A method for triggering uploading of federated learning parameters driven by medical clinical key indicator threshold

Abstract

The application discloses a medical clinical key index threshold value driven federal learning parameter uploading trigger method, belongs to the cross technical field of federal learning and medical image processing, and specifically comprises the following steps: step 1, constructing a dynamic clinical key index threshold value; step 2, obtaining the local FAM parameter after local training of the current federal participating hospital, simultaneously obtaining the global model FAM module parameter issued from the server in the last round, performing global feature consistency double checking, and obtaining the high-quality local FAM parameter or the local FAM parameter to be optimized of the local FAM parameter of this time; and step 3, generating a global new model of this round based on the dynamic clinical key index threshold value of step 1 and the high-quality local FAM parameter or the local FAM parameter to be optimized of the local FAM parameter of this time of step 2, and issuing the global new model to the receiving server of each federal participating hospital. The application combines parameter screening with diagnostic recall rate, avoids uploading invalid parameters without difference, and realizes clinical adaptation and global cooperation of medical image federal learning.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of federated learning and medical image processing, specifically to a method for triggering the upload of federated learning parameters driven by thresholds of key clinical indicators. Background Technology

[0002] In the field of medical image diagnosis, the technology involved is FedMedCLIP (Federated Medical Contrastive Language-Image Pre-trained Model), which proposes a resource-efficient federated learning framework for healthcare. This framework freezes the CLIP (Contrastive Language-Image Pre-trained Model) encoder, designs a Mask Feature Adaptation Module (FAM) as a global communication module, and introduces a local masked MLP classifier and a KL divergence mutual learning mechanism. This significantly reduces communication and computation costs while addressing the heterogeneity problem of medical data.

[0003] However, FedMedCLIP has the following drawbacks.

[0004] 1. Insufficient Adaptability to Clinical Metrics: Existing methods (including FedMedCLIP) primarily optimize the model's loss function (such as contrastive loss and cross-entropy loss) during training, without directly aligning with key metrics of clinical concern (such as recall / sensitivity). This leads to a disconnect between the federated learning process and actual clinical needs; although the model loss may decrease, unacceptable missed diagnoses may still occur.

[0005] 2. Indiscriminate parameter uploading leads to resource waste: In existing federated learning frameworks, regardless of whether local model updates make a meaningful positive contribution to clinical diagnostic performance, some parameters from all clients (such as FAM (Feature Adaptation Module) parameters in FedMedCLIP) are uploaded to the server for aggregation. This results in a large number of "invalid parameters" consuming network bandwidth and server computing resources, especially noticeable in the later training stages when model performance tends to converge.

[0006] 3. The mechanism for balancing global and local performance is simplistic: FedMedCLIP achieves mutual learning between global and local features through KL (relative entropy) divergence distillation. However, this mechanism lacks a hard check to ensure that the model's feature direction remains consistent with the global model. In highly heterogeneous medical scenarios, models on individual clients may experience excessive "feature shifts" due to local data biases. The soft constraints of KL divergence cannot completely prevent these negative impacts on the stability and generalization ability of the global model.

[0007] Based on this, the present invention designs a federated learning parameter upload triggering method driven by thresholds of key medical clinical indicators to solve the above problems. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention provides a method for triggering the upload of federated learning parameters driven by thresholds of key medical clinical indicators.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] The specific steps of the federated learning parameter upload triggering method driven by thresholds of key clinical indicators are as follows:

[0011] Step 1: Construct dynamic clinical key indicator thresholds ;

[0012] Step 2: Obtain the local FAM parameters of the participating hospitals in the current federal program after local training, and at the same time obtain the global model FAM module parameters issued from the server in the previous round. Perform a double verification of global feature consistency to obtain the high-quality local FAM parameters or the local FAM parameters to be optimized in this local FAM parameter test.

[0013] Step 3: Thresholds for dynamic clinical key indicators based on Step 1 The high-quality local FAM parameters or the local FAM parameters to be optimized from step 2 are used to generate a new global model for this round, which is then distributed to the receiving servers of each participating federal hospital.

