A radiotherapy dose prediction method and system based on a generative adversarial network

Abstract

The application discloses a radiotherapy dose prediction method and system based on a generative adversarial network, belongs to the technical field of radiotherapy dose prediction, and comprises the following steps: acquiring a training set; obtaining the projection area of a beam angle in a target region; acquiring a plurality of beam angles with large projection areas, wherein the plurality of beam angles form a beam channel; acquiring a three-dimensional dose based on the beam channel; taking the three-dimensional dose as a fourth channel of a planned dose map to obtain an input map; training the input map and a real dose map based on a generative adversarial network to obtain a prediction model; and analyzing target region planning data to be predicted based on the prediction model to obtain a predicted radiotherapy dose. The three-dimensional dose of the beam channel is taken as the fourth channel and is combined into the planned dose map, and a limitation is added based on a real physical condition, so that the generative adversarial network can be used for mixed training sets of different types of radiotherapy techniques.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy dose prediction technology, and specifically to a radiotherapy dose prediction method and system based on generative adversarial networks. Background Technology

[0002] Dose prediction in radiotherapy has been extensively studied. Generally, prediction models employ backpropagation-based neural networks and deep convolutional neural networks. Backpropagation-based neural networks are trained by finding the relationship between the dose of each voxel in the at-risk organ and the patient's geometric anatomy. Features that may affect the dose are used as input to the backpropagation neural network, and the corresponding voxel dose points are used as the model's output. The network is then trained using the backpropagation algorithm. However, 3D dose distribution prediction methods based on backpropagation neural networks require manual feature extraction, leading to strong subjectivity in feature selection. The emergence of deep convolutional neural networks (DNNs) for 3D dose distribution prediction has solved this problem, such as U-net, DenseNet, and HD U-net.

[0003] In recent years, scholars have successfully proposed using the Pix2pix model based on Generative Adversarial Networks (GANs) for dose prediction in external beam radiotherapy. Here, Pix2pix is ​​a GAN-based image translation model that can predict 3D dose distribution based on the patient's geometric anatomy and organ prediction model. For example, using GANs to predict the dose of laryngeal cancer, this study trained a GAN model on 219 cases of laryngeal cancer using a technique called intensity-modulated radiotherapy (IMRT), achieving good results. Chinese scholars, including Fan Jiawei et al., also used a single technique from IMRT cases to train a prediction model, achieving meaningful results. Furthermore, some scholars have used IMRT technology, employing datasets of 42 cervical cancer cases and 130 rectal cancer cases, to propose an end-to-end multi-constraint GAN framework (Mc-GAN) for automatic dose prediction to improve prediction accuracy.

[0004] Clinically used techniques include intensity-modulated radiotherapy (IMRT) and volumetric modulated radiotherapy (VMAT), which have specific physical conditions. Current technologies struggle to differentiate between datasets of two or more techniques with varying physical conditions. While GANs have shown success with single-technique datasets, collecting data on similar techniques in clinical settings is difficult, posing challenges for clinical application. Furthermore, datasets generated and mixed from different types of radiotherapy techniques can lead to dose prediction failures. Summary of the Invention

[0005] To address the aforementioned technical problems in the existing technology, this invention provides a radiotherapy dose prediction method and system based on generative adversarial networks. By restricting the training set through beam channels, a training set formed by two or more technologies can be utilized, which helps to increase the data sources for the training set.

[0006] This invention discloses a radiotherapy dose prediction method based on generative adversarial networks (GANs). The prediction method includes: acquiring a training set, wherein the training set includes target planning data and a true dose map, wherein the target planning data includes a planned dose map, a corresponding beam angle, and a target region; obtaining the projected area of ​​the beam angle in the target region; acquiring multiple beam angles with large projected areas, wherein the multiple beam angles form a beam channel; acquiring a three-dimensional dose based on the beam channel; using the three-dimensional dose as the fourth channel of the planned dose map to obtain an input map; training the input map and the true dose map based on a GAN to obtain a prediction model; and analyzing the target planning data to be predicted based on the prediction model to obtain the predicted radiotherapy dose.

[0007] Preferably, the three-dimensional dose is obtained based on any of the following dose calculation methods: pencil beam convolution superposition method, folded convolution superposition method, and Monte Carlo dose calculation method.

