An artificial intelligence model management system based on metadata

Abstract

The present invention discloses an artificial intelligence model management system based on metadata, which relates to the field of model management technology and includes a metadata processing module, which collects metadata including continuous metadata and discrete metadata, encodes the center point of the continuous metadata calculation field into a label distribution vector, counts the total number of categories of the discrete metadata, encodes and constructs a mapping matrix, performs vector splicing, and uses a linear transformation matrix to perform linear mapping to convert it into a uniform length latent vector. The system of the present invention can cover the full range of variables through detailed segmentation by performing Gaussian kernel label distribution encoding on continuous metadata and encoding on discrete metadata, ensures that different types of features can be integrated into the same latent space through a unified metadata latent vector, and ensures the comparability of features. A gradual denoising diffusion process is introduced into the model, so that the medical data generated by each inference has higher diversity.

Description

Technical Field

[0001] The present invention relates to the field of model management technology, and in particular to an artificial intelligence model management system based on metadata. Background Art

[0002] Metadata, as the "data" that describes data, includes information such as data sources, data types, data relationships, and data processing methods. With the continuous advancement of big data technology, enterprises and research institutions are increasingly relying on efficient metadata management systems when processing massive amounts of information. In the medical field, prediction and analysis based on metadata are also gradually being proposed and used. These systems can not only improve data query efficiency, but also reveal the deeper value hidden behind the data through metadata correlation analysis.

[0003] However, existing metadata management and AI model training methods still have certain shortcomings. Current mainstream technologies mostly focus on the processing and modeling of a single type of metadata. The difference between continuous and discrete metadata is often not sufficiently refined. In particular, it is difficult to efficiently capture the potential relationships between categories in medical information, resulting in information loss and limited model expression capabilities. In addition, existing methods usually ignore how to efficiently integrate heterogeneous data sources. Especially in the case of multimodal data in the medical field, model training is difficult to simultaneously consider the richness and completeness of the data. Therefore, existing technologies face large performance bottlenecks and application limitations when processing complex data types and large-scale data. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a metadata-based artificial intelligence model management system to solve the problem that the difference processing between continuous metadata and discrete metadata is often not sufficiently refined, especially for medical information, it is difficult to efficiently capture the potential relationship between categories, resulting in information loss and limited model expression capabilities. In addition, existing methods usually ignore how to efficiently integrate heterogeneous data sources, especially in the case of multimodal data in the medical field, and model training is difficult to simultaneously consider the richness and completeness of the data.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an artificial intelligence model management system based on metadata, comprising:

[0008] The metadata processing module collects metadata, including continuous metadata and discrete metadata. It calculates the field center point of continuous metadata and encodes it into a label distribution vector. It also counts the total number of categories of discrete metadata, encodes and constructs a mapping matrix, performs vector concatenation, and uses a linear transformation matrix to perform linear mapping to convert it into a uniform-length latent vector.

[0009] The medical data processing module extracts the image latent vector of the medical image and fuses it with the unified latent vector to generate a conditional vector. The ControlNet control module is introduced based on the external structural features to generate a control vector, which is then fused with the image latent vector to generate a fused latent vector. Noise is gradually added and a diffusion model is constructed based on the U-Net network. Inverse denoising training is performed to generate a restored latent vector.

[0010] Model deployment module: The client deploys the diffusion model to perform data standardization, aggregate model parameters, update the global model, and evaluate the global model;

[0011] The query module analyzes the uncertainty of the global model, defines the knowledge graph, records the model operation data, uploads the operation log, converts the query statements entered by the user and performs data query.

