Proton dose calculation model establishment method and proton dose calculation method

Through multi-layer long and short-term memory network, the proton dose calculation model is established, and energy information is integrated, and the problem of the balance between accuracy and speed of proton dose calculation methods in the existing technology is solved, and efficient and universal proton dose calculation is achieved. for clinical application.

CN119092047APending Publication Date: 2024-12-06XIEHE HOSPITAL ATTACHED TO TONGJI MEDICAL COLLEGE HUAZHONG SCI & TECH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411110281.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Existing proton dose calculation methods are difficult to balance between accuracy and speed, especially when dealing with tissue inhomogeneity and real-time adaptive radiotherapy, and AI-based and deep learning-based approaches cannot accomplish any arbitrary on lightweight models. The dosage calculation of energy proton beams limits its universality and clinical application.

Method used

A multi-layer long and short-term memory network is used to establish a proton dose calculation model. By constructing a training data set, including proton beam irradiation samples and corresponding dose distributions, the network is trained to integrate energy information and achieve accurate calculation of any energy proton beam.

Benefits of technology

It improves the universality, accuracy and efficiency of proton dose calculation, and can accurately calculate the dose distribution of proton beams at different human parts and angles, and is suitable for clinical applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119092047A_ABST
    Figure CN119092047A_ABST
Patent Text Reader

Abstract

The invention discloses a proton dose calculation model establishment method and a proton dose calculation method, and belongs to the field of medical assistance, and the method comprises the steps: constructing a training data set, and the training data comprises a proton beam irradiation sample, proton beam energy and the dose distribution of a proton beam in the proton beam irradiation sample after the proton beam irradiation sample with the corresponding energy is irradiated; establishing a multi-layer long-short-term memory network, and training the multi-layer long-short-term memory network by using the training data set; during training, for each piece of training data, after an initial hidden state and an initial unit state of the long-short-term memory network are set as proton beam energy, a proton beam irradiation sample is converted into a two-dimensional CT slice sequence perpendicular to the incident direction of a proton beam, each two-dimensional CT slice is flattened into one dimension, a corresponding time step is input, and the two-dimensional CT slice sequence is converted into a two-dimensional CT slice sequence; and enabling the long-short-term memory network to calculate dose distribution, and supervising a calculation result by taking the dose distribution in the training data as label data. According to the method, the universality of rapid proton dose calculation can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of medical assistance, and more specifically, relates to a proton dose calculation model establishment method and a proton dose calculation method. Background Art

[0002] Proton therapy is a type of radiation therapy. After protons enter the human body, a sharp dose peak is formed at the end of the range, called the Bragg peak. By modulating the energy to broaden the Bragg peak, the Bragg peak can be made to cover the tumor. Because proton therapy has good physical properties in depth dose distribution - the Bragg peak (referring to the high-dose peak of radiation in proton therapy reaching its highest point when it reaches the tumor site), compared with traditional radiotherapy, it can achieve higher target area dose irradiation and realize low-dose irradiation of surrounding normal tissues. After the proton beam is incident on the human body, the dose distribution in the human body is an important basis for affine treatment plans. Therefore, a high-precision and high-calculation speed proton dose calculation method is extremely important for clinical cancer radiotherapy workers.

[0003] However, achieving a computational method that strikes a balance between accuracy and speed remains a challenge. Currently, conventional pencil beam algorithms and Monte Carlo simulations offer a trade-off between speed and accuracy. Pencil beam algorithms are widely used in commercial treatment planning systems due to their ability to quickly provide dose calculation results and low computational requirements, but they suffer from poor accuracy in dealing with tissue inhomogeneities. In contrast, Monte Carlo simulations provide accurate dose calculation results by simulating particle interactions, but their computational complexity is high and time-consuming, which limits their efficiency in rapidly generating dose calculations in a variety of scenarios, making them particularly challenging when dealing with motion changes in real-time adaptive radiotherapy.

