Radiotherapy dose calculation optimization method and device based on generative adversarial network

Abstract

The invention discloses a radiotherapy dose calculation optimization method and device based on a generative adversarial network, and relates to the technical field of medical plans, and the method comprises the steps: obtaining phase space files and condition vectors of different accelerators, and making a data set; training and verifying the constructed generative adversarial network model by using the data set; determining the value of each parameter in the condition vector according to the actual operation parameter of the specific accelerator, inputting the value to the verified generative adversarial network model, and generating particle distribution data corresponding to the condition vector; and taking particle distribution data generated by the model as initial input of a Monte Carlo (MC) simulation method to obtain a radiotherapy dose distribution result simulated by the MC. Radiotherapy dose calculation is realized through a method of combining the generative adversarial network and MC simulation, and the duration of MC simulation and quality inspection processes can be greatly shortened while the dose calculation precision is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of medical planning, and in particular to a method and device for optimizing radiotherapy dose calculation based on a generative adversarial network. Background Art

[0002] Electron accelerators are important equipment for radiotherapy. Common accelerators that generate photons by electron targeting are devices that generate high-energy X / γ rays by bombarding high-atomic-number targets (such as tungsten, tantalum, etc.) with high-energy electron beams. Accurate simulation of the accelerator beam is a prerequisite for radiotherapy dose calculation. The dose calculation method is generally implemented using Monte Carlo (MC) simulation. MC simulation is a numerical calculation method based on random sampling and statistics, and is widely used in radiation transmission and dose calculation. This method accurately calculates key parameters such as energy deposition and dose distribution by simulating the random motion trajectories of a large number of particles. Due to its high precision, MC simulation is regarded as the gold standard for radiotherapy dose verification, but its computational cost is high and often requires a long running time for the following reasons:

[0003] MC-based radiotherapy dose calculations are typically divided into two phases. The first phase involves accelerator head simulation: simulating the particle transport process within the accelerator and storing the beam information as a phase space file (Phase Space File). This phase space file can be reused in subsequent radiotherapy dose calculation simulations. The second phase involves external accelerator simulation: based on the phase space file, the interactions of particles within the patient's body after leaving the accelerator head are simulated. During the accelerator head simulation phase, high-energy electrons undergo multiple Coulomb scattering, Bremsstrahlung, and positron generation within the target. Each secondary particle can trigger Compton scattering, electron pair effects, and other factors, forming an exponentially growing particle shower. Simultaneously, to ensure statistical accuracy, reconstructing the energy spectrum distribution and spatial dose distribution requires simulating hundreds of millions of particles. Therefore, generating the phase space file can take days or even a week, seriously impacting the progress of radiotherapy dose calculations. Summary of the Invention

[0004] To address these challenges and meet these technical requirements, the inventors have proposed a method and device for optimizing radiotherapy dose calculation based on a generative adversarial network (GAN). This method uses a GAN-based MC simulation and combines GAN with MC simulation to achieve radiotherapy dose calculation for dose optimization and quality assurance (QA). The technical solutions of this invention are as follows:

[0005] In a first aspect, the present application provides a method for optimizing radiotherapy dose calculation based on a generative adversarial network, comprising the following steps:

[0006] Obtain phase space files and condition vectors of different accelerators and make them into data sets;

[0007] Use the dataset to train and verify the constructed generative adversarial network model;

[0008] Determine the values of each parameter in the condition vector based on the actual operating parameters of a specific accelerator, and input them into a verified generative adversarial network model to generate particle distribution data corresponding to the condition vector;

[0009] The particle distribution data generated by the model is used as the initial input of the MC simulation method to obtain the radiotherapy dose distribution results simulated by MC.

[0010] A further technical solution is to obtain phase space files and condition vectors of different accelerators, including:

[0011] For medical linear accelerators of different models and parameters, corresponding condition vectors are constructed, including particle source position variance, initial energy mean, accelerator parameters, and measured physical data; wherein the accelerator parameters include at least the accelerator model code, handpiece angle, and irradiation field size, and the physical data include at least the percentage depth dose, off-axis ratio curve, and tissue phantom ratio;

[0012] Based on the parameters in the condition vector, the MC simulation method is used to simulate and calculate the phase space file of medical linear accelerators of different models and parameters;

[0013] During the simulation process, the spatial coordinates, motion direction, energy value and statistical weight of each particle are recorded and stored in the phase space file.