[0012] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, wherein: the vector splicing is performed, and the linear transformation matrix is ​​used to perform linear mapping into a uniform length latent vector, including:

[0013] For each field of continuous metadata, the center point of each field is calculated based on the variable range and score segment of the field sample value;

[0014] Construct a Gaussian kernel label distribution function to encode its value into a vector, and map the input value x of the field to a label distribution vector;

[0015] For each field with discrete metadata type, count the number of unique values ​​of all samples in the field as the total number of categories , perform Embedding encoding to construct a mapping matrix with a dimension of , d is a fixed embedding dimension;

[0016] Each row represents the embedding vector of a category, and the embedding vector representation of each discrete variable is obtained;

[0017] The concatenated label distribution vector and the embedded vector representation of discrete metadata have output vector dimensions of all continuous variables of p and embedding vector dimensions of all discrete variables of d. and the total number of discrete variables , define the dimension D after splicing;

[0018] The encoding vectors of all fields are sequentially concatenated into a total vector, and a linear transformation matrix is ​​used to perform linear mapping to convert it into a latent vector of uniform length.

[0019] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, wherein: the generation control vector is fused with the unified latent vector of the metadata to generate the condition vector, including:

[0020] According to the corresponding medical image data of the clinical data, the image data is paired with the unified latent vector of the metadata, and the image latent vector of the medical image data is extracted using a pre-trained image encoder, and then fused with the unified latent vector of the metadata to generate a conditional vector;

[0021] Determine the external structural features of medical image data, including key points, labels, and structural information of the image, introduce the ControlNet control module, and take the condition vector and external structural features as input to generate the control vector and fused with the image latent vector to generate a fused latent vector;

[0022] Perform forward diffusion process based on the fused latent vector by gradually adding noise;

[0023] A diffusion model is constructed based on the U-Net network, and reverse denoising training is performed to restore the latent vector of the original image from the noisy state. The mean square error loss MSE is used to train the model to predict noise.

[0024] Use the Adam optimizer for gradient descent optimization to update the model parameters. Stop the iteration if the model loss no longer decreases significantly during the continuous iteration.

[0025] The restored latent vector is generated by the trained diffusion model based on the noise input and conditional vector generation conditions in the latent space.

[0026] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, wherein: the aggregation model parameters update the global model, including:

[0027] Collect customer metadata based on the client terminal, use the Meta-Calibrator to standardize the metadata of each client, deploy the diffusion model and use local data for model training;

[0028] A weighted coefficient is determined for each client, and the model parameters of all clients are aggregated using the weighted average method, and the global model is updated according to the global model parameters.

[0029] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, wherein: the global model evaluation includes:

[0030] Perform FID and SSIM evaluation on the global model;

[0031] The sum of the mean and standard deviation of the historical distribution difference between the generated image and the real image is used as the difference threshold. If the distribution difference value is less than or equal to the difference threshold, the original features of the image are retained.

[0032] The sum of the mean and standard deviation of the historical structural difference values ​​between the generated image and the real image is used as the structural threshold. If the structural difference value is greater than or equal to the structural threshold, it is judged that the original features of the image are retained.

[0033] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, wherein: the analysis of model uncertainty of the global model includes:

[0034] A repeated sampling mechanism LAS is introduced for the global model to perform multiple random initializations on the input image latent vector and conditional vector, and obtain the corresponding repeated sampling restored latent vector. The average prediction value and average standard deviation are calculated to generate an uncertainty score.

[0035] The sum of the mean and twice the standard deviation of the historical uncertainty score is used as the scoring threshold. If the uncertainty score is less than the scoring threshold, it is judged to be of low uncertainty.

[0036] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, the collected metadata includes continuous metadata and discrete metadata, including:

[0037] By traversing each column field in the original clinical data table, the value type and value distribution range of each column are read. If the field value is a continuous numeric type, it is marked as continuous metadata. If the field value is a discrete classification label, it is marked as discrete metadata. Each row in the clinical data table is used as a sample.

[0038] As a preferred solution of the metadata-based artificial intelligence model management system described in the present invention, the definition of the knowledge graph to record model operation data refers to using the DataHub architecture to define the triple structure of the knowledge graph KG, recording the model's input / output mapping, data source, processing steps, model parameters and execution environment information.

[0039] As a preferred solution of the metadata-based artificial intelligence model management system described in the present invention, the uploading operation log refers to recording the operation log for each operation step of the model and storing it as structured metadata.

[0040] As a preferred solution of the metadata-based artificial intelligence model management system of the present invention, wherein: the conversion of the query statement input by the user and the data query include:

[0041] Use the BERT model to convert user-entered query statements into SPARQL statements to query the metadata in the knowledge graph;

[0042] Based on the user query, the system will return the corresponding model path, inference conditions and detailed information of input data.

