WO2022095167A1 - 剂量确定方法及装置 - Google Patents

剂量确定方法及装置 Download PDF

Info

Publication number
WO2022095167A1
WO2022095167A1 PCT/CN2020/132278 CN2020132278W WO2022095167A1 WO 2022095167 A1 WO2022095167 A1 WO 2022095167A1 CN 2020132278 W CN2020132278 W CN 2020132278W WO 2022095167 A1 WO2022095167 A1 WO 2022095167A1
Authority
WO
WIPO (PCT)
Prior art keywords
dose
deep learning
network model
learning network
feature information
Prior art date
Application number
PCT/CN2020/132278
Other languages
English (en)
French (fr)
Inventor
周琦超
刘耀颖
曲宝林
徐寿平
陈朝才
Original Assignee
福建自贸试验区厦门片区Manteia数据科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 福建自贸试验区厦门片区Manteia数据科技有限公司 filed Critical 福建自贸试验区厦门片区Manteia数据科技有限公司
Publication of WO2022095167A1 publication Critical patent/WO2022095167A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/003—Reconstruction from projections, e.g. tomography
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2210/00—Indexing scheme for image generation or computer graphics
    • G06T2210/41—Medical

Definitions

  • the present disclosure takes the Chinese patent document with the application number of 202011226096.5 and the title of “Dose Determination Method and Device”, which was filed on November 05, 2020, as a priority document, the entire contents of which are incorporated into the present disclosure by reference.
  • the present disclosure relates to the field of dose control, and in particular, to a dose determination method and device.
  • High-accuracy dose calculation has high precision and slow speed, while low-accuracy dose calculation has low accuracy and high speed.
  • accuracy and speed is an open question.
  • Some current Eclipse radiotherapy systems use the low-accuracy anisotropic algorithm AAA (Analytical anisotropic algorithm) for optimization, and use the high-accuracy AXB algorithm (Acuros XB) or Monca algorithm for final dose calculation to determine whether the plan passes .
  • AAA Analytical anisotropic algorithm
  • AXB Acuros XB
  • Monca Monca algorithm
  • the AAA algorithm is faster but has lower accuracy than the Mongolian card and AXB. It may be that the AAA-optimized plan deviates greatly from the prescribed dose after using AXB calculation, and it needs to be re-planned.
  • the present disclosure provides a dose determination method and device to at least solve the technical problem of low computational efficiency in the clinical dose determination method in the related art.
  • a dose determination method comprising: calculating a first dose on a plan file through a first dose calculation algorithm; acquiring feature information corresponding to the plan file, wherein the feature information includes The computer tomography CT image corresponding to the plan file; input the first dose and the feature information into a deep learning network model, and the deep learning network model outputs the corresponding second dose, wherein the depth
  • the learning network model is trained by multiple sets of training samples, and each set of training samples includes the first dose and characteristic information of the input plan file, and the corresponding second dose.
  • inputting the first dose and the characteristic information into a deep learning network model, and before outputting the corresponding second dose by the deep learning network model includes: establishing a deep learning network model, and determining training Samples and test samples, wherein the test samples include the first dose and characteristic information of the plan document calculated by the first dose calculation algorithm, and the corresponding second dose calculated by the second dose calculation algorithm; according to the The training sample trains the deep learning network model; the trained deep learning network model is detected according to the test sample, and the difference between the second dose output by the deep learning network model and the second dose in the test sample does not exceed the predetermined value.
  • the difference value it is determined that the training of the deep learning network model is effective; when the difference between the second dose output by the deep learning network model and the second dose in the test sample exceeds the preset difference value, the The deep learning network model continues to be trained.
  • the method before training the deep learning network model according to the training samples, includes: adjusting the training samples through an interpolation method, so that the first dose and the second dose are equal to The physical coordinate values corresponding to the pixels of the feature information are the same.
  • detecting the trained deep learning network model according to the test sample includes: inputting the first dose and characteristic information of the test sample into the trained deep learning network to obtain the training The output second dose of the completed deep learning network is obtained and calculated by the second dose calculation algorithm; with the calculated second dose as a reference, the output second dose and the calculated second dose are calculated.
  • Gamma analysis pass rate between, and calculate and draw the dose volume histogram of the output second dose and the dose volume histogram of the calculated second dose respectively, determine the dose volume histogram of the output second dose and The difference between the dose volume histogram of the calculated second dose; the preset passing rate between the output second dose and the gamma passing rate calculated by the calculation of the second dose, and the output of the second dose If the difference between the dose-volume histogram of the second dose and the dose-volume histogram of the calculated second dose does not exceed a preset difference, it is determined that the training of the deep learning network model is effective.
  • continuing to train the deep learning network model includes: : Obtain the historical plan files other than the training samples and test samples in the system, as well as the feature information corresponding to the historical plan files, as the first update training samples, wherein the historical plan files and the objects corresponding to the plan files
  • the types of the training samples are the same; the deep learning network model is trained by the first update training sample; and/or, the feature information of the planning file of the training sample except the CT image is obtained, and the training sample is updated.
  • the feature information is used to train the deep learning network model; and/or, the network structure and model parameters of the deep learning network model are modified, and the modified deep learning network model is trained through the training samples.
  • modifying the network structure and model parameters of the deep learning network model includes at least one of the following: modifying properties of a network layer of the deep learning network model, wherein the properties of the network layer include the following At least one of: the number of network layers, the parameters of the network layer, the type of the network layer, the connection method of the network layer, the weight of the network layer; according to the result of the middle layer of the deep learning network model and the final result of the deep learning network model Modify the loss function of the deep learning network model and the structure of the intermediate layer; modify the number of training samples of the deep learning network model, and the data size of the training samples, wherein the training samples include the first A dose, a second dose, and characteristic information.
  • continuing to train the deep learning network model includes: : Acquire multiple objects of multiple types other than the objects of the training samples and the objects of the test samples in the system; establish multiple planning files corresponding to the multiple objects according to the multiple objects, and determine the multiple planning documents corresponding to the multiple objects, and the first dose, characteristic information and second dose corresponding to the planning documents; the first dose, characteristic information and second dose of the multiple planning documents are taken as the first dose, characteristic information and second dose of the multiple planning documents. Second, update the training samples; train the deep learning network model through the second update training samples until the training is valid.
  • a dose determination method comprising: acquiring multiple objects of different types; determining multiple planning files corresponding to the multiple objects, and calculating a first dose calculation algorithm by using a first dose calculation algorithm. dose; acquiring feature information of the multiple objects, wherein the feature information includes CT images of the objects; inputting the first dose and the feature information into a deep learning network model, and the The deep learning network model outputs the corresponding second dose, wherein the deep learning network model is trained from multiple groups of training samples, and each group of training samples includes the first dose and feature information of the input plan file, and the corresponding first dose. two doses.
  • a dose determination device comprising: a first calculation module configured to calculate a first dose on a plan file through a first dose calculation algorithm; a first acquisition module configured to obtain The feature information corresponding to the plan file, wherein the feature information includes the CT image corresponding to the plan file; the first output module is configured to input the first dose and the feature information into the depth A learning network model, wherein the deep learning network model outputs the corresponding second dose, wherein the deep learning network model is trained from multiple groups of training samples, and each group of training samples includes the first dose and characteristics of the input plan file information, and the corresponding second dose.
  • a dose determination device comprising: a second acquisition module configured to acquire multiple objects of different types; a second calculation module configured to determine the amount corresponding to the multiple objects a plan file, and calculates a first dose through a first dose calculation algorithm; a third acquisition module, configured to acquire characteristic information of the multiple objects, wherein the characteristic information includes the electronic computed tomography scanner CT of the object an image; a second output module, configured to input the first dose and the feature information into a deep learning network model, and the deep learning network model outputs a corresponding second dose, wherein the deep learning network model is composed of multiple
  • the training samples are formed by training each group of training samples, and each group of training samples includes the first dose and characteristic information of the input plan file, and the corresponding second dose.
  • a processor is further provided, the processor is configured to run a program, wherein when the program runs, any one of the dose determination methods described above is executed.
  • a computer storage medium is also provided, where the computer storage medium includes a stored program, wherein when the program runs, a device where the computer storage medium is located is controlled to execute any one of the above The method for determining the dose described in the item.
  • the first dose is calculated by using the first dose calculation algorithm for the plan file; the feature information corresponding to the plan file is obtained, wherein the feature information includes the CT image of the computer tomography scanner corresponding to the plan file;
  • the dose and feature information are input into the deep learning network model, and the deep learning network model outputs the corresponding second dose, wherein the deep learning network model is trained by multiple groups of training samples, and each group of training samples includes the first dose of the input plan file and feature information, as well as the corresponding second dose, by using a deep learning network model, the first dose and the feature information containing CT images are input into the neural network, and the second dose is output correspondingly.
  • the purpose of determining the second dose by including the characteristic information of the CT image thereby achieving the technical effect of improving the efficiency of determining the second dose, thereby improving the efficiency of dose optimization, and solving the clinical dose determination method in the related art, with low calculation efficiency technical issues.
  • FIG. 1 is a flowchart of a dose determination method according to an embodiment of the present disclosure
  • FIG. 3 is a flowchart of a dose determination method according to the related art
  • FIG. 4 is a flow diagram of another dose determination according to an embodiment of the present disclosure.
  • 6-1 is a schematic diagram of a model structure according to an embodiment of the present disclosure.
  • 6-2 is a schematic diagram of details of a model structure according to an embodiment of the present disclosure.
  • FIG. 7 is a schematic diagram of model augmentation training according to an embodiment of the present disclosure.
  • FIG. 8 is a schematic diagram of a dose determination device according to an embodiment of the present disclosure.
  • Figure 9 is a schematic diagram of another dose determination device according to an embodiment of the present disclosure.
  • AAA Analytical anisotropic algorithm
  • AAA has higher accuracy in dose calculation in heterogeneous areas and has the characteristics of fast speed.
  • dose deviations of more than 5% may occur.
  • the Monte Carlo algorithm simulates the transport and energy deposition of each particle (photon, electron, etc.), and Monte Carlo can be considered to have a sufficiently high dose accuracy. However, it is extremely computationally demanding, which can significantly extend dose calculation time.
  • the AXB algorithm (Acuros XB) is a new dose calculation algorithm that introduces the Boltzmann transport equation. AXB will theoretically converge to the same solution as the Monte Carlo algorithm, with Monte Carlo level accuracy, but its The speed may be ten times slower than the AAA algorithm.
  • a method embodiment of a dose determination method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings may be executed in a computer system such as a set of computer-executable instructions, and, Although a logical order is shown in the flowcharts, in some cases steps shown or described may be performed in an order different from that herein.
