Radiotherapy plan dose distribution verification method based on deep learning

Abstract

The invention relates to the technical field of deep learning, in particular to a radiotherapy plan dose distribution verification method based on deep learning, and the method comprises the following steps: collecting historical radiotherapy plan data, generating a physical reference dose field through a Monte Carlo algorithm, unifying the voxel resolution of an anatomical structure to 1 cubic millimeter, and normalizing the dose according to a prescription, data enhancement is carried out only by adopting translation and mirror transformation, trace Gaussian noise is added, and a physical information enhanced three-dimensional training data set is constructed. According to the method, a three-dimensional convolutional network is utilized to automatically learn a dose distribution rule of a historical high-quality plan, a physical constraint module is embedded to ensure that a prediction result accords with a radiology principle, a real-time clinical rule engine is combined to instantly identify and correct a violation hot spot cold region, and an uncertainty quantification technology is assisted to position a high-risk region, so that the accuracy of a prediction result is improved. Finally, minute-level full-automatic verification is achieved, executable optimization suggestions are output, and efficiency is improved by dozens of times while safety is improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a method for verifying radiotherapy plan dose distribution based on deep learning. Background Technology

[0002] A radiotherapy plan is a detailed treatment plan developed before radiotherapy. It uses medical imaging to determine the spatial location of the tumor target area and surrounding organs at risk, and calculates the radiation beam parameters based on the principles of radiation physics. It plans a radiation dose distribution scheme that can effectively kill tumor cells while maximally protecting normal tissues. The reason for verifying the dose distribution of the radiotherapy plan is to ensure that the actual irradiation dose is highly consistent with the plan design. Any deviation may lead to failure of tumor control or irreversible damage to normal tissues. Especially in complex techniques such as intensity-modulated radiotherapy, the dose distribution pattern is variable. It is necessary to strictly verify to eliminate the risks caused by machine execution errors, limitations of calculation models, and changes in patient anatomy. Currently, the existing publication CN 110085298 B describes a method for predicting the three-dimensional dose distribution of intensity-modulated radiotherapy (IMRT) plans based on deep learning. This method includes: collecting effective IMRT planning data to form a case database; extracting the three-dimensional anatomical contour features of the region of interest (ROI) for each patient from the case database; dividing the three-dimensional anatomical contour of the ROI into several two-dimensional contour slices; extracting the dose features of each patient from the case database and dividing them into several two-dimensional dose slice distribution maps; constructing a deep convolutional network, inputting the patient's two-dimensional contour slices and the corresponding two-dimensional dose slice distribution maps, and obtaining a correlation model between the anatomical contour features and dose features through model training; and using the trained correlation model to predict the three-dimensional dose distribution of new patients. While this method can effectively obtain the correlation between anatomical features and dose features, improving the accuracy of dose prediction, it is time-consuming (lasting several hours to days) and can only sample a portion of the slices, failing to comprehensively assess the three-dimensional dose field. Furthermore, manual inspection is susceptible to subjective experience and may miss subtle abnormalities, lacking the ability to quantitatively predict risks in complex cases.

[0003] In summary, a deep learning-based method for verifying radiotherapy planning dose distribution is needed to address the aforementioned issues. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for verifying radiotherapy planning dose distribution, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention proposes a deep learning-based method for verifying radiotherapy plan dose distribution, comprising: S1. Collect historical radiotherapy plan data, generate a physical reference dose field using the Monte Carlo algorithm, unify the resolution of anatomical structure voxels to 1 cubic millimeter, normalize the dose according to the prescription dose, use only translation and mirror transformation for data augmentation, and add a small amount of Gaussian noise to construct a three-dimensional training dataset with enhanced physical information. S2. Construct a physically constrained 3D convolutional network, embed pen-shaped beam energy kernel convolutional modules in a U-shaped architecture to enhance physical consistency, use dense connection blocks to preserve detailed features, and design a hybrid loss function that incorporates clinical dose constraints; S3. Input the new patient's three-dimensional anatomical data into the training model to predict the dose distribution. After restoring the original resolution through trilinear interpolation, apply anisotropic Gaussian filtering to eliminate discontinuities between slices. S4. Establish a real-time clinical rule engine to automatically detect insufficient target coverage or organ overload violations, and trigger the built-in optimizer to iteratively correct the problem areas; S5. Monte Carlo stochastic inference is used to quantify the prediction uncertainty and generate heat maps of high-risk areas to mark the voxel locations that exceed the clinical tolerance threshold. S6. Introduce a gradient inversion layer to obfuscate data source features, and combine it with a generative adversarial network to synthesize anatomical variant samples, thereby improving the generalization ability across medical institutions; S7. Deploy dual-channel anomaly monitoring, which detects input data anomalies through contour segmentation verification and dose reconstruction error detection, automatically intercepts problematic cases and transfers them to manual review; S8. Establish a fully automated clinical workflow, issue reports for validated protocols, generate optimization suggestions for unvalidated protocols and provide feedback to physicians, and continuously track pass rates and timeliness indicators to achieve closed-loop optimization.

