An adaptive decision federated learning method for medical image recognition

Abstract

The application belongs to the technical field of medical image recognition based on computer models, and discloses a self-adaptive decision federal learning method for medical image recognition, which comprises the following steps: setting a reinforcement learning agent model; taking the parameter difference index of each client model and other client models, the prediction consistency index of each client model, the local medical image data category set of each client, and the prediction accuracy of each client model as the input state vector of the reinforcement learning agent model, taking the aggregation weight of the parameter of each client model as the action vector of the reinforcement learning agent model, and training to obtain an optimal aggregation parameter decision model; each client model respectively performs local training by using the local medical image data to obtain the update parameter and real-time state vector of the client model; inputting the real-time state vector into the optimal aggregation parameter decision model to obtain the aggregation weight of the parameter of the client model; and aggregating the update parameter of the client model to obtain the update parameter of the global model.

Description

Technical Field

[0001] This invention belongs to the field of medical image recognition technology based on computer models, and specifically relates to an adaptive decision federated learning method for medical image recognition. Background Technology

[0002] Medical imaging data is usually stored in different hospitals and medical centers. Due to factors such as patient privacy protection, medical ethics, data security policies, and restrictions on cross-institutional data sharing, it is difficult to centralize medical imaging data from multiple medical institutions onto a single server for unified training.

[0003] Federated learning provides a feasible technical approach for collaborative training of medical image analysis models across multiple medical institutions. This technology allows each institution to retain its original medical image data locally, uploading only model parameters, gradients, or statistical features to a central server for aggregation, thus achieving collaborative model optimization without requiring the data to leave the local facility. However, in medical imaging scenarios, significant differences can exist among different institutions in their acquisition equipment, scanning protocols, magnetic field strength, patient origins, lesion category ratios, and annotation standards, leading to significant non-independent and identically distributed characteristics in the client-side data. For example, in brain tumor classification tasks, the proportions of gliomas, meningiomas, pituitary adenomas, and non-tumor cases may differ across hospitals. Client-side models are prone to shifting towards their local dominant category or local optimum, causing model drift and reducing the convergence speed and diagnostic generalization ability of the global model.

[0004] The FedAvg algorithm is one of the earliest and most widely used federated learning algorithms. It provides a concise and efficient benchmarking method by aggregating the parameters of local client models. Under independent and identically distributed (IID) data settings, it exhibits strong convergence properties. However, in practical applications, the data distribution of each client is often non-independent and IID, i.e., statistical heterogeneity exists. This data heterogeneity causes each client to update towards its own local optimum during local training, resulting in model drift between the client models and the global model, severely degrading the convergence speed and generalization performance of the global model.

[0005] The existing federated learning methods for medical image recognition have several problems, including: (1) Existing federated learning methods for medical imaging are insufficient in suppressing model drift for non-independent and identically distributed data: In multi-center medical imaging applications, there are differences in case structure, scanning protocols, and lesion category ratios among different medical institutions, resulting in inconsistent update directions of client-side models; traditional federated averaging methods usually perform static weighted aggregation based on sample size, making it difficult to identify whether the update directions of different clients conform to the global medical diagnostic goals. Especially when the lesion category distribution is uneven, local models are prone to overfitting to the local dominant category, which in turn causes global model drift and convergence instability.

[0006] (2) Existing adaptive federated learning methods for medical images lack client-global target alignment at the knowledge level: In medical image classification tasks, even if different clients use the same network structure, they may learn different classification boundaries due to different local case compositions; adjusting the aggregation ratio only by gradient, participation or parameter update magnitude is still difficult to ensure that the client model is consistent with the global model at the predictive behavior level, and it is especially difficult to utilize the knowledge of similarity between lesion categories contained in the global model.