  • FIG. 1 is a flowchart of a dose determination method according to an embodiment of the present disclosure. As shown in FIG. 1 , the method includes the following steps:
  • Step S102 calculating the first dose on the plan file through the first dose calculation algorithm
  • Step S104 acquiring feature information corresponding to the plan file, wherein the feature information includes a computed tomography CT image corresponding to the plan file;
  • Step S106 the first dose and feature information are input into the deep learning network model, and the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple groups of training samples, and each group of training samples includes the input.
  • the first dose and characteristic information of the plan file, and the corresponding second dose are input into the deep learning network model, and the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple groups of training samples, and each group of training samples includes the input.
  • the first dose is calculated by using the first dose calculation algorithm for the plan file; the feature information corresponding to the plan file is obtained, wherein the feature information includes the CT image corresponding to the plan file; the first dose and the feature The information is input into the deep learning network model, and the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple sets of training samples, and each set of training samples includes the first dose and feature information of the input plan file , and the corresponding second dose, by using a deep learning network model, the first dose and the feature information containing CT images are input into the neural network, and the second dose is output correspondingly, so as to achieve a rapid
  • the purpose of determining the second dose based on the feature information of the image thus achieving the technical effect of improving the efficiency of determining the second dose, thereby improving the efficiency of dose optimization, and solving the technical problem of low optimization efficiency in clinical dose determination methods in related technologies .
  • the above-mentioned plan document can be a plan document for a single object, the above-mentioned object can be a different case, including patient information and radiotherapy implementation methods, the above-mentioned single object can be a case of a single disease, and the above-mentioned technical document can be the above-mentioned case.
  • the above-mentioned first dose calculation algorithm may be a rough calculation algorithm, such as the AAA algorithm and the pencil beam algorithm.
  • the above-mentioned AAA algorithm may be used. Since the accuracy of the first dose calculation algorithm is low, it is necessary to improve the accuracy of the dose through further calculation.
  • a second dose calculation algorithm with high precision is often used to finally determine the final dose distribution of the radiotherapy plan to ensure that the radiotherapy plan meets the prescription requirements. But its computational efficiency is low.
  • the above-mentioned characteristic information includes CT images, in addition to tissue structure delineation information, pathological stages, disease types, and other coarse dose calculation results (such as pencil beams).
  • the deep learning network model is trained by using the first dose of the known planning file and the feature information containing the CT image, and the corresponding second dose obtained through the second dose calculation algorithm, so that through the deep learning network model, Then, the corresponding second dose can be quickly obtained according to the first dose in the planning file and the characteristic information contained in the CT image.
  • the second dose in the training sample is the second dose obtained by calculating the plan file of the training sample through a high-precision second dose calculation algorithm, and the second dose calculation algorithm may be the AXB algorithm.
  • inputting the first dose and feature information into a deep learning network model, and before outputting the corresponding second dose by the deep learning network model includes: establishing a deep learning network model, and determining training samples and test samples, wherein , the test sample includes the first dose and characteristic information of the plan file calculated by the first dose calculation algorithm, and the corresponding second dose calculated by the second dose calculation algorithm; according to the training samples, train the deep learning network model; according to the test The sample detects the trained deep learning network model, and when the difference between the second dose output by the deep learning network model and the second dose in the test sample does not exceed the preset difference, it is determined that the deep learning network model training is effective ; Continue to train the deep learning network model when the difference between the second dose output by the deep learning network model and the second dose in the test sample exceeds a preset difference.
  • the deep learning network model is tested according to the test samples to determine whether the recognition accuracy of the deep learning network model meets the requirements.
  • the difference with the second dose in the test sample does not exceed the preset difference, it is determined that the training of the deep learning network model is effective; the difference between the second dose output by the deep learning network model and the second dose in the test sample, In the case of exceeding the preset difference, continue to train the deep learning network model, thereby improving the recognition accuracy of the deep learning network model.
  • the above-mentioned training samples include multiple groups, and the multiple groups of training samples are all known planning files, for example, planning files for which radiotherapy has been performed, including the first dose, characteristic information and the second dose.
  • the above test samples are also known plan files, but the test samples do not coincide with the multiple sets of training samples.
  • the method before training the deep learning network model according to the training samples, includes: adjusting the training samples by an interpolation method, so that the first dose and the second dose have physical coordinate values corresponding to the pixel points of the characteristic information. same.
  • the above-mentioned first dose and second dose may include a specific release position and the release dose at this position. It can be considered that the above-mentioned first dose and second dose are distributed on a specific image. Different positions on the image correspond to the dose at the position. .
  • the CT image and the first dose and the second dose of the above characteristic information are all related to the image.
  • the first dose and the second dose are processed by the interpolation method with the CT image, so that The first dose and the second dose are consistent with the pixel interval of the CT image, that is, the physical coordinate values corresponding to the pixels are the same, which can make it easier for the deep learning network model to associate learning with them, thereby improving the training efficiency of the deep learning network, and The recognition accuracy of deep learning network models.
  • detecting the trained deep learning network model according to the test sample includes: inputting the first dose and characteristic information of the test sample into the trained deep learning network, and obtaining the trained deep learning network Output the second dose, and obtain the calculated second dose through the second dose calculation algorithm; take the calculated second dose as a reference, calculate the gamma analysis pass rate between the output of the second dose and the calculated second dose, and calculate and draw the output separately
  • the dose volume histogram of the second dose and the dose volume histogram of the calculated second dose determine the difference between the dose volume histogram of the output second dose and the dose volume histogram of the calculated second dose;
  • the gamma pass rate obtained by the two-dose calculation reaches the preset pass rate, and in the case where the difference between the dose volume histogram of the output second dose and the dose volume histogram of the calculated second dose does not exceed the preset difference, determine the deep learning The network model training is valid.
  • the gamma analysis pass rate between the output second dose and the calculated second dose is calculated, and the output second dose is compared with the calculated second dose.
  • the dose volume histogram of the two doses and the dose volume histogram of the calculated second dose determine the difference between the dose volume histogram of the output second dose and the dose volume histogram of the calculated second dose.
  • the above-mentioned histogram difference can include data such as rate of change, numerical range, etc.
  • the difference between the two can be the difference of multiple data. If the difference of each data satisfies the preset difference, it is considered that the output second dose is the same as the calculated value. The difference of the second dose does not exceed the preset difference. On the contrary, if the difference of any one or more pieces of data does not meet the corresponding preset difference, it is considered that the difference between the output second dose and the calculated second dose exceeds the preset difference.
  • continuing to train the deep learning network model includes: obtaining the information in the system
  • the historical plan files other than the training samples and test samples, as well as the feature information corresponding to the historical plan files, are used as the first update training samples, wherein the historical plan files are of the same type as the objects corresponding to the plan files; through the first update of the training samples Train the deep learning network model; and/or, obtain the feature information of the training sample plan file except the CT image, update the feature information of the training sample, and train the deep learning network model; and/or, modify the deep learning
  • the network structure and model parameters of the network model, and the modified deep learning network model is trained through training samples.
  • the deep learning network model fails the detection of the test sample, it is necessary to continue training the deep learning network model.
  • continuous training can also be performed.
  • the plan file of a single object can be reselected as a secondary training sample to train the deep learning network model.
  • feature information other than CT images of the planning file of the training sample for example, tissue delineation, pathological staging, other dose distribution calculations such as 'pencil beam', deep learning dose calculation prediction results, stretching and rotating the input data
  • the feature information obtained by data enhancement methods is updated, and the feature information of the training sample is updated to train the deep learning network model.
  • the above-mentioned different types of objects may be plan documents of cases of different diseases, or plan documents of patients of different diseases.
  • modifying the network structure and model parameters of the deep learning network model includes at least one of the following: modifying the properties of the network layers of the deep learning network model, wherein the properties of the network layers include at least one of the following: the network layer The number of network layers, the parameters of the network layer, the type of network layer, the connection method of the network layer, the weight of the network layer; according to the result of the middle layer of the deep learning network model and the final output result of the deep learning network model, modify the deep learning network model.
  • the properties of the network layer of the deep learning network model can be modified in various ways.
  • the properties of the network layer include the following At least one of: the number of network layers, the parameters of the network layer, the type of the network layer, the connection method of the network layer, the weight of the network layer; according to the result of the middle layer of the deep learning network model and the final output result of the deep learning network model, Modify the loss function of the deep learning network model and the structure of the intermediate layer; modify the number of training samples of the deep learning network model, and the data size of the training samples, where the training samples include the first dose, the second dose, and feature information. This in turn improves the accuracy of the deep learning network model.
  • continuing to train the deep learning network model includes: obtaining the information in the system Multiple objects of multiple types except the objects of the training samples and the objects of the test samples; establish multiple plan files corresponding to the multiple objects according to the multiple objects, and respectively determine the number of the multiple plan files corresponding to the multiple objects.
  • One dose, feature information and second dose take the first dose, feature information and second dose of multiple plan files as the second update training sample; train the deep learning network model through the second update training sample until training efficient.
  • the above-mentioned plan files for selecting a plurality of different types of objects are used as the second update training samples to train the above-mentioned deep learning network model, which can be known plan files of different disease types, that is, a number of different disease types have already carried out implementation plans.
  • the document can also be an unimplemented plan document for different diseases. According to the object of different diseases, such as a case, a plan that may be implemented or can be implemented, or a theoretically feasible plan can be provided for the case, and a plan document can be generated.
  • the plan file determines the corresponding first dose and characteristic information, as well as the second dose, and uses the relevant data of the generated plan file as a secondary training sample to train the deep learning network model, thereby further enhancing the deep learning network model, Improve the recognition accuracy of deep learning network models.
