Intelligent medical AI model training method and system and storage medium

Abstract

The invention discloses an intelligent medical AI model training method and system and a storage medium, and belongs to the field of medical intelligent technology.The method comprises the steps that needed function modules and requirements are selected, and specific learning tasks are analyzed; according to the learning task, automatically identifying the required data and screening the selected data set; s2, preprocessing the data obtained in the step S2, wherein the preprocessing operation comprises cleaning and converting the original data; selecting a model according to the learning task and determining a model and a network architecture according to the data set information; building a model according to the automatically generated network architecture and model, and training the model; evaluating and deploying the model by using a test set in the data; and establishing a feedback mechanism, collecting use feedback of a clinician on the model, and performing adjustment and optimization according to the feedback. According to the method, the existing AI model training process can be optimized, and the training efficiency and updating efficiency of the model are improved.

Description

Technical Field

[0001] The present invention relates to the field of medical intelligent technology, and in particular to an intelligent medical AI model training method, system and storage medium. Background Art

[0002] With the rapid development of medical technology, artificial intelligence (AI) is increasingly used in the medical field, especially in auxiliary diagnosis, disease prediction and personalized treatment plan formulation. Developing a usable medical AI model requires close interdisciplinary cooperation, especially the need for constant communication between developers and doctors. This collaboration ensures that the AI ​​model is not only technically feasible, but also can solve practical clinical needs, meet the doctor's workflow and patient safety needs. Figure 1 This is a common flowchart for developers and doctors to collaborate on developing AI models. The specific process is as follows:

[0003] 1. Problem identification and demand definition: The development team has in-depth communication with doctors to identify the medical problems and needs that need to be solved based on the pain points in doctors’ daily work, common diagnostic problems, or the shortcomings of current AI models.

[0004] 2. Data collection and labeling: After developers clarify their needs, doctors assist developers in obtaining the medical data needed for development. Doctors are crucial in this process and need to label the data. For example, marking tumor areas in medical images or marking important symptoms and diagnostic information in medical records. Developers provide doctors with data labeling tools and work with doctors to develop labeling standards to ensure data consistency and reliability.

[0005] 3. Model training and verification: Developers build models after clarifying requirements, and train models on collected data to optimize model parameters; after model training, the model effect is verified through the test set, and developers review the results with doctors, especially the performance of the model in actual clinical scenarios. Doctors will provide feedback on the feasibility of the model in clinical practice based on the output results.

[0006] 4. User testing and model optimization: Developers give the trained model to doctors for testing in a real clinical environment, and developers further optimize the model based on the doctors' feedback.

[0007] 5. Clinical trials and real-world evaluation: After the model is optimized, developers work with doctors to conduct trials in a clinical environment. Doctors use AI models in actual diagnosis and treatment to evaluate whether they meet their needs and improve work efficiency. Developers can evaluate the model through quantitative indicators (such as sensitivity and specificity) and qualitative feedback (such as doctor experience).

[0008] 6. Model deployment and clinical integration: After the model passes clinical validation, the development team deploys the model into the doctors’ daily work environment. For example, the model is integrated into the hospital’s medical information system (HIS) or connected to an existing imaging analysis system.

[0009] 7. Continuous monitoring and feedback: After the model is launched, developers need to monitor its performance in the real environment and continuously optimize the model through feedback. Daily feedback from doctors will continue to help the development team improve the accuracy and practicality of the model. As new clinical data accumulates, developers will continue to update the model to adapt it to new medical discoveries or treatments.

[0010] Although the process of developers and doctors collaborating to develop AI models has gradually matured, there are still some obvious defects and challenges in the current process. These defects limit the breadth and depth of application of AI models in medical practice, affecting the actual effect and clinical implementation of the models, mainly reflected in:

[0011] 1. Data issues: Due to restrictions of privacy laws, medical data is difficult to obtain and the data sets are usually small and difficult to share, which will limit the training effect of the model and cause the model to overfit on small sample data sets and fail to generalize to a more diverse patient population. In addition, medical imaging and other data require a lot of manual annotation, which requires experts to spend a lot of time, and the subjectivity of annotation may also lead to unstable model performance.

