A detection method and related equipment for volumetric intensity modulated radiation therapy

Abstract

The present invention provides a detection method and related equipment for volumetric intensity modulated radiotherapy, which are applied in the field of data processing technology. The present application obtains a training data set and a target data set; pre-processes the training data set and the target data set to generate a training data set with identification information and a target data set with identification information; constructs an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; trains the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; processes the target data set based on the target deep hybrid learning model to generate a dose verification result. By constructing a deep hybrid learning model based on a generative adversarial network and a deep learning model as a general framework for multi-institution VMAT patient dose verification, and thereby eliminating the heterogeneity between variables from multiple institutions, the purpose of providing more accurate and lower computational cost for clinical PSQA is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a detection method and related equipment for volumetric intensity modulated radiation therapy. Background Art

[0002] Intensity modulated radiation therapy (IMRT) and rotational intensity modulated radiation therapy (VMAT) are commonly used cancer treatments. However, due to the modulation complexity of IMRT and VMAT, additional quality assurance must be considered. Therefore, patient dose verification (PSQA) should be performed before treatment to ensure that the implemented dose is consistent with the planned dose. Traditional PSQA usually uses measurement-based methods, which is time-consuming and labor-intensive. PSQA methods based on machine learning (ML) / deep learning (DL) technology can improve efficiency and gradually become the forefront of research. According to the applicant's survey, several scholars in the prior art extracted plan complexity features to train ML / DL models to predict gamma pass rate (GPR). Several scholars used flux maps or dosimetry features to train DL models to predict GPR. The AAPMTG-218 report recommends that intervention / tolerance limits should be set. If the dose verification results exceed the intervention limit, it means that the implemented dose deviation is large.

[0003] The new PSQA method based on ML / DL technology has been proven to be accurate in the above studies. However, some studies have shown that accelerator equipment, radiotherapy planning system (TPS) and QA equipment may significantly affect the measured GPR as well as the planning complexity indicators. Therefore, if the ML / DL model is applied to clinical practice, it will face the problem of multi-institutional data heterogeneity. Due to the widespread heterogeneity of multi-institutional data collection, the field of patient dose verification (PSQA) of rotational intensity modulated radiation therapy (VMAT) faces the challenge of a universal model across sites. Therefore, building a universal model that can handle a variety of multi-institutional data is crucial to achieve large-scale integration and analysis.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] The purpose of this application is to provide a detection method and related equipment for volumetric intensity modulated radiotherapy, which at least to some extent overcomes the problems existing in the prior art, and constructs a deep hybrid learning model based on a generative adversarial network and a deep learning model as a general framework for multi-institutional VMAT patient dose verification, thereby eliminating the heterogeneity between variables from multiple institutions, thereby achieving the purpose of providing more accurate and computationally cheaper clinical PSQA.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present invention.

[0007] According to one aspect of the present application, a detection method for volumetric intensity modulated radiotherapy is provided, including: acquiring a training data set and a target data set; preprocessing the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the mechanism data corresponding to different data; constructing an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; training the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; processing the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result.

[0008] In one embodiment of the present application, the preprocessing of the training data set and the target data set to generate a training data set with identification information and a target data set with identification information includes: extracting features from the training data set to determine an original feature library; dividing each feature data set according to the original feature library to generate a training data set and a test data set; using a classifier to predict each test data set divided from the original feature library to determine a prediction result; using a preset algorithm to train each training data set divided from the original feature library to obtain a test set class prediction result; generating a fused feature data set based on the prediction result and the test set class prediction result; processing the fused feature data set based on preset feature screening rules to generate identification information, wherein the identification information is used to identify the institutional data corresponding to different data.

[0009] In one embodiment of the present application, the initial deep hybrid learning model is trained based on the training data set with identification information to generate a target deep hybrid learning model, including: classifying the training data set with identification information based on the identification information to generate several data groups, wherein each data group is composed of data samples with the same identification information; processing several data groups respectively based on the generator of the initial deep hybrid learning model to generate fictitious data samples, wherein the fictitious data samples are the same as the data samples in the corresponding data groups, and the false data samples include the same identification information as the data samples in the corresponding data groups; dividing the false data samples into corresponding data groups based on the identification information to generate updated data groups; processing several updated data groups based on the discriminator of the initial deep hybrid learning model to generate classification results, wherein the classification results include the authenticity of the current data samples; when the discriminator completes processing of all updated data groups, generating a target deep hybrid learning model.

[0010] In one embodiment of the present application, the training data set with identification information is classified based on the identification information to generate several data groups, including: dividing the training data set with identification information into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio; extracting multiple data groups from the training set, wherein each data group contains a preset number of data samples.

[0011] In one embodiment of the present application, extracting multiple groups of data groups from the training set includes: dividing the training set based on identification information to generate a number of number of class samples; obtaining any smaller number of class samples in the training set; generating adjacent samples based on the distance between any smaller number of class samples and other smaller number of class samples of the same category, wherein the adjacent samples include a preset number of any smaller number of class samples; sampling the adjacent samples to generate a preset number of sampled samples; generating multiple groups of data groups based on any minority class sample and each sampled sample.

[0012] In one embodiment of the present application, the sampling of the adjacent samples to generate a preset number of sampling samples includes: determining a sampling ratio based on the number of samples of each category in the training set; determining a sampling ratio based on the sampling ratio; and sampling the adjacent samples based on the sampling ratio to generate a preset number of sampling samples.

[0013] In one embodiment of the present application, the discriminator based on the initial deep hybrid learning model processes several updated data groups to generate classification results, including: the discriminator processes several updated data groups based on preset model parameters to generate classification results; if the accuracy of the classification result is false, the preset model parameters are processed based on the back propagation algorithm to generate new preset model parameters; the discriminator processes several updated data groups based on the new preset model parameters to generate classification results; if the accuracy of the classification result is true, the current data sample is skipped and subsequent data samples are processed.

[0014] Another aspect of the present application is a detection device for volumetric modulated radiotherapy, characterized in that the device includes: an acquisition module, used to acquire a training data set and a target data set; a processing module, used to preprocess the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the mechanism data corresponding to different data; constructing an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; training the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; processing the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result.

