Artificial intelligence guided radiotherapy accurate positioning system
By constructing an AI-driven closed-loop system that combines deformation registration and rigid registration, the automation and individualization issues of existing radiotherapy positioning technologies in complex anatomical changes have been solved, enabling rapid and accurate radiotherapy positioning correction and improving the automation and intelligence level of radiotherapy.
Patent Information
- Application Number
- CN202511899226.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-02-24
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing radiotherapy localization techniques suffer from low automation, insufficient clinical decision support, and a lack of continuous learning capabilities when faced with complex anatomical changes, resulting in decreased localization accuracy and operational complexity, making it difficult to meet individualized needs.
We construct a closed-loop system driven by artificial intelligence throughout the entire process. By combining deformation registration and rigid registration, and using deep learning and machine learning algorithms, we achieve dynamic and adaptive positioning correction. We also introduce generative adversarial networks and closed-loop optimization mechanisms to improve positioning accuracy and automation.
It enables rapid and accurate positioning and correction before each treatment, reduces preparation time, lowers the risk of target area under-irradiation and excessive irradiation of normal tissues, provides personalized auxiliary decision support, and improves the automation and intelligence level of radiotherapy.
Smart Images
Figure CN121550601A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical technology, and in particular to an artificial intelligence-guided radiotherapy precision positioning system. Background Technology
[0002] Radiotherapy is one of the main methods of cancer treatment, and its therapeutic effect is highly dependent on the precise localization of the tumor target area before treatment. In the routine radiotherapy procedure, patients need to undergo a scan using imaging equipment such as cone-beam computed tomography (CBCT) before treatment, and the scan results are compared with the simulated localization CT images obtained during the treatment planning stage to correct for deviations in the target area position caused by daily positioning errors, organ movement, or changes in anatomical structure (such as tumor shrinkage or weight changes).
[0003] Currently, clinical practice mainly relies on two types of technologies for image-guided positioning correction: rigid registration-based positioning correction and deformation registration-based adaptive positioning. Rigid registration-based positioning correction assumes that the patient's anatomical shape during treatment only undergoes overall displacement or rotation compared to the planning stage. Image registration is performed using bony landmarks or implantation markers, and translation and rotation parameters are calculated to drive the treatment bed for position correction. However, this method cannot effectively handle non-rigid changes such as soft tissue and organ deformation and tumor volume changes, leading to a significant decrease in positioning accuracy in easily deformable areas such as the head, neck, and pelvis. Deformation registration-based adaptive positioning, on the other hand, has high computational complexity, making it difficult to meet the time requirements for rapid pre-treatment positioning. It also typically requires manual intervention to confirm the registration results, increasing the clinical workload. Furthermore, existing systems are usually static algorithms, lacking the ability to learn from clinical feedback and continuously optimize registration accuracy or decision thresholds based on actual treatment data, making it difficult to adapt to individual patient differences.
[0004] Therefore, although existing radiotherapy positioning technology can achieve basic error correction, it still suffers from defects such as low automation, insufficient clinical decision support, and lack of continuous learning ability when faced with complex anatomical changes, which restricts the further development and clinical application of precision radiotherapy. Therefore, this invention proposes an artificial intelligence-guided radiotherapy precision positioning system to solve the problems existing in the prior art. Summary of the Invention
[0005] To address the aforementioned problems, the present invention aims to propose an artificial intelligence-guided radiotherapy precision positioning system. By constructing a closed-loop system driven by artificial intelligence throughout the entire process, the present invention achieves a fundamental transformation in radiotherapy positioning from the traditional reliance on static planning and human experience to dynamic, adaptive, and intelligent guidance, thereby solving the problems in the prior art.
[0006] To achieve the objectives of this invention, the invention is implemented through the following technical solution: an artificial intelligence-guided radiotherapy precision positioning system, including a data acquisition and processing module, used to acquire the patient's initial planning CT images, initial target area contours, and cone-beam CT images acquired before treatment, and to preprocess the cone-beam CT images;
[0007] The intelligent anatomical change analysis module, connected to the data acquisition and processing module, includes:
[0008] The deformation registration unit is used to calculate the deformation field between the initial planned CT image and the cone-beam CT image using a deep learning deformation registration model based on unsupervised learning.
[0009] The intelligent judgment unit is used to use a classification model to judge the clinical significance of anatomical changes based on the deformation field and output the judgment result, which is divided into significant changes and insignificant changes.
[0010] The positioning guidance instruction generation module is used to generate different positioning guidance instructions based on the judgment results.
