Self-evolution medical artificial intelligence model optimization system and method

Through the self-evolving medical artificial intelligence model optimization system, the problem of lack of persistent learning and feedback processing in the existing technology is solved through the self-evolving medical artificial intelligence model optimization system, and the continuous update of medical models and the improvement of diagnostic accuracy are achieved, improving the improvement of the medical model. Medical process efficiency.

CN118919055AInactive Publication Date: 2024-11-08ZHIYUANHUIYI (SUZHOU) HEALTH & MEDICAL TECH CO LTD

Patent Information

Application Number
CN202410940773.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing medical imaging artificial intelligence platform lacks the ability to persistent learning and feedback processing mechanism, and cannot effectively utilize the newly added dynamic medical data, resulting in inaccuracy of diagnosis and process efficiency.

Method used

A self-evolving medical artificial intelligence model optimization system was designed, including data acquisition module, data preprocessing module and model self-evolution training module. Through incremental learning and meta-learning strategies, the model is continuously updated and optimized, and data management is available. and model management functions.

Benefits of technology

It realizes the continuous learning and anti-forgetfulness of medical models, provides more accurate diagnostic support, improves the efficiency of medical processes, and supports online reviews by doctors and model updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118919055A_ABST
    Figure CN118919055A_ABST
Patent Text Reader

Abstract

The invention discloses a self-evolution medical artificial intelligence model optimization system and method. The system comprises a data acquisition module, a data preprocessing module, a model self-evolution training module and a visual management module. The data acquisition module is used for acquiring medical data, and the medical data comprises medical image data, medical report data and corresponding annotation data; the data preprocessing module is used for preprocessing the medical data to obtain a preprocessed sample data set; the model self-evolution training module performs self-evolution training on a to-be-trained medical model based on the sample data set to obtain a trained medical model; the visual management module is used for providing a human-computer interaction interface for a user to select different to-be-trained medical models, and storing training results and training process information of the models at the same time. The trained medical model has continuous learning ability and anti-forgetting ability, and a more accurate medical diagnosis model can be continuously and efficiently provided for doctors.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging artificial intelligence and information technology, and specifically relates to a self-evolving medical artificial intelligence model optimization system and method. Background Art

[0002] With the widespread application of various medical imaging equipment, imaging data has shown a large-scale growth. At the same time, the development of big data technology has enabled a large number of artificial intelligence imaging-assisted products to be put into clinical use, further giving rise to the concept of medical imaging artificial intelligence training platform, opening a new door for assisting doctors to improve diagnosis and treatment efficiency. Since the model design schemes of artificial intelligence platforms in the current mainstream medical field are all based on static data training, they do not have the ability to incrementally train the dynamic data that is constantly added to the clinic, wasting the training potential of new medical data. In addition, artificial intelligence platforms also face considerable challenges in terms of unified data management and automated model operation. Therefore, there are two major problems in the use of artificial intelligence training platforms: lack of persistent learning capabilities and lack of feedback processing mechanisms.

[0003] Self-evolving AI is AI that learns and improves itself automatically. In order to achieve this goal, machine learning algorithms in various directions are being actively explored, such as incremental learning, online learning, meta-learning, and transfer learning. With the continuous upgrading and widespread popularization of medical equipment, medical institutions around the world generate a large amount of medical imaging data every day. Therefore, in order to improve diagnostic accuracy, accelerate medical processes, and provide greater decision-making support for doctors, the demand for self-evolving AI models is also becoming more and more vigorous driven by medical big data. Summary of the invention

[0004] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0005] A self-evolving medical artificial intelligence model optimization system, comprising: a data acquisition module, a data preprocessing module, a model self-evolution training module and a visualization management module;

[0006] The data acquisition module is used to collect medical data, and the medical data includes: medical image data, medical report data and corresponding annotation data;

[0007] The data preprocessing module is used to preprocess the medical data to obtain a preprocessed sample data set;

[0008] The model self-evolution training module performs self-evolution training on the medical model to be trained based on the sample data set to obtain a trained medical model;

[0009] The visualization management module is used to provide a human-computer interaction interface for users to select different medical models to be trained, and to store the training results and training process information of the models.

