Medical image data processing method and device and electronic equipment

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a medical image data processing method and device and electronic equipment. The method comprises the following steps: performing feature extraction on medical image data to be processed to obtain image embedding features of the medical image data; according to a pre-trained image strategy model, model reasoning is carried out on the image embedded features, a candidate strategy set is obtained, the candidate strategy set comprises a plurality of candidate operation strategies, and the candidate operation strategies are used for representing operation rules for processing the medical image data; sending the candidate strategy set to a plurality of expert clients, wherein the expert clients are used for performing advantage and disadvantage evaluation on the candidate operation strategies according to local data to obtain strategy scores; and updating model parameters of the image strategy model according to the strategy scores returned by the plurality of expert clients. The processing efficiency and accuracy of the medical image data can be improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a method for processing medical image data, a device for processing medical image data, and an electronic device. Background Technology

[0002] With the in-depth development of artificial intelligence technology, large-scale pre-trained models have driven the widespread application of cross-domain collaborative intelligent learning models. Traditional federated learning frameworks allow medical institutions to collaboratively optimize models while protecting the privacy of original data, avoiding the direct exchange of sensitive data. However, in actual deployments, the data held by each client often originates from different knowledge domains, exhibiting significant differences in feature distribution, data structure, and semantic representation, leading to prominent data heterogeneity issues. Traditional federated learning methods, such as unified parameter aggregation algorithms, employ fixed global aggregation strategies to process model parameters uploaded by clients. When faced with highly heterogeneous data distributions, this single strategy cannot dynamically adapt to the local data characteristics of each client, making it difficult for the global model to converge stably during training, resulting in a significant reduction in overall generalization performance. Specifically, conflicts may arise between local data features from different knowledge domains, causing the model update direction to deviate from the optimal solution, thereby affecting the accuracy and processing efficiency of medical image data in subsequent operations.

[0003] Furthermore, federated learning processes typically require multiple rounds of gradient or model parameter transfers between the server and client, with each round involving a significant amount of data exchange. In medical image processing scenarios, raw image data is high-dimensional and massive in size, and frequent parameter transfers lead to a sharp increase in communication overhead. This high-frequency data interaction not only consumes substantial network bandwidth resources but also significantly prolongs the model training cycle. Especially in remote medical or mobile terminal environments with limited network conditions, system response latency is exacerbated, and overall scalability is severely constrained, making it difficult to meet the timeliness requirements of real-time medical image analysis. Summary of the Invention

[0004] This application provides a method, apparatus, and electronic device for processing medical image data, with the aim of improving the efficiency and accuracy of medical image data processing.

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0006] According to one aspect of the embodiments of this application, a method for processing medical image data is provided. The method includes: extracting features from the medical image data to be processed to obtain image embedding features of the medical image data; performing model inference on the image embedding features according to a pre-trained image strategy model to obtain a candidate strategy set, the candidate strategy set including multiple candidate operation strategies, the candidate operation strategies representing operation rules for processing the medical image data; sending the candidate strategy set to multiple expert clients, the expert clients being used to evaluate the merits of the candidate operation strategies based on local data to obtain strategy scores; and updating the model parameters of the image strategy model according to the strategy scores returned by the multiple expert clients.

[0007] According to one aspect of the embodiments of this application, a medical image data processing apparatus is provided, the apparatus comprising: The extraction module is configured to extract features from the medical image data to be processed, thereby obtaining the image embedding features of the medical image data. The inference module is configured to perform model inference on the image embedding features based on a pre-trained image strategy model to obtain a candidate strategy set, which includes multiple candidate operation strategies, and the candidate operation strategies are used to represent the operation rules for processing the medical image data. The sending module is configured to send the candidate strategy set to multiple expert clients, wherein the expert clients are used to evaluate the merits of the candidate operation strategies based on local data to obtain a strategy score; The update module is configured to update the model parameters of the image strategy model based on the strategy scores returned by multiple expert clients.

[0008] In some embodiments of this application, based on the above technical solutions, the apparatus further includes: The acquisition module is configured to acquire domain embedding features sent by multiple expert clients from a candidate client set, the domain embedding features being used to represent the knowledge domain corresponding to the local data of the expert clients; The selection module is configured to select multiple expert clients from the candidate client set that match the image embedding features based on the domain embedding features.

[0009] In some embodiments of this application, based on the above technical solutions, the selection module is further configured to: obtain the semantic similarity between the domain embedding feature and the image embedding feature; and select multiple expert clients that match the image embedding feature from the candidate client set according to the semantic similarity.

[0010] In some embodiments of this application, based on the above technical solutions, the strategy scoring includes differential privacy noise added by the expert client, and the differential privacy noise follows a Gaussian distribution.

[0011] In some embodiments of this application, based on the above technical solutions, the apparatus further includes: The calibration module is configured to perform monotonically affine calibration on the strategy scores returned by multiple expert clients according to preset scale calibration parameters to obtain strategy scores with uniform dimensions; the scale calibration parameters are trained based on historical data returned by the expert clients for a common reference task, which is a general knowledge task applicable to multiple knowledge domains.

[0012] In some embodiments of this application, based on the above technical solutions, the apparatus further includes: The standardization module is configured to obtain the average and standard deviation of multiple strategy scores returned by the same expert client; and to convert the strategy scores returned by the same expert client into standardized scores based on the average and standard deviation.

[0013] In some embodiments of this application, based on the above technical solutions, the update module is further configured to: obtain a reliability weight corresponding to the expert client, the reliability weight being used to represent the accuracy of the expert client's evaluation of historical tasks; for the same candidate operation strategy, perform a weighted mapping on the strategy scores returned by different expert clients according to the reliability weight to obtain a comprehensive score for the candidate operation strategy; and perform reinforcement learning on the influence strategy model according to the comprehensive score to update the model parameters of the image strategy model.

[0014] In some embodiments of this application, based on the above technical solutions, the updating module is further configured to: obtain the average value and standard deviation of the comprehensive scores of different candidate operation strategies; and perform baseline subtraction and normalization on the comprehensive scores of different candidate operation strategies according to the average value and standard deviation.

[0015] According to one aspect of the embodiments of this application, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements the medical image data processing method as described in the above technical solutions.

[0016] According to one aspect of the embodiments of this application, an electronic device is provided, the electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the executable instructions to implement the medical image data processing method as described in the above technical solution.

[0017] According to one aspect of the embodiments of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the medical image data processing method as described in the above technical solutions.

[0018] The technical solution provided in this application embodiment effectively solves the problems of data heterogeneity and communication overhead by dynamically optimizing the model through feature extraction, strategy reasoning and expert evaluation. It has the advantages of effectively handling the heterogeneity of medical image data, improving model training efficiency and real-time performance.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0021] Figure 1 An exemplary system architecture block diagram illustrating the application of the technical solution of this application is shown schematically.

[0022] Figure 2 A flowchart illustrating a method for processing medical image data according to one embodiment of this application is shown.

[0023] Figure 3 A flowchart illustrating the processing of medical image data based on a selected expert client is shown in one embodiment of this application.

[0024] Figure 4 A flowchart illustrating the selection of an expert client is shown in one embodiment of this application.

[0025] Figure 5 A flowchart illustrating the model parameters of an image strategy model in one embodiment of this application is shown.

[0026] Figure 6 A schematic block diagram of a medical image data processing apparatus provided in an embodiment of this application is shown.

[0027] Figure 7 A schematic diagram of a computer system architecture suitable for implementing the embodiments of this application is shown. Detailed Implementation

[0028] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0029] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0030] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0031] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0032] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0033] Figure 1 An exemplary system architecture block diagram illustrating the application of the technical solution of this application is shown schematically.

[0034] like Figure 1As shown, the system architecture 100 may include terminal devices 110, a network 120, and a server 130. Terminal devices 110 may include various electronic devices such as smartphones, tablets, laptops, desktop computer smart speakers, smart wearable devices, smart in-vehicle devices, and smart payment terminals. Server 130 may be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Network 120 may be a communication medium of various connection types capable of providing a communication link between terminal devices 110 and server 130, such as a wired communication link or a wireless communication link.

