Physical constraint-combined VMAT radiotherapy plan generation method
By generating VMAT radiotherapy plans using a physical constraint-based multi-task prediction model, the problem of existing technologies failing to effectively consider physical constraints is solved, resulting in more accurate and executable radiotherapy plans.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-27
- Publication Date
- 2026-04-14
AI Technical Summary
Existing VMAT radiotherapy planning methods fail to effectively consider physical constraints, resulting in plans that are unexecutable on real machines or require additional modifications.
A multi-task prediction model trained based on physical constraints is adopted to predict the machine control parameters of VMAT radiotherapy plans by acquiring the patient's three-dimensional CT images and masks. The control parameters are optimized by combining dose prediction and physical constraint loss function to generate more reasonable and accurate radiotherapy plans.
The generated radiotherapy plans are more accurate and can be executed directly without additional corrections, thus improving the accuracy and clinical acceptability of radiotherapy plans.
Smart Images
Figure CN121846548A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical data processing technology, and in particular to a VMAT radiotherapy planning generation method that incorporates physical constraints. Background Technology
[0002] Volumetric Modulated Arc Therapy (VMAT) is a high-precision radiotherapy technique that has rapidly developed and been widely applied in clinical practice in recent years. Compared with traditional static intensity-modulated radiotherapy (IMRT), VMAT significantly shortens treatment time, reduces head leakage and scattered dose, and improves patient comfort by continuously rotating the gantry and simultaneously and dynamically adjusting the position of the multi-leaf collimator (MLC) blades, dose rate, and gantry speed during treatment, while ensuring target dose coverage and homogeneity. The quality of VMAT planning directly depends on the optimization results of machine control parameters (including MLC blade sequence, gantry angle, dose rate, collimator angle, etc.) at each control point (CP). Therefore, the core task of the planning system is to quickly search for the optimal combination of parameters that meets clinical requirements within the feasible solution space.
[0003] In recent years, artificial intelligence technologies such as deep learning and reinforcement learning have been explored for end-to-end prediction of three-dimensional dose distributions or direct output of control parameter sequences. Data-driven methods have achieved dosimetric indices comparable to manual planning on test sets, with inference times reduced to the second level, demonstrating great potential.
[0004] However, existing neural network models generally use minimizing dosimetric error or blade position error as the sole training objective, lacking explicit modeling of linear accelerator hardware and physical constraints. This results in the generated VMAT plans often being unexecutable on real machines or requiring additional post-processing corrections. Summary of the Invention
[0005] Based on the above analysis, the present invention aims to provide a VMAT radiotherapy plan generation method that incorporates physical constraints, in order to solve the problem that existing radiotherapy plans that do not consider physical constraints are unexecutable or require additional modifications.
[0006] On one hand, embodiments of the present invention provide a VMAT radiotherapy planning generation method incorporating physical constraints, comprising the following steps: Obtain the 3D CT image of the patient to be generated, as well as the corresponding target area mask and organ at risk mask; The three-dimensional CT image and the corresponding target area mask and organ at risk mask are input into the control parameter prediction model to predict the machine control parameters corresponding to each control point of the VMAT radiotherapy plan, so as to obtain the VMAT radiotherapy plan for the patient to be generated. The control parameter prediction model is a multi-task prediction model trained based on physical constraints.
[0007] Based on the further improvement of the above method, the training loss of the control parameter prediction model is calculated using the following formula: ; in, Indicates the loss of control parameters. Indicates dose prediction loss, Represents physical constraint loss. Indicates consistency loss. , , and This represents the weighting coefficient.
[0008] Based on the further improvement of the above method, the physical constraint loss is calculated using the following formula:
[0009] in, This indicates the blade velocity constraint loss. This indicates the blade clearance constraint loss. This indicates the loss due to blade position constraints. , and This represents the weighting coefficient.
[0010] Based on the further improvement of the above method, the blade velocity constraint loss is calculated in the following way: ; ; in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c+1-th control point. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c+1-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the maximum allowed variation.