[0007] (3) Existing aggregation strategies lack a continuous adaptive weight decision mechanism oriented to client state: The quality of clients in medical image federated learning is not constant. In a certain round of communication, the client model may make different contributions to the global model due to differences in local class distribution, training loss, knowledge distillation effect or parameter offset. If only static data weights, fixed participation or heuristic weights are used, the aggregation strategy cannot be adjusted in real time according to the client state in each round, which can easily cause client updates with large deviations to have a negative impact on the global medical diagnostic model.

[0008] In summary, existing federated learning methods for medical images still have technical gaps in areas such as global knowledge utilization, consistency alignment, model drift correction, and adaptive aggregation weight decision-making. Therefore, there is an urgent need for a federated learning method for medical imaging applications that can improve the robustness and generalization ability of medical image classification and diagnostic models without sharing the original medical image data. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive decision federated learning method for medical image recognition, which can improve the recognition accuracy, convergence speed and stability of the global model for medical images.

[0010] The technical solution provided by this invention is as follows: An adaptive decision federated learning method for medical image recognition includes the following steps: A federated learning system is formed by a central server and multiple medical institution clients. The central server is equipped with a global model and a reinforcement learning agent model, and each medical institution client is equipped with a client model. Each medical institution's client stores local medical image data; The reinforcement learning agent model's state vector is input using the parameter difference index between each client model and other client models, the prediction consistency index of each client model, the local medical image data category set of each client, and the prediction accuracy of each client model. The aggregate weights of the parameters of each client model are used as the action vector of the reinforcement learning agent model. The reinforcement learning agent model is trained to obtain the optimal aggregate parameter decision model. Each client model is trained locally using local medical image data to obtain updated parameters for the client model, and the real-time state vector is obtained using the training results of the client model. The real-time state vector is input into the optimal aggregation parameter decision model, and the optimal aggregation parameter decision model outputs the aggregate weights of the parameters of each client model. The update parameters of the client models are aggregated using the aggregated weights of the parameters of each client model to obtain the update parameters of the global model. The global model parameters are iteratively updated to obtain the optimal global model, which is used to identify medical images.

[0011] Preferably, the adaptive decision federated learning method for medical image recognition further includes: The central server stores the proxy dataset; During local training on each client, each client model and the global model predict samples in the proxy dataset, and the global model outputs the results as soft labels. The knowledge distillation loss is calculated based on the output results of the client models and the soft labels, and the loss function for training the client models is constructed using the knowledge distillation loss.

[0012] Preferably, the loss function for training the client-side model is: ; in, The loss function for training the client-side model. For supervised learning loss of the client model, For dynamic regularization of the client model, The knowledge distillation loss is used for the client-side model.

[0013] Preferably, the formula for calculating the knowledge distillation loss is: ; in, The cross-entropy loss is the difference between the client-side model output and the local ground truth labels. The dispersion between the client-side model output and the soft labels. This is the balance coefficient.

[0014] Preferably, the formula for calculating the parameter difference index between the client model and other client models in the real-time state vector is as follows: ; in, Indicates the first Round-based client model The parameter difference index with other client models, For the first Round-based client model and client model The pairwise cosine similarity between the updated parameters, Indicates the first The set of client models that participate in communication in rounds. Indicates the first The number of client models participating in communication rounds. This represents the iteration round of the global model.

[0015] Preferably, the formula for calculating the prediction consistency index of the client model in the real-time state vector is: ; in, Indicates the first Round-based client model The predictive consistency index; Indicates the first Round-based client model Category The predictive probability of medical images, Indicates the first Round of all client models participating in communication for each category The average predictive probability of medical images, Indicates the category of medical images. This indicates the total number of categories of medical images.

[0016] Preferably, the reward function used in the reinforcement learning agent model is: ; in, Indicates the first Performance improvement rewards for the global model in each round Indicates the first Round consistency alignment reward Indicates the first The weighting coefficient for round consistency alignment , This indicates the total number of communication rounds.