[0015] According to another aspect of the present application, an electronic device is provided, characterized in that it comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-mentioned volumetric intensity modulated radiation therapy detection method by executing the executable instructions.

[0016] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the above-mentioned volumetric intensity modulated radiation therapy detection method is implemented.

[0017] According to another aspect of the present application, a computer program product is provided, including a computer program, characterized in that when the computer program is executed by a third processor, the above-mentioned volumetric intensity modulated radiation therapy detection method is implemented.

[0018] The present application provides a detection method and related equipment for volumetric intensity modulated radiotherapy, in which a server obtains a training data set and a target data set; preprocesses the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the institutional data corresponding to different data; constructs an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; trains the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; processes the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result. By constructing a deep hybrid learning model based on a generative adversarial network and a deep learning model as a general framework for multi-institution VMAT patient dose verification, the heterogeneity between variables from multiple institutions is eliminated, thereby achieving the purpose of providing more accurate and lower computational cost for clinical PSQA.

[0019] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart showing a method for detecting volumetric intensity modulated radiation therapy provided by an embodiment of the present application;

[0021] Figure 2 A schematic structural diagram of a detection device for volumetric intensity modulated radiotherapy provided by an embodiment of the present application is shown;

[0022] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present application is shown;

[0023] Figure 4 A schematic diagram of a storage medium provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0025] Combine the following Figure 1 To describe the detection method of volumetric intensity modulated radiation therapy according to an exemplary embodiment of the present application. It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this regard. On the contrary, the embodiments of the present application can be applied to any applicable scenario.

[0026] In one embodiment, the present application also proposes a detection method and related equipment for volumetric intensity modulated radiation therapy. Figure 1 The flowchart of a detection method for volumetric intensity modulated radiotherapy according to an embodiment of the present application is schematically shown. Figure 1 As shown, the method is applied to a server, comprising:

[0027] S101, obtaining a training data set and a target data set.

[0028] In one embodiment, the present application uses 1815 dual-arc VMAT plans for continuous treatment at 8 different sites, including 876 cases of gynecological tumors, 392 cases of head and neck cancer and brain cancer, 287 cases of rectal cancer, 74 cases of prostate cancer, 102 cases of chest tumors, 55 cases of other abdominal and pelvic tumors, and 29 other cases. Among them, 80% of the total data set is used as a training data set for model training, verification and testing; 20% of the total data set is used as a target data set for clinical debugging and implementation, 20 plans at each site are used for clinical debugging, and the remaining data sets (at least 10% of the data sets) are used for clinical implementation. Both the training data set and the target data set are data sources formed by aggregating data from different institutions.

[0029] The 54 planning complexity feature parameters of each VMAT (i.e., volumetric modulated radiation therapy) plan are used as input to the model. In order to address the data imbalance problem, the ADASYN method is used, with special attention paid to the part with less data. By setting the sampling strategy to "minority", ADASYN mainly generates synthetic samples for classes that are underrepresented in the feature space. The method focuses on samples that are difficult to classify, creating more synthetic samples near these marginal instances. This adaptive approach ensures that the classifier used can better handle the problem of data imbalance, thereby potentially enhancing its generalization ability.

[0030] S102: Preprocess the training data set and the target data set to generate a training data set with identification information and a target data set with identification information.

[0031] In one implementation, identification information is used to identify the institutional data corresponding to different data. For each institutional data, a set of identification information is constructed based on the QA equipment, TPS, delivery equipment and site it uses. These identification information are represented using one-hot encoding. For example (the applicant data is more complex than this example), if there are 3 different devices and 2 sites, they may be encoded as: device A: [1,0,0], device B: [0,1,0], device C: [0,0,1], site 1: [1,0], site 2: [0,1]. If a data point uses device A and site 2, its identification information will be [1,0,0,0,1]. Among them, the identification information can be represented by a preset specific symbol. This embodiment does not limit the specific symbol as long as the present solution can be implemented.

[0032] In another embodiment, feature extraction is performed on the training data set to determine the original feature library; each feature data set is divided according to the original feature library to generate a training data set and a test data set; a classifier is used to predict each test data set divided by the original feature library to determine the prediction result; a preset algorithm is used to divide each training data set in the original feature library for training to obtain a test set class prediction result; a fused feature data set is generated according to the prediction result and the test set class prediction result; the fused feature data set is processed based on the preset feature screening rule to generate identification information. Feature extraction is performed on the training data set to determine the original feature library, wherein the feature extraction includes four features: original features, statistical features, frequency domain features, and time domain features. The extracted common features are specifically introduced as follows: Original features: The original features are the collected data, after preprocessing, using all the information in each sample data matrix for model training. In order not to lose sample information, each curve of the sample Xi is directly expanded and stretched into a row vector. Statistical features: Statistical features consider the changing trend of all data points on the curve. By extracting statistical features, the dimension of the data sample can be reduced, which is convenient for analyzing genetic data and accelerating the convergence speed during model training. The extracted statistical features include maximum value, minimum value, mean, variance, and standard deviation. Frequency domain features: Perform wavelet transform on the data set, and use the coefficients obtained by the second-order wavelet transform as new features to form frequency domain features. Time domain features: Time domain features mainly study the changing laws of signals and systems from the perspective of time. Time domain features can reflect the information of curve data in time changes. This process mainly extracts the first-order forward difference of the data set to form the first-order difference time domain features, and uses the exponential moving average feature processing method.

[0033] In another embodiment, each feature data set is divided according to the original feature library to generate a training data set and a test data set. The original feature library is divided using existing technology to generate a training data set and a test data set. In this embodiment, as long as the division can be performed, no restriction is imposed on the method used. The classifier is used to divide each test data set into the original feature library for prediction, and the prediction result is determined. The preset algorithm is used to divide each training data set into the original feature library for training to obtain the test set class prediction result. According to the prediction result and the test set class prediction result, a fusion feature data set is generated. The preset algorithm is used to divide each training data set into the original feature library for training to obtain the test set class prediction result. In the embodiment, as long as the test set class prediction result can be obtained, no restriction is imposed on the algorithm used. According to the prediction result and the test set class prediction result, a fusion feature data set is generated. The classification results of Evalfeature and Evaloriginal are judged. If Evalfeature is improved compared with Evaloriginal, the feature is incorporated into Dfusion, and the final fusion feature data set Dfusion is obtained after the traversal is completed. Specifically, the data classification result corresponding to the training data set is obtained, because the classification result of each training data set can be determined in advance, and the data can be directly obtained from the outside. Compare the prediction results with the data classification results to determine the first comparison result; compare the test set prediction results with the data classification results to determine the second comparison result; judge whether the second comparison result meets the preset requirements with the first comparison result. When the preset requirements are met, determine the corresponding features as the fused feature data set.

