A method and device for optimizing the intensity distribution of radiotherapy plans and radiotherapy equipment

Abstract

The present invention discloses a method, device and radiotherapy equipment for optimizing the intensity distribution of radiotherapy plans. The method comprises the following steps: mapping the PB dose to the MC dose through a neural network; solving the gradient of the MC dose relative to the PB dose to obtain a deposition matrix based on the MC algorithm; bringing the deposition matrix based on the MC algorithm into FMO to achieve optimization of the intensity distribution of radiotherapy plans. The present invention applies the MC algorithm to FMO to ensure that the intensity distribution obtained in the optimization process is true and reliable. Using such an intensity distribution to perform multi-leaf grating sub-field segmentation, the sub-field movement obtained is more reasonable, ensuring that the final planned dose distribution is closer to clinical requirements, and the planning effect is significantly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiotherapy, and in particular relates to a method and device for optimizing the intensity distribution of a radiotherapy plan, and radiotherapy equipment. Background Art

[0002] Inverse planning intensity modulated radiotherapy technology is the mainstream technology used by the radiotherapy department of medical institutions. Its optimization methods include: two-step method and one-step method (direct field optimization). The two-step method refers to: intensity distribution optimization and grating field segmentation. Intensity distribution optimization is the process of solving the optimal ray intensity distribution that meets clinical requirements based on patient images and clinical requirements; grating field segmentation is the process of converting the optimal intensity distribution into a series of fields that can be executed by a multi-leaf grating (Multi-Layers Collimator, MLC) physically. In intensity distribution optimization (Fluence Map Optimization, FMO), each continuous beam is divided into many small beams (beamlets), and the dose deposition of each small beam in the medium is considered separately, and the intensity value of each small beam is solved using optimization methods such as gradient descent method. Finally, for a specific point, the dose contribution of each beam to it is superimposed to obtain the total dose of the point.

[0003] An important step in FMO is to calculate the dose deposition of radiation intensity in human tissue. General software uses the pencil beam method (PB). The dose value of the PB algorithm is the two-dimensional convolution result of the pencil beam kernel and the intensity distribution. The pencil beam kernel describes the energy diffusion pattern of a beam incident in an infinite uniform medium. The methods for solving the pencil beam kernel are: MC simulation, direct acquisition from measurement data, deconvolution from the measured wide beam dose distribution, and finite pencil beam method. For the case of inhomogeneous media, the 3D density correction of the medium, the scattering of the rays before reaching the phantom, and electronic contamination are considered. Although the PB algorithm can handle irregular radiation fields better, the calculation accuracy is still insufficient for inhomogeneous human tissues and irregular surface contours, especially for hollow tissues such as the trachea and larynx. The dose difference can reach 100%. Figure 2 In order to better simulate the dose deposition of radiation in non-uniform tissues, many studies have proposed methods such as original radiation flux correction and kernel function correction, but these are essentially data fitting methods and their accuracy is far inferior to that of the MC method.

[0004] The MC algorithm is also known as the computer random simulation method. Its principle is based on the generation of random numbers. By simulating the process of a large number of particles interacting in the medium, the distribution of flux, energy spectrum, and dose in the entire medium is obtained. Since MC realistically simulates the process of energy deposition of a large number of particles (up to hundreds of millions), the simulation results are recognized to be the most accurate.

[0005] Most commercial software FMO uses the PB algorithm for computational speed. Since the PB algorithm itself is not very accurate, the optimal intensity distribution obtained by optimization is pseudo-optimal. The dose distribution obtained by the final MC algorithm does not necessarily meet the expected clinical requirements, and the result is contrary to the original intention of setting clinical requirements. Therefore, the use of the MC algorithm in plan optimization is an important research direction.