[0017] Preferably, the formula for calculating the performance improvement reward of the global model is as follows: ; in, Indicates the first Accuracy of global model validation in each round Indicates the first The change in single-round verification accuracy across rounds. Indicates from the initial round to the... Time series of changes in single-round validation accuracy. Representing time series The length.

[0018] Preferably, the formula for calculating the consistency alignment reward is: ; in, Indicates the first The average knowledge distillation loss of all client models participating in the communication round. Indicates from the initial round to the... The time series of the average knowledge distillation loss of all client models participating in the communication rounds. Representing time series The length.

[0019] The beneficial effects of this invention are: (1) This invention distills medical image soft label knowledge into the local training process of each client, so that the client can not only learn the brain tumor MRI image features of its own institution, but also learn the similarity and diagnostic distribution information between brain tumor categories expressed by the global model, thus avoiding the problem of inconsistency between local and global goals at the training goal level.

[0020] (2) The present invention incorporates the current global model parameters and the client’s historical drift parameters into the local loss function to suppress the local model shift caused by non-independent and identically distributed brain tumor image data and improve the convergence stability during multi-round communication training.

[0021] (3) This invention uses the DDPG reinforcement learning agent to model the aggregation weight allocation as a continuous control problem. It comprehensively considers multi-dimensional state information such as client parameter offset, knowledge distillation loss, prediction distribution difference, category distribution and client performance, and dynamically outputs aggregation weights to avoid low-quality update interference caused by static sample weighting.

[0022] (4) This invention constructs a closed-loop federated learning system that integrates global knowledge generation, client consistency alignment, reinforcement learning adaptive aggregation and reward feedback update, thereby improving the generalization ability, robustness and diagnostic stability of medical image recognition models without leaving the local medical image data. Attached Figure Description

[0023] Figure 1 This is a flowchart of the adaptive decision federated learning method for medical image recognition described in this invention (taking brain tumor MRI images as an example). Figure 2 The convergence curve of the BRISC medical dataset under the Dir(0.1) distribution in the experimental examples of this invention is shown. Figure 3 The convergence curve of the BRISC medical dataset under the Dir(0.5) distribution in the experimental examples of this invention is shown. Figure 4 This is a schematic diagram showing the change of aggregate weights for each client model during each round of communication under the Dir(0.1) distribution in the experimental example of this invention; Figure 5 This is a schematic diagram showing the change of aggregate weights for each client model during each round of communication under the Dir(0.5) distribution in the experimental example of this invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0025] This invention provides an adaptive decision federated learning method for medical image recognition, designed to address the heterogeneity of medical images. Without sharing the original medical image data, it can improve the robustness and generalization ability of medical image classification-assisted diagnosis models.

[0026] like Figure 1 As shown below, using brain tumor MRI image data as medical image data as an example, the adaptive decision federated learning method for medical image recognition provided by this invention will be further explained.

[0027] A central server and multiple medical institution clients constitute a federated learning system. Each medical institution client stores local brain tumor MRI image data, which includes categories such as gliomas, meningiomas, pituitary tumors, and non-tumor cases. The raw brain tumor MRI image data is not uploaded to the central server; the central server only receives the model parameters, status statistics, and necessary performance indicators trained by the clients.

[0028] Let the global model parameter vector be represented as ,in This indicates the dimension of the model. The federated learning system consists of several healthcare institution clients, and these clients are comprised of... Index. Each healthcare institution client They all have a size of Private local medical image dataset },in Indicates the client The local medical image dataset The feature input of each sample, Indicates the client The local medical image dataset The true labels of each sample are not accessible to the local datasets of other medical institution clients or the central server. The local model parameters for each medical institution client are: Because the medical institution's client and the local model correspond one-to-one, the first The local model of a medical institution client is denoted as the client model. , No. One client (client model) The experience risk is defined as: ; in, ) This indicates standard monitoring losses. Indicates the client The local medical image dataset The feature input of each sample, Indicates the client The local medical image dataset The true labels for each sample are then updated. Finally, the global model is updated while outputting a diagnosis of brain tumor imaging.

