Medical image segmentation model construction method based on federated learning

Abstract

The application discloses a medical image segmentation model construction method and system based on federated learning. The method comprises the following steps: S1, setting a personalized segmentation model and a general segmentation model in a client and initializing; S2, learning the inverse network of the personalized encoder as a reconstructor; S3, reconstructing localized reconstruction data with global information, and updating the two models with the data; S4, uploading the general encoder and the local segmenter of each client to a server, and aggregating the global encoder and the global segmenter in the server; S5, the server broadcasts the global encoder and the global segmenter to each client, and assigns the parameters of the global encoder and the global segmenter to the general encoder and the local segmenter; S6, repeating steps S2 to S5 until the performance of the two models is stable. The method can obtain a personalized performance comparable to the personalized federated learning method, and greatly improves the generalization performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, more particularly, to a medical image segmentation model construction method and system based on federated learning. BACKGROUND

[0002] Current federated medical image segmentation methods mainly include personalized federated learning (perFL) and federated domain generalization (FDG). Among them, personalized federated learning (perFL) designs a separate personalized model for each client, so that the model is most suitable for local data, but this method sacrifices the generalization ability of the model. The results of the personalized federated learning method are not satisfactory on unknown data. Federated domain generalization (FDG) uses traditional federated learning technology, and all clients jointly train a global segmentation model, which is difficult to deal with the client data drift problem. The global segmentation model often cannot meet the medical image segmentation tasks of multiple known clients. It can be seen that the current federated learning has the problems of being difficult to adapt to multi-source heterogeneous data and being unable to simultaneously consider localization and globalization in medical applications. Therefore, the present application adopts a "personalized segmentation model + general segmentation model". Dual model mode, and develops a new type of cycle model distillation (CMD) method to jointly optimize the performance of medical image segmentation of multi-source information, in order to solve the problem that the model personalization and generalization ability are difficult to simultaneously consider in federated medical image segmentation.

[0003] The Chinese invention patent with the application date of 2022.09.09: Multi-task hybrid supervision medical image system based on federated learning, the invention patent proposes a multi-task hybrid model architecture based on federated learning, aiming to improve the accuracy and robustness of the model through information complementation between different levels of features. Therefore, the invention patent does not solve the problem that the model personalization and generalization ability are difficult to simultaneously consider in federated medical image segmentation. SUMMARY

[0004] In view of the fact that the prior art does not solve the problem that the model personalization and generalization ability are difficult to simultaneously consider in federated medical image segmentation, the present application proposes a medical image segmentation model construction method and system based on federated learning. The technical scheme adopted by the present application is:

[0005] The first aspect of the present application provides a medical image segmentation model construction method based on federated learning, comprising the following steps:

[0006] S1, setting a personalized segmentation model and a general segmentation model in each client, and initializing;

[0007] S2, using local data to learn the inverse network of the personalized encoder as a reconstructor in each client;

[0008] S3, in each client, the reconstructor is spliced after the general encoder, the localized reconstruction data with global information is reconstructed, and the cyclic model distillation method is used to update the personalized segmentation model and the general segmentation model using the localized reconstruction data;

[0009] S4, upload the general encoder and the local segmenter of each client to the server, aggregate the global encoder and the global segmenter in the server;

[0010] S5, the server broadcasts the global encoder and the global segmenter to each client, and assigns the parameters of the global encoder and the global segmenter to the general encoder and the local segmenter;

[0011] S6, repeat steps S2 to S5 until the performance of the personalized segmentation model and the general segmentation model is stable.

[0012] Compared with the prior art, the present application adopts a double model mode of "personalized segmentation model + general segmentation model". In order to optimize the double model under the federal learning framework, the present application develops a new type of cyclic distillation method, which optimizes the medical image segmentation performance in each client by knowledge distillation between the personalized segmentation model and the general segmentation model, and combines multi-source information. The method of the present application greatly improves the generalization performance while obtaining the same personalized performance as the personalized federal learning method, and solves the problem that the prior art is difficult to simultaneously consider the model personalization and generalization ability for federal medical image segmentation.

[0013] As a preferred solution, the personalized segmentation model , and the general segmentation model , respectively include the personalized encoder and the general encoder , and share the same local segmenter .

