CN115270588A - Fluence spectral distribution calculation modeling and fluence spectral distribution calculation method and device - Google Patents
Fluence spectral distribution calculation modeling and fluence spectral distribution calculation method and device Download PDFInfo
- Publication number
- CN115270588A CN115270588A CN202210928213.5A CN202210928213A CN115270588A CN 115270588 A CN115270588 A CN 115270588A CN 202210928213 A CN202210928213 A CN 202210928213A CN 115270588 A CN115270588 A CN 115270588A
- Authority
- CN
- China
- Prior art keywords
- fluence
- training
- data
- spectral distribution
- neural network
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/25—Design optimisation, verification or simulation using particle-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- 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
-
- 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
- G06T1/00—General purpose image data processing
- G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/08—Probabilistic or stochastic CAD
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a fluence spectral distribution calculation modeling and fluence spectral distribution calculation method and a device, comprising the following steps: collecting image data, and completing Monte Carlo simulation calculation according to the image data to obtain training data, wherein the training data comprises incident fluence spectral distribution and emergent fluence spectral distribution of particles; preprocessing the image data according to the incident fluence spectral distribution to generate training input data; constructing an initial neural network model based on a graphic processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters; and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model. According to the invention, monte Carlo simulation calculation is used as a neural network model training target, so that the precision and efficiency of fluence spectral distribution calculation are ensured.
Description
Technical Field
The invention relates to the field of particle transport calculation, in particular to a fluence spectral distribution calculation modeling and fluence spectral distribution calculation method and device.
Background
The particle fluence is defined as the number of particles passing through a recording plane perpendicularly in unit area, and fluence spectrum distribution refers to dividing a plurality of recording grids for the recording plane and recording the fluence of different energy intervals in each grid. In the radiation therapy field or the radiation protection field, the fluence spectrum distribution emitted by the rays after penetrating through a certain structural medium needs to be calculated so as to facilitate the next calculation. For example, for radiation protection, the radiation after penetrating through a protective wall needs to be evaluated to evaluate the protection effect; for radiation therapy, it is necessary to obtain a fluence spectrum distribution after penetrating a human body for in vivo dose detection and the like.
At present, the mainstream techniques for fluence spectrum calculation are as follows: an algorithm represented by a Monte Carlo (Monte Carlo) algorithm and used for carrying out simulation sampling based on a particle interaction basic principle to obtain emergent fluence distribution; the emission fluence spectral distribution is calculated by a convolution (convolution/super-position) algorithm using a convolution principle, such as pencil beam convolution/super-position (pencil beam convolution) and tube-string convolution (cone beam convolution/super-position).
The existing Monte Carlo (Monte Carlo) algorithm is an algorithm for obtaining the distribution of the emergent fluence by carrying out analog sampling based on the basic principle of particle interaction, although the accuracy is high, the calculation efficiency is low because the particle quantity is large in the dose calculation. In addition, in the existing fluence spectrum distribution calculation method, a dose calculation model is constructed based on a deep learning model, parameters of a medical machine are used as input, the dose calculation model is trained through dose information in water, automatic modeling is achieved, but the number of neurons in dose calculation is huge, and therefore calculation efficiency is not high.
Disclosure of Invention
The invention provides a fluence spectral distribution calculation modeling and fluence spectral distribution calculation method and device, and aims to solve the technical problem that calculation efficiency and calculation accuracy are difficult to guarantee simultaneously in fluence spectral calculation.
In order to solve the above technical problem, in a first aspect, an embodiment of the present invention provides a fluence spectral distribution calculation modeling method, including:
collecting image data, and completing Monte Carlo simulation calculation according to the image data to obtain training data, wherein the training data comprises incident fluence spectral distribution and emergent fluence spectral distribution of particles;
preprocessing the image data according to the incident fluence spectral distribution to generate training input data;
constructing an initial neural network model based on a graphic processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters;
and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model.
The method obtains the incident fluence spectral distribution and the emergent fluence spectral distribution of the particles through Monte Carlo simulation calculation, takes a Monte Carlo simulation calculation result as a target for training a neural network model, optimizes the neural network model, improves the calculation precision of the model, adopts preprocessed image data, constructs the neural network model through a graphic processor to process the data, can process a large number of neuron calculation tasks generated by training or applying the neural network model in the fluence spectral distribution calculation in parallel, and improves the calculation efficiency of the fluence spectral distribution.
Further, the collecting of the image data and the completion of monte carlo simulation calculation according to the image data obtain training data, where the training data includes incident fluence spectral distribution and emergent fluence spectral distribution of particles, and specifically includes:
collecting image data, and acquiring particle information according to the image data, wherein the particle information comprises medium conditions, source conditions and cross-section data of interaction between particles and a medium;
randomly setting a field condition, performing analog sampling on the particles according to the particle information, completing Monte Carlo analog calculation, and obtaining the incident fluence spectral distribution and the emergent fluence spectral distribution of the particles; and taking the incident fluence spectral distribution and the emergent fluence spectral distribution as training data.
According to the invention, the particle information is obtained through the image data, the Monte Carlo simulation calculation is completed according to the particle information and the randomly set field conditions, the incident fluence spectral distribution and the emergent fluence spectral distribution in the Monte Carlo simulation are recorded and are used as the training sample and the training target of the neural network model, the neural network model is continuously optimized, and the precision of the neural network model in calculating the fluence spectral distribution is improved.