[0012] 2. Communication gap between doctors and developers: When doctors and developers communicate, they often have misunderstandings due to different background knowledge. Doctors may not fully understand the limitations of technology development, and developers are not familiar with the specific needs of the medical field. This poor communication will lead to deviations in the goals and needs of model development.

[0013] 3. Insufficient feedback: Physician feedback is often concentrated in the final stages of the model (such as model testing and clinical validation), resulting in the fact that the needs in the early stages of model design may not be fully met. In addition, if the participation of doctors is insufficient, developers may deviate from actual clinical needs, resulting in low practicality of the model.

[0014] 4. Continuous update and maintenance issues: Medical data changes over time, which is called "data drift". In actual use, AI models may encounter samples that are different from the distribution of training data, resulting in performance degradation, which requires the model to be continuously updated based on new data. However, the maintenance and update of medical AI models requires a lot of manpower and material resources. The development team needs to continuously obtain new medical data to update the model, and also needs to monitor the performance of the model in clinical applications. Retraining and adjusting the model may also take a lot of time, increasing costs.

[0015] 5. Clinicians have limited time: Doctors are busy with their daily work and it is difficult for them to devote a lot of time to researching and training AI models. The need to manually customize and train models will further increase their burden.

[0016] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0017] The purpose of the present invention is to provide an intelligent medical AI model training method, system and storage medium, which can optimize the existing AI model training process so that doctors can also train model A that meets their needs and automatically deploy it to their daily work environment, thereby improving the training efficiency and update efficiency of the model.

[0018] To achieve the above object, the present invention provides an intelligent medical AI model training method, comprising the following steps:

[0019] S1: Select the required functional modules and requirements, and analyze the specific learning tasks;

[0020] S2: Automatically identify the required data and filter the selected data set according to the learning task;

[0021] S3: preprocessing the data obtained in step S2, the preprocessing operation includes cleaning and converting the original data;

[0022] S4: Select the model according to the learning task and determine the model and network architecture based on the dataset information;

[0023] S5: Train the model based on the automatically generated network architecture;

[0024] S6: Evaluate and deploy the model using the test set in the data;

[0025] S7: Establish a feedback mechanism to collect clinicians’ feedback on the use of the model and make adjustments and optimizations based on the feedback.

[0026] In one embodiment of the present invention, in step S2, the screening operation mainly includes: based on the specific learning task, screening out data directly related to the task; screening out patient data covering different ages, genders, races, and health statuses; screening out data at different stages of the disease to ensure that the model can learn the characteristics of various diseases; wherein the format and standard of the data must be consistent.

[0027] In one embodiment of the present invention, in step S3, data preprocessing mainly includes the following steps:

[0028] S301: Data cleaning, specifically including: data consistency check to ensure that the format or unit of the data is consistent; label missing processing, the data used in supervised learning lacks necessary labels and needs to be supplemented or automatically labeled, and when key data lacks labels, experts are asked to label the data; duplicate data removal, the system detects duplicate data by comparing multiple fields and deletes duplicate records;

[0029] S302: Statistical data set information, specifically including statistical data spacing, data size, and maximum and minimum values ​​of the grayscale value of the data set;

[0030] S303: Data standardization: standardize the cleaned data. For image data, adjust the data from different sources to the same resolution and then normalize the grayscale value or z-score.

[0031] In one embodiment of the present invention, in step S3, selecting a model and determining the model and network architecture according to the data set information mainly include the following steps:

[0032] S401: Automatically selecting a suitable model according to the identified learning task;

[0033] S402: Dynamically adjust the network topology of the model according to the data set information.