[0043] The beneficial effects of the present invention are as follows: by performing Gaussian kernel label distribution encoding of continuous metadata and encoding of discrete metadata, the full range of variables can be covered through detailed segmentation, and a unified metadata latent vector ensures that different types of features can be integrated into the same latent space while ensuring the comparability of features. The introduction of a gradual denoising diffusion process in the model makes the medical data generated by each inference have higher diversity. The introduction of noise enables the model to generate diverse latent vector representations, thereby enhancing the flexibility of the model in actual clinical applications, especially in disease progression prediction, personalized medicine, and simulation of different clinical trials, providing more prediction options. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a schematic diagram of the structure of the metadata-based artificial intelligence model management system in Example 1;

[0046] Figure 2 This is a flow chart of the metadata-based artificial intelligence model management system in Example 1. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0050] Example 1, reference Figures 1 to 2 , which is the first embodiment of the present invention, provides an artificial intelligence model management system based on metadata, including the following steps:

[0051] S1: Collect metadata, including continuous metadata and discrete metadata. Calculate the field center of continuous metadata and encode it into a label distribution vector. Count the total number of categories of discrete metadata, encode and construct a mapping matrix, perform vector concatenation, and use a linear transformation matrix to perform linear mapping to convert it into a uniform-length latent vector.

[0052] Preferably, the collected metadata includes continuous metadata and discrete metadata, including:

[0053] By traversing each column field in the original clinical data table, the value type and value distribution range of each column are read. If the field value is a continuous numeric type, it is marked as continuous metadata. If the field value is a discrete classification label, it is marked as discrete metadata. Each row in the clinical data table is used as a sample.

[0054] By traversing the original clinical data table and classifying it according to the value type and value distribution range of each column, the data can be effectively and automatically marked as continuous or discrete metadata, which helps to clearly identify the data structure and provide accurate data type identification for subsequent data processing, modeling and analysis, ensuring consistency and efficiency in the processing process. At the same time, it helps the model to adopt appropriate processing methods based on the characteristics of the data when processing different types of data, thereby improving the performance and accuracy of the model.

[0055] Furthermore, vector concatenation is performed and linear transformation matrix is ​​used to perform linear mapping into a uniform length latent vector, including:

[0056] For each field of continuous metadata, the center point of each field is calculated based on the variable range and score segment of the field sample value, expressed as:

[0057] ;

[0058] in represents the center point of the i-th Gaussian kernel, n represents the number of segments, b represents the maximum value of the variable range, and a represents the minimum value of the variable range;

[0059] Construct a Gaussian kernel label distribution function to encode its value into a vector and map the input value x of the field to a label distribution vector, expressed as:

[0060] ;

[0061] in represents the label distribution vector of the i-th field, Represents the smoothing factor, which can be calculated according to the formula Determine, used to ensure consistent distribution coverage width;

[0062] For each field with discrete metadata type, count the number of unique values ​​of all samples in the field as the total number of categories , perform Embedding encoding to construct a mapping matrix with a dimension of , d is a fixed embedding dimension;

[0063] Each row represents the embedding vector of a category, and the embedding vector representation of each discrete variable is obtained. For each sample, the category index j is obtained by looking up the table, and the jth row is extracted as the embedding vector;

[0064] The concatenated label distribution vector and the embedded vector representation of discrete metadata have output vector dimensions of all continuous variables of p and embedding vector dimensions of all discrete variables of d. and the total number of discrete variables , define the dimension D after splicing, expressed as: ;

[0065] The encoding vectors of all fields are sequentially concatenated into a total vector, which is then linearly mapped using a linear transformation matrix into a uniform-length latent vector. This vector represents the metadata condition and is passed to the AI ​​model control path, expressed as:

[0066] ;

[0067] ;

[0068] in Represents the original vector after all metadata are concatenated, with a length of D. express, A unified latent vector representation representing metadata, represents the linear transformation matrix, which can be initialized by the standard normal distribution and then fixed during training.

