Double labeling method for man-machine collaboration

Abstract

The invention requests to protect a man-machine collaborative double labeling method, which comprises the following steps of: acquiring an image by utilizing an ABUS, and dividing the image into an unlabeled data set N and a labeled data set M; according to the method, image augmentation processing and normalization are adopted, and image features are enhanced by using a self-supervised framework based on comparative learning. Secondly, constructing a visual model by using a deep learning architecture U-Net and ResNet50, automatically labeling a region of interest in the image, and providing preliminary estimation; thereafter, manual annotation is performed on the automatically annotated mammary gland image, man-machine dual annotation is realized, and more accurate estimation is provided. And finally, learning image representation from the unlabeled data set N by using a comparative learning algorithm, and updating the visual model in combination with the supervision signal in the labeled data set M. According to the method, artificial intelligence and labeling results of professional doctors are comprehensively utilized, and a labeling quality evaluation mechanism is introduced, so that the recognizable performance of the image is improved, the expressive force of image features is enhanced, and the generalization ability of a visual model is further enhanced.

Description

Technical Field

[0001] The present invention relates to the fields of image processing, deep learning, and in particular to a human-machine collaborative double labeling method. Background Art

[0002] Breast cancer has become one of the malignant tumors that seriously affects women's health. According to the 2023 global cancer statistics released by the Global Cancer Association, the incidence of breast cancer in women has exceeded that of lung cancer, accounting for 31% of total cancer cases. Among women worldwide, breast cancer accounts for a quarter of all cancer cases and one-sixth of all cancer deaths. A large amount of evidence-based medical evidence shows that widespread screening in the population can help early diagnosis and detection of breast cancer, promote early treatment of the disease, increase the chance of successful treatment, and thus reduce mortality.

[0003] Traditional breast disease diagnosis mainly relies on doctors' interpretation of medical images, but this method has the problems of subjectivity and insufficient diagnostic accuracy. At the same time, manual labeling and analysis is a tedious and time-consuming task when faced with large-scale data sets. Therefore, automated and semi-automated analysis methods are considered to be the key to solving this problem. In recent years, artificial intelligence and machine learning technologies have been shown to improve the accuracy and efficiency of breast disease detection. However, in order to ensure the accuracy of diagnosis, large-scale medical imaging data sets still need to be combined with the expertise of human experts, so it is necessary to explore a dual labeling method of human-machine collaboration. Summary of the invention

[0004] The present invention provides a human-machine collaborative dual labeling method to solve the technical problems of inaccurate prediction results of existing automatic labeling methods and low efficiency of manual labeling, while improving the detection effect, robustness and generalization ability of the model, and is suitable for improving the accuracy of diagnosing breast diseases.

[0005] To achieve the above objectives, the present invention proposes a human-machine collaborative dual labeling method, which is specifically applied to breast disease diagnosis. The specific implementation steps are as follows:

[0006] Step 1: Generate an image using an automatic imaging system and divide the image into an unlabeled data set N and a labeled data set M. In the embodiment, an automatic breast volume ultrasound imaging (Automatic Breast Ultrasound System, ABUS) is used to collect breast images of patients and divide them into an unlabeled data set N and a labeled data set M;

[0007] Step 2: preprocess the labeled data set M; enlarge the data set through image augmentation and normalize the input data;

[0008] Step 3: The preprocessed image is subjected to image representation learning through a self-supervised framework based on contrastive learning to enhance its image features;

[0009] Step 4: For the image enhanced with features in step 3, learn it based on deep learning combined with transfer learning, build a visual model using U-Net and residual network (ResNet) deep learning architecture, automatically annotate the region of interest in the image, and obtain an image dataset M; provide a preliminary estimate of breast diseases (e.g., lumps and masses);

[0010] Step 5: manually annotate the breast images that were automatically annotated in step 4 again, so as to achieve human-machine dual annotation and obtain an updated image dataset M, providing a more accurate estimation of breast diseases;

[0011] Step 6: Use a contrastive learning algorithm to learn image representation from the unlabeled dataset N, extract additional supervisory signals from the labeled dataset M, such as disease labels or segmentation masks, use a contrastive learning model to learn feature representation, and combine the supervisory signals to update the visual model.