[0006] Preferably, the implementation steps of step S1 are as follows: S1.1. Collect historical radiotherapy plans containing CT images, three-dimensional dose distribution, and irradiation parameters. For each case, use the Monte Carlo algorithm to generate a baseline physical dose field. ; S1.2. Data Standardization: The resolution of anatomical structure voxels is standardized to [specific value]. Dosage as prescribed The interval is shown in equation (1): (1); S1.3. The enhancement strategy only allows translation and mirror transformations, with the offset as follows: ; S1.4. Add Gaussian noise to the dose field, standard deviation Output the training set: , C represents the number of anatomical channels.

[0007] Preferably, the implementation steps of step S2 are as follows: The network adopts an encoder-decoder architecture, including the input layer size. The encoder contains 4 steps with a step size of 2. The convolutional downsampling layer and the decoder contain corresponding upsampling layers. The core components include a physical constraint module, dense connection blocks, and a loss function. The physical constraint module includes: ,in for Pencil beam energy deposition core, For convolution, Hadamard product; densely connected blocks: 4 layers of convolution (kernel) Number of channels Each layer's output is concatenated with its input; the loss function is shown in equation (2): Loss (2).

[0008] Preferably, the implementation steps of step S3 are as follows: S3.1. Input the new patient's CT scan and contour data into PC-3DUnet, and output the three-dimensional dose tensor. ; S3.2. Post-processing: Trilinear interpolation recovery Original CT resolution, anisotropic Gaussian filtering to eliminate discontinuities between slices. .

[0009] Preferably, the implementation steps of step S4 are as follows: constructing a rule engine to examine clinical indicators, with target area coverage if... of This triggers a "PTV insufficient coverage" alert; organ availability is limited. [Spinal cord] If an error occurs, a "spinal cord hotspot exceeding the limit" alarm is triggered; if this occurs, the optimizer iteration is initiated, and the learning rate is adjusted. Maximum number of iterations .

[0010] Preferably, the implementation steps of step S5 are as follows: S5.1. Enable MCDropout Secondary reasoning, abandonment rate The standard deviation of voxel-level dose is calculated as shown in equation (3): (3); S5.2. Generate a risk map and mark it. The area is a high-risk area.

[0011] Preferably, the implementation steps of step S6 are as follows: S6.1. Introduce a domain adaptation module, add a gradient inversion layer after the decoder, and train a domain classifier to distinguish the data source center; S6.2. StyleGAN2 is used to synthesize rare anatomical variation data to improve the model's generalization ability and latent space dimension. .

[0012] Preferably, the implementation steps of step S7 are as follows: S7.1. Establish a dual-channel detection mechanism. For structural anomaly detection, use 3DU-Net to segment the input contour and calculate the Dice coefficient. If Dice < 0.85, trigger an alarm. S7.2. Train the autoencoder using dose anomaly detection to calculate the reconstruction error. If MSE Manual review triggered.

[0013] Preferably, the implementation steps of step S8 are as follows: S8.1. After submitting the new plan, generate the predicted dose and perform steps S4 to S7 for verification. If the verification passes, issue a verification report. S8.2. If it fails, start adaptive optimization to generate a modified plan, continuously monitor the pass rate and single case processing time, and form a clinical feedback loop with a target pass rate > 95% and a target single case processing time < 2 minutes.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes a three-dimensional convolutional network to automatically learn the dose distribution patterns of historical high-quality plans. By embedding a physical constraint module, it ensures that the prediction results conform to radiological principles. Combined with a real-time clinical rule engine, it instantly identifies and corrects hot spots and cold spots that violate regulations. With the help of uncertainty quantification technology, it locates high-risk areas and ultimately achieves fully automated verification at the minute level and outputs executable optimization suggestions, improving safety while increasing efficiency by tens of times. Attached Figure Description

[0015] Figure 1 A flowchart of the radiotherapy planning dose distribution verification method based on deep learning of the present invention is shown. Detailed Implementation

[0016] 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, and 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.