[0010] Preferably, the data acquisition module includes: an active pulling unit, a passive receiving unit and a first storage unit;

[0011] The active pulling unit is used to regularly pull the medical data from the third-party PACS / RIS system through the intermediate table and NFS protocol;

[0012] The passive receiving unit is used to provide background services to receive the medical data sent by the third-party PACS / RIS system;

[0013] The first storage unit is used to analyze the medical image data and store the analyzed medical image data, the medical report data and the annotated data.

[0014] Preferably, the data preprocessing module includes: a continuity checking unit, an image anonymity unit and a sample preprocessing unit;

[0015] The continuity checking unit is used to check the continuity of the medical image data according to the image number of the medical image data to ensure the continuity of the image;

[0016] The image anonymization unit is used to remove patient privacy information, attending physician privacy information and hospital information from the medical data;

[0017] The sample processing unit is used to process the medical data to obtain the processed sample data set.

[0018] Preferably, the sample processing unit comprises: an image data processing subunit, a report data processing subunit and a label data processing subunit;

[0019] The image data processing subunit is used to classify the medical image data according to the modality, and perform interpolation to adjust the image size of each category to complete the image data processing;

[0020] The report data processing subunit is used to extract the imaging manifestations and diagnostic conclusions in the report, respectively, to form semi-structured data, and complete the report data processing;

[0021] The annotated data processing subunit is used to verify the correspondence between the annotated data and the medical imaging data, and complete the annotated data processing.

[0022] Preferably, the model self-evolution training module includes: a training sample generation unit, an internal drive training unit, an external drive training unit and a second storage unit;

[0023] The training sample generating unit generates a training data set required for model training by using the sample data set and the historical sample set, wherein the historical sample set includes: a historical sample data set and a historical model;

[0024] The internally driven training unit establishes a plurality of parallel processing flows by using a regularization method in an incremental learning strategy, and updates the medical model to be trained based on the training data set to obtain an updated model;

[0025] The external drive training unit optimizes and learns the updated model using a meta-update strategy to obtain the trained medical model;

[0026] The second storage unit is used to move the sample data set and the model to be trained before training into the historical sample set for storage.

[0027] The present invention also provides a self-evolving medical artificial intelligence model optimization method, which is applied to any of the above-mentioned optimization systems and includes the following steps:

[0028] Collecting medical data, the medical data including: medical imaging data, medical report data and corresponding annotation data;

[0029] Preprocessing the medical data to obtain a preprocessed sample data set;

[0030] The medical model to be trained is self-evolved based on the sample data set to obtain a trained medical model.

[0031] Preferably, the method for collecting medical data includes: active pull mode and passive receive mode

[0032] The active pull mode is to regularly pull the medical data from the third-party PACS / RIS system through the intermediate table and NFS protocol;

[0033] The passive receiving mode is to provide background services to receive the medical data sent by the third-party PACS / RIS system;

[0034] The medical image data is then parsed, and the parsed medical image data, the medical report data, and the annotated data are stored.

[0035] Preferably, the pretreatment method comprises:

[0036] Checking the continuity of the medical image data according to the image number of the medical image data to ensure the continuity of the image;

[0037] Removing patient privacy information, attending physician privacy information and hospital information from the medical data;

[0038] The medical data is processed to obtain the processed sample data set.

[0039] Preferably, the method for processing the medical data includes:

[0040] For the medical image data, the medical image data is classified according to the modality, and the image size is adjusted by interpolation for each category to complete the image data processing;

[0041] Extract the imaging manifestations and diagnostic conclusions in the report separately to form semi-structured data and complete the report data processing;

[0042] Verify the correspondence between the labeled data and the medical imaging data to complete the labeled data processing.