[0011] A further improvement lies in the following: the specific method for generating different positioning guidance instructions based on the judgment result is as follows:
[0012] When the judgment result is significant, the initial target area contour is mapped onto the pre-treatment cone-beam CT image based on the deformation field to generate the mapped target area contour. Then, the current positioning error is calculated, and then a positioning guidance instruction containing the current positioning error and the warning of needing doctor intervention is generated.
[0013] When the judgment result is that the change is not significant, the rigid registration algorithm is called to register the pre-treatment cone-beam CT image with the initial planning CT image, calculate the positioning error and generate positioning guidance instructions.
[0014] A further improvement is that the classification model is a gradient boosting tree model or a lightweight convolutional neural network model, and its input features are at least one of the target area average deformation distance, maximum deformation distance and volume change rate extracted from the deformation field.
[0015] The further improvement lies in the fact that the rigid registration algorithm is a registration algorithm based on machine learning feature points.
[0016] A further improvement is that the mapping of the initial target area contour onto the pre-treatment cone-beam CT image is performed using a generative adversarial network model.
[0017] A further improvement is that the data acquisition and processing module includes an image quality assessment submodule, which is used to determine the quality of the acquired pre-treatment cone-beam CT images and perform post-processing based on the determination results.
[0018] A further improvement lies in the following: the specific method of the post-processing is as follows:
[0019] When the image is deemed to be of substandard quality, it is prevented from entering the anatomical change intelligent analysis module, and a prompt message is generated.
[0020] When the image is deemed to be of acceptable quality, it is normally entered into the intelligent anatomical change analysis module.
[0021] Further improvements include a closed-loop optimization module, which records the judgment results and subsequent actions of clinicians, and periodically performs incremental learning optimization on the classification model based on the recorded data.
[0022] The beneficial effects of this invention are as follows: By automatically comparing the differences between the current anatomical structure and the original plan before each treatment and intelligently judging the clinical significance of the changes, this invention achieves a differentiated and precise guidance strategy: When the changes are not significant, the system quickly completes high-precision positioning verification, greatly shortening the treatment preparation time and improving treatment efficiency; when the changes are significant, the system can not only accurately map the target area and calculate the current positioning error, but more importantly, it can proactively issue warnings that require physician intervention, effectively avoiding the risk of missed irradiation of the target area or excessive irradiation of normal tissues due to changes in anatomical structure, providing key decision support for the clinical implementation of adaptive radiotherapy.
[0023] Furthermore, this invention introduces a closed-loop optimization mechanism, enabling it to continuously learn from the actual operational feedback of clinicians and iteratively optimize its internal AI model. This makes it increasingly compatible with the standards and clinical habits of specific medical institutions, ultimately becoming a personalized auxiliary decision-making system that becomes more accurate and intelligent with use. This provides a solid technical foundation for promoting the precision, automation, and intelligence of radiotherapy. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation
[0025] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.
[0026] according to Figure 1As shown, this embodiment proposes an artificial intelligence-guided radiotherapy precision positioning system. The system's data flow specifically includes: first, acquiring the patient's planned CT images and real-time CBCT images; then, optimizing the image quality; next, analyzing anatomical structure changes through deformation registration; then, generating different positioning guidance strategies based on the analysis results; and finally, sending control commands to the treatment bed execution system. Throughout the process, the system records all key data for subsequent model optimization. Specifically, it includes:
[0027] The data acquisition and processing module is used to acquire the patient's initial planning CT images, initial target area contours, and pre-treatment cone-beam CT images, and to preprocess the cone-beam CT images. Further, this module is responsible for acquiring the patient's initial planning CT images and initial target area contours from the hospital information system (such as PACS) and pre-treatment cone-beam CT (CBCT) images from the treatment equipment (such as a linear accelerator). This module preprocesses the CBCT images, including denoising and correction. This module includes an image quality assessment submodule, which uses a pre-trained convolutional neural network (such as ResNet) to assess the quality of the CBCT images. This network, trained on a large amount of labeled clinical image data, can identify common image quality problems, such as motion artifacts, metal artifacts, and excessive noise, and then determine whether they meet the requirements for subsequent analysis. If the quality is unsatisfactory, the image is prevented from flowing into subsequent modules, and an alarm message is generated to alert technicians; if the quality is acceptable, the image is sent to the anatomical change intelligent analysis module.