Further, the preprocessing the image data according to the incident fluence spectrum distribution to generate training input data specifically includes:
converting the image data into an electron density distribution diagram, and constructing an incident fluence spectral distribution matrix according to incident fluence spectral distribution of particles;
constructing a space physical distance matrix based on the electron density distribution map;
and connecting and integrating the electron density distribution map, the space physical distance matrix and the incident fluence spectrum distribution matrix to obtain training input data.
According to the invention, through preprocessing the image data and integrating the image data, the data with different dimensions are in a similar range, so that the subsequent training process of the neural network model has a higher convergence speed, the model training efficiency is improved, and the efficiency of the model for calculating the fluence spectrum distribution is improved.
Further, the method for constructing the space physical distance matrix based on the electron density distribution map specifically comprises the following steps:
acquiring a physical coordinate corresponding to each grid point of an electron density grid according to the electron density distribution map, calculating the distance between each grid point and a radioactive source and the normalized distance of the radioactive source, and constructing a source distance square inverse ratio factor matrix;
acquiring a physical coordinate corresponding to each grid point of the electron density grid according to the electron density distribution diagram, calculating the vertical distance from each grid point to a central axis where the radioactive source is located, and constructing an off-axis distance matrix;
and linking and combining the source distance square inverse ratio factor matrix and the off-axis distance matrix to obtain a space physical distance matrix.
Further, before the constructing an initial neural network model based on the graphics processor, and repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters, the method further includes:
obtaining an initial training set and an evaluation set according to the training data;
expanding the initial training set, and randomly extracting a plurality of groups of incident fluence spectral distributions and corresponding emergent fluence spectral distributions in the initial training set as an expanded data set; respectively linearly superposing incident fluence spectrum distribution and emergent fluence spectrum distribution in an expanded data set to obtain expanded incident fluence spectrum distribution and corresponding expanded emergent fluence spectrum distribution, and adding the expanded incident fluence spectrum distribution and the expanded emergent fluence spectrum distribution into a target training set to complete primary data expansion;
repeating data expansion for multiple times until all incident fluence spectrum distributions and corresponding emergent fluence spectra in the initial training set are extracted, and the number of samples in the target training set reaches a preset condition, stopping data expansion, and obtaining a target training set; the target training set further comprises an initial training set.
According to the method, the training data are divided to obtain the initial training set and the evaluation set, the initial training set is expanded, overfitting of the model is avoided, and the generalization capability of the model is improved as much as possible. In addition, the samples in the training set are expanded to increase the number of the samples, the diversity of the samples is improved, and the problem that the sample acquisition efficiency is not high due to low Monte Carlo simulation calculation efficiency is solved.
Further, the method includes the steps of constructing an initial neural network model based on the graphics processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters, and specifically includes:
constructing an initial neural network model based on a graphic processor, immediately setting a model weight, and performing training iteration on the initial neural network model by adopting a self-adaptive gradient method;
randomly extracting a plurality of data samples from a training input data set, inputting the data samples into an initial neural network model, and obtaining a first emergent fluence spectral distribution output by the model;
acquiring second emergent fluence spectral distribution corresponding to the incident fluence spectral distribution of the plurality of data samples in a target training set;
calculating a loss value according to the first emergent fluence spectrum distribution and the second emergent fluence spectrum distribution, and updating the model weight based on the loss value;
and storing the model weight until all data samples in the training input data are extracted, and updating the model weight based on the evaluation set.
According to the invention, the neural network model is constructed through the graphic processor to process data, so that a large number of neuron calculation tasks can be processed in parallel, and the fluence spectrum distribution calculation efficiency is improved. And meanwhile, iterative optimization of model parameters is carried out through a self-adaptive gradient algorithm, the loss value of emergent fluence spectral distribution is calculated, the weight of the model is updated, and the precision of the model is improved.
Further, the updating the model weight based on the evaluation set specifically includes:
calculating a loss value of a current training period of the evaluation set data sample;
and comparing the loss value of the current training period with the loss value of the previous period, if the loss value of the current training period is smaller than the loss value of the previous period, updating the stored model weight, otherwise, not updating the model weight.
According to the method, the training data obtained through Monte Carlo simulation calculation is used as an evaluation set, the loss value of the model is evaluated, the weight of the model is optimized and updated, the precision of the model is improved, and the precision of fluence spectral distribution calculation is improved.
In a second aspect, an embodiment of the present invention provides a fluence spectrum distribution calculation method, including:
acquiring a neural network model constructed by applying the fluence spectral distribution calculation modeling method;
acquiring first image data to be detected, wherein the first image data comprises image information or geometric information; preprocessing the first image data to obtain first input data;
inputting the first input data into the neural network model, and calculating an emergent fluence spectral distribution matrix;
and reducing the emergent fluence spectrum distribution matrix into an absolute value to complete emergent fluence spectrum distribution calculation.
According to the invention, the model input data is obtained by preprocessing according to the image data, the fluence spectral distribution is calculated according to the neural network model, the fluence spectral distribution calculation efficiency is improved, and the neural network model takes a Monte Carlo simulation calculation sample as training data, so that the fluence spectral distribution calculation precision of the neural network model is ensured.
In a third aspect, an embodiment of the present invention provides a fluence spectrum distribution calculation modeling apparatus, including: the system comprises a simulation calculation module, a data preprocessing module and an iterative training module;
the simulation calculation module is used for collecting image data, completing Monte Carlo simulation calculation according to the image data and obtaining training data, wherein the training data comprises incident fluence spectral distribution and emergent fluence spectral distribution of particles;
the data preprocessing module is used for preprocessing the image data according to the incident fluence spectral distribution to generate training input data;
the iterative training module is used for constructing an initial neural network model based on the graphic processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters; and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model.
