A Federal Unified Segmentation Method and System with Partial Category Labeling
By pre-training a local model in federated learning and adjusting the loss function using pseudo-labels and uncertainty-aware scores, the problems of class conflict and inefficiency in partial class labeling are solved, achieving more efficient multi-class image segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2023-08-16
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies suffer from problems such as class conflicts, failure to utilize inter-class association information, and test time being proportional to the total number of classes in federated learning with partial class labeling, resulting in low segmentation efficiency when there are many classes.
We employ pre-trained local models on each client, adjust the loss function using pseudo-labels and uncertainty-aware scores, aggregate the global model through the server, dynamically adjust the learning focus, reduce the impact of noise and pseudo-labels, and enhance the model's generalization ability.
It improves the ability to segment multiple categories in input images, reduces category conflicts, and enhances the model's generalization performance across various client domains.
Smart Images
Figure CN116994061B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image segmentation technology, and in particular to a federated unified segmentation method and system with partial category labeling. Background Technology
[0002] Current research methods for federated learning with partial category labeling mainly focus on learning each category separately. To address the category conflict issue in federated learning, these methods employ marginal mutually exclusive losses for updating the model based on partially labeled categories.
[0003] However, the method based on marginal mutually exclusive loss has the following shortcomings:
[0004] (1) Federated aggregation operations can cause knowledge conflicts between foreground and background classes.
[0005] (2) The inter-class association information was not utilized.
[0006] (3) The test duration is proportional to the total number of categories, which is not suitable for situations with a large number of categories. Summary of the Invention
[0007] The purpose of this invention is to provide a federated unified segmentation method and system with partial category labeling, which improves the ability to segment multiple categories in an input image.
[0008] To achieve the above objectives, the present invention provides the following solution:
[0009] A federally unified segmentation method with partial category labeling includes:
[0010] Each client uses local sample data to pre-train a local model and then uploads the pre-trained local model to the server; the categories used to segment the input image are different in the local sample data of each client.
[0011] The server distributes the received N pre-trained local models and the initialized global model to each of the clients; N represents the number of clients.
[0012] For the nth client, the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model; the current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client, and the current trained global model is uploaded to the server.
[0013] The server aggregates the N currently trained global models it receives to obtain the current global model, and then sends the current global model to each of the clients.
[0014] For the nth client, if the prediction accuracy of the current global model is less than a first set threshold, then the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model. If the prediction accuracy of the current global model is greater than or equal to the first set threshold, then the prediction output of the current global model is used as the source of the pseudo-labels of the local sample input data of the nth client. The current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client. The current trained global model is then uploaded to the server.
[0015] The server, in the return step, aggregates the N currently trained global models received to obtain the current global model, and distributes the current global model to each client until the prediction accuracy of the current global model is greater than a second set threshold. The current global model is then used as a general global segmentation model. The general global segmentation model is used to segment the image to be segmented and outputs segmentation results for multiple categories.
[0016] Optionally, the local sample data in each client includes input data and label data. The input data in each client consists of medical CT images of different organs, and the label data consists of the outlines of the corresponding organs.
[0017] Optionally, the pseudo-label of the nth client local sample input data is obtained by merging the foreground pseudo-label and the background pseudo-label. The foreground pseudo-label is obtained by merging the predicted foreground of all pre-trained teacher models, and the background pseudo-label is the intersection of the predicted background of all pre-trained teacher models.
[0018] Optionally, the prediction output of the current global model can be used as the source of the pseudo-label for the nth client-side local sample input data, specifically including:
[0019] Calculate the foreground intersection volume of the current global model's predicted output and the local model's predicted output of the nth client. If the foreground intersection volume is greater than a set volume value, then use the foreground intersection volume as the pseudo-label of the local sample input data of the nth client; otherwise, use the current global model's predicted output as the pseudo-label of the local sample input data of the nth client.
[0020] Optionally, the current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, to obtain the currently trained global model for the nth client, specifically including:
[0021] Using formula Calculate the uncertainty perception score;
[0022] The uncertainty-aware score is used to adjust the weights of the loss function when training the current global model;
[0023] Among them, U j N represents the uncertainty perception score. c w represents the total number of categories. T This represents the set of pre-trained local models, where vox represents voxels, and E... vox This represents the average entropy of all classes of voxels. This indicates that the voxel voxel belongs to category c and is a uniquely hot pseudo-label.
