Active learning medical image annotation and diagnosis system and method

Abstract

The invention belongs to the technical field of medical image diagnosis, and particularly relates to an active learning medical image annotation and diagnosis system and method, and the system comprises a presentation layer (front end), a data processing layer (rear end), an infrastructure layer (database), and an active learning algorithm module. The presentation layer is responsible for operations of user registration and login, task selection, data visualization, manual and automatic labeling and the like; and the data processing layer is constructed based on a Flask framework, is used for storing to-be-labeled and labeled data, verifying user identities, managing user information and tasks, and is responsible for storage and front-end calling of algorithm results. The infrastructure layer adopts server software such as a MySQL database and the like, so that the durability and the security of data are ensured. Based on an active learning algorithm, the marking cost can be remarkably reduced, and the precision and efficiency of medical image diagnosis are greatly improved; the automation degree of the system and the accuracy and reliability of medical image processing are remarkably improved while the data annotation cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image diagnosis, and particularly relates to an active learning medical image annotation and diagnosis system and method. Background Art

[0002] With the booming development of artificial intelligence technology, vision technology has been widely applied in the field of medical image processing, and the application of artificial intelligence in the medical field is showing an increasing trend. The construction of algorithm models relied on by these artificial intelligence applications cannot do without the support of large-scale labeled data, so the demand for medical images by scientific research personnel continues to climb. However, compared with natural images, the data annotation work of medical images has stronger professionalism and complexity. It not only requires annotators to have medical professional knowledge, but also needs to accurately diagnose various subtle features and lesion information in the images.

[0003] Traditional medical image annotation methods mostly rely on annotators to carefully read and diagnose images manually, and then make annotations after making diagnoses and judgments, and this annotation process is often carried out offline. Especially when encountering complex and difficult images, multiple experts often need to participate in the research and judgment together, and this cooperation mode is extremely inefficient, seriously hindering the progress speed of the annotation work. At the same time, with the rapid expansion of the number of medical images, annotators need to deal with a large number of images, which undoubtedly increases the risk probability of missing labels and mislabels, and the long-term high-intensity work is likely to cause physical and mental fatigue of the annotators, thus resulting in a significant reduction in the annotation accuracy.

[0004] The medical image annotation work has a high professional knowledge threshold and is usually undertaken by medical imaging professionals in hospitals. In view of the current situation of huge workload, low efficiency and scattered dataset resources, the annotation work often requires medical imaging professionals to cooperate with research teams related to intelligent vision. Although this cooperation mode helps to improve the annotation quality, the problems of low efficiency and high cost are still severe. For example, during the annotation process, due to the cumbersome communication and coordination links between medical imaging professionals and research teams, information transmission may deviate or be delayed, further affecting the annotation efficiency.

[0005] Currently, although automatic annotation can be achieved by training an object detection model, there are still many difficulties and inconveniences in the iterative process, and there may also be significant errors in the annotation results. In this context, the active learning algorithm plays a key role. It can train a model based on the existing annotated data and, according to the prediction uncertainty of the model, accurately select the most uncertain samples from the unannotated data. These samples are often the key data points that the model currently most needs to learn and clearly annotate. Then, experts annotate these samples, enabling the model to obtain more valuable annotation information. Next, the model training is restarted based on the newly annotated data, and so on. The active learning algorithm continuously guides the model to focus on the most challenging and learning-potential data, gradually improving the model's annotation ability and accuracy, effectively reducing the dependence on large-scale manual annotation, improving the annotation efficiency, and reducing costs. However, there is still an urgent need for a complete platform to effectively integrate the active learning algorithm, continuously optimize the algorithm performance through a reasonable operation mechanism, and meet the needs of annotators to flexibly adjust the automatic annotation results, so as to achieve efficient and accurate automatic annotation of medical images.