[0006] Most researchers use tools such as EGS4 to pre-generate the MC deposition matrix of all beamlets, and then directly call the values ​​in the intensity distribution optimization. The efficiency of performing MC dose calculation directly on the CPU is very low. In order to speed up the calculation, the application of MC dose calculation on GPU has gradually increased, and GPU-based MC dose calculation software libraries have been developed, such as gDPM and GPUMCD. CPU-based MC optimization provides a solution. Its application on GPU can effectively improve the calculation speed, but for intensity distribution optimization, the speed is still too slow to meet the requirements of commercial software. In addition, these two methods require the deposition matrix to be stored in the calculation, which takes up a lot of hard disk and memory space, and each patient needs to be calculated once, which takes a lot of time and cannot meet the needs of medical institutions with a large number of patients every day. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a method and device for optimizing the intensity distribution of radiotherapy plans and radiotherapy equipment.

[0008] In order to achieve the above object, the technical solution of the present invention is as follows:

[0009] In one aspect, the present invention discloses a method for optimizing intensity distribution of a radiotherapy plan, comprising the following steps:

[0010] S1: Mapping PB dose to MC dose through neural network;

[0011] S2: Solve the gradient of MC dose relative to PB dose and obtain the deposition matrix based on MC algorithm;

[0012] S3: Bring the deposition matrix based on the MC algorithm into FMO to optimize the intensity distribution of radiotherapy plans.

[0013] Based on the above technical solution, the following improvements can be made:

[0014] As a preferred solution, S1 specifically includes the following steps:

[0015] S1.1: Data preparation: select N1 patient CTs, add N2 radiation fields to each patient, calculate the PB dose and MC dose of each radiation field, and form N1*N2 groups of data. Each group of data includes the corresponding CT, PB dose and MC dose.

[0016] S1.2: Data division: divide the N1*N2 group data into training set and test set in proportion;

[0017] S1.3: Create a convolutional neural network;

[0018] S1.4: Use the training set data to train the created convolutional neural network;

[0019] S1.5: Test the test set data on the trained convolutional neural network to obtain the predicted dose, and calculate the gamma pass rate of the predicted dose and MC dose.

[0020] As a preferred solution, the convolutional neural network includes: an input layer, a fusion layer, a convolution layer, an activation layer, a pooling layer, an upsampling layer, and an output layer.

[0021] As a preferred solution,

[0022]

[0023]

[0024] Where A is the deposition matrix based on the MC algorithm;

[0025] D PB With D MC They are the dose based on the PB algorithm and the dose based on the MC algorithm;

[0026] a jm is each value of the deposition matrix based on the MC algorithm;

[0027] represents the derivative of the MC dose at the voxel point (jth) relative to the kth PB dose;

[0028] dD PB,k / dx m represents the derivative of the kth PB dose with respect to the mth intensity value;

[0029] N dose Represents the number of doses;

[0030] From the above formula, it can be seen that by solving the gradient of the convolutional neural network mapping from PB dose to MC dose, the deposition matrix based on the MC algorithm can be obtained.

[0031] As a preferred solution, FMO specifically includes the following steps:

[0032] S3.1: Split the field into pencil beams;

[0033] S3.2: Introducing the deposition matrix based on the MC algorithm;

[0034] S3.3: Set the initial strength value as the initial value for optimization;

[0035] S3.4: Solve the derivative of the objective function using the deposition matrix based on the MC algorithm;

[0036] S3.5: Find the optimal step size under the current gradient;

[0037] S3.6: Update intensity values, doses, and objective functions;

[0038] S3.7: Repeat steps S3.4-S3.6 until the stopping condition is met and the iteration stops.

[0039] As a preferred solution, the following steps are also included:

[0040] S4: Calculate the dose in the human body and the DVH of each organ based on the intensity distribution optimization method of the radiotherapy plan, and evaluate the effect of the plan.

[0041] On the other hand, the present invention further discloses a radiotherapy plan intensity distribution optimization device, comprising:

[0042] A neural network mapping module is used to map the PB dose to the MC dose through a neural network;

[0043] The deposition matrix acquisition module is used to solve the gradient of MC dose relative to PB dose and obtain the deposition matrix based on MC algorithm;

[0044] The intensity distribution optimization module is used to bring the deposition matrix based on the MC algorithm into FMO to achieve intensity distribution optimization of radiotherapy plans.

