Nasopharyngeal carcinoma radiotherapy planning system and method based on deep learning and threshold optimization

Abstract

The invention relates to the technical field of data processing, and discloses a nasopharyngeal carcinoma radiotherapy planning system and method based on deep learning and threshold optimization, and the method comprises the steps: obtaining a CT image of a patient; identifying a planned target area and a protection area in the CT image, and respectively segmenting the planned target area and the protection area; determining a predicted three-dimensional dose and a predicted dose volume histogram according to the segmented planned target region and the protection region; obtaining dose prescription data; and performing reverse optimization on the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data, and determining a radiotherapy plan. According to the method, the prediction of the three-dimensional dose and the optimization of the prediction dose volume histogram can be realized only by roughly defining the approximate range of the dose prescription data, and the accurate radiotherapy plan can be determined, so that the calculation process is simplified, the consumption of calculation resources is reduced, and the problem of model overfitting caused by excessive accurate calculation is avoided.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a nasopharyngeal carcinoma radiotherapy planning system and method based on deep learning and threshold optimization. Background Art

[0002] Nasopharyngeal carcinoma, as a public health problem that cannot be ignored worldwide, has a profound and serious impact. In particular, the incidence of nasopharyngeal carcinoma is on the rise in some regions, and the accuracy of its treatment is crucial to improving patient prognosis.

[0003] In the comprehensive treatment strategy for nasopharyngeal carcinoma, radiotherapy is one of the most important treatment methods, and its precise implementation is directly related to the treatment effect. At present, the formulation of nasopharyngeal carcinoma radiotherapy plan depends on the doctor's manual delineation of the target area (such as the tumor itself and the surrounding tissues that may be invaded) and the organs at risk (such as the eyes, brain and other important structures), as well as the physicist's fine adjustment of complex parameters such as field layout and dose distribution. This process is complicated and time-consuming. At the same time, due to the difference in professional judgment between doctors and physicists, there may be significant differences between treatment plans, which in turn affects the patient's treatment effect and quality of life.

[0004] In the field of radiotherapy, research on automatic radiotherapy planning technology is deepening. Studies have reported on the feasibility of using predicted doses for radiotherapy. Current research mainly uses models based on back-propagation neural networks and deep convolutional neural network (CNN) models. Early research methods trained models by exploring the relationship between the dose received by each voxel in the risk organ and the patient's geometric anatomical structure. These models take the features that affect the dose as input and the corresponding voxel dose points as output, and then train them through the back-propagation algorithm of the neural network. However, the 3D dose distribution prediction method based on BP neural network requires manual feature extraction, which leads to strong subjectivity in feature selection. Therefore, the rise of 3D dose distribution prediction technology based on deep convolutional neural network effectively solves this problem. These network models include U-net, DenseNet, HD U-net, etc.

[0005] Deep learning technology also demonstrates its core value in automatically outlining organs at risk (automatic segmentation). This technology uses a deep convolutional neural network (DCNN) to automatically outline medical images. For example, the U-NetDCNN proposed by Olaf et al. has been successfully applied to biomedical image recognition, realizing the automatic outlining of biological cell images. In the field of radiotherapy, some scholars have combined the trained DCNN model with GPU hardware acceleration to quickly outline tissues and organs in radiotherapy images, thereby accurately and automatically outlining the target area and the structure of organs at risk, promoting the development of precision radiotherapy technology. Domestic research in this area started late, with only some exploratory studies, such as Men Kuo et al.'s research on automatic outlining of organs at risk in the head and neck, and Deng Jincheng et al.'s attempts in the chest.

[0006] Another core technology of automated radiotherapy is the automated radiotherapy technology that combines dose prediction and optimization algorithms. Chanyavanich et al. proposed a method based on dose volume histogram (DVH) guidance. This method established a set of DVH models by analyzing the anatomical and geometric feature mapping relationship between a large number of executed treatment plans and the patient contouring results. For new cases, the possible DVH is predicted based on its geometric features, and this result is used to guide the optimization algorithm to generate a radiotherapy plan. Fan et al. trained a dose distribution prediction model based on anatomical and prescription dose data, and generated the final radiotherapy plan through the flux map optimization algorithm (FMO). Vasant Kearney et al. proposed a new fully convolutional volume dose prediction neural network (DoseNet), which adds residual blocks of the same dimension to U-net to reduce network redundancy, providing SBRT patients with better dose prediction accuracy than U-net and fully connected methods while maintaining computational efficiency.

[0007] Murakami Y et al. proposed a method based on generative adversarial networks (GAN) to synthesize dose distribution from CT datasets to FMO. This method only requires pairing CT and dose images. In order to compare the prediction performance, a model trained on the structure image of the organ at risk was also created. Although the performance of the CT model is slightly inferior to the structure mask model, it is within an acceptable range, and the prediction of dose distribution based on CT images can skip the target segmentation step and achieve a higher degree of automation. Shao et al.

[14] designed an asymmetric network (A-Net) to predict the optimal 3D dose distribution for lung cancer patients, in which the encoder and decoder are asymmetric to retain input information and adapt to the limitations of GPU memory. The network also uses squeeze and excitation SE units to improve data fitting ability and exceeds HD-Unet and 3D-Unet in most indicators. Song et al. applied DeepLabv3+ to the plan of the VMAT optimization algorithm and introduced a prior dose information-assisted planning scheme. The prediction accuracy was evaluated by paired t-test using MSE, normalized dose difference, and DVH index, and the dose distribution of rectal cancer was successfully predicted. Li HD et al. proposed a multi-task attention adversarial network (MtAA-NET) for automatic dose planning using only CT images. The framework includes a main dose distribution prediction task and an auxiliary segmentation task to provide anatomical information of PTV and OAR. It also introduces an attention mechanism embedded cross-task feature fusion (CtFF) module to further fuse features. Cros S et al. proposed a three-dimensional deep neural network (HDAU-net) that combines a hierarchical dense architecture with an attention U-net and adds a clinical objective with a weighted MSE-DVH loss function. This model outperforms the HD U-net model in all indicators and outperforms U-net in uniformity while maintaining a high consistency level and significantly reduced maximum dose. Zhan, Bo proposed a multi-constrained dose prediction model based on a generative adversarial network in 2022, named Mc-GAN. It can automatically predict dose distribution maps from CT images as well as masks of PTV and protection areas. During the feature extraction process, a dual attention module (DAM) is embedded to force the generator to pay more attention to internal semantic correlations. To improve the prediction accuracy, two additional losses, Local Constraint Loss (LCL) and Self-Supervision Perceptual Loss (SPL), are introduced. Evaluations on two internal datasets show that Mc-GAN outperforms other state-of-the-art methods in almost all PTV and protection region criteria.