[0006] In view of this, the inventors expect to design an active learning medical image annotation and diagnosis system. Summary of the Invention

[0007] The purpose of the present invention is to provide an active learning medical image annotation and diagnosis system for the problems in the prior art of low annotation efficiency, high cost, large error in annotation results, and strong dependence of deep learning models on annotation data in medical image annotation. This system adopts a front-end and back-end separation development mode, with high maintainability and development efficiency. The front-end framework uses Vue, uses Vuex for state management, and realizes front-end and back-end communication through Axios. The front-end style library uses Element-UI to quickly build the interface. The back-end is built using the Flask framework of Python, and the database is managed using Flask-SQLAlchemy. Through this system and the introduction of the active learning algorithm, the efficiency and accuracy of medical image annotation can be effectively improved, while reducing the cost of manual annotation.

[0008] To achieve the above technical objectives and reach the above technical effects, the present invention is realized through the following technical solutions:

[0009] The present invention provides an active learning medical image annotation and diagnosis system, and the system framework design is mainly divided into the following parts:

[0010] Front-end part: It includes user login, registration, data visualization, manual annotation, and automatic annotation, providing an intuitive interface for users to operate. The front-end is built using the Vue.js framework, with Vuex for state management and the Axios library for communication with the back-end. The interface style library uses Element-UI to ensure the fluency and efficiency of the front-end user experience.

[0011] Back-end part: The back-end is built based on the Flask framework and is responsible for handling user requests, including user information verification, addition, deletion, modification, and query, collection and storage of annotation results, as well as calling and running of algorithm models. The back-end uses Flask-SQLAlchemy to interact with the database to ensure data persistence and security. In addition, the back-end is also responsible for managing the training and inference of algorithm models, providing API interfaces for the front-end to call, and feeding back model results to the front-end.

[0012] Infrastructure layer: The back-end uses a relational database (such as MySQL) to save and manage data, ensuring data persistence and security. The database is responsible for storing user information, annotation data, key model result data, and supports data backup and recovery to ensure the stability and reliability of the system.

[0013] Active learning algorithm module: According to the annotation data set passed from the front-end, the data is divided into a labeled pool and an unlabeled pool, and the most valuable samples for the model are selected from the unlabeled pool through active learning strategies (such as uncertainty sampling). The selected samples are transmitted to the front-end through Axios, and the annotation is completed by users or experts, and the annotation results are stored in the back-end database.

[0014] Furthermore, in the above-mentioned active learning medical image annotation and diagnosis system, the main modules of the system front-end include:

[0015] User login module: Depending on the user identity, different users will see different front-end interfaces.

[0016] Medical image manual annotation module: Users can upload medical image data to be annotated and use annotation tools to accurately annotate it to generate a data set for model training.

[0017] Medical image automatic annotation module: The system uses a trained model to automatically annotate medical images and displays the results on the front-end interface. Users can verify the results. The correct results are directly saved to the training set, and the incorrect results are saved after being adjusted by the correction module. The model training data set is continuously expanded during the use of the system.

[0018] Annotation result correction module: Specifically used to correct the incorrect results generated by automatic annotation and continuously optimize the performance of the model through a closed-loop feedback mechanism.

[0019] Adaptive Model Selection Module: The system supports selecting appropriate deep learning models according to medical types to meet different medical image processing requirements.

[0020] Furthermore, in the above-mentioned active learning medical image annotation and diagnosis system, the medical image manual annotation module and the medical image automatic annotation module ensure the consistency and integrity of the generated dataset through interactive switching; the system can flexibly switch between manual annotation and automatic annotation to ensure that the advantages of both annotation methods are integrated in the generated dataset; manual annotation ensures high precision and pertinence, while automatic annotation improves efficiency and speed, enabling the system to make full use of the high-quality dataset generated by the automatic annotation function and provide reliable data support for subsequent model training and optimization.

