network model;

[0036] Based on the preset initial feature extraction template, feature extraction is performed on each image in the training data and an image feature set corresponding to each image is constructed based on the extracted key feature values;

[0037] Calculate the similarity between the image feature sets corresponding to each image;

[0038] Extracting a preset number of image feature sets before the total similarity to construct a standard feature set set; wherein each image feature set of the standard feature set set is used as the standard feature set;

[0039] The matching value is configured based on the total similarity corresponding to each standard feature set in the standard feature set set.

[0040] Preferably, based on the matching results, determining the fitness of each asymmetric U-shaped network model includes:

[0041] Analyze the matching results and determine the similarity of each standard feature set in the set of standard feature sets corresponding to the asymmetric U-shaped network model;

[0042] Based on the similarity, the adaptation coefficient is determined;

[0043] Determine the degree of fitness based on the fitness coefficient and the matching value corresponding to the standard feature set;

[0044] The degree of adaptation is the sum of the products of the matching values ​​corresponding to each standard feature set and the corresponding adaptation coefficients.

[0045] The present invention has the following advantages:

[0046] 1. An asymmetric U-shaped segmentation network is used to improve the accuracy and robustness of lesion area segmentation in breast ultrasound images.

[0047] 2. An attention-based multi-branch residual encoder is introduced to enhance the representation ability of the network. In addition, a deep supervised boundary detection module is proposed to improve the segmentation accuracy of lesion boundaries.

[0048] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0049] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0051] Figure 1 Schematic diagram of a breast cancer lesion segmentation system based on an asymmetric U-shaped network in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the model architecture in an embodiment of the present invention;

[0053] Figure 3 Schematic diagram of a multi-branch residual unit in an embodiment of the present invention;

[0054] Figure 4 Schematic diagram of an external attention mechanism in an embodiment of the present invention;

[0055] Figure 5 Schematic diagram of a single boundary detection module in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0057] The embodiment of the present invention provides a breast cancer lesion segmentation system based on an asymmetric U-shaped network. Figure 1 As shown, including:

[0058] The image to be processed acquisition module 1 is used to acquire the image to be processed;

[0059] Processing module 2, used to input the image to be processed into the asymmetric U-shaped network to obtain the segmentation result;

[0060] Among them, the asymmetric U-shaped network includes: a multi-branch residual encoder, an external attention module and a boundary detection module.

[0061] Among them, the multi-branch residual encoder is used to extract features from the image to be processed using a preset number of branches, and then fuse each set of features to form a final representation feature.

[0062] Among them, the multi-branch residual encoder uses a preset number of branches to extract features from the image to be processed, and then fuses each set of features to form the final representation feature, including:

[0063] Construct the feature map corresponding to the image to be processed;

[0064] The feature map is processed through multiple identical branches to obtain the output of the model of each branch;

[0065] Based on the feature map and the output of the model of each branch, the representation feature is determined. The calculation formula is as follows:

[0066]

[0067] In the formula, f is the characterization feature, I0 is the feature map, is the output of the model of the g-th branch, and G is the total number of branches.

[0068] The feature map is processed by multiple identical branches to obtain the output of the model of each branch, including:

[0069] Each branch performs three convolutions and normalization on the feature map.

[0070] Among them, the external attention module includes: key matrix and value matrix;

[0071] The representation feature is multiplied by the transposed matrix of the key matrix and normalized, and then multiplied by the value matrix to obtain the output feature.

[0072] The boundary detection module includes: a deep supervised BD module, which is used to output multi-scale BD prediction maps at different levels of the decoder; the BD prediction maps correspond to feature maps of different resolutions;

[0073] The deep supervised BD module outputs multi-scale BD prediction maps at different levels of the decoder and performs the following operations:

[0074] Upsample the preset high-resolution image to obtain a feature map with the same size as the label image;

[0075] Performing a convolution operation to generate a first parameter set containing boundary information of a potential lesion area;

[0076] Pass the first parameter set through the maximum pooling layer to obtain the second parameter set;

[0077] The absolute difference set between the first parameter set and the second parameter set is calculated as the BD prediction map.

[0078] Among them, the total loss function used by the asymmetric U-shaped network is as follows:

[0079]

[0080] Among them, N is the total number of CNN layers with side output, Used to measure the similarity between the prediction result of the i-th lateral layer and the true label, and They represent the boundary loss and segmentation loss of the i-th layer respectively, ω1 is the preset first weight, ω2 is the preset second weight, and ω3 is the preset third weight.