[0069] By performing Gaussian kernel label distribution encoding for continuous metadata and Embedding encoding for discrete metadata, the full range of variables can be covered through detailed segmentation, avoiding information loss caused by single numerical input. This allows the model to more accurately learn the underlying patterns of the data when processing these features. For discrete data, Embedding encoding effectively reduces dimensionality while retaining category information by mapping each category to a dense vector space, improving the model's generalization ability and prediction accuracy.

[0070] The unified metadata latent vector ensures that different types of features can be integrated into the same latent space, while ensuring the comparability of features. The relative relationship between each feature can be captured by the model, thus providing the model with more comprehensive conditional input. This latent vector representation not only improves the integration of data, but also enhances the model's comprehensive understanding of various input features, thereby further improving the effectiveness of generation or prediction tasks.

[0071] S2 extracts the image latent vector of the medical image and fuses it with the unified latent vector to generate a conditional vector. The ControlNet control module is introduced according to the external structural features to generate a control vector, which is fused with the image latent vector to generate a fused latent vector. Noise is gradually added and a diffusion model is constructed based on the U-Net network. Inverse denoising training is performed to generate a restored latent vector.

[0072] Preferably, a control vector is generated and fused with the unified latent vector of metadata to generate a conditional vector, including:

[0073] According to the corresponding medical image data of the clinical data, the image data is paired with the unified latent vector of the metadata, and the image latent vector of the medical image data is extracted using the pre-trained image encoder. It is then fused with the unified latent vector of the metadata to generate a conditional vector, which is expressed as:

[0074] ;

[0075] in represents the conditional vector, represents the image latent vector output by the image encoder, represents the fusion coefficient, which is determined based on historical experience;

[0076] Determine the external structural features of medical image data, including key points, labels, and structural information of the image, introduce the ControlNet control module, and take the condition vector and external structural features as input to generate the control vector , and fused with the image latent vector to generate a fused latent vector, expressed as:

[0077] ;

[0078] in represents the fused latent vector, represents the control weight, which is obtained by training and optimizing the calibrated training data;

[0079] Based on the fusion latent vector, the forward diffusion process is performed by gradually adding noise, making each image latent vector more blurred, which is expressed as:

[0080] ;

[0081] in and denote the fused latent vectors at time steps t and t-1 respectively, represents the decay factor at step t (determined based on historical experience), represents standard Gaussian noise, with covariance being the unit matrix;

[0082] A diffusion model is constructed based on the U-Net network, and reverse denoising training is performed to restore the latent vector of the original image from the noise state. The mean square error loss MSE is used to train the model to predict noise, which is expressed as:

[0083] ;

[0084] in represents the calculation loss, Represents the U-Net network model used to predict noise;

[0085] Use the Adam optimizer for gradient descent optimization to update the model parameters. Stop the iteration if the model loss no longer decreases significantly during the continuous iteration.

[0086] According to the noise input and conditional vector generation conditions in the latent space, the restored latent vector is generated by the trained diffusion model and reconstructed into image data through the decoder.

[0087] By pairing the unified latent vector of medical data and metadata, the generated conditional vector can provide more comprehensive input information for the model, allowing the model to simultaneously consider the latent characteristics of the medical data and the individualized metadata of the patient. Controlling the conditional vector through the ControlNet module can further enhance the model's fine-grained control over medical data. In particular, in medical data generation tasks, controlling the influence of external structural features can help the model generate medical data that is more consistent with medical knowledge. For example, it can accurately reconstruct the anatomical structure of a specific part or adjust certain features of medical data (such as the lesion area). Through optimized control weights, the model can more accurately adjust the generation process based on feedback from the training data, ensuring that the generated images better meet clinical needs.

[0088] By performing a forward diffusion process on the fused latent vector and adding noise, the model learns how to gradually recover from a fuzzy state, thereby better generating high-quality medical data when faced with unknown or ambiguous inputs. The introduction of the diffusion model can further improve the diversity of generated medical data and allow the model to generate different versions of medical data, enhancing the model's exploration capabilities.

