An artificial intelligence model training method and system based on federated learning

Abstract

The application relates to the technical field of artificial intelligence model training, and discloses an artificial intelligence model training method and system based on federated learning. The method receives model parameter updates and local data characteristic information from a client, judges whether the model parameter updates contain image bottom texture features introduced by the client according to the local data characteristic information, and then adjusts the processing mode of the model parameter updates in global model aggregation according to the judgment result to integrate the updates containing the image bottom texture features. The method effectively solves the problem that, in existing federated learning, a central aggregation server cannot correctly understand and effectively absorb parameter updates containing special data characteristic information from a minority client, resulting in a decline in the generalization ability of a global model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence model training, and more specifically, to an artificial intelligence model training method and system based on federated learning. Background Technology

[0002] In today's medical field, artificial intelligence diagnostic systems are widely used to improve the accuracy and efficiency of disease diagnosis. However, due to restrictions on patient privacy and inter-institutional data sharing, traditional centralized data training models are difficult to implement. To address this, federated learning, as a distributed training paradigm, has emerged. It allows medical institutions to train models locally, uploading only the model parameters to a central aggregation server for integration, and then distributing the updated global model back to all participants, thus achieving model training while protecting privacy.

[0003] In practical applications of federated learning, a central aggregation server typically selects clients to participate in model training based on preset rules. For example, to accelerate model convergence, the system operations team might adjust the client selection rules, prioritizing hospitals with high data activity. However, this strategy can lead to training bias towards a few hospitals with concentrated data features, as high data activity does not always equate to rich data diversity. For instance, some large specialized hospitals may have a large volume of data updates but limited patient characteristics, while general hospitals or primary healthcare institutions with more diverse patient data may participate less in training due to insufficient data activity. This can result in a "Matthew effect" in the training data, where the data features of a few active hospitals dominate the model training process.

[0004] This bias causes the global model to overfit to the characteristics of a few active hospital patient populations, exhibiting higher sensitivity to imaging features of patients in specific age groups, regions, or disease stages. When this global model is distributed and deployed to hospitals with more diverse patient data, its diagnostic accuracy drops significantly, even leading to misdiagnosis or missed diagnosis, preventing the global model from effectively generalizing to all patient populations across the entire healthcare ecosystem.

[0005] Meanwhile, in hospitals that participate less in training, there may be older medical imaging equipment that introduces specific, subtle background textures into image acquisition. This texture is not a pathological feature, but rather the equipment's own "imprint," exhibiting a systematic, slight difference from data generated by mainstream equipment. When hospitals with such equipment occasionally participate in training, their local models learn and adapt to this data containing specific background textures, indirectly incorporating adjustments for these subtle data differences into the uploaded model parameters.

[0006] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0007] This invention provides a method and system for training artificial intelligence models based on federated learning. It aims to solve the problem that in federated learning scenarios, when the central aggregation server integrates model parameters, it cannot correctly understand and effectively absorb parameter updates from minority clients that contain special data feature information, which leads to a decrease in the generalization ability of the global model and even misdiagnosis or missed diagnosis.

[0008] The technical solution of this application is as follows:

[0009] Firstly, this application discloses a method for training an artificial intelligence model based on federated learning, used in a federated learning scenario to train a cross-institutional medical diagnostic model. The method includes:

[0010] Receive model parameter updates from the client, as well as local data characteristic information describing the client, which reflects the underlying image texture introduced by the client during data acquisition;

[0011] Based on the local data characteristics, determine whether the model parameter update includes low-level image texture features introduced by the client;

[0012] Based on the judgment results, the processing method of the model parameter update in the global model aggregation is adjusted to integrate the update containing the underlying texture features of the image.

[0013] Through this technical solution, this application can effectively identify and process the underlying texture features of images introduced by the client in federated learning, avoid the decline in model generalization ability caused by data heterogeneity, and thus improve the accuracy and robustness of cross-institutional medical diagnostic models.

[0014] Furthermore, based on the judgment result, the processing method of the model parameter update in the global model aggregation is adjusted to integrate the update containing the low-level texture features of the image, including:

[0015] When the client's local model is being trained, the feature extraction layer inside the local model is guided to become sensitive to small changes in the client's state.

[0016] Extract a dynamic fingerprint representing the current data style of the client from the local model;

[0017] The real-time environment and operation parameter vectors of the client for monitoring and sampling medical images are incorporated into the feature summary of the client. After the central aggregation server receives the model parameter updates, dynamic fingerprints, real-time environment and operation parameter vectors from each client, it uses a meta-aggregation network to calculate and adjust the processing method of each client model update in the global aggregation. Based on the adjustment factor output by the meta-aggregation network, the contribution of each client model update is adjusted and integrated into the global model.

[0018] The meta-aggregation network is trained through meta-learning, with the goal of learning a function that predicts aggregation weights and learning rate adjustment factors based on the dynamic client context.

[0019] Through this technical solution, this application can achieve fine-grained perception and adjustment of client data style and environmental parameters by using dynamic fingerprints and meta-aggregation networks, thereby more intelligently integrating the contributions of different clients in global model aggregation and further improving the model's adaptability and generalization ability.

[0020] More specifically, in some implementations, a dynamic fingerprint representing the current data style of the client is extracted from the local model, including:

[0021] The background texture of the input image is reconstructed by using feature maps extracted from early convolutional layers by the local model and then processed by the decoder network.

[0022] The client's local training loss function is combined with the loss from the disease diagnosis task and the loss from the texture reconstruction task;

[0023] The client extracts feature vectors from a specific layer in the local model that is responsible for extracting low-level features.

[0024] The client performs a global average pooling operation on the feature vector to obtain a fixed-dimensional feature vector, which serves as a dynamic fingerprint representing the current data style of the client.

[0025] Through this technical solution, this application can more accurately capture and characterize the client-specific data style by combining the loss functions of texture reconstruction tasks and disease diagnosis tasks, and extracting and pooling feature vectors from the low-level feature layer of the local model, thus providing a reliable basis for subsequent aggregation and adjustment.

[0026] Building upon the above, this application further proposes that, during training, the feature extraction layer within the client's local model should be guided to become sensitive to minute changes in the client's state, including:

[0027] During the training process, the local model periodically applies transformations to the input image that simulate the background texture drift of the client.

[0028] The parameters of the transformation are identified or predicted using this local model;

[0029] Based on the identified or predicted transformation parameters, the feature extraction layer inside the local model is guided to become sensitive to subtle changes in the client's state.

[0030] Through this technical solution, this application can enhance the model's sensitivity to small changes in client state by periodically applying transformations that simulate texture drift and guiding the local model to identify or predict these transformation parameters, thereby enabling it to better adapt to and handle data heterogeneity.

[0031] Preferably, during training, the local model periodically applies a transformation simulating client-side background texture drift to the input image, including:

[0032] Based on the influence of wear and tear of internal components, ambient temperature and humidity fluctuations, and imaging parameter settings on image background texture under different operating states of medical imaging clients, a parameterized texture drift generator is constructed. This texture drift generator is used to generate various combinations of transformation parameters for noise distribution patterns, weak stripe directions, contrast gradients, and geometric distortion.

[0033] During each training cycle, the local model randomly selects a set of transformation parameters from the texture drift generator and performs real-time transformation on the input image of the current training batch based on these transformation parameters to generate an image that simulates the background texture drift of the client.

[0034] During the training process of the client's local model, the potential evolution path and key influencing factors of the client's background texture drift are identified by analyzing the client's long-term running data. The parameter extraction strategy in the texture drift generator is dynamically adjusted to prioritize generating simulated transformations that better match the current drift state of the client.

[0035] Through this technical solution, this application can construct a parameterized texture drift generator and, in conjunction with long-term client data analysis, dynamically adjust the transformation parameter extraction strategy to more realistically and accurately simulate client background texture drift, thereby further improving the model's adaptability to changes in actual data.

[0036] In one implementation, during the training process of the client's local model, based on analysis of the client's long-term running data, the potential evolution path and key influencing factors of the client's background texture drift are identified, and the parameter extraction strategy in the texture drift generator is dynamically adjusted to prioritize generating simulated transformations that better match the current actual drift state of the client, including:

[0037] This local model receives data from internal sensors on the client, environmental monitoring data, and operation log information.

[0038] The local model is used to extract features and recognize patterns from the received data to distinguish texture drift patterns caused by different factors such as wear and tear on the client, fluctuations in ambient temperature and humidity, and deviations in operating parameter settings.

[0039] Based on the differentiated texture drift patterns, a multi-dimensional drift pattern representation is constructed, which is used to quantify the contribution of each factor to texture drift.

[0040] By using this multi-dimensional drift pattern characterization, the extraction probability of different transformation parameters in the texture drift generator is finely adjusted to ensure that the generated simulated transformation accurately reflects the complex drift state of the current client under the combined effect of multiple factors.

[0041] Through this technical solution, this application can construct a multi-dimensional drift pattern representation by receiving multi-source data and performing feature extraction and pattern recognition, thereby achieving a refined quantification of the causes of texture drift and adjusting the extraction probability of transformation parameters accordingly to ensure that the simulated transformation more accurately reflects the complex drift state of the client. As an optional solution, the extraction probability of different transformation parameters in the texture drift generator is finely adjusted using this multi-dimensional drift pattern representation, including:

[0042] This multi-dimensional drift pattern representation is received through the local model;

[0043] The multi-dimensional drift pattern representation is used as input via a mapping network;

[0044] By analyzing the correlation between drift pattern representation and actual texture drift in the long-term running data of the client through this mapping network, the internal connection weights are adjusted, and the extraction probability of different transformation parameters in the texture drift generator is calculated and output in real time based on the adjusted internal connection weights.

[0045] Through this technical solution, this application can analyze the correlation between drift pattern representation and actual texture drift by introducing a mapping network, and adjust the extraction probability of transformation parameters in real time, thereby achieving more intelligent and dynamic control over the texture drift generator and further improving the accuracy of simulation transformation.

[0046] In one implementation, the correlation between drift pattern representations and actual texture drift in the client's long-term runtime data is analyzed through the mapping network, and the internal connection weights are adjusted, including:

[0047] Receive the multi-dimensional drift pattern representation and the actual texture drift data;

[0048] Data quality assessment was performed on the multi-dimensional drift pattern representation and the actual texture drift data to identify noisy data, outlier data, and incomplete data.

[0049] Process the noisy data, the outlier data, and the incomplete data;

[0050] The processed multi-dimensional drift pattern representation and the actual texture drift data are input into the mapping network;

[0051] The mapping network is used to analyze the correlation between the drift pattern representation and the actual texture drift, and the internal connection weights are adjusted.

[0052] Through this technical solution, this application can ensure that the mapping network learns on cleaner and more reliable data by performing quality assessment and processing on the input data, thereby improving the accuracy of the correlation analysis between drift pattern representation and actual texture drift, and further optimizing the adjustment of internal connection weights.

[0053] In another implementation, the mapping network adjusts the internal connection weights by analyzing the correlation between drift pattern representations and actual texture drift in the client's long-term runtime data, including:

[0054] Receive the multi-dimensional drift pattern representation and the actual texture drift data;

[0055] The received multi-dimensional drift pattern representation and actual texture drift data are used to identify the client source and distinguish the data from different medical institution clients.

[0056] Based on each client source identified, a dedicated sub-mapping network is trained through this local model. This sub-mapping network is responsible for learning the correlation between the drift pattern representation of a specific client source and the actual texture drift.

[0057] This local model introduces an attention allocation mechanism that dynamically activates and weights the output of the corresponding sub-mapping network based on the client source represented by the drift pattern being processed.

[0058] This attention allocation mechanism analyzes the characteristics of the current client's source and generates a set of attention weights.

[0059] Based on this attention weight, the prediction results of different sub-mapping networks are aggregated to obtain the final transformation parameter extraction probability;

[0060] The internal connection weights are adjusted based on the final transformation parameter extraction probability.

[0061] Through this technical solution, this application can achieve personalized learning and aggregation of drift patterns from different client sources by introducing client source identification, dedicated sub-mapping networks and attention allocation mechanisms, thereby more accurately capturing and processing heterogeneity among clients and further improving the model's adaptability to complex federated learning environments.

[0062] Secondly, this application also discloses an artificial intelligence model training system based on federated learning, used for training cross-institutional medical diagnostic models in federated learning scenarios. The system includes:

[0063] The input end is used to receive model parameter updates from the client, as well as local data characteristic information describing the client, which reflects the underlying image texture introduced by the client during data acquisition.

[0064] The judgment end is used to determine, based on the local data characteristic information, whether the model parameter update includes the underlying image texture features introduced by the client;

[0065] The adjustment end is used to adjust the processing method of the model parameter update in the global model aggregation according to the judgment result, so as to integrate the update containing the underlying texture features of the image.

[0066] This application provides a system-level solution that, through the collaborative work of the input end, the judgment end, and the adjustment end, effectively identifies and processes the underlying texture features of images introduced by the client in federated learning, thereby improving the training effect and generalization ability of cross-institutional medical diagnostic models.

[0067] Beneficial effects

[0068] This application discloses a method and system for training artificial intelligence models based on federated learning. It receives model parameter updates and local data characteristic information from clients, and determines whether the model parameter updates contain low-level image texture features introduced by the client based on the local data characteristic information. Then, based on the determination result, it adjusts the processing method of model parameter updates in global model aggregation to integrate updates containing low-level image texture features. This method effectively solves the problem in existing federated learning where the central aggregation server cannot correctly understand and effectively absorb parameter updates from a minority of clients that contain special data feature information, leading to a decline in the generalization ability of the global model. By identifying and specifically processing the low-level image texture features introduced by the client, this application can avoid the global model being overly biased towards the data features of mainstream active hospitals, thereby improving the diagnostic accuracy of the model when facing marginal cases with special equipment characteristics or similar data distributions. It effectively generalizes to all patient groups in the entire medical ecosystem, significantly improving the accuracy and robustness of cross-institutional medical diagnostic models. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating an artificial intelligence model training method based on federated learning provided in an embodiment of the present invention.

[0070] Figure 2 This is a flowchart of a method for adjusting model parameter updates in global model aggregation, provided by an embodiment of the present invention.

[0071] Figure 3 This is a schematic diagram of the structure of an artificial intelligence model training system based on federated learning provided in an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Reference Figure 1 , Figure 1 This is a flowchart illustrating an artificial intelligence model training method based on federated learning provided by an embodiment of the present invention, used to train a cross-institutional medical diagnosis model in a federated learning scenario.

[0074] S11, receive model parameter updates from the client, as well as local data characteristic information describing the client, the local data characteristic information reflecting the underlying image texture introduced by the client during data acquisition;

[0075] S12, based on the local data characteristic information, determine whether the model parameter update includes the image underlying texture features introduced by the client;

[0076] S13, Based on the judgment result, adjust the processing method of the model parameter update in the global model aggregation to integrate the update containing the underlying texture features of the image.

[0077] This application aims to address the issue of how differences in client devices affect the aggregation performance of image underlying texture features in federated learning. By introducing a mechanism for perceiving and judging local data characteristics, this application can identify and process model parameter updates containing specific image underlying texture features, thereby optimizing the global model aggregation process and improving the model's generalization ability and robustness.

[0078] The "federated learning" mentioned in this application is a distributed machine learning paradigm that allows multiple clients to train models locally and send model parameter updates to a central server for aggregation, without sharing the original data. This approach achieves collaborative training while protecting data privacy.

[0079] "Client" refers to the individual medical institutions or data owners participating in federated learning, who possess medical image data locally and train their models there. "Model parameter updates" refer to the parameter differences between the client's local model and the global model after training; these differences are uploaded to the central aggregation server. "Local data characteristic information" refers to metadata describing the characteristics of the client's local dataset, reflecting the underlying image texture introduced by the client during data acquisition, such as weak, non-pathological background textures caused by equipment aging, environmental factors, or imaging parameter settings. "Image underlying texture features" refer to low-level, recurring visual patterns in the image; these patterns are typically related to the image acquisition device, environment, or processing method, rather than the image content itself.

[0080] "Global model aggregation" refers to the process by which a central aggregation server integrates model parameter updates from different clients to generate a new global model. The core of this application lies in intelligently adjusting the processing method of model parameter updates in global model aggregation by judging local data characteristics, ensuring that updates containing specific image underlying texture features are effectively integrated, rather than simply averaged or ignored.

[0081] Specifically, the implementation of this application includes the following steps:

[0082] First, the system receives model parameter updates from clients, along with information describing the client's local data characteristics. During the federated learning training cycle, each client, after completing local model training, uploads its local model parameter updates to the central aggregation server. Simultaneously, the client also uploads local data characteristic information. This information can be generated locally by the client, for example, by pre-analyzing the local dataset to extract statistical features related to the underlying texture of the image, such as texture intensity, directionality, and contrast. As one implementation, the client can configure a dedicated module that performs texture analysis on a small number of sample images in the local dataset before or after each local training session, generating a simplified texture descriptor and uploading it as part of the local data characteristic information. For example, a Gabor filter bank can be used to extract multi-scale, multi-directional texture features from the image, and then statistical analysis can be performed on these features to obtain a fixed-dimensional vector as the texture descriptor.

[0083] Secondly, based on local data characteristic information, it is determined whether the model parameter update includes low-level image texture features introduced by the client. After receiving the model parameter update and local data characteristic information uploaded by the client, the central aggregation server uses the local data characteristic information to determine whether the model parameter update includes specific low-level image texture features introduced by the client device or environment. As one implementation, the central aggregation server can pre-set a texture feature library containing known low-level image texture patterns that may be generated by older devices or specific environments. The server compares the received local data characteristic information with this texture feature library; if the matching degree reaches a preset threshold, it is determined that the model parameter update includes low-level image texture features. For example, a classifier can be trained that takes local data characteristic information as input and outputs the probability of whether the client data contains specific texture features.

[0084] Finally, based on the judgment results, the processing method of model parameter updates in global model aggregation is adjusted to integrate updates that include the underlying texture features of the image.

[0085] Once the central aggregation server determines that a client's model parameter update contains specific low-level image texture features, it will no longer simply average this update with other updates. Instead, it will adjust its processing method in the global model aggregation based on the determination result. One approach is to assign a lower aggregation weight to updates identified as containing specific low-level image texture features to avoid excessive interference with the learning of the mainstream features of the global model. For example, the contribution of the client's model update can be dynamically adjusted based on the strength or uniqueness of the texture features. Another approach is to design a dedicated aggregation strategy for these updates containing specific texture features, such as fusing them with a pre-trained texture adaptation module instead of directly aggregating them with the main model.

[0086] The overall working principle of this application lies in the fact that, by perceiving and judging the characteristics of local data on the client side, the central aggregation server can identify model parameter updates that introduce specific underlying texture features of images due to differences in devices or environments. Traditional federated learning methods often simply average the model parameter updates from all clients during aggregation, which may lead to the dilution or misinterpretation of updates containing unique texture features, thus affecting the generalization ability of the global model. This application introduces an intelligent judgment mechanism to distinguish these special updates and adjust their processing method in the global model aggregation accordingly. For example, for updates containing specific texture features, strategies such as weighted aggregation, feature separation aggregation, or meta-learning aggregation can be used to ensure that these unique feature information can be effectively integrated, rather than being ignored as "noise." As a result, the global model can not only learn mainstream disease diagnostic features, but also adapt to the subtle texture differences introduced by different client devices and environments, thereby improving the robustness and accuracy of the model in diverse medical scenarios.

[0087] The core innovation of this application lies in its breakthrough of the traditional federated learning's assumption of homogeneity in client data during model aggregation. Traditional methods often struggle to effectively integrate systematic differences in client data (such as differences in underlying image textures), leading to limited generalization ability of the global model. This application introduces a mechanism for perceiving and judging the characteristics of local client data, enabling intelligent identification and processing of model parameter updates containing specific underlying image texture features. Compared to the closest existing technology, this application's advantage lies in its differentiated processing based on the actual characteristics of the client data, rather than simply aggregating all model parameter updates indiscriminately. This differentiated processing allows the global model to better adapt to subtle texture differences introduced by different client devices and environments, significantly improving the model's diagnostic accuracy and generalization ability in diverse medical scenarios, effectively solving the problem of model performance degradation caused by device differences.

[0088] In some embodiments described above, this application proposes a method for adjusting model parameter updates based on judgment results in global model aggregation to integrate updates including image underlying texture features. However, in its implementation, simple adjustments may not adequately capture and effectively address the dynamics and diversity of image underlying texture features introduced by different clients. This is especially true in medical image data acquisition, where minor changes in client device status, fluctuations in environmental factors, and differences in operating parameter settings can all lead to subtle drifts in image underlying texture features. If these issues are not addressed, the generalization ability and robustness of the global model may be affected when facing data with dynamically changing underlying texture features from heterogeneous clients, thereby reducing the accuracy of medical diagnosis.

[0089] In this regard, refer to Figure 2 , Figure 2 This is a flowchart of a method for adjusting model parameter updates in global model aggregation according to an embodiment of the present invention. S13 includes:

[0090] S131, When the client's local model is being trained, the feature extraction layer inside the local model is guided to become sensitive to small changes in the client's state.

[0091] S132, Extract a dynamic fingerprint representing the current data style of the client from the local model;

[0092] S133, monitor and sample the real-time environment and operation parameter vectors of the medical image client, and incorporate the client's feature summary so that after the central aggregation server receives the model parameter updates, dynamic fingerprints, real-time environment and operation parameter vectors from each client, it uses a meta-aggregation network to calculate and adjust the processing method of each client's model update in the global aggregation, and adjusts the contribution of each client's model update according to the adjustment factor output by the meta-aggregation network, and integrates it into the global model;

[0093] The meta-aggregation network is trained through meta-learning, with the goal of learning a function that predicts aggregation weights and learning rate adjustment factors based on the dynamic client context.

[0094] Specifically, guiding the feature extraction layer within the local model to become sensitive to subtle changes in the client's state refers to using specific training strategies or regularization mechanisms to enable the local model to recognize and adapt to subtle changes in the underlying texture of the image caused by factors such as client device wear and tear, fluctuations in ambient temperature and humidity, or imaging parameter settings. The aim is to enhance the local model's ability to perceive drift in specific client-specific data features, thereby more accurately reflecting these changes in model updates.

[0095] Extracting a dynamic fingerprint representing the current data style of the client from the local model can be understood as extracting feature vectors from specific layers of the local model (such as early convolutional layers) that can compactly represent the unique texture features of the current client data. This dynamic fingerprint serves as a real-time representation of the client data style, aiming to provide the central aggregation server with quantitative information about the heterogeneity of the client data.

[0096] In practical applications, monitoring and sampling the real-time environment and operational parameter vectors of a client device for medical imaging specifically refers to collecting data from internal sensor sensors (such as temperature and humidity), external environmental monitoring data, and imaging parameter settings recorded in the device's operation log. These parameters provide the external context of the client's operational status, aiming to supplement the dynamic fingerprint and more comprehensively characterize the potential influencing factors in the client's data generation process.

[0097] Furthermore, the meta-aggregation network can be understood as a specially designed neural network trained through meta-learning. The network aims to learn a function that predicts the aggregation weights and learning rate adjustment factors for each client model update during global aggregation, based on the client's dynamic client context (including dynamic fingerprints, real-time environment, and operational parameter vectors). Its purpose is to achieve adaptive adjustment of the contribution of client model updates, thereby more effectively integrating updates that include low-level image texture features.

[0098] This application's solution constructs comprehensive dynamic client context information by introducing sensitive guidance for subtle changes in client state, dynamic fingerprint extraction, and real-time monitoring of environment and operational parameter vectors. Thanks to this rich context information, the central aggregation server can utilize a meta-aggregation network to perform refined processing of model updates for each client. Through meta-learning, the meta-aggregation network learns how to intelligently predict and adjust the aggregation weights and learning rate adjustment factors for each client's model update based on the client's dynamic context. This adaptive adjustment mechanism allows the global model to fully consider and effectively integrate the unique, even constantly shifting, underlying image texture features of different clients during the aggregation process, thus overcoming the limitations of traditional simple aggregation methods when processing heterogeneous and dynamic texture data.

[0099] Through the aforementioned technical solution, this application significantly enhances the ability of federated learning models to adapt to client heterogeneity and dynamic changes when processing cross-institutional medical image data. Specifically, by guiding local models to become sensitive to subtle changes in client states and extracting dynamic fingerprints, the system can more accurately capture and quantify client-specific image underlying texture features and their drift trends. Combined with real-time environment and operational parameter vectors, comprehensive client dynamic context information is provided to the central aggregation server. Therefore, the meta-aggregation network can intelligently adjust the contribution of each client model update based on this context information, thereby achieving more effective integration of updates containing image underlying texture features. This refined aggregation strategy not only enhances the robustness of the global model to client data heterogeneity but also improves the model's adaptability to image underlying texture drift when facing changes in client device states or environments, ultimately contributing to improved generalization ability and accuracy of medical diagnostic models.

[0100] Specifically, the dynamic fingerprint representing the current data style of the client can be extracted from the local model in the following way.

[0101] Extracting a dynamic fingerprint representing the current data style of the client from the local model includes:

[0102] The background texture of the input image is reconstructed by using feature maps extracted from early convolutional layers by the local model and then processed by the decoder network.

[0103] The loss function trained locally on the client is combined with the loss from the disease diagnosis task and the loss from the texture reconstruction task.

[0104] The client extracts feature vectors from a specific layer in the local model that is responsible for extracting low-level features.

[0105] The client performs global average pooling on the feature vector to obtain a fixed-dimensional feature vector, which serves as a dynamic fingerprint representing the current data style of the client.

[0106] Specifically, reconstructing the background texture of the input image using feature maps extracted from early convolutional layers in a local model and then fed into a decoder network involves utilizing the feature representations generated by the early convolutional layers in the local model, which are responsible for capturing the basic visual elements of the image, and inputting them into a specially designed decoder network. This decoder network upsamples these abstract feature maps and transforms them back into image space, aiming to accurately reconstruct the background texture of the input image that is irrelevant to disease diagnosis. Early convolutional layers are typically understood as layers near the input in a neural network, primarily responsible for extracting low-level visual features such as edges, corners, and textures. The decoder network can be understood as a neural network structure capable of mapping low-dimensional feature representations back to high-dimensional pixel space; for example, it can employ deconvolutional layers or a combination of upsampling layers and convolutional layers. Its purpose is to transform the abstract features extracted from the early convolutional layers into recognizable image textures.

[0107] The local training loss function on the client, combined with the loss functions for disease diagnosis and texture reconstruction, can be understood as a composite loss function used during local training. This composite loss function consists of two parts: one part is the loss for the core disease diagnosis task, such as cross-entropy loss or mean squared error loss, used to measure the model's performance in diagnostic accuracy; the other part is the loss for the texture reconstruction task, such as pixel-level mean squared error loss or perceptual loss, used to measure the accuracy of the decoder network in reconstructing background texture. The aim is to simultaneously optimize these two loss components, guiding the local model to effectively perceive, extract, and reconstruct the client-specific low-level image texture information while learning disease-related features, thus maintaining the model's sensitivity to these texture features.

[0108] In practical applications, the client extracts feature vectors from specific layers in the local model responsible for extracting low-level features. Specifically, this involves selecting one or more layers in the local model that are dedicated to capturing low-level visual features of the image; these could be earlier convolutional layers or specific feature extraction modules. The output extracted from these layers, i.e., feature maps, is considered as feature vectors. The purpose is to ensure that the extracted feature vectors primarily reflect the image's underlying texture, color, edges, and other general visual styles, rather than high-level semantic content, thereby accurately representing the client's data style.

[0109] Furthermore, by performing global average pooling on the feature vector through the client, a fixed-dimensional feature vector is obtained, which serves as a dynamic fingerprint representing the client's current data style. This refers to performing global average pooling (GAP) on the feature vector extracted from a specific layer (usually a multi-dimensional feature map). Global average pooling compresses the spatial dimension of each feature map into a single value by calculating the average value of each feature channel, thereby converting the entire feature map into a fixed-dimensional feature vector. Its purpose is to convert a variable-size feature map into a fixed-dimensional representation, facilitating subsequent processing and use as a dynamic fingerprint, while preserving the global information of the feature map, enabling it to stably and effectively represent the client's current data style.

[0110] This application's solution introduces a texture reconstruction task during local model training and combines it with the loss function of a disease diagnosis task. This allows the local model to learn diagnostic features while simultaneously being guided to perceive and reconstruct client-specific background textures. Through this technical solution, a dynamic fingerprint representing the current data style of the client can be effectively extracted from the local model. This dynamic fingerprint accurately captures the low-level image texture features introduced by the client during data acquisition, providing crucial information for the subsequent processing of model parameter updates by the central aggregation server. By combining the loss functions of the texture reconstruction and disease diagnosis tasks, the local model can simultaneously focus on diagnostic accuracy and client texture characteristic recognition during training, improving the representativeness and accuracy of the dynamic fingerprint. Furthermore, by extracting from the low-level feature layer and performing global average pooling, the stability and comparability of the extracted fingerprint are ensured, facilitating cross-client style analysis and aggregation adjustments in a federated learning environment. This, in turn, enhances the robustness and generalization ability of the federated learning model in the face of client data heterogeneity.

[0111] In some of the embodiments described above in this application, it is proposed that the feature extraction layer within the local model be guided to become sensitive to subtle changes in the client's state during training. However, in practical implementation, effectively and robustly making the local model sensitive to subtle, dynamic shifts in the underlying texture of the image introduced during client data acquisition is a challenging problem. Simply guiding the model may not be sufficient to cope with the complex and varied texture shift patterns in the real world, potentially leading to insufficient adaptability of the model to changes in the style of client data.

[0112] In response, this application further proposes a method to guide the feature extraction layer within the local model to become sensitive to minute changes in the client state, which includes:

[0113] During the training process, the local model periodically applies a transformation to the input image that simulates client-side background texture drift;

[0114] Identify or predict the parameters of the transformation using a local model;

[0115] Based on the identified or predicted transformation parameters, the feature extraction layer inside the local model is guided to become sensitive to subtle changes in the client's state.

[0116] Specifically, periodically applying simulated client-side background texture drift to the input image refers to artificially introducing perturbations similar to actual client-side background texture drift into the training data during each cycle or specific time interval of local model training. These perturbations can simulate changes in the underlying texture of the image caused by factors such as device aging, changes in ambient lighting, and fine-tuning of imaging parameters. The purpose is to expose the model to various simulated drift states during the training phase, enabling it to learn the commonalities and characteristics of these drift patterns.

[0117] The identification or prediction of transformation parameters through a local model can be understood as follows: after a simulated transformation is applied, the local model is trained to identify or predict the specific parameters of these transformations. For example, if the applied transformation is a type of noise or geometric distortion, the model can be trained to output the intensity, type, or specific parameters of the distortion. The aim is to enable the model not only to perceive the existence of texture drift but also to understand the nature and extent of the drift, thereby providing a basis for subsequent sensitivity guidance.

[0118] In practical applications, based on the identified or predicted transformation parameters, the feature extraction layer within the local model is guided to become sensitive to subtle changes in the client's state. Specifically, this means adjusting the learning strategy or weights of the feature extraction layer using the transformation parameters identified or predicted by the model. For example, an auxiliary loss function can be designed to encourage the feature extraction layer to generate discriminative feature representations when faced with different transformation parameters. Alternatively, a self-supervised learning task can be used to enable the feature extraction layer to predict whether the input image has undergone a certain texture shift transformation and its parameters. The goal is to enable the feature extraction layer to actively focus on and encode the underlying texture features that reflect subtle changes in the client's state, thereby improving the model's adaptability to changes in the client's data style.

[0119] The proposed solution applies simulated client-side background texture shifts periodically during local model training, enabling the model to encounter and learn to handle various potential texture shift patterns during the training phase. Because the model is forced to identify or predict the parameters of these shifts, its internal feature extraction layer actively learns discriminative features related to these texture shifts. In this way, the feature extraction layer no longer focuses solely on disease-related semantic features but also becomes sensitive to subtle changes in the underlying image texture caused by the client's state, thus capturing the true characteristics of the client data more comprehensively.

[0120] By employing the aforementioned technical solution, this application effectively addresses the potential robustness limitations of traditional methods in guiding local models to adapt to subtle changes in client state. Through proactive simulation and identification of background texture drift, the local model is trained to be more robust and better able to adapt to dynamic changes in client data styles within real-world federated learning environments. This not only improves the model's generalization ability to different client data characteristics but also enables the global model to more accurately integrate updates from various clients during aggregation, thereby enhancing the overall performance and reliability of cross-institutional medical diagnostic models.

[0121] In some preferred embodiments, when the local model periodically applies a transformation simulating client-side background texture drift to the input image during training, this can be achieved as follows: First, a texture drift generator is constructed, capable of generating various images simulating texture drift based on preset parameters (e.g., noise type, intensity, direction, contrast variation, slight geometric distortion, etc.). In each training batch, a set of parameters is randomly selected from this generator and applied to the current input medical image. For example, slight stripe noise due to X-ray machine aging, or contrast or brightness drift caused by differences in parameter settings of different imaging devices, can be simulated.

[0122] Furthermore, the identification or prediction of transformation parameters through the local model can be achieved by adding an auxiliary prediction head after the backbone network of the local model. This auxiliary prediction head is trained to predict the specific parameters of the texture drift transformation applied to the input image. For example, if Gaussian noise is applied, the auxiliary prediction head will attempt to predict the mean and variance of the Gaussian noise. This prediction task can be trained in parallel with the main disease diagnosis task as a self-supervised task, and its loss function can prompt the feature extraction layer to learn feature representations sensitive to these transformation parameters.

[0123] Therefore, based on the identified or predicted transformation parameters, the feature extraction layer within the local model can be guided to become sensitive to subtle changes in the client's state. This can be achieved by adjusting the weight update strategy of the feature extraction layer or introducing an attention mechanism. For example, a loss term can be designed to reward the feature extraction layer when the model successfully predicts the transformation parameters, or to penalize it when the prediction fails. Furthermore, the predicted transformation parameters can be used to dynamically adjust the attention weights of different channels or regions in the feature extraction layer, enabling the model to pay more attention to regions or features significantly affected by texture drift, thereby enhancing its ability to perceive subtle changes in the client's state.

[0124] In some embodiments described above, a method is proposed to periodically apply simulated client background texture shifts to the input image during local model training, in order to guide the feature extraction layer within the local model to become sensitive to subtle changes in the client state. However, in its implementation, if the simulated texture shifts cannot adequately reflect the complex and multi-factor-induced background texture changes of the medical imaging client in actual operation, it may lead to poor improvement in the local model's sensitivity to real client state changes, thereby affecting the robustness and diagnostic accuracy of the global model.

[0125] In response, this application further proposes that, during the training process of the aforementioned local model, a transformation simulating client-side background texture drift is periodically applied to the input image, including:

[0126] Based on the influence of wear and tear of internal components, ambient temperature and humidity fluctuations, and imaging parameter settings on image background texture under different operating states of medical imaging clients, a parameterized texture drift generator is constructed. The texture drift generator is used to generate various combinations of transformation parameters for noise distribution patterns, weak stripe directions, contrast gradients, and geometric distortion.

[0127] During each training cycle, the local model randomly selects a set of transformation parameters from the texture drift generator and performs real-time transformation on the input image of the current training batch according to the set of transformation parameters to generate an image simulating the background texture drift of the client.

[0128] During the training process, the local model of the client analyzes the long-term running data of the client to identify the potential evolution path and key influencing factors of the client's background texture drift, and dynamically adjusts the parameter extraction strategy in the texture drift generator to prioritize generating simulated transformations that better match the current actual drift state of the client.

[0129] Specifically, constructing a parameterized texture drift generator involves a deep analysis of how medical imaging equipment, such as X-ray machines and CT scanners, in actual use—including long-term wear and tear of internal mechanical components, seasonal or daily fluctuations in ambient temperature and humidity, and subtle adjustments or drifts in imaging parameters (e.g., exposure time, current, voltage) by operators—commonly or independently affect the background texture of the final image. Based on these influencing factors, a mathematical model or algorithm is designed to generate various simulated texture drifts. This model can output different types of noise (e.g., Gaussian noise, salt-and-pepper noise), subtle stripes of different directions and intensities, contrast gradients of local or overall image data, and parameter combinations of various geometric distortions (e.g., slight scaling, rotation, shearing, or perspective transformations). This generator aims to provide a rich and controllable texture drift simulation space.

[0130] In this model, during each training epoch, a set of transformation parameters is randomly selected from the texture drift generator. These parameters are then used to transform the input images of the current training batch in real time, generating images that simulate background texture drift on the client side. This means that new, random, but controlled texture drifts are introduced with each model iteration, ensuring the model remains robust to various potential real-world drift scenarios. This real-time transformation effectively expands the training data, increases its diversity, and thus improves the model's generalization ability.

[0131] In practical applications, during the training process, the local model on the client analyzes long-term operational data to identify the potential evolution paths and key influencing factors of background texture drift. It then dynamically adjusts the parameter extraction strategy in the texture drift generator to prioritize generating simulated transformations that better match the current drift state of the client. This is an adaptive mechanism that, by collecting and analyzing historical operational data from the client (such as device logs, environmental sensor data, maintenance records, etc.), can reveal the client's unique texture drift patterns and trends. For example, if aging of the X-ray tube on a client causes stripes in a specific direction in the image background, the system will adjust the generator to generate simulated transformations with similar stripe features more frequently. The goal is to make the simulated drift more closely resemble the client's real-world situation, thereby training the local model more effectively and making it more sensitive to changes in the client's actual state.

[0132] This application addresses the limitations of traditional methods that may not simulate texture drift realistically or adaptively enough by constructing a parameterized texture drift generator and dynamically adjusting its parameter extraction strategy based on the analysis of long-term client operation data. Through this technical solution, the application significantly improves the realism and adaptability of the local model in simulating client background texture drift. Compared to methods that only periodically apply general simulation transformations, this solution, through a refined parameterized generator and a dynamic adjustment mechanism based on long-term data analysis, enables the simulated texture drift to more accurately reflect the subtle and dynamic changes in the medical imaging client caused by various complex factors during actual operation. This not only enhances the local model's generalization ability to various potential drift patterns during training, but more importantly, it allows the local model to perform customized learning for the actual drift trends of a specific client, thereby more effectively guiding its internal feature extraction layer to become highly sensitive to subtle changes in the client's state. Therefore, in federated learning scenarios, global models can better handle and integrate model updates from different clients that are affected by their unique texture drift during aggregation, ultimately improving the robustness, accuracy, and stability of cross-institutional medical diagnostic models in real-world deployments.

[0133] In some embodiments described above, this application proposes that during the training process of the client's local model, based on the analysis of long-term operational data of the client, the potential evolution path and key influencing factors of background texture drift are identified, and the parameter extraction strategy in the texture drift generator is dynamically adjusted to prioritize generating simulated transformations that better match the current actual drift state of the client. However, in practical applications, background texture drift of medical imaging clients is often a complex phenomenon with multiple coupled factors, which may be simultaneously affected by factors such as wear and tear of internal components, fluctuations in ambient temperature and humidity, and deviations in operating parameter settings. If only macroscopic analysis of long-term operational data is relied upon, it may be difficult to accurately distinguish and quantify the respective contributions of these different factors to texture drift, resulting in simulated transformations generated by the texture drift generator that are not refined or accurate enough, failing to fully reflect the client's true complex drift state, and thus affecting the adaptability of the local model to changes in real-world data.

[0134] In response, this application further proposes that, during the training process of the client's local model, based on the analysis of long-term running data of the client, the potential evolution path and key influencing factors of the client's background texture drift are identified, and the parameter extraction strategy in the texture drift generator is dynamically adjusted to prioritize the generation of simulated transformations that better match the current actual drift state of the client, including:

[0135] The local model receives data from internal sensors on the client, environmental monitoring data, and operation log information.

[0136] The received data is processed by a local model to extract features and recognize patterns, in order to distinguish texture drift patterns caused by different factors such as client wear, ambient temperature and humidity fluctuations, and deviations in operating parameter settings.

[0137] Based on the differentiated texture drift patterns, a multi-dimensional drift pattern representation is constructed. This multi-dimensional drift pattern representation is used to quantify the contribution of each factor to texture drift.

[0138] By using multi-dimensional drift pattern characterization, the extraction probability of different transformation parameters in the texture drift generator is finely adjusted to ensure that the generated simulated transformation accurately reflects the complex drift state of the current client under the combined effect of multiple factors.

[0139] Specifically, receiving data from internal sensors on the client side, environmental monitoring data, and operation log information via a local model means that the local model is configured to acquire diverse data on the client's operational status in real time or periodically. This data includes, but is not limited to, internal sensor data reflecting hardware wear or operational status, such as internal device temperature, vibration frequency, current, and voltage. Environmental monitoring data can include real-time environmental parameters such as temperature, humidity, and air pressure of the client's environment. Operation log information can record operator adjustments to imaging parameters, equipment maintenance records, and error reports. All of this data collectively constitutes a comprehensive observation of the factors influencing background texture drift on the client side.

[0140] Furthermore, the received data is processed using a local model for feature extraction and pattern recognition to distinguish texture drift patterns caused by various factors such as client wear, ambient temperature and humidity fluctuations, and deviations in operating parameter settings. This means that the local model integrates data processing and analysis modules, enabling it to extract key features related to texture drift from raw sensor, environmental, and log data. For example, through time series analysis, spectral analysis, or machine learning classifiers, the correlation between specific data patterns and certain drift factors (such as specific noise patterns caused by equipment aging, image contrast changes caused by temperature increases, and geometric distortion caused by adjustments in operating parameters) can be identified. The aim is to decouple the complex, multi-factor coupled texture drift phenomenon into patterns caused by a single or a few dominant factors, thereby providing a basis for subsequent fine-tuning.

[0141] Based on the differentiated texture drift patterns, a multi-dimensional drift pattern representation is constructed to quantify the contribution of each factor to texture drift. Specifically, after identifying different drift patterns, the system assigns one or a set of values ​​to each pattern, forming a vector or matrix to quantify its intensity and scope of influence. For example, a wear index can be assigned to the "device wear" pattern, a temperature influence factor to the "ambient temperature" pattern, and a deviation weight to the "operating parameter deviation" pattern. This multi-dimensional representation provides a quantitative and comprehensive perspective to understand the current texture drift state of the client, rather than simply qualitatively identifying drift.

[0142] In practical applications, the sampling probability of different transformation parameters in the texture drift generator is finely adjusted through multi-dimensional drift pattern characterization to ensure that the generated simulated transformation accurately reflects the complex drift state of the current client under the combined influence of multiple factors. This means that the texture drift generator no longer simply randomly samples transformation parameters, but dynamically adjusts the sampling probability distribution of its internal parameters based on the quantized drift pattern characterization of the current client. For example, if the multi-dimensional drift pattern characterization shows that the current client is mainly affected by "device wear" and "high humidity," then the texture drift generator will prioritize sampling transformation parameter combinations that can simulate noise caused by wear and blurring or stripes caused by high humidity, thus making the simulated transformation closer to the actual situation of the client.

[0143] This application's solution, by introducing the reception and analysis of internal sensor data, environmental monitoring data, and operation log information from the client, can comprehensively perceive the client's operating status and environmental changes. Through this technical solution, this application can significantly improve the accuracy and realism of simulating client background texture drift. Compared to relying solely on long-term operational data for macroscopic adjustments, this solution achieves refined identification and quantification of texture drift causes through in-depth analysis of multi-source data. This allows the texture drift generator to generate simulated transformations that highly match the client's actual drift state. This highly matched simulated data can more effectively train the local model, making it more sensitive to subtle changes in the client's state. Therefore, even when faced with complex client background texture drift during actual deployment, it can maintain stable diagnostic performance. Furthermore, by quantifying the contribution of each factor, this solution also provides valuable reference information for client maintenance and fault diagnosis, further enhancing the robustness and practicality of the federated learning system in the field of medical diagnosis.

[0144] In some embodiments described above, this application proposes analyzing long-term client operation data to identify potential evolution paths and key influencing factors of client background texture drift, and dynamically adjusting the parameter extraction strategy in the texture drift generator to prioritize generating simulated transformations that better match the current actual drift state of the client. However, in practical applications, relying solely on multi-dimensional drift pattern representation to adjust the extraction probability of different transformation parameters in the texture drift generator may not be sufficient to capture the complex nonlinear relationships between various factors, leading to deviations between the simulated transformation and the complex drift state of the actual client, thereby affecting the sensitivity and adaptability of the local model to real texture drift.

[0145] In response, this application further proposes a method for finely adjusting the extraction probability of different transformation parameters in a texture drift generator through multi-dimensional drift pattern characterization, specifically including:

[0146] Receive multi-dimensional drift pattern representations through a local model;

[0147] A mapping network is used to represent multi-dimensional drift patterns as input;

[0148] By analyzing the correlation between drift pattern representation and actual texture drift in long-term client data through mapping network analysis, the internal connection weights are adjusted, and the extraction probability of different transformation parameters in the texture drift generator is calculated and output in real time based on the adjusted internal connection weights.

[0149] Specifically, the local model is configured to receive and process a multi-dimensional drift pattern representation constructed from data from client-side internal sensors, environmental monitoring, and operational logs. This multi-dimensional drift pattern representation quantifies the contribution of various factors, such as client wear, ambient temperature and humidity fluctuations, and deviations in operational parameter settings, to texture drift. The mapping network can be understood as a non-linear function approximator, such as a multilayer perceptron or a more complex neural network structure, aiming to learn and model the complex mapping relationship between the multi-dimensional drift pattern representation and actual texture drift. In practical applications, the mapping network analyzes historical drift pattern representations and observed texture drift data from long-term client operation data, for example through supervised learning, to continuously adjust its internal connection weights to minimize the error between the predicted extraction probability and the actual required extraction probability.

[0150] Therefore, the mapping network can calculate and output the extraction probability of various transformation parameters (such as noise distribution pattern, weak stripe direction, contrast gradient, geometric distortion, etc.) in the texture drift generator in real time and accurately according to the drift pattern representation of the current client, thereby ensuring that the generated simulated transformation is highly matched with the actual complex drift state of the current client.

[0151] This application addresses the problem of accurately capturing complex texture drift patterns using only multi-dimensional drift pattern representations by introducing a mapping network. Through this technical solution, the application enables refined and adaptive adjustment of the probability of transformation parameter extraction in the texture drift generator. Specifically, by using a mapping network for deep learning and analysis of the complex correlation between multi-dimensional drift pattern representations and actual texture drift, it overcomes the limitations of traditional methods where simple rules or linear models struggle to capture multi-factor coupling effects. Consequently, the generated simulated texture drift transformations more accurately and realistically reflect the actual operating state and environmental changes of the client, significantly improving the sensitivity and robustness of the local model to various complex texture drifts in the real world. This not only helps improve the generalization ability and diagnostic accuracy of medical diagnostic models in federated learning scenarios but also effectively reduces the risk of model performance degradation due to client data characteristic drift, thus providing a more stable and reliable artificial intelligence model training method for cross-institutional medical diagnosis.

[0152] In some embodiments described above, this application proposes analyzing the correlation between multi-dimensional drift pattern representations and actual texture drift using a mapping network, and adjusting the internal connection weights to calculate and output the extraction probabilities of different transformation parameters in the texture drift generator in real time. However, in practical applications, the long-term running data of the client, including drift pattern representations and actual texture drift data, may inevitably contain noise, outliers, or incomplete data due to factors such as sensor errors, data transmission interruptions, or environmental interference. If this unverified and unprocessed raw data is directly input into the mapping network, it may cause the network to learn incorrect correlation patterns, resulting in inaccurate adjustment of the internal connection weights, which in turn affects the accuracy of the transformation parameter extraction probabilities, ultimately reducing the realism and effectiveness of simulating client background texture drift. To address this issue and further improve the robustness and accuracy of the mapping network's correlation analysis and internal connection weight adjustment, this application proposes an optimized processing method.

[0153] In response, this application further proposes the above-mentioned method of analyzing the correlation between drift pattern representation and actual texture drift in long-term client operation data through mapping networks, and adjusting the internal connection weights, including:

[0154] Receive the multi-dimensional drift pattern representation and actual texture drift data;

[0155] Data quality assessment is performed on the multi-dimensional drift pattern representation and the actual texture drift data to identify noisy data, outlier data, and incomplete data;

[0156] The noise data, the outlier data, and the incomplete data are processed.

[0157] The processed multi-dimensional drift pattern representation and the actual texture drift data are input into the mapping network;

[0158] The mapping network is used to analyze the correlation between the drift pattern representation and the actual texture drift, and the internal connection weights are adjusted.

[0159] Specifically, receiving multi-dimensional drift pattern representations and actual texture drift data means that the local model or central aggregation server collects and acquires from the client a quantized representation (i.e., multi-dimensional drift pattern representation) describing the client's background texture drift state, as well as the actually observed texture drift data. This data is fundamental to understanding the client's texture drift patterns.

[0160] The data quality assessment of multi-dimensional drift pattern representations and actual texture drift data, identifying noisy data, outlier data, and incomplete data, can be understood as a systematic check of the reliability and completeness of the data before it is input into the mapping network. Noisy data refers to random or unsystematic errors, possibly caused by sensor interference or transmission errors; outlier data refers to observations that deviate significantly from the rest of the dataset, possibly indicating measurement errors or special events; incomplete data refers to records missing certain fields or values, possibly due to interrupted data acquisition or storage problems. Identifying these problematic data is a crucial step in ensuring the accuracy of subsequent analysis.

[0161] In practical applications, noisy, outlier, and incomplete data are processed using a series of data preprocessing techniques to correct or remove identified problematic data. For example, noisy data can be processed using smoothing filters (such as Gaussian filtering and median filtering); outlier data can be removed by truncating, replacing it with the mean / median, or using statistical models (such as Z-score and IQR); incomplete data can be processed using interpolation (such as linear interpolation and polynomial interpolation), mean / median padding, or deletion of incomplete records. The aim is to improve the purity and usability of the data, providing high-quality input for the accurate learning of the mapping network.

[0162] Furthermore, inputting the processed multi-dimensional drift pattern representation and actual texture drift data into the mapping network means ensuring that the data meets the input requirements of the mapping network after data quality assessment and processing, and then using it as the network's input features. Thus, the mapping network can perform learning and correlation analysis based on more reliable data.

[0163] Ultimately, by analyzing the correlation between drift pattern representations and actual texture drift through a mapping network and adjusting the internal connection weights, the mapping network, upon receiving high-quality input data, can more accurately learn and model the complex nonlinear relationship between drift pattern representations and actual texture drift. Through training mechanisms such as backpropagation, the network's internal connection weights are iteratively adjusted to minimize prediction errors, thereby enabling the network to more accurately predict the extraction probabilities of different transformation parameters in the texture drift generator.

[0164] This application's solution effectively addresses potential noise, outliers, and incompleteness issues in the original data by introducing a data quality assessment and processing step before the mapping network analyzes the correlation between drift pattern representations and actual texture drift and adjusts the internal connection weights. Through this technical solution, the application significantly improves the accuracy and robustness of the mapping network in analyzing the correlation between drift pattern representations and actual texture drift in long-term client data. By performing rigorous data quality assessment and processing before inputting data into the mapping network, the negative impact of noise, outliers, and incomplete data on model learning is effectively avoided. This allows the mapping network to be trained on more reliable and realistic data, thereby learning a more accurate mapping relationship between drift patterns and actual texture drift. Consequently, the internal connection weights of the mapping network can be more optimized and stable, leading to more accurate prediction of transformation parameter extraction probabilities in the texture drift generator. Ultimately, this improves the realism and effectiveness of simulating client background texture drift, providing a more stable and adaptable training environment for cross-institutional medical diagnostic models in federated learning scenarios.

[0165] This application further proposes a more refined method for adjusting the internal connection weights, which aims to more accurately reflect and adapt to the texture drift characteristics of different clients by introducing client source identification, dedicated sub-mapping networks, and attention allocation mechanisms.

[0166] The above-mentioned mapping network analyzes the correlation between drift pattern representations and actual texture drift in long-term client data, and adjusts the internal connection weights, including:

[0167] Receives multi-dimensional drift pattern representations and actual texture drift data;

[0168] Client source identification is performed on the received multi-dimensional drift pattern representation and actual texture drift data to distinguish data from different medical institution clients;

[0169] Based on each client source identified, a dedicated sub-mapping network is trained using a local model. This sub-mapping network is responsible for learning the correlation between the drift pattern representation of a specific client source and the actual texture drift.

[0170] An attention allocation mechanism is introduced through the local model. This mechanism dynamically activates and weights the output of the corresponding sub-mapping network based on the client source represented by the current drift pattern being processed. The attention allocation mechanism analyzes the features of the current client source and generates a set of attention weights.

[0171] Based on the attention weights, the prediction results of different sub-mapping networks are aggregated to obtain the final transformation parameter extraction probability;

[0172] The internal connection weights are adjusted based on the final transformation parameter extraction probability.

[0173] Specifically, after receiving multi-dimensional drift pattern representations and actual texture drift data, the first step is to identify the client's origin. Client origin identification refers to determining which healthcare institution the data originates from by analyzing metadata contained in the data, such as client identifiers, institution names, or device serial numbers. This step is fundamental to subsequent personalized processing, ensuring that data from each client can be correctly routed to its corresponding processing logic.

[0174] Furthermore, based on the identified client origin, a dedicated sub-mapping network is trained using a local model. This means that for N different medical institution clients, N independent sub-mapping networks will be trained. Each sub-mapping network is designed to specifically learn the correlation between the drift pattern representation and the actual texture drift for its corresponding client. For example, if there are three medical institution clients: Hospital A, Hospital B, and Hospital C, then sub-mapping networks A, B, and C will be trained. Each sub-mapping network can be a multilayer perceptron, convolutional neural network, or recurrent neural network, etc., and its structure and parameters will be optimized according to the data characteristics of the specific client to better capture the unique texture drift patterns of that client.

[0175] The attention allocation mechanism introduced in the local model is a key component of this scheme. This mechanism dynamically activates and weights the outputs of the corresponding sub-mapping networks based on the client source represented by the drift pattern currently being processed. Specifically, when processing data from client A, the attention allocation mechanism assigns higher weights to sub-mapping network A, while assigning lower weights or even activating other sub-mapping networks. The attention allocation mechanism generates a set of attention weights by analyzing the characteristics of the current client source, such as its historical drift data, device configuration, or environmental parameters. These weights reflect the contribution of each sub-mapping network to the probability of predicting the transformation parameters in the current context.

[0176] Therefore, based on the generated attention weights, the prediction results of different sub-mapping networks are aggregated. The aggregation process can be weighted averaging, gating mechanisms, or other fusion strategies, aiming to effectively integrate the prediction results of each sub-mapping network for its specific client. In this way, a comprehensive and more accurate final transform parameter extraction probability can be obtained.

[0177] Finally, based on the final transformation parameter extraction probability, the internal connection weights of the mapping network are adjusted. Here, the mapping network can be understood as the overall architecture containing all sub-mapping networks and their attention allocation mechanisms. Adjusting the internal connection weights means updating the parameters of each sub-mapping network and the attention allocation mechanism, enabling the entire system to more accurately predict and adapt to the client's texture drift.

[0178] This application's solution effectively addresses the challenge of a single mapping network accurately capturing and adapting to diverse texture drift patterns across different clients in multi-institution federated learning scenarios by introducing client source identification, a dedicated sub-mapping network, and an attention allocation mechanism. Through these technical solutions, this application significantly improves the accuracy and adaptability of texture drift simulation during the training of cross-institutional medical diagnostic models in federated learning scenarios. Compared to the aforementioned solutions that use a single mapping network for weight adjustment, this solution, through client source identification and a dedicated sub-mapping network, achieves refined modeling of unique texture drift patterns from different medical institution clients, avoiding prediction biases that may result from a "one-size-fits-all" approach. Furthermore, the introduced attention allocation mechanism allows the system to dynamically select and weight the most relevant sub-mapping network output based on the characteristics of the current client, thereby ensuring highly personalized and accurate predictions of transformation parameter extraction probabilities. This personalized and dynamic adaptability enables the local model to more accurately simulate client background texture drift during training, thereby more effectively guiding the feature extraction layer within the local model to become sensitive to subtle changes in the client's state. Ultimately, this improves the robustness and generalization ability of the global model, especially when dealing with medical image data from different institutions with diverse data characteristics.

[0179] refer to Figure 3 , Figure 3 This is a schematic diagram of an artificial intelligence model training system based on federated learning provided in an embodiment of the present invention. It is used to train a cross-institutional medical diagnostic model in a federated learning scenario, including:

[0180] The input terminal is used to receive model parameter updates from the client, as well as local data characteristic information describing the client, which reflects the underlying image texture introduced by the client during data acquisition.

[0181] The judgment end is used to determine, based on the local data characteristic information, whether the model parameter update includes the underlying image texture features introduced by the client;

[0182] The adjustment end is used to adjust the processing method of the model parameter update in the global model aggregation according to the judgment result, so as to integrate the update including the underlying texture features of the image.

[0183] This application presents a system-level solution to address the impact of varying low-level image texture features on model aggregation performance in federated learning environments. By configuring dedicated input, decision, and adjustment terminals, the system can automatically perceive, identify, and process model parameter updates containing specific low-level image texture features, thereby optimizing the global model aggregation process and significantly improving the model's generalization ability and robustness. The system's modular design ensures the collaborative work of its functional units, enabling refined management and optimization of the federated learning model training process.

[0184] The "federated learning" mentioned in this application is a distributed machine learning paradigm that allows multiple clients to train models locally and send model parameter updates to a central server for aggregation, without sharing the original data. This approach achieves collaborative training while protecting data privacy.

[0185] "Client" refers to the individual medical institutions or data owners participating in federated learning, who possess medical image data locally and train their models there. "Model parameter updates" refer to the parameter differences between the client's local model and the global model after training; these differences are uploaded to the central aggregation server. "Local data characteristic information" refers to metadata describing the characteristics of the client's local dataset, reflecting the underlying image texture introduced by the client during data acquisition, such as weak, non-pathological background textures caused by equipment aging, environmental factors, or imaging parameter settings. "Image underlying texture features" refer to low-level, recurring visual patterns in the image; these patterns are typically related to the image acquisition device, environment, or processing method, rather than the image content itself.

[0186] "Global model aggregation" refers to the process by which a central aggregation server integrates model parameter updates from different clients to generate a new global model. The core of this application lies in intelligently adjusting the processing method of model parameter updates in global model aggregation by judging local data characteristics, ensuring that updates containing specific image underlying texture features are effectively integrated, rather than simply averaged or ignored.

[0187] This system, through the perception and judgment of local data characteristics on the client side, can identify model parameter updates that introduce specific underlying texture features of images due to differences in device or environment. Traditional federated learning methods often simply average the model parameter updates from all clients during aggregation, which may dilute or misinterpret updates containing unique texture features, thus affecting the generalization ability of the global model. This system, by introducing an intelligent judgment mechanism, can distinguish these special updates and adjust their processing method in global model aggregation accordingly. For example, for updates containing specific texture features, strategies such as weighted aggregation, feature separation aggregation, or meta-learning aggregation can be used to ensure that these unique feature information are effectively integrated, rather than being ignored as "noise." Therefore, the global model can not only learn mainstream disease diagnostic features but also adapt to subtle texture differences introduced by different client devices and environments, thereby improving the robustness and accuracy of the model in diverse medical scenarios.

[0188] Specifically, the system in this embodiment of the present application includes the following functional modules:

[0189] First, the input is configured to receive model parameter updates from the client, as well as information describing the client's local data characteristics.

[0190] During the training cycle of federated learning, each client, after completing model training locally, uploads its local model parameter updates to the central aggregation server. Simultaneously, the client also uploads local data characteristic information. This information can be generated locally by the client, for example, by pre-analyzing the local dataset to extract statistical features related to the underlying texture of the images, such as texture intensity, directionality, and contrast. As one implementation, the client can configure a dedicated module that performs texture analysis on a small number of sample images in the local dataset before or after each local training session, generating a simplified texture descriptor and uploading it as part of the local data characteristic information. The input end can be a network interface module responsible for establishing secure communication channels with each client and receiving uploaded data packets. This module can integrate data verification functions to ensure the integrity and correctness of the received model parameter updates and local data characteristic information. For example, the input end can use standard network communication protocols (such as HTTPS) for data transmission and utilize message queue mechanisms to handle concurrent client upload requests.

[0191] Secondly, the judgment end is configured to determine, based on the local data characteristic information, whether the model parameter update includes the underlying image texture features introduced by the client.

[0192] After receiving model parameter updates and local data characteristic information uploaded by the input terminal, the judgment terminal of the central aggregation server uses the local data characteristic information to determine whether the model parameter update contains specific low-level image texture features introduced by the client device or environment. As one implementation, the judgment terminal can pre-set a texture feature library containing known low-level image texture patterns that may be generated by older devices or specific environments. The judgment terminal compares the received local data characteristic information with this texture feature library; if the matching degree reaches a preset threshold, it determines that the model parameter update contains low-level image texture features. For example, the judgment terminal can train a classifier that takes local data characteristic information as input and outputs the probability of whether the client data contains specific texture features. The judgment terminal can be an independent analysis module containing pre-set logic or a trained classification model. The output of this module is the judgment result regarding whether the model parameter update contains specific texture features.

[0193] Finally, the adjustment end is configured to adjust the processing method of the model parameter update in the global model aggregation according to the judgment result, so as to integrate the update containing the underlying texture features of the image.

[0194] Once the judgment end determines that a client's model parameter update contains specific low-level image texture features, the adjustment end will no longer simply average this update with other updates. Instead, it will adjust its processing method in the global model aggregation based on the judgment result. One implementation approach is to assign a lower aggregation weight to updates identified as containing specific low-level image texture features to avoid excessive interference with the learning of the mainstream features of the global model. For example, the contribution of the client model update can be dynamically adjusted based on the strength or uniqueness of the texture features. Another approach is to design a dedicated aggregation strategy for these updates containing specific texture features, such as fusing them with a pre-trained texture adaptation module instead of directly aggregating them with the main model. The adjustment end can be an aggregation strategy management module that dynamically selects or modifies the parameters of the global model aggregation algorithm based on the judgment result provided by the judgment end. This module ensures the flexibility and adaptability of the aggregation process, thereby effectively integrating diverse client data features.

[0195] The implementation method described in this application demonstrates significant technological advancements in addressing the impact of varying client device underlying texture features on model aggregation performance in federated learning. Traditional AI model training systems in federated learning scenarios typically employ a uniform aggregation strategy, simply averaging or weighted averaging the model parameter updates from all clients, failing to effectively identify and process specific underlying texture features introduced by client devices or the environment. This approach results in limited generalization ability of the global model when faced with medical image data from diverse device backgrounds, and may even lead to decreased diagnostic accuracy.

[0196] In contrast, the system proposed in this application constructs an intelligent federated learning model training framework by introducing an input end, a judgment end, and an adjustment end. The input end is responsible for comprehensively receiving model parameter updates and local data characteristic information from the client, laying the foundation for subsequent intelligent processing. The judgment end can accurately identify whether the model parameter updates contain low-level image texture features introduced by the client based on these characteristic information, thus avoiding the drawback of traditional systems treating all updates the same. Furthermore, the adjustment end can dynamically adjust the processing method of model parameter updates in global model aggregation based on the judgment results, ensuring that updates containing unique texture features are effectively integrated, rather than diluted or misunderstood. As a result, the system proposed in this application can significantly improve the robustness and accuracy of the global model in diverse medical scenarios, effectively solving the challenges faced by traditional systems when processing heterogeneous client data, and providing a more refined and efficient solution for training cross-institutional medical diagnostic models.

[0197] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for training artificial intelligence models based on federated learning, used in federated learning scenarios to train cross-institutional medical diagnostic models, characterized in that, include: Receive model parameter updates from the client, as well as local data characteristic information describing the client, the local data characteristic information reflecting the underlying image texture introduced by the client during data acquisition; Based on the local data characteristic information, determine whether the model parameter update includes the underlying image texture features introduced by the client; Based on the judgment result, the processing method of the model parameter update in the global model aggregation is adjusted to integrate the update including the underlying texture features of the image; When the client's local model is being trained, the feature extraction layer inside the local model is guided to become sensitive to minute changes in the client's state. Extract a dynamic fingerprint representing the current data style of the client from the local model; The real-time environment and operation parameter vectors of the client that monitors and samples medical images are incorporated into the feature summary of the client. After the central aggregation server receives the model parameter updates, dynamic fingerprints, real-time environment and operation parameter vectors from each client, it uses a meta-aggregation network to calculate and adjust the processing method of each client model update in the global aggregation. Based on the adjustment factor output by the meta-aggregation network, the contribution of each client model update is adjusted and integrated into the global model. The meta-aggregation network is trained through meta-learning, with the goal of learning a function that predicts aggregation weights and learning rate adjustment factors based on the dynamic client context.

2. The method for training an artificial intelligence model based on federated learning according to claim 1, characterized in that, Extracting a dynamic fingerprint representing the current data style of the client from the local model includes: The background texture of the input image is reconstructed by using feature maps extracted from early convolutional layers by the local model and then processed by the decoder network. The loss function trained locally on the client is combined with the loss from the disease diagnosis task and the loss from the texture reconstruction task. The client extracts feature vectors from a specific layer in the local model that is responsible for extracting low-level features. The client performs global average pooling on the feature vector to obtain a fixed-dimensional feature vector, which serves as a dynamic fingerprint representing the current data style of the client.

3. The method for training an artificial intelligence model based on federated learning according to claim 1, characterized in that, The step of guiding the feature extraction layer within the client's local model to be sensitive to minute changes in the client's state during training includes: During the training process, the local model periodically applies transformations to the input image that simulate the background texture drift of the client. The parameters of the transformation are identified or predicted using the local model; Based on the identified or predicted transformation parameters, the feature extraction layer inside the local model is guided to become sensitive to subtle changes in the client's state.

4. The method for training an artificial intelligence model based on federated learning according to claim 3, characterized in that, During training, the local model periodically applies transformations to the input image that simulate client-side background texture shifts, including: Based on the influence of wear and tear of internal components, ambient temperature and humidity fluctuations, and imaging parameter settings on image background texture under different operating states of medical imaging clients, a parameterized texture drift generator is constructed. The texture drift generator is used to generate various combinations of transformation parameters for noise distribution patterns, weak stripe directions, contrast gradients, and geometric distortion. During each training cycle, the local model randomly selects a set of transformation parameters from the texture drift generator and performs real-time transformation on the input image of the current training batch according to the set of transformation parameters to generate an image simulating the background texture drift of the client. During the training process, the local model of the client analyzes the long-term running data of the client to identify the potential evolution path and key influencing factors of the client's background texture drift, and dynamically adjusts the parameter extraction strategy in the texture drift generator to prioritize generating simulated transformations that better match the current actual drift state of the client.

5. The method for training an artificial intelligence model based on federated learning according to claim 4, characterized in that, During the training process of the client's local model, based on the analysis of the client's long-term running data, the potential evolution path and key influencing factors of the client's background texture drift are identified, and the parameter extraction strategy in the texture drift generator is dynamically adjusted to prioritize generating simulated transformations that better match the current actual drift state of the client, including: The local model receives data from internal sensors on the client, environmental monitoring data, and operation log information. The local model performs feature extraction and pattern recognition on the received data to distinguish texture drift patterns caused by different factors such as wear and tear of the client, fluctuations in ambient temperature and humidity, and deviations in operating parameter settings. Based on the differentiated texture drift patterns, a multi-dimensional drift pattern representation is constructed, which is used to quantify the contribution of each factor to texture drift. The multi-dimensional drift pattern characterization allows for fine-tuning of the extraction probability of different transformation parameters in the texture drift generator, ensuring that the generated simulated transformation accurately reflects the complex drift state of the current client under the combined influence of multiple factors.

6. The method for training an artificial intelligence model based on federated learning according to claim 5, characterized in that, The step of finely adjusting the extraction probability of different transformation parameters in the texture drift generator through the multi-dimensional drift pattern characterization includes: The multi-dimensional drift pattern representation is received through the local model; The multi-dimensional drift pattern representation is used as input through a mapping network; The mapping network is used to analyze the correlation between the drift pattern representation and the actual texture drift in the long-term running data of the client, adjust the internal connection weights, and calculate and output the extraction probability of different transformation parameters in the texture drift generator in real time based on the adjusted internal connection weights.

7. The method for training an artificial intelligence model based on federated learning according to claim 6, characterized in that, The step of analyzing the correlation between drift pattern representations and actual texture drift in the long-term running data of the client through the mapping network and adjusting the internal connection weights includes: Receive the multi-dimensional drift pattern representation and actual texture drift data; Data quality assessment is performed on the multi-dimensional drift pattern representation and the actual texture drift data to identify noisy data, outlier data, and incomplete data; The noise data, the outlier data, and the incomplete data are processed. The processed multi-dimensional drift pattern representation and the actual texture drift data are input into the mapping network; The mapping network is used to analyze the correlation between the drift pattern representation and the actual texture drift, and the internal connection weights are adjusted.

8. The method for training an artificial intelligence model based on federated learning according to claim 6, characterized in that, The step of adjusting the internal connection weights by analyzing the correlation between drift pattern representations and actual texture drift in the long-term running data of the client through the mapping network includes: Receive the multi-dimensional drift pattern representation and actual texture drift data; The received multi-dimensional drift pattern representation and actual texture drift data are used to identify the client source and distinguish data from different medical institution clients. Based on each client source identified, a dedicated sub-mapping network is trained through the local model. The sub-mapping network is responsible for learning the correlation between the drift pattern representation of a specific client source and the actual texture drift. An attention allocation mechanism is introduced through the local model. This mechanism dynamically activates and weights the output of the corresponding sub-mapping network based on the client source represented by the current drift pattern being processed. The attention allocation mechanism analyzes the features of the current client source and generates a set of attention weights. Based on the attention weights, the prediction results of different sub-mapping networks are aggregated to obtain the final transformation parameter extraction probability; The internal connection weights are adjusted based on the final transformation parameter extraction probability.

9. A federated learning-based artificial intelligence model training system for training cross-institutional medical diagnostic models in a federated learning scenario, characterized in that, include: The input terminal is used to receive model parameter updates from the client, as well as local data characteristic information describing the client, which reflects the underlying image texture introduced by the client during data acquisition. The judgment end is used to determine, based on the local data characteristic information, whether the model parameter update includes the underlying image texture features introduced by the client; The adjustment end is used to adjust the processing method of the model parameter update in the global model aggregation according to the judgment result, so as to integrate the update including the underlying texture features of the image; When the client's local model is being trained, the feature extraction layer inside the local model is guided to become sensitive to minute changes in the client's state. Extract a dynamic fingerprint representing the current data style of the client from the local model; The real-time environment and operation parameter vectors of the client that monitors and samples medical images are incorporated into the feature summary of the client. After the central aggregation server receives the model parameter updates, dynamic fingerprints, real-time environment and operation parameter vectors from each client, it uses a meta-aggregation network to calculate and adjust the processing method of each client model update in the global aggregation. Based on the adjustment factor output by the meta-aggregation network, the contribution of each client model update is adjusted and integrated into the global model. The meta-aggregation network is trained through meta-learning, with the goal of learning a function that predicts aggregation weights and learning rate adjustment factors based on the dynamic client context.

Images