Construction method of weight prediction model of proton uniform scanning radiation field dot matrix and weight prediction method

Abstract

The invention discloses an establishment method of a weight prediction model of a proton uniform scanning radiation field dot matrix and a weight prediction method, and belongs to the field of medical assistance. The method comprises the following steps: constructing a training data set, wherein training data comprises a three-dimensional CT image matrix with uniform dimensions, a dose matrix of a single aprotic robust plan, a position mask matrix of a plan dot matrix, a reference dose matrix, and a corresponding label as a weight matrix; establishing a three-dimensional attention mechanism Unet network, and training the three-dimensional attention mechanism Unet network by using the training data set; during training, each sample of training data is subjected to zero-dimension stacking after being normalized and then serves as input data to be input into the model, corresponding dot matrix weights are generated, plan dot matrix weights in the training data serve as label data, and a generation result is supervised and optimized. According to the invention, the traditional time-consuming and labor-consuming robustness plan weight generation process in proton therapy can be accelerated or replaced.

Description

Technical Field

[0001] This invention belongs to the field of medical auxiliary technology, and more specifically, relates to a method for constructing a weight prediction model for a proton uniform scanning field array and a weight prediction method. Background Technology

[0002] Proton radiotherapy demonstrates significant therapeutic efficacy by minimizing the dose to surrounding normal tissues and maximizing the dose to the tumor. However, due to its high sensitivity to patient location and anatomical changes, inter-fractional variations during treatment can significantly impact the effectiveness of the initial intensity-modulated proton therapy (IMRT) plan. To address the effects of these variations, robust optimization is employed to resolve uncertainties in location and range. Furthermore, adaptive proton therapy is implemented to overcome significant anatomical changes occurring during radiotherapy. However, in both approaches, traditional plan optimization is typically an extremely time-consuming process, requiring substantial human and resource investment.

[0003] Therefore, an increasing number of studies are exploring methods to accelerate robust intensity-modulated proton therapy (IMPT) and adaptive proton therapy planning optimization, combining these methods with automation to minimize labor and time costs, ultimately achieving rapid or even real-time robust treatment planning optimization and adjustment. With advancements in computing power and the development of deep learning methods, deep learning models have become a very promising solution due to their powerful learning and computational efficiency. Many proton therapy tasks, such as organ profiling and dose calculation, have successfully integrated appropriate deep learning models to reduce manual workload, accelerate proton therapy workflows, and achieve increasingly higher accuracy.

[0004] However, to date, no research has explored the application of deep learning methods in robust intensity-modulated proton therapy (IMPT) or adaptive proton therapy point weight generation. The only similar study was a 2022 study that investigated the feasibility of using machine learning to fine-tune point weights under the assumption that the dose-effect matrix of existing single-field optimized IMPT plans remains unchanged. However, this method only meets the need to adjust the dose according to existing plans for head and neck cancer patients. Therefore, robust IMPT and adaptive proton therapy require more automated and time-efficient plan optimization and point weight generation algorithms.

[0005] In summary, existing AI and deep learning-based planning optimization algorithms are unable to adaptively generate proton scan lattice weights to accommodate arbitrary changes in CT scan images between patient sessions, which limits planning optimization efficiency and restricts the application of related optimization methods in clinical settings such as adaptive therapy. Summary of the Invention

[0006] To address the shortcomings of related technologies, the present invention aims to provide a method for constructing a weight prediction model for a proton uniform scanning field lattice and a weight prediction method. This method is intended to solve the problem that existing planning optimization algorithms based on artificial intelligence and deep learning cannot generate adaptive lattice weights to cope with arbitrary changes between patient sessions, thus limiting the efficiency of planning optimization and its clinical application.