[0034] In one embodiment of the present invention, step S402 specifically includes the following steps:

[0035] S4021: Automatically select a suitable network structure according to the dimension of the data, and dynamically adjust the network input size and downsampling multiple according to the size of the input data;

[0036] S4022: Automatically adjust the size of the convolution kernel according to the resolution of the input data;

[0037] S4023: Dynamically adjust the number of downsampling and upsampling layers according to image size and computing resources;

[0038] S4044: Dynamically adjust the depth of the network according to the size of the feature map.

[0039] In one embodiment of the present invention, in step S5, the training process mainly includes:

[0040] S501: Split the data into training set, validation set and test set;

[0041] S502: Load the preprocessed data set, and load the training parameters according to the model and network architecture determined in step S4;

[0042] S503: Automatically allocate computing resources to ensure that the environment adapts to training requirements, start training tasks and record logs;

[0043] S504: Load the data into the memory in batches for training and verification;

[0044] S505: forward propagating the input data through the network to generate a prediction result, calculating the loss between the prediction result and the actual label, calculating the gradient based on the loss and updating the network weight;

[0045] S506: Monitor the training loss in real time, calculate the indicators using the validation set after each training cycle or a predetermined interval, and display the change curves of the training loss and the validation loss through a graphical tool.

[0046] In one embodiment of the present invention, in step S6, the evaluation and deployment process mainly includes:

[0047] S601: After the model training is completed, the model performance is evaluated on the divided test set;

[0048] S602: Test the trained model in a simulated clinical environment;

[0049] S603: Deploy the model verified by steps S601 and S602 to the application environment of the medical institution.

[0050] In one embodiment of the present invention, in step S7, the feedback process mainly includes:

[0051] S701: Collect new data generated by the model in actual applications, and then regularly evaluate the performance of the model and check whether the model needs to be updated or retrained;

[0052] S702: Adjustments are made to the original model based on the collected new data and user feedback, and the optimized model is automatically updated and deployed in the environment of the medical institution.

[0053] The present invention also provides an intelligent medical AI model training system, which is used to execute the above-mentioned intelligent medical AI model training method, including:

[0054] The requirement module is used for users to select the required functional modules and requirements and analyze the specific learning tasks;

[0055] The data screening module is used to automatically identify the required data and screen the selected data set according to the learning task;

[0056] A data preprocessing module is used to preprocess the obtained data, and the preprocessing operation includes cleaning and converting the original data;

[0057] Model selection module, used to select models according to learning tasks and determine models and network architectures based on dataset information;

[0058] Model training module, used to build models based on automatically generated network architecture and models, and to train the models;

[0059] Model evaluation module, used to evaluate and deploy the model using the test set in the data;

[0060] The feedback module is used to collect clinicians' feedback on the use of the model and make adjustments and optimizations based on the feedback.

[0061] The present invention also provides a storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned intelligent medical AI model training method are implemented.

[0062] Compared with the prior art, the intelligent medical AI model training method, system and storage medium according to the present invention enable developers to only deploy the system in the hospital, and the hospital staff can independently train the AI ​​model that meets their needs. In addition, the data security is guaranteed when training the model, the training efficiency is improved, and the time of the staff and doctors is saved. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flowchart of the existing model developed by developers and doctors in collaboration;

[0064] Figure 2 is a flowchart of an intelligent medical AI model training method according to one embodiment of the present invention;

[0065] Figure 3 is a schematic diagram of the interaction between a doctor and a training system in an intelligent medical AI model training method according to an embodiment of the present invention;

[0066] Figure 4 It is a structural diagram of an intelligent medical AI model training system according to one embodiment of the present invention. DETAILED DESCRIPTION

[0067] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation modes.

[0068] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0069] like Figures 2 to 3As shown, according to a preferred embodiment of the present invention, an intelligent medical AI model training method comprises the following steps:

[0070] S1: Select the required functional modules and requirements, and analyze the specific learning tasks.

[0071] In this step, clinicians select the required AI functional modules, such as auxiliary diagnosis, disease prediction, and risk assessment, through an intuitive user interface. Each functional module has specific sub-functions, and doctors can choose according to specific needs. Based on the doctor's selection, the system automatically identifies and parses them into specific AI tasks. For example, the auxiliary diagnosis module can select a classification task to determine whether there is a lesion, and can select a target detection or segmentation task to determine where the lesion is, thereby generating detailed learning tasks.

