A semi-supervised image segmentation method and system based on contrastive learning-assisted dual-region balanced consistency learning

Abstract

The present invention discloses a semi-supervised image segmentation method and system of dual-region balanced consistency learning assisted by contrastive learning. During training, an image segmentation network architecture is constructed, including a student model, a teacher model, a dual-region balanced consistency learning strategy module, a contrastive learning module using hard negative samples, and an EMA module; the dual-region balanced consistency learning strategy is to divide according to uncertainty, and balance the model's attention to different regions through loss weights to improve the overall effect of consistency learning. The contrastive learning module of hard negative samples can extract class samples from the segmentation results of consistency learning, learn feature representations of different classes, and use hard negative samples to further improve inter-class contrast and intra-class consistency in the segmentation results. The present invention balances the model's attention to high-uncertainty regions and low-uncertainty regions, as well as the problem of insufficient class separability in the segmentation results, thereby improving the segmentation accuracy of medical images.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and in particular relates to a semi-supervised image segmentation method and system thereof using contrastive learning-assisted dual-region balanced consistency learning. Background Art

[0002] The development of modern medical imaging technology has enabled doctors to obtain detailed images of the human body's internal anatomy, tissues, and organs through a variety of imaging methods and equipment, enabling diagnosis, treatment, and disease monitoring. Computer-assisted analysis of medical images is playing an increasingly significant role in clinical diagnosis and treatment.

[0003] Medical image segmentation aims to determine the category of each pixel in the image and to separate the required anatomical structures, such as organs and tumors, from the image. Accurate segmentation can provide more reliable anatomical information, reflect the pathological characteristics of specific diseases, and is also an important step in computer-assisted disease diagnosis and subsequent treatment planning. Medical image segmentation methods based on deep learning have made great technological progress in recent years, especially in the identification of lesions and anatomical structures. Medical image segmentation methods based on deep learning usually take the form of fully supervised learning, which means that the model requires a large amount of labeled data for training. Semi-supervised learning can learn information from a large amount of unlabeled data, and then combine it with a small amount of labeled data to train and optimize the segmentation model, achieving a segmentation effect close to that of a fully supervised model. Therefore, semi-supervised segmentation methods have become an effective way to alleviate the problem of insufficient labeled data in medical image segmentation.

[0004] In the field of semi-supervised medical image segmentation, consistency regularization is a mainstream method, primarily improving upon the Mean-Teacher (MT) method. This method consists of two models: a student model and a teacher model. Labeled images are fed into the student model, and after obtaining the segmentation results, the true labels of the images are used for supervision. Unlabeled images are subjected to various random perturbations, such as Gaussian noise, and then fed into the student and teacher models, respectively. The predicted segmentation results are then calculated based on the consistency loss. The student model backpropagates through the consistency loss and the supervision loss, with the gradient updating the model parameters. The teacher model parameters are updated using the exponential moving average (EMA) of the student model parameters.

[0005] However, there are still some shortcomings in the current Mean-Teacher-based improved methods: (1) The basic MT model can improve the model effect when only high-uncertainty regions or low-uncertainty regions are selected to calculate the consistency loss, among which the effect is better when only high-uncertainty regions are selected for consistency learning. This shows that it is meaningful to perform consistency learning in both regions separately. Therefore, how to balance the model's attention to high-uncertainty regions and low-uncertainty regions to further improve the overall effect of consistency learning has become a challenge. (2) The method based on consistency learning only calculates the MSE loss between predictions under different perturbations. The MSE loss only calculates the difference between voxels and is not used to perceive class information. Therefore, the semi-supervised model implemented only using consistency learning cannot guarantee the class separability in the segmentation results.

[0006] In summary, the current consistency method will cause the model to be unable to fully utilize the information of unlabeled images, thereby affecting the segmentation accuracy of the model. Summary of the Invention

[0007] The first purpose of the present invention is to address the problems existing in the prior art and propose a contrastive learning-assisted dual-region balanced consistency learning semi-supervised medical image segmentation method and system, aiming to solve the problem of how to balance the model's attention to high-uncertainty areas and low-uncertainty areas, and the problem of insufficient class separability in the segmentation results, thereby improving the segmentation accuracy of medical images.

