A semi-supervised medical image segmentation method based on clustering fusion cross learning

Abstract

The present invention discloses a semi-supervised medical image segmentation method based on cluster fusion cross learning. CFCT adopts a training mechanism similar to CPS. Two sub-networks are used to generate pseudo labels for cross-training learning to strengthen the consistency between the two sub-networks. A Mask Transformer network based on cluster fusion is used to replace the traditional pixel-based segmentation network. CFN introduces a target query vector into the basic architecture of the original convolutional neural network backbone. The target query vector is updated by cross-attention calculation with the multi-scale pixel features output from the CNN backbone network decoder. The updated target query vector is regarded as a cluster center. The cluster center obtains the affinity between different clusters and pixels through cross-attention calculation. The calculation process is similar to the classic k-means clustering algorithm. Finally, the updated cluster center is merged with the image pixel features output by the backbone network through a segmentation head to obtain a predicted segmentation result.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a semi-supervised medical image segmentation method based on clustering fusion cross-learning. Background Art

[0002] Medical image segmentation is a basic and critical task in computer-aided diagnosis, which aims to accurately identify the target organ, tissue or lesion area at the pixel level. In the field of computer-aided diagnosis (CAD), automatic medical image segmentation technology based on deep learning has been widely studied. Although deep neural networks have shown significant progress compared with traditional algorithms, they require a large amount of labeled data as support during the training process. In practical applications, the semantics of medical images are complex and the labeling process is time-consuming and labor-intensive, which limits the further development of deep learning algorithms in this field. It is relatively easy to directly obtain a large amount of medical imaging data. Therefore, in order to reduce the dependence on labeling and reduce costs, semi-supervised learning algorithms have received widespread attention and research.

[0003] Semi-supervised image segmentation methods use a small amount of labeled data and a large amount of unlabeled data to enhance the accuracy and generalization ability of the segmentation model. The main challenge is how to effectively make full use of the information of unlabeled data. As a typical method of semi-supervised learning, the consistency regularization-based method perturbs the image or network features to enhance the consistency of prediction under different perturbations. A common method is to use networks with different structures or different parameters for cross-training and learning. The representative method is the cross pseudo-label supervision (CPS) method. CPS inputs labeled and unlabeled images into two sub-networks with the same structure but initialized with different parameters. For labeled data, the corresponding true segmentation label is used to supervise the output result; for unlabeled data, the prediction results of the two sub-networks are used as pseudo-labels to supervise the other sub-network, thereby strengthening the consistency between the two sub-networks. However, in these methods, the two sub-networks operate independently and only rely on their final semantic segmentation results as pseudo-labels. The lack of interaction between the networks and the constraints and guidance in the whole process may lead to poor segmentation performance.

[0004] In response to the above problems, the present invention proposes a cross-learning (CFCT) mechanism based on cluster fusion, which replaces the traditional pixel-based segmentation network with a Mask Transformer network based on clustering assumptions, and adopts the same cross-training method as CPS to strengthen the consistency between the two sub-networks. From the perspective of clustering, image segmentation is equivalent to assigning pixels to different cluster centers. Common difficulties in medical image segmentation include irregular tumor or organ target shapes, fuzzy boundaries, and identical intensity distributions. Traditional convolutional neural networks often perform poorly in the face of the above difficulties. The cluster fusion network can better solve these difficulties by assigning easily confused areas to corresponding cluster centers, and can perform more accurate semantic category segmentation with the cooperation of global semantic information. In addition, the present invention introduces a training method based on cluster center exchange on the basis of the cluster fusion network, which further strengthens the consistency by exchanging cluster centers between sub-networks, thereby improving the robustness and fast convergence ability of semi-supervised learning. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned problems and to provide a semi-supervised medical image segmentation method based on clustering fusion cross-learning.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows: CFCT adopts the same training mechanism as CPS, generates pseudo labels through two sub-networks for cross-training learning to strengthen the consistency between the two sub-networks, CFCT uses a Mask Transformer network based on clustering fusion to replace the traditional pixel-based segmentation network, CFN introduces a target query vector in the basic architecture of the original convolutional neural network backbone, and updates the target query vector by cross-attention calculation with the multi-scale pixel features output from the CNN backbone network decoder, and regards the updated target query vector as the cluster center. The cluster center obtains the similarity between different clusters and pixels through cross-attention calculation, and finally, the updated cluster center is merged with the image pixel features output by the backbone network through a segmentation head to obtain the predicted segmentation result.