[0021] Furthermore, in the above-mentioned active learning medical image annotation and diagnosis system, after the active learning strategy selects samples, the selected samples are sent back to the annotation interface for users to further annotate:

[0022]

[0023] Among them, is the uncertainty value of sample x, used to measure the uncertainty of sample prediction; P M (y i |x) is the predicted value of the model for sample x; C is the total number of categories.

[0024] The present invention also provides an active learning medical image annotation and diagnosis method, which is implemented based on the above-mentioned active learning medical image annotation and diagnosis system, and includes the following steps:

[0025] S1. Data Upload and Initial Annotation

[0026] After the user logs in to the system, upload medical image data and perform manual annotation through tools such as rectangular frames and polygons to generate the initial labeled dataset required for the active learning model; after completing the annotation, set parameters in the model management interface and start the initial training task;

[0027] S2. Model Training and Sample Selection

[0028] The backend uses the labeled dataset to train the model, and according to the prediction uncertainty of the trained model for unlabeled data, adopts an active learning strategy to select the most informative unlabeled samples and return the samples to the user for annotation through the front-end interface;

[0029] S3. Loop Iterative Optimization

[0030] After the user completes the sample annotation, the new annotation data is added to the labeled dataset, and the updated dataset is used to retrain the model. By repeatedly executing the closed-loop process of sample selection, annotation, and model training, the accuracy of the model is gradually improved until the annotation budget is exhausted or the model meets the expected accuracy requirements, and the best weights are saved for each training.

[0031] Further, in step S1, after entering the account and password on the login interface, the system backend will check the existence of the user account and determine whether the user role is an administrative user or a general user. For administrative users, the front end will display all operable modules, and administrative users can perform management and operations on the entire system. For general users, the front end only presents the automatic annotation interface, and general users can select different model algorithms to automatically annotate images, and the system will automatically annotate the pictures.

[0032] Further, when an administrative user uses it for the first time, they need to upload a local dataset and perform manual annotation to generate an initial dataset for model training. The uploaded medical images need to be sent to the backend through the front end and displayed before the data is returned to the front end to prevent data loss caused by browser refresh. Administrative users can perform annotations using manual annotation tools, and the annotation results are sent to the backend Flask framework through the Axios component, sorted and saved as a model training set, and the model is trained using the just-mentioned data in the algorithm model module.

[0033] Further, when annotators and general users use the automatic annotation interface, the weights with the highest accuracy will be loaded to automatically generate annotation results, and the annotation results will be automatically verified. When the verification passes, that is, the annotation results are correct, the automatically annotated medical images and the corresponding annotation information can be loaded into the training dataset.

[0034] Further, in step S2, the method for the active learning strategy to select samples is: the system evaluates the uncertainty of the through the active learning strategy, and uses the entropy sampling method to select the most valuable samples from the unlabeled dataset.

[0035] Further, in step S3, the stop condition for cyclic iterative optimization is: by setting the number of iterations of active learning, the upper limit of annotation resources, or when the model accuracy reaches the preset requirements, the training process ends.

[0036] The beneficial effects of the present invention are:

[0037] 1. The advantages of the system of the present invention in medical image annotation and diagnosis work include:

[0038] Lightweight collaboration: Supports multiple people to participate in annotation, reducing the dependence on medical imaging professionals.

[0039] Closed-loop feedback optimization: Through the cyclic mechanism of manual annotation, automatic annotation, and result correction, continuously improve performance and annotation accuracy.

[0040] Resource conservation: Significantly reduce manpower requirements and annotation costs, improve efficiency while ensuring data quality.

[0041] 2. The medical image annotation and diagnosis system of the present invention is not only applicable to the initial annotation task, but also supports model reuse and automatic annotation, and is suitable for long-term and multi-scenario medical image processing needs.

