Adaptive planning method based on orthogonal dr images, and related device
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- MEVION MEDICAL EQUIPMENT CO LTD
- Filing Date
- 2025-11-10
- Publication Date
- 2026-05-21
Smart Images
Figure CN2025133720_21052026_PF_FP_ABST
Abstract
Description
Adaptive planning method and related equipment based on orthogonal DR
[0001] This application claims priority to Chinese Patent Application No. CN202411632465.9, filed on November 15, 2024, which is incorporated herein by reference in its entirety. Technical Field
[0002] This application relates to the field of radiotherapy technology, such as adaptive planning methods and related equipment based on orthogonal DR. Background Technology
[0003] In the field of radiotherapy, ensuring the precision of treatment is crucial, especially when dealing with the complex and variable disease of tumors. Traditional radiotherapy plans are often based on the location and shape of the tumor at the time of the patient's initial diagnosis. However, during radiotherapy, the tumor may shrink, move, or change shape due to the treatment response, which may cause the original treatment plan to become invalid or require adjustment. To address this, adaptive planning has emerged, which can adjust the treatment plan in real time according to the actual changes in the patient's tumor during the treatment process.
[0004] Adaptive planning refers to adjusting the treatment plan in real time during radiotherapy based on changes in the location and shape of the patient's tumor to ensure the accuracy of the treatment; adaptive planning allows for real-time adjustment of radiation dose and beam direction based on the latest imaging data (such as CT images).
[0005] However, implementing adaptive planning faces a series of challenges. First, to monitor tumor changes, traditional practices require regular CT scans to ensure there are no significant changes in the tumor and surrounding tissues, increasing patient radiation exposure and consuming considerable time. Second, traditional adaptive planning heavily relies on 3D imaging data and planning CT scans to assess target area changes, which not only prolongs the treatment cycle but also increases the burden on hospital equipment and patient treatment costs. Therefore, further improvements to traditional adaptive planning are necessary. Summary of the Invention
[0006] Based on the above problems, the purpose of this application is to provide an adaptive planning method and related equipment based on orthogonal DR, which can achieve real-time assessment of tumor changes by combining orthogonal 2D imaging and deformation field reconstruction technology, while reducing the number of CT scans.
[0007] The objective of this application is achieved through the following technical solution:
[0008] Firstly, this application proposes an adaptive planning method based on orthogonal DR, the method comprising:
[0009] The patient's planned CT images are automatically segmented to obtain segmentation results; the segmentation results include a delineation set of the patient's target area or organs at risk.
[0010] After positioning the patient, DR images of the patient in two orthogonal directions are acquired; and the planned CT images of the patient are projected in the space of the positioning coordinate system to obtain the corresponding orthogonal DRR images.
[0011] The DR images in the two orthogonal directions are registered with the corresponding DRR images to obtain 2D deformable meshes in the two orthogonal directions.
[0012] The 3D deformation field is reconstructed using the 2D deformation mesh;
[0013] The 3D deformation field is applied to the delineation set to obtain the deformation components of the patient's corresponding target area or organs at risk; based on the deformation components, it is determined whether to reacquire the patient's CT images.
[0014] In one possible implementation, the patient's planned CT images are automatically segmented using the 3DUNet automatic segmentation network.
[0015] In one possible implementation, the automatic segmentation further includes obtaining the loss function of the 3DUNet automatic segmentation network, wherein the loss function is determined by the Dice loss and cross-information entropy loss between the ground truth label and the predicted segmentation;
[0016] The loss function of the 3DUNet automatic segmentation network is:
[0017] Among them, L Dice () represents the Dice loss function; L CE () represents the cross-information entropy loss function; L loss () represents the total loss function; e represents the true label; e ′ For predicting segmentation; t represents the voxel index; N is the total number of voxels; a g dn is the weight of the g-th segmentation category; w is the balancing coefficient used to balance the weights of the Dice loss function and the cross-entropy loss function; g is the g-th segmentation category; dn is the total number of segmentation categories.
[0018] In one possible implementation, after positioning the patient, DR images of the patient in two orthogonal directions are acquired; and the planned CT images of the patient are projected into the positioning coordinate system to obtain corresponding orthogonal DRR images; including:
[0019] The patient's DR images are acquired using two orthogonal DR acquisition devices;
[0020] In the positioning coordinate system space, the patient's planned CT images are projected along the directions of the two DR acquisition devices to generate corresponding orthogonal DRR images.
[0021] In one possible implementation, the registration of the DR images and corresponding DRR images in two orthogonal directions to obtain 2D deformable meshes in the two orthogonal directions includes:
[0022] For each direction, the DR image and the corresponding DRR image are input as a pair into the VoxelMorph network to obtain the output velocity field;
[0023] The velocity field is converted into a differential deformation field by scaling the integral layer using a differentiable sum of squares;
[0024] For each direction, extract deformation information from the differential deformation field;
[0025] A 2D deformation mesh is generated using the deformation information.
[0026] In one possible implementation, reconstructing the 3D deformation field through the 2D deformation mesh includes:
[0027] By utilizing the invertible relationship between the components of the 2D deformable mesh on the orthogonal plane and the 3D vector field, the nodal displacements in the 3D deformable field are calculated through two-dimensional to three-dimensional back projection and geometric transformation.