[0008] In a first aspect, the present invention provides a semi-supervised image segmentation method using contrastive learning-assisted dual-region balanced consistency learning, which specifically comprises the following steps:

[0009] Step 1: Obtain medical images, including labeled images The corresponding unlabeled image ;

[0010] Step 2: Preprocess the medical images and construct the data set;

[0011] Step 3: Build an image segmentation network architecture and train it using the dataset;

[0012] The image segmentation network includes a student model , teacher model , dual-region balanced consistency learning strategy, contrastive learning module using hard negative samples, EMA module;

[0013] During the training process, the student model For the preprocessed label image and unlabeled images Process them separately to get the corresponding first prediction segmentation results and the second predicted segmentation result ; Wherein the subscript i represents the i-th image;

[0014] The teacher model For the preprocessed unlabeled images Processing is performed to obtain the third predicted segmentation result ;

[0015] The implementation method of the dual-zone balance consistency learning strategy is as follows:

[0016] 1) For unlabeled images, we need to measure the uncertainty in the prediction results. Considering that information entropy reflects the uncertainty of the prediction results, high entropy means higher uncertainty. Therefore, we use normalized entropy to quantify the uncertainty of the prediction results.

[0017] Model for students Output of the second predicted segmentation result Calculate the information entropy value of each voxel in the image ;

[0018]

[0019] in ne v ∈ [0, 1] ； p i c Represents the second predicted segmentation result The predicted probability map of class c; Represents a set of segmented object categories; Indicates the number of segmentation object categories.

[0020] According to the information entropy value of each voxel , filter out Greater than threshold high entropy region ;

[0021] Threshold It can gradually increase with the number of iterations, and its update can be seen in:

[0022]

[0023] Where t represents the current iteration number, ttotal represents the total number of iterations of model training, h represents the initial threshold, and r represents the threshold growth rate; Represents the growth coefficient.

[0024]

[0025] use Calculate the second predicted segmentation result And the third predicted segmentation result Consistency loss between high entropy regions ;

[0026]

[0027] in and They represent the predictions on the vth voxel in the second predicted segmentation result and the third predicted segmentation result respectively.

[0028] use Calculate the second predicted segmentation result And the third predicted segmentation result Consistency loss between low entropy regions ;

[0029]

[0030] In order to make the model focus more on learning the knowledge of high entropy areas, a larger weight is set for its loss, and a smaller weight is selected for the consistency loss of low entropy areas. and consistency loss in low entropy regions , calculate the total loss of consistency learning ;

[0031]

[0032] in 、 Respectively 、 The weight of > >0.

[0033] 2) For the label image, use the student model to first predict the segmentation result and the true label Calculating supervised loss ;

[0034]

[0035] in represents the cross entropy loss, represents Dice loss;

[0036] The contrastive learning module using hard negative samples includes the following steps:

[0037] 1) Create a storage container , used to store positive and negative samples;

[0038] 2) Set anchor samples;

[0039] Using feature mapping model The second predicted segmentation result of the current iteration Prediction probability plot for different classes in Mapped into the feature space, we get class samples of different classes , and As anchor samples;

[0040]

[0041] in, Represents the class sample of the c-th segmentation object;

[0042] because represents the category of the segmented object, represents the segmentation object class, Represents the background class, we can get { , }, represents the class sample of the segmented object class, Represents the class sample of the background class;

[0043] 3) Obtain positive sample pairs;

[0044] Since in each iteration, Only one class sample corresponding to each class can be obtained , cannot provide multiple samples of the same category; therefore, use To store the unlabeled images in the previous iteration Class sample As a positive sample, it can be compared with the anchor sample Forming positive sample pairs ;

[0045] 4) Obtain hard negative sample pairs;

[0046] If the consistency loss of the high entropy region in the current iteration is higher than the consistency loss of the high entropy region in the previous iteration, it indicates that the second predicted segmentation result of the unlabeled image And the third predicted segmentation result There is a large divergence between the student model and the teacher model, the feature mapping model The corresponding generated class samples of different classes from the anchor samples are regarded as hard negative samples Store in storage container In; if Already exists in , then the current Replace the original and store it in a storage container middle;

[0047] During training, from the storage container Remove hard negative samples With anchor sample Construct hard negative sample pairs For contrastive learning;

[0048] 5) Use positive and negative sample pairs for comparative learning;

[0049] Combined with the infoNCE loss in contrastive learning to calculate the loss value, the learned class information is incorporated into the model training; contrastive learning loss The calculation is as follows:

[0050]

[0051] in represents the anchor point sample; Indicates that the sample obtained from the i-1th image and the anchor sample belong to the same category, that is, the positive sample; Indicates from The hard negative samples taken from the anchor point are of different classes; τ is the temperature coefficient; Indicates that if the anchor sample is of category C, samples of other categories except category C are taken from {C};

[0052] The total loss of consistency learning calculated according to the dual-region balanced consistency learning strategy and monitoring losses , and the contrastive learning module calculates the contrastive learning loss , calculate the total loss ,After back-propagation, the gradient updates the parameters of the student model;

[0053]

[0054] The parameters of the student model update the parameters of the teacher model through the EMA module;

[0055]

[0056] represents the parameters of the teacher model in the last round of training, Represents the parameters of the student model during this round of training, represents weight;

[0057] Step 4: Use the student model trained in step 3 , realizing medical image segmentation.