[0008] Although the above-mentioned dose prediction methods have made significant technical progress, they have extremely high requirements for the consistency of the physical conditions of the data set. For example, the data set must come from a treatment plan using IMRT technology with the same number of fields. However, in clinical practice, it is extremely challenging to collect data that meets strict physical condition consistency because it is difficult to find cases with exactly the same physical conditions. In addition, similar cases bring a lot of uncertainty to the model method. The technologies used in clinical practice, such as IMRT and VMAT, have certain physical conditions. When multiple technical data sets with different physical conditions are mixed, these technologies are difficult to distinguish, resulting in dose prediction failure. Although GAN performs well in a data set of a single technology (physical condition), it is difficult to collect radiotherapy data of the same type of technology in clinical practice, which brings difficulties to clinical application.

[0009] Most existing studies adopt the mode of "dose prediction + traditional optimization algorithm", that is, the approximate distribution of DVH of the final plan is determined by dose prediction, and then the parameters are manually adjusted to make the dose distribution calculated by the optimization algorithm roughly consistent with the dose prediction, so as to determine the generated radiotherapy plan as the final radiotherapy plan. Obviously, this is a semi-automatic radiotherapy plan generation method, which requires manual repeated adjustment of parameters to ensure that the results are roughly consistent with the predicted results. Summary of the invention

[0010] The main purpose of the present invention is to provide a nasopharyngeal carcinoma radiotherapy planning system and method based on deep learning and threshold optimization, aiming to solve the problem of requiring manual repeated adjustment of parameters to ensure that the results are roughly consistent with the predicted results.

[0011] To achieve the above objectives, the present invention provides a nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization, comprising:

[0012] Obtain CT images of patients;

[0013] Identifying a planned target area and a protection area in a CT image, and segmenting the planned target area and the protection area respectively;

[0014] Determining a predicted three-dimensional dose and a predicted dose volume histogram according to the segmented planned target area and the protection area;

[0015] Obtain dosage prescription data;

[0016] The predicted three-dimensional dose and the predicted dose volume histogram are respectively reversely optimized according to the dose prescription data to determine a radiotherapy plan.

[0017] Preferably, the predicted three-dimensional dose is calculated in the following manner:

[0018] The calculation is performed by a pencil beam convolution stacking algorithm, and the calculation formula of the pencil beam convolution stacking algorithm is as follows:

[0019]

[0020] Where ρ is the density; Ψ is the probability density function distribution of the incident source, i.e., the fluence; x, y, and z are coordinates; x′ is the integral coordinate of x; y′ is the integral coordinate of y; z′ is the integral coordinate of z; D(x, y, z) is expressed as a three-dimensional dose, which is used to limit the physical conditions of the beam channel; K w is a pencil beam; F is a three-dimensional dose distribution function calculated by the pencil beam convolution superposition algorithm, and F describes the dose size at different coordinate positions (x, y, z).

[0021] Preferably, the pencil beam calculation method comprises:

[0022] Finite pencil beams were computed in a water phantom using the Monte Carlo method;

[0023] The pencil beam calculation method further includes:

[0024] Determining dose distribution data of the pencil beam in a water phantom according to a limited pencil beam;

[0025] The dose impact matrix was determined based on the dose distribution data of the pencil beam in the water phantom.

[0026] Preferably, after the step of determining the predicted three-dimensional dose and the predicted dose volume histogram according to the segmented planned target area and the protection area, the method further comprises:

[0027] Based on the formula of the pencil beam convolution superposition algorithm, a three-dimensional dose map is determined by a summation operation of the dose impact matrix.

[0028] Preferably, the step of identifying the planned target area and the protection area in the CT image and segmenting the planned target area and the protection area respectively comprises:

[0029] Identify the planned target area and protection area in the CT image;

[0030] The planned target area is divided into a plurality of sub-areas of preset sizes.

[0031] Preferably, after the step of dividing the planned target area into a plurality of sub-areas of preset sizes, the following steps are included:

[0032] Get beam channel;

[0033] Get three image channels for any of the sub-regions:

[0034] According to the firing rate channel and the three image channels, dose images of the sub-areas corresponding to the three image channels are determined.

[0035] Preferably, the firing rate channel generation method comprises:

[0036] extracting preset information through the planned target area;

[0037] According to historical cases, obtain the beam angle that meets the preset information;

[0038] Projecting the beam angle onto the planned target volume to obtain the grid weights of the beam channel data;

[0039] Select the ray directions with the grid weights arranged in the front column as the firing rate channels.

[0040] Preferably, the dosage prescription data includes:

[0041] Dose maximum objective function, dose average objective function, equivalent uniformity objective function, deviation from prescription dose degree objective and smoothness objective function.

[0042] Preferably, the step of performing reverse optimization on the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data to determine the radiotherapy plan includes:

[0043] The step of performing reverse optimization on the predicted three-dimensional dose and the predicted dose volume histogram respectively according to the dose prescription data;

[0044] Calculate the voxel dose unit based on the inverse optimization results and the dose calculation engine;

[0045] According to the voxel dose unit, optimizing a flux weight map;

[0046] The flux weight map and the machine information of the accelerator are used to optimize the leaf sequence using a leaf sequence algorithm, and a radiotherapy plan is determined according to the leaf sequence.

[0047] A nasopharyngeal carcinoma radiotherapy planning system based on deep learning and threshold optimization, wherein the system applies any of the above-mentioned nasopharyngeal carcinoma radiotherapy planning methods based on deep learning and threshold optimization, and the system includes a server and a processor;

[0048] The server is used to obtain CT images of patients; obtain dosage prescription data;

[0049] The processor is used to identify the planned target area and the protection area in the CT image, and to segment the planned target area and the protection area respectively; based on the segmented planned target area and the protection area, a predicted three-dimensional dose and a predicted dose volume histogram are determined; the system is used to reversely optimize the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data to determine the radiotherapy plan.