[0043] Preferably, the self-evolution training method comprises:

[0044] Generate a training data set required for model training using the sample data set and the historical sample set, wherein the historical sample set includes: a historical sample data set and a historical model;

[0045] Using the regularization method in the incremental learning strategy to establish several parallel processing flows, and performing internally driven training updates on the medical model to be trained based on the training data set to obtain an updated model;

[0046] Using a meta-update strategy to perform external-driven optimization learning on the updated model to obtain the trained medical model;

[0047] The sample data set and the model to be trained before training are moved into the historical sample set for storage.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The present invention designs a self-evolution optimization method for medical artificial intelligence models. The medical model using this method has continuous learning ability and anti-forgetting ability, and can continuously and efficiently provide doctors with more accurate medical models.

[0050] (2) The present invention also constructs a medical artificial intelligence model optimization platform based on self-evolution, which is equipped with a data management system and a model management system. It can deploy medical models and perform self-evolution training on the models. The platform supports doctors to access relevant information online and perform auxiliary diagnosis, and also supports engineers to upload and update medical models. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in 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 creative labor.

[0052] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of a method flow of an embodiment of the present invention;

[0054] Figure 3 This is a block diagram of a model training method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Embodiment 1

[0058] In this embodiment, if Figure 1 As shown, a self-evolving medical artificial intelligence model optimization system includes: a data acquisition module, a data preprocessing module, a model self-evolution training module and a visualization management module.

[0059] The data acquisition module is used to collect medical data, which includes: medical imaging data, medical report data and corresponding annotation data.

[0060] The data acquisition module includes: an active pulling unit, a passive receiving unit and a first storage unit; the active pulling unit is used to periodically pull medical data from a third-party PACS / RIS system through an intermediate table and an NFS protocol; the passive receiving unit is used to provide background services to receive medical data sent by a third-party PACS / RIS system; the first storage unit is used to perform DICOM standard parsing on medical image data, and store the parsed medical image data, medical report data and annotation data.

[0061] In this embodiment, the three types of data are associated through the serial number of the examination, and one examination corresponds to a set of data. The medical report data is directly extracted from the corresponding database, the medical image data is obtained and saved locally through the DICOM standard, and the annotation data is read through the JSON file.

[0062] The data preprocessing module is used to preprocess the medical data to obtain a preprocessed sample data set.

[0063] The data preprocessing module includes: a continuity check unit, an image anonymization unit and a sample preprocessing unit. The continuity check unit is used to check the continuity of the medical image data according to the image number of the medical image data to ensure the continuity of the image; the image anonymization unit is used to remove the patient privacy information, the attending physician privacy information and the hospital information in the medical data; the sample processing unit is used to process the medical data to obtain the processed sample data set.

[0064] The sample processing unit includes: an image data processing subunit, a report data processing subunit and an annotation data processing subunit. The image data processing subunit is used to classify the medical image data according to the modality, and interpolate and adjust the image size (including height, width and thickness) of each category to ensure that the image size within the modality is consistent and complete the image data processing. In this embodiment, the modality includes: digital X-ray imaging (DX), computed tomography imaging (CT), magnetic resonance imaging (MR), and ultrasound imaging (US); the report data processing subunit is used to extract the image manifestations and diagnostic conclusions in the report respectively to form semi-structured data and complete the report data processing; the annotation data processing subunit is used to verify the correspondence between the annotation data and the medical image data and complete the annotation data processing.

[0065] The model self-evolution training module performs self-evolution training on the medical model to be trained based on the sample data set to obtain the trained medical model.

[0066] The model self-evolution training module includes: a training sample generation unit, an internal drive training unit, an external drive training unit and a second storage unit. The training sample generation unit uses the sample data set and the historical sample set to generate the training data set required for model training. The historical sample set includes: a historical sample data set and a historical model; the internal drive training unit uses the regularization method in the incremental learning strategy to establish several parallel processing flows, and updates the medical model to be trained based on the training data set to obtain an updated model; the external drive training unit uses the meta-update strategy to optimize the learning of the updated model to obtain a trained medical model; the second storage unit is used to move the sample data set and the model to be trained before training into the historical sample set for storage.