[0014] As a preferred solution, the reconstructor is supervised and updated by the first Dice loss function and the reconstruction loss function.

[0015] Further, the first Dice loss function is:

[0016]

[0017] wherein, represents the number of local data of the client k, represents the predicted label obtained by using the personalized segmentation model for the i-th sample in the client k, This represents the true label of the i-th sample in client k.

[0018] Furthermore, the reconstruction loss function is:

[0019]

[0020] in, This represents the amount of local data on client k. This represents the i-th local data of client k. This represents the i-th reconstructed data of client k.

[0021] As a preferred embodiment, in step S3, the updates of the personalized segmentation model and the general segmentation model are supervised by the second Dice loss function, the third Dice loss function, and the fourth Dice loss function.

[0022] Furthermore, for:

[0023]

[0024] in, This represents the amount of local data on client k. This indicates that the i-th sample in client k uses a personalized segmentation model. The obtained prediction labels, This represents the true label of the i-th sample in client k.

[0025] Furthermore, the third Dice loss function is:

[0026]

[0027] in, This represents the amount of local data on client k. The i-th reconstructed data in client k Using a general segmentation model The obtained prediction labels, This represents the true label of the i-th sample in client k.

[0028] Furthermore, the fourth Dice loss function is:

[0029]

[0030] in, This represents the amount of local data on client k. This represents the i-th reconstructed data in client k. Using a personalized segmentation model The obtained prediction labels, denotes the true label of the i-th sample in the client k.

[0031] The second aspect of the present application also provides a medical image segmentation model construction system based on federated learning, comprising a server and a plurality of clients;

[0032] The client is configured to train the personalized segmentation model and the general segmentation model set by the local data, and upload the general segmentation model to the server;

[0033] The server is provided with a model aggregation module and an aggregated model broadcast module;

[0034] The model aggregation module is configured to aggregate the general encoder and the local segmenter uploaded by the client to the server and generate a global encoder and a global segmenter;

[0035] The aggregated model broadcast module is configured to broadcast the global encoder and the global segmenter to the clients.

[0036] The present application has the following beneficial effects:

[0037] The present application is based on federated learning, and the ability of the general segmentation model to extract features containing global information is retained through the cyclic model distillation method. By using the global model as the carrier of information transmission between clients, there is no additional communication cost, and there is no need to transmit data, effectively maintaining the privacy of the client data. Compared with existing federated medical image segmentation methods, the present application has excellent segmentation performance in known medical images. Compared with other personalized federated learning methods, the present application has obvious performance improvement in the face of unknown data sets. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The present application provides a medical image segmentation model construction method based on federated learning;

[0039] Figure 2 The present application provides a model construction process schematic diagram;

[0040] Figure 3 The present application provides a medical image segmentation experiment result diagram. DETAILED DESCRIPTION

[0041] The accompanying drawings are only used for illustrative purposes and cannot be understood as limiting the present application;

[0042] It should be noted that the embodiments described are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0043] The terms used in the embodiments of the present application are only for the purpose of describing the specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0044] The following description refers to the accompanying drawings. In the following description, identical reference numbers in different drawings represent the same or similar elements unless otherwise indicated. The following description of illustrative embodiments is not meant to limit or restrict the scope of the application, as defined in the appended claims, in which only the claims are intended to be limiting on the application. The following examples are illustrative only, and not intended to limit or restrict the scope of the application as defined in the appended claims.

[0045] In addition, in the description of the present application, "a plurality of" means two or more, unless otherwise specified. The "and / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the associated objects before and after it are in an "or" relationship. The present application will be further described below in conjunction with the drawings and examples.

[0046] The present application will be further described below in conjunction with the drawings and examples.