[0011] Based on the further improvement of the above method, the blade clearance constraint loss is calculated using the following formula: ; in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. Indicates the steepness parameter. This indicates the minimum gap.
[0012] Based on the further improvement of the above method, the blade position constraint loss is calculated using the following formula: ; in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. Minimum range of leaf position Maximum range of blade position Minimum range of leaf position Maximum range of blade position.
[0013] Based on further improvements to the above method, the dose prediction loss is calculated using the following formula: ; ; ; in, This represents the predicted three-dimensional dose distribution map. This represents the gold standard for three-dimensional dose distribution maps. Indicates DVH loss, and Indicates the weighting coefficient. Denotes the 2-norm of a matrix. Indicates the DVH loss in the target area. This represents the DVH loss of the s-th organ at risk, where S represents the number of organs at risk. This represents the number of DVH critical points for the s-th organ at risk. This represents the target dose value at the k-th critical point of the s-th organ at risk of disease (DVH). This represents the dose value predicted by the model for the k-th critical point of the s-th organ at risk of disease (DVH). Indicates the number of dose intervals. This represents the number of voxels in the s-th organ at risk. This represents the set of voxels representing the s-th organ at risk. This represents the predicted dose value of the i-th voxel for the s-th organ at risk. This represents the boundary value of the b-th dose interval. Indicates the steepness parameter. This represents the sigmoid function. Indicates the s-th organ at risk The target volume fraction at that location.
[0014] Based on further improvements to the above method, the multi-task prediction model includes: The input layer is used to map the input 3D CT image and the corresponding target area mask and organ at risk mask to obtain the initial features; Multi-stage coding unit, used to encode initial features in multiple stages based on an attention mechanism; The bottleneck layer is used to globally model the features output by the coding unit in the last stage, thus obtaining the globally modeled features. The dose distribution prediction branch is used to perform multi-stage decoding of global modeling features and predict the three-dimensional dose distribution map. The control parameter prediction branch is used to predict control parameters based on global modeling features.
[0015] Based on the further improvement of the above method, each coding unit includes a residual convolutional block, a local-global attention module and a downsampling layer connected in sequence; The residual convolutional block is used to extract convolutional features to obtain convolutional features; The local-global attention module is used to capture long-range dependencies of convolutional features to obtain spatial attention features; The downsampling layer is used to downsample spatial attention features.
[0016] Based on a further improvement of the above method, the local-global attention module includes: The feature map flattening unit is used to flatten the input features to obtain a feature sequence; Position encoding unit, used to add position encoding to each position of the feature sequence; Multi-head self-attention unit, used to perform attention calculation on feature sequence using multi-head attention mechanism to obtain attention feature sequence; Feedforward units are used to reorganize the attention feature sequences; The sequence rearrangement unit is used to restore the recombined feature sequence to the feature dimension input by the local-global attention module, thus obtaining spatial attention features.
[0017] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. By training the model based on physical constraints, more reasonable, accurate, and executable radiotherapy plans are generated without the need for additional corrections, thereby improving the accuracy and clinical acceptability of radiotherapy plans.
[0018] 2. By constructing a multi-task prediction model and using dose prediction task to assist control parameter prediction task, the generated radiotherapy plan can be closer to the clinically expected dose.
[0019] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart illustrating a specific embodiment of the VMAT radiotherapy planning generation method incorporating physical constraints, according to a specific embodiment of the present invention. Detailed Implementation
[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0022] A specific embodiment of the present invention discloses a VMAT radiotherapy planning generation method incorporating physical constraints, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the three-dimensional CT image of the patient to be generated, as well as the corresponding target area mask and organ at risk mask; S2. Input the three-dimensional CT image and the corresponding target area mask and organ at risk mask into the control parameter prediction model to predict the machine control parameters corresponding to each control point of the VMAT radiotherapy plan, and obtain the VMAT radiotherapy plan for the patient to be generated. The control parameter prediction model is a multi-task prediction model trained based on physical constraints.
[0023] It should be noted that the VMAT radiotherapy plan includes parameters such as gantry angle, LMC blade position, and dose rate (MU value) for each control point.
