A collaborative medical prediction system for heterogeneous data centers

Abstract

The application discloses a cooperative medical prediction system for heterogeneous data centers, and relates to the field of intelligent medical treatment. The system comprises a distributed client set for local training based on local private medical image and dose data; a centralized coordination node for managing and coordinating the federal training process and performing aggregation of cross-client parameters; a decoupled model architecture comprising a globally shared feature encoder and multiple client-specific feature adapters; an alternating training scheduling module configured to periodically switch between global aggregation mode and local adaptation mode; and an adaptive aggregation weighting module for dynamically calculating and assigning the weight of each client in global aggregation based on the statistical difference between the data distribution of each client and the overall distribution. The application can improve the overall generalization performance and prediction stability of the model on heterogeneous multi-center data, and alleviate the performance difference between centers caused by uneven data distribution.

Description

Technical Field

[0001] This invention relates to the field of intelligent healthcare, and in particular to a collaborative medical prediction system for heterogeneous data centers. Background Technology

[0002] Radiation therapy is one of the core treatment methods for thoracic malignancies such as lung cancer; however, radiation pneumonitis is a common and potentially life-threatening complication. Accurate prediction of radiation pneumonitis risk is crucial for optimizing radiotherapy planning, achieving individualized treatment, and improving patient prognosis. In recent years, artificial intelligence-based medical image analysis models have shown potential in this field, but their widespread clinical application faces two fundamental challenges: First, the robustness and generalization ability of these models heavily rely on large-scale, diverse training data, while medical data is difficult to centrally share and aggregate across institutions due to privacy regulations, ethical reviews, and data governance barriers. Second, different medical centers exhibit significant differences in imaging equipment, acquisition protocols, radiotherapy techniques, patient population characteristics, and clinical practice models, resulting in severely non-independent and identically distributed data. This often leads to a significant performance degradation in models trained on single-center data at other centers. Federated learning, as a distributed machine learning paradigm, provides a framework solution for utilizing multi-center data while protecting data privacy. However, how to effectively address the aforementioned data heterogeneity challenges within the framework of federated learning, and construct predictive models that protect privacy, are robust across centers, and have clinical interpretability, has become a key bottleneck for the practical application of this technology.

[0003] Traditional federated learning methods, represented by federated averaging algorithms, rely on a weighted average of local model updates based on the proportion of data volume from each client to form a global model. While effective under the assumption of homogeneous data distribution, this approach reveals significant shortcomings in real-world medical scenarios: First, it attempts to fit all distributions with a single global model, ignoring inherent differences between centers. This leads to severe performance degradation on clients with unique data distributions or smaller data volumes, exacerbating performance unfairness between centers. Second, standard methods lack effective mechanisms to explicitly decouple and separate cross-center general knowledge from center-specific knowledge, making the model prone to overfitting a few large centers that dominate the data distribution. Furthermore, existing methods largely focus on model performance itself, lacking designs for seamlessly integrating complex models into clinical workflows, particularly lacking interpretable output interfaces that help clinicians understand and trust model decisions. Therefore, based on these challenges, this invention proposes a collaborative medical prediction system for heterogeneous data centers. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a collaborative medical prediction system for heterogeneous data centers. Without requiring centralized raw data, it effectively separates and collaboratively optimizes the general knowledge representation across medical centers and the unique distribution characteristics of each center, mitigating model bias and performance imbalance caused by non-independent and identically distributed data. Ultimately, it outputs a radiation pneumonitis risk prediction tool that not only has high prediction accuracy but also whose decision-making process can be intuitively understood and verified by clinicians.

[0005] To achieve the above objectives, this invention provides a collaborative medical prediction system for heterogeneous data centers. The key operating mechanism of this system is an alternating training scheduling strategy, periodically switching between a global aggregation mode and a localized adaptation mode: In global mode, each client updates the shared encoder and uploads parameters, which are then aggregated via an adaptive weighting module that considers differences in data distribution to optimize the general representation; in local mode, the shared encoder is frozen, and each client independently updates its local adapter parameters to finely adapt to its own data distribution, and these parameters are never shared externally. Furthermore, the system integrates a privacy enhancement mechanism and a clinical auxiliary output module. The former applies controlled perturbations during parameter exchange to achieve differential privacy protection, while the latter generates a visual attention map corresponding to the input image space, transforming the model's black-box predictions into a decision-making basis that can be intuitively examined, thus completing a complete technical loop from privacy-preserving collaborative training to clinically usable auxiliary decision-making.

[0006] In a first aspect, the present invention provides a collaborative medical prediction system for heterogeneous data centers, comprising:

[0007] A distributed collection of clients, each client corresponding to a medical center, used for localized training based on local private medical image and dosage data;

[0008] A centralized coordination node is used to manage and coordinate the federated training process and perform cross-client parameter aggregation.

