Multi-modal image radiotherapy guiding method based on AI driving

Abstract

The invention relates to the technical field of radiotherapy guidance, in particular to a multi-modal image radiotherapy guidance method based on AI driving, which comprises the following steps: forming a radiotherapy area AI prediction model; collecting a real CT image; converting the real CT image into a simulated MRI image; generating a simulated dimension reduction CT image; determining the position of an artifact point; determining a scanning abnormal state of the real CT image according to the gray level break variable of the artifact point; according to the gray abrupt change spacing distance of the scanning abnormal state and the artifact point position, adjacent contrast reduction is determined to be adopted; according to the similarity between the shape of a closed area formed by connecting artifact points in the real CT image and the shape of a corresponding area, whether correction is interrupted or not is judged; if so, adjusting the correction into the correction of the artifact area; outputting target region radiotherapy guide parameters; and if the corrected radiotherapy-guided target region radiotherapy coverage rate is abnormal, re-determining the training data volume ratio of the MRI images of the training set. According to the invention, the accuracy of radiotherapy guidance is improved.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy guidance technology, and in particular to an AI-driven multimodal image-guided radiotherapy method. Background Technology

[0002] In the existing technology, the existing technology of image-guided radiotherapy mainly focuses on improving the accuracy of radiotherapy and reducing damage to healthy tissues. Traditional methods usually rely on marker positioning or static image guidance, such as using bony or metallic markers to determine the location of the tumor. Imaging technologies such as CT and MRI are used to obtain detailed images of the tumor and its surrounding structures before treatment to plan radiotherapy. However, these traditional methods have some limitations. They often cannot reflect changes in the tumor position caused by factors such as the patient's breathing and gastrointestinal motility in real time. Secondly, markers may shift or are not applicable to all patients. Relying solely on pre-treatment imaging data, it is impossible to dynamically adjust the direction and dose of the radiotherapy beam during treatment.

[0003] Chinese Patent Publication No. CN112785632A discloses a cross-modal automatic registration method for DR and DRR images in image-guided radiotherapy based on EPID, including the following steps: S1 Image Acquisition: Before radiotherapy, the patient's CT three-dimensional image is acquired first, and two two-dimensional DRR images in orthogonal directions are digitally reconstructed, namely the anterior-posterior view at 0° and the lateral view at 90°. During radiotherapy, orthogonal planar X-ray DR images are acquired using EPID, and all DICOM format data is converted to JPG format, preprocessed, and stored as training and testing datasets; S2 Deep Learning Construction The learning model: An automatic segmentation and registration network model was established, consisting of DeepLabV3+, a cross-modal attention module, and a cross-modal optimization module. The input is the DR-DRR image group acquired in S1. Deep learning training was performed on the GPU, and the optimal network model weights were saved. S3 Bone boundary segmentation: The network trained in S2 was used to segment the bone boundary and output the boundary contour for visualization. S4 Image registration: The similarity of the boundary contours in DR and DRR obtained in S3 was calculated using the classic mutual information method, and the displacement values ​​in the X and Y directions were obtained. The registration rate was calculated based on the results. S5 EPID image-guided radiotherapy clinical application: The model was applied to EPID image-guided radiotherapy. During patient radiotherapy, the bone boundary of the DR-DRR image group was segmented and visualized in real time. The displacement values ​​in the X and Y directions were calculated for registration. The positioning was adjusted according to the displacement and visualization results for treatment.

[0004] Therefore, it can be seen that the cross-modal automatic registration method for DR and DRR images in EPD-based image-guided radiotherapy has the problem of reduced target localization accuracy due to the indistinct distinction between artifacts caused by small foreign objects obstructing the CT scan and noise artifacts during CT imaging. Summary of the Invention

[0005] To address this, the present invention provides an AI-driven multimodal image-guided radiotherapy method to overcome the problem in the prior art where the indistinguishability between artifacts caused by small foreign objects obstructing the CT scan and noise artifacts during CT imaging leads to a decrease in the accuracy of target localization.