[0024] Compared with existing technologies, the VMAT radiotherapy plan generation method combined with physical constraints provided in this embodiment predicts the machine control parameters corresponding to each control point of the VMAT radiotherapy plan by inputting the patient's three-dimensional CT image and the corresponding target area mask and organ at risk mask into a multi-task prediction model trained based on physical constraints. This yields the VMAT radiotherapy plan for the patient to be generated, thus taking physical constraints into account and generating a more reasonable and accurate radiotherapy plan without the need for additional corrections, thereby improving the accuracy and clinical acceptability of the radiotherapy plan.
[0025] Specifically, the constructed multi-task prediction model includes: The input layer is used to map the input 3D CT image and the corresponding target area mask and organ at risk mask to obtain the initial features; Multi-stage coding unit, used to encode initial features in multiple stages based on an attention mechanism; The bottleneck layer is used to globally model the features output by the coding unit in the last stage, thus obtaining the globally modeled features. The dose distribution prediction branch is used to perform multi-stage decoding of global modeling features and predict the three-dimensional dose distribution map. The machine parameter prediction branch is used to predict control parameters based on global modeling features.
[0026] In practice, the multi-task prediction model adopts a symmetrical encoder-decoder structure.
[0027] During implementation, the input layer receives data including 3D CT images, corresponding target area masks, and organ-at-risk masks. The input data tensor is represented as follows: ,in This indicates the number of input channels, including CT image (one channel) + target area mask (one channel) + masks for each organ at risk (one channel per structure). D, H, and W represent depth, height, and width, respectively.
[0028] In practice, the input layer uses 3D convolutional layers to increase the number of input channels. Mapping the initial feature dimensions yields the initial features. ,in, This represents the initial feature dimension, typically 32 or 64. The kernel size of a 3D convolutional layer is... Step size 1, fill size 1.
[0029] Then, multi-stage feature extraction is performed through multi-stage coding units. In practice, L stages of coding units can be used, usually L=4 or 5.
[0030] Specifically, each coding unit includes a residual convolutional block, a local-global attention module, and a downsampling layer connected in sequence; The residual convolutional block is used to extract convolutional features to obtain convolutional features; The local-global attention module is used to capture long-range dependencies of convolutional features to obtain spatial attention features; The downsampling layer is used to downsample spatial attention features.
[0031] For the Each encoding unit, the residual convolutional block is represented as: ; in, This represents residual convolution, consisting of two convolutional layers, instance normalization, and ReLU activation, with residual connections. Indicates the first Features of the residual convolution block output of each coding unit Indicates the first The features output by each coding unit, if ,but This represents the initial features output from the input layer of the first deep learning model. Indicates the first The feature dimensions output by the residual convolutional block of each coding unit. They represent the first The depth, height, and width of the feature map output by the residual convolutional block of each coding unit.
[0032] Then, long-range dependencies are captured through a local-global attention module. Specifically, the local-global attention module includes: The feature map flattening unit is used to flatten the input features to obtain a feature sequence; Position encoding unit, used to add position encoding to each position of the feature sequence; Multi-head self-attention unit, used to perform attention calculation on feature sequence using multi-head attention mechanism to obtain attention feature sequence; Feedforward units are used to reorganize the attention feature sequences; The sequence rearrangement unit is used to restore the recombined feature sequence to the feature dimension input by the local-global attention module, thus obtaining spatial attention features.
[0033] During implementation, the feature map flattening unit will flatten the 3D feature map. Expand into feature sequences : , .
[0034] The positional encoding unit adds learnable 3D positional encoding to the feature sequence: ; Among them, the Each coding unit corresponds to This represents learnable 3D positional encoding.
[0035] Multi-head self-attention units use a multi-head attention mechanism to perform attention calculations on the feature sequence to obtain the attention feature sequence. : ; ; in, This indicates splicing, where h represents the number of attention heads. This represents the i-th attention head. Indicates the output projection matrix. , and Let these represent the query projection matrix, key projection matrix, and value projection matrix of the i-th attention head, respectively. This indicates the calculation of self-attention.