[0045] On the other hand, the present invention further discloses a radiotherapy device for executing any one of the above-mentioned radiotherapy plan intensity distribution optimization methods, or comprising: the above-mentioned radiotherapy plan intensity distribution optimization device.

[0046] The method, device and radiotherapy equipment for optimizing the intensity distribution of radiotherapy plans of the present invention have the following beneficial effects:

[0047] The present invention applies the MC algorithm in FMO to ensure that the intensity distribution obtained in the optimization process is true and reliable. Using such an intensity distribution to segment the multi-leaf grating sub-fields, the obtained sub-field movement is more reasonable, ensuring that the final planned dose distribution is closer to clinical requirements, and the planning effect is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 A flowchart of a method for optimizing intensity distribution of a radiotherapy plan provided in an embodiment of the present invention.

[0050] Figure 2 This is a comparison chart of MC dose and PB dose;

[0051] Figure 2 (a) is the MC dose diagram;

[0052] Figure 2 (b) is the PB dose map.

[0053] Figure 3 A layer of data graph in a training set provided by an embodiment of the present invention;

[0054] Figure 3 (a) is CT image;

[0055] Figure 3 (b) is the PB dose data diagram;

[0056] Figure 3 (c) is the MC dosage data diagram.

[0057] Figure 4 A flow chart for optimizing the intensity distribution diagram provided in an embodiment of the present invention.

[0058] Figure 5 A flow chart of a method for mapping a PB dose to a MC dose provided in an embodiment of the present invention.

[0059] Figure 6 A schematic diagram of the structure of a convolutional neural network provided in an embodiment of the present invention.

[0060] Figure 7 A comparison chart of the neural dose gamma pass rate predicted by the convolutional neural network provided in an embodiment of the present invention and the PB dose and gamma pass rate;

[0061] Figure 7 (a) is the MC dose diagram;

[0062] Figure 7 (b) is the MC dose distribution diagram;

[0063] Figure 7(c) is the GroundTruth graph;

[0064] Figure 7 (d) is the PB dose map;

[0065] Figure 7 (e) is the PB dose distribution diagram;

[0066] Figure 7 (f) is the PB dose gamma pass rate diagram;

[0067] Figure 7 (g) Neural dose map predicted by convolutional neural network;

[0068] Figure 7 (h) Neural dose distribution map predicted by convolutional neural network;

[0069] Figure 7 (i) Neural dose gamma pass rate diagram predicted by convolutional neural network.

[0070] Figure 8 A diagram of the convolution process of a 3*3 convolution kernel on a 4*4 matrix (with 0s padded on the outside) provided in an embodiment of the present invention.

[0071] Fig. 9 A diagram of the reverse pooling upsampling process provided by an embodiment of the present invention.

[0072] Fig.10 A diagram of the interpolation upsampling process provided by an embodiment of the present invention.

[0073] Fig.11 A diagram of the deconvolution upsampling process provided by an embodiment of the present invention.

[0074] Fig.12 A comparison chart between the optimized DVH based on a convolutional neural network and the optimized DVH based on MC provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] In order to achieve the purpose of the present invention, a method and apparatus for optimizing the intensity distribution of a radiotherapy plan and some embodiments of a radiotherapy device are provided. Figure 1 As shown, the method for optimizing the intensity distribution of radiotherapy plans comprises the following steps:

[0078] S1: Mapping PB dose to MC dose through neural network;

[0079] S2: Solve the gradient of MC dose relative to PB dose and obtain the deposition matrix based on MC algorithm;

[0080] S3: Bring the deposition matrix based on the MC algorithm into FMO to optimize the intensity distribution of radiotherapy plans.

[0081] It is worth noting that in S1, the PB dose and MC dose can be, but are not limited to, calculated by the planning system TPS.

[0082] The calculation method of PB dose is to calculate the dose deposition of main rays and scattered rays in tissues according to formulas or test results, and the calculation method of MC dose is to simulate the dose deposition of billions of particles in tissues.