[0007] Furthermore, the backbone network adopts the classic U-shaped network structure, the pixel encoder uses ResNet50, and a pixel decoder composed of residual modules is designed to process the upsampled features.

[0008] Furthermore, the clustering Transformer decoder adopts a dual attention mechanism, including cross-attention between cluster centers and pixel features and self-attention of cluster centers, as follows: first, a parameter-learnable feature vector is introduced as the clustering target query vector, and then the query vector and the multi-scale pixel features output from the backbone network decoder are cross-attention calculated respectively, and the cluster-based argmax function is used to replace the traditional softmax activation function to obtain the attention mapping matrix, and then the relevant pixel features are weighted by the attention mapping matrix to update the cluster center. The calculation process of cross-attention is expressed as follows:

[0009]

[0010] Where C∈R N×D represents the cluster center with N channels and D channels, represents the updated cluster center, Q c ∈R N×D represents the query matrix mapped from the clustering target query vector, K p ∈R HW×D and V p ∈R HW×D They represent the key and value matrices mapped from pixel features of size D×H×W respectively.

[0011] Furthermore, the segmentation head is set to further merge the image pixel features output by the backbone network with the updated cluster centers to obtain the final segmentation prediction result, as follows: First, the product of the cluster center and the image pixel features is calculated to obtain the cluster assignment feature (CA∈R N×H×W ), which represents the similarity between each pixel and the cluster center, and the expression is as follows:

[0012]

[0013] Where I∈R D×H×W Represents the image pixel features of size D×H×W output from the backbone network, and then passes the cluster centers through a multi-layer perceptron layer (MLP) and a softmax activation function based on dimension N to obtain the cluster classification features C corresponding to K categories K ∈R N×K Finally, the cluster assignment feature (CA) is combined with the cluster classification feature (C K ) to aggregate and generate the final segmentation result Z∈R K×H×W , the calculation expression is as follows:

[0014] =

[0015]

[0016] Furthermore, the two parallel sub-networks in CFCT are represented as sub-network 1 and sub-network 2, both of which adopt the clustering fusion network architecture. The difference is that they use different initial network parameters, which are respectively and , given an input image X, the prediction results of these two sub-networks are expressed as follows:

[0017]

[0018]

[0019] Among them, P1 and P2 represent the prediction results of subnetwork 1 and subnetwork 2 respectively, and two cross branches are introduced before the two subnetworks are connected to the segmentation head. Through the cross branches, the updated cluster centers in the two subnetworks are exchanged, and then the exchanged cluster centers are input into the segmentation head of the corresponding subnetwork, thereby generating two new prediction results, which are defined as follows:

[0020]

[0021]

[0022] in and They represent the network parameters of sub-network 1 and sub-network 2 after exchanging cluster centers, P 1C It indicates the predicted segmentation result after sub-network 1 and sub-network 2 exchange cluster centers. 2C It represents the segmentation result after sub-network 2 and sub-network 1 exchange cluster centers;

[0023] Furthermore, CFCT exchanges cluster centers between the two sub-networks to achieve cluster cross-validation and introduces a new hybrid loss function. In the network training phase, the traditional cross-pseudo-supervision method is first used to enhance the consistency between the two models. According to the training method in CPS, the one-hot output results of the two sub-networks are used as pseudo-labels to supervise the other sub-network. The pseudo-labels generated by the two sub-networks are defined as:

[0024]

[0025]

[0026] Since Y1 and Y2 are pseudo labels, no gradient backpropagation is set between P1 and Y1 or between P2 and Y2, so the cross-training loss for unlabeled data is defined as follows:

[0027]

[0028] Among them, L dice is the Dice loss function. Since the updated cluster centers aggregate information from feature maps of different scales, according to the principle of consistency regularization, the cluster centers of different sub-networks from the same image are consistent. On this basis, by exchanging the cluster centers between the two sub-networks, we can get a prediction result similar to that of the sub-network before the exchange, and then the CSCT loss definition of unlabeled data is as follows:

[0029]

[0030] CSCT further plays the role of clustering centers. By exchanging clustering centers, multi-scale pixel features from the initial sub-network can be aggregated and interacted with the image features output from the new sub-network. The overall objective function L total It consists of two parts: a supervised loss function on a labeled dataset and an unsupervised loss function on an unlabeled dataset:

[0031]

[0032]

[0033] in indicates that there is supervised loss, and Respectively represent the prediction results of the labeled images obtained from the original sub-network and the sub-network after cluster exchange, represents the true label, and Represent cross entropy loss and Dice loss respectively, and It is a hyperparameter and is set according to different data sets.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) A novel semi-supervised architecture based on cross-teaching is proposed, which adopts a network design based on cluster fusion and regards pixel-level semantic segmentation as a cluster assignment process, thus promoting the consistency between different sub-networks from a completely new perspective.