[0081] The working principle and beneficial effects of the above technical solution are:

[0082] The encoder-decoder architecture has excellent performance in biomedical segmentation tasks and is therefore widely used in breast lesion segmentation. Based on this, this application proposes an asymmetric U-shaped network for segmenting breast lesions from ultrasound images. The network architecture is as follows: Figure 2 As shown in the figure, we first customize a multi-branch residual encoder based on the classic UNet model as our backbone network. We then introduce an external attention module during the decoder upsampling to establish long-range dependencies in the feature maps. Finally, we use a boundary detection module to deepen the extraction of feature map boundaries, and use the feature maps output by the shallow network to guide deep supervision.

[0083] UNet-like networks and their variants often adopt deeper network architectures or modify the convolution kernel scale to capture cancer characteristics under the influence of different receptive fields. Notably, these networks tend to use two-layer convolution, batch normalization, ReLU activation, and downsampling operations in the encoder path. However, for breast ultrasound images, where the surrounding tissue is extremely similar and the boundaries are vaguely defined, such simple architectures often fail to fully extract and characterize the detailed features of breast lesions, resulting in suboptimal ultrasound image segmentation accuracy.

[0084] In order to reduce the loss of small features and edge information caused by downsampling in the encoder part and increase the possibility of network calculation, this application uses a multi-branch residual unit architecture. Figure 3 As shown in the figure, this unit contains 32 identical branches for enriching feature information, all sharing the same weights. Each branch refines lesion features through three convolutions and normalization to capture potential information at different scales. Each feature set is fused and skip connections are used to reduce the possibility of gradient vanishing. The final representation is activated by Reunit (Reinforced Unit). The mathematical expression of feature f is as follows:

[0085]

[0086] Where I0 is the feature map of the input unit, Represents the output models of different branches, G=32.

[0087] The encoder layer in this architecture has channels of [64, 256, 512, 1024, 2048] per layer. Skip connections are used to pass the feature information extracted at each level to the decoder, enabling the decoder to utilize these features to better restore image details and achieve accurate segmentation. Overall, the decoder demonstrates excellent image feature learning capabilities, capturing the complex patterns in breast ultrasound images through its multi-branch structure of convolutional operations. Rich high-level features can be extracted from the original image to better handle the diverse shapes and textures of breast lesions. Furthermore, robust initialization through pretrained weights accelerates network convergence, enabling the network to achieve satisfactory performance in a relatively short period of time.

[0088] Unlike the traditional symmetric U-shaped network that uses downsampling to compress feature maps, a convolution operation with a stride of 2 is used in the encoder to better preserve feature details. In order to further strengthen the focus on the lesion area, this paper introduces an external attention (EA) module to enhance the network's global context understanding of breast images and capture long-distance dependencies, fully considering the relationship between each pixel in the feature map. The EA module can learn the correlation of spatial positions during training, allowing the network to focus on lesion features, suppress irrelevant information such as background, and enhance the accuracy of network detection. The EA module is as follows: Figure 4 As shown:

[0089] This module can be implemented using only two linear layers and a learnable external matrix. It is represented as:

[0090]

[0091] Among them, F is the input module feature, M K With M V are learnable parameters that are independent of the input and act as memory units for the entire training dataset, serving as keys and values, respectively. K The transposed matrix is ​​multiplied and normalized, and then multiplied by M V Multiply to get the processed output feature map F out .

[0092] When observing lesion segmentation maps generated using various CNNs, we found that typical networks struggled to accurately delineate lesion boundaries in complex ultrasound images. To alleviate this problem, this application employed a deep supervised boundary detection module to improve the accuracy and stability of breast lesion segmentation. The design of this module is based on the following key ideas: through multi-scale boundary prediction and edge extraction, with corresponding rewards and penalties, the network can better capture breast lesion boundary information.

[0093] Specifically, multi-scale boundary prediction maps are output at different levels of the decoder. These prediction maps correspond to feature maps of different resolutions and can capture boundary information of different scales. Figure 5 As shown in Figure 1, the high-resolution image is first upsampled to obtain a feature map of the same size as the label image. Then, a 1×1 convolution operation is used to generate D containing the boundary information of the potential lesion area. * To reduce noise and discontinuities in the boundary map, a max pooling layer is used to obtain a smoother version. And calculate D * and The absolute difference between To emphasize the detailed feature information of the boundary. Finally, the output Expressed as:

[0094]

[0095] The boundary prediction maps generated by different levels of the decoder increase the learning opportunities of the network, and the supervisory signal prompts the network to better learn boundary features, thereby coping with the complex boundary contours of breast ultrasound images and improving the accuracy of segmentation.