[0014] Furthermore, the specific steps for step 1 are as follows:

[0015] Step 11: Collect data from recently involved federal hospitals Recall rate on local validation set, calculate recent Standard deviation of recall on local validation set Furthermore, the average decrease in recall after removing the MLP and KL divergence mutual learning modules from the standard complete FedMedCLIP model was analyzed. Calculate baseline threshold ;

[0016] Step 12: Weighting factors obtained by clinical experts based on risk level assessment of actual application scenarios. Then compared with the baseline threshold Combined, calculate clinical scenario thresholds ;

[0017] Step 13: Calculate the clinical scenario thresholds for each of the participating federal hospitals. The data is aggregated on the server, and collective calibration and correction are performed based on global multi-hospital data to calculate a global fine-tuning amount. In addition, the clinical scenario threshold of this hospital dynamic clinical key indicator thresholds were obtained. .

[0018] Furthermore, baseline threshold The formula is as follows:

[0019]

[0020] Average decline in recall rate The calculation is as follows:

[0021]

[0022] The recall rate of the standard complete FedMedCLIP model across various test scenarios and hospital validation sets. The recall rate of the standard complete FedMedCLIP model in various test scenarios / hospital validation sets after removing the MLP module. The recall rate of the standard complete FedMedCLIP model in various test scenarios / hospital validation sets after removing the KL divergence mutual learning module.

[0023] Furthermore, clinical scenario thresholds The calculation is as follows:

[0024] .

[0025] Furthermore, step 13 is performed as follows:

[0026] Step 131, Obtain Each participating hospital in the federal system calculates its own clinical scenario thresholds. And aggregated to the server;

[0027] Step 132: Assess the data quality of each hospital. and model calibration error ;

[0028] Step 133: Assess the quality of data for each hospital. and model calibration error Calculate the confidence weight of the hospital Then, normalization calculations are performed to obtain the normalized confidence weights. ;

[0029] Step 134: Normalize the confidence weights The global baseline threshold is obtained by calculating the weighted average of the global fine-tuning values. Based on the global baseline threshold and preset basic reference threshold Calculate global fine-tuning amount In addition, the clinical scenario threshold of this hospital dynamic clinical key indicator thresholds were obtained. .

[0030] Furthermore, the specific steps for step 2 are as follows:

[0031] Step 21: Obtain the local FAM parameters of the participating federal hospitals after local training, and simultaneously obtain the global model FAM module parameters from the server in the previous round; flatten the local FAM parameters and global model FAM module parameters into one-dimensional vectors, and denote them as local feature vectors. and global feature vectors ;

[0032] Step 22: Process the local feature vectors respectively and global feature vectors L2 normalization is performed, and then cosine similarity is calculated. and similarity ;

[0033] Step 23: Determine similarity If the similarity threshold is greater than or equal to the threshold, mark the local FAM parameter as a high-quality local FAM parameter and proceed to step 26. If the similarity threshold is not met, proceed to step 24.

[0034] Step 24: Determine similarity If the cumulative number of calculations is greater than the preset maximum number of retries, and if it is, mark the local FAM parameter of this time as the local FAM parameter to be optimized, and then execute step 26; if it is not, execute step 25.

[0035] Step 25: Call the KL divergence mutual learning module of the original FedMedCLIP framework, keep the parameters of CLIP encoder and MLP module frozen throughout, use KL divergence as the constraint loss to drive the local FAM feature distribution to approximate and adjust to the global FAM feature distribution, and then execute steps 21-25.

[0036] Step 26: Output the high-quality local FAM parameters or the local FAM parameters to be optimized.

[0037] Furthermore, similarity The specific calculations are as follows:

[0038]

[0039] ε is .

[0040] Furthermore, the specific steps for step 3 are as follows:

[0041] Step 31: Determine whether the increase in recall rate of the local model in this round is greater than or equal to the dynamic clinical key indicator threshold. If the determination is yes, proceed to step 32; if the determination is no, proceed to step 33.

[0042] Step 32: Determine whether the local FAM parameter is a high-quality local FAM parameter. If it is, proceed to step 34; if it is not, proceed to step 33.

[0043] Step 33: Determine whether the cumulative number of executions in 31 or 32 is greater than or equal to the maximum number of retries. If the determination is yes, mark the local FAM parameters as global aggregation parameters to be optimized, and then execute step 37. If the determination is no, call the KL divergence mutual learning module of the original FedMedCLIP framework, keep the CLIP encoder and MLP module parameters frozen throughout, use KL divergence as the constraint loss, drive the local FAM feature distribution to approach and adjust the global FAM feature distribution, and re-execute step 31.

