CN120104813B - An artificial intelligence model management system based on metadata - Google Patents

Abstract

The invention discloses an artificial intelligent model management system based on metadata, which relates to the technical field of model management and comprises a metadata processing module, wherein collected metadata comprises continuous metadata and discrete metadata, a central point of a calculation field of the continuous metadata is encoded into a label distribution vector, the total number of statistic categories of the discrete metadata is encoded to construct a mapping matrix, vector splicing is carried out, and linear mapping is carried out by using a linear transformation matrix to convert the linear mapping into a uniform length latent vector. According to the system, through carrying out Gaussian kernel tag distribution coding of continuous metadata and coding of discrete metadata, the complete range of variables can be covered through careful segmentation, different types of features can be fused in the same potential space through unified metadata latent vectors, meanwhile, the comparability of the features is guaranteed, a diffusion process of gradual denoising is introduced into a model, and medical data generated by reasoning each time has higher diversity.

Description

Artificial intelligence model management system based on metadata

Technical Field

The invention relates to the technical field of model management, in particular to an artificial intelligent model management system based on metadata.

Background

Metadata, which is the "data" describing the data, contains information such as data sources, data types, data relationships, data processing methods and the like, along with the continuous progress of big data technology, enterprises and scientific research institutions increasingly rely on efficient metadata management systems when processing massive information, prediction and analysis are gradually proposed and used according to the metadata in the medical field, and the systems not only can improve the query efficiency of the data, but also can reveal the deep value hidden behind the data through the association analysis of the metadata;

however, the existing metadata management and AI model training methods still have certain disadvantages, most of the current mainstream technologies focus on processing and modeling of single type metadata, the difference processing of continuous type metadata and discrete type metadata often fails to achieve sufficient refinement, and particularly, potential relations among categories are difficult to capture efficiently for medical information, so that information loss and model expression capacity are limited.

Disclosure of Invention

The present invention has been made in view of the above-described problems occurring in the prior art.

Therefore, the invention provides an artificial intelligent model management system based on metadata, which solves the problems that the difference processing of continuous metadata and discrete metadata is often not fully refined, and particularly the potential relation among categories is difficult to capture efficiently aiming at medical information, so that information loss and model expression capacity are limited.

In order to solve the technical problems, the invention provides the following technical scheme:

in a first aspect, the present invention provides a metadata-based artificial intelligence model management system comprising:

The metadata processing module is used for acquiring metadata, wherein the metadata comprises continuous metadata and discrete metadata, calculating field center points of the continuous metadata to be encoded into tag distribution vectors, counting the total number of categories of the discrete metadata, encoding to construct a mapping matrix, vector splicing, and linear mapping by using a linear transformation matrix to be converted into uniform length latent vectors;

The medical data processing module extracts an image latent vector of a medical image, fuses the image latent vector with a unified latent vector to generate a condition vector, introduces a control Net control module according to external structural features to generate a control vector, fuses the control vector with the image latent vector to generate a fused latent vector, gradually adds noise, constructs a diffusion model according to a U-Net network, and performs reverse denoising training to generate a restored latent vector;

the model deployment module is used for carrying out data standardization on the client deployment diffusion model, aggregating model parameters, updating the global model and evaluating the global model;

And the query module is used for analyzing the model uncertainty of the global model, defining a knowledge graph, recording model operation data, uploading an operation log, converting query sentences input by a user and carrying out data query.

As a preferred scheme of the metadata-based artificial intelligence model management system of the present invention, the vector stitching, the linear mapping using a linear transformation matrix, is converted into a uniform length latent vector, including,

For each field of which the type is continuous metadata, calculating the center point of each field according to the variable range and the fraction segment of the field sample value;

Constructing a Gaussian kernel tag distribution function, encoding the value of the Gaussian kernel tag distribution function into a vector, and mapping an input value x of a field into a tag distribution vector;

for each field with discrete metadata, counting the unique value number of all samples in the field as the total number of categories Embedding coding is carried out to construct a mapping matrix with the dimension ofD is a fixed embedding dimension;

each row represents a category of embedded vectors, and the embedded vector representation of each discrete variable is obtained;

splicing the embedded vector representation of the tag distribution vector and the discrete metadata, wherein the output vector dimension of all continuous variables is p, the dimension of the embedded vector of all discrete variables is d, and the total number of the continuous variables is calculated And the total number of discrete variablesDefining a spliced dimension D;

and sequentially splicing the coded vectors of all the fields into a total vector, and performing linear mapping by using a linear transformation matrix to convert the total length latent vector.

