Breast lesion image semi-supervised segmentation method

Abstract

The application discloses a breast lesion image semi-supervised segmentation method, and belongs to the technical field of medical image processing and semi-supervised learning. The technical problem of improving the feature extraction capability of a semi-supervised segmentation network in a complex breast lesion segmentation scene and realizing accurate segmentation of a breast lesion is solved. A double-scale image of a breast image is generated, so that the network can learn the features of the breast lesion from different scales, and the feature extraction capability of the network for the breast lesion with a large size and shape difference is improved. Meanwhile, data disturbance is performed on the double-scale image, the segmentation confidence maps of the source image and the disturbed image are generated by the generator, and the stability of the network is improved. Through cascading inverse residual coding feature fusion and decoding feature fusion, more potential features are mined from existing features, the segmentation performance of the network is improved under a small amount of labeled samples, and accurate segmentation of the breast lesion can be realized by inputting the breast image into the final segmentation network.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and semi-supervised learning technology, specifically relating to a semi-supervised segmentation method for breast lesion images. Background Technology

[0002] Breast lesion examination is one of the important means of screening for breast lesions, and early and accurate identification and assessment are crucial for prognosis. Precise segmentation of breast lesions can help doctors accurately locate and delineate these areas, providing important support for the early diagnosis and treatment of breast cancer, and improving the accuracy and efficiency of diagnosis.

[0003] Deep learning, through multi-layered nonlinear transformations, can automatically extract rich features from source images, including low-level features (such as pixel brightness and local edges) and high-level abstract features (such as the overall morphology of breast lesions, contextual relationships, and boundary contours). This eliminates the need for manually designed complex feature extractors. Its ability to handle complex, nonlinear image features is advantageous for the automatic segmentation of breast images with complex and varied shapes, sizes, and internal structures. Data-driven deep convolutional neural networks rely on large amounts of high-quality data labels for segmentation performance. However, manually labeling data is time-consuming, labor-intensive, and costly, resulting in a scarcity of clinically usable labeled data. Semi-supervised learning can train networks using a small amount of labeled data and a large amount of unlabeled data, effectively alleviating the network's dependence on pixel-level labeled data and better meeting the needs of clinical applications.

[0004] Given the scarcity of labeled clinical breast images and the complex internal structures of breast lesions, current semi-supervised segmentation networks still face challenges in extracting complex features from breast images. Improving the feature extraction capabilities of these networks in complex breast lesion segmentation scenarios with a limited number of labeled samples, and achieving accurate segmentation of breast lesions, remains a pressing issue for automated clinical breast lesion examination. Summary of the Invention

[0005] To address the technical challenge of improving the feature extraction capability of semi-supervised segmentation networks in complex breast lesion segmentation scenarios with a limited number of labeled samples, and thus achieving accurate segmentation of breast lesions, this invention proposes a semi-supervised segmentation method for breast lesion images.

[0006] The method includes the following steps:

[0007] S1. Collect breast-derived images, including labeled breast-derived images. And unlabeled breast-derived images ,right Downsampling was performed to obtain ,right Downsampling was performed to obtain ,right Data perturbation is obtained ,right Data perturbation is obtained ;

[0008] and , and , and These are used to construct three sets of dual-scale images, each containing the original scale image size and the downsampled scale image size.

[0009] S2. Construct a semi-supervised segmentation network for breast lesions. The workflow of the network is as follows:

[0010] S21. The dual-scale image is fed into the shared encoder for feature encoding to generate a dual-scale encoded feature map.

[0011] S22. The dual-scale encoded feature map is fed into the cascaded inverse residual encoded feature fusion module across layers to perform cross-layer feature fusion and obtain the dual-scale cross-layer fused feature map.

[0012] S23. The dual-scale encoded feature map is fed into two independent feature decoders, and the dual-scale cross-layer fusion feature map is fused with the decoding features of the corresponding layer through skip connections to obtain the dual-scale decoded feature map and the dual-scale segmentation image.

[0013] S24. The dual-scale decoded feature map is fed into the dual-scale decoded feature fusion unit to perform step-by-step feature fusion and decoding on the dual-scale decoded feature map, and obtain the dual-scale decoded fused feature map and the breast lesion segmentation image.

[0014] S3. The parameters of the semi-supervised segmentation network for breast lesions are jointly trained based on the supervised loss of labeled images and the consistency loss of unlabeled images to obtain the final segmentation network.

[0015] S4. Input the breast image to be segmented into the final segmentation network to achieve segmentation of breast lesions.