[0028] For preprocessing, noise reduction is performed first, employing an adaptive nonlocal mean filtering algorithm. This algorithm analyzes the local noise characteristics of the image and dynamically adjusts the filtering parameters to preserve tissue edge details to the maximum extent while eliminating noise. Specifically, the algorithm first performs wavelet decomposition on the image, then filters it at different scales, and finally reconstructs the denoised image using wavelet analysis. Next, geometric correction is performed, including ray hardening correction and scattering correction. Ray hardening correction establishes energy response curves by measuring the attenuation characteristics of phantoms of different thicknesses to compensate for hardening artifacts in the image. Scattering correction uses a Monte Carlo simulation algorithm to estimate the distribution of scattered rays and subtracts the scattered components from the original projection data. Finally, intensity normalization is performed, mapping the grayscale values of the CBCT image to the standard CT value range. This process is achieved by establishing a grayscale mapping table, which is pre-built based on calibration data from water and bone phantoms.
[0029] The intelligent anatomical change analysis module, connected to the data acquisition and processing module, includes:
[0030] The deformation registration unit is used to calculate the deformation field between the initial planned CT image and the cone-beam CT image using a deep learning deformation registration model based on unsupervised learning.
[0031] The intelligent judgment unit is used to use a classification model to judge the clinical significance of anatomical changes based on the deformation field and output the judgment result. The judgment result is divided into significant changes and insignificant changes. The classification model is a gradient boosting tree model or a lightweight convolutional neural network model. Its input features are at least one of the target area average deformation distance, maximum deformation distance and volume change rate extracted from the deformation field.
[0032] Correspondingly, the workflow of the intelligent analysis module for anatomical changes is as follows:
[0033] The deformation registration unit uses the VoxelMorph model to perform deformation registration on the planned CT and CBCT to obtain the deformation field. Then, it extracts the deformation features of the target area from the deformation field, including the average deformation distance, the maximum deformation distance, and the volume change rate. After that, the intelligent judgment unit inputs the extracted features into the gradient boosting tree model (XGBoost) or the lightweight CNN classification model (lightweight convolutional neural network model) and outputs a binary classification result indicating whether the anatomical changes are significant.
[0034] The positioning guidance instruction generation module is used to generate different positioning guidance instructions based on the judgment result. Specifically:
[0035] When the judgment result is significant, the initial target area contour is mapped onto the pre-treatment cone-beam CT image based on the deformation field to generate the mapped target area contour. Then, the current positioning error is calculated, and then a positioning guidance instruction containing the current positioning error and the warning of needing doctor intervention is generated.
[0036] When the judgment result is that the change is not significant, the rigid registration algorithm (a registration algorithm based on machine learning feature points) is called to register the pre-treatment cone-beam CT image with the initial planning CT image, calculate the positioning error and generate positioning guidance instructions.
[0037] Furthermore, the workflow of the location boot instruction generation module is as follows:
[0038] The system receives the judgment results from the anatomical change intelligent analysis module. If the change is deemed significant, the following steps are executed: The initial target region contour is mapped onto the CBCT using a Generative Adversarial Network (GAN) to generate the mapped contour. Then, the current positioning error is calculated based on deformation registration results: First, the optimal rigid transformation components (translation and rotation) are decomposed from the calculated deformation field. These components represent the overall positioning deviation after excluding tissue deformation. Finally, instructions containing positioning error and warnings requiring physician intervention are generated. Specifically, this calculation is based on a hybrid method of feature point matching and surface registration. Initial transformation estimates are first obtained through feature point matching, and then fine-tuned through surface registration. The final output positioning parameters include three translational and three rotational amounts.
[0039] If the change is determined to be insignificant, the following steps are performed: A rigid registration algorithm based on machine learning feature points (e.g., SIFT, SURF, or learned feature points) is used to register the CBCT with the planned CT, calculate the positioning error, and generate positioning guidance commands (which can be directly used for automatic adjustment of the treatment bed). Specifically, the method is based on the principle of maximizing mutual information, using an algorithm to find the spatial transformation parameters that maximize the similarity between the two images. The registration process employs a multi-resolution strategy, starting with coarse registration from low-resolution images and gradually increasing the resolution for fine registration. After registration, treatment bed control commands are directly generated. These commands conform to the communication protocol of the treatment bed control system and include parameters such as movement direction, movement distance, and movement speed.