[0036] (2) Based on the cluster fusion network, a cross-training method based on cluster center exchange is proposed to further provide guidance and constraints for the semi-supervised learning process, thereby improving the robustness and fast convergence of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1This is a schematic diagram of the CFCT main network model framework proposed by the present invention;

[0038] Figure 2 This is the structure diagram of the clustered Transformer decoder proposed in the present invention;

[0039] Figure 3 This is a comparison chart of the visual segmentation results of each network model proposed in the present invention on the ACDC dataset;

[0040] Figure 4 A comparison chart of the visual segmentation results of each network model proposed in the present invention on the BUSI dataset;

[0041] Figure 5 This is the t-SNE visualization diagram of the clustering classification features proposed in the present invention. DETAILED DESCRIPTION

[0042] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0043] The present invention proposes a clustering fusion based cross-learning method (CFCT), the network architecture of CFCT is as follows: Figure 1 As shown in the figure. CFCT uses the same training mechanism as CPS, and generates pseudo labels through two sub-networks for cross-training learning to strengthen the consistency between the two sub-networks. The difference lies in the design of the sub-network. CFCT uses a cluster fusion-based MaskTransformer network (Cluster Fusion Net, CFN) to replace the traditional pixel-based segmentation network. CFN introduces a target query vector in the basic architecture of the original convolutional neural network (CNN) backbone, and updates the target query vector by cross-attention calculation with the multi-scale pixel features output from the CNN backbone network decoder. The updated target query vector is regarded as the cluster center. The cluster center is cross-attention calculated to obtain the similarity between different clusters and pixels. The calculation process is similar to the classic k-means clustering algorithm. Finally, the updated cluster center is merged with the image pixel features output by the backbone network through a segmentation head to obtain the predicted segmentation result. In addition, CFCT realizes cluster cross-validation by exchanging cluster centers between the two sub-networks, and introduces a new hybrid loss function. The following is a detailed introduction to the CFCT method and its training steps:

[0044] (1) The steps for dividing the semi-supervised medical image dataset are as follows:

[0045] In semi-supervised semantic segmentation, training set, validation set and test set are usually included. In order to implement semi-supervised training, the training set is divided into two different subsets: D and D. Where D={Xl i, Yl i}N i=1 represents the labeled dataset, which contains a limited number of labeled data, while D={Xu j}MJ=1 represents the unlabeled dataset, which contains a large amount of unlabeled data, N< <M。

[0046] (2) Clustering Fusion Network (CFN) Architecture Design:

[0047] CFN consists of three key components: a backbone network for extracting and refining pixel-level features, a clustering Transformer decoder for updating cluster centers, and a segmentation head network for generating predicted segmentation results.

[0048] ① Backbone network

[0049] The backbone network adopts the classic U-shaped network structure, extracts features through the pixel encoder, restores the feature map resolution using the pixel decoder, and fuses features of different levels through jump connections. In order to improve the feature extraction capability, ResNet50 is used as the pixel encoder in this paper, and a pixel decoder composed of residual modules is designed to process up-sampled features.

[0050] ② Clustering Transformer Decoder

[0051] like Figure 2 As shown in the figure, the clustering Transformer decoder adopts a dual attention mechanism, including cross attention between cluster centers and pixel features and self-attention of cluster centers. First, a parameter-learnable feature vector is introduced as the clustering target query vector, and then the query vector and the multi-scale pixel features output from the backbone network decoder are cross-attention calculated respectively, and the cluster-based argmax function is used to replace the traditional softmax activation function to obtain the attention mapping matrix, and then the relevant pixel features are weighted by the attention mapping matrix to update the cluster center. The calculation process of cross attention is expressed as follows:

[0052]

[0053] Where C∈R N×D represents the cluster center with N channels and D channels, represents the updated cluster center, Q c ∈R N×D represents the query matrix mapped from the clustering target query vector, K p ∈R HW×D and V p ∈RHW×D They represent the key and value matrices mapped from pixel features of size D×H×W respectively.