[0016] Furthermore, the shared encoder performs feature encoding on the original-scale image and the downsampled-scale image to obtain the encoded feature map at the original scale. and the encoded feature map at the downsampling scale ,in, For the encoded feature map corresponding to the shared encoder, the first... The layer output uses a Res2net network for the shared encoder.

[0017] Furthermore, the cross-layer feature fusion in the cascaded inverse residual coding feature fusion module is specifically as follows:

[0018] ;

[0019] ;

[0020] in, express The fused feature map obtained after cross-layer feature fusion. express The fused feature map obtained after cross-layer feature fusion. This indicates the channel attention mechanism. Indicating spatial intention mechanism, This indicates a channel splicing operation. This indicates a bilinear interpolation upsampling operation. This refers to the cascaded inverse residual module. The specific data processing flow within the cascaded inverse residual module is as follows:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] in, Represents the input feature map, This represents the feature map after channel dimensionality upscaling. It is a 3×3 depthwise separable convolution. For batch normalization operations, For activation function, Represents a 1×1 convolution. This represents the feature map after residual connection of the upgraded feature map; The feature map is obtained by performing a residual connection operation between the dimensionality-reduced feature map and the input feature map. This represents the feature map finally generated by the cascaded inverse residual module.

[0026] Furthermore, the dual-scale encoded feature maps are fed into two independent feature decoders, and... and The dual-scale encoded feature map is decoded by passing it through a skip connection to the feature decoder, resulting in a dual-scale decoded feature map. and and dual-scale segmented images and Specifically:

[0027] ;

[0028] ;

[0029] in, Indicates transposed convolution; ;

[0030] .

[0031] Furthermore, the dual-scale decoding feature map and The data is fed into a dual-scale decoding feature fusion processor, which performs step-by-step feature fusion, clustering, and decoding on the dual-scale features to obtain a dual-scale decoding fusion feature map. Segmentation images of breast lesions : Dual-scale decoding feature fusion The specific operations performed are as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] in, GAFM This indicates the gating attention fusion module. GAFM The specific operations performed are as follows:

[0038] ;

[0039] in, This indicates an element-wise multiplication operation. This indicates a grouped convolution, where the number of groups is half the number of channels. This represents the first feature map of the input. This represents the second feature map of the input. This represents the output feature map;

[0040] Decoding and fusion features Step-by-step decoding is performed to obtain segmented images of breast lesions. Specifically:

[0041] ;

[0042] .

[0043] Furthermore, when training the parameters of the semi-supervised breast lesion segmentation network based on the combined supervised loss of labeled images and the consistency loss of unlabeled images, the total loss function is: ,in, This represents the two-scale consistency loss. This indicates the loss of consistency due to perturbation. This represents the generation of consistency loss. This indicates a loss of oversight.

[0044] further, ;in, Indicates the mean square error loss; Unlabeled breast-derived images , To Image obtained by data perturbation ;

[0045] and The solution process is as follows:

[0046] Will The data was fed into a semi-supervised segmentation network for breast lesions to obtain segmentation confidence maps of breast lesions. ,Will The data was fed into a semi-supervised segmentation network for breast lesions to obtain segmentation confidence maps of breast lesions. ;Will and Feed into generator Image generation is performed, specifically as follows:

[0047] ;

[0048] ;

[0049] and They represent and The perturbation image generated by the generator and the source image;

[0050] Will Feed into generator Generated in The specific operation is as follows:

[0051] ;

[0052] ;

[0053] ;

[0054] in, This indicates a downsampling operation based on max pooling. This indicates an upsampling operation based on transpose convolution; Generation process and same.

[0055] further,

[0056] ;

[0057] in, Cross-entropy loss and The average loss, Image of labeled breast tissue Corresponding label image , This indicates a downsampling operation based on nearest neighbor interpolation.

[0058] further, ,in, This represents the cross-entropy loss.

[0059] further,

[0060] ;

[0061] in, express The breast lesion segmentation prediction image obtained after inputting into a semi-supervised breast lesion segmentation network. express The breast lesion segmentation prediction image obtained after inputting into a semi-supervised breast lesion segmentation network. express and The breast lesion segmentation prediction image is obtained after inputting into the semi-supervised segmentation network of breast lesions and fusing decoded features.

