Self-learning model system for medical image automatic segmentation
Through the self-learning model system, the medical image data is verified and matched and model training is carried out, which solves the problems of long training time and accuracy limitations of existing systems, and realizes fast and accurate training and deployment of medical image segmentation models, which is suitable for non- Customized needs of professionals and different users.
Patent Information
- Application Number
- CN202411207210.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-12-31
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing medical image segmentation self-learning system has problems with long training time and accuracy limitations. It is difficult for non-professional personnel to train models, and different users have different accuracy standards, making it difficult to customize personalized prediction models.
Provide a self-learning model system, which verifies and filters key parameters by obtaining the marked data set uploaded by the user, matches the structural parameters and data samples specified by the user, uses the model network framework for training, and outputs the training log, and finally packages the encrypted model. Used for image segmentation prediction.
Reduces the training complexity of medical imaging segmentation models, allowing non-professional personnel to train models, realizes a fast closed loop from data upload to model deployment, solves users' needs for customized models, and improves the accuracy of models. and applicability.
Smart Images
Figure CN119229111A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a self-learning model system for automatic segmentation of medical images. Background Art
[0002] Fast and accurate image segmentation results are an important part of radiotherapy and medical imaging research. Medical image segmentation plays a connecting role in radiotherapy. In the past, manual labeling required high skills and energy from doctors. The increasing demand for radiotherapy has placed greater demands on the number of doctors and technology.
[0003] In recent years, the use of artificial intelligence to assist doctors in image segmentation technology has been promoted. Different hospitals and different users have different standards for the accuracy of image segmentation results. In order to solve the problem of different users having different expectations for the accuracy of artificial intelligence-assisted medical image segmentation results, it is necessary to provide a set of image segmentation model training based on user-provided data sources.
[0004] The entire process of model training based on big data is highly specialized, making it difficult for non-professionals to access the model training process. The self-learning system streamlines the model training process and extracts specific steps based on the unique characteristics of medical image segmentation model training, integrating the complex training process into the self-learning system. Users only need to provide the corresponding training data, and the system will train the medical image segmentation model based on the dataset. Users can provide different data for training and select the model that meets their requirements for medical image segmentation.
[0005] Although self-learning systems for medical image segmentation can provide a simplified model training process, they suffer from the following major drawbacks:
[0006] Long training time: Since model training uses many parameters, many iterations, and high complexity, in order to increase the applicability of the system, it is impossible to use more professional hardware resources for training, and the model training time will be long.
[0007] Accuracy limitations: To ensure the accuracy of the model, users must provide a certain amount of data and standardized data annotations.
[0008] To this end, the present invention provides a self-learning model system for automatic segmentation of medical images. Summary of the Invention
[0009] In response to the above-mentioned deficiencies in the prior art, the present invention provides a self-learning model system for automatic segmentation of medical images, which solves the problem that different users in artificial intelligence-assisted image segmentation technology have different accuracy definition standards and hope to customize personalized prediction models.
[0010] The present invention provides the following technical solution: a self-learning model system for automatic segmentation of medical images, comprising the following steps:
[0011] S1. Obtain several labeled data sets uploaded by users and perform key parameter verification and screening on the data;
[0012] S2. For the filtered data samples, matching the user-specified structural parameters with the data samples;
[0013] S3. In the data sample, the system will match the corresponding ROI contour data from the structure set data according to the matching rules of ROI names in the training set data based on all the ROI names to be trained. The matched contour data will be used as the parameter input of the model training and enter the next process;
[0014] S4. After confirming the training parameters and model training data, the system calls the corresponding model network framework according to the user-defined model training parameters and starts model training in the background. The training process is recorded and the training log is output;
[0015] S5. After the model training is completed, the model is packaged and encrypted. The packaged model can be directly used for image segmentation prediction.
[0016] Preferably, the key parameter verification and screening of the uploaded data in step S1 are specifically as follows: medical imaging data and corresponding manually marked data are uploaded through the system interface, the system checks whether the format, resolution, layer spacing, etc. of the data meet the training requirements, and excludes unclear, damaged or incorrectly marked data based on preset standards or machine learning algorithms. According to the parameters input by the user, the system crops the data of each case to adapt to the input size requirements of the model or emphasize specific areas.
[0017] Preferably, in step S2, the user specifies the name of the structure that needs to be segmented, and the system performs matching based on the structures and matching rules existing in the data set.
[0018] Preferably, data processing and contour extraction are performed on the data matched in step S2, and the data processing formula is: k is the height of the peak, m is the center value of the peak position, and б is the standard deviation, which determines the width of the curve.
[0019] Preferably, the system in S3 stores a ROI name database or dictionary to determine the naming and classification of different structures. Through matching rules, the system finds the contour data corresponding to the user-specified ROI name from the existing structure set. These matched contour data will be used as feature input for model training.
[0020] Preferably, the model training is performed on the data processed in S3, and the formula for pixel classification of the data during the training process is: e is the base of natural logarithms, zi is the i-th element of the input vector z, It is the sum of the exponential function values of all input elements, which is used to normalize each output yi.
[0021] Compared with the prior art, the present invention has the following beneficial effects:
[0022] This self-learning model system for automatic medical image segmentation can reduce the training complexity of medical image segmentation models, allowing non-professionals to train medical image segmentation models. Users do not need to pay attention to the complex internal algorithm implementation process, solving the key needs of users for customized models in the field of medical image segmentation. The system can be integrated with the doctor's workstation to quickly deploy the model training results to the doctor's workstation for use, completing the process from uploading training data to using the model for prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 This is a business flow chart of the system of the present invention. DETAILED DESCRIPTION
[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings of the embodiments of the present disclosure. In order to keep the following description of the embodiments of the present disclosure clear and concise, the present disclosure omits detailed descriptions of known functions and known components to avoid unnecessary confusion of the concepts of the present invention.
