Medical image classification method based on federal semi-supervised learning

Abstract

The invention discloses a medical image classification method based on federal semi-supervised learning. Firstly, labeled medical images of clients are input into a local model and a local global model for training. Secondly, calculating a threshold value of each class and issuing the threshold value to a client, and inputting a medical image of each client into a local model and a local global model for training; and then performing federal aggregation on the updated local model with the label data and the local global model to obtain a server global model, issuing the server global model to a client to update the local global model, entering an active learning process after repeated operation, and repeating the active learning process to achieve a predetermined label budget. And finally, inputting the updated label data and label-free data of the client into the local model and the local global model, and training until the global model of the server converges. According to the method, the model is adaptive to each category, the method is easy to implement, the communication cost is not increased, and meanwhile, the classification accuracy is also improved.

Description

Technical Field

[0001] The present invention belongs to the field of federated semi-supervised learning, and specifically relates to a medical image classification method based on federated semi-supervised learning. Background Art

[0002] With the rapid development of next-generation artificial intelligence (AI) technologies, the efficient utilization and standardized management of public data elements have become key to driving digital transformation. Medical data, as a vital public data resource, holds immense social value and scientific research potential. However, its application faces challenges such as strict privacy protection, high labeling costs, and difficulties in cross-institutional collaboration. Currently, medical image classification primarily relies on centralized labeled data training. However, the decentralized, private, and lack of standardization of public data limits the generalization and application scope of AI models.

[0003] In this context, the combination of federated learning and semi-supervised learning offers an innovative solution for the compliant sharing and intelligent analysis of public data elements. On the one hand, federated learning, through distributed collaborative training, ensures knowledge sharing of medical data while preserving local privacy, complying with the requirements of public data security governance. On the other hand, semi-supervised learning can fully utilize the vast amount of unlabeled public data resources in medical institutions, reducing reliance on labeled data and improving model performance.

[0004] Although federated semi-supervised learning shows great potential, there are still many challenging issues in its application in actual medical scenarios:

[0005] (1) Data heterogeneity and data distribution bias

[0006] First, the distribution of medical images varies greatly between different medical institutions. These differences are not only reflected in the differences in imaging equipment (such as Computed Tomography and Magnetic Resonance Imaging), but also in image quality, resolution, and shooting angle. This makes the data distribution of each client (medical institution) highly heterogeneous. Therefore, even in the same task, data from different institutions cannot be simply aggregated and processed uniformly. Second, in medical image data, image samples of certain diseases may be very scarce (such as rare diseases), while image samples of other diseases may be very common. Federated semi-supervised learning needs to utilize unlabeled data in the presence of such imbalanced data and ensure that the global model can maintain good performance on all classes.

[0007] (2) Trade-off between performance and efficiency

[0008] First, medical images are typically high-dimensional (for example, high-resolution CT / MRI images), and model training requires high computational requirements. Federated learning requires local training on each client (such as a hospital or medical device), which can result in a significant computational burden, especially when computing resources are limited in medical institutions. Second, federated semi-supervised learning for medical image classification typically requires larger models and higher computational accuracy, especially when large amounts of unlabeled data are available. How to efficiently utilize unlabeled data to improve the model's classification performance is a key issue that needs to be addressed. Summary of the Invention

[0009] Existing federated semi-supervised learning methods for medical image classification cannot solve the above problems well. The present invention proposes a medical image classification method based on federated semi-supervised learning, which mainly handles data heterogeneity and efficiently utilizes unlabeled data in the following three ways: (1) When selecting pseudo labels for unlabeled data, a dynamic threshold is set for each class; (2) The prediction of the global model will be biased towards the majority class, so a local model trained only on local labeled data is designed to predict pseudo labels for unlabeled data to guide the learning of the global model; (3) Active learning is introduced to screen samples that are rich in information and important to the model from unlabeled data with unreliable pseudo labels, and use them as labeled data for supervised training. The present invention is aimed at the standardization needs of public data elements for the new generation of artificial intelligence, aiming to break the contradiction between data privacy protection and efficient utilization, establish a standardized and scalable public data collaboration mechanism, and provide key technical support for the open sharing and value mining of public data elements such as medical care.