[0024] Optionally, the weights of the loss function are expressed as:
[0025]
[0026]
[0027] Among them, w(U j ) represents the weights of the loss function, norm(U j ) represents the normalized uncertainty-aware score, r represents the index of the current training epoch, R represents the total number of training epochs, μ represents the mean of the uncertainty-aware repository, and U max U represents the maximum value of the uncertainty-aware warehouse. min U represents the minimum value of the uncertainty-aware warehouse. τ This indicates the score threshold.
[0028] Optionally, the server aggregates the received N currently trained global models to obtain the current global model, and then distributes the current global model to each of the clients, specifically including:
[0029] When the server aggregates the N currently trained global models it receives, the aggregation weight of each currently trained global model is determined based on the amount of local sample data of each client.
[0030] This invention also discloses a partially categorized, federally unified segmentation system, comprising: N clients and one server;
[0031] Each client is used to pre-train a local model using local sample data and then uploads the pre-trained local model to the server; the categories used to segment the input image are different in the local sample data of each client.
[0032] The server is used to distribute the received N pre-trained local models and the initialized global model to each of the clients; N represents the number of clients;
[0033] For the nth client, the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model; the current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client; and the current trained global model is uploaded to the server.
[0034] The server is used to aggregate the received N updated local models to obtain the current global model, and then distribute the current global model to each of the clients.
[0035] For the nth client, if the prediction accuracy of the current global model is less than a first set threshold, then the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model; if the prediction accuracy of the current global model is greater than or equal to the first set threshold, then the prediction output of the current global model is used as the source of the pseudo-labels of the local sample input data of the nth client; the current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client, and the current trained global model is uploaded to the server;
[0036] The server returns to aggregate the received N updated local models to obtain the current global model, and distributes the current global model to each client until the prediction accuracy of the current global model is greater than a second set threshold. The current global model is then used as a general global segmentation model. The general global segmentation model is used to segment the image to be segmented and outputs segmentation results for multiple categories.
[0037] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0038] This invention uses data from each client to pre-train local models, and uses the pre-trained local models to provide pseudo-labels for federated training. When the current global model initially converges, the prediction output of the current global model is used as the source of pseudo-labels for the input data of the nth client's local sample. The current global model of the server is determined by the aggregation of the current global models of each client. After multiple global iterations, a general global segmentation model with good performance on the datasets of all clients is obtained, improving the ability to segment multiple categories in the input image. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0040] Figure 1 This is a schematic diagram of a federal unified segmentation method for partial category labeling provided in an embodiment of the present invention;
[0041] Figure 2 A schematic diagram illustrating the principle of determining pseudo-labels and foreground uncertainty perception scores in embodiments of the present invention. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] The purpose of this invention is to provide a federated unified segmentation method and system with partial category labeling, which improves the ability to segment multiple categories in an input image.
[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0045] The concept behind this invention, a federated unified segmentation method for partial category labeling, is as follows: The partial labeling problem in federated learning is transformed into a noisy label training problem. First, following the traditional segmentation training process, subsequent clients participating in federated learning pre-train organ-specific teacher models locally. Then, innovatively, based on the pseudo-labels and prediction confidence provided by the teacher models, all clients jointly train a global model capable of uniformly segmenting all categories. During training, the focus of learning categories at each stage is dynamically adjusted to reduce the impact of noisy pseudo-labels. Finally, the pseudo-labels are refined based on the interaction prediction between the global model and the locally pre-trained models, enhancing the generalization ability of the federated model across various client domains.
[0046] Example 1
[0047] like Figure 1 As shown in the figure, this embodiment provides a federal unified segmentation method for partial category labeling, which includes the following steps.
[0048] Step 101: Each client uses local sample data to pre-train a local model and uploads the pre-trained local model to the server; the categories used to segment the input image are different in the local sample data of each client.
[0049] In step 101, based on the local sample data in each client, a teacher model (local model) for the corresponding category is pre-trained using both dice loss and binary cross-entropy loss, while ensuring the training performance of the head and tail classes. The calculation formula is as follows:
[0050]
[0051]
[0052] L = 0.5 * L Dice +0.5*L BCE ;
[0053] Where X is the prediction output of the local model, Y is the label data, and N is the number of variables. c L represents the total number of categories, and L represents the loss function. Dice L represents the loss from dice rolls. BCE X represents the binary cross-entropy loss. i Y represents the local model's prediction output for i categories. i X represents i The corresponding tag data.