[0067] In this embodiment, first, the training sample generation unit uses the sample data set and the historical sample set to generate the training data set required for model training, and adds the training data set as a new task sample to the model training; then, k-1 samples are randomly selected from the historical sample set and combined with the new task sample Dt to form a task set of k, and an internal-driven and external-driven dual-drive training method is adopted. The internal-driven training unit adopts the regularization method in the incremental learning strategy to establish k parallel processing streams, and updates the basic model respectively; the external-driven training unit adopts the meta-update strategy to optimize the learning of the internal-driven results, so as to finally obtain the updated model; after the model update is completed, the sample data set and the model to be trained before training are moved into the historical sample set and stored in the second storage unit.

[0068] In this embodiment, the model self-evolution module can also carry out different forms of self-evolution training according to the different architectures of the model to be trained:

[0069] (1) If the model is a conventional architecture, that is, an architecture in which the entire model can be trained simultaneously, the self-evolution training method is adopted. The preprocessed image data and annotated data samples are added to the model training in batches as new tasks. Randomly extract k-1 samples from the historical sample set and the new task sample Dt to form a task set of k. The dual-drive training method of internal drive and external drive is adopted. The internal drive training unit uses the regularization method in the incremental learning strategy to establish k parallel processing streams to update the basic model respectively; the external drive training unit uses the meta-update strategy to optimize the learning of the internal drive results, so as to finally obtain the updated model. If the historical data is empty, only the newly added data is used for conventional training, and the self-evolution training method is not used, and the trained model is set as the basic model.

[0070] (2) If the model is a serial architecture, the serial self-evolution training method is used. The serial architecture means that the model needs to be trained in sequence from sub-model 1, sub-model 2, sub-model 3, etc., and the subsequent model can only be trained after the previous model is trained. The method of serial self-evolution training is similar to the architecture. First, the historical data set and the new data set are used to form a task set, and then the self-evolution training method of the conventional architecture in (1) is adopted. After obtaining the evolution model of sub-model 1, the data of the task set is input into the evolution model of sub-model 1 to obtain the task set of sub-model 2. Use the task set of sub-model 2 to carry out self-evolution training on sub-model 2. Subsequent training is carried out in this way to complete the self-evolution training of the serial architecture model.

[0071] (3) If the model is a parallel architecture, the parallel self-evolution training method is used. The parallel architecture means that the model contains multiple sub-models such as sub-model 1 and sub-model 2. Each sub-model can be trained independently; the training results of the sub-model can affect the training of other models. The method of parallel self-evolution training is similar to the architecture. First, the historical data set and the new data set are used to form a task set, and then the conventional architecture self-evolution training in (1) is adopted for multiple sub-models. According to the specific task objectives of the model, the evolution model of the sub-model can be used to perform self-evolution training on other sub-models again.

[0072] Here, we take the Generative Adversarial Network (GAN) model as an example. The network contains a generator model G and a discriminator model D. The goal of model G is to generate new data from the input data, and the goal of model D is to determine whether the data is generated data. When the generative adversarial network is trained in parallel, the same task set is used to carry out self-evolution training for model G and model D respectively. Then, according to the task goal of GAN (generation task or discrimination task), self-evolution training is carried out again. If it is a discrimination task, the evolution model of model G is added to the next self-evolution training of model D, in which the goal of model D has changed. It is necessary not only to determine whether the data in the original task set is generated data, but also to determine whether the data after the original task set is generated by model G is generated data. If it is a generation task, the evolution model of model D is added to the next self-evolution training of model G, in which the goal of model G has changed. It is necessary not only to generate new data according to the original task set, but also to prevent the generated data from being recognized by model D. Finally, the self-evolution module completes the secondary self-evolution training for the specific sub-model according to the task goal. This specific sub-model is the model used for prediction.