[0034] S103, constructing an initial deep hybrid learning model (DHL) based on the generative adversarial network and the deep learning model.

[0035] In one embodiment, a deep hybrid learning (DHL) model is proposed, which combines the classical machine learning method (i.e., generative adversarial network, StarGAN is used in this embodiment) and the deep learning model (Transformer is used in this embodiment), taking advantage of their respective advantages to achieve performance superior to that of a single method. DHL provides a more accurate and computationally cheaper solution for clinical PSQA. The StarGAN process includes two main modules: a discriminator D and a generator G. The identification information is constructed based on various parameters such as different accelerators, QA equipment, radiotherapy planning systems, and treatment institutions. It is worth noting that the representation of different sites in the identification information is achieved through one-hot encoding. For example, the identification information of Clinac21EX, ArcCHECK, and Eclipse is represented as [1,0,0,0,0,0,0,1,0,0,1,0,0], while the identification information of the 8th site is [0,0,0,0,0,0,0,0,1]. Therefore, the combined representation of the identity information becomes [1,0,0,0,0,0,0,1,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,1]. The main role of the discriminator D is to distinguish between real data and data generated by the generator G. The generator G is designed to convert the input raw data into data in the target domain. The generator takes the raw data as input and processes it through a series of layers such as convolution, deconvolution, and residual blocks. The ultimate goal of the generator is to determine the mapping relationship from the original domain to the target domain. The generator G produces data that is indistinguishable from real data and can be classified as data in the target domain by the discriminator D.

[0036] The deep hybrid learning model based on StarGAN and transformer consists of two main components, including the StarGAN workflow and the hybrid classification and regression workflow. The StarGAN flow uses only a single model to transform different types of data sets from multiple sites. A hybrid classification workflow is adopted to extract complex feature representations from unstructured data through Transformer to obtain better interpretation capabilities, and a high-precision classification and prediction model is established for unstructured data using random forests; finally, the hybrid classification and prediction results are combined with the logistic regression model (LR) and the support vector machine (SVM) model. In order to obtain the optimal hyperparameters and fit the model, 5-fold cross validation is used for model training. In summary, this application introduces a hybrid classification and regression PSQA general model based on StarGAN. The model effectively promotes multi-institutional data transmission and extracts complex features from unstructured data. In addition, it combines logistic regression and support vector machine methods to build accurate classification and prediction models. The overall design of the model is an integrated solution, including data transmission, feature extraction, accurate classification and prediction processes.

[0037] S104: Train the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model.

[0038] In one embodiment, a training data set with identification information is classified and processed based on identification information to generate several data groups, wherein each data group is composed of data samples with the same identification information; a generator based on an initial deep hybrid learning model processes several data groups respectively to generate fictitious data samples, wherein the fictitious data samples are the same as the data samples in the corresponding data group, and the fictitious data samples include the same identification information as the data samples in the corresponding data group. The generator (i.e., StarGAN) receives the original data and the corresponding identification information, and then generates new data under these identification information, with the purpose of making the generated new data consistent with the target data corresponding to the identification information as much as possible, thereby reducing the discriminative ability of the discriminator (i.e., Transformer). The identification information is constructed based on various parameters such as different accelerators, QA equipment, radiotherapy planning systems, and treatment institutions. It is worth noting that the representation of different sites in the identification information is achieved through one-hot encoding. For example, the identification information of Clinac21EX, ArcCHECK, and Eclipse is represented as [1,0,0,0,0,0,0,1,0,0,1,0,0], and the identification information of the 8th site is [0,0,0,0,0,0,0,0,1]. Therefore, the combined representation of the identification information becomes [1,0,0,0,0,0,0,1,0,0,1,0,0,0,0,0,0,0,0,0,0,1].

[0039] The generator takes the original data as input and processes it through a series of layers, such as convolution, deconvolution, and residual blocks. The ultimate goal of the generator is to determine the mapping relationship from the original domain to the target domain. The generator produces data that is indistinguishable from real data and can be classified as data in the target domain by the discriminator.

[0040] Before training the initial deep hybrid model, the applicant assumes that all sites are converted to site 2 as the target. Therefore, all target domain identification information is assigned to the identification information of site 2. In each training iteration, the applicant first samples a batch of real data from the training dataset, denoted as x, and their corresponding identification information label org Subsequently, the discriminator D is trained, i.e., the real data x is first input into the discriminator D, which tries to maximize the adversarial loss of Item, then, the generator G uses label orgSynthetic data is generated. These generated data, after being introduced into the discriminator D, aim to maximize the terms of the adversarial loss. The cumulative loss of these components serves as the final loss of the discriminator D and is optimized through back-propagation. The structure of the discriminator D consists of multiple convolutional layers paired with the LeakyReLU activation function, which aims to identify the authenticity of the input data. After the optimization of the discriminator D, the generator G is updated. This process involves using the generated synthetic data. After the introduction of the discriminator D, the goal of the generator G is to minimize the adversarial loss and the classification loss. In addition, the generator G reconstructs the previously generated data into the original domain data, and this reconstruction process generates a loss. The cumulative loss of these components constitutes the final loss of the generator G and is updated through back-propagation. The architecture of the generator G contains multiple convolutional layers, normalization layers, and upsampling / downsampling layers, which aims to generate the corresponding target domain data based on the given conditional identification information. By iteratively alternating the training of the above discriminator D and generator G, the generation performance and discrimination ability of the model can be gradually enhanced. Specifically, the adversarial loss is calculated by the discriminator D to maximize D's own objective function; it is also used by the generator G to minimize G's own objective function. Classification loss of real data Calculated by the discriminator D, it is used to optimize D itself so that it can correctly classify real images. Classification loss of fake data This loss is calculated by the generator G and is used to optimize G itself so that the images it generates can be correctly classified into the target domain. The reconstruction loss is calculated by the generator G to maintain the consistency of the generated data content, thereby optimizing the generator G itself. The complete objective of the discriminator D: a linear combination of the adversarial loss and the real data classification loss, and the discriminator D is the main execution. The complete objective of the generator G: a linear combination of the adversarial loss, the fake data classification loss, and the reconstruction loss, and the generator G is the main execution.