[0054] ③ Split head

[0055] The role of the segmentation head is to further merge the image pixel features output by the backbone network with the updated cluster centers to obtain the final segmentation prediction result. First, the product of the cluster center and the image pixel features is calculated to obtain the cluster assignment feature (CA∈R N×H×W ), which represents the similarity between each pixel and the cluster center, and the expression is as follows:

[0056] ×

[0057] Where I∈R D×H×W Represents the image pixel features of size D×H×W output from the backbone network, and then passes the cluster centers through a multi-layer perceptron layer (MLP) and a softmax activation function based on dimension N to obtain the cluster classification features C corresponding to K categories K ∈R N×K Finally, the cluster assignment feature (CA) is combined with the cluster classification feature (C K ) to aggregate and generate the final segmentation result Z∈R K×H×W , the calculation expression is as follows:

[0058] =

[0059]

[0060] (3) The implementation steps of the cluster fusion-based cross-learning mechanism (CFCT) are as follows:

[0061] CFCT consists of two parallel sub-networks, denoted as sub-network 1 and sub-network 2, both of which adopt the cluster fusion network (CFN) architecture. The difference is that they use different initial network parameters, denoted as and When given an input image X, the prediction results of these two sub-networks are expressed as follows:

[0062]

[0063]

[0064] Among them, P1 and P2 represent the prediction results of sub-network 1 and sub-network 2 respectively. In addition, Figure 1As shown by the green arrows in , two cross branches are introduced before connecting to the segmentation head. Through the cross branch, the updated cluster centers in the two sub-networks are exchanged, and then the exchanged cluster centers are input into the segmentation head of the corresponding sub-network, thereby generating two new prediction results, which are defined as follows:

[0065]

[0066]

[0067] in and They represent the network parameters of sub-network 1 and sub-network 2 after exchanging cluster centers, P 1C It indicates the predicted segmentation result after sub-network 1 and sub-network 2 exchange cluster centers. 2C It represents the segmentation result after sub-network 2 exchanges cluster centers with sub-network 1.

[0068] In the network training phase, the traditional cross pseudo supervision (CPS) method is first used to enhance the consistency between the two models. According to the training method in CPS, the one-hot output results of the two sub-networks are used as pseudo labels to supervise the other sub-network. The pseudo labels generated by the two sub-networks are defined as:

[0069]

[0070]

[0071] Since Y1 and Y2 are pseudo labels, no gradient backpropagation is set between P1 and Y1 or between P2 and Y2, so the cross-training loss for unlabeled data is defined as follows:

[0072]

[0073] Among them, L dice is the Dice loss function. Since the updated cluster centers aggregate information from feature maps of different scales, according to the principle of consistency regularization, the cluster centers of different sub-networks from the same image are consistent. On this basis, the present invention proposes a new cross-training method (CSCT) based on cluster center exchange, which obtains prediction results similar to those of the sub-networks before the exchange by exchanging cluster centers between two sub-networks. The CSCT loss of unlabeled data is defined as follows:

[0074]

[0075] CSCT further plays the role of cluster centers. By exchanging cluster centers, multi-scale pixel features from the initial sub-network can be aggregated and interacted with image features from the output of the new sub-network. This method provides more detailed guidance for the final prediction, thereby further improving the accuracy of the semantic segmentation results. The overall objective function L total It consists of two parts: a supervised loss function on a labeled dataset and an unsupervised loss function on an unlabeled dataset:

[0076]

[0077]

[0078] in indicates that there is supervised loss, and Respectively represent the prediction results of the labeled images obtained from the original sub-network and the sub-network after cluster exchange, represents the true label, and Represent cross entropy loss and Dice loss respectively, and It is a hyperparameter and is set according to different data sets.