[0012] The specific implementation steps of step one are as follows:

[0013] An automatic breast volume ultrasound imaging system is used. The ultrasound coupling agent is evenly applied to one breast. The disposable coupling film covers the scanner. The operator places the scanner on one breast and presses gently to ensure contact with the breast skin. After the scanner is in place, the operator starts scanning. Each breast is repeated three times, in the three quadrants of AP, LAT, and MED, to form a complete view of the entire breast. If necessary, other views can be added as needed. The scanning time in one direction is about 60 seconds, and the inter-layer spacing of the acquired image is 0.5 mm. After the scan is completed, the data is transmitted to the workstation to obtain automatically reconstructed cross-sectional, coronal, and sagittal images; the acquired images are divided into an unlabeled data set N and a labeled data set M;

[0014] The specific implementation steps of step 2 are as follows:

[0015] First, through image augmentation, the images in the dataset M are rotated by 0°, 90°, 180°, and 270° respectively, and then horizontal or vertical flips are randomly applied to the images to increase the dataset; the augmented images are used as input;

[0016] Second, normalize the input data, and its conversion function is:

[0017] x new =(x-μ) / σ;

[0018] Among them, μ is the mean of the sample data, σ is the standard deviation of the sample data, and x newis the image data after normalization.

[0019] The specific implementation steps of step three are as follows:

[0020] First, for the image sequence Where T represents the sequence length, D is the number of sequence layers, H and W are the length and width of a single sequence, and data augmentation operations are performed to generate two different views, denoted as I_1 and I_2.

[0021] Second, a momentum encoder is introduced to provide contrastive supervision for the online encoder to learn discriminative representations. The enhanced full view is used as input to preserve the semantic integrity and discriminability of the learned representation. Since the semantics of image tags are ambiguous, unlike the semantically determined tags in natural language processing, it is logical to use global representations for contrastive learning. The formula is as follows:

[0022]

[0023] in is the input flag of the momentum encoder, represents the momentum encoder, represents the momentum encoder output sequence, Represents the output sequence Features after global pooling.

[0024] The momentum encoder updates parameters using an exponential moving average (EMA), which can promote smooth feature changes for more stable training. and The parameters are denoted as θ s and θ t , then the parameter update formula is as follows:

[0025] θ t ←μθ t +(1-μ)θ s

[0026] Where μ is the moving average hyperparameter.

[0027] Third, in order to align with the output of the momentum encoder, a feature decoder is introduced. The feature decoder has the same structure as the pixel decoder, but it contains fewer transformer blocks and a lighter structure. The two serve different learning objectives and do not share parameters. The feature decoder aims to transform the latent space feature z s Decoded into the momentum encoder feature space, similar to the momentum encoder, the output of the feature decoder is also average pooled as a representation of the entire image, and then the feature is used for contrastive learning, the specific form is as follows:

[0028]

[0029] Where N is the number of all markers in the set, including visible and invisible markers, is the feature decoder output.

[0030] Fourth, for the contrastive learning objective, an asymmetric contrastive loss function is designed based on InfoNCE, which aims to bring semantically consistent samples in a batch closer and push inconsistent samples away. In order to enhance the learning ability of the model, "projection prediction" and "projection" layers are attached to the feature decoder and momentum encoder respectively. Projection head of momentum encoder The parameters are also updated by exponential sliding average. Specifically, the output y of the feature decoder is x It will be converted through the "projection prediction" structure Similarly, the output of the momentum encoder is represented by z t , apply the "projection" head to get Where D p Represents the dimension after projection, and then calculates the cosine similarity ρ between the two:

[0031]

[0032] Next, ρ + It is expressed as the cosine similarity of the positive pair, which is given by and To construct, represents the cosine similarity of the jth negative pair, and is obtained by using the momentum encoder for different images in a batch Construct negative sample pairs, and the loss function is as follows:

[0033]

[0034] Where τ is the temperature coefficient and K represents the batch size.