[0010] To achieve the above objectives, the present invention adopts a technical solution: a medical image classification method based on federated semi-supervised learning, comprising the following steps:

[0011] Step 1: Establish a federated learning system, initialize N clients and 1 server, each client holds a local private dataset consisting of a labeled medical image dataset and unlabeled medical image datasets Usually, the labeled dataset is smaller than the unlabeled dataset. In order to capture the local labeled data distribution of the client to guide the global model learning, each client trains a local model locally using only the labeled dataset. and a local global model learned using all datasets The server has a server-global model ω g ,ω g By all Gather together.

[0012] Step 2: Each client's labeled medical images are input into the randomly initialized local model and the local global model for training for a certain number of rounds to obtain a local model learned only locally on each client. At the same time, the local global model is federated and aggregated to obtain the server global model and sent to the client. The client uses this to update the local global model.

[0013] Step 3: Send the number of each class in each client's labeled image to the server, count the total number of each class in the server, and calculate the threshold of each class and send it to the client.

[0014] Step 4: Input all medical images of each client into the local model and the local global model for training. For labeled images, supervised learning is performed, that is, the predicted output of the local model and the predicted output of the local global model are respectively subjected to cross entropy loss with the true label.

[0015] Step 5: For unlabeled images, if the local model's prediction result exceeds the class threshold, it will be consistent with the output of the local global model and obtain the consistency loss. Otherwise, it will be added to the dataset to participate in active learning. middle, Initially empty.

[0016] Step 6: Update the local model with the cross entropy loss of labeled data, and update the local global model with the cross entropy loss of labeled data and the consistency loss of unlabeled data. Then, federate the local global model to obtain the server global model and send it to the client. The client uses this to update the local global model.

[0017] Step 7: Repeat Step 4 to Step 6 until the predetermined number of iterations are completed and the active learning process begins. Select information-rich samples and add them to the labeled data, and delete them from the unlabeled data, then recalculate and update the class threshold according to Step 3.

[0018] Step 8: Repeat Step 7 until the predetermined label budget is reached, after which the labeled and unlabeled datasets of each client no longer change.

[0019] Step 9: Input the client's labeled data and unlabeled data updated in Step 8 into the local model and the local global model. Perform cross entropy loss on the labeled data prediction outputs of the two models and the true labels to obtain their respective labeled data cross entropy losses. At the same time, for the unlabeled data prediction outputs of the local model, select the prediction outputs that exceed the class threshold and perform consistency learning with the prediction outputs of the local global model to obtain the consistency loss of the unlabeled data. Then update all models according to Step 6.

[0020] Step 10: Repeat Step 9 until the server global model converges and obtains the classification result.

[0021] In Step 2, the client only uses labeled data to train its own local model and the server global model, where the local model captures the local data distribution and the server global model captures the global data distribution.

[0022] The calculation of the class threshold in Step 3 is as follows:

[0023]

[0024] is the threshold of class c, std(β t ) is the standard deviation of the distribution of labeled medical images in this federated system, β t (c) represents the proportion of labeled medical images of category c in the total labeled medical images, is a predefined threshold cardinality.

[0025] The training method with labeled data in Step 4 can be expressed as:

[0026]

[0027] They represent the local model and local global model parameters of the i-th client in the t-th round, They represent the loss of labeled data of the i-th client in the t-th round on the local model and the local global model, CE represents the cross entropy loss, and f(.) represents the corresponding local model and local global model. is the labeled dataset of the i-th client, yes The size of They are the mth labeled data of the client and its corresponding true label, Indicates weak enhancement of labeled images.

[0028] The training method of unlabeled data in Step 5 can be expressed as:

[0029]

[0030] represents the consistency loss of the client's unlabeled data on the local global model, is the unlabeled dataset of the i-th client, yes The size of is the mth unlabeled data of the client, Respectively represent weak enhancement and strong enhancement of unlabeled images, represents the softmax prediction output of the local model for the weakly enhanced unlabeled data, Represents unlabeled data Pseudo labels, add unlabeled data below the class threshold to middle.

[0031] In Step 6, the server global model is aggregated and updated using the federated averaging algorithm. To prevent the local model from overfitting to the local labeled data, the exponential moving average (EMA) algorithm is used to update the trained local model, thereby effectively integrating global knowledge into the local model.