As a preferable scheme of the artificial intelligence model management system based on metadata, the generating control vector and the merging of the unified latent vector of the metadata to generate the condition vector comprise,

Pairing the image data with the unified latent vector of the metadata according to the corresponding medical image data of the clinical data, extracting the image latent vector of the medical image data by using a pre-trained image encoder, and fusing the image latent vector with the unified latent vector of the metadata to generate a condition vector;

Determining external structural features of medical image data, including key points, labels and structural information of the image, introducing a control Net control module to take a condition vector and the external structural features as inputs to generate a control vector Fusing the fusion latent vector with the image latent vector to generate a fusion latent vector;

Forward diffusion is carried out by gradually adding noise based on the fusion latent vector;

Constructing a diffusion model according to the U-Net network, performing reverse denoising training, recovering a latent vector of an original image from a noise state, and training model prediction noise by using a mean square error loss MSE;

gradient descent optimization is carried out by using an Adam optimizer, parameters of the model are updated, and iteration is stopped when the loss of the model is not obviously reduced in the continuous iteration process;

And generating a restored latent vector through the trained diffusion model according to noise input in the latent space and the condition vector generation condition.

As a preferred embodiment of the metadata-based artificial intelligence model management system of the present invention, the aggregate model parameter updating global model comprises,

Collecting client metadata according to client terminals, standardizing the metadata of each client by using a metadata Calibrator Meta-calizer, deploying a diffusion model and performing model training by using local data;

and determining a weighting coefficient for each client, aggregating model parameters of all clients by using a weighted average method, and updating the global model according to the global model parameters.

As a preferred embodiment of the metadata-based artificial intelligence model management system of the present invention, wherein said evaluating the global model comprises,

Performing FID evaluation and SSIM evaluation on the global model;

According to the evaluation, the sum of the mean value and the standard deviation of the historical distribution difference value between the generated image and the real image is used as a difference threshold value, and if the distribution difference value is smaller than or equal to the difference threshold value, the original characteristics of the image are judged to be reserved;

And taking the sum of the mean value and the standard deviation of the historical structure difference value between the generated image and the real image as a structure threshold value according to the evaluation, and judging that the original characteristics of the image are reserved if the structure difference value is greater than or equal to the structure threshold value.

As a preferred embodiment of the metadata-based artificial intelligence model management system of the present invention, the model uncertainty for the global model analysis includes,

Introducing a repeated sampling mechanism LAS aiming at the global model, carrying out random initialization on the input image latent vector and the condition vector for a plurality of times, obtaining a corresponding repeated sampling reduction latent vector, calculating an average predicted value and an average standard deviation, and generating an uncertainty score;

And according to the sum of the average value and the double standard deviation of the historical uncertainty score as a score threshold, if the uncertainty score is smaller than the score threshold, judging that the uncertainty is lower.

As a preferred embodiment of the metadata-based artificial intelligence model management system of the present invention, wherein the acquisition metadata includes continuous metadata and discrete metadata, including,

And reading the value type and the value distribution range of each column by traversing each column of field in the original clinical data table, marking as continuous metadata if the field value is of a continuous digital type, marking as discrete metadata if the field value is of a discrete classification label, and taking each row in the clinical data table as a sample.

As an optimal scheme of the artificial intelligent model management system based on metadata, the method comprises the steps of defining a knowledge graph, recording model operation data, namely defining a triplet structure of a knowledge graph KG by using DataHub framework, and recording input/output mapping, data sources, processing steps, model parameters and execution environment information of the model.

As a preferable scheme of the artificial intelligence model management system based on metadata, the uploading operation log is to record the operation log for each operation step of the model and store the operation log as structured metadata.