[0004] In recent years, the application of artificial intelligence and deep learning in radiotherapy has increased year by year. The technology of using neural network models for rapid dose prediction is also constantly breaking through. Although such methods can improve the accuracy of results, data processing is time-consuming and usually based on relatively fixed scenarios. It can only predict the planned dose for a specific prescription dose or part. In addition, proton beams of different energies may be involved in the intensity-modulated radiotherapy plan at the same time, and existing methods cannot be well applied to the proton dose calculation in this scenario.

[0005] In general, the existing proton dose calculation methods based on artificial intelligence and deep learning cannot complete the dose calculation of proton beams of arbitrary energy based on lightweight models, which limits the universality of proton dose calculation and the clinical application of related calculation methods. Summary of the invention

[0006] In view of the defects and improvement needs of the prior art, the present invention provides a proton dose calculation model establishment method and a proton dose calculation method, the purpose of which is to improve the universality of proton dose calculation.

[0007] To achieve the above object, according to one aspect of the present invention, a method for establishing a proton dose calculation model is provided, comprising:

[0008] A training data set is constructed; in the training data set, each piece of training data includes a proton beam irradiation sample, a proton beam energy, and a dose distribution of the proton beam in the proton beam irradiation sample after the proton beam of corresponding energy irradiates the sample; the proton beam irradiation sample is a three-dimensional CT image matrix containing a human body structure;

[0009] A multi-layer long short-term memory network is established, and the long short-term memory network is trained using a training data set. After the training, a proton beam dose calculation model is obtained;

[0010] The LSTM network is trained using a training data set, including: for each piece of training data, setting the initial hidden state and the initial unit state of the LSTM network to the proton beam energy in the training data, converting the proton beam irradiation samples in the training data into a two-dimensional CT slice sequence perpendicular to the incident direction of the proton beam, and flattening each two-dimensional CT slice into one dimension, and inputting the long short-term memory unit of the corresponding time step in the LSTM network, so that the LSTM network calculates the dose distribution of the proton beam of the corresponding energy in the proton beam irradiation sample, and using the dose distribution in the training data as label data to supervise the calculation result of the LSTM network.

[0011] Furthermore, the distribution range of human body structures included in the proton beam irradiated sample covers the dose diffusion range of the proton beam in the human body.

[0012] Furthermore, the proton beam irradiated sample also includes a portion of the air region before the proton beam enters the human body structure.

[0013] Furthermore, the method for establishing the proton beam irradiation sample includes:

[0014] Step S1: After unifying the resolution of the CT image containing the human body structure to a preset resolution, an original three-dimensional CT image matrix is ​​constructed;

[0015] Step S2: creating a mask matrix for the original three-dimensional CT image matrix, wherein each element is used to mark the voxel at the corresponding position in the original three-dimensional CT image matrix as a voxel inside the human body or a voxel outside the human body;

[0016] Step S3: Make the proton beam irradiate the mask matrix to find the voxel in the first human body irradiated by the proton beam, and record its position as (x, y, z);

[0017] Step S4: in the original three-dimensional CT image matrix, starting from (x, y, z) and translating by a preset distance along three dimensions, so as to extract a three-dimensional CT image matrix containing human body structure of a preset dimension size as a proton beam irradiation sample;

[0018] Among them, the proton beam is centered in the extracted proton beam irradiated sample; in the proton beam irradiated sample, the cross-sectional dimension perpendicular to the incident direction can cover the diffusion range of the proton beam perpendicular to the incident direction, and the depth along the incident direction can cover the diffusion range of the proton beam along the incident direction.

[0019] Furthermore, in the proton beam irradiated sample, the cross-sectional dimension perpendicular to the incident direction is 48 mm×48 mm, and the depth along the incident direction is 300 mm.

[0020] Furthermore, the method for establishing the proton beam irradiation sample also includes: for each original three-dimensional CT image matrix and its mask matrix, rotating them according to a preset angle interval; at each angle, extracting the proton beam irradiation sample at the corresponding angle using step S3 and step S4.

[0021] Furthermore, the CT image containing human body structure is a CT image containing different parts of the human body.