[0058] In a second aspect, the present invention provides a semi-supervised image segmentation system, comprising:

[0059] A data acquisition module, used to acquire medical image data with segmented targets;

[0060] Data preprocessing module, used to preprocess medical image data with segmentation targets;

[0061] Image segmentation module, using the trained student model The preprocessed medical image data is segmented to obtain the image segmentation result.

[0062] The beneficial effects of the present invention are as follows:

[0063] 1. The present invention addresses the problem that the existing medical image segmentation methods do not balance the model's attention to different deterministic regions, resulting in low segmentation accuracy. During the model training process, the information entropy of the student model's prediction results is calculated, and the results are divided into two regions with different information content by a threshold. A larger loss weight is set for the region with information entropy higher than the threshold, so that the model focuses more on learning this part of the region. This approach helps to improve the accuracy of medical image segmentation.

[0064] 2. To address the problem of insufficient class separability in the segmentation results of the current semi-supervised methods based only on consistency learning, the present invention converts the entire prediction probability map of different classes of the student model into class samples, and combines the results of consistency learning to selectively select samples that can effectively play the role of contrast learning for learning, thereby improving the accuracy of medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a schematic diagram of the semi-supervised medical image segmentation network structure of the contrastive learning-assisted dual-region balanced consistency learning in the present invention.

[0066] Figure 2 This is the effect diagram of segmentation of 10% label image and 20% label image in the present invention. DETAILED DESCRIPTION

[0067] The present invention will be further explained below with reference to specific embodiments.

[0068] The present invention provides a semi-supervised medical image segmentation method of dual-region balanced consistency learning assisted by contrastive learning. The proposed dual-region balanced consistency learning strategy can focus on the learning of high-entropy areas while not ignoring the low-entropy areas in the image, balancing the model's attention to the two regions and improving the overall effect of consistency learning. In addition, a contrastive learning module using hard negative samples is proposed, which can combine consistency learning to select samples with better contrastive learning effects and improve the class separability in the segmentation results. By constructing an overall model in which consistency learning and contrastive learning complement each other, the segmentation results of medical images are effectively improved. Experimental results show that in the 20% labeled image experiment of the LA dataset, the Dice coefficient of the present invention reached 91.26%; in the 10% labeled experiment, the Dice coefficient reached 89.50%, exceeding other comparison methods, proving its effectiveness and reliability.

[0069] The method is as Figure 1 Specifically include the following:

[0070] Step 1: Obtain MR images of the heart from the left atrium segmentation challenge public dataset (LA), and divide the dataset into labeled data with segmentation object labels and unlabeled data in proportion.

[0071] Step 2: Preprocess the MR images in the dataset. The preprocessing method includes random flipping of the 3D images and cropping them to a uniform size.

[0072] Step 3: Build Figure 1 The semi-supervised image segmentation network shown is trained using preprocessed MR image data.

[0073] The semi-supervised image segmentation network includes a student model , teacher model , dual-region balanced consistency learning strategy (DRBCL), contrastive learning module using hard negative samples (CLHNS), EMA module;

[0074] Student Model , teacher model , both have the same network structure. In this embodiment, both adopt the V-Net model to predict the segmentation results of the input image.

[0075] During the training process, the student model For the preprocessed label image and unlabeled images Process them separately to get the first predicted segmentation result and the second predicted segmentation result ; Wherein the subscript i represents the i-th image;

[0076] The teacher model For the preprocessed unlabeled images Processing is performed to obtain the third predicted segmentation result ;

[0077] For the label image, the student model is used to calculate the supervision loss between its prediction result and the true label. The supervision loss formula is as follows:

[0078] (1)

[0079] in represents the cross entropy loss, Denotes Dice loss.

[0080] For unlabeled images, we first calculate the normalized entropy of the unlabeled image prediction results output by the student model. The normalized entropy formula used is as follows:

[0081] ne v =− ∑ c ∈ C p i c log( p i c ) / log( C ) ∈ [0, 1] (2)

[0082] Represents the segmentation result of the cth category of the segmented object in the image. It represents the information entropy value of each voxel in the 3D image.