[0009] The decoupled model architecture includes a globally shared feature encoder and multiple client-specific feature adapters; the globally shared feature encoder is used to learn general image representations across data centers; the client-specific feature adapters are used to correct center-specific distribution shifts in the feature space.

[0010] The alternating training scheduling module is configured to periodically switch between global aggregation mode and local adaptation mode:

[0011] An adaptive aggregation weighting module, integrated into the centralized coordination node, is used to dynamically calculate and allocate the weights of each client's data distribution in the global aggregation based on the statistical differences between the data distribution of each client and the overall distribution.

[0012] The clinical auxiliary output module is used to output the radiation pneumonia risk prediction of the target patient and the corresponding visual decision basis based on the decoupled model architecture after training.

[0013] Furthermore, the centralized coordination node also includes a model initialization and distribution unit, which initializes a unique feature adapter for each new client and distributes the latest globally shared feature encoder, enabling new clients to quickly integrate into the federated collaboration network without contributing historical data and to initiate their localization adaptation process, thereby improving the scalability of the system.

[0014] Furthermore, the globally shared feature encoder is designed as a dual-channel three-dimensional convolutional neural network that processes multimodal medical inputs, capable of simultaneously extracting anatomical structural information and physical dose information of the corresponding radiation dose distribution from computed tomography images.

[0015] Furthermore, the client-specific feature adapter includes a trainable normalization layer or a lightweight feedforward network layer, which is configured to receive the output of the globally shared feature encoder and rescale and offset the features using learned transformation parameters to match the local data distribution, thereby achieving efficient adaptation to the local data distribution without compromising the generalization ability of the global model.

[0016] Furthermore, in the global aggregation mode, the client updates the globally shared feature encoder based on local data and securely uploads the updated parameters to the centralized coordination node for aggregation, generating a new generation of global encoder;

[0017] In the localized adaptation mode, the client freezes the globally shared feature encoder and only uses local data to update its corresponding client-specific feature adapter. The adapter parameters always reside locally and are not shared.

[0018] Furthermore, the global aggregation mode and the localization adaptation mode are automatically executed alternately in a preset number of rounds or according to changes in model performance, forming a two-stage iterative optimization loop, which enables global knowledge sharing and local personalized adaptation to work together and converge to a better balance point.

[0019] Furthermore, the model initialization and distribution unit and the alternating training scheduling module are configured collaboratively as follows:

[0020] When a new client joins, based on the characteristic distribution of its limited startup data, an adapter initialization prototype is selected or combined from the adapter parameters of existing clients.

[0021] In the early stages of training for new clients, constraints are imposed on the updates of their global model parameters, while allowing their local adapter parameters to be optimized rapidly.

[0022] As the training rounds increase, the global model update constraints on the new clients are gradually relaxed until they participate in federated collaboration on an equal footing.

[0023] Furthermore, the adaptive aggregation weighting module dynamically adjusts the weights by calculating the distribution divergence metric of each client's training data and the entire data in the feature space. Clients with more significant distribution differences are given higher weights during the aggregation process, which encourages the global model to focus on the edges of the data distribution or the centers that are not fully represented, thereby enhancing the model's overall adaptability to heterogeneous data.

[0024] Furthermore, the adaptive aggregation weighting module is also used for:

[0025] Obtain the characteristic distribution statistics of local data from each client;

[0026] Based on the comparison between the aforementioned feature distribution statistics and the global feature distribution statistics, a feature distribution similarity measure between each client and the global model is calculated;

[0027] Based on the feature distribution similarity metric and the proportion of data volume of each client, the contribution weight of each client in the global model aggregation is dynamically determined, wherein clients whose feature distribution has a moderate difference from the global model are given a higher aggregation weight.

[0028] Furthermore, the clinical auxiliary output module is also used for:

[0029] During the model training phase, a dose distribution reconstruction task is introduced as an auxiliary learning objective, enabling the feature encoder to learn dose-related feature representations.

[0030] During the model inference phase, a data-driven original attention map and a dose saliency map based on the internal dose representation of the model are generated simultaneously.

[0031] The original attention map and the dose significance map are fused to generate a fused visualization atlas that can simultaneously reflect the importance of anatomical features and dose relevance, serving as the basis for interpreting model decisions.

[0032] Furthermore, the clinical auxiliary output module integrates a gradient visualization unit to generate an attention heatmap corresponding to the input medical image space. The heatmap is generated by upsampling the spatial gradient of the input features predicted by the computational model to highlight the anatomical region that contributes the most to the prediction of radiation pneumonitis risk, thereby providing an intuitive visual explanation for clinical decision-making and improving the credibility and practicality of the model.

[0033] Furthermore, the clinical auxiliary output module spatially fuses and displays the attention heatmap with the input dose distribution map to help clinicians intuitively assess whether the model's decision logic is related to the anatomical features of the high-dose irradiation area, thereby further enhancing the credibility and depth of decision support.