[0022] Further, constructing a training data set includes: after constructing the proton beam irradiation sample, for each proton beam irradiation sample, using proton beams of different energies to be incident on the proton beam irradiation sample along the center position, calculating the dose distribution of the proton beam at each energy in the proton beam irradiation sample, and using the proton beam energy, the proton beam irradiation sample and the corresponding dose distribution as training data;

[0023] The training data in which the proton beam penetrates the body structure are removed from the training data, and the remaining training data constitute the training data set.

[0024] According to another aspect of the present invention, there is provided a proton beam dose calculation method, comprising:

[0025] After obtaining the patient's CT image, a proton beam irradiation sample is prepared for all proton beams within the radiation field of the patient's intensity modulated radiotherapy plan;

[0026] For each proton beam, after setting the initial hidden state and initial unit state of the proton dose calculation model to the proton beam energy of the proton beam, the corresponding proton beam irradiation sample is converted into a two-dimensional CT slice sequence perpendicular to the incident direction of the proton beam, and each two-dimensional CT slice is flattened into one dimension and input into the long short-term memory unit of the corresponding time step in the proton dose calculation model, so that the proton dose calculation model calculates the dose distribution of the proton beam in the proton beam irradiation sample;

[0027] The dose distribution corresponding to each proton beam is weighted and summed according to the proton beam weight to obtain the global dose distribution;

[0028] The proton dose calculation model is established by the proton dose calculation model establishment method provided by the present invention.

[0029] According to another aspect of the present invention, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the proton dose calculation model establishment method provided by the present invention and / or the proton dose calculation method provided by the present invention are implemented.

[0030] According to another aspect of the present invention, a computer-readable storage medium is provided, comprising a stored computer program; when the computer program is executed by a processor, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned proton dose calculation model establishment method provided by the present invention, and / or the above-mentioned proton dose calculation method provided by the present invention.

[0031] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0032] (1) The present invention establishes a proton dose calculation model based on a multi-layer long short-term memory network. The long short-term memory network structure can process sequence data well. The structural characteristics are just adapted to the characteristics of the gradual deposition of proton beam dose along the incident direction, and can accurately capture the diffusion characteristics of the proton beam. The calculation is accurate, and the network parameters are few, and the computing resource requirements are low. Therefore, the calculation efficiency is high. When training the model, the present invention sets the initial hidden state and unit state in the model to the proton beam energy corresponding to each training data, and integrates the energy information into the long-term memory unit. While the model remembers the energy label, it can also use the energy information from the initial stage to affect the output of the long short-term memory unit gate structure, so as to achieve the purpose of learning energy information. The model finally trained can realize the calculation of proton beam dose of any energy, and the universality is effectively improved. In general, the proton dose calculation model established by the present invention has high calculation accuracy, high calculation efficiency, and high universality.

[0033] (2) In the preferred embodiment of the present invention, the distribution range of human body structures contained in the proton beam irradiation samples in the training data set used for model training covers the dose diffusion range of the proton beam in the human body, thereby being able to fully capture the diffusion characteristics of the proton beam after it is incident on the human body, thereby ensuring the accuracy of the calculation.

[0034] (3) In a preferred embodiment of the present invention, the proton beam irradiation sample also includes a partial air region before the proton beam is incident on the human body, and the air region forms a "buffer" zone. When the actual proton beam is incident on the human body, a certain degree of reflection will occur on the skin surface, so that a certain dose of proton beam will exist in the air region before the human skin. By setting a "buffer" zone in the proton beam irradiation sample, the present invention can capture the proton beam diffusion process more completely and further improve the calculation accuracy.

[0035] (4) In a preferred embodiment of the present invention, the first voxel in the human body irradiated in the original three-dimensional CT image matrix after the proton beam is incident is determined and its position is recorded. Then, the proton beam irradiation sample is extracted by translating a preset distance along three dimensions from the position in the original three-dimensional CT image matrix. The proton beam irradiation sample thus extracted covers the proton beam diffusion range in both the perpendicular incident direction and along the incident direction, while filtering out information irrelevant to the proton dose calculation, thereby improving the calculation accuracy and reducing the calculation amount, thereby further improving the calculation efficiency. In a further preferred embodiment, in the proton beam irradiation sample, the cross-sectional dimension perpendicular to the incident direction is 48 mm×48 mm, and the depth along the incident direction is 300 mm, which can filter out information irrelevant to the proton dose calculation to the greatest extent while covering the proton beam diffusion range.