[0025] See also Figure 1 , a self-learning model system for automatic segmentation of medical images, comprising the following steps:
[0026] S1. Obtain several labeled data sets uploaded by users, verify and screen the key parameters of the data, upload medical imaging data and corresponding manually labeled data through the system interface, and the system checks whether the format, resolution, and inter-layer spacing of the data meet the training requirements. Based on preset standards or machine learning algorithms, the system excludes unclear, damaged, or incorrectly labeled data. Based on the parameters entered by the user, the system crops the data of each case to adapt to the input size requirements of the model or emphasize specific areas.
[0027] S2. For the filtered data samples, the user-specified structural parameters are matched with the data samples. The user specifies the name of the structure that needs to be segmented, and the system matches it according to the structure and matching rules existing in the data set.
[0028] The matched data is processed and contour extracted. The data processing formula is: k is the height of the peak, m is the center value of the peak position, and б is the standard deviation, which determines the width of the curve.
[0029] S3. In the data sample, the system will match the corresponding ROI contour data from the structure set data according to the matching rules of ROI names in the training set data based on all the ROI names to be trained. The matched contour data will be used as the parameter input of the model training and enter the next process;
[0030] The system stores a ROI name database or dictionary to determine the naming and classification of different structures. Through matching rules, the system finds the contour data corresponding to the user-specified ROI name from the existing structure set. These matched contour data will be used as feature input for model training.
[0031] The processed data is used for model training. The formula for pixel classification of data during training is: e is the base of natural logarithms, zi is the i-th element of the input vector z, It is the sum of the exponential function values of all input elements, which is used to normalize each output yi.
[0032] S4. After confirming the training parameters and model training data, the system calls the corresponding model network framework according to the user-defined model training parameters, and starts model training in the background. The training process is recorded and the training log is output.
[0033] S5. After the model training is completed, the model is packaged and encrypted. The packaged model can be directly used for image segmentation prediction.
[0034] Users upload patient image data containing annotations of the structures to be trained as the model training data source, where the number of model data sources is at least 10. To ensure the effectiveness of model training, users need to upload more than 10 cases of accurately annotated data.
[0035] This step is automatically handled by the system. The system checks and processes the data source to ensure that all data can be used for model training, and divides the data source into a training set and a validation set. The training set is used for model training, and the validation set is used to verify the trained model after training to check the accuracy of the model prediction.
[0036] After the data check is passed, the user can start model training by clicking the start button in the system. The system will automatically train the model through the internal algorithm and output the training log and the training-related Loss function value. The log and Loss results will be refreshed and displayed in the system for users to view.
[0037] After the model training is completed, the user can publish the specified model to the doctor's workstation. The designated doctor's workstation manufacturer must integrate medical image segmentation prediction through a standard interface.
[0038] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the scope of the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the spirit and scope of protection of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present invention.
Claims
1. A self-learning model system for automatic segmentation of medical images, characterized in that: The following steps are involved: S1. Obtain several labeled data sets uploaded by users, and verify and screen the key parameters of the data; S2. For the filtered data samples, matching the structural parameters specified by the user with the data samples; S3. In the data sample, the system will match the corresponding ROI contour data from the structure set data according to the matching rules of the ROI names in the training set data according to all the ROI names to be trained. Therefore, the matched contour data will be used as the parameter input of the model training and enter the next process; S4. After confirming the training parameters and model training data, the system calls the corresponding model network framework according to the model training parameters defined by the user, and starts model training in the background. The training process is recorded and the training log is output; S5. After the model training is completed, the model is packaged and encrypted. The packaged model can be directly used for image segmentation prediction.
2. A self-learning model system for automatic segmentation of medical images according to claim 1, characterized in that: In step S1, the key parameter verification and screening of the uploaded data are specifically processed as follows: the medical imaging data and the corresponding manually labeled data are uploaded through the system interface, and the system checks whether the format, resolution, layer spacing, etc. of the data meet the training requirements. Based on preset standards or machine learning algorithms, unclear, damaged or incorrectly labeled data are excluded, and according to the parameters entered by the user, the system crops the data of each case to adapt to the input size requirements of the model or emphasize specific areas.
3. The self-learning model system for automatic segmentation of medical images according to claim 1, characterized in that: In step S2, the user specifies the name of the structure that needs to be segmented, and the system matches it according to the structure and matching rules existing in the data set.
4. The self-learning model system for automatic segmentation of medical images according to claim 1, characterized in that: The data matched in step S2 are processed and contour extracted. The data processing formula is: k is the height of the peak, m is the center value of the peak position, and б is the standard deviation, which determines the width of the curve.
5. The self-learning model system for automatic segmentation of medical images according to claim 1, characterized in that: The ROI name database or dictionary is stored in the S3 system to determine the naming and classification of different structures. Through matching rules, the system finds the contour data corresponding to the user-specified ROI name from the existing structure set. These matched contour data will be used as feature input for model training.
6. The self-learning model system for automatic segmentation of medical images according to claim 1, characterized in that: The model is trained on the processed data in S3. The formula for pixel classification of data during training is: e is the base of natural logarithms, zi is the i-th element of the input vector z, It is the sum of the exponential function values of all input elements, which is used to normalize each output yi.