[0050] By only roughly defining the approximate range of the dose prescription data, the predicted three-dimensional dose and the optimization of the predicted dose volume histogram can be achieved, and an accurate radiotherapy plan can be determined. This not only simplifies the calculation process and reduces the consumption of computing resources, but also avoids the problem of model overfitting caused by over-precise calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of the nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization of the present invention.

[0052] Figure 2 This is a diagram of a generative adversarial network model based on the beam channel;

[0053] Figure 3 Schematic diagram of the beam channel using the pencil beam convolution model;

[0054] Figure 4 This is the generator model diagram of 8-layer 2D U-net.

[0055] Figure 5 Diagram of the discrimination process using Markov for the discriminator.

[0056] Figure 6 The leaf sequence diagram of the multi-leaf collimator is optimized for the leaf sequence algorithm.

[0057] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0058] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0059] In the subsequent description, the suffixes such as "module", "component" or "unit" used to represent elements are only used to facilitate the description of the present invention, and have no specific meanings. Therefore, "module", "component" or "unit" can be used in a mixed manner.

[0060] See also Figure 1 To achieve the above purpose, in the first embodiment of the present invention, a nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization includes:

[0061] Step S10, obtaining a CT image of the patient;

[0062] Step S20, identifying the planned target area and the protection area in the CT image, and segmenting the planned target area and the protection area respectively;

[0063] Step S30, determining a predicted three-dimensional dose and a predicted dose volume histogram according to the segmented planned target area and protection area;

[0064] Step S40, obtaining dosage prescription data;

[0065] Step S50, reversely optimizing the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data to determine the radiotherapy plan.

[0066] By only roughly defining the approximate range of the dose prescription data, the predicted three-dimensional dose and the optimization of the predicted dose volume histogram can be achieved, and an accurate radiotherapy plan can be determined. This not only simplifies the calculation process and reduces the consumption of computing resources, but also avoids the problem of model overfitting caused by over-precise calculations.

[0067] Specifically, step S50 includes:

[0068] Step S51, generating a preliminary plan based on the predicted three-dimensional dose and the predicted dose volume histogram;

[0069] Step S52, automatically evaluate and perform 3D quality assurance (3DQA) evaluation on the preliminary plan, and generate a plan report:

[0070] Step S53, reversely optimize the plan report according to the dose prescription data to determine the radiotherapy plan.

[0071] This embodiment realizes the full automation and efficiency of the process from CT scanning to the generation of personalized radiotherapy plans (i.e., radiotherapy plans for different patients). The core of this process lies in the close integration with the CT simulation positioning machine, which ensures the seamless connection and instant processing of data flow.

[0072] After the patient completes the CT scan, the system starts immediately and uses advanced image processing and machine learning technology to automatically perform subsequent steps.

[0073] First, it can accurately and automatically identify and outline the normal (endangered) organs in the head and neck region (i.e., the planned target area and the protected area), providing a solid guarantee for the safety of the radiotherapy plan. At the same time, for patients with nasopharyngeal cancer, it can also accurately identify and outline the tumor site, laying the foundation for precise attack on the tumor.

[0074] In the automatic delineation of organs throughout the body, the present invention will also adopt an innovative reverse delineation strategy. Under this strategy, the system will give priority to the detailed delineation of organs at risk, and then use the powerful analytical ability of deep learning to intelligently and automatically delineate the remaining areas based on the complex relationship between the boundary expansion of the planned target area and the gross tumor target area in different diseases of the head and neck. This strategy not only improves the comprehensiveness and accuracy of delineation, but also significantly reduces the manual operation of doctors, further improving work efficiency.

[0075] The generated plan report is then seamlessly transferred to the doctor's workstation for final review and approval by clinical experts. Doctors can intuitively view the automatic delineation results, dose distribution map and optimized radiotherapy plan on the workstation, and make necessary adjustments based on the patient's specific situation to ensure the personalization and safety of radiotherapy.

[0076] Specifically, protected areas (OARs) refer to normal tissues or organs that need special protection during treatment. Although these organs are not within the treatment target area, they may be affected by the side effects produced during treatment. Therefore, when formulating radiotherapy plans, the radiation dose to these organs must be reduced as much as possible to reduce the potential risk of damage.

[0077] In the second embodiment of the present invention, based on the first embodiment, the predicted three-dimensional dose is calculated in the following manner:

[0078] The calculation is performed using the pencil beam convolution stacking algorithm. The calculation formula of the pencil beam convolution stacking algorithm is as follows:

[0079]

[0080] Where ρ is the density; Ψ is the probability density function distribution of the incident source, i.e., the fluence; x, y, and z are coordinates; x′ is the integrated coordinate of x; y′ is the integrated coordinate of y; z′ is the integrated coordinate of z; D(x, y, z) is expressed as a three-dimensional dose, which is used to limit the physical conditions of the beam channel; Kw is a pencil beam; F is a three-dimensional dose distribution function calculated by the pencil beam convolution superposition algorithm, and F describes the dose size at different coordinate positions (x, y, z).

[0081] The pencil beam convolution superposition (PBC) algorithm is used for dose calculation. The core assumption of the pencil beam convolution superposition algorithm is that the ray is composed of a finite number of beams, and these finite beams are defined as "finite pencil beams". Under this setting, the dose distribution of any grid point in space is the result of the joint contribution of all pencil beams directly irradiated or scattered.

[0082] See also Figure 3 In a third embodiment of the present invention, based on the fourth embodiment, a pencil beam calculation method includes:

[0083] Step S60, using the Monte Carlo method to calculate a finite pencil beam in the water phantom;

[0084] After step S60, the method further includes:

[0085] Step S61, determining dose distribution data of the pencil beam in the water phantom according to the limited pencil beam;

[0086] Step S62, determining a dose impact matrix according to the dose distribution data of the pencil beam in the water phantom.