[0036] (5) In the preferred embodiment of the present invention, for each original three-dimensional CT image matrix and its mask matrix, they are rotated and proton beam irradiation samples at different angles are extracted, so that the model can learn the dose distribution when the proton beam is incident at different angles, further improving the universality of proton dose calculation.

[0037] (6) In a preferred embodiment of the present invention, the CT images containing human body structures used to construct the training data set are CT images containing different human body parts, thereby enabling the model to learn the dose distribution of the proton beam in different parts of the human body, thereby further improving the proton dose calculation.

[0038] (7) In the preferred embodiment of the present invention, a proton irradiation sample is constructed, and proton beams of different energies are incident on the proton beam irradiation sample to record the dose distribution of the proton beams of corresponding energies in the proton beam irradiation sample. Then, the training data in which the proton beam penetrates the body structure is eliminated to avoid the dose distribution caused by the proton beam penetrating the body affecting the model training effect, thereby ensuring the quality of the final data set and effectively improving the model training effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic flow chart of a method for calculating a proton dose provided by an embodiment of the present invention;

[0040] Figure 2 A diagram of a proton pencil beam dose calculation model based on a long short-term memory network provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0042] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the accompanying drawings are used to distinguish similar objects but not necessarily to describe a specific order or sequence.

[0043] In order to make the proton dose calculation more universal, more accurate and more efficient, the present invention provides a proton dose calculation model establishment method and a proton dose calculation method, the overall idea of ​​which is to establish a model that is compatible with the diffusion characteristics of the proton beam, and to train it using training data corresponding to different energies. During the training process, the proton beam energy information is integrated into the network structure so that the model learns the energy information, and the model finally trained can be used for proton beam dose calculation of any energy.

[0044] The following are examples.

[0045] Embodiment 1:

[0046] A method for establishing a proton dose calculation model, such as Figure 1 As shown, including:

[0047] A training data set is constructed; in the training data set, each piece of training data includes a proton beam irradiation sample, a proton beam energy, and a dose distribution of the proton beam in the proton beam irradiation sample after the proton beam of corresponding energy irradiates the sample; the proton beam irradiation sample is a three-dimensional CT image matrix containing a human body structure;

[0048] A multi-layer long short-term memory network is established, and the long short-term memory network is trained using a training data set. After the training, a proton beam dose calculation model is obtained.

[0049] In this embodiment, the proton beam irradiation samples in the training data set are obtained by preprocessing the imaging information of the patient, and the proton beam irradiation samples are used as input data in the model training process.

[0050] In order to make the model applicable to the calculation of proton doses of different parts of the human body, in the present embodiment, the selected patient imaging information in constructing the training data set is specifically CT images of different parts of the human body, which may include annotated tumor areas. The method for establishing the proton beam irradiation sample specifically includes:

[0051] Step S1: using a preset linear interpolation method to unify the resolution of the CT image containing the human body structure to a preset resolution, and constructing an original three-dimensional CT image matrix; the consistency of data resolution is helpful for the model to extract and learn data features; optionally, in this embodiment, the preset resolution is specifically 2mm×2mm×2mm, according to which the CT image can be divided into voxels of 2mm×2mm×2mm in size, and the original three-dimensional CT image matrix can be obtained;

[0052] Step S2: Create a mask matrix for the original three-dimensional CT image matrix, in which each element is used to mark the voxel at the corresponding position in the original three-dimensional CT image matrix as a voxel inside the human body or a voxel outside the human body; optionally, in this embodiment, the voxel value inside the human body is set to 1, and the voxel value outside the human body is set to 0, with M surface Represents the mask matrix, then the mask matrix can be expressed as follows:

[0053]

[0054] Among them, (x j ,y j ,z j ) represents the coordinates in the original three-dimensional CT image matrix, and “inbody” represents the human body; the mask matrix assists the planned CT image matrix in the subsequent proton beam irradiation sample extraction operation;