[0072] The required AI functional modules and requirements include: application scenarios (such as tumor detection, cardiovascular disease prediction, etc.), target output (such as predicting the probability of a certain disease, detecting the location of a tumor, etc.), performance indicators (such as accuracy, recall rate, F1 value, etc.), and data requirements (such as required medical imaging data, case data, etc.).

[0073] S2: According to the learning task, the required data is automatically identified and the selected data set is screened.

[0074] Since this system is an AI system based on medical imaging data, the data required is mainly medical imaging data, which comes from two sources:

[0075] 1. Extract data related to the learning task from the system's preset database or hospital information system (HIS). The main data sources include medical imaging materials (such as patients' X-rays, CT or MRI medical images); these data usually have high clinical value, but may have differences in format, type, etc., and need to be preprocessed and standardized.

[0076] 2. External data includes public data or data from other sources obtained outside the clinical environment. It is usually used to supplement and enrich the learning materials of the model. The main sources are public medical databases (such as NIH imaging datasets, etc.). These resources usually contain a large amount of well-labeled medical data and imaging data, as well as sorted and labeled medical research datasets obtained from other research institutions or academic projects. Such data can also be used to train the model to improve its generalization ability.

[0077] After selecting the appropriate data, not all data is suitable for model training. In order to ensure that the model can learn effectively, the selected data set needs to be screened. The screening operation mainly includes:

[0078] 1. According to the specific learning task, filter out the data directly related to the task. For example, if the task is to diagnose lung cancer, the system will only extract images related to the lungs and will not extract data related to other organs.

[0079] 2. In order to avoid bias in the model, the screened data needs to cover patient data of different ages, genders, races, and health statuses to ensure the model's generalization ability and ability to adapt to different groups.

[0080] 3. The screening of case data should also cover data at different stages of the disease, such as normal cases, mild cases, and severe cases, to ensure that the model can learn the characteristics of various diseases, thereby improving its accuracy and diagnostic range.

[0081] 4. Data format types may differ from source to source, so ensure consistency in data formats and standards. For example, imaging data should conform to the DICOM format, and case data should use unified terminology coding (such as ICD-0001).

[0082] 5. Ensure that the data used is strictly and accurately labeled, such as clearly labeling the lesion area in the medical image, and select data verified by clinical experts for key mission data to ensure that these data are authoritative and reliable in medicine. Key data is the data labeled by the hospital itself, which can ensure the security of the data.

[0083] 6. When screening data, ensure that the data has been desensitized (such as deleting personal information and anonymizing it) and complies with regulatory requirements for data privacy protection.

[0084] S3: Preprocess the data obtained in step S2. The preprocessing operation includes cleaning and converting the original data to make it more suitable for training and analysis of the AI ​​model.

[0085] Data preprocessing mainly includes the following steps:

[0086] S301: Data cleaning: Data cleaning aims to eliminate errors, noise, missing values ​​and other problems in the data to ensure that the model can process high-quality and consistent data, mainly including: Data consistency check: Ensure that the format or unit of the data is consistent and eliminate non-compliant data; Label missing processing: The data used in supervised learning lacks necessary annotations (such as diagnostic results or marks of lesions in images), and labels need to be completed or automatically labeled. When key data lacks labels, experts can be asked to label the data to ensure data quality; Remove duplicate data: The system detects duplicate data by comparing multiple fields, such as patient ID, examination time, diagnostic results, etc. If exactly the same record appears repeatedly in the data set, one is retained and the remaining duplicate records are deleted.

[0087] S302: Statistical data set information: Statistical data set information refers to the overview, summary and analysis of the data in the data set in order to better understand the structure, distribution and characteristics of the data, including the spacing of statistical data, the size of the data, the maximum and minimum values ​​of the grayscale value of the data set, etc. This process is very important for model training because it helps identify potential problems and ensure the quality and integrity of the data. Statistical information includes numerical statistical information of the data, statistics of missing values, outlier analysis, etc.

