CN112349383A

CN112349383A - Radiotherapy dose calculation and verification method based on machine learning and Monte Carlo algorithm

Info

Publication number: CN112349383A
Application number: CN202011310708.9A
Authority: CN
Inventors: 林小惟
Original assignee: Individual
Current assignee: Individual
Priority date: 2020-11-20
Filing date: 2020-11-20
Publication date: 2021-02-09

Abstract

The invention belongs to the technical field of radiotherapy radiation dose calculation, and relates to a radiotherapy dose verification method based on GPU parallel Monte Carlo dose calculation and machine learning. The invention comprises the following steps: (1) data input: inputting organs and organs at risk of a target area of a patient to draw a CT image, a DVH image, material density, material information, source parameters, geometric information of a die body and the like; (2) particle input simulation: using a CUDA framework of NVIDIA company, using a GPU of a display card to perform parallel calculation, and using a transport principle of Monte Carlo particles to transport the particles to obtain the simulated distribution of the dose; (3) outputting a simulation result obtained by the Monte Carlo of the parallel GPU in the step (2); (4) establishing a machine learning model by using the parameters, and verifying the output result of the step (3); the invention has the following beneficial effects: the Monte Carlo calculation speed is greatly improved through the parallel GPU hardware, the Monte Carlo calculation result is verified through machine learning, the problems that the time spent on dose verification work of patient treatment by existing radiotherapy is long and the cost of manpower and material resources is high are solved, the dose verification efficiency can be improved, and the verification cost is reduced.

Description

Radiotherapy dose calculation and verification method based on machine learning and Monte Carlo algorithm

Technical Field

The invention belongs to the technical field of radiation dose calculation, and relates to a machine learning and Monte Carlo dose calculation and verification method.

Background

The more intensive radiation therapy techniques have been in clinical use in recent years. In order to ensure the curative effect of the patient and reduce the tumor recurrence and the damage of normal organs, the dose calculation and verification work needs to be carried out on the treatment plan of the patient. Dose calculation is one of the core contents of a radiotherapy treatment planning system, provides irradiation dose data in an area of interest quickly and accurately, is vital to formulation of a radiotherapy plan, and is a main bottleneck for formulation of the radiotherapy plan by reducing dose calculation time on the premise of ensuring dose calculation precision.

There are two methods for increasing the dose calculation speed, one is to adopt different dose calculation methods, and the other is to use hardware with stronger calculation capability. At present, dose calculation algorithms researched and used in radiotherapy are basically divided into three types, namely a Monte Carlo algorithm, a differential convolution integral algorithm and a pencil beam algorithm from high to low in calculation time according to calculation accuracy. The Monte Carlo algorithm is generally used as a standard of dose calculation, simulates the whole process of interaction between particles and substances, and can calculate dose distribution under various complex conditions, so the Monte Carlo algorithm is called as the gold standard of the industry and is the method with the highest precision of all dose calculation. Because the accuracy of the monte carlo simulation is proportional to the square of the number of instances of the simulated particles, when a certain accuracy is required, the monte carlo program needs to simulate a large number of particles, which consumes a lot of time.

Traditional dose is verified by the physics teacher artifical case by case before patient's treatment, uses even die body and all kinds of detectors to carry out dose measurement mostly, needs the physics teacher to put the die body, then verifies case by case, and it is as good as once to verify the time spent and patient's normal treatment, needs longer time. A large number of studies show that the planar dose verification based on the uniform phantom cannot detect clinically significant implementation errors, and simultaneously, the undifferentiated dose verification of each treatment plan consumes a large amount of manpower and material resources. Therefore, how to improve the clinical relevance of patient dose verification and pre-screen the plan with larger implementation error so as to carry out dose verification in a targeted manner is a problem to be solved in the clinic.

Disclosure of Invention

It is an object of the present invention to overcome the above-mentioned deficiencies of the prior art by providing a machine learning and monte carlo dose calculation and verification method.

In order to achieve the purpose, the invention provides the following technical scheme:

a monte carlo dose calculation method, adapted to be executed in a computing device, comprising the steps of:

(1) data input: inputting a CT image of a patient, organ and endangered organ delineation images and DVH images of a target area, material density, material information, source parameters, geometric information of a phantom and the like;

(2) particle input simulation: using a CUDA framework of NVIDIA company, using a GPU of a display card to perform parallel calculation, and using a transport principle of Monte Carlo particles to transport the particles to obtain the simulated distribution of the dose;

(3) and (3) outputting a simulation result obtained by the Monte Carlo based on the parallel GPU in the step (2).