As a preferred embodiment of the metadata-based artificial intelligence model management system of the present invention, wherein the converting the query sentence input by the user and performing the data query includes,

Converting a query sentence input by a user into an SPARQL sentence by using the BERT big model, and querying metadata in the knowledge graph;

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

The invention has the beneficial effects that through carrying out Gaussian kernel tag distribution coding of continuous metadata and coding of discrete metadata, the complete range of variables can be covered through careful segmentation, different types of features can be fused in the same potential space through unified metadata latent vectors, meanwhile, the comparability of the features is guaranteed, a gradual denoising diffusion process is introduced into a model, medical data generated by each reasoning has higher diversity, the model can generate various latent vector representations through the introduction of noise, the flexibility of the model in actual clinical application is enhanced, and more prediction options are provided especially in the aspects of disease progress prediction, personalized medical treatment and simulation of different clinical tests.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram of an artificial intelligence model management system based on metadata in embodiment 1;

fig. 2 is a flow chart of an artificial intelligence model management system based on metadata in embodiment 1.

Detailed Description

In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings.

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways other than those described herein, and persons skilled in the art will readily appreciate that the present invention is not limited to the specific embodiments disclosed below.

Further, reference herein to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic can be included in at least one implementation of the invention. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments.

Embodiment 1, referring to fig. 1 to 2, is a first embodiment of the present invention, and provides an artificial intelligence model management system based on metadata, comprising the steps of:

S1, collecting metadata, wherein the metadata comprises continuous metadata and discrete metadata, calculating field center points of the continuous metadata to be encoded into tag distribution vectors, counting total number of categories of the discrete metadata, encoding to construct a mapping matrix, vector splicing, and linear mapping to be converted into uniform length latent vectors by using a linear transformation matrix;

Preferably, the acquisition metadata includes both continuous metadata and discrete metadata, including,

And reading the value type and the value distribution range of each column by traversing each column of field in the original clinical data table, marking as continuous metadata if the field value is of a continuous digital type, marking as discrete metadata if the field value is of a discrete classification label, and taking each row in the clinical data table as a sample.

The data can be automatically marked as continuous or discrete metadata by traversing the original clinical data table and classifying according to the value type and the value distribution range of each column, so that the data structure can be clearly identified, accurate data type identification is provided for subsequent data processing, modeling and analysis, consistency and efficiency in the processing process are ensured, and meanwhile, a proper processing method can be adopted by a model according to the characteristics of the data when the model is used for processing different types of data, so that the performance and accuracy of the model are improved.

Further, vector stitching is performed, linear mapping is performed using a linear transformation matrix to convert into a uniform length latent vector, including,

For each field of which the type is continuous metadata, calculating a center point of each field according to the variable range of the field sample value and the fractional segment, wherein the center point is expressed as:

;

Wherein the method comprises the steps of The method is characterized by comprising the steps of representing an ith Gaussian kernel center point, n represents a segmentation number, b represents the maximum value of a variable range, and a represents the minimum value of the variable range;

Constructing a Gaussian kernel tag distribution function, encoding values of the Gaussian kernel tag distribution function into vectors, mapping input values x of fields into tag distribution vectors, and representing the tag distribution vectors as follows:

;

Wherein the method comprises the steps of A tag distribution vector representing the i-th field,Representing the smoothing factor, which can be according to the formulaDetermining, for ensuring that the distribution coverage width is consistent;

for each field with discrete metadata, counting the unique value number of all samples in the field as the total number of categories Embedding coding is carried out to construct a mapping matrix with the dimension ofD is a fixed embedding dimension;

each row represents a category of embedded vector, the embedded vector representation of each discrete variable is obtained, for each sample, the category index j is obtained by looking up a table, and the j-th row is extracted as the embedded vector;

splicing the embedded vector representation of the tag distribution vector and the discrete metadata, wherein the output vector dimension of all continuous variables is p, the dimension of the embedded vector of all discrete variables is d, and the total number of the continuous variables is calculated And the total number of discrete variablesThe post-splice dimension D is defined, expressed as:;

The coding vectors of all fields are spliced into a total vector in sequence, linear mapping is carried out by using a linear transformation matrix, the vector is a uniform length latent vector, the vector is represented by metadata condition, and the control path of the incoming AI model is represented as:

;

;

Wherein the method comprises the steps of Representing the original vector after all metadata are spliced, the length is D,The representation is made of a combination of a first and a second color,A unified latent vector representation representing metadata,The representation of the linear transformation matrix may be initialized by a standard normal distribution and then fixed during the training process.