[0035] Depending on the implementation requirements, the system architecture in this application embodiment can have any number of terminal devices, networks, and servers. For example, server 130 can be a server group composed of multiple server devices. In addition, the technical solutions provided in this application embodiment can be applied to terminal device 110, or to server 130, or can be implemented jointly by terminal device 110 and server 130. This application does not impose any special limitations on this.

[0036] In federated learning frameworks, the distribution of medical image data held by different clients varies. Traditional unified parameter aggregation strategies cannot adapt to the diverse characteristics of local data, leading to difficulties in global model convergence. Furthermore, frequent gradient or model parameter transmissions incur communication overhead, impacting the overall system efficiency and scalability, especially in environments with limited network bandwidth, where system responsiveness is constrained.

[0037] For example, in a federated learning network composed of multiple medical institutions, each institution holds medical image data of different modalities, such as X-rays, MRI, and computed tomography. Due to differences in data sources and acquisition conditions, the data distribution among clients exhibits heterogeneity. When using a federated averaging algorithm for model training, the server needs to exchange parameters with clients in multiple rounds. However, due to the inconsistent data distribution, the generalization ability of the global model decreases on some clients, and the communication process consumes network resources, leading to a prolonged training cycle and a decline in the overall system efficiency.

[0038] If the aforementioned issues are not addressed, federated learning systems will be unable to effectively handle heterogeneous medical image data, leading to model errors in practical applications. Furthermore, continuous communication overhead will limit the system's deployment scope, especially in resource-constrained environments, making it difficult for federated learning to meet real-time and reliability requirements, thus impacting the system's practicality and scalability.

[0039] To address the problems existing in the above-mentioned related technologies, embodiments of this application provide a technical solution including a method for processing medical image data, a device for processing medical image data, a computer-readable medium, an electronic device, and a computer program product.

[0040] The technical solutions provided in this application will be described in detail below with reference to specific embodiments.

[0041] Figure 2 A flowchart illustrating a method for processing medical image data according to one embodiment of this application is shown. Figure 2 As shown, the method may include the following steps S210 to S240.

[0042] S210: Extract features from the medical image data to be processed to obtain the image embedding features of the medical image data.

[0043] Medical imaging data refers to image information acquired through various medical imaging devices, such as computed tomography (CT) images, magnetic resonance imaging (MRI) images, X-ray images, and ultrasound images. This data typically contains rich biomedical information and serves as a crucial basis for clinical diagnosis and treatment.

[0044] Feature extraction refers to the process of identifying and quantifying representative information or patterns from raw medical image data. Through feature extraction, high-dimensional raw image data can be transformed into low-dimensional image embedding features, thereby reducing data redundancy, highlighting key information, and providing more efficient input for subsequent model processing.

[0045] Image embedding features are numerical vectors obtained after feature extraction that can effectively represent the content and semantics of medical image data. These embedding features can be a compact representation of the original image data in a certain feature space, and can capture deep information such as the texture, shape, and structure of the image.

[0046] For example, traditional image processing algorithms, such as Gray-Level Co-occurrence Matrix (GLCM), can be used to extract texture features, or Scale-Invariant Feature Transform (SIFT) can be used to extract local keypoint features. These methods can extract visual information such as texture, shape, and edges from images and encode them into numerical vectors. As another implementation, a pre-trained Convolutional Neural Network (CNN) can be used as an encoder. The original image is input into the network, and the output of the intermediate layers of the network is used as the image embedding feature. This approach can automatically learn the deep semantic features of the image.

[0047] S220: Based on the pre-trained image strategy model, perform model inference on the image embedding features to obtain a candidate strategy set. The candidate strategy set includes multiple candidate operation strategies, which are used to represent the operation rules for processing medical image data.

[0048] An image strategy model is a pre-trained model that generates or selects a set of operational rules for processing medical image data based on the embedded features of the input image. This model can be a reinforcement learning agent, a generative network, or a complex decision-making system, with the goal of learning how to recommend the most appropriate processing strategy for different image data.

[0049] Model inference refers to the process of using a pre-trained image policy model to perform calculations based on new input data (i.e., image embedding features) to obtain prediction results or decision outputs (i.e., candidate policy sets).

[0050] A candidate strategy set refers to a set of potential operational rules generated or recommended by an imaging strategy model during a single inference process. This set includes a variety of different processing schemes to address the diversity and complexity of medical imaging data.

[0051] Candidate operation strategies refer to each specific operation rule in the set of candidate strategies. These rules can define a series of image processing steps, such as image enhancement, denoising, segmentation, registration, etc., as well as the parameter settings for these steps.

[0052] For example, the image strategy model can be a rule-based expert system that triggers predefined operational rules based on specific patterns in the image embedding features (such as features indicating the detection of a certain lesion), such as "if the image shows a high-density area, then contrast enhancement is recommended." Alternatively, the image strategy model can be a sequence generation model, such as a model based on a recurrent neural network (RNN) or Transformer. It receives image embedding features as input and generates a series of operational steps and their parameters, forming a complete processing flow. These operational strategies can be general image processing algorithms, such as Gaussian filtering, median filtering, histogram equalization, image segmentation algorithms, etc., along with the specific parameter settings of these algorithms.

[0053] S230: Send the candidate strategy set to multiple expert clients, which are used to evaluate the merits of the candidate operation strategies based on local data and obtain strategy scores.

[0054] Expert clients are computational entities that possess domain-specific knowledge and local medical imaging data. These clients can be hospitals, research institutions, or independent medical imaging analysis centers, and they are able to evaluate candidate operational strategies based on their own professional experience and data.

[0055] Local data refers to medical image datasets that are independently owned by the expert client and not shared with external parties. This data reflects the actual situation of a specific medical institution or disease area and serves as the basis for evaluating the effectiveness of operational strategies.

[0056] A strategy score is a quantitative evaluation given by an expert client after assessing candidate operational strategies based on local data. This score reflects the quality of a particular operational strategy within the local environment of a specific expert client.

[0057] For example, once the candidate strategy set is sent to the expert clients, each expert client can use its local medical image data as a test set and apply each candidate strategy sequentially. After applying the strategy, the expert client can quantify the effectiveness of each strategy using its local evaluation metrics (e.g., Dice coefficient or Jaccard index for segmentation tasks; peak signal-to-noise ratio (PSNR) for denoising tasks). Furthermore, local physicians or technicians can subjectively evaluate the processed images and provide scores. These evaluations are aggregated into a strategy score, reflecting the performance of each strategy on the specific client's local data.

[0058] S240: Update the model parameters of the image strategy model based on the strategy scores returned by multiple expert clients.

[0059] Model parameters refer to the set of numerical values ​​within an image strategy model that define its behavior and decision-making logic. By updating these parameters, the model's performance can be adjusted, enabling it to generate more optimized operational strategies.

[0060] For example, a simple averaging aggregation approach can be used to average the scores of all expert clients for the same candidate operational strategy, and then use the average score as the final evaluation of that strategy. Subsequently, the image strategy model can use these aggregated scores as feedback signals to adjust its internal parameters through optimization algorithms such as gradient descent, enabling it to generate or recommend higher-quality operational strategies in the future. As an alternative implementation, a more complex aggregation mechanism can be designed, for example, assigning different weights to the scores of each expert client based on their historical performance or data volume, and then performing weighted aggregation. The aggregated scores can be viewed as reward signals in reinforcement learning, guiding the image strategy model to optimize its strategies, thereby enabling it to better adapt to the processing needs of different types of medical image data.

[0061] The following example will provide a more detailed explanation of the above technical solution.

[0062] For example, a central server could be designed to develop a universal medical image preprocessing system that automatically recommends optimal preprocessing workflows for medical image data from different sources and modalities to improve the accuracy of subsequent diagnostic models. This server would collaborate with multiple expert clients (e.g., Hospital A, Hospital B, and Hospital C), each possessing different types and distributions of medical image data.