[0055] In order to expand the training data set and enable the model to learn the proton beam dose at different incident angles during the training process, as a preferred implementation, this embodiment rotates each original three-dimensional CT image matrix and its mask matrix according to a preset angle interval, and at each angle, uses subsequent steps S3 and S4 to extract the proton beam irradiation samples at the corresponding angle;

[0056] Optionally, in this embodiment, the angle interval of rotating the original three-dimensional CT image matrix and its mask matrix is ​​10 degrees. In practical applications, the angle interval can also be set to other values ​​according to practical application requirements, but it should not be too small, otherwise the proton beam irradiation sample extracted after one rotation is not much different from the proton beam irradiation sample extracted before the rotation, and it should not be too small, otherwise, the feature learning opportunity of a certain angle may be lost;

[0057] Steps S3 and S4 for obtaining proton beam irradiated samples at various angles are specifically as follows:

[0058] Step S3: Make the proton beam irradiate the mask matrix to find the first voxel in the human body irradiated by the proton beam, that is, the first voxel with a value of 1 on the 0 degree irradiation line in the mask matrix, and record its position as (x, y, z);

[0059] Step S4: in the original three-dimensional CT image matrix, starting from (x, y, z) and translating by a preset distance along three dimensions, so as to extract a three-dimensional CT image matrix containing human body structure of a preset dimension size as a proton beam irradiation sample;

[0060] Among them, the proton beam is centered in the extracted proton beam irradiated sample; in the proton beam irradiated sample, the cross-sectional dimension perpendicular to the incident direction can cover the diffusion range of the proton beam perpendicular to the incident direction, and the depth along the incident direction can cover the diffusion range of the proton beam along the incident direction.

[0061] In this embodiment, in the proton beam irradiated sample, the cross-section dimension perpendicular to the incident direction is 48mm×48mm, and the depth along the incident direction is 300mm. Specifically, after determining the voxel position (x, y, z), the position is translated upward by 10 voxels along the incident direction, and translated downward by 139 voxels, and the proton beam irradiated sample with the dimension (24, 150, 24) is extracted by left and right translation in the plane perpendicular to the incident direction. It should be noted that this extraction method ensures that the incident ray is as centered as possible in the extracted sample, and the cross-section dimension perpendicular to the incident ray is (24, 24), corresponding to the actual cross-section size of 48mm x 48mm, which can cover the dose diffusion range of the proton pencil beam. In the extraction method, the voxel position (x, y, z) is moved up by 10 voxels to down by 139 voxels in the proton beam incident direction, which not only ensures that the sample has a depth of 300 mm in the proton beam incident direction to receive a sufficient range of energy irradiation, but also includes the air area of ​​10 voxels in front of the incident line position, that is, 20 mm. The air area forms a "buffer" zone. When the actual proton beam is incident on the human body, a certain degree of reflection will occur on the skin surface, so that a certain dose of proton beam will exist in the air area before the human skin. In this embodiment, an air "buffer" zone is provided in the proton beam irradiated sample to assist the network in learning the incident front dose and the air conditions, thereby more completely capturing the proton beam diffusion process and further improving the calculation accuracy of the trained model.

[0062] It should be noted that in actual applications, the size of the vertical cross section of the incident beam can be adjusted according to the personalized model established for different patients, such as (20, 20) (18, 18), etc., to ensure that the size of the dimension covers the dose diffusion after the proton beam is injected, and it cannot be too large, otherwise the increase of invalid information will affect the learning effect of the imaging model. The air "buffer" area can also be adjusted to 15mm or 30mm as needed.

[0063] The proton beam irradiation samples extracted from medical images in the above manner have a dimensional size that matches the diffusion range of the proton beam, minimizing information irrelevant to the proton dose calculation and including situations at different incident angles and different parts of the human body, which helps to complete the calculation quickly and accurately and has high universality.