Through carrying out Gaussian kernel tag distribution coding of continuous metadata and Embedding coding of discrete metadata, the complete range of variables can be covered through fine segmentation, so that information loss caused by single numerical input is avoided, potential rules of data can be learned more accurately when the model processes the characteristics, and for discrete data, embedding coding can effectively reduce dimensions and simultaneously reserve category information, and generalization capability and prediction accuracy of the model are improved through mapping each category to a dense vector space;

The unified metadata latent vector ensures that different types of features can be fused in the same potential space, meanwhile, the comparability of the features is ensured, and the relative relation among the features can be captured by a model, so that more comprehensive condition input is provided for the model, and the latent vector representation not only improves the integration degree of data, but also enhances the comprehensive understanding capability of the model to various input features, so that the effect of generating tasks or predicting tasks is further improved.

S2, extracting an image latent vector of a medical image, fusing the image latent vector with a unified latent vector to generate a condition vector, introducing a control Net control module according to external structural features to generate a control vector, fusing the control vector with the image latent vector to generate a fused latent vector, gradually adding noise, constructing a diffusion model according to a U-Net network, and performing reverse denoising training to generate a restored latent vector;

Preferably, a control vector is generated, fused with a unified latent vector of metadata to generate a conditional vector, including,

Pairing the image data with the unified latent vector of the metadata according to the corresponding medical image data of the clinical data, extracting the image latent vector of the medical image data by using a pre-trained image encoder, and fusing the image latent vector with the unified latent vector of the metadata to generate a condition vector, wherein the condition vector is expressed as follows:

;

Wherein the method comprises the steps of The condition vector is represented as a vector of conditions,Representing the latent image vector output by the image encoder,Representing a fusion coefficient, and determining based on historical experience;

Determining external structural features of medical image data, including key points, labels and structural information of the image, introducing a control Net control module to take a condition vector and the external structural features as inputs to generate a control vector And fusing the fusion latent vector with the image latent vector to generate a fusion latent vector, which is expressed as:

;

Wherein the method comprises the steps of The fusion latent vector is represented as a result of the fusion,Representing control weight, training and optimizing according to the calibrated training data to obtain the control weight;

Forward diffusion process is performed by gradually adding noise based on the fusion latent vectors, so that each image latent vector becomes more blurred, expressed as:

;

Wherein the method comprises the steps of AndThe fusion potential vectors of the t and t-1 time steps are respectively shown,Representing the decay factor of step t, (determined based on historical experience),Representing standard Gaussian noise, the covariance being noise of the identity matrix;

Constructing a diffusion model according to a U-Net network, performing inverse denoising training, recovering a latent vector of an original image from a noise state, and training model prediction noise by using a mean square error loss MSE, wherein the model prediction noise is expressed as:

;

Wherein the method comprises the steps of Representing the loss of computation and,Representing a U-Net network model for predicting noise;

gradient descent optimization is carried out by using an Adam optimizer, parameters of the model are updated, and iteration is stopped when the loss of the model is not obviously reduced in the continuous iteration process;

and restoring the latent vector generated by the trained diffusion model according to noise input in the latent space and the condition vector generation condition, and reconstructing the restored latent vector into image data by a decoder.

The medical data and the unified latent vector of the metadata are paired, the generated condition vector can provide more comprehensive input information for the model, so that the model can simultaneously consider the potential characteristics of the medical data and the personalized metadata of a patient, the condition vector is controlled through the control Net module, the fine granularity control capability of the model on the medical data can be further enhanced, particularly in a medical data generation task, the influence of the external structural characteristics can be controlled to help the model to generate the medical data which accords with medical knowledge, for example, the anatomical structure of a specific part can be accurately reconstructed, or certain characteristics (such as lesion areas) of the medical data can be adjusted, the model can more accurately adjust the generation process according to the feedback of training data through optimized control weight, and the generated image can be ensured to accord with clinical demands;

The fusion latent vector is subjected to a forward diffusion process and noise is added, so that a model learns how to recover gradually from a fuzzy state, high-quality medical data can be generated better when unknown or fuzzy input is faced, the diffusion model is introduced, the diversity of generated medical data can be further improved, the model is allowed to generate medical data of different versions, and the exploration capability of the model is enhanced;