[0083] In order to further optimize the implementation effect of the present invention, in some other implementations, the remaining characteristic technologies are the same, except that S1 specifically includes the following steps:

[0084] S1.1: Data preparation: select N1 patient CTs, add N2 radiation fields to each patient, calculate the PB dose and MC dose of each radiation field, and form N1*N2 groups of data. Each group of data includes the corresponding CT, PB dose and MC dose.

[0085] S1.2: Data division: divide the N1*N2 group data into training set and test set in proportion;

[0086] S1.3: Create a convolutional neural network;

[0087] S1.4: Use the training set data to train the created convolutional neural network;

[0088] S1.5: Test the test set data on the trained convolutional neural network to obtain the predicted dose, and calculate the gamma pass rate of the predicted dose and MC dose.

[0089] Furthermore, the convolutional neural network includes: input layer, fusion layer, convolution layer, activation layer, pooling layer, upsampling layer, output layer, such as Figure 6 shown.

[0090] In a specific embodiment, the method for mapping PB dose to MC dose can be specifically divided into the following steps: Figure 5 As shown;

[0091] S1.1: Data preparation:

[0092] 50 patients were selected, and their CT scans covered the head and neck, chest, abdomen, pelvis and other parts. 72 radiation fields (with an interval of 0.5°) were added to each patient, and the isocenter positions of the radiation fields were distributed in different CT layers as much as possible, so that the radiation fields could cover more CT ranges.

[0093] The PB dose and MC dose of each field are calculated separately. It is worth noting that the dose network here remains unified and a 3 mm grid can be used;

[0094] Overall, there are 3600 sets of CT, PB dose, and MC dose data, such as Figure 3 As shown;

[0095] Since the sampling interval and size of CT data are inconsistent with the dose grid, the CT data needs to be interpolated to a size and interval consistent with the dose grid.

[0096] S1.2: Data division: Divide the 3600 sets of data into training set, validation set and test set in a ratio of 7:2:1.

[0097] S1.3: Create a convolutional neural network: A convolutional neural network consists of an input layer, a fusion layer, a convolution layer, an activation layer, a pooling layer, an upsampling layer, and an output layer.

[0098] Random cropping is performed on the CT and PB dose images to improve the generalization ability of the neural network. The processed image size is 128×128. The two are fused into a set of input data (128×128×2) and then processed by 64 convolution kernels to form a 128×128×64 tensor. The convolution process is continued to form a 128×128×64 tensor, which is then processed by maximum pooling to form a 64×64×64 tensor, and so on, until a 128×128×64 tensor is obtained after upsampling and convolution, and a 128×128 output image is obtained after convolution again.

[0099] The true output value of the convolutional neural network is the MC dose, and the loss function is defined as the mean square error (MSE) between the MC dose and the predicted dose, which is the mean of the sum of the squared distances between the target variable and the predicted value:

[0100]

[0101] Wherein: n is 128×128.

[0102] S1.4: Train convolutional neural network;

[0103] The process of network training is to continuously iterate and solve the convolution kernel, which is the process of feature extraction;

[0104] The specific steps of training are:

[0105] i. Randomly select data (batch size = 512) and perform online data augmentation on the input image. The methods include translation and 2D rotation. The purpose is to increase the amount of data and improve the generalization ability of the convolutional neural network;

[0106] ii. Use stochastic gradient descent to back-propagate the gradients of all weights and offsets of the network, using a learning rate of 10 -5 Update the weights and offsets, and on this basis, obtain the new predicted value, compare it with the MC dose, and obtain the mean square error. Repeat this process 200 times to complete one epoch.

[0107] iii. Repeat steps i) and ii) until the mean square error is less than the threshold or reaches the maximum epoch (e.g. 5000), the iteration stops and the convolutional neural network training is completed;

[0108] The above calculation process is performed on NVIDIA RTX 2080Ti.

[0109] In this embodiment, the data is divided into a training set, a validation set, and a test set. During the training process, generally after a few epochs, the validation set is run to see the effect. This allows timely discovery of model or parameter problems, timely termination of training, and re-adjustment of parameters or adjustment of the model. In addition, the generalization ability of the model can be verified. If the effect on the validation set is much worse than that on the training set, it should be considered whether the model is overfitted.