[0006] To achieve the above objectives, the present invention provides an AI-driven multimodal image-guided radiotherapy method, comprising:

[0007] Acquire multimodal image features that are the same type as the target radiotherapy area, and train the basic AI prediction model based on the training set, test set and validation set divided by the multimodal image features to form a radiotherapy area AI prediction model.

[0008] Acquire real CT images of the target radiotherapy area;

[0009] The real CT image is input into the AI ​​prediction model of the radiotherapy area to convert the real CT image into a simulated MRI image;

[0010] The simulated MRI image is re-input into the AI ​​prediction model of the radiotherapy area to generate a simulated dimensionality-reduced CT image;

[0011] The simulated dimensionality-reduced CT image and the real CT image are superimposed to determine the location of artifact points based on dimensionality reduction;

[0012] Calculate the gray-level abrupt change of each artifact point after dimensionality reduction to determine the scanning abnormality of the real CT image;

[0013] Based on the gray-scale abrupt change interval distance of the abnormal scanning state and artifact point position, the adjacent contrast restoration is used to correct the CT scan image;

[0014] Based on the similarity between the shape of the closed region formed by connecting the edge artifact points in the real CT image and the shape of the corresponding region, it is determined whether to interrupt the correction of the dimensionality-upgraded simulated MRI image.

[0015] If interrupted, the correction will be adjusted to correct artifact regions;

[0016] The features of the multimodal image are input into the AI ​​prediction model of the radiotherapy area to output the target area radiotherapy guidance parameters and complete the radiotherapy guidance.

[0017] If the corrected radiotherapy-guided target coverage is abnormal, the proportion of training data in the training set of MRI images should be redefined.

[0018] Furthermore, determining the abnormal state of the CT scan includes,

[0019] Compare the grayscale mutation amount with the preset grayscale mutation amount;

[0020] If the grayscale mutation amount is greater than the preset grayscale mutation amount, then the CT scan is determined to be abnormal.

[0021] Furthermore, based on the scan anomaly and the gray-level abrupt change interval distance of the pixel sampling points being greater than the preset interval distance, it is determined that the CT scan image is corrected by using adjacent comparison restoration.

[0022] Furthermore, the adjacent comparison restoration is to restore the CT scan image based on the predicted connection value between the target radiotherapy area and the adjacent areas.

[0023] Further, determining whether to interrupt the correction of the upgraded simulated MRI image includes:

[0024] The similarity is compared with a preset similarity.

[0025] If the similarity is greater than or equal to the preset similarity;

[0026] The simulated MRI image is corrected by interrupting the connection and comparison under the condition that the gray-scale mutation interval distance is greater than the preset interval distance, and the closed area formed by the artifact points is corrected and reconstructed.

[0027] Furthermore, the grayscale mutation amount is the maximum absolute value of the difference between the grayscale value of a single artifact point and the average grayscale value of the non-artifact points of several real CT images;

[0028] The average gray value is the average of the gray values ​​of several non-artifact pixels within a unit radius centered on the artifact point location in the real CT image.

[0029] Furthermore, the grayscale abrupt change interval distance is the distance between every two adjacent artifact point positions.

[0030] Furthermore, the artifact point location is the corresponding location on the real CT image of the position where the brightness difference in the simulated dimensionality-reduced CT image is greater than or equal to a preset difference.

[0031] The brightness difference is the difference between the average brightness of a number of its neighboring pixels.

[0032] Furthermore, since the target radiotherapy coverage rate of the modified radiotherapy-guided radiotherapy is lower than the standard coverage rate, the proportion of training data in MRI images is increased.

[0033] Furthermore, the proportion of training data for the MRI images is the ratio of the amount of training data for the MRI images to the total amount of training data for the images.