[0035] Fifth, at the feature level, information transfer in two situations is achieved through self-distillation, and the loss form is as follows:

[0036]

[0037] Where L latent Represents the loss function value.

[0038] Sixth, the loss function of the entire framework can be expressed as:

[0039] L=L r +λ c L c +λ latent L latent

[0040] in λ c represents the contrast loss weight, λ latent represents the consistency loss weight.

[0041] The specific implementation steps of step 4 are as follows:

[0042] First, the constructed visual models include the classification model ResNet50 and the segmentation model U-Net, which are pre-trained in ImageNet before construction;

[0043] Second, the structure of the U-Net network can be divided into two main parts: the contraction path and the expansion path. In the first part, the contraction path, each block consists of two consecutive 3×3 convolutional layers, followed by a ReLU activation unit and a max-pooling layer. This combination is repeated multiple times to form the architecture of the entire contraction path. The innovation of U-Net lies in the second part, the expansion path. At each stage of the expansion path, the feature map is upsampled by a 2×2 up-convolution. Then, the feature map of the same size is cropped from the corresponding contraction path layer and connected to the upsampled feature map. This is followed by two consecutive 3×3 convolutional layers and ReLU activation. In the final stage, the number of channels of the feature map is reduced to the required number through an additional 1×1 convolution to generate the final segmentation image.

[0044] The energy function of the network is given by:

[0045] E = ∑ω(x)log(p k(x) (x)

[0046] Among them, p k is the pixel-level SoftMax function applied to the final feature map, defined as:

[0047]

[0048] Among them, a k represents the activation k in the channel.

[0049] Third, ResNet50 uses the output of the previous layer directly across several layers as a shortcut link for the input of a subsequent layer. Its input function is:

[0050] F(x)=H(x)-x

[0051] Among them, F(x) is the residual, H(x) is the complete output of the previous layer, and x is the input of the previous layer.

[0052] During the training process, the loss function will swing after the parameters are updated. The network optimizer can optimize the swing amplitude of the loss function, which can be achieved by using the root mean square transfer algorithm (RMSProp), and the convergence speed of the function will be further accelerated. The learning rate is a hyperparameter that adjusts the network weights through the gradient of the loss function, and the interval adjustment learning rate (stepLR) decay strategy is adopted. The learning rate is set to 0.001, the maximum iteration step is set to 5000, and the learning rate is decayed by 1 / 2 at 2000 and 4000 steps. Since the early layers of CNN have little correlation with the specific image dataset, the relationship with the selected dataset and its task objectives becomes closer as time goes by. Therefore, fine-tuning starts from the last layer of ResNet50, and then gradually includes more layers in the update process.

[0053] The specific implementation steps of step 5 are as follows:

[0054] After the breast images are initially annotated using automatic annotation techniques (such as visual models based on U-Net and ResNet50), multiple trained human experts use specialized annotation tools or software to perform manual annotation and update the annotated data to the annotation dataset M;

[0055] The specific implementation steps of step six are as follows:

[0056] First, use the contrastive learning model to learn image feature representation from the unlabeled dataset N, which is the same as step 3;

[0057] Second, extract additional supervisory signals such as disease labels or segmentation masks from the updated annotated dataset M;

[0058] Third, the contrastive learning task is combined with the supervision signal to form a multi-task loss function as follows:

[0059] L=L c +λ s *L s

[0060] Among them, L c is the contrast loss, L s is the supervision loss, λ s is the weight of the supervision loss.

[0061] Fourth, the visual model is trained using a multi-task loss function, and the parameters of the contrastive learning and supervised learning modules are updated simultaneously. By minimizing the multi-task loss function, the visual model can simultaneously learn meaningful feature representations and task-specific knowledge, thereby improving its generalization ability and accuracy.