(4) And establishing a machine learning model by using the parameters. Verifying the output result of the step (3)

In the step (1), the source parameters comprise the size and direction of an irradiation field, the energy size and direction of a radioactive source and the type of particles;

in the step (2), the Monte Carlo database comprises a section library; the particles include one or more of photons, protons, heavy ions, or neutrons.

In step (2), the GPU parallel computing method further includes a memory application step, and the like, and specifically includes the steps of:

before the simulation begins. Enough storage space needs to be applied for each arithmetic unit, namely each thread, so as to record data in the transportation process and the final deposition result.

All the original photons participating in the transport are divided into several groups of the same number, as required by the parallel computation. The photons of each group are delivered to an operation matrix formed by one or a plurality of operation units (32 or l6 are common), the dose deposition accumulation in the same group of photon transportation process is added into a shared storage matrix representing the voxel distribution, then the data is returned to the CPU, and the result of each group is added together and output and recorded. Each particle dose deposition process in each group of photons will be handed over to one GPU thread to be responsible, i.e. each thread runs the complete dose deposition process for one photon at a time.

In the simulation process, the differentiation between threads occurs mainly in: uncertainty of the type of physical reaction that occurs during the transport of photons and differences in the transport of photons and electrons. For the latter, because the selected model directly deposits the electron energy as the dose locally, and all the calculation on the thread is the photon transportation process, the thread difference problem is partially solved, and the parallel efficiency is improved.

In the step (4), the parameters include the number of overlapped volume cubic graphs (OVH), step size, simple statistical features, Dose Volume Histogram (DVH) features, and spatial distribution features, three-dimensional dose distribution data and contour line data. Establishing a predictive model includes applying a Support Vector Machine (SVM) to establish a predictive model, and screening the plurality of features includes applying a minimum redundant maximum correlation or a minimum absolute contraction selection to extract a plurality of features from a dose distribution of the region of interest; processing the plurality of features and removing redundancy; establishing a prediction model; and training the prediction model by using the screened data and the patient treatment result data to establish a prediction model between the feature vector and the treatment result.

In the step (4), OVH is used as an input feature of machine learning to represent spatial position information features between the organs at risk and the target area, volume fractions of the organs at risk from the target area of the tumor are described, sample training is carried out to generate an SVM classifier, a classification result is obtained by voting according to the feature parameters to be verified, and verification is carried out on a clinical intensity modulated radiation therapy plan according to the classification result.

The invention also provides a dose verification system based on machine learning, comprising:

the parameter acquisition unit is used for acquiring characteristic parameters of radiotherapy plan such as OVH (over-the-horizon) and dose;

the prediction model establishing unit is used for establishing a machine learning model, performing sample training by taking the characteristic parameters as input samples of the machine learning model to generate an SVM classifier, voting according to the characteristic parameters to be verified to obtain a classification result, and verifying on a clinical intensity modulated radiation therapy plan according to the classification result;

and the training verification unit is used for carrying out sample training to obtain an optimal regression model and verifying the radiotherapy dosage calculated by the Monte Carlo according to the optimal regression model.

The invention has the following beneficial effects:

1. the Monte Carlo dose calculation method based on the GPU parallel calculation can accelerate the calculation speed of the radiotherapy dose, greatly improve the clinical applicability, has high calculation speed and is convenient for evaluating the overall precision;

2. the invention provides a radiotherapy dose verification method based on machine learning. The problems that the time spent on carrying out dose verification work on patient treatment by the existing radiotherapy is long and the cost of manpower and material resources is high are solved, the dose verification efficiency can be improved, and the verification cost is reduced.

Drawings

In order to clearly illustrate the specific embodiments of the present invention, the drawings which are required to be used in the embodiments will be briefly described below.

FIG. 1 is a flowchart of a Monte Carlo computation method based on GPU parallelism according to an embodiment of the present invention.

Fig. 2 is a schematic diagram of a machine learning-based radiotherapy dose verification method provided by the invention.

Fig. 3 is a schematic diagram of a dose verification system based on machine learning according to the present invention.

Detailed Description

In order to make the technical field of the invention better understand the scheme of the embodiment of the invention, the embodiment of the invention is further described in detail with reference to the drawings and the implementation mode.