[0063] First, when the central server receives a batch of chest X-ray images to be processed, it performs feature extraction on the medical image data. Specifically, the server can use a pre-trained deep learning model (e.g., a ResNet encoder) to transform these X-ray images into compact image embedding features. These embedding features capture key visual patterns in the images that are relevant to pathological information.

[0064] Next, the server performs model inference on these image embedding features based on a pre-trained image policy model. Upon receiving the image embedding features, the image policy model (e.g., a reinforcement learning-based agent) generates a set of candidate policies. This set may contain various preprocessing operation sequences, such as: "Policy 1: Gaussian blur (sigma=1.5), then Local Histogram Equalization (CLAHE)", "Policy 2: Median filtering (kernel size 3x3), then gamma correction (gamma=0.8)", and "Policy 3: Direct Global Histogram Equalization". These candidate operation strategies aim to address potential problems in X-ray images, such as noise and insufficient contrast.

[0065] Subsequently, the central server sends this set of candidate strategies to multiple expert clients. For example, Hospital A focuses on lung disease diagnosis, Hospital B focuses on orthopedic image analysis, and Hospital C focuses on pediatric imaging. Each expert client, upon receiving the set of candidate strategies, evaluates their merits based on its local data. For instance, Hospital A uses its local chest X-ray dataset as a test set to apply each candidate strategy. After applying the strategy, Hospital A can use its internal lung lesion detection model to detect the processed images and generate a strategy score for each strategy based on the model's performance (e.g., accuracy and recall). Similarly, Hospitals B and C also generate their own strategy scores based on their local data and evaluation criteria.

[0066] Finally, the central server receives strategy scores from Hospital A, Hospital B, and Hospital C. These scores reflect the performance of each candidate strategy in different expert client local environments. Based on these scores, the central server updates the model parameters of the imaging strategy model. For example, the server can aggregate scores from different hospitals using a weighted average, and then use the aggregated score as a reward signal to adjust the parameters of the imaging strategy model through a reinforcement learning algorithm. In this way, the imaging strategy model can learn which strategies perform better in the feedback from different expert clients, enabling it to generate more effective and adaptable processing strategies for diverse medical imaging data in the future.

[0067] The technical solution proposed in this application, through the above examples, clearly demonstrates its advantages in addressing the technical challenges of medical image data processing in federated learning. Traditional federated learning methods often struggle to achieve effective global model convergence when faced with highly heterogeneous medical image data, because a single aggregation strategy cannot adapt to the diverse characteristics of local data. For example, if the traditional federated averaging algorithm is used to directly aggregate model parameters from different hospitals, the model may not achieve optimal performance in any specific domain due to the significant differences in data distribution among hospitals A, B, and C. However, this application generates candidate operation strategies and allows them to be evaluated locally by an expert client. This enables the central model to learn from diverse feedback without directly processing heterogeneous raw data or model parameters, thereby effectively improving the model's convergence ability and generalization performance in heterogeneous environments.

[0068] Furthermore, traditional federated learning methods typically require hundreds of rounds of gradient or model parameter transmission between the server and client, resulting in significant communication overhead. For example, each iteration may require the transmission of millions or even hundreds of millions of model parameters. In contrast, the technical solution presented in this application significantly reduces this communication burden. In the example above, the central server only transmits a set of candidate policies containing a small number of operational rules, while the expert client only returns policy scores. These data volumes are far smaller than those transmitted when transmitting complete model gradients or model parameters. This lightweight communication model greatly improves the overall efficiency and scalability of the system, especially in medical environments where network bandwidth is limited. Therefore, this application effectively solves the problems of model convergence difficulties and high communication overhead in medical image data processing during federated learning through a collaborative policy evaluation and model update mechanism, providing a new approach for intelligent medical image processing.

[0069] Figure 3 A flowchart illustrating the processing of medical image data based on a selected expert client is shown in one embodiment of this application. Figure 3 As shown, the method may include the following steps S310 to S360.

[0070] S310: Extract features from the medical image data to be processed to obtain the image embedding features of the medical image data.

[0071] S320: Based on the pre-trained image strategy model, perform model inference on the image embedding features to obtain a candidate strategy set. The candidate strategy set includes multiple candidate operation strategies, which are used to represent the operation rules for processing medical image data.

[0072] S330: Obtain domain embedding features sent by multiple expert clients from the candidate client set. These domain embedding features are used to represent the knowledge domain corresponding to the local data of the expert clients.

[0073] S340: Select multiple expert clients from the candidate client set that match the image embedding features based on the domain embedding features.

[0074] S350: Send the candidate strategy set to multiple expert clients, which use local data to evaluate the merits of the candidate operation strategies and obtain strategy scores.

[0075] S360: Update the model parameters of the image strategy model based on the strategy scores returned by multiple expert clients. The strategy scores are used to represent the evaluation of the merits of the candidate operation strategies by the expert clients.

[0076] Domain embedding features are low-dimensional vector representations that capture and quantify the core semantic information of a specific knowledge domain. They are generated through the analysis and learning of local data from expert clients, thus abstractly representing the professional domains in which the client excels. Specifically, domain embedding features can be trained on local data from expert clients (such as textual or image data like medical image reports, diagnostic records, and treatment plans) using deep learning models, such as autoencoders, word embedding models, or graph neural networks. This maps the high-dimensional raw data into a low-dimensional vector space, making semantically similar domains closer together in the embedding space. Furthermore, domain embedding features can also be implemented through expert knowledge engineering, where domain experts manually define or annotate key domain concepts, construct domain knowledge representations using ontology or knowledge graph technologies, and then convert them into vector form.

[0077] After obtaining the domain embedding features, multiple expert clients matching the image embedding features can be selected from the candidate client set based on these features. This step aims to ensure that only expert clients whose professional fields are highly relevant to the medical image data being processed are selected for subsequent strategy evaluation, thereby improving the accuracy and efficiency of the evaluation. One approach is to calculate the similarity between the image embedding features and the domain embedding features of each expert client, such as cosine similarity or the reciprocal of the Euclidean distance, and then set a similarity threshold, selecting expert clients with similarities higher than this threshold. Alternatively, a ranking mechanism can be used, sorting all candidate expert clients in descending order of the similarity between their domain embedding features and image embedding features, and then selecting the top N expert clients, where N can be a preset fixed number or dynamically adjusted according to actual needs.

[0078] This application optimizes the expert client selection process by introducing domain embedding features and a matching selection mechanism. Specifically, before sending the candidate strategy set to the expert clients, the system first obtains the domain embedding features sent by each expert client in the candidate client set. These features accurately characterize their knowledge domain expertise based on local data. Simultaneously, the system already possesses the image embedding features of the medical image data to be processed. By comparing the image embedding features with the domain embedding features of each expert client, the system can intelligently filter out expert clients that highly match the current medical image data domain. Only these selected expert clients receive the candidate strategy set and perform strategy scoring. This pre-screening mechanism ensures that subsequent strategy scoring comes from truly relevant experts, thereby making the update of the image strategy model more accurate and effective, and avoiding bias and resource waste that may result from irrelevant experts participating in the evaluation.

[0079] The following is a specific example to illustrate this.

[0080] For example, a central server needs to process CT image data about lung nodules. First, the server extracts features from the lung CT image data to obtain image embedding features representing "lung nodule CT image". Simultaneously, the system maintains a candidate client set containing multiple expert clients, such as those from radiology and pulmonology departments of different hospitals. Each expert client sends its domain embedding features to the server; for example, hospital A's domain embedding features might represent "lung disease diagnosis, CT image analysis", while hospital B's might represent "cardiac disease diagnosis, MRI image analysis". After receiving these domain embedding features, the server calculates the similarity between the image embedding features of "lung nodule CT image" and the domain embedding features of each expert client. For example, hospital A's domain embedding features have a high similarity to the image embedding features, while hospital B's have a low similarity. Based on this, the server selects expert clients, such as hospital A, that match the domain of "lung nodule CT image" and sends a set of candidate strategies to these selected expert clients for evaluation.