In a fourth aspect, an embodiment of the present invention provides a fluence spectrum distribution calculating device, including: the device comprises a model building module, a data processing module, a data calculating module and a normalization processing module;
the model building module is used for obtaining a neural network model built by applying the fluence spectral distribution calculation modeling device;
the data processing module is used for acquiring image data, and the image data comprises image information or geometric information; preprocessing the image data to obtain model input data;
the data calculation module is used for inputting the model input data into the neural network model and calculating an emergent fluence spectrum distribution matrix;
and the normalization processing module is used for reducing the emergent fluence spectral distribution matrix into an absolute value to complete emergent fluence spectral distribution calculation.
Drawings
FIG. 1 is a schematic flow chart of a method for modeling fluence spectral distribution calculation according to an embodiment of the present invention;
FIG. 2 is a schematic flow chart of step 102 according to an embodiment of the present invention;
FIG. 3 is a schematic flow chart of a method for calculating fluence spectral distribution according to an embodiment of the invention;
FIG. 4 is a schematic structural diagram of a fluence spectral distribution calculation modeling apparatus provided in an embodiment of the present invention;
fig. 5 is a schematic structural diagram of a fluence spectrum distribution calculating device according to an embodiment of the invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example one
Referring to fig. 1, fig. 1 is a schematic flow chart of a fluence spectral distribution calculation modeling method according to an embodiment of the present invention, which mainly includes steps 101 to 103, and specifically includes the following steps:
step 101: collecting image data, and completing Monte Carlo simulation calculation according to the image data to obtain training data, wherein the training data comprises incident fluence spectral distribution and emergent fluence spectral distribution of particles.
In this embodiment, the conditions required for the monte carlo simulation calculation include three parts: one part is the medium conditions (including medium geometry, medium material, density of the medium), a second part is the source conditions (including species of particles: e.g. photons, charged particles, neutrons, particle direction, particle energy), and a third part is cross-sectional data of particle-medium interactions, which can be obtained by consulting published databases, such as the National Institute of Standards and Technology, NIST.
In this embodiment, for the first part, if the collected body parts are directly assigned with the medium density according to the density distribution of the image data, and then assigned with the medium material according to the density range (for example, 0.001 to 0.044 is defined as air, 0.044 to 0.302 is defined as lung tissue, 0.302 to 1.101 is defined as soft tissue, and 1.101 to 2.088 is defined as bone tissue), if the design drawing is used, the medium geometric distribution is restored according to the drawing, and the medium composition and the density are directly assigned according to the material composition specification of the drawing. For the second part, analog sampling is performed according to the source illumination range (field range, intensity distribution), source energy, and species.
In the embodiment, image data is collected, and the image data includes CT/MR image data of various parts of the human body, such as different parts of the human body, such as the head and neck, the chest, the abdomen, the pelvic cavity, and the like, or related design drawings.
As a specific example of the embodiment of the present invention, before completing monte carlo simulation calculation according to image data, the method further includes expanding the image data, and specifically includes:
numbering the collected N total original image data in sequence;
the serial numbers arranged in sequence are disordered;
b numbers are sequentially extracted from the disordered numbers, and corresponding original image data are obtained according to the extracted numbers;
acquiring new image data by linearly superposing the extracted B image data;
when all the original image data are traversed, completing one-time image data expansion;
and repeating the image data expansion for multiple times to generate new image data until the number of new image data meeting the requirement is generated.
In this embodiment, the linear superposition is performed on the original image data to obtain new image data, specifically:
wherein, I k For the extracted original sample image, x k The superposition weight of (a) is a random floating point number of 0 to 1, and x is the same k The following conditions are satisfied:
in this embodiment, the training data is calculated by simulation using a monte carlo algorithm based on the extended image data. Particle information is obtained from the image data, the particle information including media conditions, source conditions, and cross-sectional data of particle-media interactions.
In this embodiment, during the monte carlo simulation calculation, the field conditions are randomly set, and the field conditions include field shape, intensity distribution, incident angle and incident energy; the energy and direction of the simulated particles are sampled according to the energy spectrum and the irradiation range of the source, the process of the simulated particles penetrating the medium is simulated, and after the simulation of a sufficient number of particles (10 is adopted in the embodiment) 9 One sample is completed by particle simulation), and the incident beam before entering the medium is recorded respectivelyThe fluence spectral distribution and the exit fluence spectral distribution leaving the medium are used as training data.
In the embodiment, the particle information is obtained through the image data, the Monte Carlo simulation calculation is completed according to the particle information and the randomly set field conditions, the incident fluence spectral distribution and the emergent fluence spectral distribution in the Monte Carlo simulation are recorded and are used as the training sample and the training target of the neural network model, the neural network model is continuously optimized, and the precision of the neural network model in calculating the fluence spectral distribution is improved.
Step 102: and preprocessing the image data according to the incident fluence spectral distribution to generate training input data.
Referring to fig. 2, fig. 2 is a schematic flow chart of step 102 according to an embodiment of the present invention, which mainly includes steps 201 to 203, specifically as follows:
step 201: and converting the image data into an electron density distribution diagram, and constructing an incident fluence spectral distribution matrix according to the incident fluence spectral distribution of the particles.