Example 1

A GPU-based parallel monte carlo dose calculation method, as shown in fig. 1, adapted to be executed in a computing device, comprising the steps of:

Example 2

A parallel computing method based on a GPU comprises the following steps:

(41) before the simulation begins. Enough storage space needs to be applied for each arithmetic unit, namely each thread, so as to record data in the transportation process and the final deposition result. Monte Carlo simulation is random in nature and does not predict which path a given particle will travel, and therefore it is not possible to recombine particles with the same fate into the same warp. When the warpage is divergent, it is divergent

(42) All the original photons participating in the transport are divided into several groups of the same number, as required by the parallel computation. The photons of each group are delivered to an operation matrix formed by one or a plurality of operation units (32 or l6 are common), the dose deposition accumulation in the same group of photon transportation process is added into a shared storage matrix representing the voxel distribution, then the data is returned to the CPU, and the result of each group is added together and output and recorded. Each particle dose deposition process in each group of photons will be handed over to one GPU thread to be responsible, i.e. each thread runs the complete dose deposition process for one photon at a time.

(43) In the simulation process, the differentiation between threads occurs mainly in: uncertainty of the type of physical reaction that occurs during the transport of photons and differences in the transport of photons and electrons. For the latter, because the selected model directly deposits the electron energy as the dose locally, and all the calculation on the thread is the photon transportation process, the thread difference problem is partially solved, and the parallel efficiency is improved.

Finally, GPUMCD can be configured to be a multi-GPU approach with a multi-GPU system that is simple for Monte Carlo simulation, with both GPUs performing the initial particle function and simulating their own set of particles. After the simulation, the two dose arrays produced were summed. The gain in expected linearity performance is a result of copying the input data to two graphics cards instead of one when there are enough particles to model to overcome the overhead.

Example 3

The present invention also provides a computing device comprising:

one or more GPU processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more GPU processors, the one or more programs including instructions for a monte carlo dose calculation method, the method comprising the steps of:

Example 4

As shown in fig. 2, a dose verification method based on machine learning includes:

training and evaluation of ML models since the dose index is scalar, the learning problem is treated as a regression problem. A separate ML model is trained. The prediction model based on machine learning can be built by applying Support Vector Machine (SVM) or logistic regression (logistic regression) and other technologies.

Taking the logistic regression technique as an example, the observation value can be mapped to the value range of [0,1] by using Sigmoid function to realize binary classification, and the parameter of the prediction model is determined by maximum likelihood fitting.

Establishing a predictive model includes applying a Support Vector Machine (SVM) to establish a predictive model, and screening the plurality of features includes applying a minimum redundant maximum correlation or a minimum absolute contraction selection to extract a plurality of features from a dose distribution of the region of interest; processing the plurality of features and removing redundancy; establishing a prediction model; and training the prediction model by using the screened data and the patient treatment result data to establish a prediction model between the feature vector and the treatment result.

The method comprises the steps of using OVH as an input feature of machine learning, representing a spatial position information feature between an organ at risk and a target area, describing a volume fraction of the organ at risk from the target area of a tumor, carrying out sample training to generate an SVM classifier, voting according to feature parameters to be verified to obtain a classification result, verifying on a clinical intensity modulated radiation therapy plan according to the classification result to train a prediction model, and determining parameters of the prediction model through maximum likelihood fitting. In some embodiments, when the minimum absolute value contraction selection factor is used to screen features, then the feature screening and model training steps may be performed simultaneously. For example, when the feature values are screened by using the Lasso regularization factor, the number of variables can be reduced by adding the regularization term in the process of optimizing the fitting parameters.

The accuracy of the predictive model is verified using the test set. In the process of predicting model training, along with the increase of lambda, the deviation of model prediction is gradually reduced, and the acting characteristic value is also gradually reduced, which shows that two tasks of model training and parameter screening are completed simultaneously. Firstly, inputting the characteristics of a test set into a prediction model to obtain prediction probability; then setting a threshold, if the prediction probability is larger than the threshold, judging that complications occur in radiotherapy, otherwise, judging that the test case will not be subjected to a parallel method; the actual disease condition and the predicted result of the test set can be divided into 4 different types, so as to obtain true positive probability (TPR) and false positive probability (FPR), and the calculation formulas are respectively:

TPR＝TP/(TP+FN)

FPR＝FP/(FP+TN)

a series of TPR and FPR is obtained by continuously varying the threshold, with FPR as abscissa and TPR as ordinate, and connecting these points to obtain the ROC curve. The area under the ROC curve (AUC) can quantitatively reflect the prediction capability of the model. The value range of the AUC is [0,1], the larger the value of the AUC is, the stronger the prediction capability of the model is, particularly, if the value of the AUC is 0/1, the prediction is completely wrong/correct, and the value of the AUC is 0.5, the random guess of the model is represented.

To make the results of ML more accurate, we consider using the average dose-volume metric observed in the training data for baseline prediction for each cross-validation.