[0032] In Step 7, active learning is performed within each client. The active learning process of the i-th client can be expressed as follows:

[0033] Step 7-1: Import to local model Get the features of the previous layer of the last fully connected layer

[0034] Step 7-2: Use clustering algorithm to Clustering is performed to obtain K categories of data, each of which contains at least one unlabeled data.

[0035] Step 7-3: Use local global model Query the clustering results and select the sample with the largest predicted entropy value output by the local global model in each class Tag.

[0036] Step 7-4: Join from Delete and set Empty.

[0037] Beneficial effects of the present invention:

[0038] 1. The present invention counts the total number of various types of labeled medical images in the server, sets a dynamic threshold for each class on the client, and fairly trains each type of data without leaking privacy, making the model adaptive to each type of disease. It is easy to implement and does not increase communication costs.

[0039] 2. Considering that the global model's prediction of unlabeled data will be biased towards the majority class, the present invention uses a local model learned only on local labeled data to perform pseudo-label prediction on unlabeled data, and uses this to guide global model learning, thereby improving the reliability of pseudo-label prediction and the accuracy of global model classification.

[0040] 3. The present invention introduces active learning, which filters out samples that are rich in information and important to the model from unlabeled data with unreliable pseudo-labels. These samples participate in supervised training as labeled data, balancing the training of various diseases from both local and global perspectives. While efficiently utilizing unlabeled data, it also ensures that the global model can maintain relatively good performance on all classes. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the overall framework diagram of the present invention;

[0042] Figure 2 This is the trend chart of the accuracy results under the diabetic dataset;

[0043] Figure 3 This is the recall rate trend chart for the diabetic dataset;

[0044] Figure 4 This is the F1 result trend chart for the diabetic dataset. DETAILED DESCRIPTION

[0045] The present invention will be further described below with reference to the accompanying drawings and specific implementation steps:

[0046] like Figure 1 The figure shows the overall framework of a medical image classification method based on federated semi-supervised learning in the present invention, wherein ω in the middle left of the figure l 、ω g Corresponding to the local global model parameters and server global model parameters described above, the local global model at this time is used for federated aggregation. The local model and global model in the right figure correspond to the local model and local global model described above. The local global model at this time is used for local training. σ(C) is the number of classes C in the client's internal labeled data, and T(C) is the threshold of class C calculated by the server based on the category distribution of all client labeled data. The specific steps of the embodiment are as follows:

[0047] Step 1: Establish a federated learning system, initialize N clients and 1 server, and the dataset of each client i∈{1,...,N} consists of a labeled medical image dataset and unlabeled medical image datasets The model consists of a local model that is biased towards its private data distribution. and local global models Composition, all clients have the same small label ratio In this embodiment, γ is initially 0.1; the server contains a server global model ω g ,ω g By all Gather together.

[0048] Step 2: To obtain more reliable pseudo-label predictions for unlabeled data, warm-up the labeled data before training the unlabeled data. This involves inputting each client's labeled medical images into a randomly initialized local model and a local global model for training for a certain number of rounds (50 rounds in this embodiment). This process then results in a local model learned only locally on each client and a server-based global model that aggregates the local global model. This helps the model quickly converge to a good initial state, thereby improving the stability and performance of subsequent training. The specific training is as follows:

[0049]

[0050] They represent the local model and local global model parameters of the i-th client in the t-th round, They represent the loss of labeled data of the i-th client in the t-th round on the local model and the local global model, CE represents the cross entropy loss, is the labeled dataset of the i-th client, yes The size of They are the mth labeled data of the client and its corresponding true label, Indicates weak enhancement of labeled images.

[0051] The i-th client uses For the local model of round t Local global model Perform stochastic gradient update to get The server global model of the final round t+1 and local models It can be expressed as:

[0052]

[0053] Among them, the server global model is aggregated and updated through the federated average algorithm according to the size of each client's labeled data. represents the size of the labeled dataset of the i-th client. In order to prevent the local model from overfitting to the local labeled data, the global knowledge outside the local data is integrated into the local model through the exponential moving average algorithm. ema It is a hyperparameter that controls the ratio of the fusion of the local model and the local global model. In this embodiment, it is set to 0.95.