  • the deep learning network model can also be continuously trained by increasing or decreasing the above-mentioned types of objects or the number of different types of objects as the second updated training samples.
  • FIG. 2 is a flowchart of another dose determination method according to an embodiment of the present disclosure. As shown in FIG. 2 , according to another aspect of the embodiment of the present disclosure, there is also provided a dose determination method, including:
  • Step S202 acquiring multiple objects of different types
  • Step S204 determining multiple planning files corresponding to multiple objects, and calculating the first dose through a first dose calculation algorithm
  • Step S206 acquiring feature information of multiple objects, wherein the feature information includes CT images of the objects;
  • Step S208 the first dose and feature information are input into the deep learning network model, and the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple groups of training samples, and each group of training samples includes the input.
  • the first dose and characteristic information of the plan file, and the corresponding second dose are input into the deep learning network model, and the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple groups of training samples, and each group of training samples includes the input.
  • multiple objects of different types are acquired; multiple planning files corresponding to the multiple objects are determined, and the first dose is calculated by the first dose calculation algorithm; characteristic information of the multiple objects is acquired, wherein the characteristic information includes the objects
  • the CT image of the computer tomography scanner input the first dose and feature information into the deep learning network model, and the deep learning network model outputs the corresponding second dose, wherein the deep learning network model is trained by multiple groups of training samples, each The set of training samples includes the first dose and feature information of the input plan file, and the corresponding second dose.
  • the first dose and the feature information containing the CT image are input into the neural network, and the corresponding output
  • the second dose achieves the purpose of quickly determining the second dose according to the first dose and the feature information containing the CT image, thereby achieving the technical effect of improving the efficiency of determining the second dose, thereby improving the efficiency of dose optimization, and solving the problem of
  • the clinical dose determination method in the related art has the technical problem of low calculation efficiency.
  • the deep learning method is used to input features such as AAA calculated dose and CT image into the neural network, and the corresponding AXB dose is used as the output, so as to achieve the effect of rapidly mapping the AAA to the AXB accurate dose distribution. Integrating the deep learning model into the radiotherapy planning optimization system can improve the iterative dose calculation accuracy of planning optimization, and can greatly improve the optimization efficiency and planning quality.
  • the technology became "Deep Learning Enhanced AAA Algorithm Accuracy Technology".
  • Fig. 3 is a flowchart of a dose determination method according to the related art.
  • the clinical planning optimization system first calculates the dose distribution using the AAA algorithm when the MLC (multi-page collimator) is fully opened, and optimizes the MLC, DVH (Dose-Volume Histogram) is calculated. If the DVH reaches the standard, the AXB algorithm is further used to calculate the DVH. If the DVH reaches the standard again, the plan ends. Among them, AXB calculation is time-consuming and takes up most of the planning optimization time.
  • the AAA algorithm is not sensitive to changes in tissue density, resulting in different results between the AAA algorithm and the AXB algorithm. If the DVH of the AAA algorithm is qualified but the DVH of the AXB algorithm is not qualified, the re-planning time cost is large, which greatly reduces the clinical efficiency.
  • FIG. 4 is a flow chart of another dose determination according to an embodiment of the present disclosure.
  • this embodiment uses a deep learning method to improve the calculation accuracy of the AAA algorithm to AXB level, and the time is extremely short (about 5s). Propose a new radiotherapy planning process. Using "deep learning to enhance the accuracy of AAA algorithm", instead of slow calculation processes such as AXB, while taking into account the accuracy, it greatly improves the speed of radiotherapy planning.
  • Fig. 5 is a flow chart of model training according to an embodiment of the present disclosure.
  • the cases whose radiotherapy plans have been executed are extracted from the eclipse system, the RT Plan files thereof are imported into the Eclipse system, and the AAA and AXB dose distributions are calculated respectively , the grid size (grid size) of AAA and AXB is 2.5mm ⁇ 2.5mm ⁇ 2.0mm, and saved as the corresponding RT Dose file.
  • the AAA dose and the corresponding cross-sectional CT are used as input, and the corresponding AXB dose is used as the output for deep learning training.
  • the interval between dose and CT pixels was set to 1.37mm ⁇ 1.37mm ⁇ 2.0mm by interpolation method.
  • FIG. 6-1 is a schematic diagram of a model structure according to an embodiment of the present disclosure
  • FIG. 6-2 is a schematic diagram of details of a model structure according to an embodiment of the present disclosure.
  • the 3D HD U- Net is a 3D U-Net that combines the Densenet structure, which can better learn deep-level features and three-dimensional spatial information.
  • U-NET is an encoder-decoder network, which consists of a down-sampling encoder and an up-sampling decoder, and has four large connection structures that can be set to compensate for the loss of data information in the down-sampling process.
  • U-NET neural network can better complete automatic feature extraction and result prediction
  • HD U-NET is a variant network of U-NET, which is characterized by adding a residual link structure after each convolution, so as to ensure full During the training process, all the feature information is greatly preserved
  • 3D HD U-NET is a 3D version of the HD U-NET network, which is characterized by the input data, convolution kernel, and the structure of the pooling layer are all 3D.
  • the advantage of 3D networks is that 3D spatial features can be directly extracted. This allows for better feature extraction and prediction results.
  • the gamma analysis pass rate of the "enhanced AAA" dose distribution generated by deep learning and the reference dose distribution was calculated, and the DVH difference between the two was calculated as the result analysis.
  • FIG. 7 is a schematic diagram of model enhancement training according to an embodiment of the present disclosure.
  • cases of different disease types are collected, RT Struct files of the cases are exported using the Eclipse system, and beams of different angles and apertures are set in the Eclispe system Set the beam settings and the corresponding RT Struct files to calculate the AAA dose and AXB dose distribution in the Eclipse system respectively, and save them as RT Dose files.
  • Model training, model structure, and result analysis are similar to those in (1), (2), and (3) in 1.
  • FIG. 8 is a schematic diagram of a dose determination device according to an embodiment of the present disclosure. As shown in FIG. 8 , according to another aspect of the embodiment of the present disclosure, a dose determination device is further provided, including: a first calculation module 82 , The first acquisition module 84 and the first output module 86 are described in detail below.
  • the first calculation module 82 is configured to calculate the first dose through the first dose calculation algorithm for the plan file;
  • the first acquisition module 84 is connected to the above-mentioned first calculation module 82 and is configured to obtain the characteristic information corresponding to the plan file, wherein , the characteristic information includes the CT image corresponding to the plan file;
  • the first output module 86 connected to the above-mentioned first acquisition module 84, is configured to input the first dose and the characteristic information into the deep learning network model,
  • the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple groups of training samples, and each group of training samples includes the first dose and feature information of the input plan file, and the corresponding the second dose.
  • the first dose is calculated by using the first dose calculation algorithm for the plan file; the feature information corresponding to the plan file is obtained, wherein the feature information includes the CT image of the computer tomography scanner corresponding to the plan file; the first dose and the feature The information is input into the deep learning network model, and the corresponding second dose is output by the deep learning network model, wherein the deep learning network model is trained by multiple sets of training samples, and each set of training samples includes the first dose and feature information of the input plan file , and the corresponding second dose, by using the deep learning network model, the first dose and the feature information containing CT images are input into the neural network, and the second dose is output correspondingly, so as to achieve a rapid
  • the purpose of determining the second dose based on the characteristic information achieves the technical effect of improving the efficiency of determining the second dose, thereby improving the efficiency of dose optimization, thereby solving the technical problem of low computational efficiency in the clinical dose determination method in the related art.
  • FIG. 9 is a schematic diagram of another dose determination apparatus according to an embodiment of the present disclosure. As shown in FIG. 9 , according to another aspect of the embodiment of the present disclosure, there is also provided a dose determination apparatus, including: a second acquisition module 92 , the second calculation module 94 , the third acquisition module 96 and the second output module 98 , the device will be described in detail below.
  • the second acquisition module 92 is configured to acquire multiple objects of different types; the second calculation module 94 is connected to the above-mentioned second acquisition module 92 and is configured to determine multiple planning files corresponding to the multiple objects, and calculate the first dose through the first dose calculation.
  • the algorithm calculates the first dose; the third acquisition module 96, which is connected to the above-mentioned second calculation module 94, is configured to acquire characteristic information of a plurality of objects, wherein the characteristic information includes the CT images of the objects; the second output module 98, connected to the above-mentioned third acquisition module 96, and configured to input the first dose and feature information into the deep learning network model, and output the corresponding second dose by the deep learning network model, wherein the deep learning network model is trained by multiple groups of training samples Thus, each group of training samples includes the first dose and characteristic information of the input plan file, and the corresponding second dose.
  • the second acquisition module 92 is used to acquire multiple objects of different types; the second calculation module 94 determines multiple plan files corresponding to the multiple objects, and calculates the first dose through the first dose calculation algorithm; the third acquisition module 96 obtain feature information of a plurality of objects, wherein the feature information includes the CT image of the object; the second output module 98 inputs the first dose and feature information into the deep learning network model, and the corresponding output from the deep learning network model.
  • the second dose wherein the deep learning network model is trained by multiple sets of training samples, each set of training samples includes the first dose and feature information of the input plan file, and the corresponding second dose.
  • the model inputs the first dose and the characteristic information containing CT images into the neural network, and outputs the second dose correspondingly, so as to achieve the purpose of quickly determining the second dose according to the first dose and the characteristic information containing CT images, thus achieving the improvement of The efficiency of the second dose is determined, thereby improving the technical effect of the efficiency of dose optimization, thereby solving the technical problem of low calculation efficiency in the clinical dose determination method in the related art.
  • a computer storage medium includes a stored program, wherein when the program is executed, the device where the computer storage medium is located is controlled to execute any one of the above dose determination methods.
  • a processor configured to run a program, wherein when the program runs, any one of the above-mentioned dose determination methods is executed.
  • the disclosed technical content may be implemented in other manners.
  • the device embodiments described above are only illustrative, for example, the division of the units may be a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components may be combined or Integration into another system, or some features can be ignored, or not implemented.
  • the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of units or modules, and may be in electrical or other forms.
  • the units described as separate components may or may not be physically separated, and components shown as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.
  • the above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.
  • the integrated unit if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium.
  • the technical solutions of the present disclosure can be embodied in the form of software products in essence, or the part that contributes to the prior art, or all or part of the technical solutions, and the computer software product is stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present disclosure.
  • the aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes .

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)

Abstract

一种剂量确定方法及装置，方法包括：对计划文件通过第一剂量计算算法计算第一剂量（S102）；获取计划文件中的特征信息，其中，特征信息包括计划文件对应的电子计算机断层扫描仪CT影像（S104）；将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量（S106）。解决了相关技术中临床剂量确定方法计算效率较低的技术问题。

Description

剂量确定方法及装置
本公开以2020年11月05日递交的、申请号为202011226096.5且名称为“剂量确定方法及装置”的中国专利文件为优先权文件，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及剂量控制领域，具体而言，涉及一种剂量确定方法及装置。
背景技术
目前剂量计算领域遇到问题，高准确度的剂量计算精度高速度慢，低准确度的剂量计算精度低速度快。准确度和速度之间的抉择是一个尚未解决的问题。
目前的部分Eclipse放疗系统采用低准确度的各向异性算法AAA(Analytical anisotropic algorithm)算法进行优化，使用高准确度的AXB算法(Acuros XB)或蒙卡算法进行最终剂量计算以判断该计划是否通过。AAA算法速度快但是精度相比蒙卡和AXB较低，可能AAA优化的计划使用AXB计算后与处方剂量偏差较大，则需要重新计划。
针对上述的问题，目前尚未提出有效的解决方案。
发明内容
本公开提供了一种剂量确定方法及装置，以至少解决相关技术中临床剂量确定方法，计算效率较低的技术问题。
根据本公开实施例的一个方面，提供了一种剂量确定方法，包括：对计划文件通过第一剂量计算算法计算第一剂量；获取所述计划文件对应的特征信息，其中，所述特征信息包括所述计划文件对应的电子计算机断层扫描仪CT影像；将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
在公开的一些实施例中，将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量之前，包括：建立深度学习网络模型， 确定训练样本和测试样本，其中，所述测试样本包括计划文件的通过第一剂量计算算法计算得到的第一剂量和特征信息，以及对应的通过第二剂量计算算法计算得到的第二剂量；根据所述训练样本训练所述深度学习网络模型；根据测试样本对训练完成的深度学习网络模型进行检测，在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，不超过预设差值的情况下，确定所述深度学习网络模型训练有效；在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，超过预设差值的情况下，对所述深度学习网络模型继续训练。
在公开的一些实施例中，根据所述训练样本训练所述深度学习网络模型之前，包括：通过插值法对所述训练样本进行调整，以使所述第一剂量和所述第二剂量，与特征信息的像素点对应的物理坐标值相同。
在公开的一些实施例中，根据测试样本对训练完成的深度学习网络模型进行检测包括：将所述测试样本的第一剂量和特征信息，输入所述训练完成的深度学习网络，得到所述训练完成的深度学习网络的输出第二剂量，并通过所述第二剂量计算算法得到计算第二剂量；以所述计算第二剂量作为参考，计算所述输出第二剂量与所述计算第二剂量之间的伽马分析通过率，并分别计算绘制所述输出第二剂量的剂量体积直方图和所述计算第二剂量的剂量体积直方图，确定所述输出第二剂量的剂量体积直方图和所述计算第二剂量的剂量体积直方图的差异；在所述输出第二剂量与所述计算第二剂量计算得到的伽马通过率的达到预设通过率，且在所述输出第二剂量的剂量体积直方图和所述计算第二剂量的剂量体积直方图的差异不超过预设差异的情况下，确定所述深度学习网络模型训练有效。
在公开的一些实施例中，在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，超过预设差值的情况下，对所述深度学习网络模型继续训练包括：获取系统中除所述训练样本和测试样本之外的历史计划文件，以及历史计划文件对应的特征信息，作为第一更新训练样本，其中，所述历史计划文件与所述计划文件对应的对象的种类相同；通过第一更新训练样本对所述深度学习网络模型进行训练；和/或，获取所述训练样本的计划文件的除所述CT影像之外的特征信息，更新所述训练样本的特征信息，对所述深度学习网络模型进行训练；和/或，修改所述深度学习网络模型的网络结构和模型参数，通过所述训练样本对修改后的深度学习网络模型进行训练。
在公开的一些实施例中，修改所述深度学习网络模型的网络结构和模型参数包括下列至少之一：修改所述深度学习网络模型的网络层的属性，其中，所述网络层的属性包括下列至少之一：网络层的数量，网络层的参数，网络层的种类，网络层的连接 方式，网络层的权重；根据所述深度学习网络模型的中间层的结果和所述深度学网络模型最终的输出结果，修改所述深度学习网络模型的损失函数和所述中间层的结构；修改所述深度学习网络模型的训练样本的数量，以及训练样本的数据尺寸，其中，所述训练样本包括第一剂量、第二剂量和特征信息。
在公开的一些实施例中，在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，超过预设差值的情况下，对所述深度学习网络模型继续训练包括：获取系统中除所述训练样本的对象和测试样本的对象之外的多个种类的多个对象；根据所述多个对象建立所述多个对象对应的多个计划文件，并分别确定所述多个对象对应的多个计划文件，以及所述计划文对应的第一剂量，特征信息和第二剂量；将所述多个计划文件的第一剂量，特征信息和第二剂量，作为第二更新训练样本；通过第二更新训练样本对所述深度学习网络模型进行训练，直至训练有效。
根据本公开实施例的另一方面，还提供了一种剂量确定方法，包括：获取不同种类的多个对象；确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；获取所述多个对象的特征信息，其中，所述特征信息包括所述对象的电子计算机断层扫描仪CT影像；将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
根据本公开实施例的另一方面，还提供了一种剂量确定装置，包括：第一计算模块，设置为对计划文件通过第一剂量计算算法计算第一剂量；第一获取模块，设置为获取所述计划文件对应的特征信息，其中，所述特征信息包括所述计划文件对应的电子计算机断层扫描仪CT影像；第一输出模块，设置为将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
根据本公开实施例的另一方面，还提供了一种剂量确定装置，包括：第二获取模块，设置为获取不同种类的多个对象；第二计算模块，设置为确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；第三获取模块，设置为获取所述多个对象的特征信息，其中，所述特征信息包括所述对象的电子计算机断层扫描仪CT影像；第二输出模块，设置为将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息， 以及对应的第二剂量。
根据本公开实施例的另一方面，还提供了一种处理器，所述处理器设置为运行程序，其中，所述程序运行时执行上述中任意一项所述的剂量确定方法。
根据本公开实施例的另一方面，还提供了一种计算机存储介质，所述计算机存储介质包括存储的程序，其中，在所述程序运行时控制所述计算机存储介质所在设备执行上述中任意一项所述的剂量确定方法。
在本公开实施例中，采用对计划文件通过第一剂量计算算法计算第一剂量；获取计划文件对应的特征信息，其中，特征信息包括计划文件对应的电子计算机断层扫描仪CT影像；将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量的方式，通过使用深度学习网络模型，将第一剂量与包含有CT影像的特征信息输入神经网络，对应输出第二剂量，达到了快速根据第一剂量和包含有CT影像的特征信息确定第二剂量的目的，从而实现了提高确定第二剂量的效率，进而提高剂量优化的效率的技术效果，进而解决了相关技术中临床剂量确定方法，计算效率较低的技术问题。
附图说明
此处所说明的附图用来提供对本公开的进一步理解，构成本申请的一部分，本公开的示意性实施例及其说明用于解释本发明，并不构成对本公开的不当限定。在附图中：
图1是根据本公开实施例的一种剂量确定方法的流程图；
图2是根据本公开实施例的另一种剂量确定方法的流程图；
图3是根据相关技术的剂量确定方法的流程图；
图4是根据本公开实施方式的另一种剂量确定的流程图；
图5是根据本公开实施方式的模型训练的流程图；
图6-1是根据本公开实施方式的模型结构的示意图；
图6-2是根据本公开实施方式的模型结构的细节的示意图；
图7是根据本公开实施方式的模型增强训练的示意图；
图8是根据本公开实施例的一种剂量确定装置的示意图；
图9是根据本公开实施例的另一种剂量确定装置的示意图。
具体实施方式
为了使本技术领域的人员更好地理解本公开方案，下面将结合本公开实施例中的附图，对本公开实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本公开一部分的实施例，而不是全部的实施例。基于本公开中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都应当属于本公开保护的范围。
需要说明的是，本公开的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是设置为区别类似的对象，而不必设置为描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的本公开的实施例能够以除了在这里图示或描述的那些以外的顺序实施。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，例如，包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
首先对本实施例出现的名词进行解释说明：
各向异性算法AAA(Analytical anisotropic algorithm)，广泛应用于临床放疗计划优化，相比于基于校正的计划优化方法，AAA在异质区域剂量计算准确度较高，并且具有速度快的特点。但是对于非均匀区域的AAA计算，其可能出现超过5％的剂量偏差。