[0079] Example:

[0080] (1) Dataset selection and configuration

[0081] The present invention uses two public datasets, ACDC and BUSI, for performance evaluation. The ACDC dataset contains 200 short-axis cardiac magnetic resonance images from 100 patients. These images are equipped with segmentation masks for labeling the left ventricle (LV), myocardium (Myo), and right ventricle (RV). To ensure fair comparison, the present invention uses the same data preprocessing and enhancement methods as in the literature. The images are divided into 2D slices in the training phase, and the predicted slices are stacked to form a 3D image in the inference phase. A total of 140 images were randomly selected from 70 patients for training, and 60 images were selected from 30 patients as validation sets. The BUSI dataset was collected by Al-Dhabyani et al. at Baheya Hospital and includes 780 breast ultrasound images from 600 female patients. Among them, 647 cases were abnormal (210 malignant and 437 benign masses), and 133 cases were normal. The present invention uses abnormal samples for training and validation, and constructs the training set, validation set, and test set in a ratio of 6:2:2.

[0082] (2) Experimental details and related settings

[0083] The experiment of the present invention was implemented using the PyTorch framework and trained on a workstation equipped with 64 GB memory and an NVIDIA GeForce RTX 4090 GPU processor. The network was trained using a stochastic gradient descent (SGD) optimizer with a batch size of 8 for a total of 30,000 iterations. In order to effectively adjust the learning rate, a polynomial learning rate strategy was adopted, where the initial learning rate of the ACDC dataset was 0.01 and the initial learning rate of the BUSI dataset was 0.001. In order to evaluate the quality of the segmentation results, the Dice Similarity Coefficient (DSC) and 95% Hausdorff distance (HD95) were used. The backbone network used ResNet50 pre-trained on ImageNet as a pixel encoder, and a parallel pixel decoder was designed with a symmetric residual structure.

[0084] To verify the performance of the proposed method on medical image segmentation, we compared it with six state-of-the-art semi-supervised segmentation methods on the ACDC and BUSI datasets. To ensure fairness, all these methods used the same backbone network and were trained and tested under the same settings.

[0085] (3) Experimental results

[0086] Table 1 shows the comparison results of various methods on the ACDC dataset using 3 and 7 labeled cases, respectively. Our method consistently outperforms all other methods in both cases. When using 7 cases as labeled samples, our method achieves the best results in both DICE and HD95 indicators, reaching 89.77% and 10.67, respectively. Compared with the second-place method, our method improves DICE by 2.9% and reduces HD by 1.28mm. In addition, when only 3 cases are used as labeled samples, our method shows a more obvious advantage. Compared with the second-place method, our DICE increases by 20.6% and HD decreases by 7.82mm. It is worth noting that as the number of labeled cases decreases to 3, the performance of other methods drops significantly, but our method is still able to provide strong performance, even surpassing the indicators of most other methods when using 7 labeled cases.

[0087] Table 1: Comparison of evaluation indicators of segmentation results of various network models on the ACDC dataset

[0088]

[0089] Table 2 shows the comparison results on the BUSI dataset when using 10% and 20% of the labeled data. In both cases, our method significantly outperforms all other semi-supervised segmentation methods. When using 20% ​​of the labeled data, our method achieves 79.64% DICE and 19.97 HD95 metrics, respectively. Compared with the second-place method, our DICE increases by 4.2% and HD decreases by 4.47mm. When using 10% of the labeled data, our method increases DICE by 2.2% and HD decreases by 2.86mm compared to the second-place method.

[0090] Table 2: Comparison of evaluation indicators of segmentation results of various network models on the BUSI dataset

[0091]

[0092] Figure 3 The visualization results of different segmentation methods on the ACDC dataset are shown, where the first three rows are the segmentation results under 3 labeled cases, and the last three rows are the results under 7 labeled cases. The visualization results show the advantages of the present invention in terms of accuracy and reduction of false positives, especially when facing similar pixel grayscale distribution and segmenting smaller organs. Figure 4 The visualization results of different segmentation methods on the BUSI dataset are shown, where the first three rows are the segmentation results using 10% annotated cases, and the last three rows are the results using 20% ​​annotated cases. It can be seen that when faced with difficulties such as similar pixel grayscale distribution and irregular target shapes, the present invention still shows obvious advantages. Especially when annotated data is scarce, our method can still accurately segment the target contour, while the performance of other methods drops sharply.

[0093] Finally, we further analyzed the recognition ability of the cluster centers in the present invention for different categories. Specifically, we used the trained CFCT network to infer the ACDC and BUSI datasets, extracted cluster classification features, and used the t-SNE visualization method to evaluate their ability to effectively distinguish different categories. Figure 5 It can be clearly seen that in the ACDC and BUSI datasets, different categories can be clearly separated by a large metric distance, which verifies the ability of the cluster center to identify different categories.