[0028] Based on the node displacements, a complete 3D deformation field is constructed.
[0029] In one possible implementation, applying the 3D deformation field to the delineation set to obtain deformation components of the patient's corresponding target area or organs at risk; and determining whether to re-acquire the patient's CT images based on the deformation components, includes:
[0030] If the deformation component exceeds its deformation threshold, the patient's CT images are reacquired;
[0031] If the deformation component is less than or equal to its deformation threshold, the patient's planned CT and delineation set are corrected according to the deformation component, and the treatment plan is regenerated using the original optimized parameters.
[0032] Secondly, this application provides an adaptive planning apparatus based on orthogonal DR, the apparatus comprising:
[0033] The segmentation and delineation module is used to automatically segment the patient's planned CT images and obtain segmentation results; the segmentation results include a delineation set of the patient's target area or organs at risk;
[0034] The image acquisition module is used to acquire DR images of the patient in two orthogonal directions after positioning the patient; and to project the planned CT images of the patient into the space of the positioning coordinate system to obtain corresponding orthogonal DRR images.
[0035] The 2D deformation acquisition module is used to register the DR image and the corresponding DRR image in two orthogonal directions respectively to obtain the 2D deformation mesh in the two orthogonal directions.
[0036] The 3D deformation field acquisition module is used to reconstruct the 3D deformation field through the 2D deformation mesh;
[0037] The determination module is used to apply the 3D deformation field to the delineation set to obtain the deformation components of the target area or organs at risk of the patient; and to determine whether to reacquire the patient's CT images based on the deformation components.
[0038] Thirdly, this application provides a treatment system, the system comprising:
[0039] A radiation therapy device used to administer radiation therapy to a patient according to a treatment plan;
[0040] The adaptive planning device based on orthogonal DR described in this application is used to obtain an adaptive treatment plan for a patient.
[0041] Fourthly, this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned method or the function of the device described in this application.
[0042] Fifthly, this application provides a computer-readable storage medium that stores computer instructions, which, when read by a computer, execute the steps of the method described above in this application.
[0043] Compared with existing technologies, the beneficial effects of this application include at least the following: Utilizing orthogonal 2D images acquired during each treatment session, an orthogonal 2D deformation mesh is rapidly constructed using deformation registration technology, thereby reconstructing a 3D deformation field; by evaluating the changes in the displacement and threshold of each voxel in the target area and the corresponding region of the organ at risk contour, it can automatically determine whether CT data acquisition needs to be repeated or adaptive plan adjustments need to be made, significantly reducing the number of CT scans. Necessary CT image acquisition is only performed when tumor changes exceed a set threshold, thus greatly reducing the patient's radiation dose; the method of reconstructing the 3D deformation field using an orthogonal 2D deformation mesh avoids the excessive reliance on 3D image data in traditional adaptive planning, further shortening the treatment cycle and reducing treatment costs; and without introducing additional detection or acquisition equipment while ensuring the real-time performance and accuracy of the adaptive treatment process, the treatment process is more efficient and convenient. Attached Figure Description
[0044] This application will be further described below with reference to the accompanying drawings and specific embodiments.
[0045] Figure 1 is a schematic diagram of the adaptive planning method of orthogonal DR according to an embodiment of this application;
[0046] Figure 2 is a flowchart illustrating the adaptive planning method of orthogonal DR according to an embodiment of this application;
[0047] Figure 3 is a schematic diagram of the VoxelMorph end-to-end unsupervised architecture according to an embodiment of this application;
[0048] Figure 4 is a schematic diagram of an orthogonal imaging system according to an embodiment of this application;
[0049] Figure 5 is a schematic diagram of a surface in the 3D deformation field of the deformation mesh component back projection in the orthogonal plane according to an embodiment of this application;
[0050] Figure 6 is a schematic diagram of the 2D registration mesh after orthogonal DR and DRR registration according to an embodiment of this application;
[0051] Figure 7 is a schematic diagram of the adaptive planning method apparatus for orthogonal DR according to an embodiment of this application. Detailed Implementation
[0052] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided to make this application more complete and comprehensive, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar structures, and therefore repeated descriptions of them will be omitted.
[0053] The terms used in this application to express position and direction are illustrated with the accompanying drawings, but may be changed as needed, and all such changes are included within the scope of protection of this application.
[0054] Referring to Figures 1 and 2: Figure 1 is a schematic diagram of the adaptive planning method of orthogonal DR according to an embodiment of this application; Figure 2 is a flowchart of the adaptive planning method of orthogonal DR according to an embodiment of this application.
[0055] This application provides an adaptive planning method based on orthogonal DR, the method comprising:
[0056] The patient's planned CT images are automatically segmented to obtain segmentation results; the segmentation results include a delineation set of the patient's target area or organs at risk.
[0057] After positioning the patient, DR images of the patient in two orthogonal directions are acquired; and the planned CT images of the patient are projected in the space of the positioning coordinate system to obtain the corresponding orthogonal DRR images.
[0058] The DR images in the two orthogonal directions are registered with the corresponding DRR images to obtain 2D deformable meshes in the two orthogonal directions.
[0059] The 3D deformation field is reconstructed using the 2D deformation mesh;
[0060] The 3D deformation field is applied to the delineation set to obtain the deformation components of the patient's corresponding target area or organs at risk; based on the deformation components, it is determined whether to reacquire the patient's CT images.