[0034] Compared with existing technologies, the beneficial effects of this invention are as follows: The method described in this invention uses multimodal image feature fusion and AI prediction models, trained and fused using CT and MRI images. Because the artifacts caused by small foreign objects obstructing the CT scan are not clearly distinguishable from noise artifacts in CT imaging, the AI ​​is subjected to unnecessary artifact interference during processing, leading to a decrease in the accuracy of the output target area localization. By performing a cyclical process of CT, simulated MRI, and dimensionality-reduced CT for bidirectional detection and amplification of differences, the accuracy and fit of the AI ​​model training are improved, avoiding localization distortion caused by unnecessary interference conditions. Furthermore, by quantitatively judging whether abnormalities occur in the CT scan and comparing the grayscale abrupt change with preset values, if small foreign objects obstruct the CT scan, the interfered CT is corrected to improve the accuracy of the CT image data. This improves the accuracy of target area localization guidance output by the AI ​​model. Since local grayscale is easily affected by filtering and noise, making it more prominent, setting a threshold based on the grayscale abrupt change interval and performing adjacent comparison restoration enhances the targeting and effectiveness of the correction method, reducing the disruption of tissue boundary continuity caused by traditional interpolation correction. Because the difference in brightness is easily affected by processing parameters during AI model training, leading to false positives, artifact locations are screened to reduce noise and artifact interference before CT is input into the AI, thereby reducing the AI's image enhancement and filtering processing time and improving image processing efficiency. Artifacts caused by grayscale drift due to respiratory rate and intensity are determined by judging the similarity of artifacts to determine whether artifact formation is related to the target scan area, thus determining the image processing method and improving image processing accuracy.

[0035] Furthermore, the method described in this invention can identify the impact of small-volume foreign objects or noise artifacts on image quality by judging abnormal CT scan states. Sudden grayscale changes can lead to amplification of differences in the subsequent AI fusion process, resulting in inaccurate image judgment. Judging abnormal scan states and determining correction methods achieves specialized image processing, thereby improving the accuracy of image output.

[0036] Furthermore, the method described in this invention sets a threshold based on the gray-scale mutation interval distance and performs an adjacent comparison restoration method. This addresses the issues of traditional interpolation methods potentially disrupting the continuity of tissue boundaries and the need for differentiated processing of small-scale isolated artifacts and large-area continuous artifacts, which leads to decreased processing efficiency. By performing an adjacent comparison restoration method, the mutation amount is smoothed, thereby improving the accuracy of CT image processing.

[0037] Furthermore, the method described in this invention distinguishes between grayscale drift artifacts caused by respiratory rate and intensity and actual lesions in the target scanning area by judging the similarity between the location of artifact points and the shape of closed regions. This avoids image processing deviations caused by misjudgment and improves the accuracy of AI models in identifying target regions. By analyzing the similarity between the shape of the closed region formed by the artifact points and the corresponding actual anatomical structure, functional artifacts, such as those caused by respiratory motion, and pathological artifacts, such as tumors or lesion areas, are distinguished, thereby reducing unnecessary image correction operations and improving the accuracy of radiotherapy-guided target localization.

[0038] Furthermore, the method described in this invention addresses the model prediction bias caused by the imbalance of different modal feature weights during multimodal image fusion by dynamically adjusting the proportion of MRI images in the training set. When the corrected radiotherapy-guided target area coverage is abnormal, increasing the proportion of MRI images in the training data enhances the AI ​​model's sensitivity to soft tissue contrast, enabling the model to more accurately capture target area features and thus improve the overall performance of radiotherapy guidance.

[0039] Furthermore, the method described in this invention determines the location of artifact points by superimposing simulated MRI images with real CT images, thereby improving the accuracy of artifact localization. Compared with single-modal image analysis methods, multi-modal image fusion analysis can utilize the advantages of CT and MRI respectively, improving the sensitivity and specificity of artifact detection in complex scenes. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of the AI-driven multimodal image-guided radiotherapy method according to an embodiment of the present invention;

[0041] Figure 2 This is a flowchart illustrating the determination of abnormal CT scan states using an AI-driven multimodal image-guided radiotherapy method according to an embodiment of the present invention.

[0042] Figure 3 This is a flowchart illustrating the process of determining whether to interrupt the correction of the upgraded simulated MRI image using an AI-driven multimodal image radiotherapy guidance method according to an embodiment of the present invention.