[0042] Of course, it is not necessary for any product implementing the present invention to achieve all of the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0044] Figure 1 It is a framework diagram of the active learning medical image annotation and diagnosis system in the embodiment of the present invention;

[0045] Figure 2 It is an evolutionary framework diagram of the front end of the medical image annotation and diagnosis system provided in the embodiment of the present invention;

[0046] Figure 3 It is a flowchart of the active learning algorithm in the embodiment of the present invention;

[0047] Figure 4 It is a flowchart of the manual annotation function of the medical image automatic diagnosis system provided in the embodiment of the present invention;

[0048] Figure 5 It is a flowchart of the medical image automatic diagnosis provided in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0050] This embodiment provides an active learning medical image annotation and diagnosis system. By combining active learning and deep learning technologies, it can significantly reduce the time and labor costs of medical image annotation, while improving the accuracy and consistency of annotation results, providing important technical support for research and applications in the field of medical imaging.

[0051] In this embodiment, by constructing an annotation system based on active learning, it can not only reduce the model's dependence on manually annotated data, but also improve the training efficiency of the deep learning model while ensuring the annotation quality. Through the front-end and back-end separation design, combined with an intuitive and user-friendly interface and efficient back-end services, the system provides a comprehensive solution to optimize the medical image annotation and diagnosis process.

[0052] The system architecture of this embodiment includes the following three parts

[0053] Front-end system: Adopting the Vue.js framework, it has functional modules such as user login, registration, data visualization, manual annotation, and automatic annotation. Users perform annotation operations on medical images through an intuitive and user-friendly interface. The annotation tool supports various forms (such as rectangular boxes, polygons, etc.) and communicates with the back-end through Axios to ensure the efficiency and security of data transmission.

[0054] Back-end system: Based on the Flask framework, it is responsible for user information management, collection and storage of annotation data, and invocation and training of deep learning models. The back-end interacts with a relational database (such as MySQL) through Flask-SQLAlchemy to ensure data persistence and security, and at the same time supports front-end requests for model training and inference through API interfaces.

[0055] Infrastructure layer: Uses a relational database to store user information, annotation data, and model results, supports regular backup and recovery to ensure data integrity and security.

[0056] The working process of the system includes the following steps:

[0057] Upload the dataset to be annotated: Users upload the dataset X to be annotated, and the system divides it into a labeled dataset and an unlabeled dataset according to user settings. Users manually annotate the allocated labeled data through the annotation interface, and the annotation results and corresponding labels are automatically saved in a specified location at the back-end.

[0058] Initial model training: Use the annotated data as the initial training dataset for the model and conduct preliminary training.

[0059] The active learning strategy selects samples: The system evaluates the uncertainty of through the active learning strategy, and uses methods such as entropy sampling to select the most valuable samples from the unlabeled dataset. The selected samples are sent back to the annotation interface for users to further annotate.

[0060]

[0061] Among them, is the uncertain fixed value of sample x, which is used to measure the uncertainty of sample prediction; P M (y i |x) is the predicted value of the model for sample x; C is the total number of categories.

[0062] Iterative optimization and stopping conditions: By setting the number of iterations of active learning, the upper limit of annotation resources, or when the model accuracy reaches the preset requirements, the training process is ended.

[0063] Automatic annotation and verification: Use the trained model to automatically annotate the remaining unlabeled data and save the results to the corresponding storage location. The user checks the annotation results through the verification module. If errors are found, they can be modified through the correction interface and then saved to the training set.

[0064] Model performance optimization: The corrected data is added to the training set for subsequent iterative optimization to continuously improve the performance of the model.

[0065] The active learning in this embodiment combined with user interactive annotation significantly improves the efficiency and automation level of medical image annotation.

[0066] Specific design of the system function module:

[0067] Medical image manual annotation module: The user can upload medical image data to be annotated and use annotation tools to accurately annotate it to generate a data set for model training.

[0068] Medical image automatic annotation module: The system uses the trained model to automatically annotate medical images and displays the results on the front-end interface. The user can verify the results. The correct results are directly saved to the training set, and the incorrect results are adjusted through the correction module and then saved. The model training data set can be continuously expanded during the use of the system.