[0061] The working principle of the above technical solution is as follows: First, image segmentation technology is used to automatically segment the patient's planned CT images to accurately distinguish different tissue structures of the patient, especially the target area and organs at risk; after segmentation, a delineation set containing the target area and organs at risk is automatically generated; after the patient is accurately positioned, DR (digital X-ray) imaging technology is used to acquire the patient's DR images in two orthogonal directions (such as front-back and left-right); the patient's planned CT images are projected in the space of the positioning coordinate system to generate digital reconstructed images (DRR) corresponding to the DR images; these DRR images are simulated X-ray images used for registration with the actual DR images. Image registration technology is used to register DR images with their corresponding DRR images in two orthogonal directions. The purpose of registration is to find the optimal alignment between the DR and DRR images to determine the patient's actual deformation. After registration, 2D deformation meshes in two orthogonal directions are generated based on the registration results. These meshes reflect the patient's deformation in different directions. The two 2D deformation meshes are then fused and reconstructed to generate a 3D deformation field. This 3D deformation field comprehensively reflects the patient's deformation in three-dimensional space. The generated 3D deformation field is applied to the previously generated delineation set to obtain the deformation components of the patient's target area or organs at risk. Based on the obtained deformation components, the patient's deformation is analyzed in detail. If the deformation exceeds a preset threshold or meets the criteria for reassessment, further steps are required. If the deformation analysis indicates that the patient's CT images need to be reacquired to update the treatment plan, the CT images are reacquired to ensure the accuracy and effectiveness of the treatment plan.
[0062] The effects of the above technical solution are as follows: By automatically segmenting the planned CT images, the delineation sets of the patient's target area and organs at risk can be accurately obtained, providing accurate basic data for subsequent deformation analysis and treatment plan adjustments; by utilizing the registration of orthogonal DR images and DRR images, the patient's deformation during treatment can be accurately captured, thereby generating accurate 2D deformation meshes and 3D deformation fields. Applying the 3D deformation field to the delineation set allows for the accurate calculation of the deformation components of the target area and organs at risk, providing a reliable basis for adjusting the treatment plan. By analyzing the deformation components, it can be determined whether the patient's deformation level has reached the standard requiring a re-acquisition of CT images, avoiding unnecessary CT image acquisition, reducing patient radiation exposure and medical costs. A single ordinary CT scan requires 5-10 mSv, while orthogonal X-ray only requires 0.1 mSv. If a treatment plan requires 30 fractions, reducing one CT scan is equivalent to reducing the dose to 1 / 50th to 1 / 100th of the original dose, greatly reducing the harm of radiation dose to the patient. At the same time, it also reduces the time delay caused by frequent CT image acquisition, improving treatment efficiency. Originally, ART (Adaptive Radiation Therapy) required deformation field assessment using 3D data, meaning that 3D data was needed to calculate the deformation field from the adaptive treatment plan itself, and the dose of the original plan was deformed by the deformation field. This application avoids the necessity of 3D data by reconstructing the 3D deformation field using orthogonal 2D deformation meshes, while shortening the treatment cycle and reducing treatment costs. Under the premise of ensuring the real-time performance and accuracy of the adaptive treatment process, no additional detection or acquisition equipment is introduced, making the treatment process more efficient and convenient.
[0063] In one possible implementation, the planned CT images of a patient are automatically segmented using the 3DUNet automatic segmentation network; and the loss function of the 3DUNet automatic segmentation network is obtained by using the Dice loss and cross-entropy loss between the ground truth and predicted segments; the loss function of the 3DUNet automatic segmentation network is:
[0064] Among them, L Dice () represents the Dice loss function; L CE () represents the cross-information entropy loss function; L loss () represents the total loss function; e represents the true label; e ′ For predicting segmentation; t represents the voxel index; N is the total number of voxels; a g dn is the weight of the g-th segmentation category; w is the balancing coefficient used to balance the weights of the Dice loss function and the cross-entropy loss function; g is the g-th segmentation category; dn is the total number of segmentation categories.
[0065] The working principle and effects of the above technical solution are as follows:
[0066] 3DUNet is a deep learning-based 3D image segmentation network inspired by the traditional U-Net architecture, but extended and optimized for 3D image data. The network consists of an encoder and a decoder. The encoder extracts image features, while the decoder performs accurate segmentation based on these features. The 3DUNet network framework used in this embodiment possesses excellent automatic adaptation capabilities, adaptively adjusting its network structure, preprocessing methods, and training parameters according to the different characteristics of the input 3D CT image data (e.g., resolution, size, dimensions). This means that for different organs at risk (OARs) or target areas of different body parts, there is no need to redesign the network structure or adjust hyperparameters; the network can automatically adjust the U-Net depth, convolutional kernel size, pooling layers, and other core structures based solely on the unique properties of the dataset (e.g., image size, dimensions, annotation details).
[0067] 3DUNet can intelligently identify and apply the best network configuration by analyzing the statistical characteristics of the input data. During the training phase, the network learns how to extract key features from the input CT images and accurately segment the target area and organs at risk based on these features.