[0036] The feedforward unit uses the following method to process the attention feature sequence. Reorganization: ; ; in, It is a two-layer fully connected network. This indicates the characteristics of the feedforward unit output. Representation layer normalization.
[0037] Sequence rearrangement unit pairs Perform dimensional restoration and rearrangement as follows The spatial attention features are obtained from the dimensional form. .
[0038] The downsampling layer halves the resolution of spatial attention features through strided convolutions, thus achieving downsampling. The kernel size of the 3D convolutional layer is Step size 2. Indicates the first Features output by each coding unit.
[0039] In implementation, the bottleneck layer can be constructed by stacking M consecutive Transformer blocks, and the features output by the last coding unit... Perform global context modeling to obtain global features. .
[0040] In practice, the multi-task prediction model of this invention includes a three-dimensional dose prediction task and a control parameter prediction task.
[0041] The two tasks of three-dimensional dose prediction and control parameter prediction complement each other. Three-dimensional dose prediction serves as an auxiliary task, making the predicted control parameters more accurate and the generated radiotherapy plan closer to the expected clinical dose distribution requirements.
[0042] Specifically, the dose distribution prediction branch performs multi-stage decoding of global modeling features to predict a three-dimensional dose distribution map.
[0043] In practice, the dose distribution prediction branch also contains L stages of decoding units.
[0044] Each decoding unit includes an upsampling layer, skip connections, convolutional blocks, and cross-attention modules.
[0045] The upsampling layer uses 3D transposed convolution to upsample the input features.
[0046] By using gated attention, the output features of the encoding units at the corresponding stage are weighted and fused with the upsampled features to obtain the fused features. This allows for the incorporation of detailed encoder features during decoding, thereby improving decoding capabilities.
[0047] Then, further feature extraction is performed through convolutional blocks to obtain the convolutional features of the decoding unit.
[0048] Finally, a cross-attention mechanism is employed to perform cross-attention calculations between the convolutional features of the decoding unit and the global features output by the bottleneck layer. The cross-attention query originates from the convolutional features of the decoding unit, while the key and value are derived from the global features output by the bottleneck layer. This cross-attention mechanism further enhances decoding capabilities and improves prediction accuracy.
[0049] The output layer projects the decoded features output by the decoding unit in the last stage through 3D convolution to generate an image, thus obtaining the predicted three-dimensional dose map.
[0050] The control parameter prediction branch can use an RNN structure to generate control parameters for each control point based on global modeling features.
[0051] During implementation, 3D CT images of multiple historical patients, along with corresponding target area masks and organ-at-risk masks, are collected as input data for the samples. The corresponding 3D dose maps and control parameter sequences serve as labels to construct a sample set. Based on this constructed sample set, a multi-task prediction model is trained using physical constraints to obtain a control parameter prediction model.
[0052] Specifically, the training loss of the control parameter prediction model is calculated using the following formula: ; in, Indicates the loss of control parameters. Indicates dose prediction loss, Represents physical constraint loss. Indicates consistency loss. , , and These represent weighting coefficients. During implementation, the weighting coefficients are set according to different objectives; for example, if physical constraints are more important, the weights can be set larger. The sum of the four weighting coefficients is one.
[0053] Specifically, the physical constraint loss is calculated using the following formula:
[0054] in, This indicates the blade velocity constraint loss. This indicates the blade clearance constraint loss. This indicates the loss due to blade position constraints. , and This represents the weighting coefficient, which sums to 1. The coefficient value can be determined based on the importance of each loss.
[0055] The speed at which the blades move between consecutive control points cannot exceed the maximum permissible speed of the machine. Assume the maximum permissible speed for blade position change between adjacent control points is... The time interval is assumed here. (Assuming equal time intervals for each control point), the maximum permissible location change is... For each pair of blades (including the left and right blades), the position of change should be less than [a certain value]. .
[0056] Therefore, the blade velocity constraint loss is calculated using the following method: ; ; in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c+1-th control point. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c+1-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This represents the maximum allowable positional change, where x represents the blade positional change.
[0057] During implementation, Penalize speeds that exceed a threshold.