蒙特卡罗算法模拟每个粒子(光子、电子等)的传输和能量沉积，可认为蒙特卡洛具有足够高的剂量准确性。但是，其对计算性能要求极高，这可能会显著延长剂量计算时间。
AXB算法(Acuros XB)是一种引入玻尔兹曼输运方程的新剂量计算算法，AXB在理论上会收敛到与蒙特卡罗算法相同的解，具有蒙特卡洛级别的准确度，但是其速度可能比AAA算法慢十倍。
根据本公开实施例，提供了一种剂量确定方法的方法实施例，需要说明的是，在附图的流程图示出的步骤可以在诸如一组计算机可执行指令的计算机系统中执行，并且，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。
图1是根据本公开实施例的一种剂量确定方法的流程图，如图1所示，该方法包括如下步骤：
步骤S102，对计划文件通过第一剂量计算算法计算第一剂量；
步骤S104，获取计划文件对应的特征信息，其中，特征信息包括计划文件对应的电子计算机断层扫描仪CT影像；
步骤S106，将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
通过上述步骤，采用对计划文件通过第一剂量计算算法计算第一剂量；获取计划文件对应的特征信息，其中，特征信息包括计划文件对应的电子计算机断层扫描仪CT影像；将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量的方式，通过使用深度学习网络模型，将第一剂量与包含有CT影像的特征信息输入神经网络，对应输出第二剂量，达到了快速根据第一剂量和包含有CT影像的特征信息确定第二剂量的目的，从而实现了提高确定第二剂量的效率，进而提高剂量优化的效率的技术效果，进而解决了相关技术中临床剂量确定方法，优化效率较低的技术问题。
上述计划文件可以为单一对象的计划文件，上述对象可以为不同病例，包括病人信息和放疗实施方法，上述单一对象可以为单一病种的病例，上述技术文件可以为上述病例的放疗计划文件，包括放疗次数和每次的放疗实施方法。
上述第一剂量计算算法可以为粗计算算法，例如AAA算法、笔形束算法，本实施例可以采用上述AAA算法，由于第一剂量计算算法的精度较低，因此需要通过进一步计算提高剂量的精度，相关技术中常采用精度较高的第二剂量计算算法，对放疗计划最终剂量分布进行最终确定，确保放疗计划达到处方要求。但是其计算效率低。
上述特征信息包括CT影像，除此之外还可以是组织结构勾画信息、病理分期、病种类别、其他粗剂量计算结果(如笔形束)。
本实施例通过已知计划文件的第一剂量和包含有CT影像的特征信息，以及对应通过第二剂量计算算法得到的第二剂量，对深度学习网络模型进行训练，从而通过深度学习网络模型，就可以根据计划文件的第一剂量和包含有CT影像的特征信息快速得到对应的第二剂量。
训练样本中的第二剂量，是通过对训练样本的计划文件通过高精度的第二剂量计算算法进行计算得到的第二剂量，第二剂量计算算法可以为AXB算法。
需要说明的是，上述深度学习网络模型输出的第二剂量与实际上述通过第二剂量计算算法得到的第二剂量之间存在一定差距，也即是深度学习网络模型由于其本身的计算方式与第二剂量计算算法的计算方式不同，其在实施时始终存在与计算得到的第二剂量之间的误差，但是在误差满足一定要求的情况下，认为深度学习网络模型输出的第二剂量与计算的第二剂量等同。
从而达到了快速根据第一剂量和包含有CT影像的特征信息确定第二剂量的目的，从而实现了提高确定第二剂量的效率，进而提高剂量优化的效率的技术效果，进而解决了相关技术中临床剂量确定方法，优化效率较低的技术问题。
在公开的一些实施例中，将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量之前，包括：建立深度学习网络模型，确定训练样本和测试样本，其中，测试样本包括计划文件的通过第一剂量计算算法计算得到的第一剂量和特征信息，以及对应的通过第二剂量计算算法计算得到的第二剂量；根据训练样本训练深度学习网络模型；根据测试样本对训练完成的深度学习网络模型进行检测，在深度学习网络模型输出的第二剂量与测试样本中的第二剂量的差，不超过预设差值的情况下，确定深度学习网络模型训练有效；在深度学习网络模型输出的第二剂量与测试样本中的第二剂量的差，超过预设差值的情况下，对深度学习网络模型继续训练。
在通过上述训练样本对深度学习网络模型进行训练的之后，根据测试样本对深度学习网络模型进行检测，以判断深度学习网络模型的是识别精度是否达到要求，在深度学习网络模型输出的第二剂量与测试样本中的第二剂量的差，不超过预设差值的情况下，确定深度学习网络模型训练有效；在深度学习网络模型输出的第二剂量与测试样本中的第二剂量的差，超过预设差值的情况下，对深度学习网络模型继续训练，从而提高深度学习网络模型的识别精度。上述训练样本包括多组，多组训练样本均为已知计划文件，例如，已经进行放疗的计划文件，包括第一剂量，特征信息和第二剂量。上述测试样本也是已知计划文件，但是测试样本与多组训练样本不重合。
在公开的一些实施例中，根据训练样本训练深度学习网络模型之前，包括：通过插值法对训练样本进行调整，以使第一剂量和第二剂量，与特征信息的像素点对应的物理坐标值相同。
上述第一剂量和第二剂量可以包括具体的释放位置和该位置的释放剂量，可以认为上述第一剂量和第二剂量是在具体的图像上分布的，图像上不同的位置，对应位置的剂量。上述特征信息的CT影像和第一剂量与第二剂量都与图像有关，在通过深度学习网络模型将其进行关联时，通过插值法对第一剂量和第二剂量，与CT影像进行 处理，使得第一剂量和第二剂量，与CT影像的像素间隔一致，也即是像素点对应的物理坐标值相同，可以更方便深度学习网络模型对其关联学习，进而提高深度学习网络的训练效率，以及深度学习网络模型的识别精度。
在公开的一些实施例中，根据测试样本对训练完成的深度学习网络模型进行检测包括：将测试样本的第一剂量和特征信息，输入训练完成的深度学习网络，得到训练完成的深度学习网络的输出第二剂量，并通过第二剂量计算算法得到计算第二剂量；以计算第二剂量作为参考，计算输出第二剂量与计算第二剂量之间的伽马分析通过率，并分别计算绘制输出第二剂量的剂量体积直方图和计算第二剂量的剂量体积直方图，确定输出第二剂量的剂量体积直方图和计算第二剂量的剂量体积直方图的差异；在输出第二剂量与计算第二剂量计算得到的伽马通过率达到预设通过率，且在输出第二剂量的剂量体积直方图和计算第二剂量的剂量体积直方图的差异不超过预设差异的情况下，确定深度学习网络模型训练有效。
在深度学习网络模型输出的输出第二剂量与测试样本中的计算第二剂量进行比较时，计算输出第二剂量与计算第二剂量之间的伽马分析通过率，输出第二剂量与计算第二剂量越相近，伽马通过率就越大，深度学习网络模型的精确度就越好，伽马通过率达到预设通过率，则认为深度学习网络模型的精度足够好；分别计算绘制输出第二剂量的剂量体积直方图和计算第二剂量的剂量体积直方图，确定输出第二剂量的剂量体积直方图和计算第二剂量的剂量体积直方图的差异。上述直方图差异可以包括变化率，数值范围等数据，二者的差别可以为多项数据的差，需要每项数据的差均满足预设差值的情况下，认为该输出第二剂量与计算第二剂量的差不超过预设差值，反之，若有其中任意一项或者多项数据的差值不满足对应的预设差值，则认为该输出第二剂量与计算第二剂量的差超过预设差值。
在公开的一些实施例中，在深度学习网络模型输出的第二剂量与测试样本中的第二剂量的差，超过预设差值的情况下，对深度学习网络模型继续训练包括：获取系统中除训练样本和测试样本之外的历史计划文件，以及历史计划文件对应的特征信息，作为第一更新训练样本，其中，历史计划文件与计划文件对应的对象的种类相同；通过第一更新训练样本对深度学习网络模型进行训练；和/或，获取训练样本的计划文件的除CT影像之外的特征信息，更新训练样本的特征信息，对深度学习网络模型进行训练；和/或，修改深度学习网络模型的网络结构和模型参数，通过训练样本对修改后的深度学习网络模型进行训练。
在深度学习网络模型未通过测试样本的检测的情况下，需要对深度学习网络模型进行继续训练。另外，在深度学习网络模型通过测试样本的检测，但是需要进一步提 高深度学习网络模型的精度的情况下，也可以进行继续训练。在进行继续训练的时候可以重新选取单一对象的计划文件，作为二次训练样本对深度学习网络模型进行训练。还可以获取训练样本的计划文件的除CT影像之外的特征信息，例如，组织勾画，病理分期，其他剂量分布计算如‘笔形束’，深度学习剂量计算预测结果，对输入数据进行拉伸旋转等数据增强手段得到的特征信息，更新训练样本的特征信息，对深度学习网络模型进行训练。还可以修改深度学习网络模型的网络结构和模型参数，通过训练样本对修改后的深度学习网络模型进行训练。还可以选取多个不同种类的对象的计划文件，作为二次训练样本对上述深度学习网络模型进行训练。上述不同种类的对象可以为不同病种的病例的计划文件，或者不同病种的病人的计划文件。
在公开的一些实施例中，修改深度学习网络模型的网络结构和模型参数包括下列至少之一：修改深度学习网络模型的网络层的属性，其中，网络层的属性包括下列至少之一：网络层的数量，网络层的参数，网络层的种类，网络层的连接方式，网络层的权重；根据深度学习网络模型的中间层的结果和深度学网络模型最终的输出结果，修改深度学习网络模型的损失函数和中间层的结构；修改深度学习网络模型的训练样本的数量，以及训练样本的数据尺寸，其中，训练样本包括第一剂量、第二剂量和特征信息。
修改深度学习网络模型的网络结构和模型参数，通过训练样本对修改后的深度学习网络模型进行训练可以通过多种方式，修改深度学习网络模型的网络层的属性，其中，网络层的属性包括下列至少之一：网络层的数量，网络层的参数，网络层的种类，网络层的连接方式，网络层的权重；根据深度学习网络模型的中间层的结果和深度学网络模型最终的输出结果，修改深度学习网络模型的损失函数和中间层的结构；修改深度学习网络模型的训练样本的数量，以及训练样本的数据尺寸，其中，训练样本包括第一剂量、第二剂量和特征信息等。进而提高深度学习网络模型的准确度。
在公开的一些实施例中，在深度学习网络模型输出的第二剂量与测试样本中的第二剂量的差，超过预设差值的情况下，对深度学习网络模型继续训练包括：获取系统中除训练样本的对象和测试样本的对象之外的多个种类的多个对象；根据多个对象建立多个对象对应的多个计划文件，并分别确定多个对象对应的多个计划文件的第一剂量，特征信息和第二剂量；将多个计划文件的第一剂量，特征信息和第二剂量，作为第二更新训练样本；通过第二更新训练样本对深度学习网络模型进行训练，直至训练有效。