[0044] Step 34: Mark the trainable parameters of the FAM feature adaptation module and the trainable parameters of the MLP local training module generated in this local training as high-quality global aggregation parameters, compress them and upload them to the server;

[0045] Step 35: The server only receives the high-quality global aggregation parameters uploaded in this round, and performs aggregation weight allocation, global model aggregation, and dynamic threshold adjustment.

[0046] Step 36: Each participating federal hospital receives the new global model for this round from the server, and performs model updates, baseline data updates, and closed-loop connection.

[0047] Step 37: The global aggregation parameters to be optimized are uploaded late and will not participate in this round of global aggregation. They will be uploaded to the server after step 31 is executed in subsequent rounds.

[0048] Compared to existing technologies, the advantages of this invention are as follows: Based on the FedMedCLIP framework, this invention retains its core architecture of "frozen CLIP encoder, FAM feature adaptation, MLP local training, and KL divergence mutual learning." Through "dynamic clinical threshold and global feature consistency dual verification" and a "closed-loop optimization process," combined with diagnostic recall, parameter screening avoids the indiscriminate uploading of invalid parameters, achieving clinical adaptation and global collaboration in medical image federated learning. Furthermore, by analyzing baseline model fluctuations, considering actual clinical application needs, and collecting feedback from multiple hospitals for optimization, this invention further enhances the clinical adaptation and global collaboration in medical image federated learning. Dynamic adjustments are made to ensure that the thresholds are accurately matched with medical scenarios. At the same time, a closed-loop optimization process is established to form a complete mechanism of "hospital self-check - server aggregates parameters - threshold dynamic iteration". Through the verification requirement of feature similarity of no less than 90% and the limit of a maximum of 3 retries, the application effect of the model on local data of each hospital is guaranteed, and its diagnostic role is guaranteed to be stable in the global scope. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0050] Figure 1 This is a flowchart illustrating the federated learning parameter upload triggering method driven by threshold of key medical clinical indicators of the present invention.

[0051] Figure 2 This is the flowchart for step 2;

[0052] Figure 3 This is the flowchart for step 3;

[0053] Figure 4 This is a comparison of the accuracy of this invention with other federated learning strategies;

[0054] Figure 5 This is a comparison of the training loss of this invention with other federated learning strategies;

[0055] Figure 6 This describes how the accuracy of this invention and other federated learning strategies changes with training rounds.

[0056] Figure 7 This describes how the training loss of this invention and other federated learning strategies changes with training rounds. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0058] Example 1: In some embodiments, please refer to the accompanying drawings. Figure 1The specific steps of the federated learning parameter upload triggering method driven by thresholds of key clinical indicators are as follows:

[0059] Step 1: Construct dynamic clinical key indicator thresholds ;

[0060] Step 1 is performed as follows:

[0061] Step 11: Collect data from recently involved federal hospitals Recall rate on the local validation set, calculate the most recent Standard deviation of recall on local validation set The average decrease in recall after removing the MLP (Multilayer Perceptron) and KL divergence mutual learning modules from the standard complete FedMedCLIP model was further analyzed. Calculate baseline threshold ;

[0062] standard deviation of recall rate The calculation formula is as follows:

[0063]

[0064]

[0065] for Average recall on the local validation set in rounds, For the first Recall rate on the local validation set;

[0066] Baseline threshold The formula is as follows:

[0067]

[0068] Average decline in recall rate The calculation is as follows:

[0069]

[0070] The recall rate of the standard complete FedMedCLIP model across various test scenarios and hospital validation sets. The recall rate of the standard complete FedMedCLIP model in various test scenarios / hospital validation sets after removing the MLP module. Recall rates on various test scenarios / hospital validation sets after removing the KL divergence mutual learning module from the standard complete FedMedCLIP model;

[0071] Step 12: Weighting factors obtained by clinical experts based on risk level assessment of actual application scenarios. Then compared with the baseline threshold Combined, calculate clinical scenario thresholds ;

[0072] Practical applications include emergency screening, routine physical examinations, and postoperative follow-up.