[0081] Through the above technical solution, this application ensures that strategy scoring is provided by expert clients whose knowledge domain is highly relevant to the medical imaging data, significantly improving the accuracy and relevance of the feedback. Simultaneously, it avoids sending candidate strategy sets to irrelevant experts, effectively reducing unnecessary communication overhead and computational resource consumption. More importantly, this precise expert client selection mechanism enables the iterative learning process of the imaging strategy model to obtain high-quality, highly relevant expert feedback, thereby accelerating model convergence and ultimately achieving superior model performance and generalization ability, especially when processing highly heterogeneous medical imaging data, where its advantages are even more pronounced.

[0082] Figure 4 A flowchart illustrating the selection of an expert client is shown in one embodiment of this application. For example... Figure 4 As shown, based on the above embodiments, step S340 may further include the following steps S342 to S344.

[0083] S342: Obtain the semantic similarity between domain embedding features and image embedding features.

[0084] In practical implementation, obtaining the semantic similarity between domain embedding features and image embedding features aims to quantify the semantic similarity between the expert client's knowledge domain (represented by domain embedding features) and the medical image data to be processed (represented by image embedding features). This provides an objective and computable metric to avoid subjective judgment and ensure the accuracy of subsequent expert client selection. This semantic similarity can be achieved in several ways. For example, the cosine similarity between two embedding feature vectors can be calculated. Cosine similarity measures the directional consistency of two vectors in multidimensional space; the closer the value is to 1, the more semantically similar they are. Alternatively, it can be achieved by calculating the reciprocal of the Euclidean distance or an exponential decay function between the two embedding feature vectors. Euclidean distance measures the linear distance between vectors in multidimensional space; the smaller the distance, the more semantically similar they are. Furthermore, a specially trained similarity measurement model can be used, taking two embedding features as input and outputting their semantic similarity score.

[0085] S344: Select multiple expert clients from the candidate client set that match the image embedding features based on semantic similarity.

[0086] This step, based on the quantitative similarity values ​​obtained in the previous step, filters out the clients most relevant to the current medical imaging data from all potential expert clients. Its purpose is to ensure that only expert clients whose professional domains are highly aligned with the current task participate in subsequent strategy evaluation, thereby improving the accuracy and efficiency of the evaluation and avoiding interference from irrelevant clients. Specifically, a preset semantic similarity threshold can be set, and all expert clients with semantic similarity higher than this threshold are selected; alternatively, candidate clients can be sorted in descending order based on semantic similarity, and the top N expert clients can be selected, where N is a preset number of clients; furthermore, dynamic thresholds or clustering-based methods can be used, for example, clustering clients with high semantic similarity into one class and selecting representative clients from among them, or adaptively determining the selection range based on the similarity distribution.

[0087] In processing medical image data, this embodiment first extracts features from the medical image data to obtain image embedding features. Simultaneously, the system acquires the domain embedding features of each expert client in the candidate client set. These domain embedding features characterize the knowledge domains that the expert clients excel in or the characteristics of their local data.

[0088] To ensure the accuracy and efficiency of subsequent strategy evaluation, this application embodiment introduces a precise expert client screening mechanism before sending the candidate strategy set to the expert client.

[0089] Specifically, the system calculates the semantic similarity between the domain embedding features of each expert client and the image embedding features of the current medical imaging data. This semantic similarity quantifies the relevance between the expert client's professional domain and the current medical imaging task. In this way, the professional matching degree of each expert client can be objectively evaluated. Subsequently, based on these calculated semantic similarities, the system selects multiple expert clients from the candidate client set that highly match the image embedding features. For example, a similarity threshold can be set, or several clients with the highest similarity can be selected. In this way, only expert clients that are highly relevant to the current medical imaging data in their knowledge domain are selected to participate in subsequent strategy evaluation, thereby ensuring the professionalism and reliability of the strategy evaluation, avoiding interference from irrelevant clients, and improving the efficiency and accuracy of the overall processing flow.

[0090] By introducing semantic similarity to accurately match expert clients, the embodiments of this application ensure that subsequent strategy scoring by these expert clients is more targeted and professional. When updating the model parameters of the image strategy model, based on these high-quality strategy scores, the image strategy model can learn and optimize more effectively, thereby improving the model's performance and accuracy when processing specific medical image data. This precise matching mechanism enables the entire medical image data processing method to achieve more intelligent and efficient collaborative learning when facing heterogeneous data and diverse expert domains.

[0091] As a specific implementation method, when obtaining the semantic similarity between domain embedding features and image embedding features, a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model or a Word2Vec model can be used to generate domain embedding features and image embedding features. These embedding features are high-dimensional vectors that can capture their semantic information. Subsequently, their semantic relevance is quantified by calculating the cosine similarity between the two vectors. For example, if the image embedding feature represents "nodule detection in lung CT images," while the domain embedding feature of an expert client represents "respiratory disease diagnosis" or "tumor image analysis," their cosine similarity will be high. When selecting expert clients based on semantic similarity, a cosine similarity threshold can be set, such as 0.7. All expert clients with a cosine similarity greater than 0.7 with the image embedding feature are considered matching expert clients and selected. Alternatively, the top 5 expert clients with the highest cosine similarity can be directly selected to ensure that a sufficient number of highly relevant experts participate in the subsequent strategy evaluation. These selected expert clients will then receive a set of candidate strategies and be evaluated based on their local expertise and data.

[0092] Through the above technical solution, this application can accurately quantify the degree of matching between the knowledge domain of expert clients and the medical image data to be processed, effectively solving the problems of inaccurate or inefficient matching quantification in traditional methods. Screening based on semantic similarity ensures that only expert clients whose professional domains are highly compatible with the current task participate in subsequent strategy evaluation, thereby significantly improving the accuracy and reliability of strategy evaluation. This not only avoids interference from irrelevant clients and improves the efficiency of the overall processing flow, but also provides more targeted and high-quality feedback for optimizing the image strategy model, ultimately improving the overall performance and intelligence level of medical image data processing.

[0093] In one embodiment of this application, the policy scoring includes differential privacy noise added by the expert client, the differential privacy noise following a Gaussian distribution.

[0094] Specifically, the strategy score is the result generated by the expert client after evaluating candidate operation strategies based on its local data, directly reflecting the expert client's judgment on the merits of the strategies. To protect the privacy of the expert client's local data, introducing differential privacy noise into the strategy score is a key technical approach. The addition of differential privacy noise aims to obscure the direct correlation between the strategy score and the original sensitive data, making it difficult for attackers to deduce the expert client's local data information by analyzing the strategy score.

[0095] After calculating the original policy score, the expert client can utilize differential privacy mechanisms to add noise to the original score by randomizing the response or output perturbation. For example, a Laplace or Gaussian mechanism can be used, with the noise intensity determined according to a preset privacy budget, ensuring that the utility of the score is preserved as much as possible while meeting a certain level of privacy protection. Another approach is to incorporate differential privacy protection into the internal calculation process of generating the policy score within the expert client. For example, noise can be introduced when aggregating local data or calculating intermediate statistics, making the final policy score inherently privacy-preserving. This can be achieved by adding noise during gradient calculation or model updates, thereby indirectly affecting the final policy score.

[0096] The Gaussian distribution (i.e., the normal distribution) is a common probability distribution characterized by data points being symmetrically distributed around the mean, with data points having lower probabilities the further they are from the mean. In differential privacy, the Gaussian distribution is chosen as the noise source because of its favorable mathematical properties, such as ease of analysis, controllable variance, and scalability in multidimensional spaces. The introduction of Gaussian noise effectively perturbs the policy score, making it statistically indistinguishable from the original score, thus achieving the goal of privacy protection.

[0097] Expert clients can generate random numbers with a mean of 0 and a standard deviation of a specific value, following a Gaussian distribution. These random numbers are then superimposed onto the original policy score. The magnitude of the standard deviation is typically calculated based on privacy budget and sensitivity to ensure that the required level of privacy protection is met while minimizing the impact on the score's utility. In some scenarios, the generation of Gaussian noise can be combined with encryption techniques. For example, expert clients can use homomorphic encryption to add noise to the policy score in an encrypted state, or in secure multi-party computation protocols, a distributed random number generator can collaboratively generate Gaussian-distributed noise and securely add it to the policy score to further enhance privacy and security.