[0088] S303: Data standardization: The cleaned data is standardized. For image data, the data from different sources needs to be adjusted to the same resolution and then the grayscale value is normalized or z-score standardized to ensure that the model can be processed consistently and adapt to the input requirements of the model.

[0089] After data selection and preprocessing are completed, the following steps are also included:

[0090] S304: Provide the doctor with a data list, showing detailed information such as the data type, source, and quantity to be used for model training for the doctor to adjust.

[0091] Doctors can provide feedback and make adjustments based on their needs: Data volume adjustment: Doctors can increase or decrease the amount of data based on actual needs, such as selecting more imaging data to improve the model's image recognition capabilities; Data type supplement: If the data type provided by the system is incomplete, doctors can manually add other types of data (such as data annotated by the doctor himself).

[0092] S4: Select the model based on the learning task and determine the model and network architecture based on the dataset information.

[0093] The model in the present invention can select the unet structure commonly used for segmentation tasks. The topology of the model is driven by data, and the network structure of the model, including the number of convolutional layers, convolution kernel size, pooling strategy and network depth, will be automatically adjusted according to the shape, size and label distribution of the input data set. Specifically, the adjusted network structure includes: Patch size, which automatically determines the appropriate patch size by analyzing the resolution, size and video memory limit of the input data; the number of pooling layers, which is determined by the minimum size of the data and the target output spatial resolution; the number of feature maps, which dynamically adjusts the number of feature maps of each layer according to the video memory utilization, and the default is a multiple of 32 (for example, 32, 64, 128...).

[0094] Data preprocessing is completed before model training. Selecting a suitable model based on the learning task and determining the model topology based on the data set information are key steps in building an efficient and accurate AI system. They have an important impact on the performance, generalization ability and practical application effect of the model. The following is a detailed process and method for selecting a model based on the learning task and determining the model and network architecture based on the data set information:

[0095] S401: Automatically select a suitable model according to the identified learning tasks (classification task, detection task and segmentation task). For example, classification task can be selected for disease prediction in auxiliary diagnosis, target detection task can be selected for lesion location detection, and segmentation task can be selected for lesion outline drawing.

[0096] S402: Dynamically adjust the network topology of the model according to the data set information. The specific process is as follows:

[0097] S4021: Automatically select the appropriate network structure according to the dimension of the data (2D or 3D), and dynamically adjust the network input size and downsampling multiple according to the size of the input data. In order to maximize the use of video memory resources, the batch size and input network size are automatically adjusted according to the size of the video memory.

[0098] Specifically, the properties of the training dataset are analyzed, the voxel spacing of the image and the size range of the target region (ROI) are counted, and a 2D model is tended to be selected when the data resolution is very high, the target region is small and has no continuity on different slices; a 3D model is selected when the data resolution is low, the target region is large and has 3D continuity.

[0099] S4022: Automatically adjust the size of the convolution kernel according to the resolution of the input data. Specifically, analyze the resolution of the dataset and the size of the target region, and select the appropriate convolution kernel size to balance the receptive field and computational efficiency. For lower resolution data (larger voxels), a larger convolution kernel is usually selected to capture more contextual information; for high resolution data, a smaller convolution kernel is selected to better capture details to extract fine features.

[0100] S4023: Dynamically adjust the number of downsampling (MaxPooling) and upsampling (UpSampling) layers based on the image size and computing resources. The more downsampling layers, the deeper the network, which is suitable for more complex tasks. The computing resources refer to the size of the GPU memory used for computing.

[0101] This dynamic adjustment mechanism is mainly to ensure that the network can effectively utilize hardware resources when processing large-resolution images, while taking into account sufficient feature extraction and perception capabilities.

[0102] First, the resolution of the input image is analyzed, and the number of network layers is determined based on the spatial size of the image. Generally speaking, larger images require more downsampling layers to reduce the size of the feature map, making it easier for the network to aggregate features at a deeper level.