The diffusion model constructed based on the U-Net network can enable the model to gradually recover clear image latent vectors from noise through reverse denoising training, the model can accurately predict details of images in the denoising process, meanwhile, an efficient training process is maintained, and the use of the Adam optimizer ensures that parameters of the model can be stably updated, so that convergence is accelerated and the performance of the model is improved;

The trained diffusion model can generate a reduction latent vector according to noise input and a condition vector, the reduction latent vector is converted into image data through a decoder, the medical image data and the metadata are combined to generate the condition vector, the individuation information of a patient can be effectively introduced into an image generation process, the external structural characteristics can be considered in the image generation process due to the introduction of a control module (control Net), a specific anatomical structure can be accurately controlled in the generation process, the distortion of the generated image is effectively reduced, and the medical credibility of the image is improved;

The diffusion process of gradual denoising is introduced into the model, so that medical data generated by each reasoning has higher diversity, and the introduction of noise enables the model to generate various latent vector representations, so that medical data of different versions are output, the flexibility of the model in actual clinical application is enhanced, and more prediction options are provided especially in the aspects of disease progress prediction, personalized medical treatment and simulation of different clinical tests.

S3, the client deploys the diffusion model to perform data standardization, the aggregate model parameters update the global model, and the global model is evaluated;

Preferably, the aggregate model parameters update the global model, including,

Collecting client metadata according to client terminals, standardizing the metadata of each client by using a metadata Calibrator Meta-calizer, deploying a diffusion model and performing model training by using local data;

Determining a weighting coefficient for each client, and aggregating model parameters of all clients by using a weighted average method, and updating a global model according to the global model parameters, wherein the model parameters are expressed as:

;

Wherein the method comprises the steps of The parameters of the global model are represented as,Representing the total number of clients,Represents the data amount of client i, D represents the total data amount,Parameters representing the ith client model, including weights and biases for the diffusion model,As model weights for different clients.

The metadata calibrator is used to ensure that the data of all clients are processed under the same standard, so that the uniformity and consistency of the data are improved, the diffusion model is deployed and the local data are used for training, each client can be subjected to personalized adjustment according to the local data, the model performance of each client is improved, and the weighted average method is used for aggregating the model parameters of each client, so that the influence of each client is better balanced.

Further, the global model is evaluated, including,

Performing FID evaluation and SSIM evaluation on the global model;

According to the evaluation, the sum of the mean value and the standard deviation of the historical distribution difference value between the generated image and the real image is used as a difference threshold value, and if the distribution difference value is smaller than or equal to the difference threshold value, the original characteristics of the image are judged to be reserved;

And taking the sum of the mean value and the standard deviation of the historical structure difference value between the generated image and the real image as a structure threshold value according to the evaluation, and judging that the original characteristics of the image are reserved if the structure difference value is greater than or equal to the structure threshold value.

The improvement and the fidelity of the image quality can be effectively measured by using the FID and SSIM to evaluate the difference between the generated image and the real image, the FID evaluation can quantify the distribution difference between the generated image and the real image, and the SSIM evaluation focuses on the structural similarity of the images.

S4, the uncertainty of the global model analysis model is defined, the operation data of the knowledge graph record model is defined, an operation log is uploaded, and query sentences input by a user are converted and data query is carried out;

preferably, the model is analyzed for global model uncertainties, including,

Introducing a repeated sampling mechanism LAS aiming at the global model, randomly initializing an input image latent vector and a condition vector for a plurality of times, acquiring the output of the model under different initialization conditions through a plurality of times of sampling, thereby acquiring uncertainty information of a result, acquiring a corresponding repeated sampling reduction latent vector, calculating an average predicted value and an average standard deviation, and generating an uncertainty score;

And according to the sum of the average value and the double standard deviation of the historical uncertainty score as a score threshold, if the uncertainty score is smaller than the score threshold, judging that the uncertainty is lower.

By introducing a repeated sampling mechanism (LAS), the model can randomly initialize the input image latent vector and the condition vector for a plurality of times to acquire the output under different initialization conditions, thereby better reflecting the uncertainty of the generated result;

The average predicted value and standard deviation of the restored latent vector obtained by multiple sampling are calculated, an uncertainty score can be generated for each output, and the method can effectively identify which generated results have higher stability and which have larger uncertainty.