[0014] Its further technical solution is to build a generative adversarial network model, including:

[0015] A generative adversarial network model is built based on a deep learning framework, where both the generator and the discriminator adopt a multi-layer perceptron structure;

[0016] The input of the generator is a random noise vector and the condition vector, which are gradually mapped through a fully connected layer and an activation function to output simulated particle distribution data with the same dimension as the real particle distribution in the phase space file;

[0017] The discriminator receives real or simulated particle distribution data and the condition vector, extracts data features through a multi-layer neural network and performs binary classification to determine the authenticity of the input data.

[0018] A further technical solution is to use the dataset to train the constructed generative adversarial network model, including:

[0019] Input the training parts divided from the dataset into batches into the generative adversarial network model, set the batch size, and determine the number of iterations per training cycle based on the size of the training parts and the convergence speed requirements;

[0020] During the training process, the generator and the discriminator are optimized alternately, and the discriminator parameters are first updated several times iteratively, and then the generator parameters are updated until the generative adversarial network model reaches a balanced state.

[0021] Its further technical solution is that the generative adversarial network model reaches a balanced state when:

[0022] When the loss curves of the generator and the discriminator tend to be stable, the simulated samples and the real samples are highly similar in the key features of the particle distribution, and the fluctuation amplitude of the model performance indicators in multiple consecutive training cycles is less than the preset threshold, the generative adversarial network model is considered to have reached a balanced state.

[0023] Its further technical solution is that the method further comprises:

[0024] For the specific accelerator, comparing its MC simulation results with measurement results under the same conditions to verify the MC simulation results;

[0025] If the deviation exceeds the allowable range, the values of the parameters in the condition vector input to the model are adjusted, the particle distribution data are regenerated, and the updated MC simulation results are obtained;

[0026] Compare again until the deviation is within the allowable range.

[0027] In a second aspect, the present application also provides a radiotherapy dose calculation optimization device based on a generative adversarial network, comprising:

[0028] Data preparation module, used to obtain phase space files and condition vectors of different accelerators and make them into data sets;

[0029] A model training module, used to train and verify the constructed generative adversarial network model using the dataset;

[0030] A model application module is used to determine the values of each parameter in the condition vector based on the actual operating parameters of a specific accelerator, and input them into the verified generative adversarial network model to generate particle distribution data corresponding to the condition vector;

[0031] The dose calculation module is used to use the particle distribution data generated by the model as the initial input of the MC simulation method to obtain the radiotherapy dose distribution results simulated by MC.

[0032] Its further technical solution is that the device also includes:

[0033] The QA quality inspection module is used to compare the MC simulation results of a specific accelerator with the measurement results under the same conditions to verify the MC simulation results. When the deviation exceeds the allowable range, the values of each parameter in the condition vector input to the model are adjusted, and the model application module and dose calculation module are called to regenerate the particle distribution data and obtain updated MC simulation results, and then compare again until the deviation is within the allowable range.

[0034] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in the first aspect when executing the computer program.

[0035] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described in the first aspect when executed by a processor.

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

[0037] In the radiotherapy dose calculation optimization method proposed in the present application, the acquired phase space file data set is input into the generative adversarial network model for learning. The generator uses the conditional vector to generate a simulated particle distribution and compares it with the real particle distribution in the phase space file. The discriminator determines the truth of the input data and returns the adjustment parameters. After obtaining the stable model, the conditional vector can be determined according to the actual operating parameters of the specific accelerator, and the trained model can be used to directly generate the corresponding particle distribution data, which is used as the initial input in the MC simulation to quickly obtain the dose simulation results. Compared with the traditional MC simulation method, which takes a long time to generate phase space files, and the phase space files are large in size, and the storage and use efficiency are low, the present application introduces a generative adversarial network model to directly generate the particle distribution of any accelerator to replace the traditional phase space file. While ensuring the accuracy of dose calculation, the calculation efficiency can be greatly improved. At the same time, the generated particle distribution data occupies a small volume, saving storage costs.

[0038] In addition, the MC simulation results must undergo strict quality assurance (QA) verification. Compared with the traditional verification process, which requires continuous adjustment of the accelerator structure configuration parameters to simulate and generate the corresponding phase space file, this application introduces a generative adversarial network to simplify the verification process. It only needs to adjust the conditional vector input to the model to quickly obtain simulation result feedback.