In the present embodiment, the image is converted into an electron density distribution map based on image or geometric information, the image is a CT (Computed Tomography) image, and the HU value of the CT image is converted into an electron density distribution based on the HU-electron density conversion curve of the machine that acquired the CT image. (or according to MR (Magnetic Resonance) image information, electron density is given according to different media, such as air: 0.001, lung tissue: 0.4, fat tissue: 0.9, water: 1.0, muscle tissue: 1.05, bone tissue: 1.69.) or the electron density distribution can be interpolated to 100 x 100 (length x width x height) according to the known structure geometry, such as the thickness and the corresponding cement wall electron density according to the drawing of a machine room protection wall, or the structure of a medical linear accelerator head, such as the structure position of a multi-page grating, etc.), and the physical size of each grid is 0.5cm x 0.5cm.
In this embodiment, the step of constructing the incident fluence spectral distribution matrix F before the particles enter the phantom specifically includes: the particle fluence is defined as the number of particles per unit area passing perpendicularly through the recording plane, the fluenceThe spectral distribution means that a plurality of recording grids are divided for a recording plane, and the fluence of different energy intervals in each grid is recorded. A matrix with the grid number of 100 multiplied by 100 (length multiplied by width multiplied by energy) is constructed, wherein the size of a space physical dimension is 50cm multiplied by 50cm, the grid number is 100 multiplied by 100, the size of each grid in the space dimension is 0.5cm multiplied by 0.5cm, the range of the energy dimension is 0-20 MeV, 100 intervals are divided, and the interval of each energy interval is 0.2MeV. Respectively recording the fluence of particles in each energy interval of each recording grid to obtain an incident fluence spectrum distribution matrix F in . The incident fluence spectrum can be obtained by first calculating the planar incident fluence spectral distribution before the particles emitted by the source enter the phantom from the fluence definition when the source type (e.g., point source or parallel beam source), the particle type (photon or charged particle) and the energy spectrum of the source are known, directly from the determined geometric relationships. The second method is to count the fluence spectrum distribution of the recording plane according to a phase space file storing the recording plane (the phase space file stores the states of all particles in the recording plane, including the types, the moving directions, the energies and the positions of the particles).
Step 202: and constructing a space physical distance matrix based on the electron density distribution diagram.
In this embodiment, in order to describe the spatial position corresponding to the electron density distribution, a spatial physical distance matrix is constructed according to the physical coordinates corresponding to each grid point of the electron density grid, and the matrix size is 100 × 100 × 100 (length × width × height), specifically as follows:
the space physical matrix comprises two parts, one part is a source distance square inverse ratio factor, and the calculation formula is as follows:
where rijk is the (i, j, k) grid point to source distance, and r0 is the normalized distance of the source (typically 100 cm).
The other part is an off-axis distance matrix, and the calculation formula is as follows:
where d is the vertical distance from the (i, j, k) grid to the central axis of the source, d0 is half of the total length of the matrix, which is 25cm in this embodiment.
Finally, the two part matrixes are linked and combined into a space physical distance matrix with the size of 100 multiplied by 2.
Step 203: and connecting and integrating the electron density distribution map, the space physical distance matrix and the incident fluence spectrum distribution matrix to obtain training input data.
In this embodiment, after obtaining the electron density distribution map, the spatial physical distance matrix, and the incident fluence spectral distribution, the electron density distribution map, the spatial physical distance matrix, and the incident fluence spectral distribution are respectively integrated as data channels, and the final input matrix is 100 × 100 × 100 × 4, so as to achieve faster convergence speed in the training process, the data of different dimensions are normalized according to their respective corresponding maximum values, for example, between 0 and 1, the maximum value of the electron density distribution is 2, and the incident fluence spectral distribution is integrated according to the maximum value F max,in And (5) carrying out normalization.
Step 103: constructing an initial neural network model based on a graphic processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters; and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model.
In this embodiment, before the constructing an initial neural network model based on a graphics processor, performing repeated training and evaluation on the initial neural network model according to training input data and training data, and updating model parameters, the method further includes:
the sample size in the training data is as follows: the scale of 1 is divided into an initial training set and an evaluation set.
Expanding the initial training set, and randomly extracting a plurality of groups of incident fluence spectral distributions and corresponding emergent fluence spectral distributions in the initial training set as an expanded data set; respectively linearly superposing incident fluence spectrum distribution and emergent fluence spectrum distribution in an expanded data set to obtain expanded incident fluence spectrum distribution and corresponding expanded emergent fluence spectrum distribution, and adding the expanded incident fluence spectrum distribution and the expanded emergent fluence spectrum distribution into a target training set to complete primary data expansion;
repeating data expansion for multiple times until all incident fluence spectrum distributions and corresponding emergent fluence spectra in the initial training set are extracted, and the number of samples in the target training set reaches a preset condition, stopping data expansion, and obtaining a target training set; the target training set further comprises an initial training set.
As a specific example of this embodiment, the data expansion specifically includes:
numbering incident fluence spectral distributions with the total number of M in the initial training set and corresponding emergent fluence spectral distributions in sequence;
the serial numbers are disordered, B serial numbers are sequentially extracted from the disordered serial numbers, and corresponding original samples are obtained according to the serial numbers, wherein the original samples are incident fluence spectral distribution and corresponding emergent fluence spectral distribution;
obtaining a new input sample S by continuously linearly superposing the extracted B samples in And outputting the sample S out ；
And when all the original samples are traversed, completing one-time data expansion.