[0098] After the expert client evaluates the merits of candidate operational strategies based on local data and obtains a strategy score, it proactively adds differential privacy noise to the strategy score before sending it out. This noise prevents external observers from accurately deducing the expert client's original local data information from the perturbed strategy score, thus effectively protecting the privacy of medical imaging data.

[0099] The added differential privacy noise follows a Gaussian distribution. The choice of Gaussian distribution as the noise source is based on its mathematical properties, which are widely used in the field of privacy protection. Gaussian noise provides quantifiable levels of privacy protection, and by adjusting its standard deviation, a balance can be struck between the level of privacy protection and the usability of the policy score. This means that while protecting the privacy of local data on the expert client, the policy score can still effectively reflect the true merits of candidate operational strategies, thereby ensuring that the image policy model can be effectively updated based on these protected scores.

[0100] In medical image data processing, the policy scores returned by expert clients are crucial inputs for updating the model parameters of the image strategy model. By incorporating differential privacy noise into the policy scores, this solution enhances the privacy and security of the entire system without altering the overall model update process. Even during collaborative parameter updates across multiple expert clients, the local data privacy of each client is effectively protected, thus promoting the healthy development of federated learning models for medical image data processing. This approach enables federated learning to better meet data privacy and security compliance requirements when applied in the medical field, encouraging more medical institutions to participate in collaborative learning and jointly improve the intelligence level of medical image processing.

[0101] As a specific implementation method in medical image data processing, after the expert client completes the evaluation of candidate operational strategies and generates an original strategy score, the client will activate a privacy protection module. This module will calculate the standard deviation of the required Gaussian noise based on a preset privacy budget and the sensitivity of the strategy score. For example, if the original strategy score is a real value, the expert client can use a pseudo-random number generator to generate a random number that follows a Gaussian distribution with a mean of 0 and a standard deviation of a specific value. This random number is then superimposed on the original strategy score to form a perturbed strategy score. For example, if the original strategy score is 0.85 and the generated Gaussian noise is 0.03, the final sent strategy score will be 0.88. The above technical means can be implemented with reference to the following example: The expert client can use a random number generation function in a programming language to generate a single Gaussian noise value, where the mean parameter is set to 0 and the standard deviation parameter is set to the calculated standard deviation. Then, this noise value is added to the original strategy score to obtain the final privacy-protected strategy score. This perturbed strategy score is then sent to the server to update the model parameters of the image strategy model. In this way, even if the server receives a perturbed score, it is difficult to deduce the original medical image data from the expert client from that score, thus effectively protecting data privacy.

[0102] By adding differential privacy noise to the policy scoring, and ensuring that this noise follows a Gaussian distribution, this scheme effectively blurs the direct correlation between the policy scoring and sensitive local data on the expert client. This significantly reduces the risk of inferring original medical image data through policy scoring, thereby solving the problem of medical data privacy leakage in federated learning environments. Therefore, this application's embodiments, while ensuring continuous model updates for the medical image data processing method, greatly enhance privacy security during data transmission. This not only improves the compliance and credibility of the entire system but also encourages more medical institutions to participate in collaborative intelligent learning while protecting patient privacy, jointly promoting the improvement of medical image diagnosis and treatment.

[0103] In one embodiment of this application, before updating the model parameters of the image strategy model, monotonic affine calibration can be performed on the strategy scores returned by multiple expert clients according to preset scale calibration parameters to obtain a strategy score with uniform dimensions; the scale calibration parameters are trained based on historical data returned by expert clients for a common reference task, which is a general knowledge task applicable to multiple knowledge domains.

[0104] The preset scaling parameters are a set of values ​​used to adjust the scale of expert client strategy scoring, aiming to eliminate the influence of differences in scoring habits or standards among different expert clients. These parameters are not randomly set, but are derived through analysis and learning of the historical performance of expert clients on specific tasks.

[0105] For example, this parameter could include the rating offset and rating scaling factor for each expert client, used to map the original ratings to a uniform scale. Alternatively, this parameter could be a set of nonlinear transformation functions that learn the historical rating distribution of expert clients and map it to a standard normal distribution or other pre-defined uniform distribution.

[0106] Monotonic affine calibration is a mathematical transformation process that essentially involves a linear transformation of the original strategy score. This is achieved by multiplying by a scaling factor and adding an offset to adjust the score. This calibration method preserves the relative order of the scores while altering the absolute numerical range and center point of the scores, thus mapping scores from different expert clients to the same dimension or scale. For example, calibration can be performed using the formula S_calibrated = a × S_original + b, where a and b are the scale calibration parameters. Another approach is to use a piecewise linear function for calibration, applying different affine transformations within different score intervals to more precisely fit the scoring characteristics of expert clients, while still maintaining overall monotonicity.

[0107] A unified-dimensional strategy score is a strategy score obtained after monotonically affine calibration. Its characteristic is that all expert clients' scores fall within a common, comparable numerical range or distribution. This means that evaluations of the same operational strategy by different expert clients can be directly compared, aggregated, or weighted after calibration without bias caused by inconsistencies in the original score dimensions. For example, all scores may be calibrated to the interval between 0 and 1, or to a standard distribution with a mean of 0 and a variance of 1.

[0108] The scaling calibration parameters are trained using historical data returned by expert clients for common reference tasks. For example, the system collects historical data on strategy scores given by each expert client when completing a series of predefined "common reference tasks." This historical data reflects the evaluation habits and preferences of expert clients on standard tasks. By performing statistical analysis or machine learning training on this historical data, scaling calibration parameters specific to each expert client can be derived. For instance, least squares or regression analysis can be used to fit each expert client's historical scores to the true or standard scores of the common reference tasks, thereby obtaining the slope and intercept of the affine transformation. Another training method is to use Bayesian inference or maximum likelihood estimation to estimate parameters that can unify these distributions based on the distribution of expert client scores on common reference tasks.

[0109] Public reference tasks are general tasks applicable across multiple knowledge domains. They are tasks that do not rely on specific professional domain knowledge but are universally applicable to all or most expert clients, possessing clear evaluation criteria or known optimal solutions. These tasks are designed to provide a neutral, standardized evaluation benchmark to fairly assess the scoring behavior of different expert clients. For example, in the field of medical imaging, public reference tasks might be common image recognition tasks with well-defined diagnostic criteria, such as identifying normal anatomical structures and distinguishing between typical cases of benign and malignant tumors. The evaluation results of these tasks are relatively objective and easily reach a consensus. Another type of public reference task can be the processing and evaluation of synthetic image data with known properties, thereby ensuring the objectivity and quantifiability of the evaluation results.

[0110] The proposed solution introduces a key preprocessing step—monotonically affine calibration—before updating the model parameters of the image strategy model based on strategy scores returned by multiple expert clients.

[0111] Its operating logic is that, firstly, the system will pre-acquire and store the historical strategy scoring data generated by each expert client when performing public reference tasks.

[0112] These public reference tasks are carefully designed as general tasks applicable to multiple knowledge domains, ensuring the objectivity and comparability of their evaluation results. Based on this historical data, the system can train and obtain scale calibration parameters unique to each expert client. These parameters essentially capture the differences in scoring habits, scales, and preferences among different expert clients.

[0113] Upon receiving policy scores from different expert clients, these raw scores are not directly used to update the model parameters. Instead, the system performs monotonic affine calibration on each expert client's policy score using pre-trained scale calibration parameters. This calibration process, through linear transformation, effectively adjusts the absolute numerical range and center point of the scores while maintaining the relative order of the original scores, thereby unifying the scores from all expert clients to the same dimension and scale. After calibration, the inconsistency in scoring caused by differences in knowledge domains or evaluation habits is resolved, and all policy scores become policy scores with a unified dimension, allowing for direct and fair comparison and integration. Subsequently, these unified policy scores are used to update the model parameters of the image policy model.

[0114] In this way, model parameter updates are no longer affected by inconsistent scoring dimensions, enabling more accurate and reliable reflection of expert client evaluations, thereby effectively optimizing the performance of the image strategy model. This calibration mechanism is closely integrated with the basic image strategy model update process, eliminating heterogeneity before data aggregation. This allows the subsequent model learning process to receive high-quality, standardized feedback signals, significantly improving the stability and convergence of model training and ensuring that the model can effectively learn from multi-source expert knowledge.

[0115] In one specific implementation, consider three expert clients, A, B, and C. When evaluating operational rules for processing medical image data, they may employ different scoring standards or habits. For example, client A might habitually give higher scores, client B might tend to give lower scores, and client C's scoring range might be narrower. To address this inconsistency in metric, this application can first define a series of common reference tasks. For instance, identifying a set of standardized medical images containing common lesions (such as pulmonary nodules or fractures) and requiring expert clients to score preset diagnostic operational strategies. The system collects historical scoring data from clients A, B, and C on these common reference tasks. For example, client A's historical scores might be concentrated in the 80-100 range, client B in the 60-80 range, and client C in the 70-90 range. The system can utilize this historical data to train a unique scaling parameter for each client through regression analysis. For example, for client A, the scaling parameter might be a linear transformation mapping high-score intervals to standard intervals; for client B, it might be a linear transformation boosting low-score intervals to standard intervals. When new candidate operational strategies are sent to these expert clients and strategy scores are returned (e.g., client A returns 95, client B returns 70, and client C returns 85), the system performs monotonic affine calibration on these scores using their respective preset scale calibration parameters. For example, by applying the formula S_calibrated = a × S_original + b, all scores are calibrated to a uniform dimension of 0-100. After calibration, client A's 95 might become 88, client B's 70 might become 85, and client C's 85 might become 86. In this way, scores that initially appear to differ significantly become directly comparable under a uniform dimension, thus providing accurate and reliable input for subsequent parameter updates of the image strategy model.

[0116] By employing the aforementioned technical solution, this application introduces a step of monotonic affine calibration of the strategy scores returned by expert clients before updating the model parameters of the image strategy model. This calibration process utilizes preset scale calibration parameters trained based on historical data from a common reference task, effectively eliminating the problem of inconsistent strategy score dimensions caused by differences in knowledge domains or evaluation habits among different expert clients. This allows heterogeneous dimensional scores, which were previously impossible to directly compare and integrate, to be converted into strategy scores with unified dimensions, thereby ensuring that the feedback signals upon which subsequent model parameter updates are based are accurate and comparable. Therefore, the solution of this application significantly improves the accuracy and reliability of the image strategy model learning from multi-source expert knowledge, avoids model performance degradation or convergence difficulties caused by scoring bias, and enables the image strategy model to more effectively adapt to diverse local data characteristics, thereby achieving superior overall performance.

[0117] In one embodiment of this application, before updating the model parameters of the image strategy model, the average and standard deviation of multiple strategy scores returned by the same expert client can be obtained; based on the average and standard deviation, the strategy scores returned by the same expert client are converted into standardized scores.

[0118] By statistically analyzing the historical rating behavior of specific expert clients, we can capture the central tendency and dispersion of their rating distribution. Obtaining the average reflects the overall level of the expert client's ratings, while the standard deviation reflects the range of fluctuation. These statistics form the basis for subsequent rating standardization. The average and standard deviation can be calculated locally by the expert client and uploaded to the server, or the server can receive all strategy ratings, group them according to the expert client's identifier, and then calculate the average and standard deviation separately for each expert client's rating set.

[0119] Based on this, policy scores returned by the same expert client are converted into standardized scores according to the mean and standard deviation. The purpose of this step is to eliminate inconsistencies in scoring scales between different expert clients due to differences in scoring habits and scale usage, allowing policy scores from different expert clients to be compared and aggregated under a unified scale. One implementation method is to use Z-score standardization, which involves subtracting the mean from that expert client's score for each original policy score and then dividing by that expert client's standard deviation, thus converting the score into a standardized score with a mean of 0 and a standard deviation of 1. Another implementation method is to use T-score standardization, which maps the Z-score standardized score to a scale with a mean of 50 and a standard deviation of 10 through a linear transformation. This conversion makes the scores from different expert clients comparable, thus providing more reliable input for subsequent model parameter updates.

[0120] This application's embodiments address the issue of inconsistent scoring scales among different expert clients by introducing a scoring standardization mechanism, ensuring that model updates are based on uniform and comparable data. After feature extraction from medical image data to obtain image embedding features, the image strategy model performs inference to generate a set of candidate strategies. These candidate strategy sets are sent to multiple expert clients for evaluation, with each expert client returning a strategy score based on its local data. Before using these strategy scores to update the model parameters of the image strategy model, this application's embodiments first standardize the strategy scores returned by each expert client. Specifically, the system obtains the average and standard deviation of multiple strategy scores returned by the same expert client over a period of time or for a series of tasks, and then uses these statistics to convert each original strategy score from that expert client into a standardized score. This processing method ensures that scores from different expert clients can be fairly compared and utilized when updating model parameters, avoiding bias introduced by differences in scoring scales, thereby improving the accuracy and stability of model training.

[0121] As a specific implementation, a medical image processing system may include multiple expert clients, such as those from different hospitals or departments. When the system generates a set of candidate operational strategies and sends them to these expert clients for evaluation, each expert client scores these strategies based on its internal standards and data. For example, expert client A may tend to give a higher score, while expert client B may tend to give a lower score, even if they agree on the relative merits of the same strategy. To eliminate this individual scoring scale difference, before updating the model parameters of the image strategy model, the system collects multiple strategy scores from its historical or current batch for each expert client. For example, for expert client A, the system calculates the mean (μA) and standard deviation (σA) of all its strategy scores. Then, for each strategy score X returned by expert client A, the system converts it into a standardized score X' = (X - μA) / σA. Similarly, for expert client B, the system also calculates its mean (μB) and standard deviation (σB) and converts its score Y into a standardized score Y' = (Y - μB) / σB. In this way, the strategy scores of all expert clients are converted to a unified standardized scale, enabling subsequent updates to image strategy model parameters to be based on more comparable and consistent data.

[0122] Through the above technical solution, the embodiments of this application effectively solve the problem of individual scoring scale differences or biases that may exist when different expert clients evaluate candidate operation strategies. By performing personalized standardization processing on the strategy scores of each expert client, the absolute bias of the scores is eliminated, and only the relative superiority or inferiority information between candidate operation strategies is retained. This enables a fair comparison and utilization of evaluation information from different expert clients when updating the model parameters of the image strategy model, thereby significantly improving the accuracy and stability of model training and avoiding the degradation of model performance caused by inconsistent scoring scales.

[0123] Figure 5 A flowchart illustrating the updating of model parameters for an image strategy model in one embodiment of this application is shown. Figure 5 As shown, based on the above embodiments, step S240 (or step S360) may further include the following steps S242 to S246.

[0124] S242: Obtain the reliability weight corresponding to the expert client. The reliability weight is used to represent the accuracy of the expert client's evaluation of historical tasks.

[0125] Reliability weighting is a key indicator for measuring the trustworthiness of evaluation information provided by expert clients. Its implementation can be, but is not limited to, the following: One approach is to calculate it by analyzing the consistency or accuracy rate between the expert client's evaluation results and actual results (e.g., the "gold standard" verified by authoritative experts or the consensus of most experts) in similar tasks completed in the past. For example, the percentage of correct evaluations by the expert client in historical tasks can be statistically analyzed, or the error between its evaluation and the standard answer can be calculated. Another approach is to assign initial weights based on a pre-defined expert ranking system, including the expert client's professional qualifications, years of experience, and activity level in a specific field, and then dynamically adjust these weights based on its subsequent evaluation performance.

[0126] S244: For the same candidate operation strategy, the strategy scores returned by different expert clients are weighted and mapped according to the reliability weight to obtain the comprehensive score of the candidate operation strategy.