[0062] The beneficial effects of the method described in this invention are as follows:

[0063] 1. This invention downsamples the source image to generate a dual-scale image of the breast, enabling the network to learn the features of breast lesions from different scales, improving the network's ability to extract features of breast lesions with large differences in size and shape. At the same time, the dual-scale image is perturbed, and the generator performs image mutual generation between the segmentation confidence maps of the source image and the perturbed image, improving the stability of the network in learning breast lesion features under interference scenarios such as motion artifacts, blur, and noise.

[0064] 2. This invention employs cascaded inverse residual coding feature fusion and decoding feature fusion, enabling the mining of more latent features using existing features. This improves the segmentation performance of the network with a limited number of labeled samples. By inputting breast images into the final segmentation network, accurate segmentation of breast lesions can be achieved. This invention can effectively segment breast lesions, assisting doctors in accurately locating and delineating breast lesion regions, thus improving the accuracy and efficiency of clinical diagnosis. This invention solves the problem of accurate segmentation of breast lesions with a limited number of labeled clinical samples. Attached Figure Description

[0065] Figure 1 This is a flowchart of the method described in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the encoder and decoder in the network structure of an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of a dual-scale decoding feature fusion decoder in the network structure of an embodiment of the present invention;

[0068] Figure 4 This is a schematic diagram of the original-perturbation image generator in the network structure of an embodiment of the present invention;

[0069] Figure 5 This is a schematic diagram of the cascaded inverse residual module and the cascaded inverse residual coding feature fusion module in the network structure of this embodiment of the invention;

[0070] Figure 6 This is a schematic diagram of the dual-scale decoding feature fusion module and the gated attention fusion module in the network structure of this embodiment of the invention;

[0071] Figure 7 This is a visualization of the segmentation results after training the network with 30% labels on a breast ultrasound image dataset, as shown in this embodiment of the invention. Detailed Implementation

[0072] The technical solutions illustrated in the schematic diagrams of the embodiments of the present invention will be further described in detail below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, those skilled in the art can further extend the above technical solutions to application scenarios such as lesion segmentation in breast ultrasound and breast MRI images without creative effort, and all such extensions fall within the scope of protection of the present invention.

[0073] Example 1

[0074] This embodiment provides a semi-supervised segmentation method for breast lesion images, such as... Figure 1 As shown, the flow of the method is as follows:

[0075] S1. Collect breast-derived images, including labeled breast-derived images. And unlabeled breast-derived images ,right Downsampling was performed to obtain ,right Downsampling was performed to obtain ,right Data perturbation is obtained ,right Data perturbation is obtained ;

[0076] and , and , and These are used to construct three sets of dual-scale images, each containing the original scale image size and the downsampled scale image size.

[0077] The breast-derived images include breast images with lesions and breast images without lesions, and the source of the breast-derived images can be breast ultrasound images or breast MRI images.

[0078] S2. Construct a semi-supervised segmentation network for breast lesions. The workflow of the network is as follows:

[0079] S21. The dual-scale image is fed into the shared encoder for feature encoding to generate a dual-scale encoded feature map.

[0080] S22. The dual-scale encoded feature map is fed into the cascaded inverse residual encoded feature fusion module across layers to perform cross-layer feature fusion and obtain the dual-scale cross-layer fused feature map.

[0081] S23. The dual-scale encoded feature map is fed into two independent feature decoders, and the dual-scale cross-layer fusion feature map is fused with the decoding features of the corresponding layer through skip connections to obtain the dual-scale decoded feature map and the dual-scale segmentation image.

[0082] S24. The dual-scale decoded feature map is fed into the dual-scale decoded feature fusion decoder. The dual-scale decoded feature map is fused and decoded step by step to obtain the dual-scale decoded fusion feature map and the breast lesion segmentation image.

[0083] S3. The parameters of the semi-supervised segmentation network for breast lesions are jointly trained based on the supervised loss of labeled images and the consistency loss of unlabeled images to obtain the final segmentation network.

[0084] S4. Input the breast image to be segmented into the final segmentation network to achieve segmentation of breast lesions.

[0085] Example 2

[0086] This embodiment further defines embodiment 1 and provides further explanation of step S1.

[0087] The source breast ultrasound image is downsampled by using nearest neighbor interpolation, with the downsampling scale being approximately half the size of the source ultrasound image, ensuring that the size is divisible during the feature extraction process. Data perturbation is performed by randomly altering the brightness and hue of the source breast ultrasound image to generate a perturbed breast ultrasound image.

[0088] Example 3

[0089] This embodiment further defines embodiment 1 and provides further explanation of step S2.