[0096] Four boundary detection modules are added to the decoder to detect boundaries and jointly locate breast lesions. Therefore, this network generates four boundary contours D' abs And four breast lesion segmentation results D o ' ut The total loss of the network is L total express:

[0097]

[0098] Where N is the total number of CNN layers with side outputs (i.e., N=4), Used to measure the similarity between the prediction result of the i-th lateral layer and the true label, and They represent the boundary loss and segmentation loss of the i-th layer respectively. The weight ω is used to balance the relationship between them. In practice, ω1 = ω3 = 1 and ω2 = 5 can be set. It can be expressed as Equation 5:

[0099]

[0100] where N p Indicates the total number of pixels in the sample image, y i,j is the true value of the jth pixel in the i-th layer of the label, P i,j is the predicted probability of the pixel.

[0101] Used to encourage the model to produce clean and sharp boundaries. represents the j-th pixel value of the boundary map generated by the network, D i,j Is its label value. Mean square error expression:

[0102]

[0103] It is used to further evaluate the segmentation loss and is expressed as:

[0104]

[0105] In practical applications, the model must first be trained. The experiment used stochastic gradient descent to optimize the entire network, with a minimum batch size of 4, weight decay of 0.0001, and momentum of 0.9. Each method was trained for 100 epochs. The initial learning rate was set to 0.001. Data augmentation was performed on the training set using random rotations and horizontal flipping. The network was implemented in PyTorch 1.13. The development environment was Ubuntu 18.04 and Python 3.7, equipped with two NVIDIA RTX2080 GPUs.

[0106] Before the processing module inputs the image to be processed into the asymmetric U-shaped network to obtain the segmentation result, the processing module also performs the following operations:

[0107] Based on a preset initial feature extraction template, feature extraction is performed on the image to be processed and an image feature set is constructed based on the extracted feature values; the features extracted by the initial feature extraction template include: feature values ​​representing the distribution of pixels in each preset pixel value interval, feature values ​​representing the total number of contours in the image, feature values ​​representing the sizes of the top N contour areas sorted by size, etc.

[0108] Matching the image feature set with each standard feature set in the set of standard feature sets corresponding to each asymmetric U-shaped network model in the preset model repository; matching the image feature set with the standard feature set is actually calculating the similarity between the image to be processed and the image during the training of the asymmetric U-shaped network;

[0109] Based on the matching results, the fitness of each asymmetric U-shaped network model is determined;

[0110] Extract the asymmetric U-shaped network with the highest fitness.

[0111] Among them, the standard feature set corresponding to the asymmetric U-shaped network model is constructed through the following steps:

[0112] Obtain training data for the asymmetric U-shaped network model;

[0113] Based on the preset initial feature extraction template, feature extraction is performed on each image in the training data and an image feature set corresponding to each image is constructed based on the extracted key feature values;

[0114] Calculate the similarity between the image feature sets corresponding to each image;

[0115] Extracting a preset number of image feature sets before the total similarity to construct a standard feature set set; wherein each image feature set of the standard feature set set is used as the standard feature set;

[0116] The matching value is configured based on the total similarity of each standard feature set within the standard feature set. For example, if the total similarity is the sum of the individual similarities, and the matching value is configured as 1000, the matching value for each standard feature set is determined by multiplying the ratio of the similarity of each standard feature set to the total similarity by 1000.

[0117] Based on the matching results, the adaptability of each asymmetric U-shaped network model is determined, including:

[0118] Analyze the matching results and determine the similarity of each standard feature set in the set of standard feature sets corresponding to the asymmetric U-shaped network model;

[0119] Based on the similarity, the adaptation coefficient is determined. By configuring a corresponding table of adaptation coefficient and similarity, the purpose of determining the adaptation coefficient based on the similarity can be achieved. The higher the similarity, the higher the adaptation coefficient.

[0120] Determine the degree of fitness based on the fitness coefficient and the matching value corresponding to the standard feature set;

[0121] The degree of adaptation is the sum of the products of the matching values ​​corresponding to each standard feature set and the corresponding adaptation coefficients.

[0122] The working principle and beneficial effects of the above technical solution are:

[0123] This embodiment provides a calling method for specific use of the network, which extracts features from the image to be processed based on the initial feature extraction template, constructs an image feature set, matches the image feature set with each standard feature set in the standard feature set set corresponding to each model in the model library, and then analyzes the most suitable model based on the matching results to further ensure the accuracy and effectiveness of the segmentation of the image to be segmented. The system corresponding to this application in this embodiment is applied on the terminal, and communicates with the server to call the server's model repository for model adaptation.

[0124] To improve the efficiency of model calling, in one embodiment, the processing module further performs the following operations:

[0125] Matching an image feature set of the image to be processed with an image feature set in a stored quick matching table constructed based on historical segmentation records;

[0126] When the match is successful, the model corresponding to the corresponding quick match item in the quick match table is used to process the pending items;

[0127] When the matching is unsuccessful, adaptation processing is performed through the server's preset model repository.

[0128] The quick matching items include: the image feature set corresponding to the image to be processed in the historical segmentation record and the unique identification code of the model.

