WO2020244172A1

WO2020244172A1 - Plan implementation method and device based on predicted dose guidance and gaussian process optimization

Info

Publication number: WO2020244172A1
Application number: PCT/CN2019/121146
Authority: WO
Inventors: 文虎儿; 鞠垚; 姚毅
Original assignee: 苏州雷泰智能科技有限公司
Priority date: 2019-06-04
Filing date: 2019-11-27
Publication date: 2020-12-10
Also published as: CN110232964A; CN110232964B

Abstract

The present invention relates to the technical field of radiation therapy, and provides a plan implementation method and device based on predicted dose guidance and Gaussian process optimization. The method comprises: calculating the predicted dose of a case by using a dose prediction model; calculating, according to the scoring rule, the planned score of the predicted dose as the best planned score; according to organ anatomy information, determining a plurality of groups of planning parameters on the basis of a prior database; calculating the planned scores corresponding to the plurality of groups of planning parameters, and forming a Gaussian data set; and calculating a new planning parameter on the basis of the Gaussian data set by using the Gaussian process, calculating the corresponding planned score, adding same to the Gaussian data set, iteratively performing the step, finally calculating the intensity modulation optimization result in the planning parameter corresponding to the highest score in the Gaussian data set. The prediction model is used to predict the dose distribution of the case to optimize the guidance, the quality of the plan is ensured, the posterior distribution is calculated on the basis of prior data by using the Gaussian process, and the number of trials and errors is reduced, thereby accelerating the optimization speed.

Description

Plan realization method and device based on predicted dose guidance and Gaussian process optimization

Technical field

The invention relates to the technical field of radiotherapy, in particular to a plan realization method and device based on predicted dose guidance and Gaussian process optimization.

Background technique

At present, intensity-modulated radiotherapy technology has been widely used in clinical practice, but as its core treatment plan design method, there are still many problems. The current manual trial-and-error method has severely limited the hospital’s work efficiency and planned completion cost, increasing The burden of hospitals and patients. More importantly, there is great uncertainty in the quality of the treatment plan obtained by such a plan-making method, which largely depends on the experience of the plan designer and the time spent designing the treatment plan. At the same time, according to the research reports of different treatment centers, it can be found that there are big differences in the quality of the plan and the time of the plan design, whether within the treatment center or between the treatment centers. This difference also makes the evaluation of the plan quality different. The cooperation, experience exchange and data sharing of the center have caused difficulties. Therefore, the introduction and development of automatic plans have urgent practical significance and may also be another revolution in the history of radiotherapy.

There are currently three main ways to realize automatic planning:

1) Automatic planning based on prior knowledge (Knowledge-based planning, KBP), using prior knowledge and experience to predict the optimal dose distribution of new patients or the initial value for later manual planning, mainly based on atlas and Two implementation methods based on model. The main disadvantage of this implementation method is that it only predicts the dose distribution in the region of interest, and the optimal dose distribution cannot be calculated well for undrawn tissues and organs, and the planning quality of new patients is severely limited by the planning quality of previous cases .

2) Protocol-based Automatic Iterative Optimisaztion (PB-AIO), which sets constraints on the target area and organs at risk, uses the initial template to start optimization, and continuously adjusts constraints and weights in the iterative process to reach the target Optimal distribution of doses to areas and organs at risk. This method is limited by the initial setting of the template by the physicist.

3) Multi-Criteria Optimisation (MCO) aims to find a balance between the constraints of multiple organs until the optimization of any organ destroys other constraints. There are two ways to achieve it: a priori and a posteriori. The main disadvantage of this method is that the optimal plan generated is a plan based on the intensity map, and the realization of machine parameters is not considered, and there will be a loss of plan quality when generating the final position of the subfield.

Summary of the invention

The purpose of the present invention is to provide a plan realization method and device based on predicted dose guidance and Gaussian process optimization in order to solve the optimization problem of radiotherapy plan in view of the above-mentioned deficiencies in the prior art.