[0041] Adversarial loss: To make the generated data indistinguishable from real data, an adversarial loss is adopted, where the generator G generates G(x,c) conditioned on the input data x and multi-domain target identification information c, and the discriminator D attempts to distinguish between real and fake data.

[0042] The discriminator D calculates the adversarial loss using the following formula:

[0043]

[0044] Among them, D src (x) is the output of the discriminator D for the real data x, D src (x) represents the probability that the discriminator D judges x as real data. G(x,c) is the false data generated by the generator G given the input data x and the multi-domain target identification information c. It means that for the real data x, the expected value of the probability that the discriminator D judges it as real data is calculated. This part of the loss encourages the discriminator D to judge the real data as real. It means that for the fake data G(x,c) generated by the generator G, the expected value of the probability that the discriminator D will judge it as fake data is calculated. This part of the loss encourages the discriminator D to judge the fake data as fake. The generator G tries to minimize this goal, while the discriminator D tries to maximize it.

[0045] Classification loss: For a given input data x and target domain identification information c, our goal is to convert x into output data y and correctly classify it as belonging to the target domain c. The formula is defined as follows:

[0046]

[0047] Among them, D cls (c'|x) is the probability that the classifier D, given the input data x, classifies it as the target domain identification information c'. cls (c'|x) represents the probability that the classifier D judges that the target domain is c', given the input data x. It represents the expected value of the probability that the classifier D correctly classifies the input data x and the target domain identification information c' as the target domain c'. This part of the loss encourages the classifier D to correctly classify the input data as the target domain c.

[0048] By minimizing this objective, D learns to classify the real data point x to its corresponding original domain identification information c'. Here, it is assumed that the pair (x, c') consisting of the input data and its domain identification information is given by the training data. Similarly, the loss function for false data domain classification is defined as:

[0049]

[0050] Among them, D cls (c|G(x,c)) is the probability that the classifier D classifies the false data G(x,c) generated by the generator as the false data domain identification information c. cls (c|G(x,c)) represents the probability that the classifier D judges the false data domain to be c, given the false data G(x,c) generated by the generator. It is the expected value of the probability that the classifier D correctly classifies the false data G(x,c) generated by the generator and the false data domain identification information c as the false data domain c. This part of the loss encourages the classifier D to correctly classify the false data into the corresponding false data domain.

[0051] Reconstruction Loss: The applicants applied a cycle consistency loss to the generator, which is defined as follows:

[0052]

[0053] Where G(x,c) is the output data generated by the generator G given the input data x and the multi-domain target identification information c. (G(x,c),c') is the output data generated by the generator G given the input data x and the target domain identification information c, which is input into the generator G again and generated given the target domain identification information c'. ||xG(G(x,c),c')‖1 is the L1 distance between the input data x and the reconstructed data G(G(x,c),c'). The L1 distance is the average of the absolute differences between the input data and the reconstructed data. It is the expected value of the L1 distance between the input data x and the reconstructed data G(G(x,c),c') for the input data x, the target domain identification information c and the target domain identification information c'. This part of the loss encourages the generator G to generate outputs as close as possible to the data generated by the input data x under different conditions.

[0054] Complete goal: The objective function for optimizing G and D is:

[0055]

[0056] in, is the loss function of the discriminator D, which is used to measure the performance of the discriminator. The goal of the discriminator is to minimize (Adversarial Loss) and maximize (Classification loss in the real data domain). is the adversarial loss, which is used to measure the adversarial performance of the fake data generated by the generator G and the discriminator D. The goal of the generator is to maximize λ cls It is the weight coefficient of the classification loss, which is used to balance the importance of the classification loss in the real data domain and the classification loss in the false data domain. is the real data domain classification loss, which is used to measure the performance of the discriminator D in correctly classifying the real data into the target domain. The goal of the discriminator is to maximize

[0057] is the loss function of the generator G, which is used to measure the performance of the generator. The goal of the generator is to maximize (Adversarial Loss) and minimize (false data domain classification loss) and (Reconstruction loss). is the false data domain classification loss, which is used to measure the performance of the classifier in correctly classifying the false data generated by the generator as the false data domain. The goal of the classifier is to maximize is the reconstruction loss, which is used to measure the reconstruction ability of the generator G. The goal of the generator is to minimize

[0058] In another embodiment, the false data samples are divided into corresponding data groups based on the identification information to generate updated data groups; the discriminator based on the initial deep hybrid learning model processes several updated data groups to generate classification results, wherein the classification results include the authenticity of the current data samples; when the discriminator completes the processing of all updated data groups, the target deep hybrid learning model is generated. The discriminator, its task is to distinguish whether the input data is real or generated by the generator, and classify the identification information of these data, and the classification result is: false data / true data. As an example, the intervention limits of 3% / 3mm, 3% / 2mm and 2% / 2mm are set to 95%, 90% and 80% respectively to distinguish the plan as "pass" or "fail". That is, given the intervention limit of 3% / 3mm is 95%, if the gamma pass rate is 90% under the standard of 3% / 3mm, it is considered to be unpassed if it is lower than 95%, and it is considered to be passed if it is higher than 95%. The results under other standards are similar. The generator tries to create data that is realistic enough so that the discriminator cannot distinguish between true and false. The discriminator distinguishes between real data and generated data and classifies their identification information as accurately as possible. Ultimately, the generator and discriminator become stronger in the game, thereby improving the accuracy of the deep hybrid training model.

[0059] In another embodiment, the discriminator processes several updated data groups based on preset model parameters to generate a classification result, wherein the preset model parameters are mainly the current weights and biases of the discriminator; if the accuracy of the classification result is false, the preset model parameters are processed based on the back propagation algorithm to generate new preset model parameters; the discriminator processes several updated data groups based on the new preset model parameters to generate a classification result; if the accuracy of the classification result is true, the current data sample is skipped and subsequent data samples are processed.