[0127] Weighted mapping aims to integrate evaluations from different expert clients while taking into account their differences in reliability. Specifically, a weighted average method can be used, multiplying the strategy score returned by each expert client by its corresponding reliability weight, then summing all weighted scores, and finally dividing by the sum of all reliability weights to obtain a comprehensive score. For example, if expert client A has a reliability weight of W_A and its strategy score is S_A; and expert client B has a reliability weight of W_B and its strategy score is S_B, then the comprehensive score can be calculated as (W_A × S_A + W_B × S_B) / (W_A + W_B).

[0128] S246: Reinforcement learning is performed on the impact strategy model based on the comprehensive score to update the model parameters of the image strategy model.

[0129] Reinforcement learning is a machine learning paradigm that learns optimal behavioral policies through interaction with the environment. Here, the overall score is used as a reward signal in reinforcement learning to guide the image policy model in updating its parameters. Specifically, when the image policy model generates a candidate action policy and it is evaluated by an expert client, the resulting overall score serves as the "reward" for that policy. Reinforcement learning algorithms (e.g., policy gradient algorithms or Q-learning algorithms) adjust the model's internal parameters based on this reward signal, making the model more likely to generate policies that achieve higher overall scores in the future.

[0130] This application addresses the issue of inconsistent expert client rating reliability by introducing a reliability weighting mechanism and reinforcement learning, ensuring more accurate model updates. After feature extraction from medical image data and model inference based on the image strategy model to obtain a candidate strategy set, this set is sent to multiple expert clients to obtain strategy scores. Furthermore, this application obtains reliability weights corresponding to each expert client, reflecting the accuracy of the expert clients' evaluations in historical tasks. For the same candidate operation strategy, these reliability weights are used to weight and map the strategy scores returned by different expert clients, resulting in a more objective and reliable comprehensive score. This comprehensive score effectively integrates the opinions of different experts while reducing the impact of unreliable evaluations. Subsequently, this reliability-weighted comprehensive score is used as a reward signal for reinforcement learning to train and update the parameters of the image strategy model. In this way, the image strategy model can learn from high-quality feedback, gradually optimizing its ability to generate strategies, thereby providing better and more accurate operation rules in subsequent medical image data processing.

[0131] The following is a specific example to illustrate this.

[0132] Suppose that in a medical image data processing task, the image strategy model generates a candidate operation strategy and sends it to three expert clients for evaluation. Expert client 1 returns a strategy score of 80, expert client 2 returns a strategy score of 70, and expert client 3 returns a strategy score of 60. In previous tasks, the system has calculated the reliability weight of expert client 1 as 0.9, expert client 2 as 0.7, and expert client 3 as 0.5. To obtain the comprehensive score of the candidate operation strategy, the system performs a weighted mapping: Comprehensive score = (80 × 0.9 + 70 × 0.7 + 60 × 0.5) / (0.9 + 0.7 + 0.5) = (72 + 49 + 30) / 2.1 = 151 / 2.1 ≈ 71.9 points. This comprehensive score of 71.9 points will be used as a reward signal and input into the reinforcement learning module of the image strategy model. Reinforcement learning algorithms adjust the neural network parameters of the image policy model based on this reward value, enabling the model to achieve a higher overall score when generating similar policies in the future. For example, if a policy gradient algorithm is used, the model will increase the probability of generating that policy based on this positive reward signal, or adjust the policy parameters to make it closer to the optimal policy.

[0133] Through the above technical solution, the embodiments of this application can effectively solve the problem of inconsistent accuracy of evaluations from different expert clients. By introducing reliability weights and performing weighted mapping, the negative impact of low-reliability expert evaluations can be effectively filtered or weakened, ensuring that the feedback signals used for model updates are more accurate and reliable. Based on this, utilizing reinforcement learning mechanisms, the image strategy model can continuously learn and optimize from these high-quality comprehensive scores, thereby significantly improving the model's performance and robustness in medical image data processing tasks and avoiding performance degradation or convergence difficulties caused by differences in expert evaluations.

[0134] In one embodiment of this application, before performing reinforcement learning on the influence strategy model based on the comprehensive score, the mean and standard deviation of the comprehensive scores of different candidate operation strategies can be obtained; the comprehensive scores of different candidate operation strategies are then subjected to baseline subtraction and normalization based on the mean and standard deviation.

[0135] Before applying reinforcement learning to the image strategy model, it is necessary to first obtain the mean and standard deviation of the comprehensive scores for different candidate operational strategies. The mean is a statistic used to represent the central tendency of a set of data, and can be obtained by summing all the comprehensive scores and dividing by the number of scores. The standard deviation measures the dispersion of the data, i.e., the average degree to which data points deviate from the mean, and can be obtained by calculating the square root of the mean of the squares of the differences between each score and the mean. These statistics quantify the overall distribution characteristics of the comprehensive scores, providing a statistical basis for subsequent data processing.

[0136] Based on this, the composite scores of different candidate operational strategies are subjected to baseline subtraction and normalization according to the obtained mean and standard deviation. Baseline subtraction involves subtracting the mean from each composite score, thereby shifting the center point of the data to zero. Normalization involves dividing the result after baseline subtraction by the standard deviation to unify the variance of the data to 1. For example, Z-score standardization can be used. Through this process, all composite scores will be transformed into standardized scores with zero mean and unit variance, thereby eliminating scaling differences and distribution inconsistencies in scores between different strategies.

[0137] This application addresses the learning instability caused by inconsistent comprehensive scoring scales by introducing a scoring standardization step. After weighting and mapping the policy scores returned by different expert clients based on policy scores from multiple expert clients and incorporating reliability weights to obtain comprehensive scores for candidate operation policies, these comprehensive scores serve as input for reinforcement learning. However, directly using the raw comprehensive scores can lead to inconsistent sensitivity of the reinforcement learning algorithm to high and low scores due to potentially large differences in numerical range and unbiased distribution, thus affecting the stability and convergence efficiency of the learning process. Therefore, before reinforcement learning, this application first obtains the mean and standard deviation of the comprehensive scores for all candidate operation policies. The mean provides information on the central trend of the scores, while the standard deviation reflects the range of fluctuation. Based on this statistical information, baseline subtraction and normalization are performed on each comprehensive score. Baseline subtraction adjusts the center of all scores to zero, eliminating the influence of absolute numerical magnitude; normalization unifies the range of fluctuation in the scores, allowing scores of different policies to be compared on the same scale. After this preprocessing, the input received by the reinforcement learning algorithm will be standardized scores with uniform scale and distribution characteristics. This allows the algorithm to more stably evaluate the relative merits of different strategies, reduces interference caused by differences in the original scores, and thus improves the efficiency of reinforcement learning and the accuracy of model parameter updates.

[0138] As a specific implementation method, suppose that when processing medical image data, three candidate operation strategies A, B, and C are obtained through image strategy model inference. After evaluation by expert clients and weighting by reliability weights, their comprehensive scores are obtained respectively. For example, the comprehensive score of strategy A is 85 points, the comprehensive score of strategy B is 60 points, and the comprehensive score of strategy C is 95 points. Before using these comprehensive scores for reinforcement learning, their mean and standard deviation are first calculated. Assume that the calculated mean of this set of comprehensive scores is 80 points and the standard deviation is 10 points. Next, based on these mean and standard deviation, the comprehensive score of each strategy is baseline-subtracted and normalized. Specifically, for strategy A, its standardized score can be calculated as (85 - 80) / 10 = 0.5. For strategy B, its standardized score can be calculated as (60 - 80) / 10 = -2.0. For strategy C, its standardized score can be calculated as (95 - 80) / 10 = 1.5. In this way, the original composite score is transformed into a standardized score with zero mean and unit variance. These standardized scores, such as 0.5, -2.0, and 1.5, are then fed into the reinforcement learning module to update the model parameters of the image policy model.