In order to achieve the above objectives, the technical solutions adopted by the present invention are as follows:

In the first aspect, the present invention provides a plan realization method based on predicted dose guidance and Gaussian process optimization, which is used for the optimal design of intensity-modulated radiation therapy plans. The method includes:

a) Based on the computerized tomography and delineation data of the case to be treated, the trained dose prediction model is used to calculate the predicted dose of the case;

b) Calculate the plan score of the predicted dose according to the predetermined scoring rule as the best plan score;

c) According to the organ anatomy information of the case, based on the historical data in the associated prior database, determine the multiple sets of planning parameters of the case;

d) Calculate the plan scores corresponding to the multiple sets of plan parameters, and form a Gaussian data set by the plan parameters and the corresponding plan scores;

e) When there is a plan score with a score higher than the best plan score in the Gaussian data set, calculate the strength adjustment optimization result under the plan parameters corresponding to the plan score, otherwise, continue to step f);

f) Based on the Gaussian data set, use the Gaussian process to calculate a new plan parameter, and calculate the new plan score corresponding to the plan parameter, and add the new plan parameter and the corresponding new plan score to the Gaussian data set;

g) Perform step f) iteratively until the preset iteration termination condition is met, and calculate the intensity adjustment optimization result under the plan parameters corresponding to the highest plan score in the Gaussian data set.

Optionally, step a) specifically includes:

Establish a deep learning convolutional neural network model;

Extract the computerized tomography and delineation data and dose data of the preset disease type in the preset case database. The computerized tomography and delineation data include the delineation data of the skin and key organs in the computerized tomography image of the preset disease type;

Use the delineation data of the organ as the input of the model and the dose data as the output of the model, and train the model to obtain a trained dose prediction model;

Aiming at the computerized tomography and delineation data of the cases to be treated, the trained dose prediction model is used to calculate the predicted dose of the cases to be treated.

Optionally, step b) specifically includes:

Based on predetermined scoring rules, calculate the dose volume histogram index score of each organ at risk of the case;

Sum the indicator scores and get the total score as the best plan score.

Optionally, step c) specifically includes:

Based on the delineation data, extract the histogram of the overlapping volume of the organ at risk of the case;

Calculate the similarity between the overlap volume histogram of the organ at risk and the overlap volume histogram in the historical data in the prior database;

From the historical data in the prior database, a predetermined number of optimized parameters with the highest similarity to the histogram of the overlapping volume of the organ at risk are selected as the planning parameters of the case. The optimized parameters include the field angle and constraint conditions of radiotherapy.

Optionally, using Gaussian process to calculate new plan parameters in step f) specifically includes:

Calculate the probability density function of the Gaussian distribution of the plan score under any plan parameter in the Gaussian data set;

Based on the probability density function, the value of the corresponding acquisition function is calculated for multiple discrete parameters in the preset prediction parameter space, and the parameter corresponding to the largest value among the values of the acquisition function is selected as the new planning parameter. The acquisition function is Preset function.

Optionally, the preset iteration termination condition of iterative execution step f) is: if there is a plan score in the Gaussian dataset with a score higher than the best plan score or the iteration number exceeds the preset number of iterations, the iteration is terminated.

Optionally, after step g), the method further includes: outputting the obtained intensity modulation optimization result.

Optionally, the extraction of the electronic computed tomography and delineation data and dose data of the preset disease type in the preset case database includes:

Select 256*256 sampling points for a computerized tomography image of the preset disease type in the preset case database;

Extract the computed tomography values on the sampling points to form a computed tomography value matrix;

Extract the delineation data of the key organ on the sampling point as follows: For any organ, if one of the sampling points belongs to the organ, the delineation data value is 1, otherwise the delineation data value is 0;

Extract the delineation data of the skin on the sampling points as follows: if one of the sampling points belongs to the skin, the delineation data value is 1, otherwise the delineation data value is 0, thus forming a matrix of skin delineation, and then delineating the skin The matrix is multiplied by the numerical value of the corresponding position of the electronic computed tomography value matrix as the skin delineation data;

Extract the dose at the sampling point to form a dose matrix to obtain dose data.

In a second aspect, the present invention provides a plan realization device based on predicted dose guidance and Gaussian process optimization, which is used for the optimization design of intensity-modulated radiation therapy plans, and the device includes:

The predicted dose calculation module is used to calculate and obtain the predicted dose of the case based on the electronic computed tomography and delineation data of the case to be treated by using a trained dose prediction model;

The best plan score calculation module is used to calculate the predicted dose plan score as the best plan score according to predetermined scoring rules;

The planning parameter determination module is used to determine multiple sets of planning parameters of the case based on the historical data in the associated prior database according to the organ anatomy information of the case;

The Gaussian data set forming module is used to calculate the plan scores corresponding to the multiple sets of plan parameters, and the Gaussian data set is formed by the plan parameters and the corresponding plan scores;

The Gaussian data set iterative update module is used to calculate new plan parameters based on the Gaussian data set using the Gaussian process, and calculate the new plan score corresponding to the plan parameter, and add the new plan parameter and the corresponding new plan score to Gaussian data set;

The best plan result output module is used to select the plan with the highest score in the Gaussian data set when there is a plan with a score higher than the best plan score in the Gaussian data set, or when the number of iterations reaches the preset condition, and calculate the plan parameters corresponding to the plan score Optimized results of the intensity adjustment under.