[0007] To achieve the above objectives, this invention provides a method for establishing a weighted prediction model for a proton uniform scanning field lattice, comprising:

[0008] Construct a training dataset; in the training dataset, each training data example includes a three-dimensional CT image matrix with uniform dimensions, a dose matrix of a single-field proton robust plan, a position mask matrix of the plan lattice, and a reference dose matrix, with the corresponding label being a weight matrix; the weight matrix is ​​the true weight corresponding to each element in the plan lattice;

[0009] The dose matrix of the single-field proton robust plan is the dose value applied to each irradiation path in the single-field proton robust plan formulated based on the three-dimensional CT image matrix; the plan matrix is ​​obtained based on the position index of all irradiation points in the single-field proton robust plan; the reference dose matrix is ​​obtained by setting the weight corresponding to each element in the plan matrix to a preset value;

[0010] A three-dimensional attention mechanism Unet network is established, and the three-dimensional attention mechanism Unet network is trained using the training dataset. After training, a proton uniform scanning field lattice weight prediction model is obtained.

[0011] The training of the Unet network, a three-dimensional attention mechanism, using the training dataset includes:

[0012] T1. For each training data example, the three-dimensional CT image matrix, the dose matrix of the single-field proton robust plan, the position mask matrix of the plan lattice, and the reference dose matrix are normalized and then stacked in the zeroth dimension.

[0013] T2. Input the stacking results as input data into the three-dimensional attention mechanism Unet network to obtain the prediction lattice weights of the single-field proton robust plan dose;

[0014] T3. By minimizing the difference loss between the predicted lattice weights and the corresponding labels, the three-dimensional attention mechanism Unet network is trained to obtain the proton uniform scanning field lattice weight prediction model.

[0015] Optionally, the difference loss between the predicted dot matrix weight matrix and the corresponding label is: the weighted sum of the errors between each element in the predicted dot matrix weight matrix and the corresponding element in the corresponding label, so as to enhance the model's learning of low weight values.

[0016] Optionally, the three-dimensional CT image matrix is ​​obtained by processing the original CT scan image;

[0017] The original CT scan image includes the air region in front of the human body structure in the direction of proton beam irradiation.

[0018] Optionally, all data samples in the training dataset have a uniform dimension on the cross section perpendicular to the proton beam irradiation direction, and the dimension along the proton beam irradiation direction is a preset value.

[0019] Optionally, constructing the training dataset includes:

[0020] Step S1: Import the patient's original CT scan images and the corresponding structures of the radiotherapy target area and critical organs delineated by the doctor into the matRad platform to obtain a three-dimensional CT image matrix. Generate an initial single-field proton robust plan for each patient based on different preset angles, and correct it to obtain the dose matrix and plan information of the single-field proton robust plan.

[0021] Step S3: Determine the dot matrix position of the irradiation point according to the beam position and energy information in the planning information, and make a mask matrix based on the dot matrix position, setting the dot matrix position to 1 and the other positions to 0;

[0022] Step S4: Set all weights of the lattice positions to preset values, and calculate the reference dose matrix of the single-field proton robust plan dose using matrad.

[0023] Step S5: Set all weights of the dot matrix positions to the planned weights in the planned information to obtain the planned weight matrix;

[0024] Step S6: Obtain robust planning data for the patient at different preset angles. The robust planning data includes a 3D CT image matrix, a dose matrix for a single-field proton robust plan, a dot matrix mask matrix, a reference dose matrix, and a plan weight matrix. Rotate all data in the matrices counterclockwise by a preset angle so that they are 0° relative to the proton beam incident direction. Determine the boundary position of the dose range based on the rotated dose matrix of the single-field proton robust plan. Crop the robust planning data by uniformly expanding it outward to a uniform dimension based on the boundary position of the dose range to obtain a training dataset.

[0025] Optionally, the original CT scan images may include CT images of different parts of the human body.

[0026] Secondly, the present invention also provides a weight prediction method for a proton uniform scanning field lattice, comprising:

[0027] The patient's CT scan images and data of the area to be treated are obtained and input into the matRad platform to obtain the ideal dose matrix, the ideal dot matrix mask matrix, and the ideal reference dose matrix.

[0028] The ideal dose matrix, ideal lattice mask matrix, and ideal reference dose matrix are normalized and then stacked in the zeroth dimension. The stacking result is then used as input data into the proton uniform scanning field lattice weight prediction model to obtain the ideal weight matrix that achieves the ideal dose matrix.

[0029] The ideal weight matrix is ​​restored according to the dimensions and angles of the original CT scan image to obtain the restored weight matrix. According to the arrangement order of proton beam irradiation points along the incident direction from near to far in the single-field proton robust plan, the corresponding weight values ​​are extracted from the restored weight matrix to obtain a one-dimensional weight array, which is used as the target prediction weight of the proton uniform scanning field array in the single-field proton robust plan.