[0110] S1.5: Test results;

[0111] Test on the trained convolutional neural network, input the CT and PB dose data of the test set to obtain the predicted dose, save all the predicted dose data locally, and use dose verification software (such as DVS, LinaTech) to calculate the gamma pass rate of the predicted dose and MC dose (3mm, 3%, 10%). A pass rate of more than 95% is qualified.

[0112] Gamma pass rate is a quantitative measure of the difference between "measurement" and "calculation". It is a combination of Distance To Agreement (DTA) and Dose Difference (DD):

[0113]

[0114]

[0115] r(r m ,rc )=|r c -r m |;

[0116] δ(r m ,r c )=D c (r c )-D m (r m );

[0117] in:

[0118] r(r m ,r c ) represents the measurement point r m and the calculation point r c The distance between

[0119] δ(r m ,r c ) represents the dose difference between the two;

[0120] Δd M Represents the distance threshold (can be but not limited to 3mm);

[0121] ΔD M The threshold value representing the dose (which can be, but is not limited to, 3% of the maximum dose);

[0122] When γ(r m )≤1, the point passes, otherwise it fails.

[0123] If only the distance factor (δ(r m ,r c )=0), then the basis for the measurement point to pass is: within the distance threshold (such as: 3mm), there is a calculated point with a dose value equal to the measurement point;

[0124] If only the dose factor (r(r m ,r c )=0), then the basis for the passing of the measurement point is that at the same position, the difference between the calculated dose and the measured dose is not greater than the dose threshold (such as: 3% of the maximum value).

[0125] like Figure 7 As shown in the figure, it shows the gamma pass rate of a certain layer of PB dose and neural dose (dose based on neural network, that is, the predicted dose of convolutional neural network). The MC dose is the true value. It can be seen that the neural dose is closer to the MC dose. The statistical results show that the gamma pass rate of the neural dose reaches 99%, while the gamma pass rate of the PB dose is only 62%.

[0126] In order to further optimize the implementation effect of the present invention, in some other embodiments, the remaining characteristic technologies are the same, except that the deposition matrix based on the MC algorithm is obtained, and the principle is:

[0127]

[0128]

[0129] Where A is the deposition matrix based on the MC algorithm;

[0130] D PB With D MC They are the dose based on the PB algorithm and the dose based on the MC algorithm;

[0131] a jm is each value of the deposition matrix based on the MC algorithm;

[0132] represents the derivative of the MC dose at the voxel point (jth) relative to the kth PB dose;

[0133] dD PB,k / dx m represents the derivative of the kth PB dose with respect to the mth intensity value;

[0134] N dose Represents the number of doses;

[0135] From the above formula, it can be seen that by solving the gradient of the convolutional neural network mapping from PB dose to MC dose, the deposition matrix based on the MC algorithm can be obtained.

[0136] It is worth noting that the above formula transforms the problem of "calculating the deposition matrix using the MC algorithm" into "solving the derivative of the MC dose to the PB dose", where: The derivative dD of the PB dose to the intensity value x is PB,k / dx m It is the deposition matrix obtained by the original algorithm, that is, the matrix calculated by the TPS software using the method of superposition of principal rays and scattered rays.

[0137] The essence of "solving the derivative of MC dose with respect to PB dose" is to solve the gradient of "convolutional neural network mapping PB dose to MC dose". The solution method can refer to the back-propagation algorithm.

[0138] 1) Solve the gradients of each type of network. The following describes the gradient solutions of the convolutional layer (with activation layer), pooling layer, and upsampling layer respectively.

[0139] i. Gradient of convolutional layer with activation layer: Figure 8 As shown, the input is X, the weight and bias of the convolutional layer are W and b respectively, the output is Y, and the activation function is Then we have:

[0140]

[0141] conv2(W,X) represents the discrete two-dimensional convolution of W and X, which can also be written as:

[0142]

[0143] Therefore, there is

[0144]

[0145]

[0146] in:

[0147] k reflects the size of the convolution kernel (2*k+1=kernel size). When a 3×3 convolution kernel is used, k is 1.