[0043] Figure 4 This is a flowchart illustrating the correction and reconstruction process of an AI-driven multimodal image-guided radiotherapy method according to an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0045] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0046] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0047] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0048] Please see Figure 1 , Figure 2 , Figure 3 as well as Figure 4 The diagrams shown are, respectively, the overall flowchart of the AI-driven multimodal image-guided radiotherapy method according to an embodiment of the present invention, the flowchart for determining abnormal CT scan states, and the flowcharts for determining whether to interrupt the correction of the upgraded simulated MRI image and the correction and reconstruction flowchart. The present invention provides an AI-driven multimodal image-guided radiotherapy method, comprising:

[0049] Acquire multimodal image features that are the same type as the target radiotherapy area, and train the basic AI prediction model based on the training set, test set and validation set divided by the multimodal image features to form a radiotherapy area AI prediction model.

[0050] Acquire real CT images of the target radiotherapy area;

[0051] The real CT image is input into the AI ​​prediction model of the radiotherapy area to convert the real CT image into a simulated MRI image;

[0052] The simulated MRI image is re-input into the AI ​​prediction model of the radiotherapy area to generate a simulated dimensionality-reduced CT image;

[0053] The simulated dimensionality-reduced CT image and the real CT image are superimposed to determine the location of artifact points based on dimensionality reduction;

[0054] Calculate the gray-level abrupt change of each artifact point after dimensionality reduction to determine the scanning abnormality of the real CT image;

[0055] Based on the gray-scale abrupt change interval distance of the abnormal scanning state and artifact point position, the adjacent contrast restoration is used to correct the CT scan image;

[0056] Based on the similarity between the shape of the closed region formed by connecting the edge artifact points in the real CT image and the shape of the corresponding region, it is determined whether to interrupt the correction of the dimensionality-upgraded simulated MRI image.

[0057] If interrupted, the correction will be adjusted to correct artifact regions;

[0058] The features of the multimodal image are input into the AI ​​prediction model of the radiotherapy area to output the target area radiotherapy guidance parameters and complete the radiotherapy guidance.

[0059] If the corrected radiotherapy-guided target coverage is abnormal, the proportion of training data in the training set of MRI images should be redefined.

[0060] Specifically, the radiotherapy equipment is the MCL system.

[0061] Specifically, the artifact point location is the corresponding position on the real CT image of the position in the simulated dimensionality-reduced CT image where the brightness difference is greater than or equal to a preset difference.

[0062] The brightness difference is the difference between the average brightness of a number of its neighboring pixels.

[0063] Specifically, the general range of the preset difference amount is [500HU, 1000HU], and the preferred embodiment of the preset difference amount is 850HU.

[0064] In this embodiment, the basic AI prediction model is the Cycle-GAN model;

[0065] The training process of the basic AI prediction model involves using historically acquired real CT images of the target radiotherapy area as the source domain and the corresponding MRI images of the target radiotherapy area as the target domain. This is achieved by simulating planar images of the tissue layer thicknesses of the target radiotherapy area from two-dimensional real CT images and superimposing several planar images to achieve dimensionality upscaling, transforming the real CT image into a simulated MRI image that resembles an MRI image. The discriminator in the basic AI prediction model is used to determine whether the image input to the discriminator is the transformed simulated MRI image or the MRI image of the corresponding target radiotherapy area for training. The specific discriminator's identification process is existing technology well-known to those skilled in the art and will not be elaborated here. Similarly, the MRI image of the target radiotherapy area can be used as the source domain, and the corresponding real CT image of the target radiotherapy area as the target domain to achieve reverse simulation.

[0066] Specifically, the grayscale mutation is used to quantify the degree of abnormality of a single artifact point and to determine the abnormal state of the scan. Its calculation object is the pixel point on the real CT image, reflecting the degree of deviation of the artifact point from its surrounding normal tissue area. The larger the value, the more severe the grayscale difference between the point and the surrounding normal tissue, and thus determine whether the abnormal point is caused by scanning abnormalities, such as metal artifacts, motion artifacts, etc.

[0067] The brightness difference is calculated on the pixels in the simulated dimensionality-reduced CT image. It is used to identify and locate artifact points, reflecting the brightness inconsistency between a point in the simulated image and its local neighborhood. It is used to detect isolated points or small areas that are abnormally bright or dark in the simulated image. Once the artifact point is located in the simulated image, it is mapped back to the corresponding position in the real CT image.