[0054] The head class refers to the class that occupies a relatively large area, while the tail class refers to the class that occupies a relatively small area. The head class and the tail class are the categories among all categories and belong to the prior data.
[0055] Step 102: The server distributes the received N pre-trained local models and the initialized global model to each of the clients; N represents the number of clients.
[0056] By distributing an initialized global model, all clients can start local training on the same initialized model, avoiding excessive differences in local models due to different initializations, which could affect the convergence speed of subsequent federated training.
[0057] The local model and the initialized global model have the same structure.
[0058] Step 103: For the nth client, use each pre-trained local model to determine the pseudo-label of the local sample input data of the nth client; use the local sample data with determined pseudo-labels to train the current global model corresponding to the nth client, obtain the current trained global model of the nth client, and upload the current trained global model to the server.
[0059] N is an integer greater than 1, and the value of n ranges from 1 to N.
[0060] like Figure 2 As shown, the pseudo-label of the nth client-side local sample input data is obtained by merging the foreground pseudo-label and the background pseudo-label. The foreground pseudo-label is obtained by merging the predicted foregrounds of all pre-trained teacher models, and the background pseudo-label is the intersection of the predicted backgrounds of all pre-trained teacher models. Pseudo-labels are used to avoid knowledge conflicts between categories.
[0061] The foreground refers to the area of interest and the area corresponding to the category, while the background refers to the area other than the foreground.
[0062] Step 104: The server aggregates the N currently trained global models received to obtain the current global model, and then distributes the current global model to each of the clients.
[0063] Step 105: For the nth client, if the prediction accuracy of the current global model is less than the first set threshold, then the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model. If the prediction accuracy of the current global model is greater than or equal to the first set threshold, then the prediction output of the current global model is used as the source of the pseudo-labels of the local sample input data of the nth client. The current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client. The current trained global model is then uploaded to the server.
[0064] Using the prediction output of the current global model as the source of pseudo-labels for the nth client's local sample input data specifically includes: calculating the foreground intersection volume of the prediction output of the current global model and the prediction output of the nth client's local model; if the foreground intersection volume is greater than a set volume value, then the foreground intersection volume is used as the pseudo-label for the nth client's local sample input data; otherwise, the prediction output of the current global model is used as the pseudo-label for the nth client's local sample input data.
[0065] Step 106: Determine whether the prediction accuracy of the current global model is greater than the second set threshold. If not, return to step 104; if yes, proceed to step 107.
[0066] Step 107: Use the current global model as a general global segmentation model; the general global segmentation model is used to segment the image to be segmented and output segmentation results for multiple categories.
[0067] The local sample data in each client includes input data and label data. The input data in each client consists of medical CT images of different organs, and the label data consists of the outlines of the corresponding organs.
[0068] The current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, resulting in the trained global model for the nth client. Specifically, this includes:
[0069] Using formula Calculate the uncertainty perception score;
[0070] The uncertainty-aware score is used to adjust the weights of the loss function when training the current global model;
[0071] Among them, U j N represents the uncertainty perception score. c w represents the total number of categories. T This represents the set of pre-trained local models, where vox represents voxels, and E... vox This represents the average entropy of all classes of voxels. This indicates that the voxel voxel belongs to category c and is a uniquely hot pseudo-label.
[0072] The weights of the loss function are expressed as follows:
[0073]
[0074]
[0075] Among them, w(U j ) represents the weights of the loss function, norm(U j ) represents the normalized uncertainty-aware score, r represents the index of the current training epoch, R represents the total number of training epochs, μ represents the mean of the uncertainty-aware repository, and U max U represents the maximum value of the uncertainty-aware warehouse. min U represents the minimum value of the uncertainty-aware warehouse. τ The score threshold, used to distinguish between head and tail categories, is predetermined and represents the uncertainty score. The uncertainty-aware repository stores each uncertainty-aware score for each client.
[0076] In the early stages of federated training, the head classes are easier to learn and relatively less affected by noisy labels. Therefore, uncertainty-aware scores are used to adjust the loss weights of the local model to accelerate its learning speed. Once the head classes are well-learned enough to guide the learning of other classes, the loss weights of the tail classes are gradually increased to ensure their learning effectiveness.