上述选取多个不同种类的对象的计划文件，作为第二更新训练样本对上述深度学习网络模型进行训练，可以是不同病种的已知计划文件，也即是多个不同病种已经进 行实施计划文件，还可以为不同病种的未实施的计划文件，可以根据不同病种的对象，例如病例，针对该病例提供可能实施或者可以实施，或者理论可行的计划，生成计划文件，针对该生成的计划文件确定对应的第一剂量和特征信息，以及第二剂量，将该生成的计划文件的相关数据作为二次训练样本，对深度学习网络模型进行训练，从而进一步对深度学习网络模型进行增强，提高深度学习网络模型的识别精度。
在对深度学习网络模型的进行继续训练的过程中还可以通过增加或减少上述对象的种类或者不同种类的对象的数量，作为第二更新训练样本对深度学习网络模型进行继续训练。
图2是根据本公开实施例的另一种剂量确定方法的流程图，如图2所示，根据本公开实施例的另一方面，还提供了一种剂量确定方法，包括：
步骤S202，获取不同种类的多个对象；
步骤S204，确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；
步骤S206，获取多个对象的特征信息，其中，特征信息包括对象的电子计算机断层扫描仪CT影像；
步骤S208，将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
通过上述步骤，采用获取不同种类的多个对象；确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；获取多个对象的特征信息，其中，特征信息包括对象的电子计算机断层扫描仪CT影像；将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量的方式，通过使用深度学习网络模型，将第一剂量与包含有CT影像的特征信息输入神经网络，对应输出第二剂量，达到了快速根据第一剂量和包含有CT影像的特征信息确定第二剂量的目的，从而实现了提高确定第二剂量的效率，进而提高剂量优化的效率的技术效果，进而解决了相关技术中临床剂量确定方法，计算效率较低的技术问题。
需要说明的是，本公开实施例还提供了一种实施方式，下面对该实施方式进行详细说明。
本实施方式使用深度学习方法，将AAA计算剂量与CT影像等特征输入神经网络，对应AXB剂量作为输出，从而达到快速的AAA映射到AXB精确级别的剂量分布的效果。将该深度学习模型集成在放疗计划优化系统中，可提高计划优化迭代剂量计算准确度，可以极大提高优化的效率和计划质量。该技术成为“深度学习增强AAA算法准确度技术”。
图3是根据相关技术的剂量确定方法的流程图，如图3所示，临床计划优化系统首先在MLC(多页准直器)全开的情况下使用AAA算法计算剂量分布，并优化MLC，计算DVH(剂量体积直方图Dose-Volume Histogram)，DVH达标则进一步使用AXB算法计算DVH，DVH再次达标则计划结束。其中AXB计算比较耗时，占去了大部分计划优化时间。
AAA算法对于组织密度变化不敏感，导致AAA算法与AXB算法结果具有差异，如果AAA算法DVH合格而AXB算法DVH不合格，重新计划时间成本大，极大降低了临床效率。
图4是根据本公开实施方式的另一种剂量确定的流程图，如图4所示，本实施方式使用深度学习方法将AAA算法计算精度提高到AXB级别，并且用时极短(5s左右)。提出新的放疗计划流程。使用“深度学习增强AAA算法准确度技术”，替代AXB等慢速计算过程，兼顾准确度的同时，极大提高放疗计划速度。
1、训练单病种AAA增强模型：
(1)训练过程：
图5是根据本公开实施方式的模型训练的流程图，如图5所示，从eclipse系统中提取已执行放疗计划的病例，将其RT Plan文件导入Eclipse系统中，分别计算AAA与AXB剂量分布，AAA与AXB的grid size(网格尺寸)都为2.5mm×2.5mm×2.0mm，并保存为对应RT Dose文件。AAA剂量与对应横断面CT作为输入，对应AXB剂量作为输出，进行深度学习训练。训练之前采用插值法将剂量与CT像素间隔设为1.37mm×1.37mm×2.0mm。
(2)模型结构：
图6-1是根据本公开实施方式的模型结构的示意图，图6-2是根据本公开实施方式的模型结构的细节的示意图，如图6-1和6-2所示，3D HD U-Net是结合了Densenet结构的3D U-Net，可以较好地学习深层次特征与三维空间信息。U-NET是编码解码网络，由下采样的编码器和上采样的解码器组成，并且有四个大的连接结构，可以设置为补足下采样过程中的数据信息损失。U-NET神经网络可以较好的完成特征自动提取 和结果预测；HD U-NET是U-NET的变种网络，其特点在于在每一个卷积之后加入了残差链接结构，从而可以保证在全训练过程中，极大保留了全部的特征信息；3D HD U-NET是HD U-NET网络的3D版本，其特点在于输入数据，卷积核，池化层的结构都是3D的。3D网络的优势在于可以直接提取3D空间特征。从而可以更好地提取特征和预测结果。
(3)模型评估：
以Eclipse计算的AXB剂量分布作为参考剂量，计算深度学习生成的“加强AAA”剂量分布与参考剂量分布的伽马分析通过率，计算两者DVH差异，作为结果分析。
2、多病种，广适用性AAA增强模型训练：
图7是根据本公开实施方式的模型增强训练的示意图，如图7所示，收集不同病种的病例，使用Eclipse系统导出病例的RT Struct文件，在Eclispe系统中设置不同角度和口径的射束设置，将射束设置以及对应的RT Struct文件在Eclipse系统中分别计算AAA剂量与AXB剂量分布，并保存为RT Dose文件。
模型训练，模型结构，结果分析与1中(1)、(2)、(3)中类似。
图8是根据本公开实施例的一种剂量确定装置的示意图，如图8所示，根据本公开实施例的另一方面，还提供了一种剂量确定装置，包括：第一计算模块82，第一获取模块84和第一输出模块86，下面对该装置进行详细说明。
第一计算模块82，设置为对计划文件通过第一剂量计算算法计算第一剂量；第一获取模块84，与上述第一计算模块82相连，设置为获取所述计划文件对应的特征信息，其中，特征信息包括计划文件对应的电子计算机断层扫描仪CT影像；第一输出模块86，与上述第一获取模块84相连，设置为将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
通过上述装置，采用对计划文件通过第一剂量计算算法计算第一剂量；获取计划文件对应的特征信息，其中，特征信息包括计划文件对应的电子计算机断层扫描仪CT影像；将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量，通过使用深度学习网络模型，将第一剂量与包含有CT影像的特征信息输入神经网络，对应输出第二剂量，达到了快速根据第一剂量和包含有CT影像的特征信息确定第二剂量的目的，从而实 现了提高确定第二剂量的效率，进而提高剂量优化的效率的技术效果，进而解决了相关技术中临床剂量确定方法，计算效率较低的技术问题。
图9是根据本公开实施例的另一种剂量确定装置的示意图，如图9所示，根据本公开实施例的另一方面，还提供了一种剂量确定装置，包括：第二获取模块92，第二计算模块94，第三获取模块96和第二输出模块98，下面对该装置进行详细说明。
第二获取模块92，设置为获取不同种类的多个对象；第二计算模块94，与上述第二获取模块92相连，设置为确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；第三获取模块96，与上述第二计算模块94相连，设置为获取多个对象的特征信息，其中，特征信息包括对象的电子计算机断层扫描仪CT影像；第二输出模块98，与上述第三获取模块96相连，设置为将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
通过上述装置，采用第二获取模块92获取不同种类的多个对象；第二计算模块94确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；第三获取模块96获取多个对象的特征信息，其中，特征信息包括对象的电子计算机断层扫描仪CT影像；第二输出模块98将第一剂量和特征信息输入深度学习网络模型，由深度学习网络模型输出对应的第二剂量，其中，深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量的方式，通过使用深度学习网络模型，将第一剂量与包含有CT影像的特征信息输入神经网络，对应输出第二剂量，达到了快速根据第一剂量和包含有CT影像的特征信息确定第二剂量的目的，从而实现了提高确定第二剂量的效率，进而提高剂量优化的效率的技术效果，进而解决了相关技术中临床剂量确定方法，计算效率较低的技术问题。
根据本公开实施例的另一方面，还提供了一种计算机存储介质，计算机存储介质包括存储的程序，其中，在程序运行时控制计算机存储介质所在设备执行上述中任意一项的剂量确定方法。
根据本公开实施例的另一方面，还提供了一种处理器，处理器设置为运行程序，其中，程序运行时执行上述中任意一项的剂量确定方法。
上述本公开实施例序号仅仅为了描述，不代表实施例的优劣。
在本公开的上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有 详述的部分，可以参见其他实施例的相关描述。
在本公开所提供的几个实施例中，应该理解到，所揭露的技术内容，可通过其它的方式实现。其中，以上所描述的装置实施例仅仅是示意性的，例如所述单元的划分，可以为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，单元或模块的间接耦合或通信连接，可以是电性或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本公开各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本公开的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可为个人计算机、服务器或者网络设备等)执行本公开各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述仅是本公开的优选实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本公开原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本公开的保护范围。