[0047] Embodiment 1

[0048] Please refer to Figure 1 and Figure 2 A medical image segmentation model construction method based on federated learning, comprising the following steps:

[0049] S1, setting a personalized segmentation model and a universal segmentation model in each client, and initializing;

[0050] S2, in each client, learning the inverse network of the personalized encoder as a reconstructor using local data; S3, updating the personalized encoder and the universal encoder in each client according to the loss function value of the reconstructor;

[0051] S3, in each client, the reconstructor is spliced after the general encoder, the localized reconstruction data with global information is reconstructed, and the cyclic model distillation module is used to update the personalized segmentation model and the general segmentation model using the localized reconstruction data;

[0052] S4, upload the general encoder and the local segmenter of each client to the server, aggregate the global encoder and the global segmenter in the server;

[0053] S5, the server broadcasts the global encoder and the global segmenter to each client, and assigns the parameters of the global encoder and the global segmenter to the general encoder and the local segmenter;

[0054] S6, repeat steps S2 to S5 until the performance of the personalized segmentation model and the general segmentation model is stable.

[0055] Compared with the prior art, the embodiment sets two models (personalized segmentation model + general segmentation model) on each client to solve the problems of insufficient personalization and generalization ability brought by federated medical image segmentation, and has good performance in personalization and generalization ability. The cyclic model distillation module is used for local segmentation model update in each client, which weakens the influence of other clients on local training and effectively avoids model drift problem.

[0056] Embodiment 2

[0057] Please refer to Figure 1 and Figure 2 A medical image segmentation model construction method based on federated learning, comprising the following steps:

[0058] S1, set a personalized segmentation model and a general segmentation model in each client, and initialize.

[0059] In one specific embodiment, the personalized segmentation model , and the general segmentation model , respectively include the personalized encoder and the general encoder , and share the same local segmenter .

[0060] S2, learn the inverse network of the personalized encoder as a reconstructor using local data in each client.

[0061] Specifically, the data with labels of each client is used as the personalized encoder in each client Input, and personalize the encoder The output result of a local segmenter with fixed input parameters. In the process, the prediction results of the client model are obtained. Then the personalized encoder will be... The extracted semantic features are fed as input to the reconstructor. In the process, the reconstructed data is obtained. .

[0062] In one specific embodiment, the reconstructor is supervised and the personalized encoder is updated by a first Dice loss function and a reconstruction loss function.

[0063] It should be noted that the first Dice loss function is:

[0064]

[0065] in, This represents the amount of local data on client k. This represents the predicted label obtained by the personalized segmentation model for the i-th sample in client k. This represents the true label of the i-th sample in client k.

[0066] Specifically, the first Dice loss function is widely used in medical image segmentation tasks to measure the similarity between two sets. In the client k, our design goal is to make the model predict the results... Better and more realistic results Fitting to ensure personalized encoder We extract effective semantic features that can be used for segmentation. Specifically, we personalize our models for all local data. Prediction results Compared with the actual results Calculate the Dice loss, then sum and average the results. The smaller the Dice loss value, the closer the personalized segmentation model's prediction is to the true result, meaning the more effective the semantic features extracted by the personalized encoder for segmentation are.

[0067] It should be noted that the reconstruction loss function is:

[0068]

[0069] in, This represents the amount of local data on client k. This represents the i-th local data of client k. This represents the i-th reconstructed data of client k.

[0070] Specifically, the reconstruction loss function is introduced to enhance the reconstructor. The ability to recover local data from personalized features, the reconstructor It can be regarded as a personalized encoder The reverse process. We do this by reconstructing the data. With local data Calculate the Euclidean distance to apply constraints and ensure the reconstructor From personalized characteristics The ability to recover local data.

[0071] S3. In each client, the reconstructor is concatenated with the general encoder to reconstruct localized reconstructed data with global information, and the personalized segmentation model and the general segmentation model are updated using the localized reconstructed data by the cyclic model distillation module.

[0072] Specifically, the cyclic model distillation module first uses a universal encoder. and refactor In client k, global information is reconstructed for localization, and then the reconstructed data is used. Distillation optimizes both personalized and general segmentation models. This cyclic distillation method avoids the loss of generalization ability in the general segmentation model during local updates, while enabling the personalized segmentation model to indirectly learn information from other clients from the reconstructed data.

[0073] In one specific embodiment, in step S3, the updates of the personalized segmentation model and the general segmentation model are supervised by the second Dice loss function, the third Dice loss function, and the fourth Dice loss function.