[0017] Please see Figure 1This invention proposes a method for verifying radiotherapy plan dose distribution based on deep learning, comprising: S1. Collect historical radiotherapy plan data, generate a physical reference dose field using the Monte Carlo algorithm, unify the resolution of anatomical structure voxels to 1 cubic millimeter, normalize the dose according to the prescription dose, use only translation and mirror transformation for data augmentation, and add a small amount of Gaussian noise to construct a three-dimensional training dataset with enhanced physical information. S2. Construct a physically constrained 3D convolutional network, embed pen-shaped beam energy kernel convolutional modules in a U-shaped architecture to enhance physical consistency, use dense connection blocks to preserve detailed features, and design a hybrid loss function that incorporates clinical dose constraints; S3. Input the new patient's three-dimensional anatomical data into the training model to predict the dose distribution. After restoring the original resolution through trilinear interpolation, apply anisotropic Gaussian filtering to eliminate discontinuities between slices. S4. Establish a real-time clinical rule engine to automatically detect insufficient target coverage or organ overload violations, and trigger the built-in optimizer to iteratively correct the problem areas; S5. Monte Carlo stochastic inference is used to quantify the prediction uncertainty and generate heat maps of high-risk areas to mark the voxel locations that exceed the clinical tolerance threshold. S6. Introduce a gradient inversion layer to obfuscate data source features, and combine it with a generative adversarial network to synthesize anatomical variant samples, thereby improving the generalization ability across medical institutions; S7. Deploy dual-channel anomaly monitoring, which detects input data anomalies through contour segmentation verification and dose reconstruction error detection, automatically intercepts problematic cases and transfers them to manual review; S8. Establish a fully automated clinical workflow, issue reports for validated protocols, generate optimization suggestions for unvalidated protocols and provide feedback to physicians, and continuously track pass rates and timeliness indicators to achieve closed-loop optimization.

[0018] In this embodiment, it should also be noted that the implementation steps of step S1 are as follows: S1.1. Collect historical radiotherapy plans containing CT images, three-dimensional dose distribution, and irradiation parameters. For each case, use the Monte Carlo algorithm to generate a baseline physical dose field. ; S1.2. Data Standardization: The resolution of anatomical structure voxels is standardized to [specific value]. Dosage as prescribed The interval is shown in equation (1): (1); S1.3. The enhancement strategy only allows translation and mirror transformations, with the offset as follows: ; S1.4. Add Gaussian noise to the dose field, standard deviation Output the training set: , C represents the number of anatomical channels.

[0019] In this embodiment, it should also be noted that the implementation steps of step S2 are as follows: The network adopts an encoder-decoder architecture, including the input layer size. The encoder contains 4 steps with a step size of 2. The convolutional downsampling layer and the decoder contain corresponding upsampling layers. The core components include a physical constraint module, dense connection blocks, and a loss function. The physical constraint module includes: ,in for Pencil beam energy deposition core, For convolution, Hadamard product; densely connected blocks: 4 layers of convolution (kernel) Number of channels Each layer's output is concatenated with its input; the loss function is shown in equation (2): Loss (2).

[0020] In this embodiment, it should also be noted that the implementation steps of step S3 are as follows: S3.1. Input the new patient's CT scan and contour data into PC-3DUnet, and output the three-dimensional dose tensor. ; S3.2. Post-processing: Trilinear interpolation recovery Original CT resolution, anisotropic Gaussian filtering to eliminate discontinuities between slices. .

[0021] In this embodiment, it should also be noted that the implementation steps of step S4 are as follows: constructing a rule engine to check clinical indicators, and the target area coverage is... of This triggers a "PTV insufficient coverage" alert; organ availability is limited. [Spinal cord] If an error occurs, a "spinal cord hotspot exceeding the limit" alarm is triggered; if this occurs, the optimizer iteration is initiated, and the learning rate is adjusted. Maximum number of iterations .

[0022] In this embodiment, it should also be noted that the implementation steps of step S5 are as follows: S5.1. Enable MCDropout Secondary reasoning, abandonment rate The standard deviation of voxel-level dose is calculated as shown in equation (3): (3); S5.2. Generate a risk map and mark it. The area is a high-risk area.

[0023] In this embodiment, it should also be noted that the implementation steps of step S6 are as follows: S6.1. Introduce a domain adaptation module, add a gradient inversion layer after the decoder, and train a domain classifier to distinguish the data source center; S6.2. StyleGAN2 is used to synthesize rare anatomical variation data to improve the model's generalization ability and latent space dimension. .