In this embodiment, data expansion is performed a plurality of times, and new sample generation is performed until the number of new fluence samples satisfying the requirement is generated.
As a specific example of this embodiment, the original samples are linearly superimposed, which is specifically as follows:
wherein F in,k For the extracted spectral distribution of the incident fluence of the original sample, F out,k For the extracted original sample exit fluence spectral distribution, y k The superposition weight of (a) is a random floating point number of 0 to 1, and y is the same k The following conditions are satisfied:
in this embodiment, the establishing an initial neural network model based on the graphics processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters specifically includes:
constructing an initial neural network model based on a graphic processor, immediately setting a model weight, and performing training iteration on the initial neural network model by adopting a self-adaptive gradient method;
randomly extracting a plurality of data samples from a training input data set, inputting the data samples into an initial neural network model, and obtaining a first emergent fluence spectral distribution output by the model;
acquiring second emergent fluence spectral distribution corresponding to the incident fluence spectral distribution of the plurality of data samples in a target training set;
calculating a loss value according to the first emergent fluence spectrum distribution and the second emergent fluence spectrum distribution, and updating the model weight based on the loss value;
storing the model weight until all data samples in the training input data are extracted;
calculating a loss value of a current training period of the evaluation set data sample;
and comparing the loss value of the current training period with the loss value of the previous period, if the loss value of the current training period is smaller than the loss value of the previous period, updating the stored model weight, otherwise, not updating the model weight.
As a specific example of this embodiment, the neural network is constructed by using an encoder-decoder structure, in which an encoder gradually reduces data dimensions, identifies image features, and a decoder gradually restores details and data dimensions of an object, and performs prediction pixel by pixel. Considering that dose calculation is a three-dimensional calculation process, a three-dimensional Vision Transformer module is adopted as an encoder. The encoder depth is 4 layers, except that the input to the first layer encoder is the model input, each layer encoder is the output of the previous layer encoder, the output is 50 x 32 via the first layer encoder, via the second layer encoder, the output is 25 × 25 × 25 × 64, the output is 12 × 12 × 12 × 128 after passing through the third layer encoder, the output is 6 × 6 × 6 × 256 after passing through the fourth layer encoder, and three-dimensional trilinear upsampling is adopted to combine residual convolution as a decoder. The encoder-decoder in the same layer are connected by using a characteristic channel layer, so that the number of input channels of the decoder is 2 times that of the decoder. The fourth layer decoder input is 6 × 6 × 6 × 512, the output is 12 × 12 × 12 × 128, the third layer decoder input is 12 × 12 × 12 × 256, the output is 25 × 25 × 25 × 64, the second layer decoder input is 25 × 25 × 128, the output is 50 × 50 × 50 × 32, the first layer decoder input is 50 × 50 × 50 × 64, and the output is 100 × 100 × 100 × 16. The last output layer reduces the feature layer to 1 by convolution layer with convolution kernel size of 1, and in order to ensure the final output range >0, the final output layer adopts relu activation function.
In this embodiment, no matter the training and application of the neural network model, each neuron (a model minimum calculation unit, generally, a specific form is y = wx + b, where y is a neuron output, x is a neuron input, and w and b are fitting parameters) needs to be calculated, and the calculation is characterized in that a single neuron is simple to calculate but has a huge number, and for the neural network constructed by the present invention, specific parameters are: 30698775. the Graphics Processing Unit (GPU) has the greatest advantage of having a large number of processing units, and can effectively perform parallel computation to complete the computation tasks of multiple neurons without increasing the processing speed of a single processing unit. Thereby improving overall computational efficiency.
In this embodiment, the model training process adopts an adaptive gradient method to perform model parameter iterative optimization, and the initial learning rate is set to 5e-4. And (3) calculating the MSE loss value of the model output result and the training sample by adopting Mean Squared Error (MSE) as a loss function. In the training process, 2 samples are randomly extracted from a training set at each iteration and input into the model (which can be increased according to the performance of a computer) to carry out derivation on MSE, and gradient direction updating model parameters are calculated.
In this embodiment, the specific calculated loss values are specifically as follows:
wherein, y i For true values, i.e. training samples, y i ' is the model output value.
In this embodiment, after traversing all training set samples once, fixing the model parameters, calculating the MSE loss value of the evaluation set samples, and storing the model parameters, at this time, completing the first period of model training. And continuing a new round of expansion on the training set again according to the steps, and entering a second training period. Starting from the second period, only when the MSE loss value of the evaluation set is smaller than that of the last period, the saved model parameters are updated to ensure that the saved model has optimal performance.
In this embodiment, training is stopped when the training period reaches the upper limit (default is set to 500) or the evaluation set loss value does not decrease for 10 consecutive periods.
Referring to fig. 3, fig. 3 is a schematic flow chart of a fluence spectrum distribution calculation method according to an embodiment of the present invention, which mainly includes steps 301 to 304:
step 301: acquiring a neural network model constructed by applying the fluence spectral distribution calculation modeling method provided by the embodiment of the invention;
step 302: acquiring first image data to be detected, wherein the first image data comprises image information or geometric information; preprocessing the first image data to obtain first input data;
step 303: inputting the first input data into the neural network model, and calculating an emergent fluence spectral distribution matrix;
step 304: and reducing the emergent fluence spectrum distribution matrix into an absolute value to complete emergent fluence spectrum distribution calculation.