[0029] ; in, It is aggregate weight. Indicates the first The global model parameters after each iteration. It is a... The dimension vector is the solution to the optimization problem in the current round. This means finding the value that minimizes the objective function. value.

[0030] In traditional federated learning, It is defined by the amount of data. In this invention, it is dynamically adjusted through an adaptive decision aggregation strategy. At the beginning of each round of communication, the server maintains a global model. And a reinforcement learning agent model, using the reinforcement learning agent model to output aggregate weights.

[0031] Unlike traditional fixed aggregation weights based on data volume, this invention achieves adaptive dynamic decision-making for aggregation weights through a reinforcement learning agent model. This mechanism comprehensively utilizes multi-dimensional state information from the client, such as model similarity, prediction distribution consistency, and training history, to tailor the optimal aggregation strategy for each round of communication. Its core advantages lie in: breaking the single limitation of data volume, real-time perception of changes in Non-IID data distribution, guiding stable convergence with long-term cumulative rewards as the goal, and avoiding reliance on manually heuristic rules. Ultimately, this achieves a comprehensive improvement in the global model's accuracy, convergence speed, and stability.

[0032] To facilitate cross-client knowledge transfer without sharing the original MRI images of brain tumors, this invention constructs a proxy dataset for knowledge distillation, denoted as... Specifically, a knowledge distillation agent dataset is constructed using a small public validation set; based on the small public validation set... The samples in the dataset are all fed into the global model to obtain the corresponding soft predictions: ; in, For parameters The global model provides soft-label information for proxy samples. Based on this, the proxy dataset is constructed as follows: ; in, Represents the first in the agent dataset One input sample, Represents the global model for samples The original score vector generated in the output layer. Representative input sample The true label.

[0033] Each client performs local optimization. After local training, the clients upload their parameters and related statistics to form a state representation. Reinforcement learning agent models are based on state. Output a continuous aggregate weight vector Then, the server calculates the reward based on the model's performance. This updates the policy / value network of the reinforcement learning model, thereby enhancing subsequent decision weights. Finally, the server aggregates the client updates with the learned weights to obtain the next global model parameters. .

[0034] (1) Consistent alignment during local training This invention designs a consistent alignment mechanism with knowledge fusion to improve the performance of medical institution clients. Specifically, it assumes that each medical institution client can access an unlabeled proxy dataset from a central server. The local training provided for each client model aims to minimize the following total loss: ; in, The loss function for training the client-side model. For supervised learning loss of the client model, For dynamic regularization of the client model, The knowledge distillation loss is used for the client-side model.

[0035] Knowledge fusion distillation loss based on knowledge extraction This is one of the core components of the invention, responsible for fusing global knowledge with local models. The loss function is defined as: ; in, It is a balance coefficient that controls the trade-off between hard loss and soft loss. It is the output of the medical institution client model. and local real value labels The cross-entropy loss between them ensures the basic discriminative ability of the model; It is the output of the client model. And soft labels obtained from the global knowledge base and output by the global model The KL divergence between them. To capture the inter-class similarity encoded in the global model, temperature-scaled Softmax is used to generate soft labels.

[0036] Therefore, the hard-label loss is given by the standard cross-entropy: ; in, This indicates that the student model (client model) is related to the sample. Output as categories The score, This indicates that the student model (client model) is related to the sample. Output as categories The score, This is a category index, representing any category. This refers to the number of categories. To effectively transmit the inter-class similarity information captured by the global model during knowledge distillation, we obtain the similarity information for proxy samples from both the teacher model (global model) and the student model (client model). The output is then converted into a soft probability distribution using temperature-scaled Softmax. Specifically, let... and These represent the teacher model and student model for the samples, respectively. The scores on both are then defined as follows: ； ; in, The soft probability distribution for the teacher model (global model), For the soft probability distribution of the student model (client model), This indicates that the student model (client model) is related to the sample. Output as categories The score, The category index represents any category. Therefore, the soft-label loss is defined as the Kullback-Leibler (KL) divergence between the teacher and student distributions: .