[0008] Preferably, the calculation formula for the pencil-beam convolution stacking method is expressed as follows:

[0009]

[0010] Where D(x,y,z) represents the three-dimensional dose, x, y, and z are coordinates, x′ and y′ are integral coordinates, F represents the integration region, ρ is the density, Ψ is the probability density function distribution of the incident source, and K w It is a pencil-shaped beam.

[0011] Preferably, the method for obtaining the beam channel includes:

[0012] The weight of the beam angle is calculated using the following formula:

[0013]

[0014] Among them, w i The weight a is represented as the beam angle i. i It represents the projected area of ​​the field of fire at beam angle i in the target region, and n represents the total number of beam angles;

[0015] The beam angles are sorted from largest to smallest according to the weights.

[0016] Take the first M beam angles as the beam channels, where M is a natural number.

[0017] Preferably, the generative adversarial network includes a generator and a discriminator.

[0018] The generator is used to generate a predicted radiotherapy dose based on the input graph;

[0019] The discriminator is used to determine whether the predicted radiotherapy dose is true.

[0020] Preferred methods for training the input map and the true dose map include:

[0021] A discriminant sub-model is obtained by training the generator with the predicted radiotherapy dose and the actual dose map / planned dose map generated by the generator.

[0022] The generator is used to train the discriminator's judgment results and the input graph to obtain a generator sub-model.

[0023] Preferably, the generator comprises an 8-level hierarchical U-net network;

[0024] The U-net network includes a feature extraction part and an upsampling part.

[0025] Each layer of the feature extraction part performs a 3x3 convolution operation on the feature map of the previous layer, followed by a 2x2 max pooling operation to obtain the feature map of the corresponding layer.

[0026] Each layer of the upsampling part includes performing a 3x3 convolution operation on the feature map of the previous layer, followed by a 2x2 deconvolution operation; to obtain the feature map of the corresponding layer.

[0027] Preferably, the discriminator divides the input image into multiple blocks; performs a true / false judgment on each block; and takes the average of the results of all blocks as the output.

[0028] Preferably, data preprocessing methods include:

[0029] Collect cervical cancer data, wherein the radiotherapy plan for the cervical cancer data includes any of the following techniques or a combination thereof: intensity-modulated radiotherapy plan, volumetric intensity-modulated radiotherapy plan, and conformal radiotherapy plan;

[0030] The cervical cancer data was cleaned, and cases with organ incompleteness were deleted.

[0031] After cleaning, the dataset is randomly divided into training set, validation set, and test set;

[0032] The names and aliases of the regions of interest in the cervical cancer data are standardized.

[0033] Preferred methods for predicting radiotherapy dose assessment include:

[0034] If the predicted radiotherapy dose to the target area meets the prescription dose requirement, the score is S, and S is greater than 0.

[0035] If the predicted radiotherapy dose to the target area cannot meet the prescribed dose requirement, the score is 0;

[0036] The scores for each target area are combined to obtain the total score for predicting the radiotherapy dose.

[0037] The present invention also provides a system for implementing the above-described radiotherapy dose prediction method, comprising an acquisition module, a beam channel analysis module, an input map construction module, a training module, and a prediction module.

[0038] The acquisition module is used to acquire the training set;

[0039] The beam channel analysis module is used to obtain the projected area of ​​the beam angle in the target area; and to obtain multiple beam angles with large projected areas, the multiple beam angles forming a beam channel;

[0040] The input map construction module is used to obtain a three-dimensional dose based on the beam channel; and to use the three-dimensional dose as the fourth channel of the planned dose map to obtain the input map;

[0041] The training module is used to train a prediction model based on a generative adversarial network on the input map and the real dose map.

[0042] The prediction module is used to analyze the target area planning map to be predicted based on the prediction model to obtain the predicted radiotherapy dose.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: the beam channel has a larger projection area, that is, it has a larger dose calculation weight; the three-dimensional dose of the beam channel is incorporated into the planned dose map as a fourth channel, and the constraints are added based on real physical conditions, so that the generative adversarial network can be used for mixed training sets of different types of radiotherapy techniques. Attached Figure Description

[0044] Figure 1 This is a flowchart of the radiotherapy dose prediction method based on generative adversarial networks of the present invention;

[0045] Figure 2 This is a schematic diagram of the beam channel;

[0046] Figure 3 This is a schematic diagram of a generative adversarial network;

[0047] Figure 4 This is a flowchart of a method for calculating the three-dimensional dose of a beam channel based on the pencil-beam convolution superposition method;

[0048] Figure 5 This is a schematic diagram of the generator's U-net network;

[0049] Figure 6 This is a network diagram of the discriminator;

[0050] Figure 7 It is a comparison chart of predicted dose, original dose, and CT.