[0083] In order to dynamically adjust the two partitions according to the learning situation of the model, a threshold is set as the basis for partitioning. The threshold formula used is as follows:

[0084] (3)

[0085] Among them, t represents the current iteration number, t total Indicates the total number of iterations for model training.

[0086] Passing this threshold Select the mask of high entropy area:

[0087] (4)

[0088] The consistency loss is calculated with the help of mask. Formula (5) and Formula (6) are the consistency losses of high entropy region and low entropy region respectively. Formula (7) is the total loss of consistency learning:

[0089] (5)

[0090] (6)

[0091] (7)

[0092] The consistency loss is calculated as the mean square error between the prediction results of the student model and the teacher model for the unlabeled image, where β is set to 0.3. It is set to a smaller value of 0.06. The purpose is to make the model pay more attention to the information in the high entropy area during training.

[0093] The obtained supervision loss and consistency loss will be used to update the parameters of the subsequent student model.

[0094] The contrastive learning module using hard negative samples is intended to improve the class separability in the segmentation results and further improve the segmentation accuracy of the model.

[0095] The storage container Memery Bank shown in the figure is used to store positive and negative samples. Feature Mapping Model To predict the results Prediction probability plot for different classes in Mapped into the feature space, we get class samples of different classes:

[0096] (8)

[0097] in, Represents class samples of different segmentation objects, Represents the categories of different segmentation objects, represents the segmentation object class, Represents the background class. Representation feature mapping model , mainly composed of convolutional layers, pooling layers and linear layers.

[0098] First, we need to set the anchor samples. Input into the student model can get the predicted probability map of different classes .Will Input feature map network Can get class samples ( , ), and use them as anchor samples.

[0099] Secondly, obtain positive samples. Since in each iteration, Only one class sample corresponding to each class can be obtained , cannot provide multiple samples of the same category. Therefore, we use To store the unlabeled images in the previous iteration Class sample , which can be compared with the anchor sample Forming positive sample pairs .

[0100] Then, get hard negative samples. Current unlabeled image The obtained class samples are used as anchor samples. Other unlabeled images other than the anchor point can also generate class samples. Then, we can select negative samples from other images that are different from the anchor point samples. In order to improve the model's ability to distinguish features of different categories, we select negative samples that are more conducive to improving the contrastive learning effect from negative samples generated by other unlabeled images. These negative samples are called hard negative samples. We select hard negative samples based on the consistency learning situation, when the consistency loss value of the high entropy region is higher than the consistency loss value in the previous iteration ( ), indicating that the prediction results of the unlabeled image have a large difference between the student model and the teacher model, and the corresponding generated samples of different classes from the anchor samples are regarded as hard negative samples Deposit If Already exists in Then we will use the current Replace the original and store it in During the training process, Remove hard negative samples With anchor sample Construct hard negative sample pairs Used for comparative learning.

[0101] Finally, we use the constructed positive and negative sample pairs for contrastive learning. We combine the infoNCE loss in contrastive learning to calculate the loss value and incorporate the learned class information into model training. The loss function used in the contrastive learning module is as follows:

[0102] (9)

[0103] Represents the anchor sample, which is the probability map of the cth class in the prediction result of the i-th image, and the class sample obtained by the feature mapping network. It represents the sample obtained from the i-1th image that belongs to the same category as the anchor sample, that is, the positive sample. Indicates from The negative samples from τ are of different classes from the anchor samples. τ is the temperature coefficient, which is set to 1.

[0104] Finally, the model parameters are updated by the total loss. The formula for the total loss is as follows:

[0105] (10)

[0106] The value of λ is 0.2. Back propagation is performed based on the total loss value. The parameters of the student model are updated by the gradient. The parameters of the teacher model are updated by the exponential moving average (EMA) of the student model parameters. The update method is as follows:

[0107] (11)

[0108] represents the parameters of the teacher model in the last round of training, Indicates the parameters of the student model during this round of training. The value of α is 0.99.

[0109] To verify the effectiveness of the segmentation method proposed in this paper, four evaluation indicators commonly used in medical image segmentation were selected to evaluate the accuracy of medical image segmentation: Dice coefficient (Dice), Jaccard similarity coefficient (Jaccard), 95% Hausdorff distance (95HD), and average surface distance (ASD). The expressions of Dice, Jaccard, 95HD, and ASD are defined as shown in formulas (11) to (14):

[0110] (12)

[0111] (13)

[0112] (14)

[0113] (15)

[0114] in, , .