[0037] and The calculation methods are as follows: ; ; in, This refers to a small batch of data randomly sampled from the client's local medical image dataset. Indicates batch size, i.e., small batch of data. Number of samples included For the first in the small batch data One input sample, For the first in the small batch data A real label, This is the model's predicted output. It is the cross-entropy function; For client model The regularization coefficient, They are respectively Wheel and Wheel's client model parameter, These are global model parameters.

[0038] (2) Adaptive decision aggregation strategy To learn the aggregation policy, we formulate the problem as a Markov Decision Process (MDP). Since the action space is continuous, a deterministic policy gradient framework is required. Therefore, we employ Deep Deterministic Policy Gradient (DDPG), which is well-suited for continuous control tasks with an Actor-Critic architecture. One of the core innovations of this invention is the transformation of the aggregation weight assignment problem into a continuous control problem using a reinforcement learning framework. The reinforcement learning agent model can adaptively adjust the aggregation weights of different clients in each communication round by learning the policy network, thereby optimizing the training process of the global model. Specifically, the training process involves the joint optimization of the Critic network and the Actor network, supported by an experience replay buffer to achieve stable and efficient learning. The core role of the Critic network is to evaluate the performance of a given state. The following actions The long-term value of this is considered. Therefore, we employ a state-action value function based on the Bellman equation, using a cumulative reward with an approximate discount to quantify the merits of the current decision: ; in, It is determined by the parameter as The state-action value output by the Critic network represents the state from the current round. Depart and execute the action Then, based on the current strategy, the expected cumulative discount reward can be obtained through continued decision-making. The discount factor is used. The core of Critic network training lies in constructing a stable target value to calculate the temporal difference error. Therefore, a target Critic network is introduced. and target Actor network Target value The calculation method is as follows: ; in, Represented as the time-series difference objective value, Discount factor; based on target value , Representing the reward for the current round, the Critic network is trained by minimizing the following loss function (Mean Squared Error, MSE): ; in, This serves as a replay buffer. Furthermore, the Actor network aims to maximize the expected Q-value estimated by the Critic network, i.e., ; in, These are the Actor network parameters. Using the deterministic policy gradient theorem, the gradient of the objective is given by the following equation: .

[0039] Since the strategy is deterministic, exploration is achieved by adding noise to the participants' output during training: ; in, It is typically modeled as Ornstein-Uhlenbeck (OU) noise to generate time-dependent perturbations.

[0040] (a) State In each round of communication, the state representation is designed to comprehensively capture the model and data characteristics of the participating clients. Specifically, the state characteristics consist of the following parts: Parameter variance, this metric reflects the degree of deviation in client-side parameters. To quantify the differences in model parameters between clients, we extracted parameter vectors. , No. One client (client model) ) parameter vector and the One client (client model) ) parameter vector The pairwise cosine similarity between them is defined as: .

[0041] Next, the average similarity between each client and other clients is calculated to form the parameter difference index: ; in, Indicates the first Round-based client model The parameter difference index with other client models, For the first Round-based client model and client model The pairwise cosine similarity between the updated parameters, Indicates the first The set of client models that participate in communication in rounds. Indicates the first The number of client models participating in communication rounds. This represents the iteration round of the global model.

[0042] Model consistency index, calculated as follows The average predicted probability of class c for all participating client models in each round is used to characterize the consistency among client-level prediction distributions. ; in, Indicates the first Round-based client model Brain tumor imaging categories The predicted probabilities are then calculated, and the formula for calculating the prediction consistency index of each client model is given based on the average distribution of KL divergence: ; in, Indicates the first Round-based client model The predictive consistency index; Indicates the first Round-based client model Category The predictive probability of medical images, Indicates the first Round of all client models participating in communication for each category The average predictive probability of medical images, Indicates the category of medical images. This indicates the total number of categories of medical images.