[0062] The characteristics and effective effects of the present invention are:

[0063] The present invention cleverly combines the essence of automatic labeling and manual labeling, improves the image's recognizability, significantly improves the accuracy of breast disease diagnosis in practical applications, and effectively reduces the risk of missed diagnosis and misdiagnosis. At the same time, by efficiently using contrastive learning algorithms, the present invention strengthens the expressiveness of image features and further enhances the generalization ability of visual models. On this basis, the present invention also cleverly combines advanced deep learning architectures such as U-Net and ResNet50, and is supplemented by transfer learning technology, so that the system can efficiently learn and identify key areas in images, such as key areas in breast images in embodiments of the present invention, greatly reducing the time and effort required for manual labeling, thereby greatly improving diagnostic efficiency, and providing strong technical support for early screening of breast diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of an embodiment of the present invention;

[0065] Figure 2 It is the overall U-net architecture;

[0066] Figure 3 This is the ResNet-50 structure diagram;

[0067] Figure 4 It is the Chinese translation of the ResNet-50 structure diagram. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. It should be understood that the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0069] like Figure 1 In one embodiment shown, the present invention proposes a human-machine collaborative dual labeling method, which is specifically applied to breast disease diagnosis. The specific implementation steps are as follows:

[0070] Step 1: Generate images using an automatic imaging system and divide the images into an unlabeled dataset N and a labeled dataset M. In the embodiment, the patient's breast images are collected using Automatic Breast Ultrasound System (ABUS), and are divided into an unlabeled dataset N and a labeled dataset M. A new breast three-dimensional ultrasound technology is used to comprehensively observe the internal structure of the breast, overcome the problems of small handheld ultrasound area, no coronal images, and reliance on operator experience. The specific steps are as follows:

[0071] First, using an automated breast volume ultrasound imaging system, the ultrasound coupling agent is evenly applied to one breast, a disposable coupling film covers the scanner, and the operator places the scanner on one breast and presses gently to ensure contact with the breast skin;

[0072] Second, after the scanner is in place, the operator begins scanning, repeating three times for each breast, in the AP, LAT, and MED quadrants, to form a complete view of the entire breast, and other views can be added as needed. The scanning time in one direction is about 60 seconds, and the inter-layer spacing of the acquired images is 0.5 mm;

[0073] Third, after the scan is completed, the data is transferred to the workstation to obtain automatically reconstructed cross-sectional, coronal and sagittal images; the acquired images are divided into an unlabeled data set N and a labeled data set M.

[0074] Step 2: Preprocess the labeled dataset M; by augmenting the image, the model’s reliance on certain attributes is reduced while the sample size is increased, thereby improving the generalization ability and preventing the model from overfitting. Specific methods include vertical flipping, horizontal flipping, image rotation and transposition. In addition, the input data is normalized to increase the convergence speed of the model, improve the accuracy of the model, and avoid the impact of scale differences between different features on model training. The specific steps are as follows:

[0075] First, through image augmentation, the images in the dataset M are rotated by 0°, 90°, 180°, and 270° respectively, and then horizontal or vertical flips are randomly applied to the images to increase the dataset; the augmented images are used as input;

[0076] Second, normalize the input data, and its conversion function is:

[0077] x new =(x-μ) / σ

[0078] Among them, μ is the mean of the sample data, σ is the standard deviation of the sample data, and x new is the image data after normalization.

[0079] Step 3: Learn image representation for the preprocessed image through a self-supervised framework based on contrastive learning. The self-supervised learning framework based on contrastive learning is a method for learning image representation. It uses large-scale data without manual annotation for self-training. This method can learn features with good representation capabilities by comparing image samples with their corresponding transformed versions during the training process. Enhance its image features. The specific steps are as follows:

[0080] First, for the image sequence Where T represents the sequence length, D is the number of sequence layers, H and W are the length and width of a single sequence, and data augmentation operations are performed to generate two different views, denoted as I_1 and I_2.