[0064] Based on the extracted proton beam irradiation samples, the specific method of constructing training in this embodiment includes:

[0065] For each proton beam irradiation sample, after proton beams of different energies are incident on the proton beam irradiation sample along the center position, the dose distribution of the proton beam at each energy in the proton beam irradiation sample is calculated, and the proton beam energy, the proton beam irradiation sample and the corresponding dose distribution are used as training data; specifically, a calculation method with higher calculation accuracy such as a pencil beam analysis algorithm and a Monte Carlo simulation algorithm can be used to complete the calculation of the dose distribution. Optionally, this embodiment uses a pencil beam analysis algorithm with faster calculation speed to complete the relevant calculations.

[0066] In this embodiment, a consistent proton beam injection method and calculation method are used for all proton beam irradiation samples to ensure the consistency of model data. The calculated real dose is used as label data for model learning.

[0067] In order to further improve the quality of the training data set and thus improve the training effect of the model, after obtaining the training data in the above manner, this embodiment will remove the training data in which the proton beam penetrates the body structure, so as to avoid the dose distribution generated by the proton beam penetrating the body affecting the training effect and ensure the quality of the final training data set. Finally, the remaining training data constitutes the training data set, which will be used to train the established long short-term memory network to obtain a model that can realize the calculation of proton doses of any energy.

[0068] Long Short-Term Memory (LSTM) is a time-recurrent neural network that selectively remembers or forgets information by controlling the flow of information in each LSTM unit through the input gate, forget gate, and output gate of each LSTM unit in the network.

[0069] The forget gate in each long short-term memory unit determines which information to discard through the σ function. The calculation formula is as follows:

[0070] F t =σ(W xf ·X t +W hf ·h t-1 +b f ),

[0071] Where Ft represents the output of the forget gate, W xf and W hf is the weight matrix, b f is the bias vector, X t is the input of the current time step and h t-1 is the hidden state of the previous time step. The input gate in the unit controls how much new input data is added to the unit state through the σ function and the tanh function. The unit state can retain information across multiple time steps as a long-term storage unit. The calculation formula is as follows:

[0072] I t =σ(W xi ·X t +W hi ·h t-1 +b i ),

[0073]

[0074] Among them, I t represents the output of the input gate, is a candidate value vector. Finally, the output gate combines the cell state with the current input and uses the σ function and tanh function to decide which information in the cell state should be used to generate the output. The calculation formula is as follows:

[0075] O t =σ(W xo ·X t +W ho ·h t-1 +b o ),

[0076] h t =O t tanh(C t ),

[0077] Among them, O t represents the output of the output gate, h t and C t Represent the hidden state and unit state of the current time step respectively.

[0078] The long short-term memory network structure can process sequence data very well. This structural characteristic is just suitable for the characteristic that the proton beam dose is gradually deposited along the incident direction. This embodiment establishes a proton dose calculation model based on a multi-layer long short-term memory network, which can accurately capture the diffusion characteristics of the proton beam, accurately calculate, and has few network parameters and low computing resource requirements, so the calculation efficiency is high. Optionally, if Figure 2As shown, the four-layer long short-term memory network established in this embodiment corresponds to the dimension of the proton beam irradiated sample, and each layer includes 150 time steps.

[0079] In order to enable the long short-term memory network to accurately learn the diffusion characteristics and energy information of the proton beam dose, in this embodiment, the long short-term memory network is trained using a training data set, including:

[0080] For each piece of training data, after setting the initial hidden state and initial unit state of the LSTM network to the proton beam energy in the training data, the proton beam irradiation samples in the training data are converted into a two-dimensional CT slice sequence perpendicular to the proton beam incident direction, and each two-dimensional CT slice is flattened into one dimension and input into the LSTM unit of the corresponding time step in the LSTM network, so that the LSTM network calculates the dose distribution of the proton beam of the corresponding energy in the proton beam irradiation sample, and uses the dose distribution in the training data as label data to supervise the calculation results of the LSTM network.

[0081] According to the above training method, the proton beam irradiation sample in each training data will be converted into 150 two-dimensional CT slices perpendicular to the incident direction. These 150 two-dimensional CT slices will be flattened into one dimension and input into the long short-term memory units corresponding to 150 time steps respectively; and the proton beam energy information will be put into the initial hidden state and initial unit state of the model, that is, the hidden state and unit state of the long short-term memory unit corresponding to the first time step of each layer. In this way, the energy information can not only be integrated into the long-term memory unit to allow the model to remember the energy label, but also the energy information can be used to influence the output of the long short-term memory unit gate structure from the initial stage.