[0054] Step 3: Medical images are highly heterogeneous. Considering the presence of both common and rare diseases, we set different class thresholds for each disease based on the distribution of each disease class in the labeled data, thereby training each class of medical images fairly. To protect the client's privacy, the class thresholds are calculated on the server. First, the client sends the number of each class in its existing labeled images to the server. The server then counts the total number of each class and calculates the threshold for each class and sends it to the client. The specific calculation is as follows:

[0055]

[0056] in, represents the number of labeled medical images of category c for the i-th client at round t, σ t (c) represents the total number of labeled medical images of category c counted on the server in round t, and further calculates the proportion of each category of labeled medical images in the total number of labeled medical images and the standard deviation of the distribution of labeled medical images:

[0057]

[0058] β t (c) represents the proportion of labeled medical images of category c in the total labeled medical images counted on the server in round t, std(β t ) represents the standard deviation of the distribution of labeled medical images, where Finally, the threshold of class c at round t can be calculated as:

[0059]

[0060] in, is a predefined threshold base, which is set to 0.85 in this embodiment. In order to ensure that there is a certain amount of unlabeled data to participate in semi-supervised training, an upper limit is set for this threshold, which is defined as follows:

[0061]

[0062] in, is the upper limit of the predefined threshold.

[0063] The class threshold is recalculated only when the client's labeled data set changes. In this embodiment, the class threshold is recalculated after one active learning is completed.

[0064] Step 4: After Warm-Up is completed, all medical images of each client are input into the local model and the local global model for training. Supervised learning is performed on labeled images, that is, the predicted output of the local model and the predicted output of the global model are respectively subjected to cross entropy loss with the true label. The training of labeled data in this step is the same as Step 2, see formula (1)(2).

[0065] Step 5: For unlabeled images, if the local model's prediction result exceeds the class threshold, it will be consistent with the output of the global model and obtain the consistency loss. Otherwise, it will be added to the dataset to participate in active learning. middle, Initially empty, the training method of unlabeled data can be expressed as:

[0066]

[0067] represents the loss of the unlabeled data of the i-th client on its local global model, represents the softmax prediction output of the local model for the weakly enhanced unlabeled data, Represents the pseudo-label of the unlabeled data. Considering that the global model will be biased towards the majority class, the pseudo-label of the unlabeled data is predicted by the local model learned only on the local labeled dataset. If the pseudo-label predicted by the local model exceeds the class threshold set in Step 3, this pseudo-label is used as the true label to guide the learning of the local global model; otherwise, this unlabeled data is added to middle.

[0068] Step 6: Update the local model with the cross entropy loss of labeled data, and update the local global model with the cross entropy loss of labeled data and the consistency loss of unlabeled data. Then, federate the local global model to obtain the server global model and send it to the client. The client uses it to update the local global model. The losses of the local model and the local global model in this step are:

[0069]

[0070] Calculated by formula (1) and (2), they represent the cross entropy loss of the labeled data of the i-th client on its local model and local global model, respectively. Represents the consistency loss of the unlabeled data of the i-th client on its local global model. The i-th client is represented by For the local model of round t and local global models Perform stochastic gradient update to get The server global model of the final round t+1 and local models The calculation method is the same as Step 2, see formula (3)(4).

[0071] Step 7: Repeat Step 4 to Step 6 until the predetermined number of iterations are completed and then enter active learning: The samples with rich information are selected and added to the labeled data and deleted from the unlabeled data. In this embodiment, the predetermined number of iterations is set to 50, after which each client performs active learning locally. If it is not empty, the active learning process of the i-th client can be specifically expressed as follows:

[0072] Step 7-1: Import to local model Get the features of the previous layer of the last fully connected layer

[0073] Step 7-2: Use K-Means++ clustering algorithm to Clustering is performed to obtain K categories of data. In this embodiment, the size of K is the number of labels required for each client to perform active learning. Each category contains at least one unlabeled data.

[0074] Step 7-3: Use local global model Query the clustering results and select the sample with the largest predicted entropy in the output of the current local global model in each class Labeling. The experimental process shows that the data that the model ultimately selects to be labeled is data of relatively small categories based on the client's existing labeled data. In this embodiment, the distribution of labeled data and unlabeled data for each client is the same, so active learning alleviates the client's Non-IID situation and effectively and efficiently utilizes unlabeled data.

[0075] Step 7-4: Join from Delete and set Empty.

[0076] After active learning is completed, the client's labeled dataset changes, so the class threshold needs to be recalculated, so Step 3 is executed.