[0148] Activation Function If it is a ReLU function, then its derivative is:

[0149]

[0150] Activation Function If it is a Sigmoid function, then its derivative is:

[0151]

[0152] According to the above formula, the derivative of the convolutional layer (with activation function) can be obtained.

[0153] ii. Gradient of the pooling layer: Taking Max Pooling and a 2×2 convolution kernel as an example, the output of this layer is equal to the maximum value of the corresponding input, that is, y(i,j)=max{x(p,q),p=i-1,i,q=j-1,j}, then the gradient can be expressed as:

[0154]

[0155] iii. Gradient of upsampling layer: There are three methods of upsampling: interpolation, deconvolution and anti-pooling;

[0156] a) If Fig.10 As shown, interpolation method: enlarge the input image to a certain size, and use interpolation method to calculate the pixel value of each point. Taking bilinear interpolation as an example, the pixel value of a certain point in the upsampled image is expressed as follows, where the upsampled image i, j corresponds to I, J: of the original image, and the gradient of the output to the input is recorded as the interpolation weight w;

[0157]

[0158] b) If Fig.11 As shown, deconvolution: Convolution can be expressed as the product of two matrices. Deconvolution is the transpose of the convolution kernel matrix and the multiplication of the original matrix. It is a conversion method from small size to large size. In essence, it is also a convolution. Its gradient calculation method can refer to the convolution method.

[0159] c) If Fig. 9 As shown in the figure, reverse pooling: taking the maximum pooling as an example, record the coordinates of max-pooling in the corresponding kernel. In the reverse pooling process, enlarge an element according to the kernel, fill in the element according to the previous coordinates, and fill in other positions with 0.

[0160] Its gradient calculation is similar to the method of the pooling layer: the gradient at the maximum value is passed directly, and the gradient at other locations is 0.

[0161] 2) Calculate the gradient in reverse order: Consider a network with an input layer A, a hidden layer B, and an output layer C. The information transmission direction is A→B→C. For each point C in C ij Relative to every point A in A st The gradient of can be passed through B and written as:

[0162]

[0163] Consider the entire convolutional neural network, starting from the last output layer, calculate the gradient relative to the previous layer, the first two layers, and the first three layers in turn, and continue recursively forward until the gradient relative to the input layer is calculated.

[0164] In order to further optimize the implementation effect of the present invention, in some other embodiments, the remaining characteristic technologies are the same, except that FMO specifically includes the following steps:

[0165] S3.1: Split the field into pencil beams;

[0166] S3.2: Introducing the deposition matrix based on the MC algorithm;

[0167] S3.3: Set the initial strength value as the initial value for optimization;

[0168] S3.4: Solve the derivative of the objective function using the deposition matrix based on the MC algorithm;

[0169] S3.5: Find the optimal step size under the current gradient;

[0170] S3.6: Update intensity values, doses, and objective functions;

[0171] S3.7: Repeat steps S3.4-S3.6 until the stopping condition is met and the iteration stops.

[0172] In a specific embodiment, the intensity distribution optimization is the first step of a two-step method for planning optimization, and the step of performing this optimization using the gradient descent method is shown in Figure 4 ,include:

[0173] S3.1: Field segmentation into beamlets: Divide the set field tungsten gate range into N×M rectangular grids, where N represents the width of the grating and M represents the direction of grating movement. Usually, the grid in the N direction is divided according to the grating width to facilitate the next step of grating sub-field segmentation, and the M direction is divided according to equal spacing;

[0174] S3.2: Introduce the deposition matrix A based on the MC algorithm;

[0175] S3.3: Set initial strength value: as the initial value for optimization;

[0176] S3.4: Solve the derivative of the objective function: The optimization objective function defined here is:

[0177]

[0178]

[0179] in:

[0180] is the objective function, expressed as the weighted quadratic error of the dose to the target volume (PTV) and the organ at risk (OAR);