In this embodiment, the output data of the neural network model is a matrix of the emitted fluence spectrum distribution of the rays leaving the phantom after passing through the phantom described in the electron density distribution, and a matrix with the number of grids of 100 × 100 × 100 (length × width × energy) is output, where the size of the spatial physical dimension is 50cm × 50cm, the number of the grids is 100 × 100, the size of each grid in the spatial dimension is 0.5cm × 0.5cm, the range of the energy dimension is 0 to 20MeV, 100 intervals are divided, and the interval of each energy interval is 0.2MeV. Respectively recording the fluence of the particles in each energy interval in each recording grid to obtain the emergent fluence spectral distribution matrix F out . For training data, in order to keep the numerical association between the incident fluence spectral distribution in the input data and the output emergent fluence spectral distribution, according to the normalization value F of the incident fluence spectral distribution corresponding to the generated emergent fluence spectral distribution max,in As denominator, a normalization calculation is performed.
In this embodiment, the output is normalized output through model calculation as the exit fluence spectral distribution matrix F out ' is restored to an absolute value from the maximum value of the inputted incident fluence spectral distribution.
As a specific example of this embodiment, the specific reduction step of the outgoing fluence spectrum distribution matrix is:
F out ＝F out ’*F max,in (9)
wherein, F in Is the incident fluence spectral distribution of the input, F max,in Is the maximum value of the input incident fluence spectral distribution, F out ' is the normalized emergent fluence spectral distribution output by model calculation.
In this embodiment, the model input data is obtained by preprocessing the image data, and the fluence spectral distribution is calculated according to the neural network model, so that the efficiency of fluence spectral distribution calculation is improved, and the neural network model takes a sample of monte carlo simulation calculation as training data, so that the precision of fluence spectral distribution calculation by the neural network model is ensured.
Referring to fig. 4, fig. 4 is a schematic structural diagram of a fluence spectrum distribution calculation modeling apparatus according to an embodiment of the present invention, which mainly includes a simulation calculation module 401, a data preprocessing module 402, and an iterative training module 403;
the simulation calculation module 401 is configured to collect image data, complete monte carlo simulation calculation according to the image data, and obtain training data, where the training data includes incident fluence spectral distribution and emergent fluence spectral distribution of particles;
the data preprocessing module 402 is configured to preprocess the image data according to the incident fluence spectral distribution to generate training input data;
the iterative training module 403 is configured to construct an initial neural network model based on a graphics processor, perform repeated training and evaluation on the initial neural network model according to training input data and training data, and update model parameters; and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model.
Referring to fig. 5, fig. 5 is a schematic structural diagram of a fluence spectrum distribution calculating device according to an embodiment of the present invention, which mainly includes a model building module 501, a data processing module 502, a data calculating module 503, and a normalization processing module 504;
the model building module 501 is configured to obtain a neural network model built by applying the fluence spectral distribution calculation modeling apparatus;
the data processing module 502 is configured to obtain image data, where the image data includes image information or geometric information; preprocessing the image data to obtain model input data;
the data calculation module 503 is configured to input the model input data into the neural network model, and calculate an emergent fluence spectral distribution matrix;
the normalization processing module 504 is configured to reduce the outgoing fluence spectrum distribution matrix to an absolute value, thereby completing outgoing fluence spectrum distribution calculation.
In this embodiment, the model input data is obtained by preprocessing the image data, and the fluence spectral distribution is calculated according to the neural network model, so that the efficiency of fluence spectral distribution calculation is improved, and the neural network model takes a sample of monte carlo simulation calculation as training data, so that the precision of fluence spectral distribution calculation by the neural network model is ensured.
The above-mentioned embodiments are provided to further explain the objects, technical solutions and advantages of the present invention in detail, and it should be understood that the above-mentioned embodiments are only examples of the present invention and are not intended to limit the scope of the present invention. It should be understood that any modifications, equivalents, improvements and the like, which come within the spirit and principle of the invention, may occur to those skilled in the art and are intended to be included within the scope of the invention.
Claims (10)
1. A fluence spectral distribution computational modeling method is characterized by comprising the following steps:
collecting image data, and completing Monte Carlo simulation calculation according to the image data to obtain training data, wherein the training data comprises incident fluence spectral distribution and emergent fluence spectral distribution of particles;
preprocessing the image data according to the incident fluence spectral distribution to generate training input data;
constructing an initial neural network model based on a graphic processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters;
and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model.
2. The method according to claim 1, wherein the collecting image data and the performing monte carlo simulation calculation according to the image data to obtain training data, the training data includes an incident fluence spectral distribution and an emergent fluence spectral distribution of particles, and specifically includes:
collecting image data, and acquiring particle information according to the image data, wherein the particle information comprises medium conditions, source conditions and cross section data of interaction between particles and a medium;
randomly setting a field condition, performing analog sampling on the particles according to the particle information, and completing Monte Carlo analog calculation to obtain incident fluence spectral distribution and emergent fluence spectral distribution of the particles; and taking the incident fluence spectral distribution and the emergent fluence spectral distribution as training data.
3. The method according to claim 2, wherein the preprocessing is performed on the image data according to the incident fluence spectral distribution to generate training input data, specifically:
converting the image data into an electron density distribution map, and constructing an incident fluence spectral distribution matrix according to incident fluence spectral distribution of particles;
constructing a space physical distance matrix based on the electron density distribution map;
and connecting and integrating the electron density distribution map, the space physical distance matrix and the incident fluence spectrum distribution matrix to obtain training input data.