[0074] Specifically, the second Dice loss function is used to supervise the personalized segmentation model. The effectiveness of local image segmentation; the third Dice loss function is used to supervise the reconstruction data. General segmentation model The update; the fourth Dice loss function is used to supervise the reconstruction of data. Personalized segmentation model Update.

[0075] Specifically, the purpose of designing the second Dice loss function is to improve the personalized segmentation model. Adaptability to local data.

[0076] It should be noted that the second Dice loss function is:

[0077]

[0078] in, This represents the amount of local data on client k. This indicates that the i-th sample in client k uses a personalized segmentation model. The obtained prediction labels, This represents the true label of the i-th sample in client k.

[0079] Specifically, the goal of designing the third Dice loss function is to avoid generic segmentation models. The generalization ability is lost during local updates. A general encoder, after model aggregation, has better global applicability and can extract general features from local data. Therefore, data reconstruction is necessary. Personalized biases in local data were removed by reconstructing the data. Optimize general segmentation models using local data Able to retain general segmentation model The ability to extract global information. This process utilizes a refactoring tool. For each general segmentation model Knowledge distillation is also a general segmentation model. Circulating distillation.

[0080] It should be noted that the third Dice loss function is:

[0081]

[0082] in, This represents the amount of local data on client k. The i-th reconstructed data in client k Using a general segmentation model The obtained prediction labels, This represents the true label of the i-th sample in client k.

[0083] Specifically, the goal of designing the fourth Dice loss function is to improve the personalized segmentation model. Learn global information. Due to data reconstruction. General coding features derived from global information Therefore, by utilizing reconstructed data Updating the model can indirectly learn knowledge from other clients. This process utilizes a generalized segmentation model. Personalized segmentation model Knowledge distillation.

[0084] It should be noted that the fourth Dice loss function is:

[0085]

[0086] wherein, represents the number of local data of the client k, represents the i-th reconstructed data in the client k using the personalized segmentation model the predicted label obtained, represents the true label of the i-th sample in the client k.

[0087] S4, upload the general encoder and the local segmenter of each client to the server, and aggregate the global encoder and the global segmenter in the server.

[0088] S5, the server broadcasts the global encoder and the global segmenter to each client, and assigns the parameters of the global encoder and the global segmenter to the general encoder and the local segmenter.

[0089] S6, repeat the steps S2 to S5 until the performance of the personalized segmentation model and the general segmentation model is stable.

[0090] Embodiment 3

[0091] A medical image segmentation model construction system based on federated learning, comprising a server and a plurality of clients;

[0092] The client is configured to train the personalized segmentation model and the general segmentation model set by the local data, and upload the general segmentation model to the server.

[0093] The server is provided with a model aggregation module and an aggregated model broadcast module.

[0094] The model aggregation module is configured to aggregate the general encoder and the local segmenter uploaded by the client to the server and generate a global encoder and a global segmenter.

[0095] The aggregated model broadcast module is configured to broadcast the global encoder and the global segmenter to the client.

[0096] Embodiment 4

[0097] This embodiment verifies and analyzes the method, more specifically:

[0098] This embodiment uses the retinal image dataset (Fundus) to evaluate the model generalization ability on the divided unknown dataset (experiment 1), and evaluates the model personalization ability (experiment 2).

[0099] Experimental setup: There are retinal fundus images from four different clinical centers in the retinal image dataset (Fundus), and the data sets of three of the four clinical centers are randomly divided into three clients for federated learning, and the remaining one data set is used as an unknown data set to evaluate the generalization performance of the federated medical image segmentation method (Experiment 1). The trained personalized segmentation learning network is further used to evaluate the model personalization performance in the three data sets for federated learning (Experiment 2). The experimental tasks are divided into optic disc (OD) and optic cup (OC) segmentation of retinal fundus images, and the evaluation indexes are Dice Coefficient (Dice) and Hausdorff Distance (HD).

[0100] Evaluation index description:

[0101] Dice Coefficient (Dice): a statistical quantity used to evaluate the similarity of two samples, which essentially measures the overlapping part of the two samples. Therefore, the higher the Dice Coefficient evaluation index, the better the model prediction effect, indicating that the performance of the corresponding federated medical image segmentation method is better.