Further, knowledge graph record model operation data is defined, including,

The DataHub architecture is used to define a triplet structure of the knowledge graph KG, and input/output mapping, data sources, processing steps, model parameters and execution environment information of the model are recorded, wherein each triplet comprises an entity (such as a data set, the model and an inference result), an attribute (such as a data type, a calculation mode and a model state) and a relationship (such as a mapping relationship between the model and the data set).

By using DataHub architecture to define a triplet structure of a Knowledge Graph (KG), input/output mapping, data sources, processing steps, model parameters and execution environment information of a model can be systematically recorded and managed, the structured information storage mode provides a clear framework for data tracing, management and analysis, transparency of data flow and model operation is ensured, the triplet form enables definition of each entity, attribute and relation to be more definite, and cross-system or team cooperation and information sharing are facilitated, so that traceability, maintainability and high efficiency of a workflow are improved.

Further, uploading the operation log refers to recording the operation log for each operation step of the model, and storing the operation log as structured metadata for auditing and review.

All metadata are indexed and associated to specific training processes, reasoning results or abnormal events, and a user can quickly find all operation histories related to a certain model or data set by using a map query interface;

By using DataHub architecture to define a triplet structure of a Knowledge Graph (KG), input/output mapping, data sources, processing steps, model parameters and execution environment information of a model can be systematically recorded and managed, transparency of data flow and model operation is ensured, definition of each entity, attribute and relationship is more definite by the triplet, cooperation and information sharing across systems or teams are facilitated, and traceability, maintainability and high efficiency of a workflow are improved.

Further, the query statement input by the user is converted and data query is performed, including,

Converting a query sentence input by a user into an SPARQL sentence by using the BERT big model, and querying metadata in the knowledge graph;

according to the user inquiry, the system returns detailed information of corresponding model paths, reasoning conditions and input data;

By converting the query statement of the user into the SPARQL statement by using the BERT big model, more natural and intelligent user interaction can be realized, so that the user can acquire information related to metadata in the knowledge graph through simple query, the automatic conversion improves the flexibility and efficiency of query, and the system can quickly return related model paths, reasoning conditions and detailed information of input data according to the user requirements, thereby optimizing the application and user experience of the knowledge graph.

In summary, the invention ensures that different types of features can be fused in the same potential space through detailed segmentation covering the complete range of variables and unified metadata latent vectors by carrying out Gaussian kernel tag distribution coding of continuous metadata and coding of discrete metadata, ensures the comparability of the features, introduces a gradual denoising diffusion process into a model, ensures that medical data generated by each reasoning has higher diversity, leads the model to generate various latent vector representations through the introduction of noise, enhances the flexibility of the model in actual clinical application, and particularly provides more prediction options in the aspects of disease progress prediction, personalized medical treatment and simulation of different clinical tests.

It should be noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention may be modified or substituted without departing from the spirit and scope of the technical solution of the present invention, which is intended to be covered in the scope of the claims of the present invention.

Claims (9)

1. An artificial intelligence model management system based on metadata, comprising:

The metadata processing module is used for acquiring metadata, wherein the metadata comprises continuous metadata and discrete metadata, calculating field center points of the continuous metadata to be encoded into tag distribution vectors, counting the total number of categories of the discrete metadata, encoding to construct a mapping matrix, vector splicing, and linear mapping by using a linear transformation matrix to be converted into uniform length latent vectors;

the medical data processing module is used for matching the image data with the unified latent vector of the metadata according to the corresponding medical image data of the clinical data, extracting the image latent vector of the medical image data by using a pre-trained image encoder, and fusing the image latent vector with the unified latent vector of the metadata to generate a condition vector;

Determining external structural features of medical image data, including key points, labels and structural information of the image, introducing a control Net control module to take a condition vector and the external structural features as inputs to generate a control vector Fusing the fusion latent vector with the image latent vector to generate a fusion latent vector;

Forward diffusion is carried out by gradually adding noise based on the fusion latent vector;

Constructing a diffusion model according to the U-Net network, performing reverse denoising training, recovering a latent vector of an original image from a noise state, and training model prediction noise by using a mean square error loss MSE;

gradient descent optimization is carried out by using an Adam optimizer, parameters of the model are updated, and iteration is stopped when the loss of the model is not obviously reduced in the continuous iteration process;

Generating a restored latent vector through the trained diffusion model according to noise input in the latent space and the condition vector generation condition;

the model deployment module is used for carrying out data standardization on the client deployment diffusion model, aggregating model parameters, updating the global model and evaluating the global model;

And the query module is used for analyzing the model uncertainty of the global model, defining a knowledge graph, recording model operation data, uploading an operation log, converting query sentences input by a user and carrying out data query.