[0115] To verify the effectiveness of the present invention, the method of the present invention is compared with a variety of existing semi-supervised medical image segmentation methods, such as Mean-Teacher (MT), UA-MT, SASSNet, CPS, URPC, MC-Net+, SCC, AC-MT, MCF and DAE-MT. The experiment is trained and tested based on the LA public dataset.

[0116] Table 1 Results of various methods and full supervision when the labeled images account for 10% and 20% in the LA dataset.

[0117]

[0118] The experimental results are shown in Table 1. The proposed dual-region balanced consistency learning strategy enables the model to focus on different regions in a balanced manner and fully learn the information in different regions. The contrastive learning module using hard negative samples can improve the class separability in the segmentation results and effectively improve the segmentation effect of the model. When the labeled data accounts for 10% and 20%, the final Dice coefficient results of the present invention are 89.50% and 91.26% respectively, which are better than the above methods. The segmentation effect of the LA dataset is shown in Figure 1. Figure 2 shown.

[0119] To validate the effectiveness of the dual-region balanced consistency learning strategy and the contrastive learning module utilizing hard negative samples, we conducted an ablation study. Using the Mean-Teacher method as the baseline, experiments were conducted on the LA dataset using 20% ​​labeled data and 80% unlabeled data. The experimental results are detailed in Table 2. As can be seen, incorporating the partition consistency learning strategy into the baseline improves the Dice score by 1.87 percentage points, as it enables the model to focus on different regions with greater focus. When the baseline is combined with the contrastive learning module, it supplements the learned class information with consistency learning, resulting in a 0.31 percentage point improvement in the Dice score over the baseline. When the partition consistency learning strategy and the contrastive learning module are combined, the contrastive learning module selects more effective samples based on the consistency learned, resulting in a 2.35 percentage point improvement in the Dice score. Overall, integrating the partition consistency learning strategy and the contrastive learning module into the baseline improves the accuracy of semi-supervised segmentation, demonstrating their effectiveness in enhancing model performance.

[0120] Table 2 Ablation experiment results, the proportion of labeled data is 20%

[0121]

[0122] In order to evaluate the advantages of the positive and negative samples sampled by our proposed method in terms of contrastive learning effects, we conducted ablation experiments using two different methods: obtaining positive and negative samples from the prediction results of labeled images and unlabeled images. The results are shown in Table 3. Exp 1: We use positive and negative samples obtained from the prediction results of labeled images for contrastive learning. The Dice score is only 91.03%. Due to the limited number of labeled images, it is impossible to learn more diverse class information from the labeled images. Exp 2: Due to the The class samples obtained in are only used as anchor samples and cannot provide redundant class samples, so we We obtained positive and negative samples from the prediction results of the previous unlabeled image and conducted experiments. The Dice score was only 90.95%. Exp 3: We selected hard negative samples to improve the model's ability to learn features. We selected positive samples from the prediction results of the previous unlabeled image and negative samples from the prediction results of the unlabeled image with the highest consistency loss in the high entropy region to obtain the best results. The data in Table 3 shows the importance of positive and negative sample sampling methods in contrastive learning and the effectiveness of our proposed positive and negative sample sampling methods.

[0123] Table 3 Ablation experiments on sampling methods, with labeled data accounting for 20%

[0124]

[0125] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A semi-supervised image segmentation method based on contrastive learning-assisted dual-region balanced consistency learning, characterized by The method comprises the following steps: Step 1: Obtain medical images, including labeled images The corresponding unlabeled image ; Step 2: Preprocess the medical images and construct the data set; Step 3: Build an image segmentation network architecture and train it using the dataset; The image segmentation network architecture includes a student model , teacher model , dual-region balanced consistency learning strategy, contrastive learning module using hard negative samples, EMA module; During the training process, the student model For the preprocessed label image and unlabeled images Process them separately to get the corresponding first prediction segmentation results and the second predicted segmentation result ; Wherein the subscript i represents the i-th image; The teacher model For the preprocessed unlabeled images Processing is performed to obtain the third predicted segmentation result ; The implementation method of the dual-zone balance consistency learning strategy is as follows: For unlabeled images, the student model Output of the second predicted segmentation result Calculate the information entropy value of each voxel in the image ; According to the information entropy value of each voxel , filter out Greater than threshold high entropy region ; use Calculate the second predicted segmentation result And the third predicted segmentation result Consistency loss between high entropy regions ;use Calculate the second predicted segmentation result And the third predicted segmentation result Consistency loss between low entropy regions ; According to the consistency loss of high entropy region and consistency loss in low entropy regions , calculate the total loss of consistency learning ; For the label image, use the student model to first predict the segmentation result and the true label Calculating supervised loss ; The contrastive learning module using hard negative samples is based on anchor samples. , obtain positive sample pairs and hard negative sample pairs; use positive sample pairs and negative sample pairs for contrastive learning to obtain contrastive learning loss ; The total loss of consistency learning calculated according to the dual-region balanced consistency learning strategy during training and monitoring losses , and the contrastive learning module calculates the contrastive learning loss , calculate the total loss , after back propagation, the gradient updates the student model Parameters; The EMA module is based on the Student model Update the teacher model with the updated parameters Parameters; Step 4: Use the student model trained in step 3 , realizing medical image segmentation.