[0087] The Monte Carlo method in step S60 is a statistical simulation method used to simulate the process of photon-matter interaction, thereby obtaining the dose distribution data of the pencil beam in the water phantom. These data provide the basis for subsequent dose calculations. Steps S61 and S62 inputs include the patient's three-dimensional CT image, as well as the contour information of the organ and the planned target area. The pencil beam data in step S60 is combined with the patient's specific anatomical information to predict the dose distribution in actual treatment.

[0088] Specifically, after step S62, the following steps are further included:

[0089] Step S63, using the photon energy spectrum of a 6MV Varian radiotherapy machine to calibrate the dose impact matrix;

[0090] Step S64, performing three-dimensional human body non-uniformity correction on the dose impact matrix based on the CT density relationship.

[0091] The energy spectrum of the 6MV photons commonly used in radiotherapy does not need to be measured too accurately, so the corresponding 6MV Varian radiotherapy machine photon energy spectrum can be used to correct the dose impact matrix. This step is to ensure the accuracy of dose calculation and simplify the energy spectrum processing process.

[0092] Step S64 considers the effect of different tissue densities in the human body on the dose distribution, but does not require the use of a specific CT density relationship. This step is critical because it ensures that the dose calculation takes into account the density differences of different tissues in the human body, thereby providing a more accurate dose prediction.

[0093] In the fourth embodiment of the present invention, based on the third embodiment, after step S30, the following steps are included:

[0094] Step S31 , based on the formula of the pencil beam convolution superposition algorithm, a three-dimensional dose map is determined by summing the dose impact matrix.

[0095] See also Figure 5 In the fifth embodiment of the present invention, based on any one of the first embodiment to the fourth embodiment, step S20 includes:

[0096] Step S21, identifying the planned target area and the protection area in the CT image;

[0097] Step S22, dividing the planned target area into a plurality of sub-areas of preset sizes.

[0098] Introduce the Generative Adversarial Network (GAN) model. The GAN model consists of two main components: the generator and the discriminator. The core task of the generator is to generate an image G(z) that is as realistic as possible based on the input random noise vector z. At the same time, the responsibility of the discriminator is to receive the images produced by the generator and judge the authenticity of these images, outputting 1 for real images and 0 for fake images.

[0099] During the training process of the GAN model, a game is played between the generator and the discriminator: the generator continuously improves the quality of its generated images to make them closer and closer to real pictures, while the discriminator continuously improves its discrimination ability to better distinguish the authenticity of images. This process can be formalized as a min-max problem: we aim to maximize the ability of the discriminator while minimizing its judgment of the images produced by the generator. Through this adversarial training, the generator can gradually learn to produce high-quality images, while the discriminator can accurately identify the subtle differences between real and generated images. This game ultimately leads to a significant improvement in the quality of the images produced by the generator, so that even the discriminator has difficulty distinguishing between generated images and real images, at which point the equilibrium state of adversarial training is reached.

[0100] Figure 5 The core of the design idea of ​​the discriminator (i.e. PatchGAN) shown in the paper is to determine whether the input image is generated by the generator. Unlike traditional discriminators, PatchGAN does not make a single authenticity judgment on the entire image, but makes independent judgments on each sub-region (i.e. "patch") in the image. This design is based on the observation that different parts of the image can be regarded as independent of each other, so the accuracy and efficiency of the discriminator can be improved by judging these parts separately.

[0101] How PatchGAN works:

[0102] Image segmentation: The input image is cut into multiple small sub-regions, each sub-region has a size of N×N pixels.

[0103] Independent judgment: The discriminator makes a true or false judgment on each sub-region, that is, it determines whether this sub-region comes from a real image or an image generated by the generator.

[0104] Result integration: The discriminator takes the average of the judgment results of all sub-regions in an image as the final output. This averaging operation helps to balance the feature differences of different parts of the image, thereby obtaining an overall evaluation.

[0105] Parameter selection:

[0106] For an input image of 256×256 pixels, the study found that the discriminator performs best when the sub-region size is 70×70 pixels. This size of sub-region can capture enough local information and cover enough areas of the image to achieve effective judgment.

[0107] See also Figure 2 In the sixth embodiment of the present invention, based on the fifth embodiment, after step S22, the following steps are included:

[0108] Step S23, obtaining a beam channel;

[0109] Step S24, obtaining three image channels of any sub-region:

[0110] Step S25, determining the dose images of the sub-areas corresponding to the three image channels according to the firing rate channel and the three image channels.

[0111] The generative adversarial network model based on the beam channel, with CT images as the conditional basis, can generate virtual images that have a direct correspondence with the conditional images, thus completing the conversion process from image to image. The key to this method is to use the beam channel technology to conditionally restrict the GAN, ensuring that the generated image is not only visually close to reality, but also closely related to the actual radiotherapy situation in terms of dose distribution. This GAN model combined with the characteristics of the beam channel improves the accuracy and reliability of radiation dose prediction.

[0112] The traditional GAN ​​model consists of a generator and a discriminator. The generator receives a single conditional input Y, where Y represents the input CT image imgA. The generator uses a U-net structure to encode and decode the input Y into the target real image imgB'. This process achieves the conversion from one modality to another while maintaining the key features of the CT image.

[0113] The discriminator processes two inputs: one is the generated image imgB' or the corresponding real image realx (i.e., imgB), with the same condition Y; the other is the beam-channel, which is combined with the input Y to obtain the so-called fakey (i.e., imgC) by using a pencil beam convolution stacking model. Here, imgC is generated based on a physical model. It utilizes the concept of beam channels and calculates the projection weight of the field angle on the PTV (i.e., the planned target volume), thereby limiting the physical range and ensuring the rationality of the dose distribution, which helps to achieve accurate image-to-image conversion. Through this dual-channel approach, the beam-channel not only provides the discriminator with fake data based on a physical model, but also combines with the images generated by the generator to deceive the discriminator. This complex adversarial process prompts the generator to continuously optimize to generate images that are anatomically and physically reasonable, while the discriminator also evolves to better distinguish between true and false image pairs. It effectively combines the physical characteristics of radiotherapy and the powerful capabilities of deep learning to improve the accuracy and practicality of CT image processing.