[0068] Dice loss is used to measure the similarity between the predicted segmentation result and the true label; the higher the Dice coefficient, the closer the predicted result is to the true situation; the Dice loss function is the negative value of the Dice coefficient, so the optimization objective is to minimize this loss.
[0069] Cross-information entropy loss is used to evaluate the difference between the probability distribution of the predicted segmentation result and the true label; the smaller the cross-information entropy loss, the more accurate the prediction result.
[0070] By combining Dice loss and cross-information entropy loss, and introducing weight coefficients and balancing coefficients, the weights between different segmentation categories and the two loss functions are balanced. The total loss function is the key to optimizing the network model. By minimizing this loss, the accuracy and robustness of network segmentation can be improved.
[0071] The specific segmentation process is as follows:
[0072] The patient's planned CT images are used as input data, and the images contain the patient's three-dimensional anatomical information;
[0073] The 3DUNet network extracts features from the input CT image, gradually abstracting key information from the image through layer-by-layer convolution and pooling operations.
[0074] During the decoding stage, the network performs segmentation prediction based on the extracted features and generates predicted segmentation results;
[0075] The predicted segmentation results are compared with the ground truth labels, and the Dice loss and cross-information entropy loss are calculated. The network is then optimized based on the total loss function. Through multiple iterations of training, the segmentation accuracy can be gradually improved. The total loss function, by introducing segmentation region weights and a balancing coefficient, can flexibly adapt to multi-region segmentation tasks. The weight 'a'... g This allows different levels of emphasis to be given to different segmentation categories, which is especially important for handling OARs or target regions with different importance. For example, when segmenting the head and neck region, there are multiple OARs, each of which is a category and may include multiple regions of interest (ROIs). The balancing coefficient w is used to adjust the relative importance between Dice loss and cross-entropy loss, ensuring that the two can work together during training to optimize segmentation performance.
[0076] Ultimately, the 3DUNet network outputs accurate segmentation results, including delineated sets of the patient's target area and organs at risk; these segmentation results will serve as the basis for subsequent adaptive planning methods.
[0077] In one possible implementation, after positioning the patient, DR images of the patient in two orthogonal directions are acquired; and the planned CT images of the patient are projected into the positioning coordinate system to obtain corresponding orthogonal DRR images; including:
[0078] The patient's DR images are acquired using two orthogonal DR acquisition devices;
[0079] In the positioning coordinate system space, the patient's planned CT images are projected along the directions of the two DR acquisition devices to generate corresponding orthogonal DRR images.
[0080] The working principle of the above technical solution is as follows:
[0081] Based on the patient's specific condition and the requirements of the radiotherapy plan, a suitable position is selected for patient positioning, such as supine or prone. After the patient is positioned, DR images of the patient are acquired in two orthogonal directions using a DR acquisition device. The DR acquisition device can be an imaging panel installed on the ceiling of the treatment room, with the isocenter of the treatment room as the origin. DR images of the patient are acquired through two orthogonal imaging panels. Based on the direction of the acquired DR images, the projection direction of the planned CT image in the positioning coordinate system is determined. The planned CT image is projected using a ray projection algorithm (or other similar algorithm) to simulate the image formed after X-rays pass through the patient's body. This process is similar to "flattening" (converting) a three-dimensional image into a two-dimensional image, but retaining key structural information. The planned CT image is projected along the two orthogonal directions to generate corresponding orthogonal DRR images.
[0082] The effects of the above technical solution are as follows: by acquiring DR images in two orthogonal directions, image information of the patient from different perspectives can be obtained; it provides a data foundation for subsequent image registration, helps to complete the construction from 2D deformation mesh to 3D deformation field, and improves the registration accuracy; by projecting the planned CT image into the positioning coordinate system space, orthogonal DRR images are generated, making the registration between DR images and DRR images more intuitive and accurate.
[0083] In one possible implementation, the registration of the DR images and corresponding DRR images in two orthogonal directions to obtain 2D deformable meshes in the two orthogonal directions includes:
[0084] For each direction, the DR image and the corresponding DRR image are input as a pair into the VoxelMorph network (deformable medical image registration learning network) to obtain the output velocity field;
[0085] The velocity field is converted into a differential deformation field by scaling the integral layer using a differentiable sum of squares;
[0086] For each direction, extract deformation information from the differential deformation field;
[0087] A 2D deformation mesh is generated using the deformation information.
[0088] The working principle of the above technical solution is as follows:
[0089] As shown in Figure 3: Figure 3 is a schematic diagram of the VoxelMorph end-to-end unsupervised architecture; VoxelMorph is a framework that combines traditional methods and learning-based methods. It is based on a probabilistic generative model and derives an unsupervised learning-based inference algorithm. In Figure 3, the first part shows the input image to the network... ψ(m,f), the output represents the mean velocity field value μ z|x:y Sum of variance Σ z|x:y The approximate posterior probability parameters are obtained; the velocity field z is sampled and converted into a differential deformation field φ using a differentiable sum-of-squares scaling integral layer. z Finally, a spatial transformation is performed on m to obtain the transformed m.