[0082] After obtaining the dose distribution output by the long short-term memory network, the loss function can be calculated based on the model calculation results and the label data, and the partial derivative can be obtained to obtain the updated gradient parameters of the model. Optionally, in this embodiment, the loss function selects the mean square error (MSE), and the calculation formula is as follows:

[0083]

[0084] Among them, D lstm is the dose output by the model, D label is the actual dose in the corresponding data set, (i, j, k) represents the voxel position in the proton beam irradiated sample, and M is the number of voxels contained in the proton beam irradiated sample.

[0085] Parameters are updated according to the gradient descent direction of the updated gradient parameters until the value of the loss function is minimized or the preset maximum number of training times is reached, thereby obtaining a proton pencil beam dose calculation model that can be used for any energy.

[0086] In general, this embodiment converts the acquired patient medical imaging into proton beam irradiation samples as input data in the training data set, and applies the pencil beam analysis algorithm to obtain the real pencil beam dose as the label data in the training data set, and records the corresponding proton beam energy. The training data set is input into the proton dose calculation model built based on the long short-term memory network structure for training. During training, the proton beam energy information is put into the initial hidden state and initial unit state of the model. After the training is completed, a dose calculation model for proton pencil beams at any energy is obtained. The proton dose calculation model established in this embodiment has high universality, a lightweight model, and fast calculation, which greatly improves the efficiency of accurate proton dose calculation.

[0087] Embodiment 2:

[0088] A proton beam dose calculation method, comprising:

[0089] After obtaining the patient's CT image, a proton beam irradiation sample is prepared for all proton beams within the radiation field of the patient's intensity modulated radiotherapy plan;

[0090] For each proton beam, after setting the initial hidden state and initial unit state of the proton dose calculation model to the proton beam energy of the proton beam, the corresponding proton beam irradiation sample is converted into a two-dimensional CT slice sequence perpendicular to the incident direction of the proton beam, and each two-dimensional CT slice is flattened into one dimension and input into the long short-term memory unit of the corresponding time step in the proton dose calculation model, so that the proton dose calculation model calculates the dose distribution of the proton beam in the proton beam irradiation sample;

[0091] The dose distribution corresponding to each proton beam is weighted and summed according to the proton beam weight to obtain the global dose distribution; the proton beam weight is set by the intensity modulated radiotherapy plan;

[0092] The proton dose calculation model is established by the proton dose calculation model establishment method provided in the above-mentioned embodiment 1.

[0093] Embodiment 3:

[0094] A computer program product includes a computer program. When the computer program is executed by a processor, it implements the proton dose calculation model establishment method provided in the above embodiment 1 and / or the proton dose calculation method provided in the above embodiment 2.

[0095] Embodiment 4:

[0096] A computer-readable storage medium includes a stored computer program; when the computer program is executed by a processor, the device where the computer-readable storage medium is located is controlled to execute the proton dose calculation model establishment method provided in the above-mentioned embodiment 1, and / or the proton dose calculation method provided in the above-mentioned embodiment 2.

[0097] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for establishing a proton dose calculation model. It is characterized in that include: Build a training dataset; In the training data set, each piece of training data includes a proton beam irradiation sample, a proton beam energy, and a dose distribution of the proton beam in the proton beam irradiation sample after the proton beam of corresponding energy irradiates the sample; the proton beam irradiation sample is a three-dimensional CT image matrix containing a human body structure; Establishing a multi-layer long short-term memory network, and using the training data set to train the long short-term memory network, after the training is completed, obtaining the proton beam dose calculation model; The LSTM network is trained using the training data set, including: for each piece of training data, after the initial hidden state and the initial unit state of the LSTM network are set to the proton beam energy in the training data, the proton beam irradiation samples in the training data are converted into a two-dimensional CT slice sequence perpendicular to the proton beam incident direction, and each two-dimensional CT slice is flattened into one dimension, and the long short-term memory unit of the corresponding time step in the LSTM network is input, so that the LSTM network calculates the dose distribution of the proton beam of the corresponding energy in the proton beam irradiation sample, and the dose distribution in the training data is used as label data to supervise the calculation result of the LSTM network.