[0039] For different accelerator scenarios, MC simulations require different structures and multiple rounds of debugging to achieve matching simulation results. However, the generative adversarial network introduced in this application has become versatile and robust after training on a large number of different accelerator datasets, and can be adapted to the corresponding accelerator by adjusting the model parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of the radiotherapy dose calculation optimization method based on the generative adversarial network in the first embodiment of the present application.

[0041] Figure 2 This is a schematic diagram of the generative adversarial network model training process in the first embodiment of the present application.

[0042] Figure 3 It is a schematic diagram of the QA verification process for MC simulation results in the first embodiment of this application.

[0043] Figure 4 This is a structural diagram of the radiotherapy dose calculation optimization device based on the generative adversarial network in the second embodiment of the present application.

[0044] Figure 5 This is a schematic diagram of the structure of a computer device in the third embodiment of the present application. DETAILED DESCRIPTION

[0045] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0046] Generative Adversarial Network (GAN) is a deep learning model consisting of a generator and a discriminator. The generator generates realistic samples by learning the distribution of real data, while the discriminator is responsible for distinguishing between real samples and generated samples. Both are continuously optimized in adversarial training. GAN has a wide range of applications in image generation, data enhancement, and medical image synthesis. For example, GAN can be used to generate high-precision simulated CT images and dose distribution maps, thereby accelerating the formulation of treatment plans. Therefore, in response to the shortcomings of the prior art, this application considers using a method combining GAN and MC simulation to optimize radiotherapy dose calculation and QA verification processes, and provides the following several embodiments to explain the technical solution of this application in detail.

[0047] Example 1:

[0048] This embodiment provides a method for optimizing radiotherapy dose calculation based on a generative adversarial network. Figure 1 As shown, the specific steps include:

[0049] Step 1: Obtain the phase space files and condition vectors of different accelerators and make them into data sets.

[0050] Specifically, for medical linear accelerators of different models and parameters, corresponding condition vectors are constructed. The condition vectors include the characteristics of the accelerator itself and certain parameters measured in the calibration experiment. The condition vector c given in this embodiment includes key parameters such as particle source position variance, initial energy mean, accelerator parameters and measured physical data. Among them, the particle source position variance (satisfying Gaussian distribution) and the initial energy mean can be selected from the range given in the accelerator open source database; the accelerator parameters include at least the accelerator model code, the head angle, and the irradiation field size. These parameter values can be obtained from the medical linear accelerator manual; the physical data include at least the percentage depth dose PDD, the off-axis ratio curve Profile, and the tissue phantom ratio TPR. These parameter values can be obtained by measuring and analyzing the medical linear accelerator. Then, based on the parameters in the condition vector c, the MC simulation method is used to simulate and calculate the phase space file of medical linear accelerators of different models and parameters. During the simulation process, the spatial coordinates (x, y, z, normalized to [-1, 1]), the direction of motion (polar angle θ, azimuth angle θ) of each particle are recorded in detail. The phase space files (expressed in spherical coordinates), energy values (E), and statistical weights (w) are stored in the phase space file. Finally, the phase space files and condition vectors c of different accelerators simulated using MC are compiled into a dataset for subsequent model training and validation.

[0051] In this embodiment, to ensure the comprehensiveness and representativeness of the data set, the simulation covers a variety of commonly used clinical accelerator configurations and operating parameters, and generates a large-scale phase space file data set to reduce the uncertainty of particle distribution and improve the model training effect.

[0052] Before model training, the generated phase space file must be preprocessed, including outlier removal, data smoothing, and feature optimization, to improve data quality and model training efficiency. Subsequently, the dataset can be divided into a training set and a validation set at an 80% to 20% ratio, ensuring consistent parameter distributions in both parts and providing a reliable data foundation for model training and validation.

[0053] Step 2: Use the dataset to train and validate the constructed generative adversarial network model.

[0054] The GAN model is built based on the deep learning framework, in which both the generator and the discriminator adopt the multi-layer perceptron (MLP) structure. Figure 2 As shown in Figure 1, the generator inputs a random noise vector (following a Gaussian distribution) and a conditional vector c. Through a fully connected layer and an activation function, the generator gradually maps the input data to simulated particle distribution data with the same dimensions as the actual particle distribution in the phase space file. The discriminator receives the actual or simulated particle distribution data and the conditional vector c, extracts data features through a multi-layer neural network, and performs a binary classification to determine the authenticity of the input data.