[0051] Figure 8 It is a training loss curve;

[0052] Figure 9 This is the system logic block diagram of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The present invention will now be described in further detail with reference to the accompanying drawings:

[0055] A radiotherapy dose prediction method based on generative adversarial networks, such as Figure 1 As shown, the prediction method includes:

[0056] Step 101: Obtain the training set, which includes case / training data such as target volume planning data (PTV) and true dose maps. The target volume planning data includes the planned dose map, as well as related beam angles, computed tomography (CT) images, organ delineation (organs at risk and target area outlines / target regions), etc.

[0057] Step 102: Obtain the projected area of ​​the beam angle in the target region. In one specific embodiment, the beam falling within the target region is set to 1, and the beam falling at other locations is set to 0.

[0058] Step 103: Obtain multiple beam angles with large projected areas, which together form a beam channel (Bc). For example, 9-20 beam angles with large projected areas can be selected. It should be noted that the computational load of the beam channel will increase.

[0059] Step 104: Obtain the three-dimensional dose based on the beam channel. The three-dimensional dose can be calculated using methods such as pencil beam convolution stacking (PBC), folded convolution stacking, or Monte Carlo dose calculation.

[0060] Step 105: Use the three-dimensional dose as the fourth channel of the planned dose map to obtain the input map. That is, the input map has 4 channels.

[0061] Step 106: Train the prediction model based on the input map and the real dose map using a generative adversarial network (GAN).

[0062] Step 107: Based on the prediction model, analyze the target area planning data to be predicted to obtain the predicted radiotherapy dose.

[0063] Step 108: Evaluate the predicted radiotherapy dose.

[0064] The beam channel has a large projected area, meaning it has a large dose calculation weight. Incorporating the three-dimensional dose of the beam channel as a fourth channel into the planned dose map adds constraints based on real physical conditions, allowing the generative adversarial network (GAN) to be used on mixed training sets of different types of radiotherapy techniques. These radiotherapy techniques can include intensity-modulated radiotherapy (IMRT), volumetric intensity-modulated radiotherapy (VMAT), and conformal radiation therapy (CRT). In this invention, the beam channel-based adversarial network is simply referred to as Bc-GAN.

[0065] Example 1

[0066] The process of predicting radiotherapy dose is illustrated using cervical cancer as an example.

[0067] Step 201: Data Acquisition and Preprocessing. 351 cases of stage I cervical cancer were collected for testing; after data cleaning, 56 cases of organ dysfunction were removed, leaving 295 cases (IMRT:VMAT = 135:160); the training, validation, and test sets were randomly selected in a ratio of approximately 7:2:1 (219:52:24). Here, the mixed datasets were exported by a physicist using the Raystation system in DICOM format. Due to the sensitive nature of the medical information, these mixed datasets are not available to the public. The use of patient data has been approved by the hospital's Review and Oversight Committee and Institutional Review Committee, and all patient data has been fully anonymized.

[0068] The names and aliases of regions of interest (ROIs) in the cervical cancer data were standardized. A unified naming dictionary was created to extract information by standardizing the names and aliases of ROIs for organs and target areas. Each 3D CT image and its planned dose were segmented into 2D slices of 256×256 pixels. A single slice was used to predict the dose distribution along the same plane; this process was repeated for each slice until a complete 3D dose distribution was produced. The training and validation sets consisted of all 2D slices of 3D CT images from 219 and 52 patients, respectively; CT images from the remaining 25 patients were used as test data for dose prediction.

[0069] Step 202: Three-dimensional dose calculation of the beam channel.

[0070] The weight of the beam angle is calculated by projecting the beam angle onto the target region:

[0071]

[0072] Among them, w i The weight a is represented as the beam angle i. i Let w represent the projected area of ​​the target region at beam angle i, and n represent the total number of beam angles. For a specific target area plan, the total projected area of ​​all beam angles is a fixed value, therefore w i With a i Proportional. Figure 2 The specific beam channel is shown.

[0073] The beam angles can be sorted from largest to smallest according to their weights; the top M beam angles are selected as beam channels, where N is a natural number, such as 9. That is, the top M beam angles with the largest weights are selected as beam channels.