Optionally, the predicted dose calculation module is specifically used for:

Establish a deep learning convolutional neural network model;

The delineation data of the organ is used as the input of the model, and the dose data is used as the output of the model, and the model is trained to obtain a trained dose prediction model;

Optionally, the optimal plan score calculation module is specifically used for:

Sum the indicator scores and get the total score as the best plan score.

Optionally, the plan parameter determination module is specifically used for:

Optionally, the Gaussian data set iterative update module is specifically used to:

Optionally, the device further includes an intensity modulation optimization result output module for outputting the obtained intensity modulation optimization result.

Optionally, the predicted dose calculation module is specifically used to:

Extract the computed tomography values at the sampling points to form a computed tomography value matrix;

The beneficial effects of the present invention include:

The plan realization method provided by the present invention includes: based on the electronic computed tomography and delineation data of the case to be treated, a trained dose prediction model is used to calculate the predicted dose of the case; and the planned score of the predicted dose is calculated according to a predetermined scoring rule as the best plan Score; according to the organ anatomy information of the case, based on the historical data in the associated prior database, determine the multiple sets of planning parameters of the case; calculate the plan scores corresponding to the multiple sets of plan parameters, and the plan parameters and the corresponding plan scores Form a Gaussian data set; when there is a plan score with a score higher than the best plan score in the Gaussian data set, calculate the intensity adjustment optimization result under the plan parameters corresponding to the plan score; otherwise, continue with the following steps; based on the Gaussian data set, use The Gaussian process calculates a new plan parameter, and calculates the new plan score corresponding to the plan parameter, and adds the new plan parameter and the corresponding new plan score to the Gaussian data set; iteratively executes the previous step until the preset iteration termination is met Conditions so far, and calculate the intensity adjustment optimization results under the plan parameters corresponding to the highest plan score in the Gaussian data set. By adopting the dose prediction model to predict the dose distribution of new cases, it can be used to optimize guidance, ensuring the quality of the plan to a certain extent, and then using the Gaussian process to calculate the posterior distribution based on the prior data and observations, and predict the best parameter calculation Point, reduce the number of trial and error, thereby speeding up the optimization speed.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present invention more clearly, the following will briefly introduce the drawings that need to be used in the embodiments. It should be understood that the following drawings only show certain embodiments of the present invention and therefore do not It should be regarded as a limitation of the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative work.

FIG. 1 shows a schematic flowchart of a plan realization method based on predicted dose guidance and Gaussian process optimization according to an embodiment of the present invention;

Figure 2 shows a schematic structural diagram of a U-net 2D delineation-dose model provided by an embodiment of the present invention;

FIG. 3 shows a schematic diagram of a training process of a deep convolutional neural network model provided by an embodiment of the present invention;

Fig. 4 shows a schematic diagram of a dose prediction process of a deep convolutional neural network model provided by an embodiment of the present invention;

FIG. 5 shows a schematic diagram of a process for obtaining k-nearest neighbor optimization parameters according to an embodiment of the present invention;

FIG. 6 shows a schematic flowchart of a method for obtaining a new prediction point by a Gaussian process according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the protection scope of the present invention.

IMRT technology has been widely used in clinical practice, but as its core treatment plan design method, there are still many problems. The current iterative manual trial and error method is time-consuming and laborious, which severely limits the hospital’s work efficiency and planning Completion costs increase the burden on hospitals and patients. More importantly, there is great uncertainty in the quality of the treatment plan obtained by such a plan-making method, which largely depends on the experience of the plan designer and the time spent designing the treatment plan.