[0069] Annotation result correction module: Specifically used to correct the error results generated by automatic annotation and continuously optimize the performance of the model through a closed-loop feedback mechanism.

[0070] Adaptive model selection module: The system supports selecting appropriate deep learning models according to medical types to meet different medical image processing requirements.

[0071] In this embodiment, the process of active learning and iterative optimization of the deep learning model:

[0072] Annotation and training: The annotator uses the manual annotation tool to generate a training data set and starts the initial model training.

[0073] Automatic annotation and verification: The system automatically annotates unlabeled data using a trained model. The annotator verifies and corrects the results, and the corrected data is added to the training set.

[0074] Iterative training: Continuously select high-value samples for annotation through an active learning strategy, and cycle through training to optimize the model until the preset accuracy or the upper limit of annotation resources is reached.

[0075] The system supports teamwork and provides a shared backend data storage location. The annotation and training work of different users jointly improve the scale and quality of the dataset.

[0076] To meet the diverse needs of medical image diagnosis, the system allows users to select a pending deep learning model to correspond to different types of medical image annotation and processing tasks.

[0077] As Figure 1 shown, this embodiment also provides an active learning medical image annotation and diagnosis method, and the specific steps are as follows:

[0078] User login and upload of medical image data: The annotator first logs in through the system login module, enters the annotation module to upload the medical image data to be annotated, and precisely annotates the image through annotation tools (such as rectangular boxes, polygons, etc.) to generate the basic training dataset required by the algorithm model.

[0079] Model training and feedback: When the training dataset is ready, the annotator can start model training on the model management interface. The deep learning framework at the system backend will perform model training based on the training dataset. After training is completed, the training results of the model are fed back to the front-end interface through the data processing layer for the annotator to view and evaluate the model performance. Accuracy and the ROC curve are the main criteria for measuring the quality of the model.

[0080]

[0081] Among them, the TP (True Positive) sample prediction value matches the true value and both are positive, that is, true positive;

[0082] FP (False Positive): The sample prediction value is positive while the true value is negative, that is, false positive;

[0083] FN (FalseNegative): The sample prediction value is negative while the true value is positive, that is, false negative;

[0084] TN (True Negative): The sample prediction value matches the true value and both are negative, that is, true negative;

[0085] The FPR is the abscissa of the ROC curve, and the TPR is the ordinate of the ROC curve.

[0086] Automatic annotation and verification: Use the trained model to automatically annotate new medical image data. The system will display the automatic annotation results on the front-end interface for the annotators to check their accuracy. If the annotation results are correct, the user can directly save the results to the back-end training set; if there are errors in the annotation results, the user can use the correction tool on the front-end interface to adjust the annotation results and save the modified data to the training set.

[0087] Continuous iteration and optimization: After completing the initial annotation, the annotator can return to the algorithm model management interface and start the model training again. The training data for this time will include manually annotated data and automatically annotated data that has been checked and adjusted, thereby further improving the performance and annotation accuracy of the model.

[0088] During the initial use of this embodiment, the user needs to upload relevant medical image data and complete manual annotation to generate an initial training data set. In subsequent use, the user can directly use the automatic annotation function to skip the steps of data set upload and manual annotation, thereby greatly simplifying the operation process and improving work efficiency.

[0089] As Figure 2 shown, the medical image annotation and diagnosis system of this embodiment adopts a modular design and includes four core modules: a manual annotation module, a model training module, an automatic annotation module, and an annotation correction module. Each module realizes data interaction through a standardized interface to ensure the stability and efficiency of the system.

[0090] As Figure 3 shown, this embodiment provides a medical image annotation and diagnosis system based on an active learning strategy, and its overall process framework is divided into the following steps: division of the labeled data set (L) and the unlabeled data set (U), model training and optimization, sample selection and expert annotation, submission and update of annotation results, and cyclic iteration and model optimization.