[0024] In this embodiment, it should also be noted that the implementation steps of step S7 are as follows: S7.1. Establish a dual-channel detection mechanism. For structural anomaly detection, use 3DU-Net to segment the input contour and calculate the Dice coefficient. If Dice < 0.85, trigger an alarm. S7.2. Train the autoencoder using dose anomaly detection to calculate the reconstruction error. If MSE Manual review triggered.

[0025] In this embodiment, it should also be noted that the implementation steps of step S8 are as follows: S8.1. After submitting the new plan, generate the predicted dose and perform steps S4 to S7 for verification. If the verification passes, issue a verification report. S8.2. If it fails, start adaptive optimization to generate a modified plan, continuously monitor the pass rate and single case processing time, and form a clinical feedback loop with a target pass rate > 95% and a target single case processing time < 2 minutes.

[0026] For examples, please refer to Figure 1 In practical applications, deep learning-based methods for verifying radiotherapy plan dose distribution specifically include: S1. Collect historical radiotherapy planning data, generate a physical reference dose field using the Monte Carlo algorithm, unify the voxel resolution of anatomical structures to 1 cubic millimeter, normalize the dose according to the prescription dose, perform data augmentation only using translation and mirror transformations, and add a small amount of Gaussian noise to construct a physically augmented 3D training dataset. Specifically: S1.1. Collect historical radiotherapy plans containing CT images, three-dimensional dose distribution, and irradiation parameters. For each case, use the Monte Carlo algorithm to generate a baseline physical dose field. ; S1.2. Data Standardization: The resolution of anatomical structure voxels is standardized to [specific value]. Dosage as prescribed The interval is shown in equation (1): (1); S1.3. The enhancement strategy only allows translation and mirror transformations, with the offset as follows: ; S1.4. Add Gaussian noise to the dose field, standard deviation Output the training set: , C represents the number of anatomical channels; Step S1 solves the problem of data augmentation damaging the correlation between dose and anatomical space by constructing a physical augmentation dataset, adding translation / mirror augmentation and noise injection to the Monte Carlo baseline dose, and achieves physical rationality and noise robustness of the training data, thereby reducing the model's generalization error. S2. Construct a physically constrained 3D convolutional network, embedding pencil-shaped beam energy kernel convolutional modules in a U-shaped architecture to enhance physical consistency, employing dense connective blocks to preserve detailed features, and designing a hybrid loss function that incorporates clinical dose constraints. Specifically: The network adopts an encoder-decoder architecture, including the input layer size. The encoder contains 4 steps with a step size of 2. The convolutional downsampling layer and the decoder contain corresponding upsampling layers. The core components include a physical constraint module, dense connection blocks, and a loss function. The physical constraint module includes: ,in for Pencil beam energy deposition core, For convolution, Hadamard product; densely connected blocks: 4 layers of convolution (kernel) Number of channels Each layer's output is concatenated with its input; the loss function is shown in equation (2): Loss (2); Step S2 solves the problem of traditional networks ignoring the physical rules and clinical constraints of radiotherapy by building a physically constrained three-dimensional network, adding a pencil beam kernel module, dense blocks, and clinical hybrid loss. This achieves deep coupling between dose prediction and clinical needs, and improves the accuracy of target gradient. S3. Input the new patient's three-dimensional anatomical data into the training model to predict the dose distribution. After restoring the original resolution through trilinear interpolation, apply anisotropic Gaussian filtering to eliminate discontinuities between slices. Specifically: S3.1. Input the new patient's CT scan and contour data into PC-3DUnet, and output the three-dimensional dose tensor. ; S3.2. Post-processing: Trilinear interpolation recovery Original CT resolution, anisotropic Gaussian filtering to eliminate discontinuities between slices. ; Step S3 performs trilinear interpolation and anisotropic filtering after predicting the dose, which solves the spatial discontinuity problem caused by two-dimensional slice reconstruction, achieves voxel-level smooth transition of the three-dimensional dose field, and reduces the dose error of small organs. S4. Establish a real-time clinical rule engine to automatically detect insufficient target coverage or organ overload violations, triggering the built-in optimizer to iteratively correct problematic areas. Specifically: Build a rule engine to examine clinical indicators and target coverage. of This triggers a "PTV insufficient coverage" alert; organ availability is limited. [Spinal cord] If an error occurs, a "spinal cord hotspot exceeding the limit" alarm is triggered; if this occurs, the optimizer iteration is initiated, and the learning rate is adjusted. Maximum number of iterations ; Step S4 automatically verifies violations and triggers optimization through a real-time clinical rule engine, which solves the problem that the prediction results are mathematically accurate but clinically infeasible, and achieves 100% automatic correction of violation