[0081] Second, a momentum encoder is introduced to provide contrastive supervision for the online encoder to learn discriminative representations. The enhanced full view is used as input to preserve the semantic integrity and discriminability of the learned representation. Since the semantics of image tags are ambiguous, unlike the semantically determined tags in natural language processing, it is logical to use global representations for contrastive learning. The formula is as follows:

[0082]

[0083] in is the input flag of the momentum encoder, represents the momentum encoder, represents the momentum encoder output sequence, Represents the output sequence Features after global pooling.

[0084] Third, the momentum encoder updates parameters through an exponential moving average (EMA), which can promote smooth feature changes for more stable training. and The parameters are denoted as θ s and θ t , then the parameter update formula is as follows:

[0085] θ t ←μθ t +(1-μ)θ s

[0086] Where μ is the moving average hyperparameter.

[0087] Third, in order to align with the output of the momentum encoder, a feature decoder is introduced. The feature decoder has the same structure as the pixel decoder, but it contains fewer transformer blocks and a lighter structure. The two serve different learning objectives and do not share parameters. The feature decoder aims to transform the latent space feature z s Decoded into the momentum encoder feature space, similar to the momentum encoder, the output of the feature decoder is also average pooled as a representation of the entire image, and then the feature is used for contrastive learning, the specific form is as follows:

[0088]

[0089] Where N is the number of all markers in the set, including visible and invisible markers, is the feature decoder output.

[0090] Fourth, for the contrastive learning objective, an asymmetric contrastive loss function is designed based on InfoNCE, which aims to bring semantically consistent samples in a batch closer and push inconsistent samples away. In order to enhance the learning ability of the model, "projection prediction" and "projection" layers are attached to the feature decoder and momentum encoder respectively. Projection head of momentum encoder The parameters are also updated by exponential sliding average. Specifically, the output y of the feature decoder is s It will be converted through the "projection prediction" structure Similarly, the output of the momentum encoder is represented by z t , apply the "projection" head to get Where D p Represents the dimension after projection, and then calculates the cosine similarity ρ between the two:

[0091]

[0092] Next, ρ + It is expressed as the cosine similarity of the positive pair, which is given by and To construct, represents the cosine similarity of the jth negative pair, and is obtained by using the momentum encoder for different images in a batch Construct negative sample pairs, and the loss function is as follows:

[0093]

[0094] Where τ is the temperature coefficient and K represents the batch size.

[0095] Fifth, at the feature level, information transfer in two situations is achieved through self-distillation, and the loss form is as follows:

[0096]

[0097] Where L latent Represents the loss function value.

[0098] Sixth, the loss function of the entire framework can be expressed as:

[0099] L=L r +λ c L c +λ latent L latent

[0100] where λ c represents the contrast loss weight, λ latent represents the consistency loss weight.

[0101] Step 4: For the image (image tile) with enhanced features in step 3, we learn it based on deep learning combined with transfer learning, build a visual model using the U-Net and Residual Network (ResNet) deep learning architecture, automatically annotate the regions of interest in the image, and obtain the image dataset N; provide a preliminary estimate of breast diseases (e.g., lumps and masses). Figure 2 The UNet shown in the figure is a symmetrical image segmentation algorithm based on convolutional neural networks. It was originally designed to solve biomedical image problems. Due to its outstanding effect, it is widely used in various directions of semantic segmentation. Figure 3 As shown in the figure, Res-Net50 is a residual learning framework proposed on the basis of the existing training deep network, which has the advantages of easy optimization and low computational burden. The specific steps are as follows:

[0102] First, the constructed visual models include the classification model ResNet50 and the segmentation model U-Net, which are pre-trained in ImageNet before construction;

[0103] Second, the structure of the U-Net network can be divided into two main parts: the contraction path and the expansion path. In the first part, the contraction path, each block consists of two consecutive 3×3 convolutional layers, followed by a ReLU activation unit and a max-pooling layer. This combination is repeated multiple times to form the architecture of the entire contraction path. The innovation of U-Net lies in the second part, the expansion path. At each stage of the expansion path, the feature map is upsampled by a 2×2 up-convolution. Then, the feature map of the same size is cropped from the corresponding contraction path layer and connected to the upsampled feature map. This is followed by two consecutive 3×3 convolutional layers and ReLU activation. In the final stage, the number of channels of the feature map is reduced to the required number through an additional 1×1 convolution to generate the final segmentation map.