[0030] The weight prediction model of the proton uniform scanning field array is obtained by the method for establishing the weight prediction model of the proton uniform scanning field array as described in any one of the first aspects.

[0031] Thirdly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a method for constructing a weight prediction model of a proton uniform scanning field lattice as described in any of the first aspects, and / or a weight prediction method of a proton uniform scanning field lattice as described in the second aspect.

[0032] Fourthly, the present invention also provides a computer-readable storage medium, including a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to perform the proton uniform scanning field matrix weight prediction model establishment method as described in any one of the first aspects, and / or the proton uniform scanning field matrix weight generation method as described in the second aspect.

[0033] Compared with the prior art, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0034] 1. This invention provides a method for establishing a weight prediction model for a proton uniformly scanned firing field lattice. Based on a three-dimensional attention mechanism Unet network, a weight prediction model for a proton uniformly scanned firing field lattice is established. This network can effectively process 3D matrix information and integrates an attention gate mechanism to further help the network fuse shallow and deep features, effectively suppressing irrelevant background responses in skip connections, while automatically focusing on significant feature regions. Only a minimal increase in additional model parameters is needed to significantly improve model accuracy. During model training, the proton robust planned dose, lattice position mask, and reference dose are normalized and stacked in the zeroth dimension, then used as input data into the 3D Attention Unet network to assist the model in extracting multiple effective information components to finally obtain the predicted weight matrix. The sub-robust planned firing field lattice weight prediction model established by this invention has high computational accuracy, high computational efficiency, and high universality.

[0035] 2. This invention provides a method for establishing a weighted prediction model for a proton uniform scanning field array. The training dataset has a unified matrix range covering all planned dose ranges and retains a portion of the air region before the radiation reaches the human body structure. This ensures that it can handle the vast majority of patient samples during testing and practical applications, and fully captures the characteristic relationship between planned dose and weights. The original CT and structural data of patients used to construct the training dataset include different human body parts, and the planned dose is a robust single-field optimized plan, further improving its universality and applicability. For each patient, a single-field optimized plan is created from multiple perspectives, and corresponding data is obtained. This expands the dataset, strengthens data learning, and is compatible with future practical applications.

[0036] 3. This invention provides a method for establishing a weighted prediction model for a proton uniform scanning field lattice. In addition to the patient's expected dose, the model's input data includes two auxiliary data sets: a lattice position matrix obtained based on the patient's structural information, and a reference dose calculated by setting all lattice weights to 1. The lattice position matrix helps the network suppress irrelevant background regions and improves the learning of lattice weight data. The combination of the reference dose and the expected dose helps the network extract the corresponding features of dose and weight to improve the calculation. Furthermore, the model only requires the patient's existing expected dose and structural information, without needing any prior planning information, to obtain the predicted weights corresponding to the expected dose. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a method for constructing a weighted prediction model for a proton uniform scanning field lattice provided by the present invention.

[0038] Figure 2This is a schematic diagram of the weight prediction model for a proton uniform scanning field lattice based on the 3D Attention Unet network. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0041] Example 1

[0042] like Figure 1 As shown, a method for establishing a weighted prediction model for a proton uniform scanning field lattice includes:

[0043] The dose matrix of the single-field proton robust plan is the dose value applied to each irradiation path in the single-field proton robust plan formulated based on the three-dimensional CT image matrix; the plan matrix is ​​obtained based on the position index of all irradiation points in the single-field proton robust plan; the reference dose matrix is ​​obtained by setting the weight corresponding to each element in the plan matrix to a preset value;

[0044] A three-dimensional attention mechanism Unet network is established, and the three-dimensional attention mechanism Unet network is trained using the training dataset. After training, a proton uniform scanning field lattice weight prediction model is obtained.

[0045] The training of the Unet network, a three-dimensional attention mechanism, using the training dataset includes:

[0046] T1. For each training data example, the three-dimensional CT image matrix, the dose matrix of the single-field proton robust plan, the position mask matrix of the plan lattice, and the reference dose matrix are normalized and then stacked in the zeroth dimension.