[0091] As Figure 4 shown, the specific process of the manual annotation function includes the following steps:

[0092] After the user starts the system, upload the medical image file to be annotated through the front-end interface;

[0093] The user uses the annotation tool provided by the front-end to perform annotation operations on the image and assign corresponding labels to the annotation objects;

[0094] After the annotation is completed, the system will automatically save the annotated image data to the specified training data set folder at the back end;

[0095] The user switches to the next image to be annotated, and the annotation process continues until all images are annotated. After completion, the back-end training dataset folder will contain the complete annotated dataset for subsequent model training.

[0096] As Figure 5 shown, the automatic annotation interface shows the specific operation process of the system's automatic annotation function:

[0097] After the user starts the system, upload the medical image file to be annotated;

[0098] According to the file type and the characteristics of the medical image, the user selects a suitable deep learning model;

[0099] The back-end of the system automatically loads the weight parameters of the selected model and automatically annotates the uploaded image data;

[0100] The annotation results are fed back to the front-end interface through the data processing layer for the user to view and verify. For the correct annotation results, the user can directly save them to the back-end training dataset; for the annotation results with errors, the user can adjust them through the correction tool on the front-end interface and save the corrected results.

[0101] Through the implementation of the above method in this embodiment, not only is the manual annotation workload significantly reduced, and the efficiency of medical image annotation is improved, but also through continuous model iterative training and active learning mechanism, the accuracy of the annotation results is continuously improved. In addition, the system supports multi-user collaboration, and users can share training data and model results, thus greatly improving the scale and quality of the annotated dataset.

[0102] Through the front-back end separation architecture and modular design, this embodiment realizes high efficiency, intelligence and collaboration in medical image annotation and diagnosis, and can significantly improve the efficiency and accuracy of medical image processing.

[0103] The above-disclosed preferred embodiments of the present invention are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation manners. Obviously, according to the content of this specification, many modifications and variations can be made. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the present invention, so that those skilled in the art in the relevant technical field can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. An active learning medical image annotation and diagnosis system, characterized in that: The system framework design is mainly divided into the following parts: Front-end part: including user login, registration, data visualization, manual annotation and automatic annotation, providing an intuitive interface for user operation; the front-end is built with Vue.js framework, using Vuex for state management, using Axios library for communication with the back-end, and the interface style library uses Element-UI; Backend: The backend is built on the Flask framework and is responsible for processing user requests, including verification of user information, addition, deletion, modification, and query, collection and storage of annotation results, and calling and running of algorithm models. The backend uses Flask-SQLAlchemy to interact with the database to ensure data persistence and security. In addition, the backend is also responsible for managing the training and reasoning of the algorithm model, providing an API interface for the front end to call, and feeding back the model results to the front end. Infrastructure layer: The backend uses a relational database to save and manage data to ensure data persistence and security. The database is responsible for storing user information, annotation data, key data of model results, and supports data backup and recovery. Active learning algorithm module: According to the labeled data set delivered by the front end, the data is divided into a labeled pool and an unlabeled pool, and the most valuable samples for the model are selected from the unlabeled pool through an active learning strategy; the selected samples are delivered to the front end through Axios, labeled by users or experts, and the labeled results are stored in the back-end database.

2. The active learning medical image annotation and diagnosis system according to claim 1, characterized in that: The main modules of the system front end include: User login module: Different users will see different front-end interfaces according to their identities; Medical image manual annotation module: users can upload medical image data to be annotated and use annotation tools to accurately annotate them to generate a data set for model training; Medical image automatic annotation module: The system uses the trained model to automatically annotate medical images and displays the results on the front-end interface; users can verify the results, and correct results are directly saved to the training set, while incorrect results are adjusted and saved through the correction module; the model training data set is continuously expanded during the use of the system; Annotation result correction module: dedicated to correcting erroneous results generated by automatic annotation, and continuously optimizing the performance of the model through a closed-loop feedback mechanism; Adaptive model selection module: The system supports selecting appropriate deep learning models according to medical types to meet different medical image processing requirements.