[0089] The diffusion model built on the U-Net network uses reverse denoising training to gradually recover clear image latent vectors from noise. The model can accurately predict image details during the denoising process while maintaining an efficient training process. The use of the Adam optimizer ensures stable parameter updates, accelerating convergence and improving model performance.

[0090] The trained diffusion model can generate a restored latent vector based on the noise input and the conditional vector, and convert it into image data through a decoder. By combining medical image data and metadata to generate the conditional vector, the patient's individualized information can be effectively introduced into the image generation process. The introduction of the control module (ControlNet) allows the image generation process to take into account external structural features and precisely control specific anatomical structures during the generation process, effectively reducing the distortion of the generated image and improving the medical credibility of the image.

[0091] The introduction of a gradual denoising diffusion process into the model makes the medical data generated in each inference more diverse. The introduction of noise enables the model to generate diverse latent vector representations, thereby outputting different versions of medical data, enhancing the flexibility of the model in actual clinical applications, especially in disease progression prediction, personalized medicine and simulation of different clinical trials, providing more prediction options.

[0092] S3, the client deploys the diffusion model to normalize data, aggregates model parameters to update the global model, and evaluates the global model;

[0093] Preferably, the aggregation model parameters update the global model, including,

[0094] Collect customer metadata based on the client terminal, use the Meta-Calibrator to standardize the metadata of each client, deploy the diffusion model and use local data for model training;

[0095] Determine the weight coefficient for each client, and use the weighted average method to aggregate the model parameters of all clients, and update the global model based on the global model parameters, which can be expressed as:

[0096] ;

[0097] in represents the global model parameters, Indicates the total number of clients, represents the data volume of client i, D represents the total data volume, represents the parameters of the i-th client model, including the weights and biases of the diffusion model, As model weights for different clients.

[0098] By using a metadata calibrator, we can ensure that the data of all clients is processed under the same standard, thereby improving the uniformity and consistency of the data. Deploying a diffusion model and using local data for training can enable each client to make personalized adjustments based on its own local data, improving the model performance of each client. Using the weighted average method to aggregate the model parameters of each client helps to better balance the influence of each client.

[0099] Furthermore, the global model evaluation includes,

[0100] Perform FID and SSIM evaluation on the global model;

[0101] The sum of the mean and standard deviation of the historical distribution difference between the generated image and the real image is used as the difference threshold. If the distribution difference value is less than or equal to the difference threshold, the original features of the image are retained.

[0102] The sum of the mean and standard deviation of the historical structural difference values ​​between the generated image and the real image is used as the structural threshold. If the structural difference value is greater than or equal to the structural threshold, it is judged that the original features of the image are retained.

[0103] By using FID and SSIM to evaluate the difference between generated images and real images, the improvement in image quality and fidelity can be effectively measured. FID evaluation can quantify the distribution difference between generated images and real images, while SSIM evaluation focuses on the structural similarity of images.

[0104] S4: Analyze the uncertainty of the global model, define a knowledge graph to record model operation data, upload operation logs, convert user-entered query statements, and perform data queries;

[0105] Preferably, the global model is analyzed for model uncertainty, including,

[0106] A repeated sampling mechanism (LAS) is introduced for the global model. The input image latent vector and conditional vector are randomly initialized multiple times. The output of the model under different initialization conditions is obtained through multiple sampling, thereby obtaining the uncertainty information of the result and the corresponding repeated sampling restored latent vector. The average prediction value and average standard deviation are calculated to generate an uncertainty score.

[0107] The sum of the mean and twice the standard deviation of the historical uncertainty score is used as the scoring threshold. If the uncertainty score is less than the scoring threshold, it is judged to be of low uncertainty.

[0108] By introducing the repeated sampling mechanism (LAS), the model can obtain outputs under different initialization conditions by randomly initializing the input image latent vector and conditional vector multiple times, thereby better reflecting the uncertainty of the generated results;

[0109] By calculating the average prediction value and standard deviation of the restored latent vector obtained by multiple samplings, an uncertainty score can be generated for each output, which can effectively identify which generated results have higher stability and which have greater uncertainty.