cases, thus improving the clinical pass rate. S5. Monte Carlo stochastic inference is used to quantify prediction uncertainty and generate heat maps of high-risk areas to mark voxel locations exceeding clinical tolerance thresholds. Specifically: S5.1. Enable MCDropout Secondary reasoning, abandonment rate The standard deviation of voxel-level dose is calculated as shown in equation (3): (3); S5.2. Generate a risk map and mark it. The area is a high-risk area; Step S5 generates an uncertainty heatmap using MC Dropout, which solves the problem that traditional methods cannot quantify and predict risks, enables visual early warning of high-risk areas, and improves the efficiency of physician review. S6. Introduce a gradient inversion layer to obfuscate data source features, and combine it with a generative adversarial network to synthesize anatomical variant samples, thereby improving the generalization ability across medical institutions. Specifically: S6.1. Introduce a domain adaptation module, add a gradient inversion layer after the decoder, and train a domain classifier to distinguish the data source center; S6.2. StyleGAN2 is used to synthesize rare anatomical variation data to improve the model's generalization ability and latent space dimension. ; Step S6 solves the generalization failure problem caused by cross-center data differences by combining domain confusion with GAN dissection mutation synthesis, and reduces the model performance volatility to a low level in heterogeneous data scenarios; S7. Deploy dual-channel anomaly monitoring, detecting input data anomalies through contour segmentation verification and dose reconstruction error detection, automatically intercepting problematic cases and transferring them to manual review. Specifically: S7.1. Establish a dual-channel detection mechanism. For structural anomaly detection, use 3DU-Net to segment the input contour and calculate the Dice coefficient. If Dice < 0.85, trigger an alarm. S7.2. Train the autoencoder using dose anomaly detection to calculate the reconstruction error. If MSE Triggering manual review; Step S7, through dual-channel anomaly monitoring, contour segmentation and dose reconstruction, solves the risk of misjudgment caused by input errors or abnormal data, and achieves a higher anomaly interception rate for the system in data pollution scenarios. S8. Establish a fully automated clinical workflow, issue reports for validated protocols, generate optimization suggestions for unvalidated protocols and provide feedback to physicians, continuously track pass rates and timeliness indicators to achieve closed-loop optimization, specifically: S8.1. After submitting the new plan, generate the predicted dose and perform steps S4 to S7 for verification. If the verification passes, issue a verification report. S8.2. If it fails, start adaptive optimization to generate a modified plan, continuously monitor the pass rate and single case processing time, and form a clinical feedback loop with a target pass rate > 95% and a target single case processing time < 2 minutes. Step S8 automatically issues / optimizes recommendations and tracks indicators through a closed-loop workflow, which solves the problem of disconnect from clinical feedback, significantly shortens the validation cycle, and increases the adoption rate of optimization recommendations.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for verifying radiotherapy planned dose distribution, characterized in that, include; S1. Collect historical radiotherapy plan data, generate a physical reference dose field using the Monte Carlo algorithm, unify the resolution of anatomical structure voxels to 1 cubic millimeter, normalize the dose according to the prescription dose, use only translation and mirror transformation for data augmentation, and add a small amount of Gaussian noise to construct a three-dimensional training dataset with enhanced physical information. S2. Construct a physically constrained 3D convolutional network, embed pen-shaped beam energy kernel convolutional modules in a U-shaped architecture to enhance physical consistency, use dense connection blocks to preserve detailed features, and design a hybrid loss function that incorporates clinical dose constraints; S3. Input the new patient's three-dimensional anatomical data into the training model to predict the dose distribution. After restoring the original resolution through trilinear interpolation, apply anisotropic Gaussian filtering to eliminate discontinuities between slices. S4. Establish a real-time clinical rule engine to automatically detect insufficient target coverage or organ overload violations, and trigger the built-in optimizer to iteratively correct the problem areas; S5. Monte Carlo stochastic inference is used to quantify the prediction uncertainty and generate heat maps of high-risk areas to mark the voxel locations that exceed the clinical tolerance threshold. S6. Introduce a gradient inversion layer to obfuscate data source features, and combine it with a generative adversarial network to synthesize anatomical variant samples, thereby improving the generalization ability across medical institutions; S7. Deploy dual-channel anomaly monitoring, which detects input data anomalies through contour segmentation verification and dose reconstruction error detection, automatically intercepts problematic cases and transfers them to manual review; S8. Establish a fully automated clinical workflow, issue reports for validated protocols, generate optimization suggestions for unvalidated protocols and provide feedback to physicians, and continuously track pass rates and timeliness indicators to achieve closed-loop optimization.