Claims (11)

  1. 一种剂量确定方法，包括：
    对计划文件通过第一剂量计算算法计算第一剂量；
    获取所述计划文件对应的特征信息，其中，所述特征信息包括所述计划文件对应的电子计算机断层扫描仪CT影像；
    将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
  2. 根据权利要求1所述的方法，其中，将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量之前，包括：
    建立深度学习网络模型，确定训练样本和测试样本，其中，所述测试样本包括计划文件的通过第一剂量计算算法计算得到的第一剂量和特征信息，以及对应的通过第二剂量计算算法计算得到的第二剂量；
    根据所述训练样本训练所述深度学习网络模型；
    根据测试样本对训练完成的深度学习网络模型进行检测，在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，不超过预设差值的情况下，确定所述深度学习网络模型训练有效；
    在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，超过预设差值的情况下，对所述深度学习网络模型继续训练。
  3. 根据权利要求2所述的方法，其中，根据所述训练样本训练所述深度学习网络模型之前，包括：
    通过插值法对所述训练样本进行调整，以使所述第一剂量和所述第二剂量，与特征信息的像素点对应的物理坐标值相同。
  4. 根据权利要求2所述的方法，其中，根据测试样本对训练完成的深度学习网络模型进行检测包括：
    将所述测试样本的第一剂量和特征信息，输入所述训练完成的深度学习网络，得到所述训练完成的深度学习网络的输出第二剂量，并通过所述第二剂量计算算 法得到计算第二剂量；
    以所述计算第二剂量作为参考，计算所述输出第二剂量与所述计算第二剂量之间的伽马分析通过率，并分别计算绘制所述输出第二剂量的剂量体积直方图和所述计算第二剂量的剂量体积直方图，确定所述输出第二剂量的剂量体积直方图和所述计算第二剂量的剂量体积直方图的差异；
    在所述输出第二剂量与所述计算第二剂量计算得到的伽马通过率达到预设通过率，且在所述输出第二剂量的剂量体积直方图和所述计算第二剂量的剂量体积直方图的差异不超过预设差异的情况下，确定所述深度学习网络模型训练有效。
  5. 根据权利要求2所述的方法，其中，在深度学习网络模型输出的第二剂量与所述测试样本中的第二剂量的差，超过预设差值的情况下，对所述深度学习网络模型继续训练包括：
    获取系统中除所述训练样本和测试样本之外的历史计划文件，以及历史计划文件对应的特征信息，作为第一更新训练样本，其中，所述历史计划文件与所述计划文件对应的对象的种类相同；通过第一更新训练样本对所述深度学习网络模型进行训练；
    和/或，获取所述训练样本的计划文件的除所述CT影像之外的特征信息，更新所述训练样本的特征信息，对所述深度学习网络模型进行训练；
    和/或，修改所述深度学习网络模型的网络结构和模型参数，通过所述训练样本对修改后的深度学习网络模型进行训练。
  6. 根据权利要求5所述的方法，其中，修改所述深度学习网络模型的网络结构和模型参数包括下列至少之一：
    修改所述深度学习网络模型的网络层的属性，其中，所述网络层的属性包括下列至少之一：网络层的数量，网络层的参数，网络层的种类，网络层的连接方式，网络层的权重；
    根据所述深度学习网络模型的中间层的结果和所述深度学习网络模型最终的输出结果，修改所述深度学习网络模型的损失函数和所述中间层的结构；
    修改所述深度学习网络模型的训练样本的数量，以及训练样本的数据尺寸，其中，所述训练样本包括第一剂量、第二剂量和特征信息。
  7. 根据权利要求2所述的方法，其中，在深度学习网络模型输出的第二剂量与所述 测试样本中的第二剂量的差，超过预设差值的情况下，对所述深度学习网络模型继续训练包括：
    获取系统中除所述训练样本的对象和测试样本的对象之外的多个种类的多个对象；
    根据所述多个对象建立所述多个对象对应的多个计划文件，并分别确定所述多个对象对应的多个计划文件，以及所述计划对应的第一剂量，特征信息和第二剂量；
    将所述多个计划文件的第一剂量，特征信息和第二剂量，作为第二更新训练样本；
    通过第二更新训练样本对所述深度学习网络模型进行训练，直至训练有效。
  8. 一种剂量确定方法，包括：
    获取不同种类的多个对象；
    确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；
    获取所述多个对象的特征信息，其中，所述特征信息包括所述对象的电子计算机断层扫描仪CT影像；
    将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
  9. 一种剂量确定装置，包括：
    第一计算模块，设置为对计划文件通过第一剂量计算算法计算第一剂量；
    第一获取模块，设置为获取所述计划文件对应的特征信息，其中，所述特征信息包括所述计划文件对应的电子计算机断层扫描仪CT影像；
    第一输出模块，设置为将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
  10. 一种剂量确定装置，包括：
    第二获取模块，设置为获取不同种类的多个对象；
    第二计算模块，设置为确定多个对象对应的多个计划文件，并通过第一剂量计算算法计算第一剂量；
    第三获取模块，设置为获取所述多个对象的特征信息，其中，所述特征信息包括所述对象的电子计算机断层扫描仪CT影像；
    第二输出模块，设置为将所述第一剂量和所述特征信息输入深度学习网络模型，由所述深度学习网络模型输出对应的第二剂量，其中，所述深度学习网络模型由多组训练样本训练而成，每组训练样本包括输入的计划文件的第一剂量和特征信息，以及对应的第二剂量。
  11. 一种处理器，所述处理器设置为运行程序，其中，所述程序运行时执行权利要求1至8中任意一项所述的剂量确定方法。
PCT/CN2020/132278 2020-11-05 2020-11-27 剂量确定方法及装置 WO2022095167A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011226096.5 2020-11-05
CN202011226096.5A CN112233200A (zh) 2020-11-05 2020-11-05 剂量确定方法及装置