This paper proposes a plan realization method based on predictive dose guidance and Gaussian process optimization for the optimization design of intensity-modulated radiation treatment plans. The method includes: a) Based on the computer tomography and delineation data of the case to be treated, using training A good dose prediction model calculates the predicted dose of the case; b) calculates the planned score of the predicted dose according to the predetermined scoring rule as the best plan score; c) according to the case’s organ anatomy information, based on historical data in the associated prior database , Determine the multiple sets of planning parameters of the case; d) Calculate the plan scores corresponding to the multiple sets of plan parameters, and form a Gaussian data set by the plan parameters and the corresponding plan scores; e) There are scores higher than the best plan in the Gaussian data set When scoring the plan score, calculate the intensity adjustment optimization result under the plan parameter corresponding to the plan score, otherwise, continue to step f); f) Based on the Gaussian data set, use the Gaussian process to calculate the new plan parameter, and calculate the plan parameter Corresponding to the new plan score, and adding the new plan parameters and the corresponding new plan score to the Gaussian data set; g) Iteratively execute step f) until the preset iteration termination conditions are met, and calculate the highest plan in the Gaussian data set The result of intensity adjustment optimization under the plan parameters corresponding to the score.

The present invention uses a dose prediction model to predict the dose distribution of a new case, which can be used for optimizing guidance, ensuring the quality of the plan to a certain extent, and then using the Gaussian process to calculate the posterior distribution based on the prior data and observation values, and predict the best Parameter calculation points reduce the number of trial and error, thus speeding up optimization.

The plan implementation method proposed by the present invention will be described in detail below with reference to FIG. 1.

Firstly, based on the prior database, the optimal dose distribution of the current case is predicted by the dose prediction model. Specifically, a priori database can be used to train the model, and then, based on the CT scan and delineation data of the case to be treated, the trained dose prediction model is used to calculate the predicted dose of the case.

The specific steps of training the dose prediction model are as follows: establish a deep learning convolutional neural network model. For example, the embodiment of the present invention can use a U-net 2D network to establish a deep learning convolutional neural network model (it should be understood that it can also be used in the field. Know other networks to build a deep learning convolutional neural network model), the model structure is shown in Figure 2; extract the {CT outline, dose} data of any disease in the prior database, and outline the outline mainly including the skin and key organs . For a CT, select 256*256 sampling points, 1) extract the CT value on the sampling point and save it as a CT matrix; 2) extract the organ delineation data on the sampling point, for any organ, if a certain point belongs to the organ , The value is 1, otherwise it is 0; 3) Multiply the matrix of the skin delineation with the value of the corresponding position of the CT value matrix to form the matrix of the skin delineation; 4) Extract the dose of the sampling point and save it as a dose matrix. The organ delineation matrix is used as the model input, and the dose matrix is used as the model output to train the network model, as shown in Figure 3. Among them, skin delineation is used to reduce the data range, and the combination of skin delineation and CT matrix not only preserves the effective data range, but also takes into account the CT values of different parts.

After the model training is completed, the CT outline data of the new case can be obtained and used as the input of the model. After the trained network, the dose can be predicted and output, as shown in Figure 4.

In the embodiment of the present invention, a priori database is formed by collecting a large number of historical cases, and a convolutional neural network structure is adopted to predict the optimal dose distribution of a new case, which can be used for optimizing guidance and ensuring the quality of the plan to a certain extent.

Then, according to the predicted dose, the plan score of the predicted dose can be calculated as the best plan score Score_best. The plan scoring rules are shown in Table 1, and the scoring method is as follows: Calculate the doses of each organ at risk according to the scoring standard in Table 1. The scores of the volume histogram (DVH) indicators are then summed to obtain the total score of DVH of all organs at risk.

Table 1 Plan scoring rules table

Taking the Rectum DVH index V75 in the prostate cancer plan as an example, the method of calculating the score is as follows: 1) If V75<10%, the score is 5; 2) If V75>15%, the score is 0; 3) If V75 In the interval [10%, 15%], use linear interpolation to obtain the score.

According to the anatomical information of the organ of the new case, it is associated with several cases with the closest anatomical information in the prior database, and the initial optimization parameter set (including multiple sets of planning parameters) for the new case is predicted and determined. See Figure 5. The implementation method is as follows: Information, extract the Overlap Volume Histogram (OVH) of the organ, the calculation method is as follows:

Among them, T is the target area; O is the organ at risk; |O| is the volume of the organ at risk; p is a subset of O; d(p, T) is the distance from p to the tumor boundary; {p∈O|d (p,T)≤t} represents the collection of voxels whose distance to the target area T is less than the distance t in the endangered organ O. The histogram function of the overlapping volume of the target area T and the organ at risk O is the volume fraction of the distance between the organ at risk O and the target area T less than t.