[0114] The generator (U-net structure) is used to receive four channel images as input, including three channel dose images and one beam channel. The generator adopts the U-net architecture, which is a network structure commonly used for image segmentation tasks. It consists of an encoder and a decoder, which can extract image features in the encoding stage and gradually restore image details in the decoding stage. In this way, the generator can generate a predicted dose image, namely the false _x (image B'). This predicted dose image is the model's understanding and reconstruction of the input dose image.

[0115] The discriminator is used to judge the difference between the generated predicted dose image Fake_x and the original dose image Image_A. The discriminator takes Fake_x and Image A as input at the same time, and outputs the judgment result by comparing the difference between the two. The key of this step is to evaluate the authenticity of the predicted dose image, that is, to judge the consistency between the dose image generated by the generator and the actual dose image.

[0116] During the training process of the beam channel-based generative adversarial network model, a part of the training data is used as a validation set (i.e., real_x, rreal_x, i.e., image B) to monitor the performance of the beam channel-based generative adversarial network model during the training process and calculate the error. This helps to adjust the parameters of the beam channel-based generative adversarial network model, prevent overfitting, and ensure that the beam channel-based generative adversarial network model has good generalization ability. After the training is completed, this validation set is no longer used.

[0117] The detailed structural design of the generator and discriminator is crucial to the performance of the model. The U-net structure of the generator effectively combines deep semantic information and shallow position information through skip connections, which helps to generate more accurate dose distribution maps. The discriminator needs to be complex enough to capture and distinguish subtle differences to ensure that the quality of the generated images can be accurately evaluated. The combined effect of this design enables the model to provide accurate and reliable dose predictions in radiotherapy planning, providing strong support for clinical decision-making.

[0118] The generator uses an 8-level hierarchical U-net to achieve the task from image to dose mapping. Figure 4 As shown in the figure, the entire network structure can be divided into two parts: feature extraction part and upsampling part.

[0119] The generator receives as input a 256×256 pixel image as a feature map, which has three channels of dose images and one beam channel.

[0120] The feature map is reduced in size layer by layer by alternating 3×3 convolution operations and 2×2 maximum pooling layers. Specifically, the feature map starts from 256×256 pixels, and after each layer of convolution and pooling, the size of the feature map is halved until it reaches 1×1 pixels.

[0121] After feature extraction, the generator uses a 3×3 convolution operation and a 2×2 maximum pooling layer again, combined with a 2×2 deconvolution layer to gradually expand the pixel size of the feature map until the feature map is restored to the original image size (i.e., 256×256 pixels).

[0122] In order to preserve the underlying information and prevent the loss of details, the network integrates high-level features from the feature extraction stage during the upsampling process. This skip connection method helps the network better restore image details and improve the accuracy of dose mapping.

[0123] Generator training phase

[0124] Optimizer: During the training phase, the Adam algorithm was used as the optimizer to effectively update the network weights.

[0125] Learning rate and training cycle: Set the learning rate to 2×10 -3 , and set the training cycle (epochs) to 1000 to ensure that the network fully learns the features in the dataset.

[0126] Finally, the generator outputs a 256×256×1 dose map that represents the radiation dose for each pixel in the input image.

[0127] In this way, the generator of the 8-level hierarchical U-net can accurately learn the complex mapping relationship from input image to radiation dose, which is crucial for the formulation of radiotherapy treatment plans. In addition, the use of the Adam optimizer and appropriate learning rate settings ensured the stability and efficiency of the network during training, thereby achieving accurate dose prediction on the nasopharyngeal carcinoma dataset.

[0128] In a seventh embodiment of the present invention, based on the sixth embodiment, a method for generating a firing rate channel includes:

[0129] Step S70, extracting preset information through the planned target area;

[0130] Step S71, obtaining a beam angle that meets preset information according to historical cases;

[0131] Step S72, projecting the beam angle onto the planned target area to obtain the grid weight of the beam channel data;

[0132] Step S73, selecting the ray directions with the grid weights arranged in the front row as the firing rate channels.

[0133] Beam channel is a method to calculate beam angle weight. Beam channel uses the patient's CT image to calculate the beam dose with a larger weight and uses the result as the input data of the generator. The grid weight is calculated as follows: the grid weight is equal to the projected area of ​​the beam angle in the planned target area under the beam direction divided by the total area of ​​all field angles. The fields falling within the area are set to 1, and the rest are set to 0; the ray directions with larger grid weights (i.e., the grid weights are arranged in the front row) are selected for dose calculation.

[0134] In an eighth embodiment of the present invention, based on any one of the first to fourth embodiments, the dosage prescription data includes:

[0135] Dose maximum objective function, dose average objective function, equivalent uniformity objective function, deviation from prescription dose degree objective and smoothness objective function.

[0136] The predicted 3D dose is reversely solved using a threshold optimization algorithm to automatically obtain a radiotherapy plan. The threshold optimization method is based on the predicted 3D dose distribution and solves it by constructing a series of parameter threshold ranges (i.e., dose prescription data) adjusted by the doctor. The dose prescription data only needs to be set once.

[0137] A series of linear objective functions are used to formulate the inverse planning optimization problem: a target is set for each organ at risk.

[0138] The calculation formula of the maximum dose objective function is as follows:

[0139]

[0140] Among them, y i It represents the maximum dose allowed for a specific organ at risk in the i-th treatment. i It is set based on clinical safety standards and previous studies to ensure the safety and effectiveness of the treatment.

[0141] It means selecting an optimal plan from all possible treatment plans v to minimize a specific goal (i.e., the maximum dose). It is usually hoped to minimize the dose to organs at risk to reduce potential side effects.

[0142] d v is the dose to the organ at risk associated with the selected treatment plan v. During the optimization process, a corresponding dose d is calculated for each treatment plan v. v , dose d v Reflects the amount of radiation that the organ at risk will receive under treatment plan v.

[0143] It indicates that the above optimization process needs to be performed for all treatment times i, where I is the set of all treatment times i. This means that an optimal plan needs to be selected independently for each treatment to ensure that the dose to the organ at risk is properly managed throughout the course of treatment. The maximum dose objective function is used in radiotherapy plan optimization to determine the best plan for each treatment to minimize the dose to the organ at risk. This is achieved by comparing the dose distribution under different plans, with the aim of minimizing damage to healthy tissues while ensuring the treatment effect.