[0090] In this embodiment, as Figure 2 The encoder shown in the figure: A shared encoder is used to encode features of the original scale image and the downsampled scale image to obtain the encoded feature map at the original scale. and the encoded feature map at the downsampling scale ,in, For the encoded feature map corresponding to the shared encoder, the first... The layer output uses a Res2net network for the shared encoder.

[0091] In this embodiment, Figure 2 encoder and Figure 5 The cascaded inverse residual module and the cascaded inverse residual coding feature fusion module in the middle As shown:

[0092] Cross-layer feature fusion in the cascaded inverse residual coding feature fusion module is specifically as follows:

[0093] ;

[0094] ;

[0095] in, express The fused feature map obtained after cross-layer feature fusion. express The fused feature map obtained after cross-layer feature fusion. This indicates the channel attention mechanism. Indicating spatial intention mechanism, This indicates a channel splicing operation. This indicates a bilinear interpolation upsampling operation. This refers to the cascaded inverse residual module, which reduces network parameters while preserving low-level features, thereby improving the feature extraction capability of the breast lesion segmentation network. The specific data processing flow in the cascaded inverse residual module is as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] in, Represents the input feature map, This represents the feature map after channel dimensionality upgrade, where the number of channels becomes twice the original number, in order to improve the model's feature extraction capability. It is a 3×3 depthwise separable convolution. For batch normalization operations, For activation function, Represents a 1×1 convolution. This represents the feature map after residual connection of the upgraded feature map; The feature map is obtained by performing a residual connection operation between the dimensionality-reduced feature map and the input feature map. This represents the feature map finally generated by the cascaded inverse residual module.

[0101] In this embodiment, as Figure 2 decoder and Figure 5 The cascaded inverse residual module in the diagram shows how the dual-scale encoded feature maps are fed into two independent feature decoders. and and will and The dual-scale encoded feature map is decoded by passing it through a skip connection to the feature decoder, resulting in a dual-scale decoded feature map. and and dual-scale segmented images and The feature decoder uses transposed convolution for upsampling to restore the image size and utilizes a cascaded inverse residual module for feature decoding. Cross-layer fused features are passed to the decoder via skip connections and fused with the decoded features using channel concatenation operations. Specifically:

[0102] ;

[0103] ;

[0104] in, Indicates transposed convolution; ;

[0105] .

[0106] In this embodiment, as Figure 3 Dual-scale decoding feature fusion decoder and Figure 6 The dual-scale decoding feature fusion module and gated attention fusion module in the figure are shown: the dual-scale decoding feature map and Feed into dual-scale decoding feature fusion decoder The dual-scale features are fused, clustered, and decoded stepwise to obtain the dual-scale decoded and fused feature map of breast ultrasound. Segmentation images of breast lesions : Dual-scale decoding feature fusion decoder The specific operations performed are as follows:

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] in, GAFM This indicates the gating attention fusion module. GAFM The specific operations performed are as follows:

[0113] ;

[0114] in, This indicates an element-wise multiplication operation. This indicates a grouped convolution, where the number of groups is half the number of channels. This represents the first feature map of the input. This represents the second feature map of the input. This represents the output feature map;

[0115] Decoding and fusion features Step-by-step decoding is performed to obtain segmented images of breast lesions. Specifically:

[0116] ;

[0117] .

[0118] Example 4

[0119] This embodiment further defines embodiment 1 and provides further explanation of step S3.

[0120] When training the parameters of a semi-supervised breast lesion segmentation network using a combination of supervised loss based on labeled images and consistency loss based on unlabeled images, the total loss function is: ,in, This represents the two-scale consistency loss. This indicates the loss of consistency due to perturbation. This represents the generation of consistency loss. This indicates a loss of oversight.

[0121] ;in, Indicates the mean square error loss; Unlabeled breast ultrasound images , To Image obtained by data perturbation ;

[0122] and The solution process is as follows:

[0123] Will The data was fed into a semi-supervised segmentation network for breast lesions to obtain segmentation confidence maps of breast lesions. ,Will The data was fed into a semi-supervised segmentation network for breast lesions to obtain segmentation confidence maps of breast lesions. The difference between the segmentation confidence map and the segmentation prediction image is that the segmentation prediction image undergoes binarization processing on top of the segmentation confidence map. After being fed into a semi-supervised segmentation network for breast lesions, the segmentation confidence map of the breast lesions is obtained after encoding and decoding. .