Calculate the similarity between the OVH of each organ of the new case and the case OVH in the database. The similarity measurement formula used in this article is the angle cosine distance, also called cosine similarity, which uses the cosine value of the angle between two vectors in the vector space as A measure of the size of the difference between two individuals. The closer the cosine value is to 1, the closer the angle is to 0 degrees, that is, the more similar the two vectors are. Calculated as follows:

Among them, a·b represents the dot product of two vectors, and |a| and |b| represent the length of the vector.

Then, the optimized parameters of the k groups of cases with the highest similarity are selected to form an optimized parameter set of k neighboring cases. The optimization parameters here include the field angle and constraint conditions.

Calculate the planned scores under the current k groups of optimization parameters, add {optimization parameter Para, score Score} to the data set T, and set the number of iterations iter=0.

If there is an optimization parameter with a score higher than Score_best in the data set T, use this parameter in a new case and calculate the intensity modulation optimization result; otherwise, continue with the following steps.

The optimized parameters of the new prediction point can be calculated through the Gaussian process, and will be added to the data set T, see Figure 6. The implementation method is as follows: Assuming that the plan score and the optimization parameters meet the multi-dimensional Gaussian distribution, given the data set T = {x = Para, y = Score}, the probability density of the Gaussian distribution under any x _* = Para, y _* = Score can be calculated function:

p(y _* |x _* ,x,y)=N(y _* |μ _* ,∑ _* )

K=k(x, x)

K _* ＝k(x, x _* )

K _** ＝k(x _* ，x _* )

x=[x ₁ , x ₂ ,...x _n ] ^T , y=[y ₁ , y ₂ ,...y _n ] ^T , y _i = f(x _i )

Among them, m(x) represents the mean value of the x sequence, and m(x _i )=0; N(y _* |μ _* , ∑ _* ) indicates that y _* satisfies the Gaussian distribution, the mean is μ _* , and the variance matrix is ∑ _* ; k represents the kernel function used to calculate the mean and variance, which can be used for curve smoothing. When x=x', the kernel function k(x,x') is equal to 1, and the greater the difference between x and x', the more k tends to 0. The acquisition function is defined as follows: f _a (x)=μ+∑. Select a number of discrete parameter points x _* in the prediction parameter space, calculate the collection function values respectively, and select the parameter corresponding to the largest value of f _a as the new prediction point x _t . Calculate the plan score Score_t under the parameter x _t . Add {parameter x _t , Score_t} into the Gaussian data set T.

If Score_t is higher than Score_best, or the number of iterations iter>predetermined number (for example, 100), use the parameter with the highest planned score in the data set T for the new case, and calculate the intensity modulation optimization result; otherwise, iter will increase by 1 and continue to execute The above-mentioned Gaussian process is the step of calculating the optimized parameters of the new prediction point.

The embodiment of the present invention adopts optimization based on the Gaussian process. The Gaussian process can calculate the posterior distribution based on the prior data and the observation value, and then predict the optimal calculation point, reduce the number of trial and error, and accelerate the optimization speed.

Finally, the optimized results of intensity modulation can be output.

In summary, the above embodiments of the present invention can automatically determine the plan parameters to be explored by applying the predicted dose and the Gaussian process in the intensity modulation optimization. It is expected that the optimal plan parameters can be found within a limited time, and the plan quality will be significantly improved. Speed up plan production efficiency.

In addition, the embodiment of the present invention also provides a plan realization device based on predicted dose guidance and Gaussian process optimization, which is used for the optimization design of intensity-modulated radiation therapy plans. Specifically, the device is used to implement the above-mentioned embodiments of the present invention. Methods. The device includes: a predicted dose calculation module, which is used to calculate the predicted dose of the case based on the electronic computed tomography and delineation data of the case to be treated using a trained dose prediction model; the best plan score calculation module is used to Calculate the plan score of the predicted dose according to the predetermined scoring rule as the best plan score; the plan parameter determination module is used to determine the multiple sets of plan parameters of the case according to the organ anatomy information of the case and the historical data in the associated prior database; The Gaussian data set forming module is used to calculate the plan scores corresponding to the multiple groups of plan parameters, and the Gaussian data set is formed by the plan parameters and the corresponding plan scores; the Gaussian data set iterative update module is used to use Gaussian data sets based on the Gaussian data set The process calculates a new plan parameter, and calculates the new plan score corresponding to the plan parameter, and adds the new plan parameter and the corresponding new plan score to the Gaussian data set; the best plan result output module is used in the Gaussian data When there is a plan with a score higher than the best plan score, or when the number of iterations reaches a preset condition, the plan with the highest score in the Gaussian data set is selected, and the strength adjustment optimization result under the plan parameters corresponding to the plan score is calculated.