[0073] High-risk scenario weighting factor The weighting factor for low-risk scenarios is 1.2. It is 0.8;

[0074] The high-risk scenario is emergency screening;

[0075] Low-risk scenarios include routine physical examinations and postoperative follow-ups;

[0076] Clinical scenario threshold The calculation is as follows:

[0077]

[0078] Step 13: Calculate the clinical scenario thresholds for each of the participating federal hospitals. The data is aggregated on the server, and collective calibration and correction are performed based on global multi-hospital data to calculate a global fine-tuning amount. In addition, the clinical scenario threshold of this hospital dynamic clinical key indicator thresholds were obtained. ;

[0079]

[0080] Step 13 is performed as follows:

[0081] Step 131, Obtain Each participating hospital in the federal system calculates its own clinical scenario thresholds. And aggregated to the server;

[0082] Step 132: Assess the data quality of each hospital. and model calibration error ;

[0083] Data quality score The server assigns scores based on hospital sample size, annotation accuracy, image clarity, and data integrity, with values ​​ranging from [0,1]. Higher scores indicate better data quality.

[0084] Model calibration error : Expected calibration error, with a value range of [0,1]. The smaller the ECE, the more reliable the model prediction and the better the calibration effect.

[0085] Step 133: Assess the quality of data for each hospital. and model calibration error Calculate the confidence weight for each hospital Then, normalization calculations are performed to obtain the normalized confidence weights. ;

[0086] Confidence weight The calculation formula is as follows:

[0087]

[0088] for ;

[0089] Normalized confidence weights The calculation formula is as follows:

[0090]

[0091] Step 134: Normalize the confidence weights The global baseline threshold is obtained by calculating the weighted average of the global fine-tuning values. Based on the global baseline threshold and preset basic reference threshold Calculate global fine-tuning amount In addition, the clinical scenario threshold of this hospital dynamic clinical key indicator thresholds were obtained. ;

[0092] Global fine-tuning amount The calculation is as follows:

[0093]

[0094]

[0095] Step 2: Obtain the local FAM parameters of the participating hospitals in the current federal program after local training, and at the same time obtain the global model FAM module parameters issued from the server in the previous round. Perform a double verification of global feature consistency to obtain the high-quality local FAM parameters or the local FAM parameters to be optimized in this local FAM parameter test.

[0096] Global feature consistency double verification includes vectorization of local FAM module parameters and global model FAM module parameters, normalization and cosine similarity calculation, threshold verification and feature alignment processing;

[0097] like Figure 2 As shown, the specific steps for step 2 are as follows:

[0098] Step 21: Obtain the local FAM parameters of the participating federal hospitals after local training, and simultaneously obtain the global model FAM module parameters from the server in the previous round; flatten the local FAM parameters and global model FAM module parameters into one-dimensional vectors, and denote them as local feature vectors. and global feature vectors ;

[0099] Step 22: Process the local feature vectors respectively and global feature vectors L2 normalization is performed, and then cosine similarity is calculated. and similarity ;

[0100] Normalization eliminates the interference of model parameter scale differences on similarity calculation;

[0101] Similarity The specific calculations are as follows:

[0102]

[0103] ε is ;

[0104] Step 23: Determine similarity If the similarity threshold is greater than or equal to the threshold, mark the local FAM parameter as a high-quality local FAM parameter and proceed to step 26. If the similarity threshold is not met, proceed to step 24.

[0105] Step 24: Determine similarity If the cumulative number of calculations is greater than the preset maximum number of retries, and if it is, mark the local FAM parameter of this time as the local FAM parameter to be optimized, and then execute step 26; if it is not, execute step 25.

[0106] Step 25: Call the KL divergence mutual learning module of the original FedMedCLIP framework, keep the parameters of CLIP encoder and MLP module frozen throughout, use KL divergence as the constraint loss to drive the local FAM feature distribution to approximate and adjust to the global FAM feature distribution, and then execute steps 21-25.

[0107] The similarity threshold is 0.90;

[0108] The maximum number of retries is preset to 3.

[0109] Step 26: Output the high-quality local FAM parameters or the local FAM parameters to be optimized in this local FAM parameter analysis.

[0110] Step 3: Thresholds for dynamic clinical key indicators based on Step 1 The high-quality local FAM parameters or the local FAM parameters to be optimized from step 2 are used to generate a new global model for this round, and then distributed to the receiving servers of each participating federal hospital.