[0090] In this embodiment, the VoxelMorph network is used to register the DR images and corresponding DRR images in two orthogonal directions to obtain 2D deformable meshes in two orthogonal directions. Specifically:
[0091] For each orthogonal direction, the DR image and the corresponding DRR image are input as a pair into the VoxelMorph network. The VoxelMorph network adopts the U-Net architecture and introduces a differential homeomorphic integration layer and a spatial transformation layer. The network processes the input image pair into a parameterized function, which outputs a velocity field. The velocity field represents the velocity and direction that each pixel or voxel in the image needs to move during the registration process. In the VoxelMorph architecture, the first output is the approximate posterior probability parameters of the mean and variance of the velocity field, which are used to sample the velocity field. The sampled velocity field is then transformed through a differentiable square sum scaling integration layer. This layer integrates the velocity field into a differential deformation field, which is the actual displacement of each pixel or voxel during the registration process. The differential deformation field represents the position that each pixel in the DR image needs to move to. The differential deformation field is then used in the spatial transformation layer, which transforms the input DR image to align it spatially with the DRR image. Deformation information is extracted from the differential deformation field, which describes the deformation process from the DR image to the DRR image.
[0092] For each orthogonal direction, the deformation information extracted from the differential deformation field is used to generate a 2D deformation mesh.
[0093] The grid represents the displacement of each pixel in the DR image when it is registered to the DRR image. Finally, for each orthogonal direction, a 2D deformable grid is obtained, which represents the accurate registration result from the DR image to the DRR image.
[0094] The effects of the above technical solution are as follows:
[0095] The VoxelMorph network combines traditional and learning-based methods, leveraging the powerful learning capabilities of convolutional neural networks (CNNs) to automatically extract image features and perform precise registration. By outputting a velocity field and converting it into a differential deformation field, it achieves fine-grained displacement control of each pixel or voxel in the image, thereby improving registration accuracy. Using DR and DRR images for registration reduces reliance on traditional CT scans, thus lowering the purchase and maintenance costs of medical equipment. Simultaneously, the simple and quick DR image acquisition process also helps reduce patient examination costs.
[0096] In one possible implementation, reconstructing the 3D deformation field through the 2D deformation mesh includes:
[0097] By utilizing the invertible relationship between the components of the 2D deformable mesh on the orthogonal plane and the 3D vector field, the nodal displacements in the 3D deformable field are calculated through two-dimensional to three-dimensional back projection and geometric transformation.
[0098] Based on the node displacements, a complete 3D deformation field is constructed.
[0099] The working principle of the above technical solution is as follows:
[0100] The two planes are orthogonal in the 3D patient coordinate system; because the two planes are orthogonal, it is possible to capture different views of the object in three-dimensional space, and through these two views, the three-dimensional shape and deformation of the object can be reconstructed.
[0101] Referring to Figure 4, which is a schematic diagram of an orthogonal imaging system according to an embodiment of this application;
[0102] Orthogonal imaging system with two X-ray sources (S A ,S B The system consists of two X-ray detectors (A, B), located on two orthogonal planes; and two 2D projected coordinate systems (x, y, y). A ,y A ) and (x B, y B The 3D patient coordinate system is (x, y, z); because (x A ,y A ) and (x B ,y B In the 3D patient coordinate system, the coordinates are orthogonal. The 3D deformed mesh can be reconstructed by using the orthogonal 2D deformed mesh. For example, each voxel in the vector field represents a displacement component. The component on the orthogonal plane and the projection directly onto the 3D vector field have an invertible relationship.
[0103] Referring to Figure 5, which is a schematic diagram of a surface in the 3D deformation field of the back projection of the deformation mesh component in the orthogonal plane of this application embodiment; the nodes of the projection and the center line of the middle mesh (deformation mesh) in Figure 5 define a plane H, i.e., a three-dimensional patient coordinate system; through two-dimensional to three-dimensional back projection and geometric transformation, the three-dimensional displacement of the nodes on the plane H is calculated.
[0104] First, the DR images and corresponding DRR images in two orthogonal directions are registered using the VoxelMorph network to obtain 2D deformable meshes in two orthogonal directions, as shown in Figure 6; Figure 6 is a schematic diagram of the 2D registration mesh after orthogonal DR and DRR registration in an embodiment of this application.
[0105] For each 2D deformable mesh, it is considered as a displacement field on a corresponding orthogonal plane, where each node represents a displacement vector. Using a 2D-to-3D backprojection technique, these displacement vectors are projected from the 2D plane into 3D space. During the backprojection process, geometric transformations are performed on the projected points to calculate their actual positions in 3D space. These geometric transformations include operations such as rotation, translation, and scaling. Through these geometric transformations, the displacement vector of each projected point in 3D space is calculated. These displacement vectors constitute the nodal displacements in the 3D deformable field. Based on the calculated nodal displacements, a complete 3D deformable field can be constructed.
[0106] The advantages of the above technical solution are as follows: By utilizing the invertible relationship between the 2D deformation mesh and the 3D vector field on two orthogonal planes, the nodal displacements in the 3D deformation field can be accurately calculated, thereby constructing a complete 3D deformation field and ensuring the accuracy of subsequent analysis and operations. Traditional deformation analysis is usually limited to a two-dimensional plane, while this method, by introducing an orthogonal imaging system and a three-dimensional reconstruction method, elevates deformation analysis to a three-dimensional level, enabling doctors and researchers to have a more comprehensive and intuitive understanding of the three-dimensional shape and deformation of objects (such as tumors, organs at risk, etc.), providing richer information for clinical diagnosis and treatment. Through two-dimensional to three-dimensional back projection and geometric transformation, the nodal displacements in the 3D deformation field can be calculated efficiently, avoiding complex mathematical models and calculation processes, which not only improves computational efficiency but also reduces computational costs.