[0103] The number of downsampling layers is dynamically calculated to ensure that the spatial size of the final feature map does not fall below a certain threshold (usually 4×4 or larger). This means that for ultra-high-resolution images, the number of downsampling times will be increased to reduce the feature map size; while for smaller-resolution images, the number of downsampling layers will be reduced to retain more detail information. The downsampling depth is dynamically adjusted according to the size of the current hardware video memory. When the video memory is limited, the depth of the network is limited to avoid exceeding the video memory limit. On the contrary, when the video memory is sufficient, more downsampling layers are allowed to increase the receptive field. This mechanism ensures a balance between network depth and hardware resources and improves the adaptability of the model.

[0104] S4044: Dynamically adjust the depth of the network according to the size of the feature map.

[0105] Specifically, for simple data sets (low image resolution, simple tasks), the network depth (i.e., the number of convolutional layers) is automatically reduced to avoid overfitting and unnecessary computational complexity; for complex high-resolution data sets, the network depth is increased to ensure sufficient feature extraction. At different levels of the network, the appropriate convolution kernel and upsampling kernel size are automatically selected according to the size of the feature map, so that the network can complete downsampling and upsampling operations while ensuring feature validity and hardware efficiency.

[0106] The above dynamic adjustment strategy can adapt to images of different sizes and hardware resources, and provide a network architecture with efficient downsampling and upsampling layers, thereby improving segmentation performance and optimizing resource usage.

[0107] S5: Train the model based on the automatically generated network architecture.

[0108] Specifically, the training process mainly includes:

[0109] S501: Split the data into training set, validation set and test set to ensure that the model can be effectively trained and evaluated. The training set is used in the learning phase of the model, usually accounting for 60-80% of the data; the validation set is used for model tuning to help verify the effect of the model during the training process, usually accounting for 10-20%; the test set is used for the final model evaluation to ensure that the model performs well on unseen data, usually accounting for 10-20%.

[0110] S502: Load the preprocessed data set (training set, validation set and test set), and load the training parameters such as learning rate, batch size, etc. according to the model and network architecture determined in step S4.

[0111] When the optimizer is SGD, the initial learning rate can be set to 0.01, and the learning rate decay is polynomial decay; the training batch has fast, regular and deep models to choose from, and the default is regular.

[0112] S503: Automatically allocate computing resources to ensure that the environment adapts to training requirements, start the training task and record logs, including training progress, loss function value and performance indicators.

[0113] S504: Load the data into the memory in batches for training and verification; perform real-time data enhancement (such as rotating, flipping, and cropping images) during the training process to improve the generalization ability of the model.

[0114] S505: forward propagate the input data through the network to generate a prediction result, calculate the loss between the prediction result and the actual label, calculate the gradient based on the loss and update the network weight.

[0115] S506: Monitor the training loss in real time, use the validation set to calculate indicators after each training cycle or a predetermined interval, monitor the validation loss, accuracy, etc.; use graphical tools to display the changing curves of the training loss and validation loss to help understand the learning progress of the model.

[0116] The training process generally has fast (100 rounds), regular (500 rounds), and deep (1000 rounds) options. Selecting the training mode will set the corresponding number of training rounds, and the training will stop when the corresponding number of rounds is reached.

[0117] S6: After training, the model is evaluated and deployed using the test set in the data. This step is critical to ensure that the model can provide reliable and effective support in actual clinical settings.

[0118] Specifically, the evaluation and deployment process mainly includes:

[0119] S601: Test set evaluation: After the model training is completed, the model performance is evaluated on the divided test set. The main purpose of the evaluation is to confirm that the model not only performs well on the training data, but also has high prediction accuracy on unseen data.

[0120] S602: Clinical Validation: Test the trained model in a simulated clinical environment to ensure that its performance in actual application scenarios meets expectations, or invite medical experts to review the model's prediction results to further confirm whether the model meets clinical needs and standards.