[0060] About the learning and recognition process of the discriminator: 1. Judgment of the discriminator: The discriminator receives input samples (whether real or generated) and makes a judgment based on its internal network structure. This judgment is determined based on the current weights and biases of the discriminator network. 2. Loss calculation: If the discriminator correctly identifies the real sample as true or the generated sample as false, the loss is small. This means that the discriminator performs well. If the discriminator mistakenly judges the real sample as false, or the generated sample as true, the loss is large. This indicates that the discriminator did not complete its task correctly.

[0061] Weight update and iterative training: 1. Weight update: According to the size of the loss, the weight of the discriminator is adjusted through the back propagation algorithm. If the loss is large, it means that the discriminator needs greater adjustment to improve its accuracy. 2. Iterative training: Through multiple iterations, the weight of the discriminator is continuously adjusted to reduce its loss when identifying samples, thereby improving its ability to distinguish between real and generated samples.

[0062] The core task of the discriminator is to accurately distinguish between real samples and generated samples. It evaluates its performance through the loss function, and the size of the loss directly affects the adjustment of the weights. Through this continuous learning and adjustment process, the discriminator gradually improves the accuracy of distinguishing between real and fake samples. After distinguishing between real data and fake data, the discriminator calculates the loss function and updates the weights. Once the weights are updated, the data will be skipped, and then new data will be loaded and the above process will be repeated. Of course, these two data may be loaded again later, depending on the training cycle, because in multiple training cycles, the data is likely to be reused. But further explained is that real data can be reused, but the so-called fake data here does not exist. The saying of reuse is that as the capacity of the generator increases, the quality of the data generated after giving the same data to the generated data again is different. Unless the generator weights are frozen and no longer updated, the generated fake data will be different from the previous fake data and will become more and more like real data.

[0063] S105: Process the target data set based on the target deep hybrid learning model to generate a dose verification result.

[0064] In one embodiment, the dose verification results include gamma pass rate and classification results. The framework of the hybrid classification regression model mainly includes a hybrid classification module and a hybrid prediction module. The transformed data obtained from StarGAN is input into the two modules. The hybrid classification module uses Transformer and random forest to independently predict the pass / fail probability. Similarly, the hybrid prediction module uses Transformer and support vector machine to independently predict the probability and value of GPR. The GPR prediction results obtained by the random forest regressor are merged with the GPR prediction results output by the Transformer model. The support vector machine model is then used to perform GPR hybrid prediction on the spliced ​​output. Combining deep learning and machine learning models can achieve better performance than a single model. In the clinical debugging phase, 20 VMAT plans for each site are prospectively collected for independent end-to-end (E2E) testing. The plan type used in the E2E test is consistent with the plan type commonly used at each site.

[0065] After model commissioning, the applicant has at least 10% of the data set from each site for commissioning the model, i.e., clinical implementation, including 35 plans from site 1, 45 plans from site 2, 20 plans from site 3, 15 plans from site 4, 25 plans from site 5, 30 plans from site 6, 15 plans from site 7, and 20 plans from site 8. If the model output is "passed", the new plan is validated using a secondary independent dose calculation algorithm and approved for delivery. If the model output is "fail", measurement-based QA is performed.

[0066] A total of 205 new plans from eight sites (at least 10% of the dataset from each site) were implemented using the targeted deep hybrid learning model as a general PSQA framework. A total of 199 plans were predicted as "passed" and 6 plans were predicted as "failed" based on the recommended 3% / 3mm and 3% / 2mm gamma criteria. The 205 failed plans were re-measured using QA equipment at all eight sites, with MAEs of 0.67% for the 3% / 3mm criterion, mean absolute error of 1.28% for the 3% / 2mm criterion, RMSE of 0.92% for the 3% / 3mm criterion, and root mean square error of 1.76% for the 3% / 2mm criterion. All failed plans were correctly identified by the model after re-measurement. One plan had predicted GPRs of 92.87% (failed) and 88.26% (failed) for the 3% / 3mm and 3% / 2mm criteria, respectively, and the re-measured GPRs were 95.8% (passed) and 90.6% (passed), respectively.

[0067] In this application, a server obtains a training data set and a target data set; preprocesses the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the institutional data corresponding to different data; constructs an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; trains the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; processes the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result. By constructing a deep hybrid learning model based on a generative adversarial network and a deep learning model as a general framework for multi-institution VMAT patient dose verification, the heterogeneity between variables from multiple institutions is eliminated, thereby achieving the purpose of providing more accurate and computationally cheaper clinical PSQA.

[0068] Optionally, in another embodiment of the above method of the present application, the training data set with identification information is classified based on the identification information to generate several data groups, including:

[0069] Dividing the training data set with identification information into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio;

[0070] A plurality of data sets are extracted from the training set, wherein each data set includes a preset number of data samples.

[0071] In one embodiment, the training sample set with identification information is divided into a training set for training an initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model according to a preset ratio, wherein the ratio can be set as needed, such as 1:8 or 1:5, etc.; multiple data groups are extracted from the training set, wherein each data group contains a preset number of data samples.

[0072] Optionally, in another embodiment of the above method of the present application, extracting multiple data sets from the training set includes:

[0073] Dividing the training set based on the identification information to generate a number of numerical samples;

[0074] Obtain any smaller number of samples from the training set;

[0075] Generate adjacent samples based on the distance between any of the smaller number of samples and other smaller number of samples of the same category, wherein the adjacent samples include a preset number of any of the smaller number of samples;

[0076] Sampling the adjacent samples to generate a preset number of sampling samples;

[0077] Based on any minority class sample and each sampling sample, multiple data sets are generated.

[0078] In one embodiment, the distance between any number of smaller number of class samples and any other number of smaller number of class samples of the same category can be Euclidean distance or other cosine distance, which is not specifically limited here. All other number of smaller number of class samples can be arranged in order of distance from any other number of smaller number of class samples from near to far, and then the data in the first few positions of the arrangement result are selected as adjacent samples. The purpose of oversampling is to increase the number of minority class samples to achieve a balance between the number of minority class samples and the number of majority class samples in the initial data set. Here, oversampling can be achieved by a sampling algorithm, which is not specifically limited here. After oversampling, multiple groups of data groups can be determined based on the results obtained by oversampling and combined with the training set. At this time, the number of descriptive data samples corresponding to the attribute category identification information in the data group is equivalent, and there are no number of smaller number of class samples or more number of class samples. In this embodiment, the training set is first preliminarily divided, and then the number of smaller number of class samples is merged with the number of larger number of class samples.