[0073] The visualization management module is used to provide a human-computer interaction interface for users to select different medical models to be trained, and to store the model's training results and training process information.

[0074] In this embodiment, the visualization management module includes an acquisition center, a model training center, a system management unit and a database. The acquisition center is used to display the list of collected data, which is divided into a patient list, a report list and an image list. The patient list displays the number of reports and images of the same patient ID. Click the report number to enter the report list of the patient, and click the image number to enter the image list of the patient. If there is no automatically corresponding report for an image of a certain examination, manual verification and matching are performed. At the same time, the list has a re-collection button to re-collect the relevant data in the active mode. The model training center is used to display the model group list and the data set list, where the model group is a group of models defined for serial execution to complete a certain diagnostic task. The model group list has the functions of model group definition, model import, model training startup, training progress, and training result display. The data set list displays the data set corresponding to the model and the data usage. The system management unit includes model configuration, system parameter configuration, user management, and storage configuration. The database consists of a resource library and a feature training library. The resource library is modeled with patients as the center. Medical report data, medical imaging data and corresponding annotation data are matched and saved. The feature training library saves the key information or feature values ​​of various data and the management information of the model and training process.

[0075] Embodiment 2

[0076] In this embodiment, if Figure 2 As shown, a self-evolving medical artificial intelligence model optimization method includes the following steps:

[0077] S1. Collect medical data, which includes: medical imaging data, medical report data and corresponding annotation data.

[0078] The methods for collecting medical data include: active pull mode and passive reception mode. The active pull mode is to regularly pull medical data from the third-party PACS / RIS system through the intermediate table and NFS protocol; the passive reception mode is to provide background services to receive medical data sent by the third-party PACS / RIS system; then parse the medical image data, and store the parsed medical image data, medical report data and annotation data.

[0079] In this embodiment, the three types of data are associated through the serial number of the examination, and one examination corresponds to a set of data. The medical report data is directly extracted from the corresponding database, the medical image data is obtained and saved locally through the DICOM standard, and the annotation data is read through the JSON file.

[0080] S2. Preprocess the medical data to obtain a preprocessed sample data set.

[0081] The preprocessing method includes: checking the continuity of the medical image data according to the image number of the medical image data to ensure the continuity of the image; removing the patient privacy information, the attending physician privacy information and the hospital information in the medical data; processing the medical data to obtain a processed sample data set.

[0082] The method for processing medical data includes: for medical imaging data, classifying the medical imaging data according to the modality, and interpolating and adjusting the image size (including height, width and thickness) of each category to ensure that the image size within the modality is consistent, and completing the image data processing. In this embodiment, the modalities include: digital X-ray imaging (DX), computed tomography imaging (CT), magnetic resonance imaging (MR), and ultrasound imaging (US); extracting the image manifestations and diagnostic conclusions in the report respectively to form semi-structured data, and completing the report data processing; verifying the correspondence between the labeled data and the medical imaging data, and completing the labeled data processing.

[0083] S3. Perform self-evolution training on the medical model to be trained based on the sample data set to obtain a trained medical model.

[0084] The method of self-evolutionary training includes: using a sample data set and a historical sample set to generate a training data set required for model training, the historical sample set includes: a historical sample data set and a historical model; using the regularization method in the incremental learning strategy to establish several parallel processing streams, and based on the training data set, performing internally driven training updates on the medical model to be trained to obtain an updated model; using a meta-update strategy to perform externally driven optimization learning on the updated model to obtain a trained medical model; and moving the sample data set and the model to be trained before training into the historical sample set for storage.

[0085] In this embodiment, if Figure 3 As shown, firstly, the sample data set and the historical sample set are used to generate the training data set required for model training, and the training data set is added to the model training as a new task sample; then, k-1 samples are randomly selected from the historical sample set and the new task sample Dt is combined to form a task set of k, and a dual-drive training method of internal drive and external drive is adopted. The internal drive is to establish k parallel processing streams by using the regularization method in the incremental learning strategy, and update the basic model respectively; the external drive is to optimize the learning of the internal drive result by using the meta-update strategy, so as to finally obtain the updated model; after the model update is completed, the sample data set and the model to be trained before training are moved into the historical sample set and stored in the second storage unit.