[0077] The adjusted loss function value is L multiplied by w(U) j ).
[0078] This paper proposes a global model-based approach to refine pseudo-labels. While adjusting the loss function ensures stable learning of coarse knowledge across all categories, each teacher model, trained only on its local dataset, struggles to generalize effectively to all categories. Specifically, pseudo-labels on other clients' datasets may exhibit greater noise. Therefore, a global model, aggregated from local models across all clients, is used to further refine global pseudo-labels to compensate for missing labels caused by domain offset issues.
[0079] Furthermore, since the federated global model is based on datasets from all clients (local sample data), it may suffer from oversmoothing predictions, potentially leading to oversegmentation of detailed regions such as edges and small objects. These regions are characterized by high uncertainty, which can be corrected using teacher predictions with relatively lower uncertainty. Specifically, the foreground intersection volume of the global model's predictions and the teacher model's predictions is first calculated. If this volume is large enough, it indicates that the teacher prediction's uncertainty is relatively low, and the intersection is used as the pseudo-label for subsequent federated training; otherwise, the global model's prediction is used as the pseudo-label. The calculation process is as follows:
[0080]
[0081] in, It's a fake tag. It is the prediction output of the global model. It is the predicted output of the pre-trained teacher model (the set of pre-trained local models). Set a percentage threshold for volume intersection (set volume value). Let c represent the c-th pre-trained teacher model.
[0082] The server aggregates the received N currently trained global models to obtain the current global model, and then distributes the current global model to each of the clients, specifically including:
[0083] When the server aggregates the N currently trained global models it receives, the aggregation weight of each currently trained global model is determined based on the amount of local sample data of each client.
[0084] The purpose of this invention is to address the shortcomings of marginal mutually exclusive losses by providing a federated unified segmentation method based on partial class labeling. This method leverages pseudo-labeling techniques from semi-supervised learning, enhancing the model's learning performance across all classes by integrating knowledge learned from the categories of each partially labeled part. Furthermore, a loss reweighting function based on confidence (uncertainty-aware score) ensures robustness to noisy labels during the learning process. This method can be used in conjunction with other pseudo-label-based label denoising methods to further improve the global model's segmentation capability for each class.
[0085] Example 2
[0086] This embodiment specifically provides a federally unified segmentation method for partial category labeling.
[0087] Assuming the WORD, AbdomenCT-1K, and AMOS segmentation datasets are used as training data for three clients, class a is selected as the labeling category for WORD, class b as the labeling category for AMOS, and classes c and d as the labeling categories for AbdomenCT-1K. The goal is to obtain segmentation results for all classes across the three clients using a single global model. First, based on the traditional centralized training paradigm, each client pre-trains a labeling-specific teacher segmentation network locally using dice loss and binary cross-entropy loss. All clients send this teacher network to the server, where it is then distributed along with the initialized global model. Next, according to the proposed pseudo-label processing method, the post-processed prediction results of all teacher models are used as pseudo-labels during training. These pseudo-labels are fed into the network for forward propagation to obtain the prediction results, which are then used to calculate the segmentation loss along with the pseudo-labels. When all clients calculate the loss based on the pseudo-labels, the loss is further adjusted according to the current training process using the proposed uncertainty-aware loss reweighting function, and the local global model is updated after gradient backpropagation. Then, all clients upload their updated local global models. The server uses the total number of local samples as the model aggregation weight and distributes the aggregated global model. Once the performance of the global model is consistently higher than that of the teacher model (local model), the predictions of the global model or the intersection of their predictions are used as pseudo-labels, also for local training in the federated phase. After several global iterations, a general global segmentation model with good performance on all clients' datasets is finally obtained. The specific steps of this embodiment's partially labeled federated unified segmentation method are as follows:
[0088] Step S1: Locally pre-train the teacher model based on the partially labeled dataset, that is, pre-train the teacher model using only partially labeled labels on the WORD, AbdomenCT-1K, and AMOS training sets.
[0089] Step S2: Upload the pre-trained teacher model to the server and distribute it together with the initialized global model. The two models together serve as the models required for federated training.