4. The fluence spectrum distribution computational modeling method of claim 3, wherein the spatial physical distance matrix is constructed based on the electron density distribution map, in particular:
acquiring a physical coordinate corresponding to each grid point of an electron density grid according to the electron density distribution map, calculating the distance between each grid point and a radioactive source and the normalized distance of the radioactive source, and constructing a source distance square inverse ratio factor matrix;
acquiring a physical coordinate corresponding to each grid point of the electron density grid according to the electron density distribution diagram, calculating the vertical distance from each grid point to a central axis where the radioactive source is located, and constructing an off-axis distance matrix;
and linking and combining the source distance square inverse ratio factor matrix and the off-axis distance matrix to obtain a space physical distance matrix.
5. The method of fluence spectrum distribution computational modeling as defined in claim 1, further comprising, before the graph-processor-based construction of an initial neural network model, and repeated training and evaluation of the initial neural network model based on training input data and training data, and updating of model parameters:
obtaining an initial training set and an evaluation set according to the training data;
expanding the initial training set, and randomly extracting a plurality of groups of incident fluence spectral distributions and corresponding emergent fluence spectral distributions in the initial training set as an expanded data set; respectively linearly superposing incident fluence spectrum distribution and emergent fluence spectrum distribution in an expanded data set to obtain expanded incident fluence spectrum distribution and corresponding expanded emergent fluence spectrum distribution, and adding the expanded incident fluence spectrum distribution and the expanded emergent fluence spectrum distribution into a target training set to complete primary data expansion;
repeating data expansion for multiple times until all incident fluence spectrum distributions and corresponding emergent fluence spectrums in the initial training set are extracted, and the number of samples in the target training set reaches a preset condition, stopping data expansion, and obtaining a target training set; the target training set further comprises an initial training set.
6. The fluence spectrum distribution computational modeling method of claim 5, wherein the graph-based processor constructs an initial neural network model, and repeatedly trains and evaluates the initial neural network model according to training input data and training data, and updates model parameters, specifically:
constructing an initial neural network model based on a graphic processor, immediately setting a model weight, and performing training iteration on the initial neural network model by adopting a self-adaptive gradient method;
randomly extracting a plurality of data samples from a training input data set, inputting the data samples into an initial neural network model, and obtaining a first emergent fluence spectral distribution output by the model;
acquiring second emergent fluence spectral distribution corresponding to the incident fluence spectral distribution of the plurality of data samples in a target training set;
calculating a loss value according to the first emergent fluence spectral distribution and the second emergent fluence spectral distribution, and updating the model weight based on the loss value;
and storing the model weight until all data samples in the training input data are extracted, and updating the model weight based on the evaluation set.
7. The fluence spectrum distribution computational modeling method of claim 6 wherein the updating of the model weights based on the evaluation set specifically comprises:
calculating a loss value of a current training period of the evaluation set data sample;
and comparing the loss value of the current training period with the loss value of the previous period, if the loss value of the current training period is smaller than the loss value of the previous period, updating the stored model weight, otherwise, not updating the model weight.
8. A fluence spectrum distribution calculation method, characterized by comprising:
obtaining a neural network model constructed by applying the fluence spectral distribution computational modeling method as defined in any one of claims 1 to 7;
acquiring first image data to be detected, wherein the first image data comprises image information or geometric information; preprocessing the first image data to obtain first input data;
inputting the first input data into the neural network model, and calculating an emergent fluence spectral distribution matrix;
and reducing the emergent fluence spectrum distribution matrix into an absolute value to complete emergent fluence spectrum distribution calculation.
9. A fluence spectral distribution computational modeling apparatus, comprising: the device comprises a simulation calculation module, a data preprocessing module and an iterative training module;
the simulation calculation module is used for collecting image data, completing Monte Carlo simulation calculation according to the image data and obtaining training data, wherein the training data comprises incident fluence spectral distribution and emergent fluence spectral distribution of particles;
the data preprocessing module is used for preprocessing the image data according to the incident fluence spectral distribution to generate training input data;
the iterative training module is used for constructing an initial neural network model based on the graphic processor, repeatedly training and evaluating the initial neural network model according to training input data and training data, and updating model parameters; and stopping training iteration until a preset condition is reached, and storing model parameters to complete the construction of the neural network model.