[0086] In this embodiment, the model self-evolution module can also carry out different forms of self-evolution training according to the different architectures of the model to be trained:

[0087] (1) If the model is a conventional architecture, that is, an architecture in which the entire model can be trained simultaneously, the self-evolution training method is adopted. The preprocessed image data and annotated data samples are added to the model training in batches as new tasks. Randomly extract k-1 samples from the historical sample set and the new task sample Dt to form a task set of k. The dual-drive training method of internal drive and external drive is adopted. The internal drive training unit uses the regularization method in the incremental learning strategy to establish k parallel processing streams to update the basic model respectively; the external drive training unit uses the meta-update strategy to optimize the learning of the internal drive results, so as to finally obtain the updated model. If the historical data is empty, only the newly added data is used for conventional training, and the self-evolution training method is not used, and the trained model is set as the basic model.

[0088] (2) If the model is a serial architecture, the serial self-evolution training method is used. The serial architecture means that the model needs to be trained in sequence from sub-model 1, sub-model 2, sub-model 3, etc., and the subsequent model can only be trained after the previous model is trained. The method of serial self-evolution training is similar to the architecture. First, the historical data set and the new data set are used to form a task set, and then the self-evolution training method of the conventional architecture in (1) is adopted. After obtaining the evolution model of sub-model 1, the data of the task set is input into the evolution model of sub-model 1 to obtain the task set of sub-model 2. Use the task set of sub-model 2 to carry out self-evolution training on sub-model 2. Subsequent training is carried out in this way to complete the self-evolution training of the serial architecture model.

[0089] (3) If the model is a parallel architecture, the parallel self-evolution training method is used. The parallel architecture means that the model contains multiple sub-models such as sub-model 1 and sub-model 2. Each sub-model can be trained independently; the training results of the sub-model can affect the training of other models. The method of parallel self-evolution training is similar to the architecture. First, the historical data set and the new data set are used to form a task set, and then the conventional architecture self-evolution training in (1) is adopted for multiple sub-models. According to the specific task objectives of the model, the evolution model of the sub-model can be used to perform self-evolution training on other sub-models again.

[0090] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A self-evolving medical artificial intelligence model optimization system, It is characterized in that include: Data acquisition module, data preprocessing module, model self-evolution training module and visualization management module; The data acquisition module is used to collect medical data, and the medical data includes: medical image data, medical report data and corresponding annotation data; The data preprocessing module is used to preprocess the medical data to obtain a preprocessed sample data set; The model self-evolution training module performs self-evolution training on the medical model to be trained based on the sample data set to obtain a trained medical model; The visualization management module is used to provide a human-computer interaction interface for users to select different medical models to be trained, and to store the training results and training process information of the models.

2. According to the self-evolving medical artificial intelligence model optimization system of claim 1, It is characterized in that The data acquisition module includes: an active pulling unit, a passive receiving unit and a first storage unit; The active pulling unit is used to periodically pull the medical data from the third-party PACS / RIS system through the intermediate table and NFS protocol; The passive receiving unit is used to provide background services to receive the medical data sent by the third-party PACS / RIS system; The first storage unit is used to parse the medical image data and store the parsed medical image data, the medical report data and the annotated data.

3. According to the self-evolving medical artificial intelligence model optimization system of claim 1, It is characterized in that The data preprocessing module includes: a continuity checking unit, an image anonymity unit and a sample preprocessing unit; The continuity checking unit is used to check the continuity of the medical image data according to the image number of the medical image data to ensure the continuity of the image; The image anonymization unit is used to remove patient privacy information, attending physician privacy information and hospital information from the medical data; The sample processing unit is used to process the medical data to obtain the processed sample data set.