[0104] The energy function of the network is given by:

[0105] E = ∑ω(x)log(p k(x) (x)

[0106] Among them, p k is the pixel-level SoftMax function applied to the final feature map, defined as:

[0107]

[0108] Among them, a k represents the activation k in the channel.

[0109] Third, ResNet50 uses the output of the previous layer directly across several layers as a shortcut link for the input of a subsequent layer. Its input function is:

[0110] F(x)=H(x)-x

[0111] Among them, F(x) is the residual, H(x) is the complete output of the previous layer, and x is the input of the previous layer.

[0112] During the training process, the loss function will swing after the parameters are updated. The network optimizer can optimize the swing amplitude of the loss function, which can be achieved by using the root mean square transfer algorithm (RMSProp), and the convergence speed of the function will be further accelerated. The learning rate is a hyperparameter that adjusts the network weights through the gradient of the loss function, and the interval adjustment learning rate (stepLR) decay strategy is adopted. The learning rate is set to 0.001, the maximum iteration step is set to 5000, and the learning rate is decayed by 1 / 2 at 2000 and 4000 steps. Since the early layers of CNN have little correlation with the specific image dataset, the relationship with the selected dataset and its task objectives becomes closer as time goes by. Therefore, fine-tuning starts from the last layer of ResNet50, and then gradually includes more layers in the update process.

[0113] Step 5: Manually annotate the breast images that were automatically annotated in step 4 again, achieving human-machine dual annotation, and obtaining an updated image dataset M, which provides a more accurate estimate of breast diseases while also reducing the workload of doctors;

[0114] After the breast images are initially labeled using automatic labeling technology (such as visual models based on U-Net and ResNet50), multiple trained human experts use specialized labeling tools or software to perform manual labeling and update the labeled data to the labeled dataset M.

[0115] Step 6: Use contrastive learning algorithm to learn image representation from unlabeled dataset N, extract additional supervisory signals from labeled dataset M, such as disease labels or segmentation masks, use contrastive learning model to learn feature representation, and combine supervisory signals to update visual model, improve model feasibility, and further tune the model.

[0116] First, use the contrastive learning model to learn image feature representation from the unlabeled dataset N, which is the same as step 3;

[0117] Second, extract additional supervisory signals, such as disease labels or segmentation masks, from the updated annotated dataset M, and jointly train and then update the visual model to better understand and identify breast diseases;

[0118] Third, the contrastive learning task is combined with the supervision signal to form a multi-task loss function as follows:

[0119] L=L c +λ s ·L s

[0120] Among them, L c is the contrast loss, L s is the supervision loss, λ s is the weight of the supervision loss.

[0121] Fourth, use the multi-task loss function to train the visual model and update the parameters of the contrastive learning and supervised learning modules at the same time. By minimizing the multi-task loss function, the visual model can simultaneously learn meaningful feature representations and task-specific knowledge, thereby improving its generalization ability and accuracy. The formula for minimizing the multi-task loss function is as follows:

[0122] L_multi=α*L_cls+β*L_seg

[0123] Among them, L_cls is the loss function of the disease classification task, L_seg is the loss function of the segmentation task, and α and β are their corresponding weight coefficients.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A human-machine collaborative double labeling method, characterized in that: The following steps are included: Step 1: Generate images using an automatic imaging system and divide the images into an unlabeled dataset N and a labeled dataset M; Step 2: preprocess the labeled data set M; By augmenting the image, the data set is enlarged and the input data is normalized; Step 3: The preprocessed image is subjected to image representation learning through a self-supervised framework based on contrastive learning to enhance its image features; Step 4: For the image enhanced with features in step 3, learn it based on deep learning combined with transfer learning, build a visual model using the U-Net and residual network deep learning architecture, automatically annotate the region of interest in the image, and obtain the image dataset M; Step 5: manually annotate the images that were automatically annotated in step 4 again to obtain an updated image dataset M; Step 6: Use the contrastive learning algorithm to learn image representation from the unlabeled dataset N, extract additional supervisory signals from the labeled dataset M, use the contrastive learning model to learn feature representation, and update the visual model in combination with the supervisory signals.

2. A human-machine collaborative double labeling method according to claim 1, characterized in that: The step 1 includes the following specific steps when using automatic breast volume ultrasound imaging: First, apply the ultrasound coupling agent evenly on one breast, cover the scanner with a disposable coupling film, and the operator places the scanner on one breast and presses gently to ensure contact with the breast skin; Second, start scanning, repeat three times on each breast to form a complete view of the entire breast, the scanning time in one direction is 60 seconds, and the layer spacing of the acquired image is 0.5mm; Third, after scanning, the data is transferred to the workstation to obtain cross-sectional, coronal, and sagittal images, and the acquired images are divided into an unlabeled dataset N and a labeled dataset M.

3. The human-machine collaborative double labeling method according to claim 1, characterized in that: The step 2 includes the following specific steps: First, through image augmentation, the images in the dataset M are rotated by 0°, 90°, 180°, and 270° respectively, and then horizontal or vertical flips are randomly applied to the images to increase the dataset, and the augmented images are used as input; Second, normalize the input data.

4. The human-machine collaborative double labeling method according to claim 1, characterized in that: The step three includes the following specific steps: First, data augmentation is performed on the image sequence to generate two different views, denoted as I_1 and I_2; Second, a momentum encoder is introduced to provide contrastive supervision for the online encoder to learn discriminative representations, and the augmented full view is used as input; Third, a feature decoder is introduced to perform average pooling calculation as the representation of the entire image. Fourth, attach the "projection prediction" and "projection" layers to the feature decoder and momentum encoder respectively; Fifth, information transfer between the two situations is achieved through self-distillation at the feature level.

5. The human-machine collaborative double labeling method according to claim 1, characterized in that: The step 4 includes the following specific steps: First, the constructed visual models include the classification model ResNet50 and the segmentation model U-Net, which are pre-trained in ImageNet before construction; Second, the structure of the U-Net network can be divided into two main parts: the contraction path and the expansion path. Each block in the contraction path consists of two consecutive 3×3 convolutional layers, followed by a ReLU activation unit and a max-pooling layer. At each stage of the expansion path, the feature map is upsampled by a 2×2 up-convolution. Then, the feature map of the same size is cropped from the corresponding contraction path layer and connected to the upsampled feature map. This is followed by two consecutive 3×3 convolutional layers and ReLU activation. In the final stage, the number of channels of the feature map is reduced to the required number through an additional 1×1 convolution to generate the final segmentation image; Third, ResNet50 uses a shortcut connection method to directly connect the output of the previous layer across several layers as the input of a subsequent layer.

6. The human-machine collaborative double labeling method according to claim 1, characterized in that: The step five includes the following specific steps: After the automatic annotation technology has preliminarily annotated the images, a number of trained human experts use special annotation tools or software to perform manual annotation and update the annotated data to the annotation dataset M.

7. The human-machine collaborative double labeling method according to claim 4, characterized in that: The step six includes the following specific steps: First, use the contrastive learning model to learn image feature representation from the unlabeled dataset N, which is the same as step 3; Second, extract additional supervisory signals from the updated labeled dataset M; Third, the contrastive learning task is combined with the supervision signal to form a multi-task loss function; Fourth, the visual model is trained using a multi-task loss function, while updating the parameters of the contrastive learning and supervised learning modules.

Images