[0094] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0095] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A semi-supervised medical image segmentation method based on clustering fusion cross-learning, comprising: S1: construct two sub-networks, both using clustering fusion network architecture but with different initial network parameters; S2: For the input image, the two sub-networks generate prediction results respectively; S3: Before the two sub-networks are connected to the segmentation head, a cross branch is introduced to exchange the updated cluster centers in the two sub-networks, and then the exchanged cluster centers are input into the segmentation head to generate new prediction results; The clustering fusion network includes: The backbone network for extracting and refining pixel-level features adopts a classic U-shaped network structure, extracts features through a pixel encoder, restores feature map resolution using a pixel decoder, and fuses features at different levels through skip connections; The clustering Transformer decoder used to update the cluster center first introduces a parameter-learnable feature vector as the clustering target query vector, then performs cross-attention calculations on the query vector and the multi-scale features output from the backbone network decoder, and uses the clustering-based argmax function to replace the traditional softmax activation function to obtain the attention mapping matrix. The pixel features are weighted by the attention mapping matrix to update the cluster center. The segmentation head used to generate the predicted segmentation result further merges the pixel features output by the backbone network with the updated cluster centers. Specifically, the product of the cluster center and the pixel features is first calculated to obtain the cluster assignment feature, which represents the similarity between each pixel and the cluster center. The cluster centers are then passed through a multi-layer perceptron layer and a dimension-based softmax activation function to obtain the cluster classification feature of the corresponding category. Finally, the cluster assignment feature is aggregated with the cluster classification feature to generate the final segmentation prediction result.

2. The semi-supervised medical image segmentation method based on clustering fusion cross-learning according to claim 1 is characterized in that: The following steps are also included: S4: In the network training phase, the traditional cross-pseudo-supervision method is first used to enhance the consistency between the two models, that is, the one-hot output results of the two sub-networks are used as pseudo-labels to supervise the other sub-network. The pseudo-labels generated by the two sub-networks do not set gradient back propagation, thereby obtaining the cross-training loss of unlabeled data; S5: By exchanging cluster centers between the two sub-networks, we obtain prediction results similar to those of the sub-networks before the exchange, thereby obtaining the cluster cross-training loss for unlabeled data; S6: The overall objective function is obtained by the supervised loss on the labeled dataset, the cross-training loss on the unlabeled data, and the clustering cross-training loss.

3. The semi-supervised medical image segmentation method based on clustering fusion cross-learning according to claim 1 is characterized in that: The calculation process of cross attention is expressed as follows: （1） Where C∈R N×D represents the cluster center with N channels and D channels, represents the updated cluster center, Q c ∈R N×D represents the query matrix mapped from the cluster query vector, K p ∈R HW×D and V p ∈R HW×D They represent the key and value matrices mapped from pixel features of size D×H×W respectively.

4. The semi-supervised medical image segmentation method based on clustering fusion cross-learning according to claim 1 is characterized in that: The calculation expression of the cluster allocation feature is: （2） Where I∈R D×H×W Represents the image pixel features of size D×H×W output from the backbone network.

5. The semi-supervised medical image segmentation method based on clustering fusion cross-learning according to claim 2 is characterized in that: The cross-training loss for the unlabeled data is defined as: （11） Among them, L dice is the Dice loss function, Y1 and Y2 are pseudo labels, and P1 and P2 represent the prediction results of the two sub-networks respectively.

6. The semi-supervised medical image segmentation method based on clustering fusion cross-learning according to claim 2 is characterized in that: The clustering cross-training loss of the unlabeled data is defined as: （12） Among them, P 1C P represents the predicted segmentation result after the first sub-network and the second sub-network exchange cluster centers. 2C Represents the predicted segmentation result after the second sub-network exchanges cluster centers with the first sub-network.

7. The semi-supervised medical image segmentation method based on clustering fusion cross-learning according to claim 2 is characterized in that: The overall objective function L total The calculation expression is: （13） （14） in indicates that there is supervised loss, and Respectively represent the prediction results of the labeled images obtained from the original sub-network and the sub-network after cluster exchange, represents the true label, and They represent cross entropy loss and Dice loss respectively, and λ1 and λ2 are hyperparameters.

Images