[0102] Hausdorff Distance (HD): the maximum distance from one set to the nearest point in another set. Therefore, the smaller the index, the closer the two sets obtained, that is, the more accurate the experimental prediction value, indicating that the model prediction effect is better, indicating that the performance of the corresponding federated medical image segmentation method is better.

[0103] Experiment 1: Model generalization ability evaluation

[0104] Experiment result 1: The numerical analysis of the results of different federated medical image segmentation methods is shown in Table 1. The experimental results are compared with the current mainstream federated medical image segmentation methods (FedAVG [1], LGFed [2], FedPer [3], ELCFS [4]). Among them, FedAvg and ELCFS are federated domain generalization methods (FDG), which are mainly used to improve the generalization ability of the segmentation model, and LGFed and FedPer are personalized federated learning methods (perFL), which aim to improve the personalization level of the segmentation model. The results of the model trained by the client alone (Sep) are used as the performance benchmark (BL), and the results of the centralized training (Central) of the data sets of all known clients are used as the reference (UPP.B) together. The larger the Dice index score in Table 1, the better the model performance; the smaller the HD index score, the better the model performance.

[0105] Table 1 Experimental results of different federated medical image segmentation methods

[0106]

[0107] Results analysis: please refer to Figure 3 and Table 1, as can be seen from Table 1, in the performance (generalization performance) experiment of the model on the unknown data set, the federal domain generalization method (FedAvg and ELCFS) has relatively excellent generalization performance compared with the individualized federated learning method (LGFed and FedPer), and in the Dice index, the model index is at least improved by more than 23%; in the HD index, the federal domain generalization method is at least 2.49 times better than the individualized federated learning method. In addition, from Table 1, it can be seen that the method of the present application has better generalization performance than the current federal domain generalization method (FedAvg and ELCFS), and is the method with the best generalization performance of the model in this experiment except for the centralized training (Central).

[0108] Experiment 2: Model individualization performance evaluation

[0109] The experimental setup is similar to that of Experiment 1, the only difference being that in Experiment 1, the data collected by the clinical centers not used for federated learning is used as the unknown data set to evaluate the generalization performance of the federated medical image segmentation method. The performance of the federated medical image segmentation method in the known client's local data, i.e. individualization performance, is evaluated in Experiment 2.

[0110] Experimental results: the numerical analysis of the results of different federated medical image segmentation methods is shown in Table 2. The experimental results are compared with the current mainstream federated medical image segmentation methods (FedAVG [1], LGFed [2], FedPer [3], ELCFS [4]), among which FedAvg and ELCFS are federal domain generalization methods, mainly used to improve the generalization ability of the segmentation model, and LGFed and FedPer are individualized federated learning methods, aiming to improve the individualization level of the segmentation model. The results of the model trained by the client alone (Sep) are used as a reference, and the results of the centralized training (Central) of all known client data sets are used as a performance benchmark for comparison. In Table 2, the larger the Dice index score, the better the model performance; the smaller the HD index score, the better the model performance.

[0111] Table 2 Experimental results on known data sets

[0112]

[0113] Results analysis: as can be seen from Table 2, in the known data set, the individualized federated learning method (LGFed and FedPer) has relatively excellent performance compared with the federal domain generalization method (FedAvg and ELCFS). In addition, the individualization segmentation ability of the method of the present application is comparable to that of the individualized federated learning method.

[0114] From the comprehensive experiments 1 and 2, it can be concluded that: in the experiment 1 (generalization performance experiment), the method of the application is superior to the federated domain generalization method and the poor individualized federated learning method; in the experiment 2 (individualized performance experiment), the method of the application is equivalent to the individualized federated learning method (LGFed and FedPer) in performance. It can be seen that the generalization performance of the method of the application is much better than the individualized federated learning method, the individualized performance is better than the federated domain generalization method, and the model generalization ability and individualization ability can be considered at the same time.

[0115] Obviously, the above embodiments of the application are only examples for clearly illustrating the application, and are not intended to limit the embodiments of the application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments are not required to be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the claims of the application.