[0144] The calculation formula of the dose mean objective function is as follows:

[0145]

[0146] Among them, z i It represents the average dose achieved by a certain radiotherapy plan in the i-th treatment. The average dose may refer to the average radiation dose received by the tumor or organs at risk (OARs). i is the average radiation dose to organs at risk.

[0147] It means to select an optimal plan from all feasible radiotherapy plans v so that a specific target (i.e., average dose) is maximized or minimized. The optimal plan maximizes the dose to the tumor and minimizes the dose to the organ at risk.

[0148] Each treatment requires an independent selection of an optimal plan to ensure the best therapeutic effect and safety throughout the course of treatment.

[0149] The calculation formula of the equivalent uniformity objective function is as follows:

[0150]

[0151] Among them, h f It represents the equivalent uniformity index for a specific organ at risk or region f. f It is used to measure the uniformity of dose distribution in the area. The higher the value, the more uniform the dose distribution.

[0152] mean{} is an operation for finding the average value, which requires calculating the average value of the expression in the brackets. mean{} means taking the average of all possible dose differences.

[0153] max{0, d v -f} calculates the planned dose d at each voxel v The positive difference between the tolerance dose threshold f and the organ at risk. v -If f is less than 0, it is taken as 0 and the difference below the tolerated dose is not calculated.

[0154] The above calculation needs to be performed for all different organs or regions at risk, F i is the set of all organs or regions at risk considered in treatment plan i.

[0155] The equivalent uniformity objective function calculation formula is used to evaluate the dose distribution uniformity of the organs at risk in the radiotherapy plan. By maximizing this indicator, it can be ensured that the dose at all voxels is as close as possible to the tolerance dose threshold of the organs at risk, while avoiding excessive dose at any voxel, thereby reducing the risk of potential side effects.

[0156] The formula for calculating the deviation from the prescribed dose target is as follows:

[0157]

[0158] In radiotherapy plan optimization, the dose control of organs at risk (OARs) is very important. The calculation formula of the deviation from the prescribed dose degree target is used to evaluate and optimize the radiotherapy plan to ensure that the irradiated dose of the organs at risk is as close as possible to the predetermined safety threshold.

[0159] Among them, l t Indicates the deviation from the prescribed dose for a specific organ or area at risk. t It is used to measure the difference between the actual dose distribution and the prescribed dose, especially when the dose is lower than the prescribed dose.

[0160] meam{},v∈O tThe operation of averaging the dose differences in all possible radiotherapy plans v indicates that the average value of the expression in the brackets needs to be calculated to obtain a comprehensive deviation index.

[0161] max{0,θt-d v} is the planned dose d calculated for each voxel v and the tolerance dose threshold θ for organs at risk t If θ t -d v If the value is less than 0, it is taken as 0, and the difference above the tolerance dose is not calculated.

[0162] The above calculations need to be performed for all different organs or regions at risk t, where t is the set of all organs or regions at risk considered in the treatment plan.

[0163] U t is the deviation of "exceeding the prescribed dose" for a specific organ at risk or area t. t Similar to U t It is also used to measure the difference between the actual dose distribution and the prescribed dose, but U t The focus is on those situations where the prescribed dose is exceeded.

[0164] mean{max{0,d v -θ t}}: This expression calculates the planned dose d at each voxel v and the tolerance dose threshold of organs at risk θ t If d v -θ t If it is less than 0, it is taken as 0 and the difference below the tolerated dose is not calculated.

[0165] In summary, the target mid-life formulas for the degree of deviation from the prescribed dose are used to calculate the degree to which the dose distribution of the organs at risk in the radiotherapy plan is "under" and "over" compared to the prescribed dose. By minimizing these deviation indicators, it can be ensured that the radiotherapy plan adheres to the prescribed dose limit as much as possible, thereby protecting the organs at risk from unnecessary radiation damage.

[0166] The calculation formula of the smoothing objective function is as follows:

[0167]

[0168] In radiotherapy plan optimization, the smoothing objective function is used to evaluate and optimize the smoothness of dose distribution to ensure that the dose variation is not too drastic, thereby reducing damage to surrounding normal tissues.

[0169] Among them, mk Represents the smoothness index for a specific organ at risk or region k. k It is used to measure the smoothness of the dose distribution in the area. The higher the value, the smoother the dose distribution.

[0170] It is an operation to find the maximum value of the dose difference in all possible radiotherapy plans v. It means that the maximum value of the expression in the brackets needs to be calculated to obtain a smoothness index in the worst case.

[0171] ∑b∈B r , max{0, w b -w b′} is to calculate the weight w between all adjacent voxel pairs (b, b′) under a certain radiotherapy plan r b and w b′ If w b -w b′ If it is less than 0, it is taken as 0 and the differences with lower weights are not calculated.

[0172] The above calculations need to be performed for all different organs or regions at risk k, where K is the set of all organs or regions at risk considered in the treatment plan.

[0173] The smooth objective function calculation formula is used to evaluate the smoothness of the dose distribution of the organs at risk in the radiotherapy plan. By minimizing this indicator, it can be ensured that the dose changes will not be too drastic, thereby reducing potential damage to the surrounding normal tissues. This helps to achieve more accurate and safe radiotherapy treatment.

[0174] The calculation formula for reverse optimization is as follows:

[0175]

[0176] In radiotherapy planning optimization, inverse optimization is a commonly used method to automatically generate the optimal radiation dose distribution according to clinical goals.

[0177] Here, z is some global dose index, such as average dose or maximum dose, which is used to control the overall dose distribution.

[0178] y is related to the dose to specific areas (such as tumors or organs at risk) and is used to ensure that the doses to these areas are within a safe range.

[0179] x is the dose weight of different voxels, which is used to adjust the dose distribution to meet the clinical goals.

[0180] Both l and u are related to linear constraints, which are used to limit the range or proportion of certain decision variables.

[0181] m is a template or model parameter used to define the shape or characteristics of the dose distribution.

[0182] g is related to other terms in the objective function and is used to balance different optimization objectives.

[0183] w is the weight coefficient, which is used to adjust the relative importance of different items in the objective function.