[0111] like Figure 3 As shown, the specific steps for step 3 are as follows:

[0112] Step 31: Determine whether the increase in recall rate of the local model in this round is greater than or equal to the dynamic clinical key indicator threshold. If the determination is yes, proceed to step 32; if the determination is no, proceed to step 33.

[0113] Step 32: Determine whether the local FAM parameter is a high-quality local FAM parameter. If it is, proceed to step 34; if it is not, proceed to step 33.

[0114] Step 33: Determine whether the cumulative number of executions in 31 or 32 is greater than or equal to the maximum number of retries. If the determination is yes, mark the local FAM parameters as global aggregation parameters to be optimized, and then execute step 37. If the determination is no, call the KL divergence mutual learning module of the original FedMedCLIP framework, keep the CLIP encoder and MLP module parameters frozen throughout, use KL divergence as the constraint loss, drive the local FAM feature distribution to approach and adjust to the global FAM feature distribution, and re-execute step 31.

[0115] The maximum number of retries is 3;

[0116] Step 34: Mark the trainable parameters of the FAM feature adaptation module and the trainable parameters of the MLP local training module generated in this local training as high-quality global aggregation parameters, compress them and upload them to the server;

[0117] Step 35: The server only receives the high-quality global aggregation parameters uploaded in this round, and performs aggregation weight allocation, global model aggregation, and dynamic threshold adjustment;

[0118] Aggregate weight allocation: Aggregate weights are allocated based on the "recall improvement" corresponding to the high-quality global aggregate parameters of each hospital. The greater the recall improvement, the higher the weight ratio, ensuring that the better-performing parameters contribute more to the global model.

[0119] Global model aggregation: The high-quality global aggregation parameters uploaded by all participating federal hospitals are weighted and averaged to obtain global FAM module parameters and global MLP module parameters; the global FAM parameters and global MLP parameters are combined with the frozen shared CLIP encoder to generate a new global model for this round.

[0120] The new global model includes global FAM module parameters and global MLP module parameters.

[0121] Dynamic threshold adjustment: Based on the feedback from each hospital's verification, the thresholds for the next round of dynamic clinical key indicators will be dynamically optimized and adjusted. and similarity threshold;

[0122] If ≥70% of the federally participating hospitals have a local model recall rate improvement exceeding the current Δθ, then the next round of Δθ will be reduced.

[0123] If ≥70% of the federally participating hospitals have similarity If all values ​​are ≥0.95: Increase the similarity threshold for the next round (up to a maximum of 0.98).

[0124] Step 36: Each participating federal hospital receives the new global model for this round from the server, and performs model updates, baseline data updates, and closed-loop connection.

[0125] Local model update: Replace the original local baseline model with the new global model and use it as the initial model for the next round of local training;

[0126] Baseline data update: Update local Historical recall rate data and verification logs;

[0127] The verification log contains the verification results and retry count for step 2, and the verification results and retry count for steps 31-34.

[0128] Closed-loop connection: The updated local baseline data and verification logs are directly used for the calculation of the dynamic clinical key indicator threshold Δθ in the next round of step 1, forming a complete optimization closed loop;

[0129] Step 37: The global aggregation parameters to be optimized are uploaded late and will not participate in this round of global aggregation. They will be uploaded to the server after step 31 is executed in subsequent rounds.

[0130] This invention uses FedMedCLIP as its basic framework, retaining its core architecture of "frozen CLIP encoder, FAM feature adaptation, MLP local training, and KL divergence mutual learning." Through "dynamic clinical threshold and global feature consistency dual verification" and a "closed-loop optimization process," it combines diagnostic recall for parameter screening to avoid indiscriminate uploading of invalid parameters, achieving clinical adaptation and global collaboration in medical image federated learning. By analyzing baseline model fluctuations, considering actual clinical application needs, and collecting feedback from multiple hospitals for optimization, it further refines the model. Dynamic adjustments are made to ensure that the thresholds are accurately matched with medical scenarios. At the same time, a closed-loop optimization process is established to form a complete mechanism of "hospital self-check - server aggregates parameters - threshold dynamic iteration". Through the verification requirement of feature similarity of no less than 90% and the limit of a maximum of 3 retries, the application effect of the model on local data of each hospital is guaranteed, and its diagnostic role is guaranteed to be stable in the global scope.