[0077] Step 8: Step 7 is an active learning cycle. After each active learning is completed, the client's labeled dataset and unlabeled dataset will change. The number of active learning cycles set in this embodiment is 2, so Step 7 is executed again with the changed labeled dataset and unlabeled dataset. After two active learning cycles, the labeled dataset and unlabeled dataset no longer change. The active learning related experimental settings are shown in Table 2.

[0078] Step 9: Input the client's labeled data and unlabeled data into the local model and the local global model, and perform cross entropy loss on the labeled data prediction outputs of the two models with the true labels to obtain the respective labeled data cross entropy losses. The calculation method is shown in formula (1) (2); at the same time, for the unlabeled data prediction output of the local model, the prediction output that exceeds the class threshold is selected for consistency learning with the local global model to obtain the consistency loss of the unlabeled data The calculation method is shown in formula (10), and then all models are updated according to Step 6.

[0079] Step 10: Repeat Step 9 until the server global model converges.

[0080] The performance evaluation of the present invention uses the medical datasets Skin Lesion, RSNA ICH, diabetic, and fundus. Table 1 below shows the four datasets:

[0081] Table 1 Dataset attribute table

[0082]

[0083] The active learning related settings are shown in Table 2 below:

[0084] Table 2 Active learning settings

[0085]

[0086]

[0087] The experiment uses ACC, Recall, and F1 as evaluation indicators for the classification model trained using the method proposed in this invention. ACC represents the model's accuracy, which is the ratio of the number of samples correctly predicted by the classification model to the total number of samples; Recall represents the model's recall, which is the ratio of the number of samples correctly predicted as positive by the classification model to the actual number of positive samples; and F1 is the harmonic mean of Precision and Recall, used to comprehensively measure the performance of the model. The formulas are defined as follows:

[0088]

[0089] Among them, the definition of precision is:

[0090]

[0091] TP (True Positive): True positive examples, samples that the model correctly predicts as positive.

[0092] TN (True Negative): True negative examples, samples that the model correctly predicts as negative.

[0093] FP (False Positive): False positive examples, samples that the model incorrectly predicts as positive.

[0094] FN (False Negative): False negative examples, samples that the model incorrectly predicts as negative.

[0095] ACC focuses on the accuracy of the overall prediction, Recall focuses on the model's ability to identify positive samples, and F1 is a comprehensive indicator of precision and recall, which is suitable for datasets with imbalanced categories.

[0096] Table 4 and Table 4 below show the experimental results of the present invention on the above four data sets:

[0097] Table 3 Experimental results record

[0098]

[0099] Table 4 Experimental results record

[0100]

[0101] FSSAL is the method proposed in this invention. As can be seen from the table, FSSAL performs better than other federated semi-supervised methods in terms of ACC, Recall, and F1 classification model evaluation indicators.

[0102] Figure 2 、 Figure 3 、 Figure 4These are the ACC, Recall, and F1 experimental results for the diabetic dataset. FedAvg is the baseline, performing only federated supervised learning, meaning each client's training data consists of only 20% labeled data. RSCFed, FedLoKe, and CBAFed are federated semi-supervised learning methods, with their training data consisting of 20% labeled data and 80% unlabeled data. FSSAL, the method proposed in this paper, initially trains data consisting of 10% labeled data and 90% unlabeled data. After the 150th round of active learning, the training data is reduced to 20% labeled data and 80% unlabeled data. It can be observed from the figure that RSCFed, FedLoKe, CBAFed, and FSSAL all perform better than FedAvg, indicating the effectiveness of utilizing unlabeled data. It can be further observed that after the 150th round, FSSAL significantly increases in ACC, Recall, and F1, indicating that when the ratio of labeled data to unlabeled data is the same, compared with other federated semi-supervised methods in the figure, the method proposed in this invention can enable labeled data and unlabeled data to better collaborate to improve the performance of the classification model, and also utilize unlabeled data more effectively and efficiently.