[0034] Furthermore, the system also includes a privacy enhancement unit that, in the global aggregation mode, applies random perturbations to the model parameter updates uploaded by the client, satisfying privacy budget constraints.

[0035] Secondly, a collaborative medical prediction method for heterogeneous data centers is also provided, the method being based on the system described in the first aspect above, comprising:

[0036] In the distributed client set, each client uses local medical image and dosage data, and under the control of the alternating training scheduling module, alternately executes global aggregation mode and localization adaptation mode to collaboratively train the decoupled model architecture.

[0037] In the global aggregation mode, the adaptive aggregation weighting module dynamically adjusts the aggregation weights according to the differences in client data distribution, and the centralized coordination node performs secure aggregation and redistribution of the globally shared feature encoder parameters.

[0038] In the localization adaptation mode, each client independently updates and maintains its unique feature adapter parameters;

[0039] After training is completed, the new medical images and dose data of the target patient are input into the localized model of the corresponding client through the clinical auxiliary output module to obtain a probability prediction of the risk of radiation pneumonitis, and at the same time generate a visual attention map to interpret the prediction.

[0040] This invention provides a collaborative medical prediction system for heterogeneous data centers. Its core lies in constructing a decoupled model architecture that separates the globally shared feature encoder from the unique feature adapters of each client, and employs an alternating training mechanism to periodically switch between a global aggregation mode and a localized adaptation mode. In the global aggregation mode, the updated shared encoder parameters from each client are aggregated through adaptive weighting to learn cross-center common features; in the localized adaptation mode, the shared encoder is frozen, and only the unique adapter parameters are updated locally on each client, achieving personalized distribution correction. The system further integrates a privacy enhancement mechanism, implementing perturbations during parameter exchange to protect data privacy, and is equipped with a clinical auxiliary output module capable of generating a visual attention map reflecting the model's decision-making basis, thus forming a complete technical closed loop from privacy-preserving distributed collaborative training to interpretable clinical decision support.

[0041] This solution effectively addresses the challenges of non-independent, identically distributed data prevalent in medical scenarios while strictly protecting the data privacy of all participants. Through a decoupled architecture and alternating training mechanism, the system balances collaborative learning of cross-center knowledge with specific adaptation to local data distribution, significantly improving the model's overall generalization performance and predictive stability on heterogeneous, multi-center data. It also mitigates inter-center performance differences caused by uneven data distribution, enhancing the model's fairness and robustness. Furthermore, its visualized decision interpretation output transforms model predictions into auxiliary information that clinicians can intuitively understand and verify, significantly enhancing the transparency, credibility, and clinical applicability of AI models and promoting their safe and reliable integration into actual diagnostic and treatment workflows.

[0042] Beneficial effects

[0043] By implementing the collaborative medical prediction system for heterogeneous data centers provided by the present invention, the following technical effects are achieved:

[0044] (1) The decoupled model architecture achieves forced decoupling and collaborative optimization of cross-center general knowledge representation and center-specific distribution features at the model structure level. This separation allows general features to be efficiently shared and refined globally, while ensuring that sensitive or unique distribution biases are accurately corrected locally, thus fundamentally avoiding the performance trade-offs and compromises caused by direct mixed training of heterogeneous data or a single global model.

[0045] (2) The alternating training scheduling mechanism organizes the federated training process into two alternating phases: global aggregation and local adaptation, forming a dynamically balanced training paradigm. This alternating rhythm not only ensures the co-evolution of general knowledge and local knowledge in terms of optimization process, but also effectively alleviates the common optimization objective conflict problem in federated learning under non-independent and identically distributed data by periodically freezing and updating different parameter groups, thereby improving the overall training stability and the convergence quality of the final model.

[0046] (3) The dynamic aggregation weight generation method based on feature distribution similarity effectively solves the client drift problem caused by simply relying on data volume weighting in traditional federated learning by introducing a similarity measure at the feature distribution level. It can guide the global model to pay more attention to clients that have moderate differences from the mainstream distribution during the aggregation process, thereby significantly enhancing the overall generalization ability of the model under heterogeneous data distribution and the performance fairness among clients, especially showing more stable prediction performance on clients with small data volume or unique distribution.

[0047] (4) The multi-level model interpretability enhancement method that integrates dose information fundamentally improves the clinical credibility of model decisions by explicitly incorporating dose physics information into the interpretability framework. It combines the activation of black-box features within the model with intuitively understandable dose distributions, and the resulting visualization allows clinicians to directly assess whether the model's decision logic is consistent with key dose-anatomical relationships, thus achieving a leap from simple feature visualization to in-depth interpretation with clinical pathophysiological basis.