10. A fluence spectrum distribution calculation apparatus, characterized by comprising: the device comprises a model construction module, a data processing module, a data calculation module and a normalization processing module;
the model construction module for obtaining a neural network model constructed by applying the fluence spectrum distribution calculation modeling apparatus as defined in claim 9;
the data processing module is used for acquiring image data, and the image data comprises image information or geometric information; preprocessing the image data to obtain model input data;
the data calculation module is used for inputting the model input data into the neural network model and calculating an emergent fluence spectral distribution matrix;
and the normalization processing module is used for reducing the emergent fluence spectral distribution matrix into an absolute value to complete emergent fluence spectral distribution calculation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210928213.5A CN115270588B (en) | 2022-08-03 | 2022-08-03 | Method and device for calculating and modeling fluence spectrum distribution and calculating fluence spectrum distribution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210928213.5A CN115270588B (en) | 2022-08-03 | 2022-08-03 | Method and device for calculating and modeling fluence spectrum distribution and calculating fluence spectrum distribution |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115270588A true CN115270588A (en) | 2022-11-01 |
| CN115270588B CN115270588B (en) | 2023-06-23 |
Family
ID=83748574
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210928213.5A Active CN115270588B (en) | 2022-08-03 | 2022-08-03 | Method and device for calculating and modeling fluence spectrum distribution and calculating fluence spectrum distribution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115270588B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110133703A (en) * | 2019-05-15 | 2019-08-16 | 中国辐射防护研究院 | A kind of gamma-rays Dosimetry based on BP neural network and Monte carlo algorithm |
| CN112086172A (en) * | 2020-09-14 | 2020-12-15 | 广州瑞多思医疗科技有限公司 | Three-dimensional dose calculation method, computer equipment and readable medium |
| CN112086173A (en) * | 2020-09-14 | 2020-12-15 | 广州瑞多思医疗科技有限公司 | Three-dimensional dose calculation method and device, computer equipment and readable medium |
| US20220088410A1 (en) * | 2020-09-21 | 2022-03-24 | Elekta, Inc. | Machine learning optimization of fluence maps for radiotherapy treatment |
| CN114298177A (en) * | 2021-12-16 | 2022-04-08 | 广州瑞多思医疗科技有限公司 | Expansion enhancement method and system suitable for deep learning training data and readable storage medium |
-
2022
- 2022-08-03 CN CN202210928213.5A patent/CN115270588B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110133703A (en) * | 2019-05-15 | 2019-08-16 | 中国辐射防护研究院 | A kind of gamma-rays Dosimetry based on BP neural network and Monte carlo algorithm |
| CN112086172A (en) * | 2020-09-14 | 2020-12-15 | 广州瑞多思医疗科技有限公司 | Three-dimensional dose calculation method, computer equipment and readable medium |
| CN112086173A (en) * | 2020-09-14 | 2020-12-15 | 广州瑞多思医疗科技有限公司 | Three-dimensional dose calculation method and device, computer equipment and readable medium |
| US20220088410A1 (en) * | 2020-09-21 | 2022-03-24 | Elekta, Inc. | Machine learning optimization of fluence maps for radiotherapy treatment |
| CN114298177A (en) * | 2021-12-16 | 2022-04-08 | 广州瑞多思医疗科技有限公司 | Expansion enhancement method and system suitable for deep learning training data and readable storage medium |
Non-Patent Citations (2)
| Title |
|---|
| 陈朝斌等: "蒙特卡罗方法在放疗计划中的应用", 《核技术》 * |
| 陈朝斌等: "蒙特卡罗方法在放疗计划中的应用", 《核技术》, no. 01, 10 January 2006 (2006-01-10) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115270588B (en) | 2023-06-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109272443B (en) | PET and CT image registration method based on full convolution neural network | |
| Jia et al. | GPU-based fast low-dose cone beam CT reconstruction via total variation | |
| Mao et al. | RapidBrachyDL: rapid radiation dose calculations in brachytherapy via deep learning | |
| CN109843377A (en) | System and method for predicting the learning model of the radiotherapeutic treatment plan of radiation therapy dose distribution | |
| US20060259282A1 (en) | Deterministic computation of radiation transport for radiotherapy dose calculations and scatter correction for image reconstruction | |
| CN110363797B (en) | PET and CT image registration method based on excessive deformation inhibition | |
| TWI646945B (en) | Deconstruction method of tissue element mass ratio based on medical image and establishment method of geometric model | |
| Peng et al. | MCDNet–a denoising convolutional neural network to accelerate Monte Carlo radiation transport simulations: A proof of principle with patient dose from x-ray CT imaging | |
| CN111565797A (en) | Electronic modeling of radiation treatment planning or administration | |
| CN102096939B (en) | Medical mass data-oriented multi-resolution volume rendering method | |
| CN109859817A (en) | Oesophagus radiotherapy treatment planning risk organ dose learns appraisal procedure, assessment system | |
| CN113178242B (en) | Automatic plan optimization system based on coupled generation countermeasure network | |
| Yousefi et al. | Novel Method for More Efficient Optimizing the Knowledge-Based Planning: Specific Voxels of each Structure Influenced by Dominant Beamlets (SVSIDB) | |
| Vazquez et al. | A deep learning-based approach for statistical robustness evaluation in proton therapy treatment planning: a feasibility study | |
| Denisova et al. | Development of anthropomorphic mathematical phantoms for simulations of clinical cases in diagnostic nuclear medicine | |
| van Dijk et al. | A novel multichannel deep learning model for fast denoising of Monte Carlo dose calculations: preclinical applications | |
| CN111494815B (en) | Three-dimensional dose calculation method, device and medium based on mixed variable-scale model | |
| CN115270588B (en) | Method and device for calculating and modeling fluence spectrum distribution and calculating fluence spectrum distribution | |
| CN112086173A (en) | Three-dimensional dose calculation method and device, computer equipment and readable medium | |
| Fu et al. | Fast Monte Carlo dose calculation based on deep learning | |
| CN114038535A (en) | Radiotherapy dose prediction device and method based on deep learning | |
| Shin et al. | Dose Super-Resolution in Prostate Volumetric Modulated Arc Therapy Using Cascaded Deep Learning Networks | |
| Witt et al. | A proton simulator for testing implementations of proton CT reconstruction algorithms on GPGPU clusters | |
| Qureshi et al. | Determination of optimal number of projections and parametric sensitivity analysis of operators for parallel‐ray transmission tomography using hybrid continuous genetic algorithm | |
| Melodia | Deep Learning Estimation of Absorbed Dose for Nuclear Medicine Diagnostics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |