Federal learning method, medical data processing method, system and related device

Abstract

The invention provides a federal learning method, a medical data processing method, a medical data processing system and a related device. The method comprises the following steps: acquiring a prediction label carrying a medical sample identifier sent by each participant and a prediction deviation between the prediction label and an original real label; for each medical sample, under the condition that the medical sample is detected to be an abnormal medical sample according to the prediction labels of all participants and the prediction deviation, generating a target real label of the medical sample according to the prediction labels of all participants; and sending a target real label carrying the medical sample identifier to each participant, so that each participant updates a local medical large model according to the target real label corresponding to the medical sample identifier and the prediction label. According to the method, collaborative correction and global optimization of the abnormal medical samples are realized, the interference of the abnormal medical samples on model training is avoided, and the generalization ability and accuracy of the finally generated model are improved.

Description

Federated learning methods, medical data processing methods, systems and related devices Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a federated learning method, a medical data processing method, a system, and related devices. Background Technology

[0002] In the field of multicenter medical research, federated learning has become a key technology for training higher-performing global models by integrating data from different medical institutions while protecting data privacy and security. Vertical federated learning, in particular, can effectively combine and analyze different dimensions of data on the same patient population held by different institutions, thereby forming a more comprehensive diagnostic basis.

[0003] To achieve the above goals, a federated learning architecture with central coordination and multi-party training is typically employed. In practical applications, it is usually assumed that the sample labels provided by the data participants are completely accurate and reliable. However, in complex scenarios such as multi-center collaborative medical analysis, due to differences in diagnostic criteria among different institutions or physicians, varying levels of annotation experience, or operational oversights, sample labels often exhibit errors or inconsistencies. Therefore, when label noise exists, it not only interferes with the convergence direction of the local models of each participant but also affects the generalization ability and accuracy of the final generated global model. Summary of the Invention

[0004] This invention provides a federated learning method, a medical data processing method, a system, and related apparatus to address the problem that in current federated learning, the presence of label noise not only interferes with the convergence direction of the local models of each participant but also affects the generalization ability and accuracy of the final generated global model.

[0005] This invention provides a federated learning method applied to a central service provider in a federated learning system. The central service provider communicates with at least two participating parties, and the at least two participating parties include at least one medical sample with the same sample but different feature dimensions. The method includes: acquiring a predicted label carrying a medical sample identifier sent by each participating party, and the prediction deviation between the predicted label and the original true label; for each medical sample, if the medical sample is detected as an abnormal medical sample based on the predicted labels of all participating parties and the prediction deviation, generating a target true label for the medical sample based on the predicted labels of all participating parties; and sending the target true label carrying the medical sample identifier to each participating party, so that each participating party updates its local large medical model based on the target true label and the predicted label.

[0006] According to an embodiment of the present invention, a federated learning method is provided in which abnormal medical samples are determined by: for each medical sample, determining the prediction variance based on the prediction labels of all participants; and determining the average prediction deviation based on the prediction deviations of all participants; determining the global deviation based on the prediction variance and the average prediction deviation; and determining the medical sample as an abnormal medical sample if the global deviation is greater than a threshold.

[0007] According to an embodiment of the present invention, a federated learning method is provided, wherein generating a target true label for a medical sample based on the predicted labels of all participants includes: determining the weight corresponding to each participant; and generating the target true label for the medical sample based on the weight corresponding to each participant and the predicted label.

[0008] This invention provides a federated learning method, wherein the weight of each participant is obtained by: determining the contribution of each participant in at least one historical federated training; determining the confidence level of the predicted label corresponding to each participant in the current federated training; and determining the weight of each participant based on the contribution and the confidence level.

[0009] According to an embodiment of the present invention, a federated learning method is provided, the method further comprising: for each medical sample, if the medical sample is detected as a normal medical sample based on the predicted labels and the prediction bias of all participants, sending a normal medical sample notification carrying a medical sample identifier to each participant, so as to notify each participant to update its local medical big model based on the original true labels and the predicted labels.

[0010] According to an embodiment of the present invention, a federated learning method is provided, the method further comprising: before performing collaborative training, sending a container image and a deployment description file to each participant, wherein the container image includes a training program and required dependency libraries for performing collaborative training, and the deployment description file includes a network resource configuration policy and a storage resource configuration policy for the container image; and before performing collaborative training, configuring a directed acyclic graph workflow engine, wherein the directed acyclic graph workflow engine is used to split the federated learning task into multiple sequentially executed subtasks, and triggering each participant to execute multiple subtasks sequentially.

[0011] This invention also provides a federated learning method applied to participants in a federated learning system. Each participant communicates with at least one other participant via a central service provider. The participants and the other participants share at least one medical sample with the same sample characteristics but different feature dimensions. The method includes: for each medical sample, inputting the medical sample into a local medical big model of the participant to obtain a predicted label output by the medical big model, and acquiring the prediction deviation between the predicted label and the corresponding original true label; sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; receiving the target true label of the medical sample returned by the central service provider; wherein the target true label is generated by the central service provider based on the predicted labels of all participants when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants; and updating the local medical big model of the participant based on the target true label and the predicted label.

[0012] According to an embodiment of the present invention, a federated learning method is provided, the method further comprising: receiving a normal medical sample notification of the medical sample returned by a central service provider; wherein the normal medical sample notification is sent by the central service provider when the medical sample is detected as a normal medical sample based on the predicted labels and the prediction bias of all participants; and updating the large medical model locally of each participant based on the predicted labels and the original true labels.

[0013] According to an embodiment of the present invention, a federated learning method further includes, before sending the prediction label carrying the medical sample identifier and the prediction bias to the central service provider, performing differential privacy noise processing or homomorphic encryption processing on the prediction label carrying the medical sample identifier and the prediction bias.

[0014] This invention also provides a medical data processing method based on federated learning, applied to participants in a federated learning system. The participants are connected to at least one other participant via a central service provider, and the participants and at least one other participant include at least one medical sample with the same sample but different feature dimensions. The method includes: acquiring medical data to be processed; inputting the medical data to be processed into a large medical model on the participant's local machine, and obtaining the diagnostic results output by the large medical model; the large medical model is obtained through any of the federated learning methods described above.

[0015] This invention also provides a federated learning device applied to a central service provider in a federated learning system. The central service provider has a communication connection with at least two participating parties, and the at least two participating parties include at least one medical sample with the same sample but different sample feature dimensions. The device includes: a data acquisition module, used to acquire a predicted label carrying a medical sample identifier sent by each participating party, and the prediction deviation between the predicted label and the original true label; a data correction module, used to generate a target true label for each medical sample based on the predicted labels of all participating parties if the medical sample is detected as an abnormal medical sample based on the predicted labels of all participating parties and the prediction deviation; and a first data transmission module, used to send the target true label carrying the medical sample identifier to each participating party, so that each participating party updates its local medical big model based on the target true label and the predicted label.

[0016] This invention also provides a federated learning device applied to participants in a federated learning system. Each participant communicates with at least one other participant via a central service provider. The participants and the at least one other participant share at least one medical sample with the same sample but different feature dimensions. The device includes: a prediction module, used to input each medical sample into a local medical big model of the participant, obtain a predicted label output by the medical big model, and acquire the prediction deviation between the predicted label and the corresponding original true label; a second data sending module, used to send the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; a data receiving module, used to receive the target true label of the medical sample returned by the central service provider; wherein the target true label is generated by the central service provider based on the predicted labels of all participants when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants; and a parameter update module, used to update the local medical big model of the participant based on the target true label.

[0017] This invention also provides a federated learning system, including a central service provider and at least two participating parties, wherein the central service provider is configured to implement any of the federated learning methods described above, and the participating parties are configured to implement any of the federated learning methods described above or any of the medical data processing methods described above.

[0018] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements either the federated learning method or the medical data processing method described above.

[0019] This invention also provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements either the federated learning method or the medical data processing method described above.

[0020] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements either the federated learning method or the medical data processing method described above.

[0021] The federated learning method provided in this invention collects the predicted labels and corresponding prediction biases of each participant carrying medical sample identifiers to detect abnormal medical samples. After detecting abnormal medical samples, it generates globally corrected target true labels based on the predicted labels of all participants corresponding to the abnormal medical samples and synchronizes them to all participants to update the local medical big model. This achieves collaborative correction and global optimization of abnormal medical samples, avoids interference from abnormal medical samples to model training, and thus improves the generalization ability and accuracy of the final generated global model. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 is a schematic diagram of the structure of the federated learning system provided in an embodiment of the present invention.

[0024] Figure 2 is one of the flowcharts of the federated learning method provided in the embodiment of the present invention.

[0025] Figure 3 is a second schematic flowchart of the federated learning method provided in an embodiment of the present invention.

[0026] Figure 4 is one of the structural schematic diagrams of the federated learning device provided in the embodiments of the present invention.

[0027] Figure 5 is a second schematic diagram of the structure of the federated learning device provided in an embodiment of the present invention.

[0028] Figure 6 is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] Federated learning, as a distributed collaborative modeling paradigm, originated from the need for data security and multi-party collaboration in the field of artificial intelligence. Its core idea is to collaboratively build a global model through local training and encrypted parameter transmission, while keeping data within the local environment of different institutions or terminals. This mechanism effectively alleviates the problems of data silos and privacy leaks, and has therefore rapidly gained attention and application in fields such as finance and healthcare.

[0031] Most existing federated learning practices adopt a typical architecture of centralized coordination and multi-party training. In this architecture, the central coordinator is responsible for model initialization and parameter aggregation, while each participating party independently completes local training and uploads updated information.

[0032] Vertical federated learning is highly compatible with the medical field. Hospitals often serve partially overlapping patient groups, but the dimensions of the medical data they collect differ significantly. For example, some institutions focus on imaging information, while others focus more on pathological examinations or clinical records. Through vertical federated learning, cross-institutional, multimodal feature joint analysis can be achieved without exposing the original images and medical records, thereby forming a more comprehensive diagnostic basis.

[0033] In practical applications, it is usually assumed that the sample labels provided by the data participants are completely accurate and reliable. However, in complex scenarios such as multi-center collaborative medical analysis, due to differences in diagnostic criteria among different institutions or physicians, varying levels of labeling experience, or operational oversights, sample labels often become incorrectly labeled or inconsistent. Therefore, when label noise exists, it not only interferes with the convergence direction of the local models of each participant, but also affects the generalization ability and accuracy of the final generated global model.

[0034] Based on this, this invention proposes a federated learning method. Specifically, it proposes a federated learning method for multi-institutional federated learning systems. Referring to Figure 1, the system integrates modules such as data governance, cross-hospital heterogeneous adaptation, federated training scheduling, label anomaly correction, model integration, and visual auditing. It constructs an intelligent anomaly detection system with a closed-loop process. Through containerized deployment, microservice decoupling, and trusted computing environment assurance, it achieves secure, reliable, and auditable training and model deployment of cross-hospital image data.

[0035] Specifically, the federated learning system proposed in this embodiment of the invention typically includes a central service provider and at least two participating parties. The central service provider can be one or a group of cloud servers or physical server clusters, and its main role is to coordinate key tasks in the entire federated learning process, but it does not access the raw medical data of the participating parties. Participating parties typically refer to institutions that possess local data and computing resources. These participating parties collaborate on training, and they possess different dimensions of features for the same medical sample.

[0036] Figure 2 is a flowchart of one of the federated learning methods provided in the embodiments of the present invention. The federated learning method is applied to the central service provider in the federated learning system. The central service provider has communication connections with at least two participants, and the at least two participants include at least one medical sample with the same sample but different sample feature dimensions. As shown in Figure 2, the method includes the following steps 210, 220 and 230.

[0037] Step 210: Obtain the predicted label carrying the medical sample identifier sent by each participant, and the prediction deviation between the predicted label and the original true label.

[0038] Specifically, in each iteration of federated learning, each participant uses its locally trained medical big model to perform local calculations on the medical samples it holds. After completing the local calculations, each participant sends its calculation results to the central service provider. These results include at least: Medical sample identifier: used to uniquely identify each medical sample, ensuring that the central service provider can correctly associate calculation results from different participants with the same medical sample; Prediction label: the prediction result given by the participant's local medical big model for the corresponding medical sample; it should be noted that the prediction label and the original true label in this embodiment can be classification labels, such as disease type, or regression labels, such as risk score or indicator prediction value, which can be flexibly set according to needs; Prediction bias: prediction bias is used to measure the difference or error between the participant's prediction label and its locally stored original true label; here, the original true label refers to the label labeled by doctors or other professionals for the medical samples before training begins. The calculation method of prediction bias can be selected according to the specific task. For example, for classification tasks, cross-entropy loss can be used as a bias measure; for regression tasks, mean squared error can be used as a bias measure, etc., without limitation.

[0039] Furthermore, to ensure the security of data transmission, the participants can send encrypted prediction labels carrying medical sample identifiers and prediction deviations using a negotiated data encryption method. After receiving the corresponding encrypted data, the central service provider will decrypt it to parse out the prediction labels carrying medical sample identifiers sent by each participant, as well as the prediction deviations between the prediction labels and the original true labels.

[0040] Step 220: For each medical sample, if the medical sample is detected as an abnormal medical sample based on the predicted labels of all participants and the predicted bias, generate the target true label of the medical sample based on the predicted labels of all participants.

[0041] After collecting the predicted labels and prediction biases of all medical samples sent by all participants, the central service provider will, based on the medical sample identifier of each medical sample, traverse the predicted labels and prediction biases of each participant under the same medical sample identifier in order to detect abnormal medical samples.

[0042] Here, "abnormal medical samples" refers to medical samples whose original true labels are highly suspected of being incorrect. It should be noted that in the field of medical imaging, labeling errors can occur due to factors such as differences in physician experience, inconsistencies in image modalities, or biases in annotation tools. Therefore, directly using these incorrect labels for training will introduce noise pollution into the model, leading to a decline in the performance of the trained model.

[0043] Specifically, for each medical sample identified by a medical sample identifier, the central service provider will perform abnormal medical sample detection based on the consensus reached by all participating parties. In other words, if the prediction results of the local medical models of multiple independent participating parties for the same medical sample are highly inconsistent, or if the prediction bias of the local medical models of all participating parties on the medical sample is generally large, then the central service provider will determine that the original true label of the sample may be incorrect and classify it as an abnormal medical sample.

[0044] In one example, the variance of the predicted labels for all participants can be calculated, and an anomaly is identified when the variance exceeds a preset threshold. The average prediction bias of all participants can also be calculated, and an anomaly is identified when the average exceeds a threshold. Furthermore, a comprehensive set of pre-defined indicators can be used, combining the consistency of the predicted labels and the magnitude of the prediction bias, to make a joint judgment; the specific criteria can be flexibly selected as needed.

[0045] Once a medical sample is identified as an abnormal medical sample, and the original true label representing that medical sample may be incorrect, the central service provider generates a new, reliable target true label based on the predicted labels of all participants to replace the unreliable original true label, thereby achieving dynamic correction of label noise and robust training.

[0046] In one example, a majority vote can be used, where the predicted label that is the same for most participants is taken as the target true label. Alternatively, a weighted average can be used, where the predicted labels of all participants are weighted and the result is taken as the target true label. The specific method can be flexibly configured according to needs, and there are no restrictions on this.

[0047] Step 230: Send the target real label carrying the medical sample identifier to each participant, so that each participant updates its local medical big model based on the target real label and the predicted label.

[0048] After completing the above-mentioned abnormal medical sample detection and generating a new target truth label for the detected abnormal medical sample, the central service provider will distribute the processing results back to all participating parties. Specifically, for a medical sample determined to be abnormal, the central service provider will send the newly generated target truth label along with its corresponding medical sample identifier to all participating parties.

[0049] Upon receiving this information, during the current iteration, participants will no longer use the locally stored original real labels for these abnormal medical samples. Instead, they will use the target real labels issued by the central service provider as supervision signals to calculate the loss and update the model parameters. Specifically, the loss and model parameters can be calculated based on the following formula: ;in, Indicates the first The participating party in the first Model parameters at +1 iteration; Indicates the first The participating party in the first Model parameters at the next iteration; This represents the learning rate, used to control the step size for parameter updates; This represents the loss function, such as cross-entropy, mean squared error, etc. Indicates the first Medical sample x from each participating party in the first... The prediction results at the next iteration; This indicates that when medical sample x is an abnormal medical sample, the target true label generated by the central coordinator is used to replace the original true label.

[0050] It should be understood that through the above closed-loop process, it is possible to dynamically identify and correct label noise that may exist in the dataset during the collaborative training process of federated learning, without human intervention or the need for the central service provider to access any raw medical data.

[0051] The federated learning method provided in this invention collects the predicted labels and corresponding prediction biases of each participant carrying medical sample identifiers to detect abnormal medical samples. After detecting abnormal medical samples, it generates globally corrected target true labels based on the predicted labels of all participants corresponding to the abnormal medical samples and synchronizes them to all participants to update the local medical big model. This achieves collaborative correction and global optimization of abnormal medical samples, avoids interference from abnormal medical samples to model training, and thus improves the generalization ability and accuracy of the final generated global model.

[0052] In some embodiments, the abnormal medical samples are determined by: for each medical sample, determining the prediction variance based on the prediction labels of all participants; and determining the average prediction deviation based on the prediction deviations of all participants; determining the global deviation based on the prediction variance and the average prediction deviation; and determining the medical sample as the abnormal medical sample if the global deviation is greater than a threshold.

[0053] Here, prediction variance is used to quantify the dispersion of predicted labels for the same medical sample among all participants. A larger prediction variance indicates greater differences in predicted labels among different participants for the same medical sample; a smaller prediction variance indicates smaller differences in predicted labels among participants.

[0054] It should be noted that in practical applications, the prediction labels may be classification labels or regression labels. Therefore, in this embodiment, before calculating the prediction variance, the prediction labels of all participants are uniformly mapped to a quantifiable numerical form.

[0055] Specifically, for regression labels, the predicted values ​​from the original predicted labels reported by the participants are directly used to calculate the prediction variance. For classification labels, they are uniformly mapped to quantifiable numerical forms according to their type. For example, for binary labels, they are directly mapped to binary values, such as positive = 1 and negative = 0. For multi-class labels, one-hot encoding or label sequence mapping is used to map the corresponding classification labels to the corresponding numerical values.

[0056] After obtaining quantifiable predictive labels, for each medical sample corresponding to a medical sample identifier, the mean of the predictive labels of all participants is calculated. Then, the sum of squared deviations between the mean of the predictive labels of each participant and the preset labels is calculated. Finally, the prediction variance is calculated based on the sum of squared deviations of all participants.

[0057] Similarly, for each medical sample corresponding to a medical sample identifier, after collecting the prediction biases reported by all participants, the average prediction bias of that sample is calculated by arithmetic mean or weighted average.

[0058] Finally, the prediction variance and average prediction bias are fused into a single quantified value, namely the global deviation, using a preset algorithm. In one example, the global deviation can be calculated using the following formula: ;in, d represents the global deviation of medical sample x; i Indicates the first Prediction bias of medical sample x from each participating party, The predicted label for the medical sample x of the kth participant; This represents the variance between the predicted labels of different participants. Refers to the conventional coefficient.

[0059] After calculating the global deviation of each medical sample corresponding to its identifier, the global deviation is compared with a pre-set threshold. This threshold can be a fixed value or a dynamic value; for example, it could be a specific multiple of the average global deviation of all medical samples in the current batch. When the calculated global deviation of a medical sample exceeds this threshold, the central service provider classifies it as an abnormal medical sample and triggers the subsequent label generation and correction process.

[0060] The federated learning method provided in this invention provides an objective detection method for abnormal medical samples by calculating the global deviation of the sample based on the prediction variance and the average prediction deviation. It does not rely on information from a single dimension, but integrates the prediction results of all participants, making the identification of label noise more accurate.

[0061] In some embodiments, generating the target true label of the medical sample based on the predicted labels of all participants includes: determining the weight corresponding to each participant; and generating the target true label of the medical sample based on the weight corresponding to each participant and the predicted label.

[0062] In this embodiment, when a medical sample is determined to be an abnormal medical sample, its original true label is no longer reliable. At this time, it is necessary to jointly generate a new target true label based on the predicted labels of all participants.

[0063] Specifically, when aggregating the predicted labels of all participants, each participant is first assigned a weight. This weight represents the credibility of that participant in the current prediction. The weight can be determined based on a variety of factors. For example, it can be based on a participant's historical performance, the performance of its local model on a validation set, or certain real-time metrics of its current prediction.

[0064] After determining the weight of each participant, the target true label is generated by weighted aggregation of the predicted labels of each participant. Similarly, before calculating the target true label, the predicted labels of all participants are uniformly mapped into a quantifiable numerical form.

[0065] Specifically, for regression labels, the predicted values ​​from the original predicted labels reported by the participants are directly used to calculate the target true label. For classification labels, they are uniformly mapped to quantifiable numerical forms according to their type. For example, for binary labels, they are directly mapped to binary values, such as positive = 1 and negative = 0. For multi-class labels, one-hot encoding or label sequence mapping is used to map the corresponding classification labels to their corresponding values. Then, a weighted aggregation method is used to calculate the final result, which is then mapped back to the corresponding classification label through a pre-defined mapping method to obtain the target true label.

[0066] In some embodiments, the weight of each participant is obtained by: determining the contribution of each participant in at least one historical federated training; determining the confidence level of the predicted label corresponding to each participant in the current federated training; and determining the weight of each participant based on the contribution and the confidence level.

[0067] Here, contribution refers to the quantified value of a participant's contribution to the performance improvement of the global model or each participant's local model in at least one historical federated training round. In practical applications, it can be determined based on factors such as the quality of the participant's local data, the training effect of the participant's local medical model, the amount of data contributed, and training stability in the historical federated training rounds, without any restrictions. For example, the greater the performance improvement of the global model or the collaborative local medical model provided by a participant in the historical federated training rounds, the higher its contribution.

[0068] It should be noted that confidence level refers to the quantitative value of the reliability of the predicted labels generated by the participants for the current medical sample to be processed in the current federated training round. In practical applications, it can be determined based on relevant factors such as the prediction bias of the participants in the current round. For example, the smaller the prediction bias, the closer the predicted label is to the true label, and the higher the confidence level.

[0069] After determining the contribution and confidence level of each participant, the final weights are generated through a preset algorithm, such as by weighted summation or direct multiplication.

[0070] In one example, the corresponding weights can be calculated using the following formula: ;in, Indicates the first Confidence of predicted labels for medical samples x from each participant; Indicates the first The credibility of each participating party, based on the first... The contribution of each participant in at least one historical federal training program was determined. These are pre-set conventional coefficients used to control the smoothness of the weight distribution; when When the confidence level approaches 0, the weights tend to be uniformly distributed, meaning the difference in confidence levels is ignored; when... Approaching In this case, the participant with the highest confidence and the highest credibility will have a greater weight.

[0071] The federated learning method provided in this invention assigns weights to each participant that are appropriate to their historical contributions and current prediction reliability. Based on a weighted fusion rule, it aggregates the prediction labels of all participants to generate target prediction labels for abnormal medical samples, thus achieving both accuracy and fairness in the generation of target prediction labels.

[0072] In some embodiments, the method further includes: for each medical sample, if the medical sample is detected as a normal medical sample based on the predicted labels and the predicted bias of all participants, sending a normal medical sample notification carrying a medical sample identifier to each participant, so as to notify each participant to update its local medical big model based on the original true labels and the predicted labels.

[0073] Specifically, when the central service provider tests a medical sample, for example, if it calculates that the global deviation of the sample is less than or equal to a preset threshold according to the method in the above embodiment, it is classified as a normal medical sample. That is, the original true label of the medical sample is reliable and does not require correction.

[0074] Subsequently, the central service provider needs to notify all participants of this determination. In practice, the central service provider sends a list to each participant, clearly indicating which medical sample identifiers correspond to medical samples that have been determined to be normal medical samples. Alternatively, the central service provider can send only a list of abnormal medical samples and their corresponding target true labels. For any medical sample not appearing in this abnormal list, the participant can assume it is a normal medical sample.

[0075] Upon receiving the notification, the participating parties, for normal medical samples, follow the standard federated learning training process, using the original ground truth labels corresponding to the medical sample identifiers stored locally to calculate the loss and backpropagation gradients, and update their local large-scale medical model. Specifically, the loss and model parameters can be calculated and updated based on the following formula: ;in, Indicates the first The participating party in the first Model parameters at +1 iteration; Indicates the first The participating party in the first Model parameters at the next iteration; This represents the learning rate, used to control the step size for parameter updates; This represents the loss function, such as cross-entropy, mean squared error, etc. Indicates the first Medical sample x from each participating party in the first... The prediction results at the next iteration; This indicates that when medical sample x is an abnormal medical sample, the target true label generated by the central coordinator is used to replace the original true label.

[0076] The federated learning method provided in this embodiment of the invention, through the above approach, not only retains the full utilization of high-quality labeled information, but also avoids unnecessary computational overhead and communication burden, thus ensuring the efficiency and stability of the entire federated learning method.

[0077] In some embodiments, the method further includes: before performing collaborative training, sending a container image and a deployment description file to each participant, wherein the container image includes a training program and required dependency libraries for performing collaborative training, and the deployment description file includes a network resource configuration policy and a storage resource configuration policy for the container image; and before performing collaborative training, configuring a directed acyclic graph workflow engine, wherein the directed acyclic graph workflow engine is used to split the federated learning task into multiple sequentially executed subtasks, and trigger each participant to execute multiple subtasks sequentially.

[0078] In this embodiment, in order to solve the problems of heterogeneous computing environments and inconsistent dependency library versions among different participants, this solution adopts containerization technology, such as packaging each component into Docker containers for deployment and configuring elastic scaling strategies to ensure that it can automatically expand or shrink according to the computing load.

[0079] Specifically, the central service provider will package all the objects required for collaborative training, including training programs such as local model training and update code written in Python, all required dependent libraries such as specific versions of NumPy, and operating system environments, into a standardized container image, and send the container image to all participants so that they can deploy all the objects required for collaborative training locally based on the container image.

[0080] In addition, the central service provider will send a deployment description file to all participants. This deployment description file defines in detail the container's operating rules, such as network resource configuration policies, including allowed ports and network bandwidth limits, as well as storage resource configuration policies, such as the mount path of local data volumes and the size of temporary storage space. It may also include allocation policies for computing resources such as CPU, memory, and GPU, which can be flexibly configured according to requirements.

[0081] Based on this, each participant only needs to install the container engine and then use the received container image and deployment description file to deploy a completely consistent training environment.

[0082] It should be understood that each iteration of federated learning involves multiple tasks with sequential dependencies, such as downloading the model, local computation, uploading results, waiting for aggregation, and downloading new labels. Manually managing these tasks is tedious and error-prone. Therefore, in this embodiment, a directed acyclic graph (DAG) workflow engine is pre-configured before collaborative training. This engine allows a complete federated learning task to be broken down into a series of independent subtasks. These subtasks are then organized into a DAG according to their logical dependencies—for example, all participants must complete their local computations before aggregation can begin. After training begins, the DAG workflow engine automatically and sequentially triggers each participant to execute the subtasks in the DAG, monitors their execution status, handles retry failures, and ensures the entire process proceeds smoothly according to predefined logic.

[0083] Furthermore, before conducting collaborative training, the corresponding national cryptographic algorithms are negotiated in advance. Based on this, after the training begins, the participants and the central coordinator can communicate encryptedly according to the negotiated national cryptographic algorithms.

[0084] In addition, before conducting collaborative training, the central service provider in this embodiment can also deploy a monitoring and log collection module to monitor node status, resource utilization and task progress in real time, providing support for operation and maintenance and anomaly handling.

[0085] The federated learning method provided in this invention pre-distributes container images containing training programs and dependency libraries, along with resource configuration strategy files. It then deploys a directed acyclic graph workflow engine to split federated learning tasks and triggers participants to execute sub-tasks sequentially. This achieves standardized deployment, process-oriented scheduling, and efficient collaboration in federated learning collaborative training, ensuring the consistency and orderliness of the federated training process.

[0086] This invention also provides a medical data processing method. Figure 3 is a second flowchart of the medical data processing method provided by this invention. It is applied to a participant in a federated learning system. The participant is connected to at least one other participant through a central service provider. The participant and at least one other participant include at least one medical sample with the same sample but different sample feature dimensions. As shown in Figure 3, the method includes the following steps 310 and 340.

[0087] Step 310: For each medical sample, input the medical sample into the local medical big model of the participant to obtain the predicted label output by the medical big model, and obtain the prediction deviation between the predicted label and the corresponding original true label.

[0088] In a federated learning training iteration, participants first process their local medical samples. Here, a medical sample refers to preprocessed feature data used as model input; for example, for a hospital with image data, the feature data for a medical sample could be a feature matrix extracted from raw CT images. Each medical sample is distinguished by a unique medical sample identifier to ensure accurate correspondence during multi-party interactions.

[0089] Specifically, medical samples are input into the participant's local medical model, which then performs forward propagation calculations and outputs a prediction label. This prediction label can be a classification label, such as a disease type, or a regression label, such as a risk score or predicted indicator value; the specific label can be flexibly set according to needs.

[0090] Simultaneously, each participant locally stores the original ground truth labels corresponding to the medical samples. These original ground truth labels are the annotation results of the medical samples before training begins. Subsequently, each participant calculates the prediction bias between the predicted labels of its local medical big data model and these original ground truth labels. This prediction bias is used to measure the difference or error between the participant's predicted labels and the locally stored original ground truth labels. For example, for classification tasks, the prediction bias can be calculated as the cross-entropy loss between the predicted labels and the original ground truth labels; for regression tasks, the prediction bias can be calculated as the mean squared error between the predicted labels and the original ground truth labels, and so on. There are no restrictions on this.

[0091] Step 320: Send the prediction label carrying the medical sample identifier and the prediction bias to the central service provider.

[0092] After completing local calculations, the participants upload their results to the central service provider for subsequent aggregation and analysis. Specifically, the participants package the medical sample identifier, the corresponding prediction label, and the calculated prediction bias and send them to the central service provider.

[0093] Step 330: Receive the target true label of the medical sample returned by the central service provider; wherein, the target true label is generated by the central service provider based on the predicted labels of all participants and the prediction deviation when the medical sample is detected as an abnormal medical sample.

[0094] After the central service provider collects the predicted labels and prediction biases of all medical samples sent by all participants, it iterates through the predicted labels and prediction biases of each participant under the same medical sample identifier to detect abnormal medical samples.

[0095] Here, "abnormal medical samples" refers to medical samples whose original true labels are highly suspected of being incorrect. It should be noted that in the field of medical imaging, labeling errors frequently occur due to factors such as differences in physician experience, inconsistencies in image modalities, or biases in annotation tools. Therefore, directly using these incorrect labels for training will introduce noise pollution into the model, leading to a decline in model performance.

[0096] Specifically, for each medical sample identified by a medical sample identifier, the central service provider will perform abnormal medical sample detection based on the consensus reached by all participating parties. In other words, if the prediction results of the local medical models of multiple independent participating parties for the same medical sample are highly inconsistent, or if the prediction bias of the local medical models of all participating parties on the medical sample is generally large, then the central service provider will determine that the original true label of the sample may be incorrect and classify it as an abnormal medical sample.

[0097] In one example, the variance of the predicted labels for all participants can be calculated, and an anomaly is identified when the variance exceeds a preset threshold. The average prediction bias of all participants can also be calculated, and an anomaly is identified when the average exceeds a threshold. Furthermore, a comprehensive set of pre-defined indicators can be used, combining the consistency of the predicted labels and the magnitude of the prediction bias, to make a joint judgment; the specific criteria can be flexibly selected as needed.

[0098] Once a medical sample is identified as an abnormal medical sample, the central service provider needs to generate a target truth label for it. It should be noted that the target truth label is a new label considered more reliable than the original truth label, generated based on the predicted labels from all participating models.

[0099] In one example, a majority vote can be used, with the predicted label from the majority of participants becoming the target true label. Alternatively, a weighted average can be used, where the predicted probability distribution vectors of all participants' predicted labels are weighted to obtain a target true label.

[0100] Here, the detection method for abnormal medical samples and the generation method for the target true label in this embodiment are the same as in the above embodiments, and will not be described in detail here.

[0101] Step 340: Update the local medical big model of the participant based on the target real label.

[0102] Upon receiving this information, in the next iteration, for these abnormal medical samples, the participants will no longer use their locally stored original real labels, but instead will use the target real labels issued by the central service provider as supervision signals to calculate the loss and update the model parameters. For example, the loss and model parameters can be calculated and updated based on the following formula: ;in, Indicates the first The participating party in the first Model parameters at +1 iteration; Indicates the first The participating party in the first Model parameters at the next iteration; This represents the learning rate, used to control the step size for parameter updates; This represents the loss function, such as cross-entropy, mean squared error, etc. Indicates the first Medical sample x from each participating party in the first... The prediction results at the next iteration; This indicates that when medical sample x is an abnormal medical sample, the target true label generated by the central coordinator is used to replace the original true label.

[0103] The federated learning method provided in this invention collects the predicted labels and corresponding prediction biases of each participant carrying medical sample identifiers to detect abnormal medical samples. After detecting abnormal medical samples, it generates globally corrected target true labels based on the predicted labels of all participants corresponding to the abnormal medical samples and synchronizes them to all participants to update the local medical big model. This achieves collaborative correction and global optimization of abnormal medical samples, avoids interference from abnormal medical samples to model training, and thus improves the generalization ability and accuracy of the final generated global model.

[0104] In some embodiments, the method further includes: receiving a normal medical sample notification corresponding to the medical sample returned by a central service provider; wherein the normal medical sample notification is sent when the central service provider detects that the medical sample is a normal medical sample based on the predicted labels and the predicted bias of all participants; and updating the local medical big model of the participants based on the predicted labels and the original true labels.

[0105] Specifically, when the central service provider tests a medical sample, for example, if it calculates that the global deviation of the sample is less than or equal to a preset threshold according to the method in the above embodiment, it is classified as a normal medical sample. That is, the original true label of the medical sample is reliable and does not require correction.

[0106] Subsequently, the central service provider needs to notify all participants of this determination. In practice, the central service provider sends a list to each participant, clearly indicating which medical sample identifiers correspond to medical samples that have been determined to be normal medical samples. Alternatively, the central service provider can send only a list of abnormal medical samples and their corresponding target true labels. For any medical sample not appearing in this abnormal list, the participant can assume it is a normal medical sample.

[0107] Upon receiving the notification, the participants, for normal medical samples, follow the standard federated learning training process, using the original real labels corresponding to the medical sample identifiers stored locally to calculate the loss and backpropagation gradients, and update their local large medical model.

[0108] The federated learning method provided in this embodiment of the invention, through the above approach, not only retains the full utilization of high-quality labeled information, but also avoids unnecessary computational overhead and communication burden, thus ensuring the efficiency and stability of the entire federated learning method.

[0109] In some embodiments, the medical sample and the corresponding original real label are constructed by: acquiring original medical data and the corresponding original medical label; standardizing and enhancing the original medical data to obtain the medical sample; and converting the original medical label into a standardized medical label according to a label mapping table to serve as the corresponding original real label of the medical sample.

[0110] Here, raw medical data refers to the most original medical data obtained by the participants from their business systems, such as patient imaging data, laboratory indicators, medical record texts, physiological signals, etc. Similarly, raw medical labels refer to the most original medical labels obtained by the participants from their business systems based on clinical diagnosis, such as disease categories and risk levels.

[0111] It should be understood that the original medical data from different participants may have differences in collection equipment, data dimensions, data distribution, and data volume. Therefore, in this embodiment, the original medical data is standardized to convert it into standard data with a unified format and scale. Here, the standardization process includes, but is not limited to, format standardization and scale standardization.

[0112] Furthermore, in order to improve the generalization ability of the model, this embodiment also performs data augmentation processing on the standardized raw medical data. For example, for numerical raw medical data, noise can be randomly added, missing values ​​can be filled in, etc.; for image-type raw medical data, rotation, flipping, cropping, brightness adjustment, etc. can be performed; and for text-type raw medical data, synonym replacement, sentence transformation, etc. can be performed.

[0113] After the above processing, the high-quality data obtained can be used as medical samples and directly used for training the local medical large-scale model of the participating parties.

[0114] It should be noted that the original medical labels from different participants may have inconsistent expressions but the same semantic meaning. For example, Hospital A might use "positive" or "negative" to represent the diagnosis, Hospital B might use "1" or "0," and Hospital C might use "confirmed" or "not confirmed." Inconsistency in these labels will prevent subsequent federated learning from proceeding. Therefore, in this embodiment, the original medical labels are converted into standardized medical labels based on a label mapping table to serve as the corresponding original, true labels for the medical samples.

[0115] Here, the label mapping table is a set of standardized correspondence rules predefined by the central service provider of the federated learning system to unify the original medical label formats of all participants. It is sent synchronously by the central service provider to all participants before federated learning to ensure the consistency of label conversion.

[0116] Here, standardized medical labels refer to uniformly formatted labels obtained after conversion according to a label mapping table, typically numerical or standardized text labels. Label standardization ensures that all participants consistently represent the true results of the same medical sample, avoiding subsequent collaborative training failures due to differences in label formats.

[0117] In some embodiments, before sending the prediction label carrying the medical sample identifier and the prediction bias to the central service provider, the method further includes: performing differential privacy noise processing or homomorphic encryption processing on the prediction label carrying the medical sample identifier and the prediction bias.

[0118] In this embodiment, the calculation results to be uploaded are subjected to privacy protection processing to prevent the central service provider or other potential attackers from inferring sensitive information about the participants' local original data from these calculation results. Specifically, at least one of the following processing methods can be adopted: Differential privacy noise processing: The participants add calculated random noise to the prediction labels and prediction biases to be sent.

[0119] Homomorphic encryption: Participants use a homomorphic public key to encrypt their prediction labels and prediction biases. Homomorphic encryption allows the central service provider to perform specific mathematical operations directly on the ciphertext without decryption. For example, the central service provider can calculate the encrypted sum or encrypted average of all participants' prediction biases without knowing the specific prediction bias of any single participant.

[0120] The federated learning method provided in this invention ensures that the original data of the participants is not leaked by performing differential privacy processing or homomorphic encryption processing before the participants upload data.

[0121] This invention also provides a medical data processing method based on federated learning, applied to participants in a federated learning system. These participants are connected to at least one other participant via a central service provider, and the participants and the other participant share at least one medical sample with the same sample but different feature dimensions. Specifically, the method includes the following steps: acquiring medical data to be processed; inputting the medical data to be processed into a large-scale medical model on the participant's local machine, and obtaining the diagnostic results output by the large-scale medical model.

[0122] Here, medical data to be processed refers to medical data collected or received locally by participating parties that has not been processed by the model, such as imaging data of newly admitted patients, laboratory reports, medical record texts, or historical data that requires secondary review.

[0123] After receiving the medical data to be processed, the participants do not need to transmit the medical data to the central service provider. They can obtain the diagnostic results output by the medical big data model by simply calling the pre-trained medical big data model locally.

[0124] Here, the large medical model is obtained according to the federated learning method described in any of the above embodiments, and will not be elaborated further here.

[0125] The federated learning-based medical data processing method provided in this invention achieves localized intelligent processing of medical data by acquiring the medical data to be processed locally and inputting it into a locally trained medical big model for inference. While ensuring data privacy and security, the medical big model obtained through federated learning can quickly output accurate diagnostic results.

[0126] Based on any of the above embodiments, this invention also provides a federated learning device applied to a central service provider in a federated learning system. The central service provider communicates with at least two participating parties, and the at least two participating parties include at least one medical sample with the same sample but different sample feature dimensions. Referring to FIG4, the device includes: a data acquisition module 410, used to acquire a predicted label carrying a medical sample identifier sent by each participating party, and the prediction deviation between the predicted label and the original true label; a data correction module 420, used to generate a target true label for each medical sample based on the predicted labels of all participating parties if the medical sample is detected as an abnormal medical sample based on the predicted labels and the prediction deviation of all participating parties; and a first data sending module 430, used to send the target true label carrying the medical sample identifier to each participating party, so that each participating party updates its local medical big model based on the target true label and the predicted label.

[0127] The federated learning device provided in this invention collects the predicted labels and corresponding prediction deviations of each participant carrying medical sample identifiers to detect abnormal medical samples. After detecting abnormal medical samples, it generates globally corrected target true labels based on the predicted labels of all participants corresponding to the abnormal medical samples and synchronizes them to all participants to update the local medical big model. This achieves collaborative correction and global optimization of abnormal medical samples, avoids interference from abnormal medical samples to model training, and thus improves the generalization ability and accuracy of the final generated global model.

[0128] Based on any of the above embodiments, this embodiment of the invention also provides a federated learning device applied to participants in a federated learning system. The participants are connected to at least one other participant via a central service provider, and the participants and at least one other participant include at least one medical sample with the same sample but different sample feature dimensions. Referring to Figure 5, the device includes: a prediction module 510, used to input the medical sample into a local medical big model of the participant for each medical sample, obtain a predicted label output by the medical big model, and acquire the prediction deviation between the predicted label and the corresponding original true label; a second data sending module 520, used to send the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; a data receiving module 530, used to receive the target true label of the medical sample returned by the central service provider; wherein the target true label is generated by the central service provider based on the predicted labels of all participants when it detects that the medical sample is an abnormal medical sample based on the predicted labels and prediction deviations of all participants; and a parameter update module 540, used to update the local medical big model of the participant based on the target true label and the predicted label.

[0129] The medical data processing device provided in this embodiment of the invention collects the predicted labels and corresponding prediction deviations of each participant carrying medical sample identifiers, performs abnormal medical sample detection, and after detecting abnormal medical samples, generates globally corrected target true labels based on the predicted labels of all participants corresponding to the abnormal medical samples and synchronizes them to all participants to update the local medical big model. This realizes the collaborative correction and global optimization of abnormal medical samples, avoids the interference of abnormal medical samples on model training, and thus improves the generalization ability and accuracy of the final generated global model.

[0130] Based on any of the above embodiments, this invention also provides a federated learning system, including a central service provider and at least two participating parties.

[0131] The central service provider is configured to implement the following federated learning method, which includes: acquiring a predicted label carrying a medical sample identifier sent by each participant, and the prediction deviation between the predicted label and the original true label; for each medical sample, if the medical sample is detected as an abnormal medical sample based on the predicted labels and the prediction deviation of all participants, generating a target true label for the medical sample based on the predicted labels of all participants; and sending the target true label carrying the medical sample identifier to each participant, so that each participant updates its local large medical model based on the target true label and the predicted label.

[0132] The participating parties are configured to implement the following federated learning method, which includes: for each medical sample, inputting the medical sample into the participating party's local medical big model to obtain the predicted label output by the medical big model, and obtaining the prediction deviation between the predicted label and the corresponding original true label; sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; receiving the target true label of the medical sample returned by the central service provider; wherein the target true label is generated by the central service provider based on the predicted labels of all participating parties when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participating parties; and updating the participating party's local medical big model based on the target true label and the predicted label.

[0133] Figure 6 illustrates a schematic diagram of the physical structure of an electronic device. As shown in Figure 6, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a federated learning method. The method includes: acquiring a predicted label carrying a medical sample identifier sent by each participant, and the prediction deviation between the predicted label and the original true label; for each medical sample, if the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants, generating a target true label for the medical sample based on the predicted labels of all participants; and sending the target true label carrying the medical sample identifier to each participant, so that each participant updates its local large medical model based on the target true label and the predicted label.

[0134] The processor 610 can also call logic instructions in the memory 630 to execute a federated learning method, the method comprising: for each medical sample, inputting the medical sample into a local medical big model of the participants, obtaining a predicted label output by the medical big model, and obtaining the prediction deviation between the predicted label and the corresponding original true label; sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; receiving the target true label of the medical sample returned by the central service provider; wherein the target true label is generated by the central service provider based on the predicted labels of all participants when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants; and updating the local medical big model of the participants based on the target true label and the predicted label of the medical sample.

[0135] The processor 610 can also call logic instructions in the memory 630 to execute a federated learning-based medical data processing method, the method including: acquiring medical data to be processed; inputting the medical data to be processed into a local medical big model of the participants, and obtaining the diagnostic results output by the medical big model.

[0136] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0137] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the federated learning method provided by the above methods. The method includes: acquiring a predicted label carrying a medical sample identifier sent by each participant, and the prediction deviation between the predicted label and the original true label; for each medical sample, if the medical sample is detected as an abnormal medical sample based on the predicted labels of all participants and the prediction deviation, generating a target true label for the medical sample based on the predicted labels of all participants; and sending the target true label carrying the medical sample identifier to each participant, so that each participant updates its local medical big model based on the target true label and the predicted label.

[0138] When the computer program is executed by the processor, the computer can also execute the federated learning method provided by the above methods, the method including: for each medical sample, inputting the medical sample into the local medical big model of the participants, obtaining the predicted label output by the medical big model, and obtaining the prediction deviation between the predicted label and the corresponding original true label; sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; receiving the target true label of the medical sample returned by the central service provider; wherein, the target true label is generated by the central service provider based on the predicted labels of all participants when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants; updating the local medical big model of the participants based on the target true label and the predicted label.

[0139] When the computer program is executed by the processor, the computer can also execute the medical data processing methods provided by the above methods, the methods including: acquiring medical data to be processed; inputting the medical data to be processed into a local medical big model of the participant, and obtaining the diagnostic results output by the medical big model.

[0140] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the federated learning method provided by the methods described above. The method includes: acquiring a predicted label carrying a medical sample identifier sent by each participant, and a prediction deviation between the predicted label and the original true label; for each medical sample, if the medical sample is detected as an anomalous medical sample based on the predicted labels of all participants and the prediction deviation, generating a target true label for the medical sample based on the predicted labels of all participants; and sending the target true label carrying the medical sample identifier to each participant, so that each participant updates its local large medical model based on the target true label and the predicted label.