[0124] Will and Feed into generator Image generation is performed, specifically as follows:

[0125] ;

[0126] ;

[0127] and They represent and The perturbation image generated by the generator and the source image;

[0128] In this embodiment, as Figure 4 As shown, generator Its structure uses a simplified Unet network, retaining only the U-shaped structure for encoding and decoding images.

[0129] Will Feed into generator Generated in The specific operation is as follows:

[0130] ;

[0131] ;

[0132] ;

[0133] in, This indicates a downsampling operation based on max pooling. This indicates an upsampling operation based on transpose convolution; Generation process and same.

[0134] ;

[0135] in, Cross-entropy loss and The average loss, Breast ultrasound source image Corresponding label image , This indicates a downsampling operation based on nearest neighbor interpolation.

[0136] ,in, This represents the cross-entropy loss.

[0137] ;

[0138] in, express The breast lesion segmentation prediction image obtained after inputting into a semi-supervised breast lesion segmentation network. express The breast lesion segmentation prediction image obtained after inputting into a semi-supervised breast lesion segmentation network. express and The breast lesion segmentation prediction image is obtained after inputting into the semi-supervised segmentation network of breast lesions and fusing decoded features.

[0139] Example 5

[0140] This embodiment further defines Embodiment 1. To verify the effectiveness and superiority of this method, a method comparison was conducted. URPC, DCNet, SSNet, and MCNet semi-supervised segmentation methods were additionally selected for parallel comparison. Each method was used to perform segmentation operations on a breast ultrasound dataset with 30% labels, and quantitative comparisons were performed. The mDice, mIoU, and other metrics for each method were calculated. , And the MAE evaluation index values, the segmentation results are as follows: Figure 7 As shown in Table 1, the comparison results are as follows: As can be seen from the table, the evaluation index values ​​of the method in this embodiment are better than those of similar comparative methods on the breast ultrasound dataset with 30% labels. Furthermore, the segmentation results of benign and malignant lesions are closer to the true segmentation labels. It can be seen that the prediction results of the segmentation method in this embodiment effectively improve the segmentation performance of breast ultrasound lesions with a small number of labeled samples.

[0141] Table 1 compares the segmentation results of each method on the breast ultrasound dataset under 30% labeling.

[0142]

[0143] Example 6

[0144] This embodiment also provides a semi-supervised segmentation system for breast lesion images, applied to the semi-supervised segmentation methods for breast lesion images in embodiments 1-4, including:

[0145] The network building unit is used to construct a semi-supervised segmentation network for breast lesion images;

[0146] The training unit is used to train the parameters of the semi-supervised segmentation network for breast lesion images;

[0147] The output unit is used to output the visualization results of breast lesion image segmentation.

[0148] The present invention also provides a semi-supervised segmentation device for breast lesion images.

[0149] The aforementioned device includes: a processor and a memory storing computer program instructions;

[0150] When the processor executes the computer program instructions, it implements the semi-supervised segmentation method for breast lesion images.

Claims

1. A breast lesion image semi-supervised segmentation method, characterized in that, The method comprises the following steps: S1. Collect breast-derived images, including labeled breast-derived images. And unlabeled breast-derived images ,right Downsampling was performed to obtain ,right Downsampling was performed to obtain ,right Data perturbation is obtained ,right Data perturbation is obtained ; and , and , and respectively constitute three groups of bi-scale images comprising original scale image sizes and down-sampled scale image sizes; S2, constructing a breast lesion semi-supervised segmentation network, the workflow in the network is: S21, sending the double-scale image into a shared encoder for feature coding to generate a double-scale coding feature map; S22, sending the double-scale coding feature map into a cascaded inverse residual coding feature fusion module across layers for cross-layer feature fusion to obtain a double-scale cross-layer fusion feature map; S23, sending the double-scale coding feature map into two independent feature decoders, and fusing the double-scale cross-layer fusion feature map with the decoding features of the corresponding layers through a jump connection to obtain a double-scale decoding feature map and a double-scale segmentation image; S24, sending the double-scale decoding feature map into a double-scale decoding feature fusion decoder for step-by-step feature fusion and decoding of the double-scale decoding feature map to obtain a double-scale decoding fusion feature map and a breast lesion segmentation image; S3, jointly training breast lesion semi-supervised segmentation network parameters based on a supervised loss of a labeled image and a consistency loss of an unlabeled image to obtain a final segmentation network; S4, inputting a breast image to be segmented into the final segmentation network to realize breast lesion segmentation.