[0068] Specifically, the superposition and overlap process involves aligning the edges of the target radiotherapy area in the simulated dimensionality-reduced CT image and the real CT image and overlapping them vertically.

[0069] In implementation, the method of this invention uses multimodal image feature fusion and AI prediction models, trained and fused using CT and MRI images. Because artifacts caused by small foreign objects obstructing the CT scan are not easily distinguishable from noise artifacts in CT imaging, the AI ​​is subjected to unnecessary artifact interference during processing, leading to a decrease in the accuracy of the output target area localization. By performing a cycle of CT, simulated MRI, and dimensionality-reduced CT for bidirectional detection and amplification of differences, the accuracy and fit of the AI ​​model training are improved, avoiding localization distortion caused by unnecessary interference conditions. By quantitatively judging whether there are abnormalities in the CT scan and comparing the grayscale abrupt change with preset values, if small foreign objects obstruct the CT scan, the interfered CT is corrected to improve the accuracy of the CT image data and thus improve the AI's accuracy. The accuracy of target area localization guidance output by the I model is improved. Since local grayscale is easily affected by filtering and noise, becoming increasingly prominent, a threshold is set based on the grayscale abrupt change interval distance, and adjacent comparison restoration is performed, enhancing the targeting and effectiveness of the correction method and reducing the disruption of tissue boundary continuity caused by traditional interpolation correction. Because the difference in brightness is easily affected by processing parameters during AI model training, leading to false positives, artifact locations are screened to reduce noise and artifact interference before CT is input into AI, thereby reducing the AI's image enhancement and filtering processing time and improving image processing efficiency. Due to the influence of respiratory rate and intensity, grayscale drift artifacts appear in CT scan images. The similarity of artifacts is judged to determine whether artifact formation is related to the target scan area, thus determining the image processing method and improving the accuracy of image processing.

[0070] Specifically, determining the abnormal state of the CT scan includes,

[0071] Compare the grayscale mutation amount with the preset grayscale mutation amount;

[0072] If the grayscale mutation amount is greater than the preset grayscale mutation amount, then the CT scan is determined to be abnormal.

[0073] In practice, the method described in this invention can identify the impact of small foreign objects or noise artifacts on image quality by judging abnormal CT scan states. Sudden grayscale changes can lead to amplification of differences in the subsequent AI fusion process, resulting in inaccurate image judgment. Judging abnormal scan states and determining correction methods achieves specialized image processing, thereby improving the accuracy of image output.

[0074] Specifically, based on the scanning anomaly and the gray-level abrupt change interval of the pixel sampling points being greater than the preset interval distance, it is determined that the CT scan image will be corrected by using adjacent comparison restoration.

[0075] Specifically, the adjacent contrast restoration is to restore the CT scan image based on the predicted connection value between the target radiotherapy area and the adjacent areas.

[0076] Specifically, the process of reconstructing CT scan images based on the predicted connectivity values ​​between the target radiotherapy area and adjacent areas is as follows:

[0077] Extract image features from adjacent regions. Image features include grayscale values ​​and texture information.

[0078] Image features are used as training values ​​and input into the regression model, which then outputs predicted connection values.

[0079] The grayscale values ​​of artifact points within the target radiotherapy area are smoothed using linear interpolation. The input data for linear interpolation includes the predicted transition value and the grayscale values ​​of several pixels in adjacent areas.

[0080] Specifically, the adjacent region is the area extending from the edge contour of the target radiotherapy area to outside the 5-pixel bandwidth line.

[0081] Specifically, the grayscale mutation amount is the maximum absolute value of the difference between the grayscale value of a single artifact point and the average grayscale value of the non-artifact points of several real CT images;

[0082] The average gray value is the average of the gray values ​​of several non-artifact pixels within a unit radius centered on the artifact point location in the real CT image.

[0083] Specifically, the grayscale abrupt change interval distance is the distance between every two adjacent artifact point positions.

[0084] Specifically, the unit radius is 1 cm.

[0085] Specifically, under the conditions of a CT image scan layer thickness of 6mm and a tube voltage of 120kV, the general range of the preset grayscale mutation amount is [4HU, 8HU], and the preferred embodiment of the preset grayscale mutation amount is 5HU.