[0055] There are many types of GAN models. In this embodiment, the WGAN model is used as an example. The WGAN model uses Wasserstein distance as the loss function and stabilizes training through gradient penalty. The original optimization goal of the WGAN model is:

[0056]

[0057] Among them, D(x|c) is the score of the discriminator for the real sample x, P r is the true data distribution; D(G(z|c)|c) is the score of the discriminator for the generated sample G(z|c), p (z) is the probability distribution of noise vector; is the expected value.

[0058] To ensure the stability and generative effect of the WGAN training process, a gradient penalty term is introduced into the loss function to constrain the gradient norm of the discriminator:

[0059]

[0060] in, For the discriminator to generate samples (i.e., the score of G(z|c)), P g To generate sample distribution; is the gradient of the discriminator with respect to the input, and the Lipschitz continuity constraint is used in the third term: Ensure that the score function of the discriminator is smooth to avoid gradient explosion; P int is the distribution obtained by random interpolation between the generated sample and the real sample, λ is the gradient penalty coefficient, and the interpolated sample for:

[0061]

[0062] The loss function of the generator is set as:

[0063]

[0064] Among them, KL is the KL divergence of the statistical weight, α is the hyperparameter; w is the statistical weight of the particle distribution, which is divided into the simulated particle distribution weight w g and the real particle distribution weight w r .in To counter the loss, maximize the score of the discriminator on the generated samples; α·KL(w g ||w r ) is a physical constraint term that forces the weight distribution to conform to physical laws.

[0065] The divided training set is input into the WGAN model constructed above, the batch size is set, and the number of iterations for each training cycle is determined according to the scale of the training part and the convergence speed requirements. During the training process, the generator and the discriminator are optimized alternately. In this example, the discriminator parameters are updated several times iteratively, and then the generator parameters are updated. The WGAN training strategy is followed to prevent the insufficient generator capacity in the early stage of model training from affecting the overall effect. By monitoring the changing trend of the loss curves of the generator and the discriminator, generating multi-dimensional information such as sample quality evaluation indicators, it is determined whether the WGAN model has reached a balanced state. When the loss curves of the generator and the discriminator tend to be stable, the simulated samples and the real samples are highly similar in the key features of the particle distribution (such as the statistical characteristics of the spatial coordinate distribution, the shape of the energy spectrum, etc.), and the fluctuation amplitude of the model performance indicators in multiple consecutive training cycles is less than the preset threshold (such as 1%), then the WGAN model is considered to have reached a balanced state, and the entire training process is like Figure 2 shown.

[0066] After reaching equilibrium, a separate validation set is used to conduct a comprehensive performance evaluation of the trained WGAN model. The evaluation metrics cover multiple dimensions, including the accuracy of the generated particle distribution (the error range compared to the actual data in the phase space file), generation efficiency (the number of particles generated per unit time), and model generalization (adaptability to different accelerator parameter combinations), ensuring reliable performance in real-world applications.

[0067] At this point, a mature WGAN model has been obtained, which can be directly applied to MC simulation methods and QA quality inspection processes. Refer to the following steps.

[0068] Step 3: Determine the values of each parameter in the conditional vector based on the actual operating parameters of the specific accelerator, and input them into the verified generative adversarial network model to generate particle distribution data corresponding to the conditional vector. This data is directly used as the initial input of the MC simulation method to obtain the radiotherapy dose distribution result simulated by MC.

[0069] Specifically, in the actual radiotherapy planning process, when dose calculation or beam characteristic analysis is required for a specific accelerator configuration, the values of the parameters in the conditional vector are determined based on the actual operating parameters of the specific accelerator and input into the verified WGAN model. The WGAN model quickly generates particle distribution data corresponding to the conditional parameters. This data directly serves as the initial input source for the MC simulation, replacing the traditional cumbersome process of simulating the accelerator head through a complete MC simulation to obtain a phase space file. This significantly shortens simulation calculation time and improves treatment planning efficiency.