[0131] Figure 4 The accuracy of five federated learning algorithms in medical imaging tasks was compared. As can be seen from the red bars in the figure, the proposed method (corresponding to the proposed red bars) achieves significantly higher accuracy than the compared algorithms FedAvg, FedCLIP, MOON, and FACMIC. This is attributed to the method's use of dynamic clinical key indicator thresholds to select high-quality parameters that substantially contribute to diagnosis, and its use of global feature consistency double verification (similarity ≥ 0.90) to eliminate invalid parameters that easily cause feature shifts. This allows the model to achieve the highest diagnostic accuracy even with multi-center heterogeneous data.

[0132] Figure 5 The training losses of five federated learning algorithms were compared. The proposed method (red bars) has the lowest loss value, indicating its optimal convergence performance. This is because: in each training round, the local FAM parameters need to be checked for cosine similarity with the global FAM parameters. Those that fail to meet the standard are retried a maximum of three times using the KL divergence module for feature alignment, effectively suppressing the interference of noisy parameters on the global model; simultaneously, the closed-loop optimization mechanism dynamically adjusts the aggregation weights based on the hospital data quality and model calibration error, further reducing the training loss.

[0133] Figure 6The validation performance of five federated learning algorithms in medical imaging tasks was compared. As can be seen from the red curve in the figure, the proposed method significantly outperforms FedAvg (Federated Averaging Algorithm), FedCLIP (Federated CLIP Algorithm), MOON (Model Contrastive Federated Learning), and FACMIC (Federated Adaptive CLIP Medical Image Classification Model) in terms of model convergence speed, final recognition accuracy, and operational stability. At the beginning of training, we validated the similarity of parameter features (requiring a value no lower than 0.90) and set a maximum of three rounds of iterative optimization, selecting only qualified model parameters for integration to allow the model to quickly enter a good training state. During training, we set dynamic judgment criteria based on different clinical scenarios: high-risk scenarios were weighted at 1.2, and low-risk scenarios at 0.8; simultaneously, we allocated weights for parameter integration based on the data quality and model error of each hospital, allowing the model accuracy to steadily increase. In the later stages of training, the entire process automatically iterates and optimizes: when the model recall improvement reaches the target for more than 70% of hospitals, the judgment criteria are appropriately relaxed; when the parameter similarity for more than 70% of hospitals reaches 0.95, the similarity standard is raised to a maximum of 0.98, and parameters with poor performance are temporarily suspended from uploading and integration. Ultimately, our model stabilizes first, and the training process is the smoothest. Experimental results show that the solution combining clinical scenario judgment, parameter validation, weight integration, and dynamic optimization can not only filter out unqualified model parameters and adapt to different medical use scenarios, but also solve the problem of data differences among multiple hospitals, effectively improving the overall performance and practical application value of the medical imaging federated learning model.

[0134] Figure 7We compared the training loss changes of five federated learning algorithms in medical imaging tasks. As shown by the red curve in the figure, the proposed method significantly outperforms FedAvg, FedCLIP, MOON, and FACMIC in terms of loss reduction speed, final convergence value, and training stability. At the beginning of training, we validated the feature similarity between local and global FAM parameters, requiring a value no lower than 0.90. Parameters that did not meet this requirement were adjusted using the KL divergence module, with a maximum of three rounds of testing. Only qualified parameters that passed the validation were included in the integration. This effectively reduced the interference of noisy parameters early on, resulting in a significantly faster model loss reduction compared to other algorithms. During training, we set dynamic judgment criteria based on different clinical scenarios: a weight of 1.2 for high-risk scenarios and 0.8 for low-risk scenarios. Simultaneously, we allocated aggregation weights based on the data quality (sample size, annotation accuracy, image clarity, etc.) and model calibration error of each hospital, allowing the model to continuously and stably optimize and reduce loss. In the later stages of training, the entire process enters a closed-loop automatic iteration state: when the recall rate of models from more than 70% of hospitals meets the target, the judgment threshold for the next round is appropriately lowered; when the parameter similarity of more than 70% of hospitals reaches 0.95, the similarity standard is raised to a maximum of 0.98. For parameters with unsatisfactory optimization effects, the system will temporarily suspend uploading and not participate in this round of integration to avoid dragging down the overall model. Ultimately, our model converged first, with the lowest training loss and minimal fluctuations throughout the entire training process. Experimental results show that this scheme, which combines clinical scenario judgment, parameter validation, weight allocation, and closed-loop dynamic optimization, can not only filter out invalid parameters and adapt to different medical use scenarios, but also effectively improve the training speed and final performance of the medical imaging federated learning model.