[0079] Optionally, in another embodiment of the method of the present application, the step of sampling the adjacent samples to generate a preset number of sampled samples includes:

[0080] Determining a sampling ratio based on the number of samples of each category in the training set;

[0081] Based on the sampling ratio, determining a sampling ratio;

[0082] The adjacent samples are sampled based on the sampling ratio to generate a preset number of sampling samples.

[0083] In one implementation, first obtain which data groups are included in the training set, confirm how many number class samples there are in each data group, and determine the sampling ratio according to the number of number class samples in each data group, and then use the sampling ratio to determine the sampling ratio. The sampling ratio can be set as needed, for example, it can be 50% or other ratios, which are not specifically limited here. Finally, use the sampling ratio to sample a preset number of adjacent samples to obtain a specified number of sampled samples, and all other smaller number of number class samples can be arranged in order from near to far in terms of distance from any smaller number of number class samples, and then select the preset number of samples ranked first in the arrangement result as neighboring samples, and the preset number can be set according to actual conditions, which is not specifically limited here. Based on any smaller number of number class samples and each sampled sample, a new sample is generated and used as a new data group, so that the balance effect between the smaller number of data groups and the other larger number of data groups can be further improved.

[0084] By applying the above technical solution, the server obtains a training data set and a target data set; performs feature extraction on the training data set to determine the original feature library; divides each feature data set according to the original feature library to generate a training data set and a test data set; uses a classifier to predict each test data set divided by the original feature library to determine the prediction result; uses a preset algorithm to divide each training data set in the original feature library for training to obtain a test set class prediction result; generates a fused feature data set based on the prediction result and the test set class prediction result; processes the fused feature data set based on the preset feature screening rules to generate identification information, wherein the identification information is used to identify the institutional data corresponding to different data.

[0085] An initial deep hybrid learning model is constructed based on a generative adversarial network and a deep learning model; a training data set with identification information is divided into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio; the training set is divided based on the identification information to generate a number of number of class samples; any smaller number of class samples in the training set are obtained; adjacent samples are generated based on the distance between any smaller number of class samples and other smaller number of class samples of the same category, wherein the adjacent samples include a preset number of any smaller number of class samples; a sampling ratio is determined based on the number of class samples in the training set; a sampling ratio is determined based on the sampling ratio; adjacent samples are sampled based on the sampling ratio to generate a preset number of sampled samples; multiple data groups are generated based on any minority class sample and each sampled sample, wherein each data group contains a preset number of data samples, wherein each data group consists of data samples with the same identification information.

[0086] A generator based on an initial deep hybrid learning model processes several data groups respectively to generate fictitious data samples, wherein the fictitious data samples are the same as the data samples in the corresponding data groups, and the fictitious data samples include the same identification information as the data samples in the corresponding data groups; the fictitious data samples are divided into corresponding data groups based on the identification information to generate updated data groups; the discriminator processes several updated data groups based on preset model parameters to generate classification results; if the accuracy of the classification result is false, the preset model parameters are processed based on the back propagation algorithm to generate new preset model parameters; the discriminator processes several updated data groups based on the new preset model parameters to generate classification results; if the accuracy of the classification result is true, the current data sample is skipped and subsequent data samples are processed, wherein the classification results include the authenticity of the current data sample; when the discriminator completes the processing of all updated data groups, a target deep hybrid learning model is generated; the target data set is processed based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result. By constructing a deep hybrid learning model based on generative adversarial networks and deep learning models as a common framework for multi-institutional VMAT patient dose verification, the heterogeneity between variables from multiple institutions was eliminated, achieving the goal of providing more accurate and computationally cheaper clinical PSQA.

[0087] In one embodiment, if Figure 2 As shown, the present application also provides a detection device for volumetric intensity modulated radiotherapy, comprising:

[0088] An acquisition module 201 is used to acquire a training data set and a target data set;

[0089] The processing module 202 is used to preprocess the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the mechanism data corresponding to different data; construct an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; train the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; process the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result.

[0090] In this application, a server obtains a training data set and a target data set; preprocesses the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the institutional data corresponding to different data; constructs an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; trains the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; processes the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result. By constructing a deep hybrid learning model based on a generative adversarial network and a deep learning model as a general framework for multi-institution VMAT patient dose verification, the heterogeneity between variables from multiple institutions is eliminated, thereby achieving the purpose of providing more accurate and computationally cheaper clinical PSQA.

[0091] In another embodiment of the present application, the processing module 202 is configured to pre-process the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, including:

[0092] Extracting features from the training data set to determine an original feature library;

[0093] Divide each feature data set according to the original feature library to generate a training data set and a test data set;

[0094] Use the classifier to divide the original feature library into various test data sets for prediction and determine the prediction results;

[0095] Use the preset algorithm to divide each training data set in the original feature library for training, and obtain the prediction results of the test set class;

[0096] Generate a fusion feature data set based on the prediction results and the test set class prediction results;

[0097] The fused feature data set is processed based on preset feature screening rules to generate identification information, wherein the identification information is used to identify the institution data corresponding to different data.

[0098] In another embodiment of the present application, the processing module 202 is configured to train the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model, including:

[0099] Based on the identification information, the training data set with identification information is classified and processed to generate a plurality of data groups, wherein each data group is composed of data samples with the same identification information;

[0100] The generator based on the initial deep hybrid learning model processes several data groups respectively to generate fictitious data samples, wherein the fictitious data samples have the same data as the data samples in the corresponding data groups, and the fictitious data samples include the same identification information as the data samples in the corresponding data groups;

[0101] Dividing the false data samples into corresponding data groups based on the identification information to generate updated data groups;

[0102] Processing a number of updated data groups based on the discriminator of the initial deep hybrid learning model to generate classification results, wherein the classification results include the authenticity of the current data sample;

[0103] When the discriminator completes processing of all updated data groups, a target deep hybrid learning model is generated.