[0048] (5) The incremental knowledge transfer and fusion method for new participating clients ensures that the federated learning system has dynamic scalability and robustness. It allows newly joined clients to quickly start personalized adaptation using existing knowledge when data is limited, while minimizing the performance disturbance to the already converged collaborative system by constraining their initial updates to the global model. Attached Figure Description

[0049] To make the above-described collaborative medical prediction system for heterogeneous data centers of the present invention more apparent and understandable, the accompanying drawings used in the specific embodiments of the present invention 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 from these drawings without creative effort.

[0050] Figure 1 This diagram illustrates the partitioning of a multi-center heterogeneous dataset.

[0051] Figure 2 This diagram illustrates the parameter upload mechanism for the federated learning client.

[0052] Figure 3 This diagram illustrates feature extraction and parameter distribution.

[0053] Figure 4 This chart shows a performance comparison between heterogeneous federated learning and centralized models. Detailed Implementation

[0054] Example 1:

[0055] This embodiment provides a specific implementation of a personalized federated learning system for predicting the risk of radiation pneumonitis across multiple centers. This system aims to address core challenges in multi-center collaborations in the medical field, including data privacy protection, data heterogeneity, and model clinical interpretability.

[0056] During implementation, the system constructed a collaborative network simulating a complex real-world environment. This network consisted of four independent medical centers and a central coordination server. These four centers differed significantly in patient populations, image acquisition equipment, radiotherapy techniques, and whether immunotherapy was used in conjunction with other treatments. Their data distribution exhibited typical non-independent, identically distributed characteristics, such as... Figure 1 As shown. For example, two of the centers primarily received Asian patients and used intensity-modulated radiotherapy (IMRT) or volumetric rotational IMRT, with some patients receiving immunotherapy; the data from the other two centers came from an international multicenter clinical trial, where patients were mainly from Europe and America, using three-dimensional conformal radiotherapy without immunotherapy, and the prescribed doses were divided into two different levels. This deliberately constructed heterogeneous environment provided a foundation for validating the system's robustness in real-world scenarios.

[0057] The core of the system lies in its decoupled model architecture and corresponding collaborative training mechanism. The prediction model deployed locally on each client is not a complete monolithic network, but rather consists of two parts: a globally shared feature encoder and a client-specific feature adapter. The feature encoder uses a 3D residual neural network as its backbone, designed with dual-channel input, and can simultaneously process the patient's planned CT images and corresponding absolute physical dose distribution maps, aiming to extract deep radiomics features related to the pathophysiology of radiation pneumonitis from multimodal data. The parameters of this encoder are jointly maintained and optimized by all clients. The cascaded adapter is a lightweight, trainable module, such as a normalization layer containing learnable parameters or a small feedforward network, whose responsibility is to transform and calibrate the general features extracted by the encoder to adapt to the center-specific data distribution shift. This architecture fundamentally separates cross-center transferable general knowledge from the center-specific distribution characteristics.

[0058] The system's training process is controlled by a sophisticated alternating scheduling protocol, which periodically distinguishes between two operating modes: global aggregation mode and local adaptation mode. In global aggregation mode, each client uses local data to perform one round of gradient updates on the shared feature encoder parameters, and then only uploads the encoder parameter updates to the central server. Figure 2As shown, the server does not perform a simple averaging, but rather intelligent aggregation through an adaptive aggregation weighting module. This module requires each client to calculate the feature distribution statistics of its local data under the current encoder before uploading parameters. The server aggregates these statistics to estimate the global feature distribution, and then calculates the similarity metric between each client's distribution and the global distribution. Ultimately, the weight of each client in this aggregation is dynamically determined by its data volume proportion and the aforementioned feature similarity. This strategy allows clients with unique data distributions that can bring diverse information to the global model, but whose distributions are not outliers, to gain greater influence in the aggregation, thereby effectively guiding the global model to learn more comprehensive and robust general representations, and alleviating the client drift problem caused by uneven data volume in traditional methods.

[0059] After aggregating and generating the next-generation global encoder parameters, the system switches to local adaptation mode. In this mode, the server distributes the updated global encoder to all clients and freezes their parameters. The distribution process is as follows: Figure 3 As shown, each client then focuses solely on training and updating its unique feature adapter parameters using only local data. These adapter parameters, containing crucial information for calibrating the local distribution, are considered the private assets of each center, remaining entirely local and never participating in any form of sharing or uploading. This alternating update cycle of the globally shared encoder and the local private adapter forms the backbone of the system training, enabling the co-evolution of general knowledge and personalized adaptation to proceed in parallel and mutually reinforce each other.