4. According to the self-evolving medical artificial intelligence model optimization system of claim 3, It is characterized in that The sample processing unit includes: an image data processing subunit, a report data processing subunit and a label data processing subunit; The image data processing subunit is used to classify the medical image data according to the modality, and perform interpolation to adjust the image size of each category to complete the image data processing; The report data processing subunit is used to extract the imaging manifestations and diagnostic conclusions in the report, respectively, to form semi-structured data, and complete the report data processing; The annotated data processing subunit is used to verify the correspondence between the annotated data and the medical imaging data, and complete the annotated data processing.

5. According to the self-evolving medical artificial intelligence model optimization system of claim 1, It is characterized in that The model self-evolution training module includes: a training sample generation unit, an internal drive training unit, an external drive training unit and a second storage unit; The training sample generating unit generates a training data set required for model training by using the sample data set and the historical sample set, wherein the historical sample set includes: a historical sample data set and a historical model; The internally driven training unit establishes a plurality of parallel processing flows by using a regularization method in an incremental learning strategy, and updates the medical model to be trained based on the training data set to obtain an updated model; The external drive training unit optimizes and learns the updated model using a meta-update strategy to obtain the trained medical model; The second storage unit is used to move the sample data set and the model to be trained before training into the historical sample set for storage.

6. A self-evolving medical artificial intelligence model optimization method, the optimization method is applied to the optimization system according to any one of claims 1 to 5, It is characterized in that The following steps are involved: Collecting medical data, the medical data including: medical imaging data, medical report data and corresponding annotation data; Preprocessing the medical data to obtain a preprocessed sample data set; The medical model to be trained is self-evolved based on the sample data set to obtain a trained medical model.

7. According to the self-evolving medical artificial intelligence model optimization method of claim 6, It is characterized in that The method for collecting medical data includes: active pull mode and passive reception mode The active pull mode is to regularly pull the medical data from the third-party PACS / RIS system through the intermediate table and NFS protocol; The passive receiving mode is to provide background services to receive the medical data sent by the third-party PACS / RIS system; The medical image data is then parsed, and the parsed medical image data, the medical report data, and the annotated data are stored.

8. According to the self-evolving medical artificial intelligence model optimization method of claim 6, It is characterized in that The pretreatment method comprises: Checking the continuity of the medical image data according to the image number of the medical image data to ensure the continuity of the image; Removing patient privacy information, attending physician privacy information and hospital information from the medical data; The medical data is processed to obtain the processed sample data set.

9. According to the self-evolving medical artificial intelligence model optimization method of claim 8, It is characterized in that The method for processing the medical data includes: For the medical image data, the medical image data is classified according to the modality, and the image size is adjusted by interpolation for each category to complete the image data processing; Extract the imaging manifestations and diagnostic conclusions in the report separately to form semi-structured data and complete the report data processing; Verify the correspondence between the labeled data and the medical imaging data to complete the labeled data processing.

10. According to the self-evolving medical artificial intelligence model optimization method of claim 6, It is characterized in that The self-evolution training method comprises: Generate a training data set required for model training using the sample data set and the historical sample set, wherein the historical sample set includes: a historical sample data set and a historical model; Using the regularization method in the incremental learning strategy to establish several parallel processing flows, and performing internally driven training updates on the medical model to be trained based on the training data set to obtain an updated model; Using a meta-update strategy to perform external-driven optimization learning on the updated model to obtain the trained medical model; The sample data set and the model to be trained before training are moved into the historical sample set for storage.

Citation Information

Patent Citations

  • Information management system integrating image and scientific research service

    CN113178250A

  • Data update detection system and method based on machine learning

    CN113392118A

  • Joint hospital data migration method and system based on few-sample model learning

    CN113986890A

  • Personalized medical image diagnosis path generation system and method

    CN116469542A

  • Radio frequency fingerprint identification method and device for small sample open-set incremental learning based on meta-learning

    CN118245872A

Cited By

  • Remote auxiliary diagnosis method and system based on multi-modal medical data

    CN120878187A