[0139] Through the above technical solution, the embodiments of this application effectively solve the problem that the scale differences and inconsistent distributions of the comprehensive scores of different candidate operation strategies lead to instability and low convergence efficiency in the reinforcement learning process, thus affecting the accuracy of model parameter updates. By performing baseline subtraction and normalization on the comprehensive scores, the reinforcement learning algorithm can receive inputs with uniform scale and distribution characteristics, thereby significantly improving the stability and convergence efficiency of reinforcement learning. This ensures that the model parameters of the image strategy model can be updated more accurately and effectively, thereby optimizing the processing effect of medical image data.

[0140] It should be noted that although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0141] The following describes an embodiment of the apparatus of this application, which can be used to execute the medical image data processing method in the above embodiments of this application.

[0142] Figure 6 A schematic block diagram of a medical image data processing apparatus provided in an embodiment of this application is shown. Figure 6 As shown, the medical image data processing device 600 includes: The extraction module 610 is configured to extract features from the medical image data to be processed, and obtain the image embedding features of the medical image data. The inference module 620 is configured to perform model inference on the image embedding features based on the pre-trained image strategy model to obtain a candidate strategy set. The candidate strategy set includes multiple candidate operation strategies, which are used to represent the operation rules for processing medical image data. The sending module 630 is configured to send a set of candidate strategies to multiple expert clients, which are used to evaluate the merits of the candidate operation strategies based on local data to obtain a strategy score. Update module 640 is configured to update the model parameters of the image strategy model based on the strategy scores returned by multiple expert clients.

[0143] In some embodiments of this application, based on the above technical solutions, the medical image data processing device 600 further includes: The acquisition module is configured to acquire domain embedding features sent by multiple expert clients from a candidate client set. These domain embedding features are used to represent the knowledge domain corresponding to the local data of the expert clients. The selection module is configured to select multiple expert clients from a candidate client set that match the image embedding features based on the domain embedding features.

[0144] In some embodiments of this application, based on the above technical solutions, the selection module is further configured to: obtain the semantic similarity between the domain embedding features and the image embedding features; and select multiple expert clients that match the image embedding features from the candidate client set according to the semantic similarity.

[0145] In some embodiments of this application, based on the above technical solutions, the strategy scoring includes differential privacy noise added by the expert client, and the differential privacy noise follows a Gaussian distribution.

[0146] In some embodiments of this application, based on the above technical solutions, the medical image data processing device 600 further includes: The calibration module is configured to perform monotonically affine calibration on the strategy scores returned by multiple expert clients according to preset scale calibration parameters to obtain strategy scores with uniform dimensions. The scale calibration parameters are trained based on historical data returned by expert clients for a common reference task, which is a general task applicable to multiple knowledge domains.

[0147] In some embodiments of this application, based on the above technical solutions, the medical image data processing device 600 further includes: The standardization module is configured to obtain the mean and standard deviation of multiple strategy scores returned by the same expert client; and to convert the strategy scores returned by the same expert client into standardized scores based on the mean and standard deviation.

[0148] In some embodiments of this application, based on the above technical solutions, the update module 640 is further configured to: obtain the reliability weight corresponding to the expert client, the reliability weight being used to represent the accuracy of the expert client's evaluation of historical tasks; for the same candidate operation strategy, perform a weighted mapping on the strategy scores returned by different expert clients according to the reliability weight to obtain a comprehensive score for the candidate operation strategy; and perform reinforcement learning on the influence strategy model according to the comprehensive score to update the model parameters of the image strategy model.

[0149] In some embodiments of this application, based on the above technical solutions, the update module 640 is further configured to: obtain the average value and standard deviation of the comprehensive scores of different candidate operation strategies; and perform baseline subtraction and normalization on the comprehensive scores of different candidate operation strategies based on the average value and standard deviation.

[0150] The specific details of the medical image data processing apparatus provided in the various embodiments of this application have been described in detail in the corresponding method embodiments, and will not be repeated here.

[0151] Figure 7 A schematic block diagram of a computer system architecture for implementing an electronic device according to embodiments of the present application is shown.

[0152] It should be noted that, Figure 7 The computer system 700 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0153] like Figure 7 As shown, the computer system 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 702 or programs loaded from storage section 708 into random access memory (RAM) 703. The RAM 703 also stores various programs and data required for system operation. The CPU 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output interface 705 (I / O interface) is also connected to the bus 704.

[0154] The following components are connected to the input / output interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a local area network card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the input / output interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 710 as needed so that computer programs read from it can be installed into the storage section 708 as needed.

[0155] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711. When the computer program is executed by central processing unit 701, it performs various functions defined in the system of this application.

[0156] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0157] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0158] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0159] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0160] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0161] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for processing medical image data, characterized in that, The method includes: Feature extraction is performed on the medical image data to be processed to obtain the image embedding features of the medical image data; Based on a pre-trained image strategy model, model inference is performed on the image embedding features to obtain a candidate strategy set. The candidate strategy set includes multiple candidate operation strategies, which are used to represent the operation rules for processing the medical image data. The candidate strategy set is sent to multiple expert clients, which are used to evaluate the merits of the candidate operation strategies based on local data to obtain strategy scores. The model parameters of the image strategy model are updated based on the strategy scores returned by multiple expert clients.

2. The method according to claim 1, characterized in that, Before sending the candidate strategy set to multiple expert clients, the method further includes: Obtain domain embedding features sent by multiple expert clients from a candidate client set, the domain embedding features being used to represent the knowledge domain corresponding to the local data of the expert clients; Based on the domain embedding features, select multiple expert clients from the candidate client set that match the image embedding features.

3. The method according to claim 2, characterized in that, The step of selecting multiple expert clients from the candidate client set that match the image embedding features based on the domain embedding features includes: Obtain the semantic similarity between the domain embedding features and the image embedding features; Based on the semantic similarity, multiple expert clients that match the image embedding features are selected from the candidate client set.

4. The method according to claim 1, characterized in that, The strategy score includes differential privacy noise added by the expert client, and the differential privacy noise follows a Gaussian distribution.

5. The method according to claim 1, characterized in that, Before updating the model parameters of the image strategy model, the method further includes: According to preset scale calibration parameters, monotonically affine calibration is performed on the strategy scores returned by multiple expert clients to obtain strategy scores with uniform dimensions; the scale calibration parameters are trained based on historical data returned by the expert clients for a common reference task, which is a general knowledge task applicable to multiple knowledge domains.

6. The method according to claim 1, characterized in that, Before updating the model parameters of the image strategy model, the method further includes: Obtain the average and standard deviation of multiple strategy scores returned by the same expert client; Based on the mean and standard deviation, the strategy scores returned by the same expert client are converted into standardized scores.

7. The method according to claim 1, characterized in that, The step of updating the model parameters of the image strategy model based on the strategy scores returned by multiple expert clients includes: Obtain the reliability weight corresponding to the expert client, the reliability weight being used to represent the accuracy of the expert client's evaluation of historical tasks; For the same candidate operation strategy, the strategy scores returned by different expert clients are weighted and mapped according to the reliability weight to obtain the comprehensive score of the candidate operation strategy. The influence strategy model is subjected to reinforcement learning based on the comprehensive score to update the model parameters of the image strategy model.

8. The method according to claim 7, characterized in that, Before performing reinforcement learning on the influence strategy model based on the comprehensive score, the method further includes: Obtain the average and standard deviation of the comprehensive scores for different candidate operational strategies; Baseline subtraction and normalization are performed on the comprehensive scores of different candidate operational strategies based on the mean and standard deviation.

9. A medical image data processing device, characterized in that, The device includes: The extraction module is configured to extract features from the medical image data to be processed, thereby obtaining the image embedding features of the medical image data. The inference module is configured to perform model inference on the image embedding features based on a pre-trained image strategy model to obtain a candidate strategy set, which includes multiple candidate operation strategies, and the candidate operation strategies are used to represent the operation rules for processing the medical image data. The sending module is configured to send the candidate strategy set to multiple expert clients, wherein the expert clients are used to evaluate the merits of the candidate operation strategies based on local data to obtain a strategy score; The update module is configured to update the model parameters of the image strategy model based on the strategy scores returned by multiple expert clients.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the executable instructions to implement the medical image data processing method as described in any one of claims 1 to 8.

Images