[0181] is the dose error value of the critical organ, and α is its weighting;

[0182] is the dose error value of the critical organ, and β is its weight;

[0183] is the variable to be optimized, which is the intensity flux diagram of the pencil beams discretized from all the fields; is the quadratic error of the OAR, which is the second norm of the difference between the calculated dose and the ideal dose for all voxel points in all organs;

[0184] N OAR is the number of organs at risk;

[0185] N i is the number of voxel points of the i-th crisis organ;

[0186] w j is the weight of the j-th voxel;

[0187] δ j is the penalty factor, which is 0 when the calculated dose meets the preset constraints, otherwise it is 1;

[0188] is the dose value at point j calculated based on the field intensity flux map, which is the calculated dose;

[0189] p j is the ideal dose value;

[0190] Similarly, is the quadratic error of the target area, where N PTV is the number of voxel points in the target area;

[0191] a jm is the dose contribution of the mth pencil beam to the jth voxel, and the matrix composed of them is the deposition matrix A based on the MC algorithm.

[0192] Therefore, the derivative of the objective function is:

[0193]

[0194] S3.5: Solve for the optimal step size: Solve for the optimal step size under the current gradient;

[0195] S3.6: Update intensity values, doses, and objective functions;

[0196] S3.7: Determine whether the stop condition is met: when the independent variable (intensity value x) or the objective function J obj When the absolute change value or relative change value of (x) is less than the threshold ε, the iteration stops, otherwise repeat steps S3.4-S3.4.

[0197] In order to further optimize the implementation effect of the present invention, in some other embodiments, the remaining characteristic technologies are the same, except that the method for optimizing the intensity distribution of radiotherapy plans further includes the following steps:

[0198] S4: Calculate the dose in the human body and the DVH of each organ based on the intensity distribution optimization method of the radiotherapy plan, and evaluate the effect of the plan.

[0199] In a specific embodiment, gDPM is directly used to calculate the deposition matrix of all pencil beams (the total number of particles is about 2×10 9 ), and substitute it into the optimization process to obtain the optimal intensity distribution, and further obtain the dose distribution as the baseline for intensity distribution optimization.

[0200] The deposition matrix based on the MC algorithm is obtained by using the method disclosed in the present invention, and is added into the optimization process to obtain the optimized intensity distribution. The dose in the human body is calculated by using the MC algorithm, and the DVH of the organ is statistically calculated.

[0201] The DVH of the two are compared to verify the planning effect of the present invention.

[0202] The results of both can be seen in Fig.12 , Fig.12 In Chinese: neural dose represents the result based on neural network (i.e. the result of the present invention), and MC dose represents the result based on random calculation. It can be seen that the two results are almost completely consistent, which verifies the feasibility of the present invention.

[0203] On the other hand, an embodiment of the present invention further discloses a radiotherapy plan intensity distribution optimization device, comprising:

[0204] A neural network mapping module is used to map the PB dose to the MC dose through a neural network;

[0205] The deposition matrix acquisition module is used to solve the gradient of MC dose relative to PB dose and obtain the deposition matrix based on MC algorithm;

[0206] The intensity distribution optimization module is used to bring the deposition matrix based on the MC algorithm into FMO to achieve intensity distribution optimization of radiotherapy plans.

[0207] On the other hand, an embodiment of the present invention further discloses a radiotherapy device for executing the radiotherapy plan intensity distribution optimization method disclosed in any of the above embodiments, or comprising: a radiotherapy plan intensity distribution optimization device disclosed in the above embodiments.

[0208] The method, device and radiotherapy equipment for optimizing the intensity distribution of radiotherapy plans of the present invention have the following beneficial effects:

[0209] The present invention applies the MC algorithm in FMO to ensure that the intensity distribution obtained in the optimization process is true and reliable. Using such an intensity distribution to segment the multi-leaf grating sub-fields, the obtained sub-field movement is more reasonable, ensuring that the final planned dose distribution is closer to clinical requirements, and the planning effect is significantly improved.