[0070] The resulting MC simulation results must undergo rigorous quality assurance (QA) verification. However, during the quality assurance process of traditional MC algorithms, the accelerator's physical data (such as percent depth dose and off-axis ratio curves) must first be measured under standard test conditions. Then, during the MC simulation under the same conditions, the accelerator parameters are rapidly found by iterating through the MC simulation parameters or employing an iterative optimization algorithm. Phase-space files are generated several or even dozens of times. Finally, the MC simulation results are verified by comparing the measured and simulated data to obtain accelerator parameters that meet clinical standards. If significant discrepancies are found, the accelerator's structural parameters must be repeatedly adjusted to ensure that the simulation environment is consistent with the actual situation. This process is often time-consuming and seriously affects accelerator utilization efficiency.

[0071] In this embodiment, MC simulation based on the WGAN model can accelerate the quality inspection process and significantly improve the quality inspection efficiency. Figure 3 As shown, this method also includes step 4: For a specific accelerator, its MC simulation results are compared with the measurement results under the same conditions to verify the MC simulation results. If the deviation exceeds the allowable range, the values of the parameters in the condition vector input to the model are adjusted. That is, by adjusting key parameters such as the particle source position variance, initial energy mean, percentage depth dose (PDD), off-axis ratio profile, and tissue phantom ratio (TPR), the particle distribution data is regenerated and an updated MC simulation result is obtained. The comparison is then repeated until the deviation is within the allowable range. This method uses the WGAN model to quickly generate particle distribution data consistent with the measurement results, replacing the traditional phase space file, thereby achieving efficient QA verification.

[0072] Example 2:

[0073] Based on the same inventive concept, this embodiment provides a device for implementing the aforementioned method for optimizing radiotherapy dose calculation based on a generative adversarial network. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of the embodiment of the device for optimizing radiotherapy dose calculation based on a generative adversarial network provided below can be found in the above limitations of the method for optimizing radiotherapy dose calculation based on a generative adversarial network, and will not be repeated here. Figure 4As shown, the device includes a data preparation module, a model training module, a model application module, and a dose calculation module, which are connected in sequence, and optionally also includes a quality assurance (QA) module. The data preparation module is used to obtain phase space files and condition vectors for different accelerators and generate a data set. The model training module is used to use the data set to train and verify the constructed generative adversarial network model. The model application module is used to determine the values of each parameter in the condition vector based on the actual operating parameters of a specific accelerator, and input them into the verified generative adversarial network model to generate particle distribution data corresponding to the condition vector. The dose calculation module is used to use the particle distribution data generated by the model as the initial input to the MC simulation method to obtain the radiotherapy dose distribution results simulated by MC. The QA module is used to compare the MC simulation results of a specific accelerator with the measured results under the same conditions to verify the MC simulation results. If the deviation exceeds the allowable range, the value of each parameter in the condition vector input to the model is adjusted. The model application module and the dose calculation module are then called to regenerate the particle distribution data and obtain an updated MC simulation result. The comparison is repeated until the deviation is within the allowable range.

[0074] Each module in the aforementioned generative adversarial network-based radiotherapy dose calculation and optimization device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0075] Example 3:

[0076] This embodiment provides a computer device, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, it includes a processor 101 and a memory 102 connected via a system bus, and a computer program stored in the memory 102 and executable on the processor 101. The processor 101 of the computer device is used to provide computing and control capabilities. The memory 102 of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 101 executes the computer program, steps 1 to 4 in the embodiment of the above-mentioned method for optimizing the calculation of radiotherapy doses based on a generative adversarial network are implemented. Alternatively, when the processor 101 executes the program, the functions of the various modules in the embodiment of the device for optimizing the calculation of radiotherapy doses based on a generative adversarial network are implemented, which will not be repeated here.

[0077] Example 4:

[0078] If the integrated module of the device for optimizing the calculation of radiotherapy doses based on a generative adversarial network is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the embodiment of the method for optimizing the calculation of radiotherapy doses based on a generative adversarial network, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement steps 1 to 4 in the embodiment of the method for optimizing the calculation of radiotherapy doses based on a generative adversarial network. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0079] The above description is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiment. It is understood that other improvements and variations directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the scope of protection of the present invention.