[0043] Client-side data distribution, and the distribution of local medical image data categories stored on each client. It is used directly as part of the feature vector to reflect data heterogeneity.

[0044] Client-side performance refers to the accuracy of each client-side model on the local validation set. Standardization is used as a performance indicator.

[0045] Finally, the state vector is constructed as Each component of the state is Min-Max normalized to ensure comparability and robustness in terms of scale.

[0046] (b) Action In this invention, the role of the reinforcement learning agent model is to obtain a specific aggregation weight. Specifically, by Parameterized Actor networks directly output deterministic actions. , indicating the current state Aggregate weights.

[0047] (c) Rewards To effectively guide the federated aggregation process, we designed a composite reward function that simultaneously considers the effectiveness of model performance improvement and consistency alignment. Then, in the... In round-robin communication, the reward function is defined as: ; in, Indicates the first Performance improvement rewards for the global model in each round Indicates the first Round consistency alignment reward Indicates the first The weighting coefficient for round consistency alignment , This represents the total number of communication rounds. A lower average distillation loss implies better alignment between the teacher and student models, resulting in a higher reward. The overall reward function integrates performance improvement rewards and consistency alignment rewards. It provides an effective assessment of the quality of global model improvement and knowledge fusion.

[0048] The first term of the reward function ( This reflects the performance improvement of the global model and is defined as: ; in, Indicates the first Accuracy of global model validation in each round Indicates the first The change in single-round verification accuracy across rounds. ; Indicates from the initial round to the... Time series of changes in single-round validation accuracy. Representing time series The length. It significantly suppressed unstable fluctuations during training and promoted a steady improvement in model performance.

[0049] The second term of the reward function is expressed as It is used to evaluate the effectiveness of the knowledge fusion process. It is defined as: ; in, Indicates the first The average knowledge distillation loss of all client models participating in the communication round. Indicates from the initial round to the... The time series of the average knowledge distillation loss of all client models participating in the communication rounds. Representing time series The length of the average historical loss is used to quantify the convergence behavior of the distillation process. A smaller average loss corresponds to a higher reward, indicating more successful knowledge fusion.

[0050] Once the model training reaches the preset number of communication rounds or the global loss convergence condition, the central server obtains the final brain tumor image classification model. This model can be used to input brain tumor MRI images and output category prediction results, providing a reference for brain tumor auxiliary diagnosis. This embodiment is mainly aimed at brain tumor MRI image classification tasks, but it can also be extended to brain tumor image segmentation, detection, or multi-task auxiliary diagnosis without changing the core federated training framework.

[0051] This invention employs a dual-mechanism collaborative framework of consistency alignment and adaptive decision-making, organically combining knowledge fusion and reinforcement learning aggregation mechanisms to collaboratively mitigate model drift from two dimensions: training target alignment and aggregation weight optimization. The client-global target alignment method based on knowledge fusion distills the soft-label knowledge of the global model into the client training process, enabling the client to simultaneously learn the distribution characteristics of local and global data, achieving proactive alignment between local and global targets, and allowing the client to actively gravitate towards the global optimization target during training. The adaptive aggregation weight decision-making method based on reinforcement learning models the aggregation weight allocation as a Markov decision process, using a deep deterministic policy gradient algorithm to train the agent and achieve optimal weight output in dynamic environments. A composite reward function is designed, simultaneously considering performance improvement and knowledge fusion effects, and a dynamic decay coefficient is designed to balance the weights of both. Regarding global knowledge utilization, existing technologies require indirect construction of global knowledge through the client model, resulting in limited representativeness. This invention, however, directly constructs a teacher knowledge base based on the global model output, achieving higher knowledge accuracy.

[0052] Test case The performance of the adaptive decision federated learning method for medical image recognition provided in this invention is verified by simulation. The specific process is as follows.