[0040] The closed-loop optimization module records the judgment results and subsequent actions of clinicians, and periodically performs incremental learning optimization on the classification model based on the recorded data. Furthermore, its workflow is as follows:
[0041] The system records the output results (significant / insignificant changes) of the intelligent judgment unit and the instructions generated by the positioning guidance instruction generation module. It also records the clinicians' feedback to the system, such as: whether the doctor accepts the warning and replans when the system issues a warning; whether the doctor modifies the mapped target area contour generated by the system (e.g., manually adjusting the displacement vector of the contour boundary through the graphical interface, adding or deleting key contour points, or translating / scaling the contour); and whether the doctor confirms or modifies the positioning error calculated by the system. Finally, the recorded data is used periodically (e.g., every weekend) to incrementally train the classification model. Specifically, the doctor's actions are used as labels (e.g., if the doctor modifies the mapped contour, it indicates that the original judgment was significantly correct, but the mapping result needs correction; if the doctor rejects the warning, the original judgment may be incorrect), to retrain the classification model and make it more consistent with clinical practice.
[0042] Specifically, through the implementation of a closed-loop optimization module, this system establishes a complete data logging mechanism, including the following data records: raw data records for each processing step, intermediate processing result records, final decision records, and clinician feedback records. All records are timestamped and have version information for easy subsequent analysis and traceability.
[0043] Accordingly, the system periodically initiates a model update process, with the update cycle adjustable based on clinical workload. The update process includes the following steps: first, new training samples are extracted from the database; then, the existing model is incrementally trained; and finally, the performance of the new model is evaluated through cross-validation. A conservative strategy is adopted for model updates; the new model is only deployed to the production environment when its performance on the validation set is significantly better than the existing model. Simultaneously, the system retains historical model versions and supports a rapid rollback mechanism to ensure stability for clinical use.
[0044] Finally, various performance indicators are monitored in real time, including data processing time, algorithm accuracy, and system stability. Monitoring results are visualized through a dashboard and support historical trend analysis. When performance anomalies are detected, the system issues multi-level alerts and notifies relevant personnel via email, SMS, and other means.
[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its framework and scope of application, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. An artificial intelligence-guided radiotherapy precision positioning system, characterized in that: It includes a data acquisition and processing module, which is used to acquire the patient's initial planning CT images, initial target area contours, and cone-beam CT images acquired before treatment, and to preprocess the cone-beam CT images; The intelligent anatomical change analysis module, connected to the data acquisition and processing module, includes: The deformation registration unit is used to calculate the deformation field between the initial planned CT image and the cone-beam CT image using a deep learning deformation registration model based on unsupervised learning. The intelligent judgment unit is used to use a classification model to judge the clinical significance of anatomical changes based on the deformation field and output the judgment result, which is divided into significant changes and insignificant changes. The positioning guidance instruction generation module is used to generate different positioning guidance instructions based on the judgment results.
2. The AI-guided radiotherapy precision positioning system according to claim 1, characterized in that: The specific steps for generating different positioning guidance instructions based on the judgment result are as follows: When the judgment result is significant, the initial target area contour is mapped onto the pre-treatment cone-beam CT image based on the deformation field to generate the mapped target area contour. Then, the current positioning error is calculated, and then a positioning guidance instruction containing the current positioning error and the warning of needing doctor intervention is generated. When the judgment result is that the change is not significant, the rigid registration algorithm is called to register the pre-treatment cone-beam CT image with the initial planning CT image, calculate the positioning error and generate positioning guidance instructions.
3. The AI-guided radiotherapy precision positioning system according to claim 1, characterized in that: The classification model is a gradient boosting tree model or a lightweight convolutional neural network model, and its input features are at least one of the target area average deformation distance, maximum deformation distance and volume change rate extracted from the deformation field.
4. The AI-guided radiotherapy precision positioning system according to claim 2, characterized in that: The rigid registration algorithm is a registration algorithm based on machine learning feature points.
5. The AI-guided radiotherapy precision positioning system according to claim 2, characterized in that: The mapping of the initial target area contour onto the pre-treatment cone-beam CT image is performed using a generative adversarial network model.
6. The AI-guided radiotherapy precision positioning system according to claim 1, characterized in that: The data acquisition and processing module includes an image quality assessment submodule, which is used to determine the quality of the acquired pre-treatment cone-beam CT images and perform post-processing based on the assessment results.
7. The AI-guided radiotherapy precision positioning system according to claim 6, characterized in that: The specific method of post-processing is as follows: When the image is deemed to be of substandard quality, it is prevented from entering the anatomical change intelligent analysis module, and a prompt message is generated. When the image is deemed to be of acceptable quality, it is normally entered into the intelligent anatomical change analysis module.
8. The AI-guided radiotherapy precision positioning system according to claim 1, characterized in that: It also includes a closed-loop optimization module, which records the judgment results and subsequent actions of clinicians, and performs incremental learning optimization of the classification model on a regular basis based on the recorded data.