3. The active learning medical image annotation and diagnosis system according to claim 1, characterized in that: The two modules, the medical image manual annotation module and the medical image automatic annotation module, ensure the consistency and completeness of the generated data set through interactive switching; the system can flexibly switch between manual annotation and automatic annotation to ensure that the advantages of both annotation methods are integrated into the generated data set; manual annotation ensures high accuracy and pertinence, while automatic annotation improves efficiency and speed, enabling the system to make full use of the high-quality data set generated by the automatic annotation function, providing reliable data support for subsequent model training and optimization.

4. The active learning medical image annotation and diagnosis system according to claim 1, characterized in that: After the active learning strategy selects samples, the selected samples are sent back to the annotation interface for users to further annotate: in, is the uncertainty value of sample x, which is used to measure the uncertainty of sample prediction; P M (y i |x) is the model's predicted value for sample x; C is the total number of categories.

5. An active learning medical image annotation and diagnosis method, implemented based on the active learning medical image annotation and diagnosis system according to any one of claims 1 to 4, characterized in that: The steps include: S1. Data upload and initial annotation After logging into the system, users upload medical image data and use the tool to manually annotate to generate the initial labeled data set required for the active learning model. After the annotation is completed, the parameters are set in the model management interface and the initial training task is started. S2. Model training and sample selection The backend uses the labeled data set to train the model. Based on the prediction uncertainty of the trained model for the unlabeled data, an active learning strategy is used to select the most informative unlabeled samples, and the samples are returned through the front-end interface for users to label. S3, loop iteration optimization After the user completes sample labeling, the newly labeled data is added to the labeled dataset, and the updated dataset is used to retrain the model. By repeatedly executing the closed-loop process of sample selection, labeling, and model training, the accuracy of the model is gradually improved until the labeling budget is exhausted or the model reaches the expected accuracy requirement, and the best weights are saved for each training.

6. The active learning medical image annotation and diagnosis method according to claim 5, characterized in that: In step S1, after entering the account and password on the login interface, the system backend will check the existence of the user account and determine whether the user role is an administrative user or a general user; For management users, the front end will display all operable modules, and management users can perform management and operations of the entire system; For general users, the front end only presents an automatic annotation interface. General users can select different model algorithms to automatically annotate images, and the system will automatically annotate the images.

7. The active learning medical image annotation and diagnosis method according to claim 6, characterized in that: When using it for the first time, the management user needs to upload the local dataset and manually annotate it to generate the initial dataset for model training; the uploaded medical images need to be sent to the backend through the frontend and displayed before the data is returned to the frontend to prevent data loss caused by browser refresh; the management user can annotate images using the manual annotation tool, and the annotation results are sent to the backend Flask framework through the Axios component, sorted and saved as a model training set, and the model is trained using the data just now in the algorithm model module.

8. The active learning medical image annotation and diagnosis method according to claim 7, characterized in that: When using the automatic annotation interface, annotators and general users will load the weight with the highest accuracy to automatically generate annotation results, and will automatically verify the annotation results. When the verification passes, the annotation results are correct, and the automatically annotated medical images and corresponding annotation information can be loaded into the training data set.

9. The active learning medical image annotation and diagnosis method according to claim 7, characterized in that: In step S2, the active learning strategy selects samples in the following way: the system evaluates the uncertainty of through the active learning strategy and selects the samples with the highest value from the unlabeled data set using the entropy sampling method.

10. The active learning medical image annotation and diagnosis method according to claim 7, characterized in that: In step S3, the stopping condition of the loop iterative optimization is: ending the training process by setting the number of iterations of active learning, the upper limit of the annotation resources, or when the model accuracy reaches the preset requirements.

Images