[0060] To transform the model's black-box predictions into decision support information that clinicians can understand and verify, the system integrates a multi-level interpretability framework that incorporates dose constraints. During training, a parallel dose distribution reconstruction branch network is appended to the shared feature encoder, with the reconstructed dose map of the input serving as an auxiliary learning objective. Through multi-task learning, the feature encoder is forced to learn dose-sensitive feature representations. During inference, for a new patient, the system first generates a data-driven initial attention heatmap using a standard gradient-weighted class activation mapping method, highlighting the anatomical regions that contribute most to the prediction. Simultaneously, using the trained dose reconstruction branch, a dose saliency map reflecting the model's internal cognition is generated. Finally, the two are spatially fused to generate a fused visualization atlas. This atlas not only uses color intensity to indicate anatomically relevant areas but also visually reflects the overlap between these areas and high-dose regions through fusion logic. Clinicians can use this atlas to intuitively determine whether the model's decisions are based on abnormal tissue texture or the expected response to high-dose irradiation, greatly enhancing the model's credibility and clinical applicability.

[0061] When a new medical center wants to join the collaborative network, it only needs to provide a small local start-up dataset. The system employs a progressive knowledge transfer and fusion mechanism: the central server first selects several adapters with the most similar feature distributions from the existing adapter library based on the feature distribution of the new center's start-up data. These adapters are then weighted and synthesized into an initial adapter prototype, thus achieving knowledge transfer. In subsequent training, the system initially imposes a strong constraint on the global encoder updates uploaded by the new center to prevent its not-yet-fully-adapted data from impacting the already stable global model, while allowing its local adapters to optimize rapidly. As training epochs increase, this constraint is gradually and linearly relaxed until the new center participates in the federation on a completely equal footing. This mechanism ensures that new centers can start up quickly and achieve good performance, while maximizing the stability of the existing federated system.

[0062] The system trained through the above process deploys a highly personalized prediction model locally on each client. This model incorporates general knowledge from multiple centers while being finely adapted to local data characteristics. To quantitatively evaluate the system's performance advantages, comprehensive comparative experiments were conducted on four heterogeneous datasets. As shown in Tables 1-4, the model exhibits robust and leading prediction performance across all test sets. Specifically, on Datasets 1 and 2, representing real-world clinical environments, the test set AUC values ​​reached 0.77 and 0.71, respectively; on Datasets 3 and 4, originating from clinical trials, the test set AUC values ​​were 0.76 and 0.75, respectively. The real-world clinical data originates from routine medical practice, with diverse patients, equipment, and treatment plans, exhibiting strong data heterogeneity but relatively lenient quality control; while the clinical trial data comes from rigorous standardized studies, with controlled patient populations, treatment plans, and image acquisition, resulting in high data quality but limited representativeness. Compared to the single-center model, this system achieves significant performance improvements across all external test sets. Compared to the standard federated averaging algorithm, this system demonstrates superior generalization ability and stability, particularly on clients with smaller datasets or unique distributions, showing AUC improvements of 0.10 and 0.10 respectively on datasets 3 and 4. These comparative results empirically demonstrate the effectiveness of the decoupled architecture and alternating training mechanism in integrating heterogeneous multi-center data, improving the model's cross-center generalization performance, and ensuring performance fairness among clients.

[0063] Table 1. Performance of the FCAAM Model

[0064]

[0065] Table 2. Performance of Model-sep (SEP)

[0066]

[0067] Table 3. Performance of Model COM

[0068]

[0069] Table 4. Performance of Model FL (FedAvg)

[0070]

[0071] In the table, AUC stands for Area Under the Curve; CI stands for Confidence Interval; FCAAM stands for Federated Cross-Center Adaptive Alternating Method, which is based on a decoupled model architecture that separates the globally shared feature encoder from the client-specific feature adapter and implements alternating training scheduling; Model-sep (SEP) stands for Isolated Single-Center Training Baseline, which represents a separate training baseline where each medical center trains a complete monolithic prediction model entirely independently using its local data, without any form of cross-center knowledge sharing or parameter exchange; COM stands for Idealized Centralized Training Baseline, which represents the ideal case of centralized training, assuming that all centers' private data can be centralized on a central server and a unified monolithic prediction model can be trained on this entire dataset; FL (FedAvg) stands for Traditional Federated Learning Baseline, which uses a standard federated averaging algorithm as a comparison baseline. In this model, all clients jointly maintain a complete global model and aggregate parameters through a simple weighted average.

[0072] In practical applications, doctors can upload a patient's CT scan and dose data through an integrated web platform interface. Within seconds, the system can return the patient's risk probability of developing grade 2 or higher radiation pneumonitis and simultaneously provide the aforementioned visualized attention map incorporating dose information to assist in treatment plan evaluation and optimization. For example... Figure 4 As shown, compared with single-center models trained independently in each center, the model obtained by the system shows significantly better and more stable predictive performance in all external test centers; compared with the standard federated averaging algorithm, the performance improvement of this system is particularly obvious on clients with small data volume and unique distribution, and the output decision interpretation has a higher degree of consistency with clinical prior knowledge.