Claims

1. A medical image classification method based on federated semi-supervised learning, characterized in that: The following steps are involved: S1: Establish a federated learning system and input each client's labeled medical images into the randomly initialized local model and the local global model for training; S2: Send the number of each class in each client's labeled image to the server, count the total number of each class in the server, and calculate the threshold of each class and send it to the client; S3: Input each client’s medical images into the local model and the local global model for training; S4: Use the updated local model and the local global model with labeled data, federate the local global model, obtain the server global model, and send it to the client. The client updates the local global model. S5: Repeat S3 and S4 until the predetermined number of iterations is completed and then enter the active learning process. The active learning process is repeated until the predetermined label budget is reached. After that, the labeled dataset and unlabeled dataset of each client no longer change. S6: Input the client labeled data and unlabeled data updated in S5 into the local model and the local global model for training until the server global model converges and obtains the classification results.

2. The medical image classification method based on federated semi-supervised learning according to claim 1, characterized in that: In the federated learning system, N clients and 1 server are initialized. Each client holds a local private dataset consisting of a labeled medical image dataset. and unlabeled medical image datasets The labeled dataset is smaller than the unlabeled dataset, and each client trains a local model locally using only the labeled dataset. and a local global model learned using all datasets The server has a server-global model ω g ,ω g By all Gather together.

3. The medical image classification method based on federated semi-supervised learning according to claim 2, characterized in that: In step 1, the labeled medical images of each client are input into the randomly initialized local model and the local global model for training, thereby obtaining a local model learned only locally on each client. At the same time, the local global model is federated and aggregated to obtain a server global model, which is then sent to the client. The client uses this model to update the local global model. The client only uses labeled data to train its own local model and the server global model, where the local model captures local data distribution and the server global model captures global data distribution.

4. The medical image classification method based on federated semi-supervised learning according to claim 3, characterized in that: The calculation of the threshold value of each class in step S2 is as follows: is the threshold of class c, std(β t ) is the standard deviation of the distribution of labeled medical images in this federated system, β t (c) represents the proportion of labeled medical images of category c in the total labeled medical images, is a predefined threshold cardinality.

5. The medical image classification method based on federated semi-supervised learning according to claim 4, characterized in that: The specific implementation process of step S3 is as follows: S3.1: For labeled images, supervised learning is performed, where the predicted output of the local model and the predicted output of the local global model are subjected to cross entropy loss with the true label. S3.2: For unlabeled images, if the local model's prediction result exceeds the class threshold, it will be consistent with the output of the local global model and obtain the consistency loss. Otherwise, it will be added to the dataset to participate in active learning. middle, Initially empty.

6. The medical image classification method based on federated semi-supervised learning according to claim 5, characterized in that: Step S4 is specifically implemented as follows: updating the local model with the cross entropy loss of the labeled data, updating the local global model with the cross entropy loss of the labeled data and the consistency loss of the unlabeled data, and then federating the local global model to obtain the server global model and send it to the client, which uses it to update the local global model; The training method of unlabeled data is expressed as: represents the consistency loss of the client's unlabeled data on the local global model, is the unlabeled dataset of the i-th client, CE represents the cross entropy loss, yes The size of is the mth unlabeled data of the client, They represent strong enhancement of unlabeled images, f(.) represents the local global model, represents the softmax prediction output of the local model for the weakly enhanced unlabeled data, Represents unlabeled data Pseudo labels, add unlabeled data below the class threshold to middle, represents the local global model parameters of the i-th client in the t-th round.

7. The medical image classification method based on federated semi-supervised learning according to claim 6, characterized in that: In S5, the active learning process is performed inside each client, that is, Select samples to add to the labeled data and delete them from the unlabeled data, then recalculate the class threshold and update it. The active learning process of the i-th client is expressed as follows: S5-1: Import to local model Get the features of the previous layer of the last fully connected layer S5-2: Use clustering algorithm to Clustering is performed to obtain K categories of data, each of which contains at least one unlabeled data; S5-3: Using local global models Query the clustering results and select the sample with the largest predicted entropy value output by the local global model in each class Tagging; S5-4: Join from Delete and set Empty.

8. The medical image classification method based on federated semi-supervised learning according to claim 7, characterized in that: The step S6 is specifically implemented as follows: inputting the client's labeled data and unlabeled data into the local model and the local global model, performing cross entropy loss on the labeled data prediction outputs of the two models and the true labels respectively to obtain the respective labeled data cross entropy losses; at the same time, for the unlabeled data prediction outputs of the local model, selecting the ones that exceed the class threshold and the prediction outputs of the local global model for consistency learning to obtain the consistency loss of the unlabeled data, and then updating all models according to S4 until the server global model converges.

Images