[0135] 1. Improve clinical adaptability and meet the actual needs of medical scenarios.

[0136] By introducing a dynamic clinical key indicator threshold based on recall rate. This approach directly links the optimization objective of federated learning with key indicators for clinical diagnosis. For example, in high-risk scenarios such as skin cancer screening (corresponding to risk_weight=1.2), this method drives the model to prioritize improving recall, effectively reducing the risk of missed diagnoses. Simultaneously, it dynamically adjusts the thresholds for key clinical indicators. Its dynamic adjustment capability allows the model to flexibly adapt to different clinical needs from emergency care to follow-up, solving the problem of existing methods being out of touch with clinical practice.

[0137] 2. Optimize resource consumption and reduce communication and computing costs.

[0138] By employing a global feature consistency dual-verification mechanism to screen "high-quality parameters," this approach avoids the indiscriminate uploading of a large number of "invalid parameters" that contribute no contribution to clinical indicators or cause feature shifts. Compared to FedMedCLIP, this scheme further reduces the amount of parameter transmission in communication rounds (as described in the patent innovation document, communication overhead can be further reduced). While ensuring model performance, it improves bandwidth utilization and server aggregation efficiency, enabling remote medical institutions with limited computing power and network conditions to effectively participate in federated learning.

[0139] 3. Enhance the stability of the global model and ensure the collaborative effect of multiple centers.

[0140] By employing a hard validation mechanism of "feature similarity ≥ 90%" and a retry alignment mechanism, the local models of each client are effectively constrained, preventing excessive feature bias caused by data heterogeneity. This allows for effective control of global model accuracy fluctuations as the number of participating hospitals increases or data distribution becomes more diverse (e.g., expanding from 5 centers to 15 centers), ensuring the stability and reliability of multi-center collaborative diagnosis.

[0141] In summary, the technical solution of this invention achieves triple optimization in "clinical practical value, resource utilization efficiency, and global collaborative stability," providing key support for the clinical application of medical image federated learning technology.

[0142] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such modifications or substitutions will 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 method for triggering the upload of federated learning parameters driven by thresholds of key clinical indicators, characterized in that: The specific steps are as follows: Step 1: Construct dynamic clinical key indicator thresholds ; Step 2: Obtain the local FAM parameters of the participating hospitals in the current federal program after local training, and at the same time obtain the global model FAM module parameters issued from the server in the previous round. Perform a double verification of global feature consistency to obtain the high-quality local FAM parameters or the local FAM parameters to be optimized in this local FAM parameter test. Step 3: Thresholds for dynamic clinical key indicators based on Step 1 The high-quality local FAM parameters or the local FAM parameters to be optimized from step 2 are used to generate a new global model for this round, which is then distributed to the receiving servers of each participating federal hospital.

2. The method for triggering the upload of federated learning parameters driven by thresholds of key medical clinical indicators according to claim 1, characterized in that, Step 1 is performed as follows: Step 11: Collect data from recently involved federal hospitals Recall rate on local validation set, calculate recent Standard deviation of recall on local validation set Furthermore, the average decrease in recall after removing the MLP and KL divergence mutual learning modules from the standard complete FedMedCLIP model was analyzed. Calculate baseline threshold ; Step 12: Weighting factors obtained by clinical experts based on risk level assessment of actual application scenarios. Then compared with the baseline threshold Combined, calculate clinical scenario thresholds ; Step 13: Calculate the clinical scenario thresholds for each of the participating federal hospitals. The data is aggregated on the server, and collective calibration and correction are performed based on global multi-hospital data to calculate a global fine-tuning amount. In addition, the clinical scenario threshold of this hospital dynamic clinical key indicator thresholds were obtained. .

3. The method for triggering the upload of federated learning parameters driven by thresholds of key medical clinical indicators according to claim 1, characterized in that, Baseline threshold The formula is as follows: ； Average decline in recall rate The calculation is as follows: ； The recall rate of the standard complete FedMedCLIP model across various test scenarios and hospital validation sets. The recall rate of the standard complete FedMedCLIP model in various test scenarios / hospital validation sets after removing the MLP module. The recall rate of the standard complete FedMedCLIP model in various test scenarios / hospital validation sets after removing the KL divergence mutual learning module.