[0053] (1) Dataset This experimental example uses the BRISC dataset, a recently released large-scale annotated set of brain tumor MRI scans. It comprises 6000 high-resolution, contrast-enhanced T1-weighted MRI images covering axial, coronal, and sagittal planes, with manual annotations provided by certified radiologists. This dataset includes three major tumor categories (glioma, meningioma, and pituitary adenoma) as well as non-tumor cases, providing a balanced and diverse sample for robust model evaluation. BRISC is a valuable benchmark for developing and validating deep learning methods in medical image analysis.

[0054] To comprehensively evaluate the performance of the proposed method on non-independent identically distributed (Non-IID) data, a Dirichlet distribution is used to generate label distributions on each client. Specifically, the first strategy utilizes the Dirichlet distribution to simulate label skew in real-world scenarios, setting concentration parameters accordingly. (denoted as Dir(0.1)) and (denoted as Dir(0.5)). Dir(0.1) corresponds to extremely heterogeneous scenarios (each client contains only a few categories), and Dir(0.5) corresponds to moderately heterogeneous scenarios. Both are used to verify the robustness of the method under different degrees of skewness.

[0055] The medical image dataset is divided into three functional subsets according to a preset ratio: training set, test set, and validation set, with the ratios being 70%, 20%, and 10%, respectively; wherein, the validation set is used to construct a proxy dataset in the knowledge distillation process.

[0056] (2) Performance comparison with existing methods The proposed FedCAW method is systematically compared with eleven widely adopted federated learning algorithms to comprehensively evaluate its performance. These comparison methods cover a variety of representative strategies for dealing with data heterogeneity, including: classic benchmark methods FedAvg, FedProx, FedNova, and FedDyn; MOON based on contrastive learning; FedDC based on distribution correction; and FedSoup, Fed-RoD, FedCSD, FedUC, and FedCD for personalized federated learning. By covering a diverse range of comparison methods from classic aggregation strategies to the latest advancements, we are able to comprehensively verify the effectiveness and superiority of FedCAW from multiple dimensions. The experimental results are shown in Table 1.

[0057] Table 1 Experimental results on the medical dataset

[0058] As can be seen from Table 1, in terms of adaptability to heterogeneous scenarios, the performance of existing technologies degrades significantly in highly heterogeneous scenarios, while the present invention maintains excellent performance even in extreme heterogeneous scenarios such as Dir(0.1) and ill-conditioned distributions. Figures 2-3 As shown, in terms of convergence speed, this invention converges faster, exhibits less fluctuation, and demonstrates higher stability. In medical applications, existing technologies achieve an accuracy of approximately 85% to 92% on real-world medical datasets such as BRISC, while this invention achieves an accuracy exceeding 98%, demonstrating a significant advantage. In summary, this invention outperforms existing technologies in terms of aggregation adaptability, target consistency, knowledge utilization efficiency, robustness to heterogeneous scenarios, convergence performance, and medical application effectiveness.

[0059] (3) Aggregate weight analysis This experimental example also analyzes the dynamics of aggregation weights in the FedCAW framework proposed on CIFAR-10 under different heterogeneous settings. For example... Figures 4-5 As shown, under the strong heterogeneity of Dir (0.1), the aggregation weights among clients fluctuate greatly, with some clients repeatedly receiving higher weights in consecutive communication rounds. In contrast, under Dir (0.5), the aggregation weights exhibit a more balanced and stable distribution. This phenomenon highlights the effectiveness of our adaptive aggregation strategy. By utilizing the combined state and reward function, the system can dynamically identify clients whose update behavior is more in line with the global optimization objective, thereby effectively mitigating model drift.