Claims

1. A radiotherapy dose calculation optimization method based on generative adversarial networks, characterized in that: The method comprises: Obtain phase space files and condition vectors of different accelerators and make them into data sets; Use the dataset to train and verify the constructed generative adversarial network model; Determine the values of each parameter in the condition vector based on the actual operating parameters of a specific accelerator, and input them into a verified generative adversarial network model to generate particle distribution data corresponding to the condition vector; The particle distribution data generated by the model is used as the initial input of the MC simulation method to obtain the radiotherapy dose distribution results simulated by MC.

2. The method for optimizing radiotherapy dose calculation based on generative adversarial networks according to claim 1, characterized in that: The step of obtaining phase space files and condition vectors of different accelerators includes: For medical linear accelerators of different models and parameters, corresponding condition vectors are constructed, including particle source position variance, initial energy mean, accelerator parameters, and measured physical data; wherein the accelerator parameters include at least the accelerator model code, handpiece angle, and irradiation field size, and the physical data include at least the percentage depth dose, off-axis ratio curve, and tissue phantom ratio; Based on the parameters in the condition vector, the MC simulation method is used to simulate and calculate the phase space file of medical linear accelerators of different models and parameters; During the simulation process, the spatial coordinates, motion direction, energy value and statistical weight of each particle are recorded and stored in the phase space file.

3. The method for optimizing radiotherapy dose calculation based on generative adversarial networks according to claim 1, characterized in that: Build a generative adversarial network model, including: A generative adversarial network model is built based on a deep learning framework, where both the generator and the discriminator adopt a multi-layer perceptron structure; The input of the generator is a random noise vector and the condition vector, which are gradually mapped through a fully connected layer and an activation function to output simulated particle distribution data with the same dimension as the real particle distribution in the phase space file; The discriminator receives real or simulated particle distribution data and the condition vector, extracts data features through a multi-layer neural network and performs binary classification to determine the authenticity of the input data.

4. The method for optimizing radiotherapy dose calculation based on generative adversarial networks according to claim 3, characterized in that: The constructed generative adversarial network model is trained using the dataset, including: Input the training parts divided from the dataset into batches into the generative adversarial network model, set the batch size, and determine the number of iterations per training cycle based on the size of the training parts and the convergence speed requirements; During the training process, the generator and the discriminator are optimized alternately, and the discriminator parameters are first updated several times iteratively, and then the generator parameters are updated until the generative adversarial network model reaches a balanced state.

5. The method for optimizing radiotherapy dose calculation based on generative adversarial networks according to claim 4, characterized in that: The generation of the adversarial network model reaches a balanced state when: When the loss curves of the generator and the discriminator tend to be stable, the simulated samples and the real samples are highly similar in the key features of the particle distribution, and the fluctuation amplitude of the model performance indicators in multiple consecutive training cycles is less than the preset threshold, the generative adversarial network model is considered to have reached a balanced state.

6. The method for optimizing radiotherapy dose calculation based on generative adversarial networks according to claim 1, characterized in that: The method further comprises: For the specific accelerator, comparing its MC simulation results with measurement results under the same conditions to verify the MC simulation results; If the deviation exceeds the allowable range, the values of the parameters in the condition vector input to the model are adjusted, the particle distribution data are regenerated, and the updated MC simulation results are obtained; Compare again until the deviation is within the allowable range.

7. A radiotherapy dose calculation optimization device based on generative adversarial networks, characterized in that: include: Data preparation module, used to obtain phase space files and condition vectors of different accelerators and make them into data sets; A model training module, used to train and verify the constructed generative adversarial network model using the dataset; A model application module is used to determine the values of each parameter in the condition vector based on the actual operating parameters of a specific accelerator, and input them into the verified generative adversarial network model to generate particle distribution data corresponding to the condition vector; The dose calculation module is used to use the particle distribution data generated by the model as the initial input of the MC simulation method to obtain the radiotherapy dose distribution results simulated by MC.

8. The radiotherapy dose calculation optimization device based on generative adversarial network according to claim 7, characterized in that: The device further comprises: The QA quality inspection module is used to compare the MC simulation results of a specific accelerator with the measurement results under the same conditions to verify the MC simulation results. When the deviation exceeds the allowable range, the values of each parameter in the condition vector input to the model are adjusted, and the model application module and the dose calculation module are called to regenerate the particle distribution data and obtain updated MC simulation results, and then compare again until the deviation is within the allowable range.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images