The method further comprises the following steps:

in order to study the dependency relationship between the number of training sets and the prediction accuracy included in the model training process, an improved model training method is used. During the training of the model, the number of training sets varied from 1 to the entire data set minus the validation set. For each validation set, each combination of training sets generates multiple models to eliminate bias in the contained data set. Predict and average the data for each model before extracting the next data set as valida

In clinical validation, the model trained by the validation process is validated using an independent queue of intensity modulated radiation therapy plans to further validate the reliability and feasibility of the machine learning model as a clinically viable tool to reduce QA workload.

The specific steps of clinical verification are as follows:

1. by using the method, the patient image comprises parameters such as organ delineation of GTV and OAR, the field direction, initial field weight averaging and the like;

2. and establishing a machine learning model, performing regression prediction and classification prediction on the test set, and obtaining a final prediction result.

Further, the performing sample training to obtain an optimal regression model, and predicting the feature parameter to be verified according to the optimal regression model includes:

1. sequentially selecting 1 sample from the samples as a test sample, and taking the rest 100 samples as training samples;

2. training 100 samples by adopting a ten-fold cross-validation method for the model, and searching for the optimal hyper-parameter beta to obtain an optimized model;

3. after obtaining the optimized model, inputting the characteristic parameters of 1 test sample into the model, and predicting which classification the input data belongs to by using the model;

4. repeating the above process N times;

5. regression results were obtained for all samples.

And obtaining the optimal regression model, and carrying out test verification.

Therefore, the radiotherapy dose calculation and verification method based on machine learning and Monte Carlo provided by the invention,

the GPU is used for carrying out calculation by using a parallel Monte Carlo algorithm, the calculation speed is accelerated, and the verification result is realized by training the intensity modulated radiation therapy plan data. The problems that the time spent on carrying out dose verification work on patient treatment by the existing radiotherapy is long and the cost of manpower and material resources is high are solved, the dose verification efficiency can be improved, and the verification cost is reduced.

The above description is only a preferred embodiment of the present application and is not intended to limit the present application, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A radiotherapy dose calculation and verification method based on machine learning and Monte Carlo algorithm is characterized in that: the method comprises the following steps:

(1) data input: inputting organs and organs at risk of a target area of a patient to draw a CT image, a DVH image, material density, material information, source parameters, geometric information of a die body and the like;

(3) outputting a simulation result obtained by the Monte Carlo based on the parallel GPU in the step (2);

(4) and establishing a machine learning model by using the parameters.

2. The Monte Carlo dose calculation method of claim 1, wherein: in the step (1), the source parameters include the size and direction of an irradiation field, the energy size and direction of a radioactive source, and the type of particles.

3. The Monte Carlo dose calculation method of claim 1, wherein: in the step (2), the Monte Carlo database comprises a section library; the particles include one or more of photons, protons, heavy ions, or neutrons.

4. The Monte Carlo dose calculation method of claim 1, wherein: in step (2), the GPU parallel computing method further includes a memory application step, and the like, and specifically includes the steps of:

(41) before the simulated particle transportation starts, each arithmetic unit, namely each thread, needs to apply for enough storage space to record data in the transportation process and a final deposition result;

(42) all the photons participating in transportation are divided into a plurality of groups with the same quantity according to the requirement of parallel computation, the photons in each group are delivered to an operation matrix formed by one or a plurality of operation units, a shared storage matrix representing voxel distribution is added in the dose deposition accumulation in the transportation process of the same group of photons, then the data is returned to a CPU, and the results of each group are added together for output and recording; each particle dose deposition process in each group of photons is handed over to a GPU thread to be responsible, i.e. each thread runs the entire dose deposition process for one photon at a time;

(43) in the simulation process, the differentiation between threads occurs mainly in: uncertainty of reaction types of photons in the transportation process and difference of the photon and electron transportation processes; for the latter, because the selected model directly deposits the electron energy as the dose locally, and all the calculation on the thread is the photon transportation process, the thread difference problem is partially solved, and the parallel efficiency is improved.

5. A dose verification method based on machine learning, characterized by:

in the step (4), the parameters include the number of overlapped volume cubic graphs (OVH), step length, simple statistical characteristics, Dose Volume Histogram (DVH) characteristics, spatial distribution characteristics, three-dimensional dose distribution data and contour line data;

6. The machine-learning based dose verification method of claim 5, wherein:

the method comprises the steps of using OVH as an input feature of machine learning, representing a spatial position information feature between an organ at risk and a target area, describing a volume fraction of the organ at risk from a tumor target area, carrying out sample training to generate an SVM classifier, voting according to feature parameters to be verified to obtain a classification result, and verifying on a clinical intensity modulated radiation therapy plan according to the classification result.

7. A machine learning based dose verification system, comprising:

the data acquisition unit is used for acquiring characteristic parameters of radiotherapy plan such as OVH (over-the-horizon) and dose;