[0073] Example 2:

[0074] Building upon the aforementioned embodiments, this paper addresses the shortcomings of traditional federated averaging or data-volume-based weighted aggregation, which fail to adequately consider the distributional differences in features among clients. A dynamic aggregation strategy based on feature distribution similarity is proposed. During global aggregation, this strategy considers not only the data volume of each client but also the consistency or complementarity between the local data feature distribution and the current global model feature distribution. Clients with high consistency indicate that their data distribution is close to the global consensus, and their updates can strengthen the mainstream pattern. Clients with high complementarity introduce diversity into their updates, preventing the model from overfitting to the mainstream center. A dynamic weighting function calculates the contribution weight of each client during each aggregation, enabling the aggregation process to adaptively balance the goals of consolidating consensus and exploring diversity. This more intelligently guides the global model to learn robust and comprehensive representations in heterogeneous data.

[0075] In the Before the round of global aggregation begins, each client Locally, the feature encoder of the current global model is used to perform a forward pass on its local dataset, extracting a batch of feature vectors from the encoder output layer. The first-order statistic (mean vector) of this batch of feature vectors is then calculated. ) and second-order statistics (variance vector) This serves as a lightweight description of the local feature distribution of the client.

[0076] The central server collects all client uploads. And estimate the statistic of the current global feature distribution by weighted average. and ,in, This is a superscript identifier. Initial weighting can be based on the proportion of data volume.

[0077] For each client, the weighted similarity between its local feature distribution and the global feature distribution is calculated as follows:

[0078] In the formula, Weighted similarity; A hyperparameter that balances the importance of mean variance and variance variance; It is a dissociative function; This is the average distance temperature parameter; This refers to the variance distance temperature parameter.

[0079] The client is in the Final aggregate weight of the round Based on its data volume ratio and feature similarity Joint decision:

[0080]

[0081] In the formula, A hyperparameter used to balance data volume and feature similarity.

[0082] The central server uses the calculated Update the global feature encoder uploaded by each client. A weighted average is performed to obtain a new global model. In the formula, This represents the global learning rate.

[0083] Aggregate weight Not directly based on similarity Decision. Calculation. The aim is to evaluate the value of each client from the perspective of data distribution; adjustable parameters are introduced. This is to allow for a flexible trade-off between the two criteria of data volume and data distribution value. Crucially, by normalizing the overall score using the Softmax function, its exponential properties significantly amplify the relative advantages between scores. Therefore, a high-weighted result is ultimately obtained. Clients that are both relevant to the global model and provide unique value, thus making their overall scores stand out in comparisons, are typically those that are relevant to the global model and provide unique value.

[0084] In a comparative experiment with the traditional federated average algorithm based on data volume weighting, under the simulated settings of four heterogeneous medical centers (client data volume ratio of 5:3:1:1, and significant covariate shift in feature distribution), the model trained using this dynamic aggregation strategy improved the AUC value for radiation pneumonitis prediction from 0.68±0.04 (federated average algorithm) to 0.76±0.03 on the test set of the client with the smallest data volume, while the variance of AUC among the four clients decreased by approximately 60%. This result demonstrates that dynamic weighting based on feature distribution similarity effectively mitigates client drift caused by non-independent and identically distributed data, significantly improving model generalization performance and cross-center fairness for small sample and marginally distributed clients.

[0085] Example 3:

[0086] Building upon the aforementioned embodiments, existing gradient-based interpretability methods primarily target image features. In radiotherapy prediction scenarios, they fail to explicitly incorporate radiation dose—a crucial physical decision factor—leading to insufficient clinical relevance in the generated attention maps. This paper proposes a multi-level interpretability framework that integrates dose constraints. The principle is to introduce an auxiliary, dose-map-supervised decoder branch during model training, in addition to the main classification task. This branch attempts to reconstruct the dose distribution from features, forcing the feature encoder to learn dose-related semantic information. During inference, a fusion map is generated by comparing the standard Grad-CAM activation map with the dose-guided attention modulation map. This map not only highlights anatomically sensitive areas but also visually indicates whether these areas overlap with high-dose regions using different colors or transparency, thereby linking and comparing the model's data-driven findings with clinical physical dose knowledge, providing a deeper level of decision interpretation.

[0087] During the training phase, two branches are connected in parallel after the globally shared feature encoder: the main classification branch, which is the original network structure and outputs the probability of radiation pneumonia risk; and the dose reconstruction branch, which consists of a lightweight three-dimensional deconvolutional network, whose input is the intermediate feature map of the feature encoder and whose output is a dose distribution reconstruction map of the same size as the input CT.

[0088] The total loss function is defined as:

[0089]

[0090] In the formula, This represents the total loss during model training. The cross-entropy loss is used for the primary classification task. To balance the hyperparameters of the two losses; To account for the loss in the dose reconstruction branch, calculate the smoothing L1 loss or mean square error between the true dose map and the reconstructed map.

[0091] Two-layer attention generation during reasoning:

[0092] First layer: Data-driven attention. Using the standard Grad-CAM method, the gradient of the main classification output with respect to the last layer feature map of the feature encoder is calculated to generate an initial attention heatmap.

[0093] The second layer: dose-guided attention. Using the trained dose reconstruction branch, the value of each voxel in the reconstructed dose map is calculated and normalized to the [0,1] interval to obtain the dose saliency map. This map directly reflects the dose distribution information encoded by the features within the model.

[0094] The final output fused attention map is generated by the following formula:

[0095]

[0096] In the formula, The final visual attention map after fusion; The attention fusion enhancement factor is a hyperparameter greater than 0. This is a dose significance map generated from the dose reconstruction branch within the model; Element-wise multiplication of matrices; This is a data-driven attention heatmap generated using the standard Grad-CAM method.

[0097] This operation makes it possible to... High value position, The response is enhanced. In the clinical support output module, A semi-transparent color is overlaid on CT images, while the real or reconstructed dose map is overlaid with contour lines or another semi-transparent color, forming a two-layer visualization. Physicians can intuitively judge the spatial relationship between the area of ​​interest in the model and the high-dose area, thereby assessing whether the model's judgment is based on abnormal anatomical texture or tissue response under high-dose irradiation, greatly enhancing the depth of interpretability and clinical value.

[0098] Compared to interpretability methods that rely solely on standard Grad-CAM to generate attention maps, this fusion dose-constraint framework underwent a blinded evaluation on an independent test set comprising 50 cases of radiation pneumonitis with a clear clinical diagnosis. Three senior radiation oncologists rated the model's decision-making process (1-5 points, with higher scores indicating stronger clinical relevance). Results showed that the fusion framework generated a visual representation with an average physician score of 4.2, significantly higher than the standard Grad-CAM score of 2.8. Quantitative analysis revealed an average 35% increase in the spatial overlap between the model's high-attention regions and the high-dose volumes outside the clinically relevant target area. This demonstrates that the framework deeply integrates data-driven features with key physical dose information, producing decision interpretations with significantly higher clinical relevance and credibility.

[0099] Example 4:

[0100] In dynamic federated learning systems, the addition of new medical centers is commonplace. Existing solutions typically require new centers to train their local adapters from scratch or directly use the global model, which is either slow to start, performs poorly, or may disrupt the already converged global model. This embodiment proposes a progressive knowledge transfer and personalized startup mechanism. When a new center joins, the system first initializes an adapter prototype for it on the server side using the existing feature adapter parameter sets of each center through a rapid prototype matching based on a small amount of startup data from the new center, rather than random initialization. Subsequently, in the initial rounds of federated training, a progressive unfreezing and regularization training strategy is introduced for the new center: first, the update magnitude of its shared encoder is strictly constrained to prevent its unfamiliar data from causing a drastic impact on the stable global model; at the same time, its adapter is allowed to be quickly personalized. As the rounds increase, the constraints are gradually relaxed, allowing it to smoothly and progressively integrate into the federated collaborative network.

[0101] When a new center joins, it needs to provide a small-scale local starter dataset. The central server distributes the current global feature encoder to the new center, which then uses the current global feature encoder to extract feature statistics from the local starter dataset.

[0102] The server maintains an archive of adapter parameters and corresponding feature statistics for each existing center. The similarity between the feature statistics and each corresponding feature statistics in the archive is calculated. The top few centers with the highest similarity are selected, and their adapter parameters are weighted and averaged. The weights are the normalized similarity scores, resulting in the initial adapter.

[0103] In the initial rounds of global aggregation involving the new center, a scaling factor greater than 0 and less than or equal to 1 is applied to its uploaded global encoder updates. This significantly mitigates the potential negative impact of outlier data from the new center. Simultaneously, in localization adaptation mode, the new center optimizes its adapter with a high learning rate, quickly adapting to local data.

[0104] As the performance of the local model at the new center stabilizes, the scaling factor is gradually increased to allow it to have a more significant impact on the global model.

[0105] When the scaling factor is set to 1, the new centers participate in federated training on an equal footing with the existing centers. Throughout the process, an L2 regularization term, similar to that used in initializing the adapter, can be added to the local loss of the new centers to encourage them not to deviate too far from the knowledge prototype in the initial stage, and then the strength of this regularization term is gradually weakened.

[0106] Once a new center completes its launch phase, its characteristic statistics and adapter parameters are archived on the server to serve any other new centers that may join in the future. This mechanism forms a sustainably expanding federal ecosystem.