[0110] Furthermore, we define the knowledge graph to record model operation data, including:

[0111] The DataHub architecture is used to define the triple structure of the knowledge graph KG, recording the model's input / output mapping, data source, processing steps, model parameters, and execution environment information. Each triple contains: entities (such as datasets, models, and inference results), attributes (such as data types, calculation methods, and model status), and relationships (such as the mapping relationship between models and datasets).

[0112] By using the DataHub architecture to define the triple structure of the knowledge graph (KG), the model's input / output mapping, data sources, processing steps, model parameters, and execution environment information can be systematically recorded and managed. This structured information storage method provides a clear framework for data traceability, management, and analysis, ensuring the transparency of data flow and model operation. The triple format makes the definition of each entity, attribute, and relationship clearer, facilitating collaboration and information sharing across systems or teams, thereby improving the traceability, maintainability, and efficiency of the workflow.

[0113] Furthermore, uploading operation logs means recording operation logs for each operation step of the model and storing them as structured metadata for auditing and review.

[0114] All metadata is indexed and associated with specific training processes, inference results, or abnormal events. Using the graph query interface, users can quickly find all operation histories related to a model or dataset.

[0115] By using the DataHub architecture to define the triple structure of the knowledge graph (KG), the model's input / output mapping, data sources, processing steps, model parameters, and execution environment information can be systematically recorded and managed to ensure transparency of data flow and model operation. The triple format makes the definitions of various entities, attributes, and relationships clearer, facilitating collaboration and information sharing across systems or teams, thereby improving the traceability, maintainability, and efficiency of the workflow.

[0116] Furthermore, the query statement entered by the user is converted and data query is performed, including:

[0117] Use the BERT model to convert user-entered query statements into SPARQL statements to query the metadata in the knowledge graph;

[0118] Based on the user's query, the system will return the corresponding model path, inference conditions and detailed information of input data;

[0119] By utilizing the BERT large model to convert user query statements into SPARQL statements, more natural and intelligent user interactions can be achieved, allowing users to obtain metadata-related information in the knowledge graph through simple queries. This automated conversion improves the flexibility and efficiency of queries, enabling the system to quickly return relevant model paths, reasoning conditions, and detailed information on input data based on user needs, thereby optimizing the application of the knowledge graph and user experience.

[0120] In summary, the present invention performs Gaussian kernel label distribution encoding of continuous metadata and encoding of discrete metadata, which can cover the full range of variables through detailed segmentation. A unified metadata latent vector ensures that different types of features can be integrated into the same latent space while ensuring the comparability of features. The introduction of a gradual denoising diffusion process in the model makes the medical data generated by each inference have higher diversity. The introduction of noise enables the model to generate diverse latent vector representations, enhancing the flexibility of the model in actual clinical applications, especially in disease progression prediction, personalized medicine, and simulation of different clinical trials, providing more prediction options.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence model management system based on metadata, characterized in that: include: The metadata processing module collects metadata, including continuous metadata and discrete metadata. It calculates the field center point of continuous metadata and encodes it into a label distribution vector. It also counts the total number of categories of discrete metadata, encodes and constructs a mapping matrix, performs vector concatenation, and uses a linear transformation matrix to perform linear mapping to convert it into a uniform-length latent vector. The medical data processing module pairs the image data with the unified latent vector of the metadata based on the corresponding medical image data of the clinical data, extracts the image latent vector of the medical image data using a pre-trained image encoder, and fuses it with the unified latent vector of the metadata to generate a conditional vector; Determine the external structural features of medical image data, including key points, labels, and structural information of the image, introduce the ControlNet control module, and take the condition vector and external structural features as input to generate the control vector and fused with the image latent vector to generate a fused latent vector; Perform forward diffusion process based on the fused latent vector by gradually adding noise; A diffusion model is constructed based on the U-Net network, and reverse denoising training is performed to restore the latent vector of the original image from the noisy state. The mean square error loss MSE is used to train the model to predict noise. Use the Adam optimizer for gradient descent optimization to update the model parameters. Stop the iteration if the model loss no longer decreases significantly during the continuous iteration. According to the noise input and conditional vector generation conditions in the latent space, the restored latent vector is generated through the trained diffusion model; Model deployment module: The client deploys the diffusion model to perform data standardization, aggregate model parameters, update the global model, and evaluate the global model; The query module analyzes the uncertainty of the global model, defines the knowledge graph, records the model operation data, uploads the operation log, converts the query statements entered by the user and performs data query.