[0104] In another embodiment of the present application, the processing module 202 is configured to perform classification processing on the training data set with identification information based on the identification information to generate several data groups, including:

[0105] Dividing the training data set with identification information into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio;

[0106] A plurality of data sets are extracted from the training set, wherein each data set includes a preset number of data samples.

[0107] In another embodiment of the present application, the processing module 202 is configured to extract multiple data sets from the training set, including:

[0108] Dividing the training set based on the identification information to generate a number of numerical samples;

[0109] Obtain any smaller number of samples from the training set;

[0110] Generate adjacent samples based on the distance between any of the smaller number of samples and other smaller number of samples of the same category, wherein the adjacent samples include a preset number of any of the smaller number of samples;

[0111] Sampling the adjacent samples to generate a preset number of sampling samples;

[0112] Based on any minority class sample and each sampling sample, multiple data sets are generated.

[0113] In another embodiment of the present application, the processing module 202 is configured to sample the adjacent samples to generate a preset number of sample samples, including:

[0114] Determining a sampling ratio based on the number of samples of each category in the training set;

[0115] Based on the sampling ratio, determining a sampling ratio;

[0116] The adjacent samples are sampled based on the sampling ratio to generate a preset number of sampling samples.

[0117] In another embodiment of the present application, the processing module 202 is configured to process the discriminator based on the initial deep hybrid learning model on the updated data groups to generate classification results, including:

[0118] The discriminator processes several updated data sets based on preset model parameters to generate classification results;

[0119] If the accuracy of the classification result is false, the preset model parameters are processed based on a back propagation algorithm to generate new preset model parameters;

[0120] The discriminator processes the updated data groups based on the new preset model parameters to generate classification results;

[0121] If the accuracy of the classification result is true, the current data sample is skipped and subsequent data samples are processed.

[0122] By applying the above technical solution, the server obtains a training data set and a target data set; performs feature extraction on the training data set to determine the original feature library; divides each feature data set according to the original feature library to generate a training data set and a test data set; uses a classifier to predict each test data set divided by the original feature library to determine the prediction result; uses a preset algorithm to divide each training data set in the original feature library for training to obtain a test set class prediction result; generates a fused feature data set based on the prediction result and the test set class prediction result; processes the fused feature data set based on the preset feature screening rules to generate identification information, wherein the identification information is used to identify the institutional data corresponding to different data.

[0123] An initial deep hybrid learning model is constructed based on a generative adversarial network and a deep learning model; a training data set with identification information is divided into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio; the training set is divided based on the identification information to generate a number of number of class samples; any smaller number of class samples in the training set are obtained; adjacent samples are generated based on the distance between any smaller number of class samples and other smaller number of class samples of the same category, wherein the adjacent samples include a preset number of any smaller number of class samples; a sampling ratio is determined based on the number of class samples in the training set; a sampling ratio is determined based on the sampling ratio; adjacent samples are sampled based on the sampling ratio to generate a preset number of sampled samples; multiple data groups are generated based on any minority class sample and each sampled sample, wherein each data group contains a preset number of data samples, wherein each data group consists of data samples with the same identification information.

[0124] A generator based on an initial deep hybrid learning model processes several data groups respectively to generate fictitious data samples, wherein the fictitious data samples are the same as the data samples in the corresponding data groups, and the fictitious data samples include the same identification information as the data samples in the corresponding data groups; the fictitious data samples are divided into corresponding data groups based on the identification information to generate updated data groups; the discriminator processes several updated data groups based on preset model parameters to generate classification results; if the accuracy of the classification result is false, the preset model parameters are processed based on the back propagation algorithm to generate new preset model parameters; the discriminator processes several updated data groups based on the new preset model parameters to generate classification results; if the accuracy of the classification result is true, the current data sample is skipped and subsequent data samples are processed, wherein the classification results include the authenticity of the current data sample; when the discriminator completes the processing of all updated data groups, a target deep hybrid learning model is generated; the target data set is processed based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result. By constructing a deep hybrid learning model based on generative adversarial networks and deep learning models as a common framework for multi-institutional VMAT patient dose verification, the heterogeneity between variables from multiple institutions was eliminated, achieving the goal of providing more accurate and computationally cheaper clinical PSQA.

[0125] The present application embodiment provides an electronic device, such as Figure 3As shown, the electronic device 3 includes a first processor 300, a memory 301, a bus 302 and a communication interface 303. The first processor 300, the communication interface 303 and the memory 301 are connected via the bus 302; the memory 301 stores a computer program that can be run on the first processor 300, and when the first processor 300 runs the computer program, the detection method for volumetric intensity modulated radiotherapy provided in any of the aforementioned embodiments of the present application is executed.

[0126] The memory 301 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 303 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0127] The bus 302 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The memory 301 is used to store a program, and the first processor 300 executes the program after receiving an execution instruction. The volumetric intensity modulated radiotherapy detection method disclosed in any implementation of the embodiment of the present application may be applied to the first processor 300, or implemented by the first processor 300.

[0128] The first processor 300 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the first processor 300. The above-mentioned first processor 300 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a readily available programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be embodied as a hardware decoding processor to execute, or a combination of hardware and software modules in the decoding processor to execute. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 301, and the first processor 300 reads the information in the memory 301 and completes the steps of the above method in combination with its hardware.

[0129] The electronic device provided in the above-mentioned embodiment of the present application and the detection method for volumetric intensity modulated radiation therapy provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0130] The present application embodiment provides a computer-readable storage medium, such as Figure 4 As shown, the computer-readable storage medium stores 401 a computer program, and when the computer program is read and executed by the second processor 402, the aforementioned volumetric intensity modulated radiation therapy detection method is implemented.

[0131] The technical solution of the embodiment of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling an electronic device (which may be an air conditioner, a refrigeration device, a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method described in the embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

[0132] The computer-readable storage medium provided in the above-mentioned embodiment of the present application and the detection method for volumetric intensity modulated radiation therapy provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0133] An embodiment of the present application provides a computer program product, including a computer program, wherein the computer program is executed by a third processor to implement the method described above.