[0074] The pencil-beam convolution stacking (PBC) method uses the following calculation formula:

[0075]

[0076] Where D(x,y,z) represents the three-dimensional dose, x, y, and z are coordinates, x′ and y′ are integration coordinates, F represents the integration region, ρ is the density, Ψ is the probability density function distribution of the incident source, and Kw is the pencil beam. For example, F=[x'-x0,y'-y0,z-z0],[x'+x0,y'+y0,z'+z0] means that integration is performed in this region.

[0077] like Figure 4 As shown, the calculation of the three-dimensional dose of the beam channel involves the following four steps:

[0078] Step 221: Calculate the pencel beam kernel. The finite pencel beam Kw is calculated in the water phantom using the Monte Carlo method.

[0079] Step 222: Summation of beam kernels. Based on the discrete form of Equation 2, the 3D dose map is calculated by summing the beam kernels.

[0080] Step 223: 3D energy spectrum correction. Assuming 6MV photons are used in radiotherapy, since the beam channel does not require a precise energy spectrum, the calculations in the above steps can be corrected using the corresponding 6MV Varian photon energy spectrum.

[0081] Step 224: 3D nonuniform correction. 3D nonuniform correction of the human body is performed based on the density relationships of CT scans, without using specific CT density relationships.

[0082] Step 203: Prediction model training.

[0083] like Figure 3 As shown, the Beam Channel-Based Generative Adversarial Network (Bc-GAN) comprises two networks: a generator (G) and a discriminator (D). The generator generates a predicted radiotherapy dose based on the input image, while the discriminator determines whether the predicted radiotherapy dose is real. For example, after inputting a random code / image, G outputs a fake image. D accepts the fake image output by G as input and determines whether the image is real or fake; it outputs 1 for real and 0 for fake. During the game between the two networks, their abilities increase: the images generated by G become increasingly similar to real images, and D becomes increasingly able to determine the authenticity of images. Finally, while maximizing the ability of D, the ability of D to judge the authenticity of G is minimized. The generator is based on the Pix2pix GAN architecture, specifically using a U-net structure.

[0084] Specifically, a discriminator is trained on the predicted radiotherapy dose generated by the generator, as well as the actual dose map / planned dose map, to obtain a discriminator sub-model; the generator is trained on the discriminator's judgment results and the input map to obtain a generator sub-model. The discriminator sub-model and the generator sub-model constitute the prediction module.

[0085] The planned dose map Y is used as imgA, and the three-dimensional dose map imgC of the beam channel is synthesized into an input map. The input map is processed by G to generate the predicted dose fake x (imgB'). During the training phase, D is trained using the real dose map real x (imgB) and the predicted dose fake x. During the validation phase, the real dose map real x (imgB) is used to verify the accuracy of the prediction model and calculate the error. After training, the real dose map real x (imgB) is no longer needed for the target area planning data to be predicted.

[0086] During the prediction phase, the predicted dose fake x and the planned dose map imag A are fed into the discriminator D for judgment. The difference between the predicted dose and the actual dose is judged, and the judgment result is output. If the output is 1 / true, the predicted dose is the final predicted dose. If the output is 0 / false, the predicted dose is an invalid predicted dose, and the calculation is terminated.

[0087] Figure 5 An 8-level hierarchical U-net network of the generator is shown, which includes a feature extraction part and an upsampling part.

[0088] Each layer of the feature extraction part performs a 3x3 convolution operation on the feature map of the previous layer, followed by a 2x2 max pooling operation to obtain the feature map of the corresponding layer, ultimately reducing the feature size from 256x256 pixels to 1x1 pixels.

[0089] Each layer of the upsampling part includes performing a 3x3 convolution operation on the feature map of the previous layer, followed by a 2x2 deconvolution operation to obtain the feature map of the corresponding layer; restoring the 1x1 pixel feature to 256x256 pixels, and the final output image is a 256x256x1 predicted dose map.

[0090] During the training phase, the Adam algorithm is used as the optimizer to minimize the loss function. The learning rate is set to 2×10⁻⁶. -3 The number of epochs is 1000.

[0091] Figure 6 The discriminator network is shown. The discriminator divides the input image into multiple N x N patches; it determines whether each patch is true or false; and it takes the average of the results from all patches as the output. For example, for a 256x256 input, the best results are obtained when the patch size is 70x70.