[0141] When the computer program is executed by the processor, the computer can also execute the federated learning method provided by the above methods, the method including: for each medical sample, inputting the medical sample into the local medical big model of the participants, obtaining the predicted label output by the medical big model, and obtaining the prediction deviation between the predicted label and the corresponding original true label; sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; receiving the target true label of the medical sample returned by the central service provider; wherein, the target true label is generated by the central service provider based on the predicted labels of all participants when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants; updating the local medical big model of the participants based on the target true label and the predicted label.

[0142] When the computer program is executed by the processor, the computer can also execute the medical data processing methods provided by the above methods, the methods including: acquiring medical data to be processed; inputting the medical data to be processed into a local medical big model of the participant, and obtaining the diagnostic results output by the medical big model.

[0143] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A federated learning method, characterized in that, A central service provider is applied to a federated learning system. The central service provider communicates with at least two participants, and the at least two participants include at least one medical sample with the same sample but different feature dimensions. The method includes: acquiring a predicted label carrying a medical sample identifier sent by each participant, and the prediction deviation between the predicted label and the original true label; for each medical sample, if the medical sample is detected as an abnormal medical sample based on the predicted labels of all participants and the prediction deviation, generating a target true label for the medical sample based on the predicted labels of all participants; sending the target true label carrying the medical sample identifier to each participant, so that each participant updates its local large medical model based on the target true label and the predicted label.

2. The federated learning method according to claim 1, characterized in that, The abnormal medical samples are determined as follows: for each medical sample, the prediction variance is determined based on the prediction labels of all participants; and the average prediction deviation is determined based on the prediction deviations of all participants; the global deviation is determined based on the prediction variance and the average prediction deviation; if the global deviation is greater than a threshold, the medical sample is determined to be an abnormal medical sample.

3. The federated learning method according to claim 1, characterized in that, The step of generating the target true label of the medical sample based on the predicted labels of all participants includes: determining the weight corresponding to each participant; and generating the target true label of the medical sample based on the weight corresponding to each participant and the predicted label.

4. The federated learning method according to claim 3, characterized in that, The weight of each participant is obtained by: determining the contribution of each participant in at least one historical federated training; determining the confidence level of the predicted label corresponding to each participant in the current federated training; and determining the weight of each participant based on the contribution and the confidence level.

5. The federated learning method according to claim 1, characterized in that, The method further includes: for each medical sample, if the medical sample is detected as a normal medical sample based on the predicted labels and the predicted bias of all participants, sending a normal medical sample notification carrying a medical sample identifier to each participant, so as to notify each participant to update its local medical big model based on the original true labels and the predicted labels.

6. The federated learning method according to claim 1, characterized in that, The method further includes: before co-training, sending a container image and a deployment description file to each participant, wherein the container image includes a training program and required dependency libraries for performing co-training, and the deployment description file includes a network resource configuration policy and a storage resource configuration policy for the container image; and before co-training, configuring a directed acyclic graph workflow engine, wherein the directed acyclic graph workflow engine is used to split the federated learning task into multiple sequentially executed subtasks, and trigger each participant to execute multiple subtasks sequentially.

7. A federated learning method, characterized in that, A method for using participants in a federated learning system, wherein each participant communicates with at least one other participant via a central service provider, and the participants and at least one other participant share at least one medical sample with the same sample but different feature dimensions, includes: for each medical sample, inputting the medical sample into a local medical big model of the participant to obtain a predicted label output by the medical big model, and obtaining the prediction deviation between the predicted label and the corresponding original true label; sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; receiving the target true label of the medical sample returned by the central service provider; wherein, if the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants, the target true label is generated by the central service provider based on the predicted labels of all participants; and updating the local medical big model of the participant based on the target true label and the predicted label.

8. The federated learning method according to claim 7, characterized in that, The method further includes: receiving a normal medical sample notification of the medical sample returned by the central service provider; wherein the normal medical sample notification is sent when the central service provider detects that the medical sample is a normal medical sample based on the predicted labels and the predicted bias of all participants; and updating the local medical big model of the participants based on the predicted labels and the original true labels.

9. The federated learning method according to claim 7, characterized in that, Before sending the prediction tag carrying the medical sample identifier and the prediction deviation to the central service provider, the method further includes: performing differential privacy noise processing or homomorphic encryption processing on the prediction tag carrying the medical sample identifier and the prediction deviation.

10. A medical data processing method based on federated learning, characterized in that, A participant in a federated learning system, wherein the participant and at least one other participant are connected via a central service provider, and the participant and at least one other participant include at least one medical sample with the same sample but different sample feature dimensions, the method comprising: acquiring medical data to be processed; inputting the medical data to be processed into a large medical model on the participant's local machine, and obtaining a diagnostic result output by the large medical model; wherein the large medical model is obtained by the federated learning method as described in any one of claims 1 to 6 or the federated learning method as described in any one of claims 7 to 9.

11. A federated learning device, characterized in that, A central service provider is applied in a federated learning system. The central service provider communicates with at least two participants, and the at least two participants include at least one medical sample with the same sample but different sample feature dimensions. The device includes: a data acquisition module for acquiring a predicted label carrying a medical sample identifier sent by each participant, and the prediction deviation between the predicted label and the original true label; a data correction module for generating a target true label for each medical sample based on the predicted labels of all participants if the medical sample is detected as an abnormal medical sample based on the predicted labels of all participants and the prediction deviation; and a first data transmission module for sending the target true label carrying the medical sample identifier to each participant, so that each participant updates its local medical big model based on the target true label and the predicted label.

12. A federated learning device, characterized in that, A participant in a federated learning system communicates with at least one other participant through a central service provider. The participant and the other participants share at least one medical sample with the same sample but different feature dimensions. The device includes: a prediction module for inputting the medical sample into a local medical big model of the participant for each medical sample, obtaining a predicted label output by the medical big model, and acquiring the prediction deviation between the predicted label and the corresponding original true label; a second data sending module for sending the predicted label carrying the medical sample identifier and the prediction deviation to the central service provider; a data receiving module for receiving the target true label of the medical sample returned by the central service provider; wherein the target true label is generated by the central service provider based on the predicted labels of all participants when the medical sample is detected as an abnormal medical sample based on the predicted labels and prediction deviations of all participants; and a parameter update module for updating the local medical big model of the participant based on the target true label and the predicted label.

13. A federated learning system, characterized in that, It includes a central service provider and at least two participating parties, wherein the central service provider is configured to implement the federated learning method as described in any one of claims 1 to 6, and the participating parties are configured to implement the federated learning method as described in any one of claims 7 to 9 or the medical data processing method as described in claim 10.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the federated learning method as described in any one of claims 1 to 6, or the federated learning method as described in any one of claims 7 to 9, or the medical data processing method as described in claim 10.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the federated learning method as described in any one of claims 1 to 6, or the federated learning method as described in any one of claims 7 to 9, or the medical data processing method as described in claim 10.

Images