2. The method for verifying radiotherapy planned dose distribution based on deep learning according to claim 1, characterized in that, The implementation steps of step S1 are as follows: S1.

1. Collect historical radiotherapy plans containing CT images, three-dimensional dose distribution, and irradiation parameters. For each case, use the Monte Carlo algorithm to generate a baseline physical dose field. ; S1.

2. Data Standardization: The resolution of anatomical structure voxels is standardized to [specific value]. Dosage as prescribed The interval is shown in equation (1): （1）； S1.

3. The enhancement strategy only allows translation and mirror transformations, with the offset as follows: ; S1.

4. Add Gaussian noise to the dose field, standard deviation Output the training set: , C represents the number of anatomical channels.

3. The method for verifying radiotherapy planning dose distribution based on deep learning according to claim 2, characterized in that, The implementation steps of step S2 are as follows: The network adopts an encoder-decoder architecture, including the input layer size. The encoder contains 4 steps with a step size of 2. The convolutional downsampling layer and the decoder contain corresponding upsampling layers. The core components include a physical constraint module, dense connection blocks, and a loss function. The physical constraint module includes: ,in for Pencil beam energy deposition core, For convolution, Hadamard product; densely connected blocks: 4 layers of convolution, kernel Number of channels Each layer's output is concatenated with its input; the loss function is shown in equation (2): Loss (2)。 4. The method for verifying radiotherapy planned dose distribution based on deep learning according to claim 3, characterized in that, The implementation steps of step S3 are as follows: S3.

1. Input the new patient's CT scan and contour data into PC-3DUnet, and output the three-dimensional dose tensor. ; S3.

2. Post-processing: Trilinear interpolation recovery Original CT resolution, anisotropic Gaussian filtering to eliminate discontinuities between slices. .

5. The deep learning-based radiotherapy planning dose distribution verification method according to claim 4, characterized in that, The implementation steps of step S4 are as follows: Construct a rule engine to examine clinical indicators, and the target area coverage is... of This triggers a "PTV insufficient coverage" alert; organ availability is limited. [Spinal cord] If an error occurs, a "spinal cord hotspot exceeding the limit" alarm is triggered; if this occurs, the optimizer iteration is initiated, and the learning rate is adjusted. Maximum number of iterations .

6. The method for verifying radiotherapy planned dose distribution based on deep learning according to claim 5, characterized in that, The implementation steps of step S5 are as follows: S5.

1. Enable MCDropout Secondary reasoning, abandonment rate The standard deviation of voxel-level dose is calculated as shown in equation (3): （3）； S5.

2. Generate a risk map and mark it. The area is a high-risk area.

7. The method for verifying radiotherapy planned dose distribution based on deep learning according to claim 6, characterized in that, The implementation steps of step S6 are as follows: S6.

1. Introduce a domain adaptation module, add a gradient inversion layer after the decoder, and train a domain classifier to distinguish the data source center; S6.

2. StyleGAN2 is used to synthesize rare anatomical variation data to improve the model's generalization ability and latent space dimension. .

8. The deep learning-based radiotherapy planning dose distribution verification method according to claim 7, characterized in that, The implementation steps of step S7 are as follows: S7.

1. Establish a dual-channel detection mechanism. For structural anomaly detection, use 3DU-Net to segment the input contour and calculate the Dice coefficient. If Dice < 0.85, trigger an alarm. S7.

2. Train the autoencoder using dose anomaly detection to calculate the reconstruction error. If MSE Manual review triggered.

9. The method for verifying radiotherapy planned dose distribution based on deep learning according to claim 8, characterized in that, The implementation steps of step S8 are as follows: S8.

1. After submitting the new plan, generate the predicted dose and perform steps S4 to S7 for verification. If the verification passes, issue a verification report. S8.

2. If it fails, start adaptive optimization to generate a modified plan, continuously monitor the pass rate and single case processing time, and form a clinical feedback loop with a target pass rate > 95% and a target single case processing time < 2 minutes.

Images