[0107] To verify the effectiveness of the incremental knowledge transfer mechanism, a scenario was simulated where a new center with only 80 local data points was added, based on a converged three-center federated model. Compared with the conventional method of randomly initializing adapters and directly participating in federated training, the training epochs required for the new center's local model to achieve comparable performance to mature centers was reduced from over 30 epochs to less than 15 epochs, an acceleration of approximately 50%. More importantly, during the 10-epoch process of integrating the new center, the performance fluctuation of the original three mature centers was controlled within 0.02, while the conventional method caused fluctuations as high as 0.05. This demonstrates that the mechanism can ensure the rapid startup of new centers while maximizing the stability and performance of the existing federated system, achieving smooth and low-disturbance expansion of the system.

Claims

1. A collaborative medical prediction system for heterogeneous data centers, characterized in that, include: A distributed collection of clients, each client corresponding to a medical center, used for localized training based on local private medical image and dosage data; A centralized coordination node is used to manage and coordinate the federated training process and perform cross-client parameter aggregation. The decoupled model architecture includes a globally shared feature encoder and multiple client-specific feature adapters; the globally shared feature encoder is used to learn general image representations across data centers; the client-specific feature adapters are used to correct center-specific distribution shifts in the feature space. The alternating training scheduling module is configured to periodically switch between a global aggregation mode and a local adaptation mode: in the global aggregation mode, the client updates the globally shared feature encoder based on local data and securely uploads the updated parameters to the centralized coordination node for aggregation to generate a new generation of global encoder; in the local adaptation mode, the client freezes the globally shared feature encoder and only uses local data to update its corresponding client-specific feature adapter, and the adapter parameters always reside locally and are not shared. An adaptive aggregation weighting module, integrated into the centralized coordination node, is used to dynamically calculate and allocate the weights of each client's data distribution in the global aggregation based on the statistical differences between the data distribution of each client and the overall distribution. The adaptive aggregation weighting module dynamically adjusts the weights by calculating the distribution divergence metric of each client's training data and the overall data in the feature space, wherein clients with more significant distribution differences are given higher weights during the aggregation process. The clinical auxiliary output module is used to output the radiation pneumonia risk prediction of the target patient and the corresponding visual decision basis based on the decoupled model architecture after training.

2. The system according to claim 1, characterized in that: The centralized coordination node also includes a model initialization and distribution unit, which is used to initialize the unique feature adapter for new clients and distribute the latest globally shared feature encoder.

3. The system according to claim 1, characterized in that: The globally shared feature encoder is designed as a dual-channel three-dimensional convolutional neural network that processes multimodal medical inputs, capable of simultaneously extracting anatomical structural information and physical dose information of the corresponding radiation dose distribution from computed tomography images.

4. The system according to claim 1, characterized in that: The client-specific feature adapter includes a trainable normalization layer or a lightweight feedforward network layer, which is configured to receive the output of the globally shared feature encoder and rescale and offset the features using learned transformation parameters.

5. The system according to claim 1, characterized in that: The global aggregation mode and the localization adaptation mode are executed alternately in a preset number of rounds or automatically according to changes in model performance, forming a two-stage iterative optimization loop.

6. The system according to claim 1, characterized in that, The adaptive aggregation weighting module is also used for: Obtain the characteristic distribution statistics of local data from each client; Based on the comparison between the aforementioned feature distribution statistics and the global feature distribution statistics, a feature distribution similarity measure between each client and the global model is calculated; Based on the feature distribution similarity metric and the proportion of data volume of each client, the contribution weight of each client in the global model aggregation is dynamically determined, wherein clients whose feature distribution has a moderate difference from the global model are given a higher aggregation weight.

7. The system according to claim 1, characterized in that, The clinical auxiliary output module is also used for: During the model training phase, a dose distribution reconstruction task is introduced as an auxiliary learning objective, enabling the feature encoder to learn dose-related feature representations. During the model inference phase, a data-driven original attention map and a dose saliency map based on the internal dose representation of the model are generated simultaneously. The original attention map and the dose significance map are fused to generate a fused visualization atlas that can simultaneously reflect the importance of anatomical features and dose relevance, serving as the basis for interpreting model decisions.

8. A collaborative medical prediction method for heterogeneous data centers, characterized in that: The method is implemented based on the system described in any one of claims 1-7: The method includes: In the distributed client set, each client uses local medical image and dosage data, and under the control of the alternating training scheduling module, alternately executes global aggregation mode and localization adaptation mode to collaboratively train the decoupled model architecture. In the global aggregation mode, the adaptive aggregation weighting module dynamically adjusts the aggregation weights according to the differences in client data distribution, and the centralized coordination node performs secure aggregation and redistribution of the globally shared feature encoder parameters. In the localization adaptation mode, each client independently updates and maintains its unique feature adapter parameters; After training is completed, the new medical images and dose data of the target patient are input into the localized model of the corresponding client through the clinical auxiliary output module to obtain a probability prediction of the risk of radiation pneumonitis, and at the same time generate a visual attention map to interpret the prediction.

Images