Publications (1)

Publication Number Publication Date
WO2022095167A1 true WO2022095167A1 (zh) 2022-05-12

Family

ID=74123329

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/132278 WO2022095167A1 (zh) 2020-11-05 2020-11-27 剂量确定方法及装置

Country Status (2)

Country Link
CN (1) CN112233200A (zh)
WO (1) WO2022095167A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116844734A (zh) * 2023-09-01 2023-10-03 福建自贸试验区厦门片区Manteia数据科技有限公司 剂量预测模型的生成方法、装置、电子设备及存储介质
CN117456315A (zh) * 2023-12-26 2024-01-26 福建自贸试验区厦门片区Manteia数据科技有限公司 剂量预测模型的训练方法、装置及计算机可读存储介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113101548B (zh) * 2021-04-20 2023-04-28 中山大学肿瘤防治中心 一种降低皮肤剂量的光子调强放疗控制方法
CN113797450A (zh) * 2021-05-27 2021-12-17 苏州雷泰医疗科技有限公司 一种放疗计划强度分布优化方法、装置及放射治疗设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107875525A (zh) * 2017-11-10 2018-04-06 上海联影医疗科技有限公司 计算放射剂量的方法、装置、计算机设备及可读存储介质
CN109069858A (zh) * 2015-12-21 2018-12-21 医科达股份有限公司 用于优化治疗计划的系统和方法
CN109621228A (zh) * 2018-12-12 2019-04-16 上海联影医疗科技有限公司 放射剂量的计算设备、装置和存储介质
CN109843377A (zh) * 2016-09-07 2019-06-04 医科达有限公司 用于预测放射疗法剂量分布的放射疗法治疗计划的学习模型的系统和方法
CN110124214A (zh) * 2019-05-24 2019-08-16 南方医科大学 基于预测剂量分布引导的调强放疗计划优化方法及应用
US20200075148A1 (en) * 2018-08-31 2020-03-05 The Board Of Regents Of The University Of Texas System Deep learning based dosed prediction for treatment planning and quality assurance in radiation therapy

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109979564A (zh) * 2017-12-28 2019-07-05 北京连心医疗科技有限公司 一种智能放疗计划方法、设备和存储介质
KR102220029B1 (ko) * 2018-10-12 2021-02-25 한국과학기술원 뉴럴 네트워크를 이용한 비매칭 저 선량 엑스선 전산단층 촬영 영상 처리 방법 및 그 장치
CN111462916B (zh) * 2020-04-03 2024-03-26 戴建荣 一种放疗计划危及器官剂量体积直方图的预测方法和装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109069858A (zh) * 2015-12-21 2018-12-21 医科达股份有限公司 用于优化治疗计划的系统和方法
CN109843377A (zh) * 2016-09-07 2019-06-04 医科达有限公司 用于预测放射疗法剂量分布的放射疗法治疗计划的学习模型的系统和方法
CN107875525A (zh) * 2017-11-10 2018-04-06 上海联影医疗科技有限公司 计算放射剂量的方法、装置、计算机设备及可读存储介质
US20200075148A1 (en) * 2018-08-31 2020-03-05 The Board Of Regents Of The University Of Texas System Deep learning based dosed prediction for treatment planning and quality assurance in radiation therapy
WO2020047536A1 (en) * 2018-08-31 2020-03-05 Board Of Regents, University Of Texas System Deep learning based dosed prediction for treatment planning and quality assurance in radiation therapy
CN109621228A (zh) * 2018-12-12 2019-04-16 上海联影医疗科技有限公司 放射剂量的计算设备、装置和存储介质
CN110124214A (zh) * 2019-05-24 2019-08-16 南方医科大学 基于预测剂量分布引导的调强放疗计划优化方法及应用

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116844734A (zh) * 2023-09-01 2023-10-03 福建自贸试验区厦门片区Manteia数据科技有限公司 剂量预测模型的生成方法、装置、电子设备及存储介质
CN116844734B (zh) * 2023-09-01 2024-01-16 福建自贸试验区厦门片区Manteia数据科技有限公司 剂量预测模型的生成方法、装置、电子设备及存储介质
CN117456315A (zh) * 2023-12-26 2024-01-26 福建自贸试验区厦门片区Manteia数据科技有限公司 剂量预测模型的训练方法、装置及计算机可读存储介质
CN117456315B (zh) * 2023-12-26 2024-04-19 福建自贸试验区厦门片区Manteia数据科技有限公司 剂量预测模型的训练方法、装置及计算机可读存储介质

Also Published As

Publication number Publication date
CN112233200A (zh) 2021-01-15

Similar Documents

Publication Publication Date Title
WO2022095167A1 (zh) 剂量确定方法及装置
CN109598727B (zh) 一种基于深度神经网络的ct图像肺实质三维语义分割方法
Gao et al. Classification of CT brain images based on deep learning networks
CN110503630B (zh) 一种基于三维深度学习模型的脑出血分类、定位与预测方法
Joshi et al. Classification of brain cancer using artificial neural network
CN108615237A (zh) 一种肺部图像处理方法及图像处理设备
CN107480702B (zh) 面向hcc病理图像识别的特征选择与特征融合方法
CN110956626B (zh) 一种基于图像的预后评估方法及装置
CN108766563A (zh) 基于剂量组学的放射治疗结果预测方法和系统
WO2020199836A1 (zh) 食管放射治疗计划中风险器官剂量学评估方法、评估系统
CN111462916B (zh) 一种放疗计划危及器官剂量体积直方图的预测方法和装置
CN110751187B (zh) 异常区域图像生成网络的训练方法和相关产品
CN110337670B (zh) 一种基于磁共振图像的脑龄测试方法及脑龄测试装置
JP7054278B1 (ja) 深層学習に基づくエッジ識別方法
CN108648182A (zh) 一种基于分子亚型的乳腺癌核磁共振图像肿瘤区域分割方法
CN108550131B (zh) 基于特征融合稀疏表示模型的sar图像车辆检测方法
CN109920512B (zh) 一种三维剂量分布网络模型的训练方法及装置
Zhang et al. ARPM‐net: a novel CNN‐based adversarial method with Markov random field enhancement for prostate and organs at risk segmentation in pelvic CT images
CN115439654B (zh) 动态约束下的弱监督农田地块精细化分割方法及系统
CN112560710A (zh) 一种用于构建指静脉识别系统的方法及指静脉识别系统
Yu et al. A recognition method of soybean leaf diseases based on an improved deep learning model
Zhao et al. BASCNet: Bilateral adaptive spatial and channel attention network for breast density classification in the mammogram
CN115295154A (zh) 肿瘤免疫治疗疗效预测方法、装置、电子设备及存储介质
CN112669319B (zh) 一种多视角多尺度淋巴结假阳性抑制建模方法
CN106770007B (zh) 一种支持向量机模型的近红外光谱特征波长选择方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20960624

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20960624

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 20.10.2023)