[0210] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it. They cannot be used to limit the protection scope of the present invention. All equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for optimizing intensity distribution of radiotherapy plan, characterized in that: The following steps are involved: S1: Mapping PB dose to MC dose through neural network; S2: Solve the gradient of MC dose relative to PB dose to obtain the deposition matrix based on MC algorithm; S3: Bring the deposition matrix based on the MC algorithm into FMO to achieve the optimization of the intensity distribution of radiotherapy plan; Obtain the deposition matrix based on the MC algorithm, the principle is: Where A is the deposition matrix based on the MC algorithm; D PB With D MC They are the dose based on the PB algorithm and the dose based on the MC algorithm; x is the intensity value; a jm is each value of the deposition matrix based on the MC algorithm; represents the derivative of the MC dose at the jth voxel with respect to the kth PB dose; represents the derivative of the kth PB dose with respect to the mth intensity value; N dose Represents the number of doses; From the above formula, it can be seen that by solving the gradient of the convolutional neural network mapping from PB dose to MC dose, the deposition matrix based on the MC algorithm can be obtained.

2. The method for optimizing the intensity distribution of radiotherapy plans according to claim 1, characterized in that: S1 specifically includes the following steps: S1.1: Data preparation: select N1 patient CTs, add N2 radiation fields to each patient, calculate the PB dose and MC dose of each radiation field, and form N1*N2 groups of data. Each group of data includes the corresponding CT, PB dose and MC dose. S1.2: Data division: divide the N1*N2 group data into training set and test set in proportion; S1.3: Create a convolutional neural network; S1.4: Use the training set data to train the created convolutional neural network; S1.5: Test the test set data on the trained convolutional neural network to obtain the predicted dose, and calculate the gamma pass rate of the predicted dose and MC dose.

3. The method for optimizing the intensity distribution of radiotherapy plans according to claim 2, characterized in that: The convolutional neural network includes: an input layer, a fusion layer, a convolution layer, an activation layer, a pooling layer, an upsampling layer, and an output layer.

4. The method for optimizing the intensity distribution of radiotherapy plans according to claim 1, characterized in that: FMO specifically includes the following steps: S3.1: Split the field into pencil beams; S3.2: Introducing the deposition matrix based on the MC algorithm; S3.3: Set the initial strength value as the initial value for optimization; S3.4: Solve the derivative of the objective function using the deposition matrix based on the MC algorithm; S3.5: Find the optimal step size under the current gradient; S3.6: Update intensity values, doses, and objective functions; S3.7: Repeat steps S3.4-S3.6 until the stopping condition is met and the iteration stops.

5. The method for optimizing the intensity distribution of radiotherapy plans according to any one of claims 1 to 4, characterized in that: The following steps are also included: S4: Calculate the dose in the human body and the DVH of each organ according to the radiotherapy plan intensity distribution optimization method, and evaluate the effect of the plan.

6. A device for optimizing the intensity distribution of radiotherapy plans, characterized in that: include: A neural network mapping module is used to map the PB dose to the MC dose through a neural network; The deposition matrix acquisition module is used to solve the gradient of MC dose relative to PB dose and obtain the deposition matrix based on MC algorithm; Intensity distribution optimization module, which is used to bring the deposition matrix based on the MC algorithm into FMO to achieve intensity distribution optimization of radiotherapy plan; Obtain the deposition matrix based on the MC algorithm, the principle is: Where A is the deposition matrix based on the MC algorithm; D PB With D MC They are the dose based on the PB algorithm and the dose based on the MC algorithm; x is the intensity value; a jm is each value of the deposition matrix based on the MC algorithm; represents the derivative of the MC dose at the jth voxel with respect to the kth PB dose; represents the derivative of the kth PB dose with respect to the mth intensity value; N dose Represents the number of doses; From the above formula, it can be seen that by solving the gradient of the convolutional neural network mapping from PB dose to MC dose, the deposition matrix based on the MC algorithm can be obtained.

7. Radiotherapy equipment, characterized in that include: The radiotherapy plan intensity distribution optimization device as described in claim 6.

Images