4. The method for triggering the upload of federated learning parameters driven by thresholds of key clinical indicators according to claim 3, characterized in that, Clinical scenario threshold The calculation is as follows: 。 5. The method for triggering the upload of federated learning parameters driven by thresholds of key medical clinical indicators according to claim 4, characterized in that, Step 13 is performed as follows: Step 131, Obtain Each participating hospital in the federal system calculates its own clinical scenario thresholds. And aggregated to the server; Step 132: Assess the data quality of each hospital. and model calibration error ; Step 133: Assess the quality of data for each hospital. and model calibration error Calculate the confidence weight of the hospital Then, normalization calculations are performed to obtain the normalized confidence weights. ; Step 134: Normalize the confidence weights The global baseline threshold is obtained by calculating the weighted average of the global fine-tuning values. Based on the global baseline threshold and preset basic reference threshold Calculate global fine-tuning amount In addition, the clinical scenario threshold of this hospital dynamic clinical key indicator thresholds were obtained. .

6. The method for triggering the upload of federated learning parameters driven by thresholds of key clinical indicators according to claim 5, characterized in that, The specific steps for step 2 are as follows: Step 21: Obtain the local FAM parameters of the participating federal hospitals after local training, and simultaneously obtain the global model FAM module parameters from the server in the previous round; flatten the local FAM parameters and global model FAM module parameters into one-dimensional vectors, and denote them as local feature vectors. and global feature vectors ; Step 22: Process the local feature vectors respectively and global feature vectors L2 normalization is performed, and then cosine similarity is calculated. and similarity ; Step 23: Determine similarity If the similarity threshold is greater than or equal to the threshold, mark the local FAM parameter as a high-quality local FAM parameter and proceed to step 26. If the similarity threshold is not met, proceed to step 24. Step 24: Determine similarity If the cumulative number of calculations is greater than the preset maximum number of retries, and if it is, mark the local FAM parameter of this time as the local FAM parameter to be optimized, and then execute step 26; if it is not, execute step 25. Step 25: Call the KL divergence mutual learning module of the original FedMedCLIP framework, keep the parameters of CLIP encoder and MLP module frozen throughout, use KL divergence as the constraint loss to drive the local FAM feature distribution to approximate and adjust to the global FAM feature distribution, and then execute steps 21-25. Step 26: Output the high-quality local FAM parameters or the local FAM parameters to be optimized.

7. The method for triggering the upload of federated learning parameters driven by thresholds of key clinical indicators according to claim 6, characterized in that, Similarity The specific calculations are as follows: ； ε is .

8. The method for triggering the upload of federated learning parameters driven by thresholds of key clinical indicators according to claim 6, characterized in that, The specific steps for step 3 are as follows: Step 31: Determine whether the increase in recall rate of the local model in this round is greater than or equal to the dynamic clinical key indicator threshold. If the determination is yes, proceed to step 32; if the determination is no, proceed to step 33. Step 32: Determine whether the local FAM parameter is a high-quality local FAM parameter. If it is, proceed to step 34; if it is not, proceed to step 33. Step 33: Determine whether the cumulative number of executions in 31 or 32 is greater than or equal to the maximum number of retries. If the determination is yes, mark the local FAM parameters as global aggregation parameters to be optimized, and then execute step 37. If the determination is no, call the KL divergence mutual learning module of the original FedMedCLIP framework, keep the CLIP encoder and MLP module parameters frozen throughout, use KL divergence as the constraint loss, drive the local FAM feature distribution to approach and adjust the global FAM feature distribution, and re-execute step 31. Step 34: Mark the trainable parameters of the FAM feature adaptation module and the trainable parameters of the MLP local training module generated in this local training as high-quality global aggregation parameters, compress them and upload them to the server; Step 35: The server only receives the high-quality global aggregation parameters uploaded in this round, and performs aggregation weight allocation, global model aggregation, and dynamic threshold adjustment. Step 36: Each participating federal hospital receives the new global model for this round from the server, and performs model updates, baseline data updates, and closed-loop connection. Step 37: The global aggregation parameters to be optimized are uploaded late and will not participate in this round of global aggregation. They will be uploaded to the server after step 31 is executed in subsequent rounds.

Images