[0129] In addition, the server performs optimization and comparison processing on the models stored in the model repository; this is achieved by configuring the model competition module; the model competition module performs the following operations:

[0130] Based on the standard feature set of each model in the model repository, two target models participating in the competition are selected;

[0131] Extracting a preset number of images to be segmented from a preset library of images to be segmented;

[0132] Using the extracted image to be segmented, the template model is tested to determine the optimal model;

[0133] A system that builds competition result information based on the optimization model and target model and sends it to the terminal;

[0134] The system's processing model updates the quick matching table based on the competition results.

[0135] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A breast cancer lesion segmentation system based on an asymmetric U-shaped network, characterized in that: include: The image to be processed acquisition module is used to acquire the image to be processed; A processing module, configured to input the image to be processed into an asymmetric U-shaped network to obtain a segmentation result; The asymmetric U-shaped network includes a multi-branch residual encoder, an external attention module and a boundary detection module; The boundary detection module includes: multi-level boundary detection, which is used to output lesion boundary prediction maps of different resolutions at different levels of the decoder; the prediction maps at each level serve as supervisory signals to focus on complex boundary contours; The multi-level boundary detection module outputs lesion boundary prediction maps of different resolutions at different levels of the decoder and performs the following operations: Upsample the preset high-resolution image to obtain a feature map with the same size as the label image; Performing a convolution operation to generate a first parameter set containing boundary information of a potential lesion area; Passing the first parameter set through a maximum pooling layer to obtain a second parameter set; An absolute difference set between the first parameter set and the second parameter set is calculated as a boundary prediction map.

2. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 1, characterized in that: The multi-branch residual encoder is used to extract features from the image to be processed using a preset number of branches, and then fuse each set of features to form a final representation feature.

3. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 2, characterized in that: The multi-branch residual encoder uses a preset number of branches to extract features from the image to be processed, and then fuses each set of features to form a final representation feature, including: Constructing a feature map corresponding to the image to be processed; Extract details from the feature map through multiple branches with shared weights to obtain outputs of models from each branch; Based on the feature map and the output of the model of each branch, the characterization feature is determined, and the calculation formula is as follows: Where, f is the characterization feature, I0 is the feature map, is the output of the model of the g-th branch, and G is the total number of branches.

4. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 1, characterized in that: Extracting details from the feature map through multiple branches with shared weights to obtain the output of the model of each branch includes: Each branch performs three convolution and normalization processing on the feature map.

5. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 1, characterized in that: The external attention module includes: a key matrix and a value matrix; The representation feature is multiplied by the transposed matrix of the key matrix and normalized, and then multiplied by the value matrix to obtain the output feature.

6. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 1, characterized in that: The total loss function used by the asymmetric U-shaped network is as follows: Among them, N is the total number of CNN layers with side output, Used to measure the similarity between the prediction result of the i-th lateral layer and the true label, and They represent the boundary loss and segmentation loss of the i-th layer respectively, ω1 is the preset first weight, ω2 is the preset second weight, and ω3 is the preset third weight.

7. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 1, characterized in that: Before the processing module inputs the image to be processed into the asymmetric U-shaped network to obtain the segmentation result, the processing module further performs the following operations: Based on a preset initial feature extraction template, feature extraction is performed on the image to be processed and an image feature set is constructed based on the extracted feature values; Matching the image feature set with each standard feature set in the set of standard feature sets corresponding to each asymmetric U-shaped network model in the preset model repository; Based on the matching results, the fitness of each asymmetric U-shaped network model is determined; The asymmetric U-shaped network with the highest adaptability is extracted.

8. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 7, characterized in that: The standard feature set associated with the asymmetric U-shaped network model is constructed by the following steps: Obtaining training data for the asymmetric U-shaped network model; Based on the preset initial feature extraction template, feature extraction is performed on each image in the training data and an image feature set corresponding to each image is constructed based on the extracted key feature values; Calculate the similarity between the image feature sets corresponding to each image; Extracting a preset number of image feature sets before the total similarity to construct the standard feature set set; wherein each image feature set of the standard feature set set is constructed as the standard feature set; The matching value is configured based on the total similarity corresponding to each standard feature set in the standard feature set set.

9. The asymmetric U-shaped network breast cancer lesion segmentation system according to claim 7, characterized in that: The determination of the adaptability of each asymmetric U-shaped network model based on the matching result includes: Analyzing the matching results to determine the similarity of each standard feature set in the set of standard feature sets corresponding to the asymmetric U-shaped network model; determining an adaptation coefficient based on the similarity; Determining a degree of adaptation based on the adaptation coefficient and a matching value corresponding to the standard feature set; The degree of adaptation is the sum of the products of the matching values ​​corresponding to each standard feature set and the corresponding adaptation coefficients.

Images