[0058] To prevent blade collisions, a minimum clearance must be maintained between the left and right blades. Meanwhile, in order to form an effective opening, the left blade position should be smaller than the right blade position; therefore, the blade clearance loss penalty clearance is smaller than... The situation.
[0059] To make the penalty smoother, the blade clearance constraint loss is calculated using the following formula: ; in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. Indicates the steepness parameter. This indicates the minimum gap.
[0060] During implementation, steepness parameters An experience value of 5-10 is recommended.
[0061] At the same time, the blades must move within physically permissible limits; the range of movement for the left blade is […]. The position range of the right blade. The blade position constraint loss is calculated using the following formula: ; in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. Minimum range of leaf position Maximum range of blade position Minimum range of leaf position Maximum range of blade position.
[0062] In practice, the difference between the predicted control parameter sequence and the labeled control parameter sequence can be calculated using existing loss calculation functions, such as the MSE loss function and the cross-entropy loss function.
[0063] Consistency loss during implementation This refers to the dose distribution map calculated based on predicted control parameters. Consistency between the dose distribution map and the model prediction. During implementation, existing functions such as MSE loss and cross-entropy loss can be used for calculation.
[0064] Specifically, the dose prediction loss is calculated using the following formula: ; ; ; in, This represents the predicted three-dimensional dose distribution map. This represents the gold standard for three-dimensional dose distribution maps. Indicates DVH loss, and Indicates the weighting coefficient. Denotes the 2-norm of a matrix. Indicates the DVH loss in the target area. This represents the DVH loss of the s-th organ at risk, where S represents the number of organs at risk. This represents the number of DVH critical points for the s-th organ at risk. This represents the target dose value at the k-th critical point of the s-th organ at risk of disease (DVH). This represents the dose value predicted by the model for the k-th critical point of the s-th organ at risk of disease (DVH). Indicates the number of dose intervals. This represents the number of voxels in the s-th organ at risk. This represents the set of voxels representing the s-th organ at risk. This represents the predicted dose value of the i-th voxel for the s-th organ at risk. This represents the boundary value of the b-th dose interval. Indicates the steepness parameter. This represents the sigmoid function. Indicates the s-th organ at risk The target volume fraction at that location.
[0065] DVH loss consists of DVH loss in the target area and DVH loss in organs at risk.
[0066] For each organ at risk of DVH loss, the difference at each key clinical endpoint was calculated. and DVH curve shape loss .
[0067] During implementation, key points are provided by clinicians, such as D5 (representing the dose at 5% volume) and Dmax (maximum dose).
[0068] During implementation, if the dose range is divided into B intervals, then there are B+1 dose boundaries. }, This represents the boundary value of the b-th dose interval. At each dose boundary... Calculate the predicted volume fraction (i.e., dosage greater than or equal to) (volume fraction) and target volume fraction The difference lies in the steepness of the approximate step function. A larger α results in a closer approximation to the step function, but the gradient may also be larger, requiring attention to the gradient explosion problem during training. α is typically set between 10 and 100.
[0069] The formula for calculating target volume dose loss is the same as that for calculating organ-at-risk dose loss, except that the definition of key points may differ. For example, key points include D95 (representing the dose at 95% volume), D50 (representing the dose at 5% volume), and D5 (representing the dose at 5% volume).
[0070] According to training Update the model parameters. When the loss converges and the model performance (such as accuracy, recall, F1 score, etc.) reaches the preset indicators, stop training and obtain the trained multi-task prediction model, i.e., the control parameter prediction model.
[0071] The three-dimensional CT image of the patient to be generated, along with the corresponding target area mask and organ at risk mask, is input into the control parameter prediction model to predict the machine control parameters corresponding to each control point of the VMAT radiotherapy plan, thereby obtaining an executable VMAT radiotherapy plan that is closer to the clinically desired dose.