[0134] The computer program product provided in the above-mentioned embodiments of the present application and the detection method for volumetric intensity modulated radiation therapy provided in the embodiments of the present application are based on the same inventive concept and have the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0135] It should be noted that, in this application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or still includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0136] Each embodiment in the present application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the detection method, electronic device, electronic device, and readable storage medium embodiment for evaluating volumetric intensity modulated radiation therapy, since they are basically similar to the above-mentioned detection method embodiment for volumetric intensity modulated radiation therapy, the description is relatively simple, and the relevant parts can be referred to the partial description of the above-mentioned detection method embodiment for volumetric intensity modulated radiation therapy.

[0137] Although the present application is disclosed as above, the present application is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application shall be subject to the scope defined by the claims.

Claims

1. A method for detecting volumetric intensity modulated radiation therapy, characterized in that: include: Get the training data set and the target data set; Preprocessing the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the institution data corresponding to different data; Build an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; Training the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; Processing the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result; The training of the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model includes: classifying the training data set with identification information based on the identification information to generate a plurality of data groups, wherein each data group is composed of data samples with the same identification information; processing the plurality of data groups respectively based on the generator of the initial deep hybrid learning model to generate fictitious data samples, wherein the fictitious data samples are the same as the data samples in the corresponding data groups, and the fictitious data samples include the same identification information as the data samples in the corresponding data groups; dividing the fictitious data samples into corresponding data groups based on the identification information to generate updated data groups; processing the plurality of updated data groups based on the discriminator of the initial deep hybrid learning model to generate classification results, wherein the classification results include the authenticity of the current data samples; when the discriminator completes the processing of all updated data groups, generating a target deep hybrid learning model; The classifying and processing the training data set with identification information based on the identification information to generate a plurality of data groups includes: dividing the training data set with identification information into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio; extracting a plurality of data groups from the training set, wherein each data group includes a preset number of data samples; The extracting of multiple data groups from the training set includes: dividing the training set based on identification information to generate a number of number class samples; obtaining any number of number class samples in the training set whose number is lower than a preset threshold; generating adjacent samples based on the distance between any number of number class samples lower than the preset threshold and other number class samples of the same category whose number is lower than the preset threshold, wherein the adjacent samples include a preset number of any number class samples lower than the preset threshold; sampling the adjacent samples to generate a preset number of sampled samples; generating multiple data groups based on any number of number class samples lower than the preset threshold and each sampled sample; The sampling of the adjacent samples to generate a preset number of sampling samples includes: determining a sampling ratio based on the number of samples of each category in the training set; determining a sampling ratio based on the sampling ratio; and sampling the adjacent samples based on the sampling ratio to generate a preset number of sampling samples.

2. The method according to claim 1, characterized in that The preprocessing of the training data set and the target data set to generate a training data set with identification information and a target data set with identification information includes: Extracting features from the training data set to determine an original feature library; Divide each feature data set according to the original feature library to generate a training data set and a test data set; Use the classifier to divide the original feature library into various test data sets for prediction and determine the prediction results; Use the preset algorithm to divide each training data set in the original feature library for training, and obtain the prediction results of the test set class; Generate a fusion feature data set based on the prediction results and the test set class prediction results; The fused feature data set is processed based on preset feature screening rules to generate identification information, wherein the identification information is used to identify the institution data corresponding to different data.

3. The method according to claim 1, characterized in that The discriminator based on the initial deep hybrid learning model processes the updated data groups to generate classification results, including: The discriminator processes several updated data sets based on preset model parameters to generate classification results; If the accuracy of the classification result is false, the preset model parameters are processed based on a back propagation algorithm to generate new preset model parameters; The discriminator processes the updated data groups based on the new preset model parameters to generate classification results; If the accuracy of the classification result is true, the current data sample is skipped and subsequent data samples are processed.

4. A detection device for volumetric intensity modulated radiation therapy, characterized in that: The device comprises: An acquisition module is used to acquire training data sets and target data sets; A processing module, used to preprocess the training data set and the target data set to generate a training data set with identification information and a target data set with identification information, wherein the identification information is used to identify the mechanism data corresponding to different data; construct an initial deep hybrid learning model based on a generative adversarial network and a deep learning model; train the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model; process the target data set based on the target deep hybrid learning model to generate a dose verification result, wherein the dose verification result includes a gamma pass rate and a classification result; The training of the initial deep hybrid learning model based on the training data set with identification information to generate a target deep hybrid learning model includes: classifying the training data set with identification information based on the identification information to generate a plurality of data groups, wherein each data group is composed of data samples with the same identification information; processing the plurality of data groups respectively based on the generator of the initial deep hybrid learning model to generate fictitious data samples, wherein the fictitious data samples are the same as the data samples in the corresponding data groups, and the fictitious data samples include the same identification information as the data samples in the corresponding data groups; dividing the fictitious data samples into corresponding data groups based on the identification information to generate updated data groups; processing the plurality of updated data groups based on the discriminator of the initial deep hybrid learning model to generate classification results, wherein the classification results include the authenticity of the current data samples; when the discriminator completes the processing of all updated data groups, generating a target deep hybrid learning model; The classifying and processing the training data set with identification information based on the identification information to generate a plurality of data groups includes: dividing the training data set with identification information into a training set for training the initial deep hybrid learning model and a test set for testing the initial deep hybrid learning model based on a preset ratio; extracting a plurality of data groups from the training set, wherein each data group includes a preset number of data samples; The extracting of multiple data groups from the training set includes: dividing the training set based on identification information to generate a number of number class samples; obtaining any number of number class samples in the training set whose number is lower than a preset threshold; generating adjacent samples based on the distance between any number of number class samples lower than the preset threshold and other number class samples of the same category whose number is lower than the preset threshold, wherein the adjacent samples include a preset number of any number class samples lower than the preset threshold; sampling the adjacent samples to generate a preset number of sampled samples; generating multiple data groups based on any number of number class samples lower than the preset threshold and each sampled sample; The sampling of the adjacent samples to generate a preset number of sampling samples includes: determining a sampling ratio based on the number of samples of each category in the training set; determining a sampling ratio based on the sampling ratio; and sampling the adjacent samples based on the sampling ratio to generate a preset number of sampling samples.

5. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the volumetric intensity modulated radiation therapy detection method according to any one of claims 1 to 3 by executing the executable instructions.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the method for detecting volumetric intensity modulated radiation therapy according to any one of claims 1 to 3 is implemented.

Images