[0107] In one possible implementation, applying the 3D deformation field to the delineation set to obtain deformation components of the patient's corresponding target area or organs at risk; and determining whether to re-acquire the patient's CT images based on the deformation components, includes:
[0108] If the deformation component exceeds its deformation threshold, the system will prompt the user to reacquire the patient's CT images.
[0109] If the deformation component is less than or equal to its deformation threshold, the patient's planned CT and delineation set are corrected according to the deformation component, and the treatment plan is regenerated using the original optimized parameters.
[0110] The working principle of the above technical solution is as follows: First, the acquired 3D deformation field is fused with the patient's delineation set; by applying the 3D deformation field to each structure in the delineation set, the new position and shape of these structures after deformation can be calculated; during the fusion process, the deformation components of each key structure (such as the target area and organs at risk) are extracted; the deformation components describe the displacement, rotation and deformation changes of these structures in 3D space; the extracted deformation components are evaluated to determine whether the deformation exceeds a preset threshold or standard; the threshold or standard can be set according to clinical experience and treatment requirements.
[0111] Based on the assessment results of the deformation component, a decision is made as to whether the patient's CT images need to be reacquired: if the deformation component is small and its impact on the treatment plan is within an acceptable range, then reacquiring CT images is not necessary; if the deformation component is large and has a significant impact on the treatment plan, or if the deformation exceeds the safe range, then reacquiring CT images is necessary to update the treatment plan and ensure the accuracy and safety of the treatment.
[0112] If CT images are reacquired, the delineation set is updated based on the reacquired CT images, and an adaptive treatment plan is regenerated. If it is determined that CT images do not need to be reacquired, the delineation set is deformed based on the current deformation field, the treatment plan is optimized, and treatment is performed according to the optimized treatment plan.
[0113] In one possible implementation, the deformation component includes changes in position, area, and perimeter of the patient's target area or organs at risk; if the deformation component exceeds its deformation threshold, a prompt is made to reacquire the patient's CT images; including:
[0114] The weighted average of the changes in location, area, and perimeter is used to obtain the weighted deformation.
[0115] If any deformation component exceeds its preset threshold, the system will prompt the user to reacquire the patient's CT images.
[0116] If all deformation components are within their preset threshold range, but the weighted deformation exceeds its preset threshold, then the system will prompt the user to reacquire the patient's CT images.
[0117] If any deformation component exceeds a preset threshold, or the weighted deformation exceeds its preset threshold, the system prompts the patient to reacquire CT images. This decision logic helps to promptly detect significant deformations that may occur in the patient's body, thereby avoiding potential risks during radiotherapy and ensuring the patient's safety.
[0118] Automated deformation component monitoring and decision-making logic can reduce manual intervention during radiotherapy and improve the efficiency of the radiotherapy process. Precise deformation component monitoring and decision-making logic can avoid unnecessary CT image acquisition, which helps reduce patients' radiation exposure and lower potential health risks.
[0119] In one possible implementation, the method further includes:
[0120] By using the patient's historical data, the system obtains multiple deformation components of the same target area or organs at risk based on time series, the amount of change in deformation components, and treatment records. Through machine learning, it predicts the deformation components at the time of the next treatment. If the deformation components exceed their deformation threshold, it prompts the system to reacquire the patient's CT images.
[0121] The working principle and effect of the above technical solution are as follows: Multiple deformation components of the same target area or organ at risk are extracted from the patient's historical data. These deformation components include changes in position, area, and perimeter. Simultaneously, the corresponding treatment records for each deformation component measurement are recorded, including key information such as radiotherapy dose and duration. The collected data undergoes preprocessing, including data cleaning, format conversion, and standardization, to ensure data quality and consistency.
[0122] By using time series analysis, deformation components of the same target area or organs at risk can be modeled over time. Through time series analysis, the deformation component values at a future point in time (i.e., at the next treatment time) can be predicted.
[0123] Based on preprocessed historical data and time-series analysis results, a machine learning model is trained. The model's goal is to predict the deformation components at the next treatment time based on historical deformation components and treatment records. The machine learning model can be a support vector machine, random forest, or neural network. After training, the model is used to predict the deformation components at the next treatment time. The predicted deformation component values are compared with a preset deformation threshold. If the predicted deformation component exceeds the threshold, significant deformation is considered to have occurred in the target area or at-risk organs. When the predicted deformation component exceeds the threshold, the patient's CT images are reacquired. During actual radiotherapy, the predictive accuracy of the machine learning model is evaluated based on the reacquired CT images and new deformation component measurements. If the model's predictive accuracy needs improvement, further optimization and adjustments can be made to enhance its predictive ability.
[0124] At the same time, new data can be added to the historical dataset to enrich the training samples and improve the model's generalization ability.
[0125] In summary, by integrating patients' historical data, time-series analysis, machine learning predictions, and the relationship between deformation components and decisions to reacquire CT images, accurate prediction and timely response to deformation components in the target area or organs at risk are achieved. This avoids unnecessary CT image acquisition, helps reduce patients' radiation exposure, lowers potential health risks, and improves the accuracy and safety of radiotherapy planning.