[0090] Step S3: Obtain pseudo-labels and uncertainty perception scores based on the teacher models; the foreground pseudo-labels are directly obtained by merging the predicted foregrounds of all teacher models, while the background pseudo-labels need to be obtained by taking the intersection to avoid recognition conflicts between categories; the calculation process of the uncertainty perception score is as follows:
[0091]
[0092] Among them, U j N represents the uncertainty perception score. c w represents the total number of categories. T This represents the set of pre-trained local models, where vox represents voxels, and E... vox This represents the average entropy of all classes of voxels. This indicates that the voxel voxel belongs to category c and is a uniquely hot pseudo-label.
[0093] Step S4: Calculate the loss using the input and pseudo-label pairs. The calculation process is as follows:
[0094]
[0095]
[0096] L = 0.5 * L Dice +0.5*L BCE ;
[0097] Where X is the model prediction, Y is the true label, and N is the ground truth. c It represents the total number of categories.
[0098] Step S5: Adjust the loss weights based on the uncertainty perception score and the loss reweighting function. The calculation process is as follows:
[0099]
[0100]
[0101] Where, μ,U max U min Let r represent the mean, maximum, and minimum values of the uncertainty-aware warehouse, r represent the index of the current training round, and R represent the total number of training rounds.
[0102] Step S6: Upload the local global model to the server for global model aggregation; the aggregation weight of the global model is determined by the amount of data in each customer's local dataset.
[0103] Step S7: When the global model initially converges, it serves as an additional source of pseudo-labels. First, calculate the volume of the foreground intersection between the global model's predictions and the teacher model's predictions. If this volume is large enough, use the intersection as the pseudo-label for subsequent federated training; otherwise, use the global model's prediction as the pseudo-label. The calculation process is as follows:
[0104]
[0105] in, It is a prediction from a global model. It is a prediction from a pre-trained teacher model. This is the percentage threshold for volume intersection.
[0106] Step S8: After several rounds of iteration, the global training converges, and finally a general global segmentation model with good performance on the entire customer dataset is obtained.
[0107] Example 3
[0108] This embodiment provides a partially categorized, federally unified segmentation system, comprising: N clients and one server.
[0109] Each client is used to pre-train a local model using local sample data and then uploads the pre-trained local model to the server; the categories used to segment the input image are different in the local sample data of each client.
[0110] The server is used to distribute the received N pre-trained local models and the initialized global model to each of the clients; N represents the number of clients.
[0111] For the nth client, the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model; the current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client, and the current trained global model is uploaded to the server.
[0112] The server is used to aggregate the received N updated local models to obtain the current global model, and then distribute the current global model to each of the clients.
[0113] For the nth client, if the prediction accuracy of the current global model is less than a first set threshold, then each pre-trained local model is used to determine the pseudo-label of the local sample input data of the nth client. If the prediction accuracy of the current global model is greater than or equal to the first set threshold, then the prediction output of the current global model is used as the source of the pseudo-label of the local sample input data of the nth client. The local sample data with determined pseudo-labels is used to train the current global model corresponding to the nth client to obtain the current trained global model of the nth client, and the current trained global model is uploaded to the server.
[0114] The server returns to aggregate the received N updated local models to obtain the current global model, and distributes the current global model to each client until the prediction accuracy of the current global model is greater than a second set threshold. The current global model is then used as a general global segmentation model. The general global segmentation model is used to segment the image to be segmented and outputs segmentation results for multiple categories.
[0115] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0116] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A federally unified segmentation method with partial category labeling, characterized in that, include: Each client uses local sample data to pre-train a local model and then uploads the pre-trained local model to the server; The categories used to segment the input image differ in the local sample data of each client; The server distributes the received N pre-trained local models and the initialized global model to each of the clients; N represents the number of clients. For the nth client, the pseudo-labels of the local sample input data of the nth client are determined using each pre-trained local model; the current global model corresponding to the nth client is trained using the local sample data with determined pseudo-labels to obtain the current trained global model of the nth client, and the current trained global model is uploaded to the server. The server aggregates the N currently trained global models it receives to obtain the current global model, and then sends the current global model to each of the clients. For the nth client, if the prediction accuracy of the current global model is less than the first set threshold, then each pre-trained local model is used to determine the pseudo label of the local sample input data of the nth client. If the prediction accuracy of the current global model is greater than or equal to the first set threshold, then the prediction output of the current global model is used as the source of the pseudo label of the local sample input data of the nth client. The current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, and the current trained global model of the nth client is obtained. The current trained global model is then uploaded to the server. The server, as described in the return step, aggregates the N currently trained global models received to obtain the current global model, and distributes the current global model to each client until the prediction accuracy of the current global model is greater than a second set threshold. The current global model is then used as a general global segmentation model. The general global segmentation model is used to segment the image to be segmented and outputs segmentation results for multiple categories. The current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, resulting in the trained global model for the nth client. Specifically, this includes: Using formula Calculate the uncertainty perception score; The uncertainty-aware score is used to adjust the weights of the loss function when training the current global model; in, The score represents the perceived uncertainty. Indicates the total number of categories. This represents the set of pre-trained local models. Represents voxels, Voxel representation The average entropy of all categories, Voxel representation A unique hot pseudo-tag belonging to category c.