[0184] γ i and β i They are weights used to adjust the dose contribution of different regions or voxels in the objective function.

[0185] K f is the weight associated with the organ at risk f, which is used to control the dose limit of this organ.

[0186] φ t and ψ t The weights are all related to the constraints or targets at time point t, which are used to take into account the variation of dose over time.

[0187] ∑ i∈I (·) is the sum of all elements with index i in some set I. Here, it may represent some sum operation such as total dose or mean dose for all patients, all treatment areas, or all radiotherapy plans.

[0188] F i is the set of organs at risk associated with index i.

[0189] O i is the set of optimized regions associated with index i.

[0190] T represents a collection of time points used to account for changes in dose over time.

[0191] K and R k Both represent a collection of templates or models that are used to define the shape or characteristics of the dose distribution.

[0192] B τ Represents the set of radiotherapy regions associated with index τ.

[0193] represents the dose or weight at voxel v and organ at risk f.

[0194] y i is the dose index associated with index i.

[0195] l t and u t is a variable, I t and u t The linear constraints associated with the time point t.

[0196] m k Represents the parameters of the template or model k.

[0197] g and w are weights, which are related to other terms in the objective function.

[0198] Constraints section:

[0199]

[0200] The constraint means that for each index i, the corresponding variable z i is through a set O i The calculation is done by taking a weighted average of the voxels in the image. This helps ensure that the dose distribution is uniform while taking into account the importance of different elements.

[0201]

[0202] The constraint means that for each index i and each voxel v, the corresponding variable y i It must be greater than or equal to a certain calculated result. This is to ensure that the dose does not exceed a certain safety threshold, thereby protecting normal tissues from the harm of excessive doses.

[0203] In summary, this inverse optimization formula generates the optimal radiation dose distribution by adjusting different parameters and weights while satisfying a series of constraints. This method helps to achieve personalized treatment plans, improve treatment effects and reduce side effects.

[0204] After transformation, optimizing the inverse optimization model is a standard linear programming problem. The inverse optimization model can optimize the value of α and substitute the value into the original inverse planning optimization model as follows:

[0205]

[0206] subject to Ax≥b,

[0207] x ≥ 0.

[0208] It is known that the above inverse optimization model is a standard linear programming problem, which can optimize x (Fluencemap) and finally generate a high-quality plan.

[0209] In the ninth embodiment of the present invention, based on the second embodiment, step S50 includes:

[0210] Step S54, reversely optimizing the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data;

[0211] Step S55, calculating a voxel dose unit according to the inverse optimization result and the dose calculation engine;

[0212] Step S56, optimizing a flux weight map according to the voxel dose unit;

[0213] Step S57, the flux weight map and the machine information of the accelerator are used to optimize the leaf sequence using a leaf sequence algorithm, and a radiotherapy plan is determined based on the leaf sequence.

[0214] Based on the threshold optimization algorithm, combined with the voxel dose unit calculated by a specific dose calculation engine, the flux weight map is optimized. Combined with the machine information of the specific accelerator, the leaf sequence algorithm is used to optimize the leaf sequence and automatically obtain the final treatment plan (such as Figure 6 ). The generated preliminary plan is automatically evaluated, including scores for each item and a total score, and is provided to the physician for approval.

[0215] The beam channel-based generative adversarial network model solves the challenges of mixed datasets in dose prediction and has multiple beneficial effects:

[0216] 1. Improved the accuracy of dose prediction. The GAN model decomposes the complex radiation dose problem into a series of simpler sub-problems by introducing the concept of beam channels, each of which focuses on a specific beam channel. This approach helps the model more accurately capture the impact of different beams on dose distribution, thereby improving the accuracy of the overall prediction.

[0217] 2. Enhanced model adaptability to mixed data sets. Traditional GAN ​​models face difficulties in processing mixed data sets (such as IMRT and VMAT) because their data distribution and characteristics are quite different. By defining beam channels and using these channels as input, the complexity of mixed data sets is effectively reduced, allowing the model to better adapt to data from different treatment techniques.

[0218] 3. The pre-dose calculation process is simplified, and only the approximate range of dose prediction needs to be roughly defined. This not only simplifies the calculation process and reduces the consumption of computing resources, but also avoids the problem of model overfitting caused by overly precise calculations.

[0219] 4. Improved the rationality of dose prediction. By limiting the physical range of GAN prediction, the generative adversarial network model based on the beam channel can generate a more reasonable dose distribution that conforms to physical laws. This helps to ensure the safety and effectiveness of radiotherapy plans and reduce adverse effects on patient health.

[0220] 5. It promotes the development of personalized treatment. The beam dose with a larger weight is calculated based on the patient's medical images, and the results are input into the model, making the dose prediction more personalized. This personalized prediction helps doctors develop more accurate and effective treatment plans for patients, improving treatment effects and patient satisfaction.

[0221] 6. Accelerated the process of clinical application. The generative adversarial network model based on beam channels solves the problem of mixed data sets in dose prediction, making the technology more convenient and efficient in clinical applications. This helps to accelerate the popularization and promotion of radiotherapy technology and improve the overall level of medical services.

[0222] 7. It has promoted the application of AI in the medical field and demonstrated the great potential of artificial intelligence technology in the medical field, especially in the application of radiotherapy dose prediction. It has promoted the deep integration of AI technology and medical practice and provided new ideas and methods for future medical innovation and development.

[0223] In summary, by introducing the beam channel-based generative adversarial network model, significant progress has been made in the field of dose prediction, which not only improves the accuracy and rationality of the prediction, but also promotes the development of personalized treatment, accelerates the process of clinical application, and promotes the widespread application of AI in the medical field.

[0224] A nasopharyngeal carcinoma radiotherapy planning system based on deep learning and threshold optimization, the system applies any of the above-mentioned nasopharyngeal carcinoma radiotherapy planning methods based on deep learning and threshold optimization, and the system includes a server and a processor;

[0225] The server is used to obtain CT images of patients and obtain dose prescription data;

[0226] The processor is used to identify the planned target area and the protection area in the CT image, and segment the planned target area and the protection area respectively; determine the predicted three-dimensional dose and the predicted dose volume histogram based on the segmented planned target area and the protection area; the system is used to reversely optimize the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data to determine the radiotherapy plan.