[0060] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. An adaptive decision-making federated learning method for medical image recognition, characterized in that, include: A federated learning system is formed by a central server and multiple medical institution clients. The central server is equipped with a global model and a reinforcement learning agent model, and each medical institution client is equipped with a client model. Each medical institution's client stores local medical image data; The reinforcement learning agent model's state vector is input using the parameter difference index between each client model and other client models, the prediction consistency index of each client model, the local medical image data category set of each client, and the prediction accuracy of each client model. The aggregate weights of the parameters of each client model are used as the action vector of the reinforcement learning agent model. The reinforcement learning agent model is trained to obtain the optimal aggregate parameter decision model. Each client model is trained locally using local medical image data to obtain updated parameters for the client model, and the real-time state vector is obtained using the training results of the client model. The real-time state vector is input into the optimal aggregation parameter decision model, and the optimal aggregation parameter decision model outputs the aggregate weights of the parameters of each client model. The update parameters of the client models are aggregated using the aggregated weights of the parameters of each client model to obtain the update parameters of the global model. The global model parameters are iteratively updated to obtain the optimal global model, which is used to identify medical images.

2. The adaptive decision-making federated learning method for medical image recognition according to claim 1, characterized in that, Also includes: The central server stores the proxy dataset; During the local training process on each client, each client model and the global model predict the samples in the proxy dataset, and the global model outputs the results as soft labels. The knowledge distillation loss is calculated based on the output of the client model and the soft labels, and the loss function for training the client model is constructed using the knowledge distillation loss.

3. The adaptive decision-making federated learning method for medical image recognition according to claim 2, characterized in that, The loss function for training the client-side model is: ； in, The loss function for training the client-side model. For supervised learning loss of the client model, For dynamic regularization of the client model, The knowledge distillation loss is used for the client-side model.

4. The adaptive decision-making federated learning method for medical image recognition according to claim 3, characterized in that, The formula for calculating the knowledge distillation loss is as follows: ； in, The cross-entropy loss is the difference between the client-side model output and the local ground truth labels. The dispersion between the client-side model output and the soft labels. This is the balance coefficient.

5. The adaptive decision-making federated learning method for medical image recognition according to any one of claims 2-4, characterized in that, In the real-time state vector, the formula for calculating the parameter difference index between the client model and other client models is as follows: ； in, Indicates the first Round-based client model The parameter difference index with other client models, For the first Round-based client model and client model The pairwise cosine similarity between the updated parameters, Indicates the first The set of client models that participate in communication in rounds. Indicates the first The number of client models participating in communication rounds. This represents the iteration round of the global model.

6. The adaptive decision-making federated learning method for medical image recognition according to claim 5, characterized in that, In the real-time state vector, the formula for calculating the prediction consistency index of the client model is: ； in, Indicates the first Round-based client model The predictive consistency index; Indicates the first Round-based client model Category The predictive probability of medical images, Indicates the first Round of all client models participating in communication for each category The average predictive probability of medical images, Indicates the category of medical images. This indicates the total number of categories of medical images.

7. The adaptive decision-making federated learning method for medical image recognition according to claim 6, characterized in that, The reward function used in the reinforcement learning agent model is: ； in, Indicates the first Performance improvement rewards for the global model in each round Indicates the first Round consistency alignment reward Indicates the first The weighting coefficient for round consistency alignment , This indicates the total number of communication rounds.

8. The adaptive decision-making federated learning method for medical image recognition according to claim 7, characterized in that, The formula for calculating the performance improvement reward of the global model is as follows: ； in, Indicates the first Accuracy of global model validation in each round Indicates the first The change in single-round verification accuracy across rounds. Indicates from the initial round to the... Time series of changes in single-round validation accuracy. Representing time series The length.

9. The adaptive decision-making federated learning method for medical image recognition according to claim 8, characterized in that, The formula for calculating the consistency alignment reward is as follows: ； in, Indicates the first The average knowledge distillation loss of all client models participating in the communication round. Indicates from the initial round to the... The time series of the average knowledge distillation loss of all client models participating in the communication rounds. Representing time series The length.

Images