2. The method for establishing a proton dose calculation model according to claim 1, It is characterized in that The distribution range of the human body structures included in the proton beam irradiated sample covers the dose diffusion range of the proton beam in the human body.

3. The method for establishing a proton dose calculation model according to claim 2, It is characterized in that The proton beam irradiated sample also includes a portion of the air region before the proton beam is incident on the human body structure.

4. The method for establishing a proton dose calculation model according to claim 3, It is characterized in that The method for establishing the proton beam irradiation sample comprises: Step S1: After unifying the resolution of the CT image containing the human body structure to a preset resolution, an original three-dimensional CT image matrix is ​​constructed; Step S2: creating a mask matrix for the original three-dimensional CT image matrix, wherein each element of the mask matrix is ​​used to mark the voxel at the corresponding position in the original three-dimensional CT image matrix as a voxel inside the human body or a voxel outside the human body; Step S3: Make the proton beam irradiate the mask matrix to find the voxel in the first human body irradiated by the proton beam, and record its position as (x, y, z); Step S4: in the original three-dimensional CT image matrix, starting from (x, y, z) and translating by a preset distance along three dimensions, so as to extract a three-dimensional CT image matrix containing human body structure of a preset dimension size as the proton beam irradiation sample; The proton beam is centered in the extracted proton beam irradiated sample; in the proton beam irradiated sample, the cross-sectional dimension perpendicular to the incident direction can cover the diffusion range of the proton beam perpendicular to the incident direction, and the depth along the incident direction can cover the diffusion range of the proton beam along the incident direction.

5. The method for establishing a proton dose calculation model according to claim 4, It is characterized in that In the proton beam irradiated sample, the cross-sectional dimension perpendicular to the incident direction is 48 mm×48 mm, and the depth along the incident direction is 300 mm.

6. The method for establishing a proton dose calculation model according to claim 4 or 5, It is characterized in that The method for establishing the proton beam irradiation sample also includes: for each original three-dimensional CT image matrix and its mask matrix, rotating them according to a preset angle interval; at each angle, extracting the proton beam irradiation sample at the corresponding angle by using step S3 and step S4.

7. The method for establishing a proton dose calculation model according to claim 4 or 5, It is characterized in that The CT image containing human body structure is a CT image containing different parts of the human body.

8. The method for establishing a proton dose calculation model according to claim 7, It is characterized in that The constructing of the training data set comprises: after constructing the proton beam irradiation sample, for each proton beam irradiation sample, after using proton beams of different energies to be incident on the proton beam irradiation sample along the center position, calculating the dose distribution of the proton beam at each energy in the proton beam irradiation sample, and using the proton beam energy, the proton beam irradiation sample and the corresponding dose distribution as training data; The training data in which the proton beam penetrates the body structure are removed from the training data, and the remaining training data constitute the training data set.

9. A proton beam dose calculation method, It is characterized in that include: After obtaining the patient's CT image, a proton beam irradiation sample is prepared for all proton beams within the radiation field of the patient's intensity modulated radiotherapy plan; For each proton beam, after setting the initial hidden state and the initial unit state of the proton dose calculation model to the proton beam energy of the proton beam, the corresponding proton beam irradiation sample is converted into a two-dimensional CT slice sequence perpendicular to the incident direction of the proton beam, and each two-dimensional CT slice is flattened into one dimension, and input into the long short-term memory unit of the corresponding time step in the proton dose calculation model, so that the proton dose calculation model calculates the dose distribution of the proton beam in the proton beam irradiation sample; The dose distribution corresponding to each proton beam is weighted and summed according to the proton beam weight to obtain the global dose distribution; Wherein, the proton dose calculation model is established by the proton dose calculation model establishment method described in any one of claims 1 to 8.

10. A computer program product, It is characterized in that It comprises a computer program, which, when executed by a processor, implements the method for establishing a proton dose calculation model according to any one of claims 1 to 8, and / or the proton dose calculation method according to claim 9.