Optionally, the predicted dose calculation module is specifically used to: establish a deep learning convolutional neural network model. For example, the embodiment of the present invention can use a U-net 2D network to establish a deep learning convolutional neural network model (it should be understood that it can also be used Other networks known in the art are used to build deep learning convolutional neural network models); the computer tomography and delineation data and dose data of the preset disease types in the preset case database are extracted, and the computer tomography and delineation data include pre Set the delineation data of the skin and key organs in the electronic computed tomography image of the disease; use the delineation data of the organ as the input of the model, and the dose data as the output of the model, and train the model to obtain the trained dose prediction Model: Aiming at the computerized tomography and delineating data of the case to be treated, the trained dose prediction model is used to calculate the predicted dose of the case to be treated.

Optionally, the best plan score calculation module is specifically used to calculate the dose volume histogram index score of each organ at risk of the case based on a predetermined scoring rule; sum the index scores to obtain the total score as the best plan score.

Optionally, the planning parameter determination module is specifically used to: extract the overlap volume histogram of the organ-at-risk of the case based on the delineation data; calculate the overlap volume histogram of the organ-at-risk and the overlap volume histogram in the historical data in the prior database Similarity: From the historical data in the prior database, a predetermined number of optimized parameters with the highest similarity to the histogram of the overlapping volume of the organ at risk are selected as the planning parameters of the case. The optimized parameters include the radiation field angle and constraint conditions.

Optionally, the Gaussian data set iterative update module is specifically used to: calculate the probability density function of the Gaussian distribution of the plan score under any plan parameter in the Gaussian data set; based on the probability density function, for multiple preset prediction parameter spaces The discrete parameters are respectively calculated for the value of the corresponding acquisition function, and the parameter corresponding to the largest value among the values of the acquisition function is selected as the new planning parameter, and the acquisition function is the preset function.

The above-mentioned embodiments are only to illustrate the technical concept and characteristics of the present invention, and their purpose is to enable those of ordinary skill in the art to understand the content of the present invention and implement it, and cannot limit the protection scope of the present invention. All equivalent changes or modifications should be covered by the protection scope of the present invention.

Claims

A plan realization method based on predicted dose guidance and Gaussian process optimization, which is used for the optimal design of intensity-modulated radiation therapy plans, characterized in that the method includes:

a) Based on the computerized tomography and delineation data of the case to be treated, the trained dose prediction model is used to calculate the predicted dose of the case;

b) Calculate the planned score of the predicted dose according to the predetermined scoring rule as the best plan score;

c) According to the organ anatomy information of the case, based on the historical data in the associated prior database, determine the multiple sets of planning parameters of the case;

d) Calculate the plan scores corresponding to the multiple sets of plan parameters, and form a Gaussian data set by the plan parameters and the corresponding plan scores;

e) When there is a plan score with a score higher than the best plan score in the Gaussian data set, calculate the intensity adjustment optimization result under the plan parameters corresponding to the plan score; otherwise, continue to step f);

f) Based on the Gaussian data set, use the Gaussian process to calculate a new plan parameter, and calculate the new plan score corresponding to the plan parameter, and add the new plan parameter and the corresponding new plan score to the Gaussian Data collection

g) Perform step f) iteratively until the preset iteration termination condition is met, and calculate the intensity adjustment optimization result under the plan parameters corresponding to the highest plan score in the Gaussian data set.
The method according to claim 1, wherein the step a) specifically comprises:

Establish a deep learning convolutional neural network model;

Extract the computerized tomography and delineation data and dose data of the preset disease type in the preset case database. The computerized tomography and delineation data include the skin and key organs in the computerized tomography image of the preset disease type Delineation data;

Using the delineation data of the organ as the input of the model, and using the dose data as the output of the model, and training the model to obtain a trained dose prediction model;