[0126] Referring to Figure 7, which is a schematic diagram of an adaptive planning method apparatus for orthogonal DR according to an embodiment of this application.
[0127] This application provides an adaptive planning device based on orthogonal DR, the device comprising:
[0128] The segmentation and delineation module is used to automatically segment the patient's planned CT images and obtain segmentation results; the segmentation results include a delineation set of the patient's target area or organs at risk;
[0129] The image acquisition module is used to acquire DR images of the patient in two orthogonal directions after positioning the patient; and to project the planned CT images of the patient into the space of the positioning coordinate system to obtain corresponding orthogonal DRR images.
[0130] The 2D deformation acquisition module is used to register the DR image and the corresponding DRR image in two orthogonal directions respectively to obtain the 2D deformation mesh in the two orthogonal directions.
[0131] The 3D deformation field acquisition module is used to reconstruct the 3D deformation field through the 2D deformation mesh;
[0132] The determination module is used to apply the 3D deformation field to the delineation set to obtain the deformation components of the target area or organs at risk of the patient; and to determine whether to reacquire the patient's CT images based on the deformation components.
[0133] In one possible implementation, the patient's planned CT images are automatically segmented using the 3DUNet automatic segmentation network.
[0134] In one possible implementation, the automatic segmentation further includes obtaining the loss function of the 3DUNet automatic segmentation network, wherein the loss function of the 3DUNet automatic segmentation network is obtained through the Dice loss and cross-information entropy loss between the ground truth label and the predicted segmentation.
[0135] The loss function of the 3DUNet automatic segmentation network is:
[0136] Among them, L Dice () represents the Dice loss function; L CE () represents the cross-information entropy loss function; Lloss () represents the total loss function; e represents the ground truth label; e′ represents the predicted segment; t represents the voxel index; N represents the total number of voxels; a g dn is the weight of the g-th segmentation category; w is the balancing coefficient used to balance the weights of the Dice loss function and the cross-entropy loss function; g is the g-th segmentation category; dn is the total number of segmentation categories.
[0137] In one possible implementation, the image acquisition module includes:
[0138] The DR image acquisition unit is used to acquire DR images of the patient using two orthogonal DR acquisition devices with the isocenter point of the treatment room as the origin.
[0139] The DDR image generation unit is used to project the patient's planned CT image along the directions of the two DR acquisition devices in the positioning coordinate system space to generate corresponding orthogonal DRR images.
[0140] In one possible implementation, the 2D deformation acquisition module includes:
[0141] The first output unit is used to input the DR image and the corresponding DRR image as a pair into the VoxelMorph network for each direction to obtain the output velocity field.
[0142] Differential deformation field element, used to convert velocity field into differential deformation field by scaling integral layer with differentiable sum of squares;
[0143] The deformation mesh acquisition unit is used to extract deformation information from the differential deformation field for each direction; and to generate a 2D deformation mesh using the deformation information.
[0144] In one possible implementation, the 3D deformation field acquisition module includes:
[0145] The nodal displacement acquisition unit is used to calculate the nodal displacement in the 3D deformation field by utilizing the invertible relationship between the components of the 2D deformed mesh on the orthogonal plane and the 3D vector field, through two-dimensional to three-dimensional back projection and geometric transformation.
[0146] The 3D deformation field reconstruction unit is used to construct a complete 3D deformation field based on the node displacement.
[0147] In one possible implementation, the determination module includes:
[0148] The first determination unit is used to re-acquire the patient's CT image if the deformation component exceeds its deformation threshold.
[0149] The second determination unit is used to correct the patient's planned CT and delineation set according to the deformation component if the deformation component is less than or equal to its deformation threshold, and to regenerate the treatment plan using the original optimized parameters.
[0150] In one possible implementation, the deformation components include changes in position, area, and perimeter of the patient's target area or organ at risk; the first determination unit includes:
[0151] The weighted deformation acquisition unit is used to weight the position change, area change and perimeter change to obtain the weighted deformation.
[0152] The first determination subunit is used to prompt the patient's CT image to be reacquired if any deformation component exceeds its preset threshold.
[0153] The second determination subunit is used to prompt the patient to reacquire CT images if each deformation component is within its preset threshold range, but the weighted deformation exceeds its preset threshold.
[0154] In one possible implementation, the device further includes a prediction module for acquiring, based on the patient's historical data, multiple deformation components of the same target area or organ at risk over time, the amount of change in the deformation components, and treatment records, and predicting the deformation components at the next treatment through machine learning; if the deformation components exceed their deformation threshold, the device prompts the patient to reacquire CT images.
[0155] The working principle and effect of the above technical solution are the same as those in the method embodiments of this application, and will not be repeated here.
[0156] This application provides a treatment system, the system comprising:
[0157] A radiation therapy device used to administer radiation therapy to a patient according to a treatment plan;
[0158] The adaptive planning device based on orthogonal DR described in any of the embodiments of this application is used to obtain an adaptive treatment plan for a patient.
[0159] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in the foregoing embodiments of this application or the function of the device described in the embodiments of this application.