[0121] S603: Deployment: Deploy the model verified by steps S601 and S602 to the application environment of the medical institution, that is, integrate the model into the existing system of the medical institution, such as the electronic medical record system, image analysis system, etc., so that clinicians can use it conveniently.

[0122] S7: Establish a feedback mechanism to collect clinicians’ feedback on the use of the model and make adjustments and optimizations based on the feedback.

[0123] Specifically, the feedback process mainly includes:

[0124] S701: Collect new data generated by the model in actual applications, analyze this data to identify possible problems or improvements in the model, and then regularly evaluate the performance of the model and check whether the model needs to be updated or retrained.

[0125] S702: Based on the collected new data and user feedback, fine-tune the original model so that the model output results are more in line with user needs. The optimized model is automatically updated and deployed to the medical institution environment to ensure that the medical institution uses the latest version of the model.

[0126] S703: Track the latest technological advances and apply new technologies or methods to the system to improve the performance and reliability of the model.

[0127] An intelligent medical AI model training method of the present invention is committed to solving the problems of complexity and difficulty in personalized customization in traditional AI model training. The training process of the AI ​​model is simplified through intelligent processes, and the technical requirements for clinicians and technicians are reduced so that they can quickly and autonomously train the model and reduce manual intervention; the system automatically completes tasks such as data processing, model selection, and parameter adjustment to improve work efficiency. Clinicians select the corresponding module through the system interface, and the system automatically identifies the functions to be implemented and the relevant data and generates detailed learning tasks, setting clear goals for the training model. The system automatically performs data preprocessing, including data cleaning, enhancement, and standardization, prepares high-quality data for the training model, and automatically selects the most suitable model architecture according to the requirements and data set information. The system starts model training and monitors the training indicators in real time during the training process, and automatically adjusts the training parameters (such as learning rate, batch size, etc.) to optimize the model effect. Through intelligent processes and automated operations, the efficiency of model training is greatly improved, manual intervention is reduced, and clinicians can more conveniently train and customize models.

[0128] like Figure 4 As shown, an intelligent medical AI model training system according to a preferred embodiment of the present invention is used to execute an intelligent medical AI model training method according to the preferred embodiment of the present invention mentioned above.

[0129] The intelligent medical AI model training system includes:

[0130] Requirements module 1 is used for users to select required functional modules and requirements and analyze specific learning tasks.

[0131] The data screening module 2 is used to automatically identify the required data and screen the selected data set according to the learning task.

[0132] The data preprocessing module 3 is used to preprocess the data. The preprocessing operation includes cleaning and converting the original data to make it more suitable for the training and analysis of the AI ​​model.

[0133] The model selection module 4 is used to select a model according to the learning task and determine the model and network architecture according to the data set information.

[0134] The model training module 5 is used to build a model based on the automatically generated network architecture and model, and to train the model.

[0135] The model evaluation module 6 is used to evaluate and deploy the model using the test set in the data to ensure that the model can provide reliable and effective support in the actual clinical environment.

[0136] Feedback module 7 is used to collect clinicians' feedback on the use of the model and make adjustments and optimizations based on the feedback.

[0137] Based on the same inventive concept, an embodiment further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the data adaptation method for data interaction as described in the above embodiment are implemented.

[0138] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0139] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0140] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0142] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A method for training an intelligent medical AI model, characterized in that: The following steps are involved: S1: Select the required functional modules and requirements, and analyze the specific learning tasks; S2: Automatically identify the required data and filter the selected data set according to the learning task; S3: preprocessing the data obtained in step S2, the preprocessing operation includes cleaning and converting the original data; S4: Select the model according to the learning task and determine the model and network architecture based on the dataset information; S5: Train the model based on the automatically generated network architecture; S6: Evaluate and deploy the model using the test set in the data; S7: Establish a feedback mechanism to collect clinicians’ feedback on the use of the model and make adjustments and optimizations based on the feedback.