Claims

1. A method for constructing a medical image segmentation model based on federated learning, characterized in that, Includes the following steps: S1, Set up and initialize the personalized segmentation model and the general segmentation model in each client; S2 uses the inverse network of the personalized encoder learned from local data in each client as a reconstructor; S3, in each client, the reconstructor is spliced ​​after the general encoder to reconstruct localized reconstructed data with global information, and the cyclic model distillation module is used to update the personalized segmentation model and the general segmentation model with the localized reconstructed data. S4 uploads the general encoder and local segmenter from each client to the server, where they are aggregated to obtain the global encoder and global segmenter. S5, the server broadcasts the global encoder and the global segmenter to each client, and assigns the parameters of the global encoder and the global segmenter to the general encoder and the local segmenter; S6. Repeat steps S2 to S5 until the performance of the personalized segmentation model and the general segmentation model is stable. The cyclic model distillation module first uses a universal encoder. and refactor In client k, global information is reconstructed for localization, and then the reconstructed data is used. Distillation optimizes both personalized and general segmentation models.

2. The method for constructing a medical image segmentation model based on federated learning according to claim 1, characterized in that, The personalized segmentation model , and the general segmentation model , Each includes the personalized encoder and the general encoder and share the same local splitter .

3. The method for constructing a medical image segmentation model based on federated learning according to claim 1, characterized in that, The reconstructor is supervised and the personalized encoder is updated using the first Dice loss function and the reconstruction loss function.

4. The method for constructing a medical image segmentation model based on federated learning according to claim 3, characterized in that, The first Dice loss function is: in, This represents the amount of local data on client k. This represents the predicted label obtained by the personalized segmentation model for the i-th sample in client k. This represents the true label of the i-th sample in client k.

5. The method for constructing a medical image segmentation model based on federated learning according to claim 3, characterized in that, The reconstruction loss function is: in, This represents the amount of local data on client k. This represents the i-th local data of client k. This represents the i-th reconstructed data of client k.

6. The method for constructing a medical image segmentation model based on federated learning according to claim 1, characterized in that, In step S3, the updates of the personalized segmentation model and the general segmentation model are supervised by the second Dice loss function, the third Dice loss function, and the fourth Dice loss function.

7. The method for constructing a medical image segmentation model based on federated learning according to claim 6, characterized in that, The second Dice loss function is: in, This represents the amount of local data on client k. This indicates that the i-th sample in client k uses a personalized segmentation model. The obtained predicted labels, This represents the true label of the i-th sample in client k.

8. The method for constructing a medical image segmentation model based on federated learning according to claim 6, characterized in that, The third Dice loss function is: in, This represents the amount of local data on client k. This represents the i-th reconstructed data in client k. Using a general segmentation model The obtained predicted labels, This represents the true label of the i-th sample in client k.

9. The method for constructing a medical image segmentation model based on federated learning according to claim 6, characterized in that, The fourth Dice loss function is: in, This represents the amount of local data on client k. This represents the i-th reconstructed data in client k. Using a personalized segmentation model The obtained predicted labels, This represents the true label of the i-th sample in client k.

10. A medical image segmentation model construction system based on federated learning, characterized in that, Includes one server and multiple clients; The client is used to train its personalized segmentation model and general segmentation model based on local data, and to upload the general segmentation model to the server. The server is equipped with a model aggregation module and an aggregated model broadcasting module; The model aggregation module is used to aggregate the general encoder and local segmenter uploaded by the client to the server and generate a global encoder and a global segmenter. The aggregation model broadcast module is used to broadcast the global encoder and the global segmenter to the client; Set up and initialize the personalized segmentation model and the general segmentation model in each client; In each client, the inverse network of the personalized encoder, learned from local data, is used as the reconstructor. In each client, the reconstructor is spliced ​​after the general encoder to reconstruct localized reconstructed data with global information, and the personalized segmentation model and the general segmentation model are updated using the localized reconstructed data by the cyclic model distillation module. The general encoders and local segmenters of each client are uploaded to the server, and then aggregated on the server to obtain the global encoder and global segmenter. The server broadcasts the global encoder and the global segmenter to each client, and assigns the parameters of the global encoder and the global segmenter to the general encoder and the local segmenter; The cyclic model distillation module first uses a universal encoder. and refactor In client k, global information is reconstructed for localization, and then the reconstructed data is used. Distillation optimizes both personalized and general segmentation models.

Images