[0086] Those skilled in the art will understand that the selectable range of the preset grayscale mutation amount and the preferred embodiment provided in this embodiment are the values ​​that are most effective in solving the technical problem of the present invention under the conditions of a CT image scanning layer thickness of 6 mm and a tube voltage of 120 kV. In actual applications or experiments, those skilled in the art can make adaptive adjustments to the preset grayscale mutation amount according to the actual application environment and application scenario.

[0087] In practice, the method of the present invention sets a threshold based on the gray-scale mutation interval distance and performs adjacent comparison restoration. Due to the problem that traditional interpolation methods may destroy the continuity of tissue boundaries and the problem that small-scale isolated artifacts and large-area continuous artifacts need to be differentiated, resulting in a decrease in processing efficiency, the adjacent comparison restoration method is used to smooth the mutation amount, thereby improving the accuracy of CT image processing.

[0088] Specifically, determining whether to interrupt the correction of the upgraded simulated MRI image includes:

[0089] The similarity is compared with a preset similarity.

[0090] If the similarity is greater than or equal to the preset similarity;

[0091] The simulated MRI image is corrected by interrupting the connection and comparison under the condition that the gray-scale mutation interval distance is greater than the preset interval distance, and the closed area formed by the artifact points is corrected and reconstructed.

[0092] Specifically, the process of correction and reconstruction is as follows:

[0093] U-Net is used to extract the mask of the closed region and its continuous homology features are calculated.

[0094] Extract artifact points and surrounding 3 mm of normal tissue;

[0095] The generator is used to repair artifact loss at artifact points.

[0096] Specifically, the similarity is calculated by using the Douglas-Peucker algorithm to calculate the polygon area overlap rate by comparing the shape of the closed region connected by the artifact points with the shape of the corresponding region.

[0097] Specifically, under the conditions of a CT image scan layer thickness of 6 mm and a tube voltage of 120 kV, the general range of the preset similarity value is [0.65, 0.8], and the preferred embodiment of the preset similarity value is 0.75.

[0098] Those skilled in the art will understand that the range of preset similarity and the preferred embodiment provided in this embodiment are the values ​​that are most effective in solving the technical problem of the present invention under the conditions of a CT image scan layer thickness of 6 mm and a tube voltage of 120 kV. In actual applications or experiments, those skilled in the art can make adaptive adjustments to the preset similarity according to the actual application environment and application scenario.

[0099] In practice, the method described in this invention distinguishes between grayscale drift artifacts caused by respiratory rate and intensity and actual lesions in the target scanning area by judging the similarity between the location of artifact points and the shape of closed regions. This avoids image processing deviations caused by misjudgment and improves the accuracy of AI models in identifying target regions. By analyzing the similarity between the shape of the closed region formed by the artifact points and the corresponding actual anatomical structure, functional artifacts, such as those caused by respiratory motion, and pathological artifacts, such as tumors or lesion areas, are distinguished, thereby reducing unnecessary image correction operations and improving the accuracy of target localization in radiotherapy-guided radiotherapy.

[0100] Specifically, the target area radiotherapy coverage based on the modified radiotherapy guidance is lower than the standard coverage, thus increasing the proportion of training data from MRI images.

[0101] In practice, the method described in this invention addresses the model prediction bias caused by the imbalance of different modal feature weights during multimodal image fusion by dynamically adjusting the proportion of MRI images in the training set. When the corrected radiotherapy-guided target area coverage is abnormal, increasing the proportion of MRI images in the training data enhances the AI ​​model's sensitivity to soft tissue contrast, enabling the model to more accurately capture target area features and thus improve the overall performance of radiotherapy guidance.

[0102] Specifically, the proportion of training data for the MRI images is the ratio of the amount of training data for the MRI images to the total amount of training data for the images.

[0103] In practice, the method described in this invention determines the location of artifact points by superimposing simulated MRI images with real CT images, thereby improving the accuracy of artifact localization. Compared with single-modal image analysis methods, multi-modal image fusion analysis can utilize the advantages of CT and MRI respectively to improve the sensitivity and specificity of artifact detection in complex scenes.