2. The artificial intelligence model management system based on metadata as set forth in claim 1 wherein said performing vector concatenation, linear mapping using a linear transformation matrix, translates into uniform length latent vectors comprises,

For each field of which the type is continuous metadata, calculating the center point of each field according to the variable range and the fraction segment of the field sample value;

Constructing a Gaussian kernel tag distribution function, encoding the value of the Gaussian kernel tag distribution function into a vector, and mapping an input value x of a field into a tag distribution vector;

for each field with discrete metadata, counting the unique value number of all samples in the field as the total number of categories Embedding coding is carried out to construct a mapping matrix with the dimension ofD is a fixed embedding dimension;

each row represents a category of embedded vectors, and the embedded vector representation of each discrete variable is obtained;

splicing the embedded vector representation of the tag distribution vector and the discrete metadata, wherein the output vector dimension of all continuous variables is p, the dimension of the embedded vector of all discrete variables is d, and the total number of the continuous variables is calculated And the total number of discrete variablesDefining a spliced dimension D;

and sequentially splicing the coded vectors of all the fields into a total vector, and performing linear mapping by using a linear transformation matrix to convert the total length latent vector.

3. The artificial intelligence model management system based on metadata as set forth in claim 2 wherein said aggregate model parameters update a global model comprising,

Collecting client metadata according to client terminals, standardizing the metadata of each client by using a metadata Calibrator Meta-calizer, deploying a diffusion model and performing model training by using local data;

and determining a weighting coefficient for each client, aggregating model parameters of all clients by using a weighted average method, and updating the global model according to the global model parameters.

4. The artificial intelligence model management system based on metadata as claimed in claim 3, wherein said evaluating the global model comprises,

Performing FID evaluation and SSIM evaluation on the global model;

According to the evaluation, the sum of the mean value and the standard deviation of the historical distribution difference value between the generated image and the real image is used as a difference threshold value, and if the distribution difference value is smaller than or equal to the difference threshold value, the original characteristics of the image are judged to be reserved;

And taking the sum of the mean value and the standard deviation of the historical structure difference value between the generated image and the real image as a structure threshold value according to the evaluation, and judging that the original characteristics of the image are reserved if the structure difference value is greater than or equal to the structure threshold value.

5. The artificial intelligence model management system based on metadata as claimed in claim 4, wherein said model uncertainty for the global model analysis comprises,

Introducing a repeated sampling mechanism LAS aiming at the global model, carrying out random initialization on the input image latent vector and the condition vector for a plurality of times, obtaining a corresponding repeated sampling reduction latent vector, calculating an average predicted value and an average standard deviation, and generating an uncertainty score;

And according to the sum of the average value and the double standard deviation of the historical uncertainty score as a score threshold, if the uncertainty score is smaller than the score threshold, judging that the uncertainty is lower.

6. The artificial intelligence model management system based on metadata according to claim 5, wherein the collected metadata includes continuous metadata and discrete metadata, including,

And reading the value type and the value distribution range of each column by traversing each column of field in the original clinical data table, marking as continuous metadata if the field value is of a continuous digital type, marking as discrete metadata if the field value is of a discrete classification label, and taking each row in the clinical data table as a sample.

7. The system of claim 6, wherein defining a knowledge graph, recording model operation data refers to defining a triplet structure of the knowledge graph KG using DataHub architecture, recording input/output mapping, data source, processing steps, model parameters and execution environment information of the model.

8. The artificial intelligence model management system based on metadata as set forth in claim 7 wherein uploading an oplog is referred to as logging the oplog for each operation step of the model and storing as structured metadata.

9. The artificial intelligence model management system based on metadata as claimed in claim 8, wherein said converting the query sentence input by the user and performing the data query includes,

Converting a query sentence input by a user into an SPARQL sentence by using the BERT big model, and querying metadata in the knowledge graph;

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