2. The method according to claim 1, characterized in that The student model and teacher model Use V-Net model.

3. The method according to claim 1, characterized in that The information entropy value of each voxel in the dual-region balance consistency learning strategy The calculation of is as follows: (1) in Represents the second predicted segmentation result The predicted probability map of class c; Represents a set of segmented object categories; Indicates the number of segmentation object categories.

4. The method according to claim 1, characterized in that The threshold in the dual-region balance consistency learning strategy The updates are as follows: (2) Among them, t represents the current iteration number, t total represents the total number of iterations of model training, h represents the initial threshold, and r represents the threshold growth rate; Represents the growth coefficient.

5. The method according to claim 1, characterized in that The consistency loss in the high entropy region in the dual-region balanced consistency learning strategy The calculation of is as follows: (3) in and Respectively represent the predictions on the vth voxel in the second predicted segmentation result and the third predicted segmentation result; Consistency loss in low entropy regions The calculation of is as follows: (4) Total loss for consistency learning The calculation of is as follows; (5) in 、 Respectively 、 The weight of > >

0.

6. The method according to claim 1, characterized in that The supervised loss in the dual-region balanced consistency learning strategy The calculation of is as follows: (6) in represents the cross entropy loss, represents the Dice loss, represents the true label.

7. The method according to claim 1, characterized in that: The process of obtaining anchor samples in the contrastive learning module using hard negative samples is as follows: Using feature mapping model The second predicted segmentation result of the current iteration Prediction probability plot for different classes in Mapped into the feature space, we get class samples of different classes , and As anchor samples; (7) in, Represents the class sample of the c-th segmentation object; because represents the category of the segmented object, represents the segmentation object class, Represents the background class, we can get { , }, represents the class sample of the segmented object class, Represents the class sample of the background class; The process of obtaining positive sample pairs is as follows: Since in each iteration, Only one class sample corresponding to each class can be obtained , cannot provide multiple samples of the same category; therefore, use To store the unlabeled images in the previous iteration Class sample As a positive sample, it can be compared with the anchor sample Forming positive sample pairs ; The process of obtaining hard negative sample pairs is as follows: If the consistency loss of the high entropy region in the current iteration is higher than the consistency loss of the high entropy region in the previous iteration, it indicates that the second predicted segmentation result of the unlabeled image And the third predicted segmentation result There is a large divergence between the student model and the teacher model, the feature mapping model The corresponding generated class samples of different classes from the anchor samples are regarded as hard negative samples Store in storage container In; if Already exists in , then the current Replace the original and store it in a storage container middle; During training, from the storage container Remove hard negative samples With anchor sample Construct hard negative sample pairs Used for comparative learning.

8. The method according to claim 1, characterized in that: The contrastive learning loss in the contrastive learning module using hard negative samples The calculation is as follows: (8) in represents the anchor point sample; Indicates that the sample obtained from the i-1th image belongs to the same category as the anchor sample, that is, the positive sample; Indicates from The hard negative samples taken from the anchor point are of different classes; τ is the temperature coefficient; It means that if the anchor sample is of category C, samples of other categories except category C are taken from {C}.

9. The method according to claim 1, characterized in that: Total loss The calculation of is as follows: (9) in Represents weight.

10. A semi-supervised image segmentation system implementing the method according to any one of claims 1 to 9, characterized in that include: A data acquisition module, used to acquire medical image data with segmented targets; Data preprocessing module, used to preprocess medical image data with segmentation targets; Image segmentation module, using the trained student model The preprocessed medical image data is segmented to obtain the image segmentation result.

Images