[0104] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An AI-driven multimodal image-guided radiotherapy method, characterized in that, include: Acquire multimodal image features that are the same type as the target radiotherapy area, and train the basic AI prediction model based on the training set, test set and validation set divided by the multimodal image features to form a radiotherapy area AI prediction model. Acquire real CT images of the target radiotherapy area; The real CT image is input into the AI ​​prediction model of the radiotherapy area to convert the real CT image into a simulated MRI image; The simulated MRI image is re-input into the AI ​​prediction model of the radiotherapy area to generate a simulated dimensionality-reduced CT image; The simulated dimensionality-reduced CT image and the real CT image are superimposed to determine the location of artifact points based on dimensionality reduction; Calculate the gray-level abrupt change of each artifact point after dimensionality reduction to determine the scanning abnormality of the real CT image; Based on the gray-scale abrupt change interval distance of the abnormal scanning state and artifact point position, the adjacent contrast restoration is used to correct the CT scan image; Based on the similarity between the shape of the closed region formed by connecting the edge artifact points in the real CT image and the shape of the corresponding region, it is determined whether to interrupt the correction of the dimensionality-upgraded simulated MRI image. If interrupted, the correction will be adjusted to correct artifact regions; The features of the multimodal image are input into the AI ​​prediction model of the radiotherapy area to output the target area radiotherapy guidance parameters and complete the radiotherapy guidance. If the corrected radiotherapy-guided target coverage is abnormal, the proportion of training data in the training set of MRI images should be redefined.

2. The AI-driven multimodal image-guided radiotherapy method according to claim 1, characterized in that, The determination of abnormal CT scan conditions includes, Compare the grayscale mutation amount with the preset grayscale mutation amount; If the grayscale mutation amount is greater than the preset grayscale mutation amount, then the CT scan is determined to be abnormal.

3. The AI-driven multimodal image-guided radiotherapy method according to claim 2, characterized in that, Based on the scanning anomaly and the gray-level abrupt change interval of the pixel sampling points being greater than the preset interval distance, it is determined that the CT scan image will be corrected by using adjacent comparison restoration.

4. The AI-driven multimodal image-guided radiotherapy method according to claim 3, characterized in that, The adjacent contrast restoration is to restore the CT scan image based on the predicted connection value between the target radiotherapy area and the adjacent area.

5. The AI-driven multimodal image-guided radiotherapy method according to claim 4, characterized in that, Determining whether to interrupt the correction of the upgraded simulated MRI image includes: The similarity is compared with a preset similarity. If the similarity is greater than or equal to the preset similarity; The simulated MRI image is corrected by interrupting the connection and comparison under the condition that the gray-scale mutation interval distance is greater than the preset interval distance, and the closed area formed by the artifact points is corrected and reconstructed.

6. The AI-driven multimodal image-guided radiotherapy method according to claim 5, characterized in that, The grayscale mutation amount is the maximum absolute value of the difference between the grayscale value of a single artifact point and the average grayscale value of the non-artifact points of several real CT images. The average gray value is the average of the gray values ​​of several non-artifact pixels within a unit radius centered on the artifact point location in the real CT image.

7. The AI-driven multimodal image-guided radiotherapy method according to claim 6, characterized in that, The grayscale abrupt change interval is the distance between every two adjacent artifact point positions.

8. The AI-driven multimodal image-guided radiotherapy method according to claim 7, characterized in that, The artifact point location is the corresponding position on the real CT image of the location in the simulated dimensionality-reduced CT image where the brightness difference is greater than or equal to a preset difference. The brightness difference is the difference between the average brightness of a number of its neighboring pixels.

9. The AI-driven multimodal image-guided radiotherapy method according to claim 8, characterized in that, The modified radiotherapy-guided target coverage is lower than the standard coverage, so the proportion of training data in MRI images is increased.

10. The AI-driven multimodal image-guided radiotherapy method according to claim 9, characterized in that, The proportion of training data for the MRI images is the ratio of the amount of training data for the MRI images to the total amount of training data for the images.

Images