[0227] Example:

[0228] At the Xiangnan College Affiliated Hospital, we successfully applied the beam channel-based generative adversarial network model to the dose prediction of 358 patients undergoing radiotherapy. This example not only verified the feasibility and effectiveness of the beam channel-based generative adversarial network model in clinical practice, but also provided valuable experience for the further promotion and application of the model.

[0229] The implementation results are as follows:

[0230] The prediction accuracy is high. The generative adversarial network model based on the beam channel performed well in dose prediction for 358 patients. The difference between the predicted dose and the actual dose was small, and indicators such as the mean square error and correlation coefficient were better than traditional methods.

[0231] Strong adaptability: The model can effectively process mixed data sets (such as IMRT and VMAT), showing strong adaptability and generalization capabilities.

[0232] High computational efficiency. Compared with complex pre-dose calculation methods, the pre-dose calculation strategy adopted by the generative adversarial network model based on the beam channel is simpler and more efficient, reducing the consumption of computing resources.

[0233] The clinical application effect is good, and the prediction results have been recognized by clinicians, providing strong support for the formulation of radiotherapy plans. At the same time, the model also helps to achieve personalized treatment, improve treatment effects and patient satisfaction.

[0234] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a computer-readable storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device to enter the method described in each embodiment of the present invention.

[0235] In the description of this specification, the description with reference to the terms "an embodiment", "another embodiment", "other embodiments", or "first embodiment to Xth embodiment" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, method steps or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0236] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0237] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0238] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A radiotherapy planning method for nasopharyngeal carcinoma based on deep learning and threshold optimization, characterized in that: include: Obtain CT images of patients; Identifying a planned target area and a protection area in a CT image, and segmenting the planned target area and the protection area respectively; Determining a predicted three-dimensional dose and a predicted dose volume histogram according to the segmented planned target area and the protection area; Obtain dosage prescription data; The predicted three-dimensional dose and the predicted dose volume histogram are respectively reversely optimized according to the dose prescription data to determine a radiotherapy plan.

2. The nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization according to claim 1, characterized in that: The predicted three-dimensional dose is calculated as follows: The calculation is performed by a pencil beam convolution stacking algorithm, and the calculation formula of the pencil beam convolution stacking algorithm is as follows: Where ρ is the density; Ψ is the probability density function distribution of the incident source, i.e., the fluence; x, y, and z are coordinates; x′ is the integral coordinate of x; y′ is the integral coordinate of y; z′ is the integral coordinate of z; D(x, y, z) is expressed as a three-dimensional dose, which is used to limit the physical conditions of the beam channel; K w is a pencil beam; F is a three-dimensional dose distribution function calculated by the pencil beam convolution superposition algorithm, and F describes the dose size at different coordinate positions (x, y, z).

3. The nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization according to claim 2, characterized in that: The pencil beam calculation method includes: Finite pencil beams were calculated in a water phantom using the Monte Carlo method; The pencil beam calculation method further includes: Determining dose distribution data of the pencil beam in a water phantom according to a limited pencil beam; The dose impact matrix was determined based on the dose distribution data of the pencil beam in the water phantom.

4. The nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization according to claim 3, characterized in that: After the step of determining the predicted three-dimensional dose and the predicted dose volume histogram according to the segmented planned target area and the protection area, the method further comprises: Based on the formula of the pencil beam convolution superposition algorithm, a three-dimensional dose map is determined by a summation operation of the dose impact matrix.

5. The nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization according to any one of claims 1 to 4, characterized in that: The step of identifying the planned target area and the protection area in the CT image and segmenting the planned target area and the protection area respectively comprises: Identify the planned target area and protection area in the CT image; The planned target area is divided into a plurality of sub-areas of preset sizes.

6. The method for nasopharyngeal carcinoma radiotherapy planning based on deep learning and threshold optimization according to claim 5, characterized in that: After the step of dividing the planned target area into a plurality of sub-areas of preset sizes, the method further comprises: Get beam channel; Get three image channels for any of the sub-regions: According to the firing rate channel and the three image channels, dose images of the sub-areas corresponding to the three image channels are determined.

7. The nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization according to claim 6, characterized in that: The firing rate channel generation method comprises: extracting preset information through the planned target area; According to historical cases, obtain the beam angle that meets the preset information; Projecting the beam angle onto the planned target volume to obtain the grid weights of the beam channel data; Select the ray directions with the grid weights arranged in the front column as the firing rate channels.

8. The method for nasopharyngeal carcinoma radiotherapy planning based on deep learning and threshold optimization according to any one of claims 1 to 4, characterized in that: The dosage prescription data includes: Dose maximum objective function, dose average objective function, equivalent uniformity objective function, deviation from prescription dose degree objective and smoothness objective function.

9. The method for radiotherapy planning for nasopharyngeal carcinoma based on deep learning and threshold optimization according to claim 8, characterized in that: The step of performing reverse optimization on the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data to determine the radiotherapy plan includes: The step of performing reverse optimization on the predicted three-dimensional dose and the predicted dose volume histogram respectively according to the dose prescription data; Calculate the voxel dose unit based on the inverse optimization results and the dose calculation engine; According to the voxel dose unit, optimizing a flux weight map; The flux weight map and the machine information of the accelerator are used to optimize the leaf sequence using a leaf sequence algorithm, and a radiotherapy plan is determined according to the leaf sequence.

10. A nasopharyngeal carcinoma radiotherapy planning system based on deep learning and threshold optimization, wherein the system applies the nasopharyngeal carcinoma radiotherapy planning method based on deep learning and threshold optimization described in any one of claims 1 to 9, characterized in that: The system includes a server and a processor; The server is used to obtain CT images of patients; obtain dosage prescription data; The processor is used to identify the planned target area and the protection area in the CT image, and segment the planned target area and the protection area respectively; Determining a predicted three-dimensional dose and a predicted dose volume histogram according to the segmented planned target area and the protection area; The system is used to perform reverse optimization on the predicted three-dimensional dose and the predicted dose volume histogram according to the dose prescription data to determine the radiotherapy plan.

Images