2. The metadata-based artificial intelligence model management system according to claim 1, characterized in that: The vector concatenation is performed by using a linear transformation matrix to perform linear mapping into a uniform length latent vector, including: For each field of continuous metadata, the center point of each field is calculated based on the variable range and score segment of the field sample value; Construct a Gaussian kernel label distribution function to encode its value into a vector, and map the input value x of the field to a label distribution vector; For each field with discrete metadata type, count the number of unique values ​​of all samples in the field as the total number of categories , perform Embedding encoding to construct a mapping matrix with a dimension of , d is a fixed embedding dimension; Each row represents the embedding vector of a category, and the embedding vector representation of each discrete variable is obtained; The concatenated label distribution vector and the embedded vector representation of discrete metadata have output vector dimensions of all continuous variables of p and embedding vector dimensions of all discrete variables of d. and the total number of discrete variables , define the dimension D after splicing; The encoding vectors of all fields are sequentially concatenated into a total vector, and a linear transformation matrix is ​​used to perform linear mapping to convert it into a latent vector of uniform length.

3. The metadata-based artificial intelligence model management system according to claim 2, characterized in that: The aggregation model parameters update the global model, including: Collect customer metadata based on the client terminal, use the Meta-Calibrator to standardize the metadata of each client, deploy the diffusion model and use local data for model training; A weighted coefficient is determined for each client, and the model parameters of all clients are aggregated using the weighted average method, and the global model is updated according to the global model parameters.

4. The metadata-based artificial intelligence model management system according to claim 3, characterized in that: The global model evaluation includes: Perform FID and SSIM evaluation on the global model; The sum of the mean and standard deviation of the historical distribution difference between the generated image and the real image is used as the difference threshold. If the distribution difference value is less than or equal to the difference threshold, the original features of the image are retained. The sum of the mean and standard deviation of the historical structural difference values ​​between the generated image and the real image is used as the structural threshold. If the structural difference value is greater than or equal to the structural threshold, it is judged that the original features of the image are retained.

5. The metadata-based artificial intelligence model management system according to claim 4, characterized in that: The analysis of model uncertainty of the global model includes: A repeated sampling mechanism LAS is introduced for the global model to perform multiple random initializations on the input image latent vector and conditional vector, and obtain the corresponding repeated sampling restored latent vector. The average prediction value and average standard deviation are calculated to generate an uncertainty score. The sum of the mean and twice the standard deviation of the historical uncertainty score is used as the scoring threshold. If the uncertainty score is less than the scoring threshold, it is judged to be of low uncertainty.

6. The metadata-based artificial intelligence model management system according to claim 5, characterized in that: The collected metadata includes continuous metadata and discrete metadata, including: By traversing each column field in the original clinical data table, the value type and value distribution range of each column are read. If the field value is a continuous numeric type, it is marked as continuous metadata. If the field value is a discrete classification label, it is marked as discrete metadata. Each row in the clinical data table is used as a sample.

7. The metadata-based artificial intelligence model management system according to claim 6, characterized in that: Defining the knowledge graph and recording model operation data refers to using the DataHub architecture to define the triple structure of the knowledge graph KG, and recording the model's input / output mapping, data source, processing steps, model parameters, and execution environment information.

8. The metadata-based artificial intelligence model management system according to claim 7, characterized in that: The uploading operation log refers to recording the operation log for each operation step of the model and storing it as structured metadata.

9. The metadata-based artificial intelligence model management system according to claim 8, characterized in that: The conversion of the query statement input by the user and performing data query includes: Use the BERT model to convert user-entered query statements into SPARQL statements to query the metadata in the knowledge graph; Based on the user query, the system will return the corresponding model path, inference conditions and detailed information of input data.

Images