[0072] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0073] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A VMAT radiotherapy planning generation method incorporating physical constraints, characterized in that, Includes the following steps: Obtain the 3D CT image of the patient to be generated, as well as the corresponding target area mask and organ at risk mask; The three-dimensional CT image and the corresponding target area mask and organ at risk mask are input into the control parameter prediction model to predict the machine control parameters corresponding to each control point of the VMAT radiotherapy plan, so as to obtain the VMAT radiotherapy plan for the patient to be generated. The control parameter prediction model is a multi-task prediction model trained based on physical constraints.
2. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 1, characterized in that, The training loss of the control parameter prediction model is calculated using the following formula: ； in, Indicates the loss of control parameters. Indicates dose prediction loss, Represents physical constraint loss. Indicates consistency loss. , , and This represents the weighting coefficient.
3. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 2, characterized in that, The physical constraint loss is calculated using the following formula: in, This indicates the blade velocity constraint loss. This indicates the blade clearance constraint loss. This indicates the loss due to blade position constraints. , and This represents the weighting coefficient.
4. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 3, characterized in that, The blade velocity constraint loss is calculated using the following method: ； ； in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c+1-th control point. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c+1-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the maximum allowed variation.
5. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 3, characterized in that, The blade clearance constraint loss is calculated using the following formula: ； in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. Indicates the steepness parameter. This indicates the minimum gap.
6. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 3, characterized in that, The blade position constraint loss is calculated using the following formula: ； in, Indicates the number of control points. Indicates the number of blade pairs. This indicates the position of the left leaf in the i-th pair of leaves corresponding to the predicted c-th control point. This indicates the position of the right leaf in the i-th pair of leaves corresponding to the predicted c-th control point. Minimum range of leaf position Maximum range of leaf position. Minimum range of leaf position Maximum range of blade position.
7. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 2, characterized in that, The dose prediction loss is calculated using the following formula: ； ； ； in, This represents the predicted three-dimensional dose distribution map. This represents the gold standard for three-dimensional dose distribution maps. Indicates DVH loss, and Indicates the weighting coefficient. Denotes the 2-norm of a matrix. Indicates the DVH loss in the target area. This represents the DVH loss of the s-th organ at risk, where S represents the number of organs at risk. This represents the number of DVH critical points for the s-th organ at risk. This represents the target dose value at the k-th critical point of the s-th organ at risk of disease (DVH). This represents the dose value predicted by the model for the k-th critical point of the s-th organ at risk of disease (DVH). Indicates the number of dose intervals. This represents the number of voxels in the s-th organ at risk. This represents the set of voxels representing the s-th organ at risk. This represents the predicted dose value of the i-th voxel for the s-th organ at risk. This represents the boundary value of the b-th dose interval. Indicates the steepness parameter. This represents the sigmoid function. Indicates the s-th organ at risk The target volume fraction at that location.
8. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 1, characterized in that, The multi-task prediction model includes: The input layer is used to map the input 3D CT image and the corresponding target area mask and organ at risk mask to obtain the initial features; Multi-stage coding unit, used to encode initial features in multiple stages based on an attention mechanism; The bottleneck layer is used to globally model the features output by the coding unit in the last stage, thus obtaining the globally modeled features. The dose distribution prediction branch is used to perform multi-stage decoding of global modeling features and predict the three-dimensional dose distribution map. The control parameter prediction branch is used to predict control parameters based on global modeling features.
9. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 8, characterized in that, Each coding unit consists of a residual convolutional block, a local-global attention module, and a downsampling layer connected in sequence; The residual convolutional block is used to extract convolutional features to obtain convolutional features; The local-global attention module is used to capture long-range dependencies of convolutional features to obtain spatial attention features; The downsampling layer is used to downsample spatial attention features.
10. The VMAT radiotherapy planning generation method incorporating physical constraints according to claim 9, characterized in that, The local-global attention module includes: The feature map flattening unit is used to flatten the input features to obtain a feature sequence; Position encoding unit, used to add position encoding to each position in the feature sequence; Multi-head self-attention unit, used to perform attention calculation on feature sequence using multi-head attention mechanism to obtain attention feature sequence; Feedforward units are used to reorganize the attention feature sequences; The sequence rearrangement unit is used to restore the recombined feature sequence to the feature dimension input by the local-global attention module, thus obtaining spatial attention features.