2. The federalized unified segmentation method for partial category labeling according to claim 1, characterized in that, The local sample data in each client includes input data and label data. The input data in each client consists of medical CT images of different organs, and the label data consists of the outlines of the corresponding organs.
3. The federalized unified segmentation method for partial category labeling according to claim 1, characterized in that, The pseudo-label of the nth client local sample input data is obtained by merging the foreground pseudo-label and the background pseudo-label. The foreground pseudo-label is obtained by merging the predicted foreground of all pre-trained teacher models, and the background pseudo-label is the intersection of the predicted background of all pre-trained teacher models.
4. The federalized unified segmentation method for partial category labeling according to claim 1, characterized in that, The prediction output of the current global model is used as the source of the pseudo-label for the nth client local sample input data, specifically including: Calculate the foreground intersection volume of the current global model's predicted output and the local model's predicted output of the nth client. If the foreground intersection volume is greater than a set volume value, then use the foreground intersection volume as the pseudo-label of the local sample input data of the nth client; otherwise, use the current global model's predicted output as the pseudo-label of the local sample input data of the nth client.
5. The federalized unified segmentation method for partial category labeling according to claim 1, characterized in that, The weights of the loss function are expressed as follows: ； ； in, The weights represent the weights of the loss function. This represents the normalized uncertainty perception score. r This represents the index of the current training round. R Indicates the total number of training rounds. This represents the mean of the uncertainty-perceived warehouse. This represents the maximum value of the uncertainty-aware warehouse. This represents the minimum value of the uncertainty-aware warehouse. This indicates the score threshold.
6. The federalized unified segmentation method for partial category labeling according to claim 1, characterized in that, The server aggregates the received N currently trained global models to obtain the current global model, and then distributes the current global model to each of the clients, specifically including: When the server aggregates the N currently trained global models it receives, the aggregation weight of each currently trained global model is determined based on the amount of local sample data of each client.
7. A federally unified segmentation system with partial category labeling, characterized in that, include: N clients and one server; Each client is used to pre-train a local model using local sample data and then upload the pre-trained local model to the server. The categories used to segment the input image differ in the local sample data of each client; The server is used to distribute the received N pre-trained local models and the initialized global model to each of the clients; N represents the number of clients; For the nth client, pseudo-labels are used to determine the local sample input data of the nth client using each pre-trained local model; The current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, and the current trained global model of the nth client is obtained. The current trained global model is then uploaded to the server. The server is used to aggregate the received N updated local models to obtain the current global model, and then distribute the current global model to each of the clients. For the nth client, if the prediction accuracy of the current global model is less than the first set threshold, then each pre-trained local model is used to determine the pseudo label of the local sample input data of the nth client; if the prediction accuracy of the current global model is greater than or equal to the first set threshold, then the prediction output of the current global model is used as the source of the pseudo label of the local sample input data of the nth client. The current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, and the current trained global model of the nth client is obtained. The current trained global model is then uploaded to the server. The server returns to aggregate the received N updated local models to obtain the current global model, and distributes the current global model to each client until the prediction accuracy of the current global model is greater than a second set threshold. The current global model is then used as a general global segmentation model. The general global segmentation model is used to segment the image to be segmented and outputs segmentation results for multiple categories. The current global model corresponding to the nth client is trained using local sample data with determined pseudo-labels, resulting in the trained global model for the nth client. Specifically, this includes: Using formula Calculate the uncertainty perception score; The uncertainty-aware score is used to adjust the weights of the loss function when training the current global model; in, The score represents the perceived uncertainty. Indicates the total number of categories. This represents the set of pre-trained local models. Represents voxels, Voxel representation The average entropy of all categories, Voxel representation A unique hot pseudo-tag belonging to category c.