2. The breast lesion image semi-supervised segmentation method of claim 1, wherein, The shared encoder performs feature encoding on the original-scale image and the downsampled-scale image to obtain the encoded feature map at the original scale. and the encoded feature map at the downsampling scale ,in, For the encoded feature map corresponding to the shared encoder, the first... The layer output uses a Res2net network for the shared encoder.

3. The breast lesion image semi-supervised segmentation method of claim 2, wherein, The cross-layer feature fusion in the cascaded inverse residual coding feature fusion module is specifically: ； ； wherein, represents a fusion feature map obtained after cross-layer feature fusion, represents a fusion feature map obtained after cross-layer feature fusion, represents a channel attention mechanism, represents a spatial attention mechanism, represents a channel concatenation operation, represents a bilinear interpolation up-sampling operation, represents a cascaded inverse residual module, and a data processing procedure performed in the cascaded inverse residual module is specifically: ； ； ； ； wherein, represents the input feature map, represents the feature map after channel dimensionality increase, is a 3x3 depth separable convolution, is a batch normalization operation, is an activation function, represents a 1x1 convolution, represents the feature map after residual connection of the dimensionality-increased feature map; is a feature map obtained after residual connection operation of the dimensionality-reduced feature map and the input feature map, represents the feature map finally generated by the cascaded inverse residual module.

4. The breast lesion image semi-supervised segmentation method of claim 3, wherein, The double-scale encoded feature maps are respectively sent into two independent feature decoders, and and The double-scale encoded feature maps are decoded by the feature decoders passed in through the skip connection to obtain double-scale decoded feature maps and , and a double-scale segmentation image and , specifically as follows: ； ； wherein denotes a transposed convolution; ; 。 5. The breast lesion image semi-supervised segmentation method of claim 4, wherein, The double-scale decoding feature map is input into a double-scale decoding feature fusion decoder to perform hierarchical feature fusion and decoding on the double-scale features to obtain a double-scale decoding fusion feature map and The double-scale decoding feature fusion decoder and the breast lesion segmentation image : The double-scale decoding feature fusion decoder The operations performed in the double-scale decoding feature fusion decoder are specifically: ； ； ； ； ； In the formula, A represents a substituted or unsubstituted aryl group, and B represents a substituted or unsubstituted heteroaromatic group. GAFM denotes a gated attention fusion module, GAFM The operations performed in the step 1300 include: ； wherein, represents an element-wise multiplication operation, represents a grouped convolution with a group number being half of the channel number, represents a first input feature map, represents a second input feature map, represents an output feature map; Decoding fused features Step-by-step decoding to obtain a breast lesion segmentation image Specifically, ； 。 6. The breast lesion image semi-supervised segmentation method of claim 5, wherein, When the breast lesion semi-supervised segmentation network parameters are jointly trained based on the supervised loss of the labeled image and the consistency loss of the unlabeled image, the total loss function is: wherein, denotes the double-scale consistency loss, denotes the perturbation consistency loss, denotes the generation consistency loss, denotes the supervised loss.

7. The breast lesion image semi-supervised segmentation method of claim 6, wherein, ; wherein, represents a mean squared error loss; is a breast source image without a label , is an image obtained by performing data perturbation on ;​ and The solving process of the above equation is as follows: Will The data was fed into a semi-supervised segmentation network for breast lesions to obtain segmentation confidence maps of breast lesions. ,Will The data was fed into a semi-supervised segmentation network for breast lesions to obtain segmentation confidence maps of breast lesions. ;Will and Feed into generator Image generation is performed, specifically as follows: ； ； and respectively represent and the perturbed image and the source image generated by the generator The operation of sending into the generator is specifically:​​ ； ； ； wherein, denotes a down-sampling operation based on max-pooling, denotes an up-sampling operation based on transposed convolution; the generation process is the same as the one described above.

8. The breast lesion image semi-supervised segmentation method of claim 7, wherein, ； wherein, is a cross-entropy loss and is the average loss of the losses, is a labeled breast-derived image is a corresponding label map , denotes a down-sampling operation based on nearest-neighbor interpolation.

9. The breast lesion image semi-supervised segmentation method of claim 8, wherein, wherein, denotes the cross-entropy loss.

10. The breast lesion image semi-supervised segmentation method according to claim 9, characterized in that, ； wherein, denotes a breast lesion segmentation prediction image obtained after inputting the breast lesion semi-supervised segmentation network, denotes a breast lesion segmentation prediction image obtained after inputting the breast lesion semi-supervised segmentation network, denotes and a breast lesion segmentation prediction image obtained after inputting the breast lesion semi-supervised segmentation network,

Images