[0160] This application also provides a computer-readable storage medium for storing a computer program. When the computer program is executed, it implements the steps of the method in this application. The specific implementation method is consistent with the implementation method and the technical effect achieved in the above method embodiments, and some contents will not be repeated.
[0161] In this application, a readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The program product can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.
[0162] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or any suitable combination thereof. Program code for performing the operations of this application may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as C or similar programming languages. The program code may be executed entirely on a user computing device, partially on an associated device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).
Claims
1. An adaptive planning method based on orthogonal DR, the method comprising: Automatic segmentation of the patient's planned CT images is performed to obtain segmentation results; The segmentation results include a delineation set of the patient's target area or organs at risk; After positioning the patient, DR images of the patient in two orthogonal directions are acquired; and the planned CT images of the patient are projected in the space of the positioning coordinate system to obtain the corresponding orthogonal DRR images. The DR images in the two orthogonal directions are registered with the corresponding DRR images to obtain 2D deformable meshes in the two orthogonal directions. The 3D deformation field is reconstructed using the 2D deformation mesh; The 3D deformation field is applied to the delineation set to obtain the deformation components of the patient's corresponding target area or organs at risk; based on the deformation components, it is determined whether to reacquire the patient's CT images.
2. The orthogonal DR based adaptive planning method of claim 1, wherein, The planned CT images of the patient were automatically segmented using the 3DUNet automatic segmentation network.
3. The orthogonal DR based adaptive planning method of claim 2, wherein, The automatic segmentation also includes obtaining the loss function of the 3DUNet automatic segmentation network, wherein the loss function is determined by the Dice loss and cross-information entropy loss between the ground truth label and the predicted segmentation; The loss function of the 3DUNet automatic segmentation network is: Among them, L Dice L is the Dice loss function; CE L is the cross-information entropy loss function; loss Let be the total loss function; e be the true label; e ′ For predicting segmentation; t represents the voxel index; N is the total number of voxels; a g dn is the weight of the g-th segmentation category; w is the balancing coefficient used to balance the weights of the Dice loss function and the cross-entropy loss function; g is the g-th segmentation category; dn is the total number of segmentation categories.
4. The orthogonal DR based adaptive planning method of claim 1, wherein, After positioning the patient, DR images of the patient are acquired in two orthogonal directions; The planned CT images of the patient are projected into the space of the positioning coordinate system to obtain corresponding orthogonal DRR images; including: The patient's DR images are acquired using two orthogonal DR acquisition devices; In the positioning coordinate system space, the patient's planned CT images are projected along the directions of the two DR acquisition devices to generate corresponding orthogonal DRR images.
5. The orthogonal DR based adaptive planning method of claim 1, wherein, The step of registering the DR images and corresponding DRR images in two orthogonal directions to obtain 2D deformable meshes in two orthogonal directions includes: For each direction, the DR image and the corresponding DRR image are input as a pair into the VoxelMorph network to obtain the output velocity field; The velocity field is converted into a differential deformation field by scaling the integral layer using a differentiable sum of squares; For each direction, extract deformation information from the differential deformation field; A 2D deformation mesh is generated using the deformation information.
6. The orthogonal DR based adaptive planning method of claim 1, wherein, The reconstruction of the 3D deformation field through the 2D deformation mesh includes: By utilizing the invertible relationship between the components of the 2D deformable mesh on the orthogonal plane and the 3D vector field, the nodal displacements in the 3D deformable field are calculated through two-dimensional to three-dimensional back projection and geometric transformation. Based on the node displacements, a complete 3D deformation field is constructed.
7. The orthogonal DR based adaptive planning method of claim 1, wherein, The 3D deformation field is applied to the delineation set to obtain the deformation components of the patient's corresponding target area or organs at risk; Determining whether to reacquire the patient's CT images based on the deformation components includes: If the deformation component exceeds its deformation threshold, the patient's CT images are reacquired; If the deformation component is less than or equal to its deformation threshold, the patient's planned CT and delineation set are corrected according to the deformation component, and the treatment plan is regenerated using the original optimized parameters.
8. An adaptive planning device based on orthogonal DR, the device comprising: The segmentation and delineation module is used to automatically segment the patient's planned CT images and obtain segmentation results; The segmentation results include a delineation set of the patient's target area or organs at risk; The image acquisition module is used to acquire DR images of the patient in two orthogonal directions after positioning the patient; and to project the planned CT images of the patient into the space of the positioning coordinate system to obtain corresponding orthogonal DRR images. The 2D deformation acquisition module is used to register the DR image and the corresponding DRR image in two orthogonal directions respectively to obtain the 2D deformation mesh in the two orthogonal directions; The 3D deformation field acquisition module is used to reconstruct the 3D deformation field through the 2D deformation mesh; The determination module is used to apply the 3D deformation field to the delineation set to obtain the deformation components of the target area or organs at risk of the patient; and to determine whether to reacquire the patient's CT images based on the deformation components.
9. A treatment system, the system comprising: A radiation therapy device used to administer radiation therapy to a patient according to a treatment plan; The adaptive planning device based on orthogonal DR as described in claim 8 is used to obtain an adaptive treatment plan for a patient.
10. An electronic device comprising a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method according to any one of claims 1-7.
11. A computer-readable storage medium storing computer instructions that, when read by a computer, perform the steps of the method as described in any one of claims 1-7.