Aiming at the computerized tomography and delineation data of the case to be treated, the trained dose prediction model is used to calculate the predicted dose of the case to be treated.
The method according to claim 1, wherein the step b) specifically comprises:

Based on a predetermined scoring rule, calculate the dose volume histogram index score of each organ at risk of the case;

Sum the indicator scores to obtain the total score as the best plan score.
The method according to claim 1, wherein the step c ) specifically comprises:

Based on the delineation data, extract a histogram of overlapping volumes of the organ at risk of the case;

Calculating the similarity between the overlapping volume histogram of the organ at risk and the overlapping volume histogram in the historical data in the prior database;

A predetermined number of optimized parameters with the highest similarity to the histogram of overlapping volumes of the organ-at-risk are selected from the historical data in the prior database as the planning parameters of the case, and the optimized parameters include the field angle of radiation therapy And constraints.
The method according to claim 1, wherein the step f) using Gaussian process to calculate new plan parameters specifically comprises:

Calculating the probability density function of the Gaussian distribution of the plan score under any plan parameter in the Gaussian data set;

Based on the probability density function, the value of the corresponding acquisition function is calculated for multiple discrete parameters in the preset prediction parameter space, and the parameter corresponding to the largest value among the values of the acquisition function is selected as the new planning parameter , The collection function is a preset function.
The method according to claim 1, wherein the preset iterative termination condition for iteratively executing step f) is: if there is a plan score in the Gaussian data set with a score higher than the best plan score or the number of iterations exceeds the preset The iteration is terminated when the number is set.
The method according to claim 1, characterized in that, after the step g), it further comprises: outputting the obtained intensity modulation optimization result.
The method according to claim 2, wherein the extracting the electronic computed tomography and delineation data and dose data of the preset disease type in the preset case database comprises:

Select 256*256 sampling points for a computerized tomography image of the preset disease type in the preset case database;

Extracting the computed tomography values on the sampling points to form a computed tomography value matrix;

Extract the delineation data of the key organ on the sampling point as follows: for any organ, if one of the sampling points belongs to the organ, the delineation data value is 1, otherwise the delineation data value is 0;

Extract the delineation data of the skin on the sampling points as follows: if one of the sampling points belongs to the skin, the delineation data value is 1, otherwise the delineation data value is 0, thus forming a matrix of skin delineation, and then Multiply the matrix of the skin delineation by the numerical value of the corresponding position of the electronic computed tomography value matrix, as the delineation data of the skin;

The dose at the sampling point is extracted to form a dose matrix to obtain dose data.
A plan realization device based on predicted dose guidance and Gaussian process optimization, used for the optimization design of intensity-modulated radiation therapy plans, characterized in that the device includes:

The predicted dose calculation module is used to calculate and obtain the predicted dose of the case based on the electronic computed tomography and delineation data of the case to be treated by using a trained dose prediction model;

The best plan score calculation module is used to calculate the predicted dose plan score as the best plan score according to a predetermined scoring rule;

A plan parameter determination module, configured to determine multiple sets of plan parameters of the case based on the historical data in the associated prior database according to the organ anatomy information of the case;

The Gaussian data set forming module is used to calculate the plan scores corresponding to the multiple groups of plan parameters, and the Gaussian data set is formed by the plan parameters and the corresponding plan scores;

The Gaussian data set iterative update module is used to calculate a new plan parameter based on the Gaussian data set using the Gaussian process, and calculate the new plan score corresponding to the plan parameter, and combine the new plan parameter and the corresponding new plan parameter The planned score is added to the Gaussian data set;

The best plan result output module is used to select the plan with the highest score in the Gaussian data set when there is a plan with a score higher than the best plan score in the Gaussian data set, or when the number of iterations reaches a preset condition, and calculate the plan The optimization result of intensity adjustment under the plan parameters corresponding to the plan score.
The device according to claim 9, wherein the predicted dose calculation module is specifically configured to:

Establish a deep learning convolutional neural network model;

Extract the computerized tomography and delineation data and dose data of the preset disease type in the preset case database. The computerized tomography and delineation data include the skin and key organs in the computerized tomography image of the preset disease type Delineation data;

Using the delineation data of the organ as the input of the model, and using the dose data as the output of the model, and training the model to obtain a trained dose prediction model;

Aiming at the computerized tomography and delineation data of the case to be treated, the trained dose prediction model is used to calculate the predicted dose of the case to be treated.