[0092] The specific training computer environment was an Ubuntu 20.04 system with a CPU of [CPU details missing]. The system used a Core™ i3-8350K at 4.00GHz, a GTX 1080 Ti graphics card, Python 3.6, and TensorFlow 1.0. The training set contained 219 training samples, spanning 40 epochs (cycles on a single case) and 131,980 iterations (batch size per training iteration). The process took approximately 48 hours. Notably, no further parameter tuning was required to achieve good results, whereas previous deep learning methods might have required significantly more time to fine-tune these parameters to reach satisfactory accuracy. Figure 8 The loss curve during training is shown, exhibiting a significant decrease before reaching 700 iterations, followed by a gradual decrease. After 40 iterations, the curve converges smoothly, and further training results in minimal fluctuations in the loss curve.

[0093] Step 204: Assessment of predicted radiotherapy dose.

[0094] Methods for predicting radiotherapy dose assessment include:

[0095] If the predicted radiotherapy dose to the target area meets the prescribed dose requirement, a score of S is awarded, where S is greater than 0 (e.g., 12.5). If the predicted radiotherapy dose to the target area does not meet the prescribed dose requirement, a score of 0 is awarded. Missing organs are assigned a default score of S. The scores for each target area are then combined to obtain the total predicted radiotherapy dose score.

[0096] The total scores of predicted and manually predicted radiotherapy doses were compared among the 24 cases in the test set. Six cases had excellent predicted radiotherapy doses, 14 had comparable doses, and 4 had slightly worse predicted doses. Overall, predicted radiotherapy doses are significantly superior to manually predicted doses.

[0097] Figure 7 A comparative example of predicted dose (I, J, K, L), original dose (E, F, G, H), and CT images (A, B, C, D) is shown, in which... Figure 7 A, E, and I are the same organ. Figure 7 B, F, and J refer to the same organ. Figure 7 C, G, and K refer to the same organ. Figure 7 D, H, and L represent the same organ, indicating that the predicted dose and the original dose are quite close.

[0098] Example 2

[0099] This embodiment provides a system for implementing the above-described radiotherapy dose prediction method, such as... Figure 9 As shown, it includes an acquisition module 1, a beam channel analysis module 2, an input graph construction module 3, a training module 4, and a prediction module 5.

[0100] Acquisition module 1 is used to acquire the training set;

[0101] The beam channel analysis module 2 is used to obtain the projected area of ​​the beam angle in the target region; and to obtain multiple beam angles with large projected areas, which together form a beam channel; the input map construction module 3 is used to obtain a three-dimensional dose based on the beam channel; and to use the three-dimensional dose as the fourth channel of the planned dose map to obtain an input map; the training module 4 is used to train the input map and the real dose map based on a generative adversarial network to obtain a prediction model; the prediction module 5 is used to analyze the target area planning map to be predicted based on the prediction model to obtain the predicted radiotherapy dose.

[0102] The radiotherapy dose prediction method of the present invention can be prepared as a program to implement the radiotherapy dose prediction method, and the program can be saved to a storage medium or computer. The present invention only discloses the radiotherapy dose prediction method, and the predicted radiotherapy dose can provide auxiliary decision-making for physicians.

[0103] This invention discloses a radiation dose prediction method based on the Bc-GAN model, addressing the issue of datasets for different radiotherapy techniques, specifically utilizing a mixed dataset of cervical cancer data. In dose prediction for cervical cancer, it was found that I MRT and VMAT in radiotherapy can be used as mixed data input models, rather than being limited to a single radiotherapy technique. This invention offers the possibility of realizing large-scale engineering and commercial applications of artificial intelligence. This invention not only reduces the amount of training data but also accelerates the generation of prediction models. This invention has the potential to extend dose prediction to more tumor types.