2. The intelligent medical AI model training method according to claim 1, characterized in that: In step S2, the screening operation mainly includes: based on the specific learning task, screening out data directly related to the task; screening out patient data covering different ages, genders, races, and health statuses; screening out data at different stages of the disease to ensure that the model can learn the characteristics of various diseases; among which, the format and standard of the data must be consistent.

3. The intelligent medical AI model training method according to claim 1, characterized in that: In step S3, data preprocessing mainly includes the following steps: S301: Data cleaning, specifically including: data consistency check to ensure that the format or unit of the data is consistent; label missing processing, the data used in supervised learning lacks necessary labels and needs to be supplemented or automatically labeled, and when key data lacks labels, experts are asked to label the data; duplicate data removal, the system detects duplicate data by comparing multiple fields and deletes duplicate records; S302: Statistical data set information, specifically including statistical data spacing, data size, and maximum and minimum values ​​of the grayscale value of the data set; S303: Data standardization: standardize the cleaned data. For image data, adjust the data from different sources to the same resolution and then normalize the grayscale value or z-score.

4. The intelligent medical AI model training method according to claim 1, characterized in that: In step S3, selecting a model and determining the model and network architecture according to the data set information mainly includes the following steps: S401: Automatically selecting a suitable model according to the identified learning task; S402: Dynamically adjust the network topology of the model according to the data set information.

5. The intelligent medical AI model training method according to claim 4, characterized in that: Step S402 specifically includes the following steps: S4021: Automatically select a suitable network structure according to the dimension of the data, and dynamically adjust the network input size and downsampling multiple according to the size of the input data; S4022: Automatically adjust the size of the convolution kernel according to the resolution of the input data; S4023: Dynamically adjust the number of downsampling and upsampling layers according to image size and computing resources; S4044: Dynamically adjust the depth of the network according to the size of the feature map.

6. The intelligent medical AI model training method according to claim 1, characterized in that: In step S5, the training process mainly includes: S501: Split the data into training set, validation set and test set; S502: Load the preprocessed data set, and load the training parameters according to the model and network architecture determined in step S4; S503: Automatically allocate computing resources to ensure that the environment adapts to training requirements, start training tasks and record logs; S504: Load the data into the memory in batches for training and verification; S505: forward propagating the input data through the network to generate a prediction result, calculating the loss between the prediction result and the actual label, calculating the gradient based on the loss and updating the network weight; S506: Monitor the training loss in real time, calculate the indicators using the validation set after each training cycle or a predetermined interval, and display the change curves of the training loss and the validation loss through a graphical tool.

7. The intelligent medical AI model training method according to claim 1, characterized in that: In step S6, the evaluation and deployment process mainly includes: S601: After the model training is completed, the model performance is evaluated on the divided test set; S602: Test the trained model in a simulated clinical environment; S603: Deploy the model verified by steps S601 and S602 to the application environment of the medical institution.

8. The intelligent medical AI model training method according to claim 1, characterized in that: In step S7, the feedback process mainly includes: S701: Collect new data generated by the model in actual applications, and then regularly evaluate the performance of the model and check whether the model needs to be updated or retrained; S702: Adjustments are made to the original model based on the collected new data and user feedback, and the optimized model is automatically updated and deployed in the environment of the medical institution.

9. An intelligent medical AI model training system, used to execute the intelligent medical AI model training method according to any one of claims 1 to 8, characterized in that: include: The requirement module is used for users to select the required functional modules and requirements and analyze the specific learning tasks; The data screening module is used to automatically identify the required data and screen the selected data set according to the learning task; A data preprocessing module is used to preprocess the obtained data, and the preprocessing operation includes cleaning and converting the original data; Model selection module, used to select models according to learning tasks and determine models and network architectures based on dataset information; Model training module, used to build models based on automatically generated network architecture and models, and to train the models; Model evaluation module, used to evaluate and deploy the model using the test set in the data; The feedback module is used to collect clinicians' feedback on the use of the model and make adjustments and optimizations based on the feedback.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the intelligent medical AI model training method according to any one of claims 1-8 are implemented.

Images