[0104] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting radiotherapy dose based on generative adversarial networks, characterized in that, The prediction method includes: Obtain a training set, the cases in the training set including target planning data and actual dose maps, the target planning data including the planned dose map, as well as the corresponding beam angle and target area; Obtain the projected area of ​​the beam angle in the target region; Obtain multiple beam angles with a large projection area, and the multiple beam angles form a beam channel; Based on the beam channel, a three-dimensional dose is obtained; The three-dimensional dose is used as the fourth channel of the planned dose map to obtain the input map; Based on a generative adversarial network, a prediction model is obtained by training on the input map and the actual dose map; wherein, the aforementioned generative adversarial network is used for a mixed training set of different types of radiotherapy techniques. Based on the prediction model, the target area planning data to be predicted is analyzed to obtain the predicted radiotherapy dose; The methods for obtaining the beam channel include: The weight of the beam angle is calculated using the following formula: ； in, w i Expressed as beam angle i The weight, a i Represented as the beam angle i The projected area of ​​the firing field in the target area. n Expressed as the total beam angle; The beam angles are sorted from largest to smallest according to the weights. Take before M A beam angle, serving as the beam channel, in which M It is a natural number; It also includes methods for data preprocessing and predicting radiotherapy dose assessment: Data preprocessing methods include: Collect cervical cancer data, wherein the radiotherapy plan for the cervical cancer data includes any of the following techniques or a combination thereof: intensity-modulated radiotherapy plan, volumetric intensity-modulated radiotherapy plan, and conformal radiotherapy plan; The cervical cancer data was cleaned, and cases with organ incompleteness were deleted. After cleaning, the dataset is randomly divided into training set, validation set, and test set; Standardize the names and aliases of the regions of interest in the cervical cancer data; Methods for predicting radiotherapy dose assessment include: If the predicted radiotherapy dose to the target area meets the prescription dose requirement, the score is S, where S is greater than 0; If the predicted radiotherapy dose to the target area cannot meet the prescribed dose requirement, the score is 0; The scores for each target area are combined to obtain the total score for predicting the radiotherapy dose.

2. The radiotherapy dose prediction method according to claim 1, characterized in that, The three-dimensional dose is obtained based on any of the following dose calculation methods: Pencil-beam convolution superposition method, folded convolution superposition method, and Monte Carlo dose calculation method.

3. The radiotherapy dose prediction method according to claim 2, characterized in that, The calculation formula for the pencil-beam convolution superposition method is expressed as follows: ； in, Represented as three-dimensional dose, x, y, z As coordinates, For the integral coordinates, F Represented as the integration region, ρ For density, Ψ Let be the probability density function distribution of the incident source. K w It is a pencil-shaped beam.

4. The radiotherapy dose prediction method according to claim 1, characterized in that, Generative adversarial networks consist of a generator and a discriminator. The generator is used to generate a predicted radiotherapy dose based on the input graph; The discriminator is used to determine whether the predicted radiotherapy dose is true.

5. The radiotherapy dose prediction method according to claim 4, characterized in that, Methods for training on the input map and the true dose map include: A discriminant sub-model is obtained by training the generator with the predicted radiotherapy dose and the actual dose map / planned dose map generated by the generator. The generator is used to train the discriminator's judgment results and the input graph to obtain a generator sub-model. The discriminator sub-model and the generator sub-model constitute a prediction model.

6. The radiotherapy dose prediction method according to claim 4, characterized in that, The generator comprises an eight-level hierarchical U-net network; The U-net network includes a feature extraction part and an upsampling part. Each layer of the feature extraction part performs a 3 x 3 convolution operation on the feature map of the previous layer, followed by a 2 x 2 max pooling operation to obtain the feature map of the corresponding layer. Each layer of the upsampling part includes performing a 3 x 3 convolution operation on the feature map of the previous layer, followed by a 2 x 2 deconvolution operation; to obtain the feature map of the corresponding layer.

7. The radiotherapy dose prediction method according to claim 4, characterized in that, The discriminator divides the input image into multiple blocks; performs a true / false judgment on each block; and takes the average of the results of all blocks as the output.

8. A system for implementing the radiotherapy dose prediction method as described in any one of claims 1-7, characterized in that, It includes an acquisition module, a beam channel analysis module, an input graph construction module, a training module, and a prediction module. The acquisition module is used to acquire the training set; The beam channel analysis module is used to obtain the projected area of ​​the beam angle in the target area; and to obtain multiple beam angles with large projected areas, the multiple beam angles forming a beam channel; The input graph construction module is used to obtain three-dimensional dose based on the beam channel; The three-dimensional dose is then used as the fourth channel of the planned dose map to obtain the input map; The training module is used to train a prediction model based on a generative adversarial network on the input map and the real dose map. The prediction module is used to analyze the target area planning map to be predicted based on the prediction model to obtain the predicted radiotherapy dose.

Images