[0047] T2. Input the stacking results as input data into the three-dimensional attention mechanism Unet network to obtain the prediction lattice weights of the single-field proton robust plan dose;

[0048] T3. By minimizing the difference loss between the predicted lattice weights and the corresponding labels, the three-dimensional attention mechanism Unet network is trained to obtain the proton uniform scanning field lattice weight prediction model.

[0049] In this embodiment, the reference dose is calculated by setting all weights of the planned matrix to 1; the three-dimensional CT image matrix is ​​obtained by processing the original CT scan image; wherein, the original CT scan image includes the air region in the direction of proton beam irradiation, before the proton beam incident on the human body structure. All data samples in the training dataset have a uniform dimension; the dimension of all data samples on the cross section perpendicular to the proton beam irradiation direction is uniformly 256×256, and the dimension along the proton beam irradiation direction is 384mm.

[0050] The difference loss between the predicted dot matrix weight matrix and the corresponding label is the weighted sum of the errors between each element in the predicted dot matrix weight matrix and the corresponding element in the corresponding label, to enhance the model's learning of low weight values. Positions with values ​​less than or equal to 50 in the label data are assigned a weight of 5, and positions with values ​​greater than 50 are assigned a weight of 2.

[0051] In this embodiment, the robust planned dose, the position mask of the planned dot matrix, and the reference dose in the training dataset are normalized and stacked in the zeroth dimension. Then, the whole dataset is used as input data in the model training process and input into the 3D Attention Unet network. This allows the 3D Attention Unet network to back-calculate the dot matrix weights of the proton robust planned dose. The planned dot matrix weights in the training data are used as label data to supervise and optimize the generation results of the 3D Attention Unet network.

[0052] To enable the model to handle the vast majority of patient samples in practical applications and to fully capture the characteristic relationship between planned dose and weights, the dimensions of the matrices included in the training dataset were standardized when constructing the training dataset. For example, the dimensions in this embodiment are (192, 128, 128). The dataset with this dimension can cover the range of all planned doses and retain part of the air region before it is incident on the human body structure.

[0053] Specifically, constructing the training dataset includes:

[0054] Step S1: Import the patient's original CT scan images and the corresponding structures of the radiotherapy target area and critical organs delineated by the doctor into the matRad platform to obtain a three-dimensional CT image matrix. Generate an initial single-field proton robust plan for each patient based on different preset angles, and correct it to obtain the dose matrix and plan information of the single-field proton robust plan.

[0055] Specifically, the patient's original CT scan images and corresponding structural files are imported into the matRad platform, and a preset linear interpolation method is used to unify the resolution of CT images containing human structures to a preset resolution, constructing an original three-dimensional CT image matrix. Consistent data resolution helps the model extract and learn data features. Optionally, in this embodiment, the preset resolution is specifically 2mm×2mm×2mm. According to this resolution, the CT images can be divided into voxels of size 2mm×2mm×2mm to obtain the original three-dimensional CT image matrix.

[0056] In this embodiment, the selected patient images include different body parts, and a robust single-field optimization plan is created for each patient from multiple preset angles, expanding the dataset while aligning with future practical applications. The preset angles are selected based on the patient's tumor location and condition, choosing appropriate irradiation angles. The patient and angle can be switched and set according to actual application needs. Robustness is considered for range uncertainty ±3.5% and positioning uncertainty ±3 mm.

[0057] Step S2: Determine the dot matrix position of the irradiation point based on the beam position and energy information in the planning information, and create a mask matrix based on the dot matrix position, setting the dot matrix position to 1 and the other positions to 0.

[0058] Specifically, based on the planning information obtained in step S1, the energy of all beams at each beam position is converted to the corresponding equivalent water depth, and compared with the equivalent water depth matrix calculated based on the CT image matrix at that position. The specific location indices of all irradiation points in the plan are then used to create a planning dot matrix, which serves as the dot matrix location information for model learning. Optionally, in this embodiment, voxels at the dot matrix positions are set to 1, and voxels outside the dot matrix positions are set to 0, with M... spot Let the mask matrix be represented as follows:

[0059]

[0060] Among them, (x j ,y j ,z j () represents the coordinates within the original 3D CT image matrix, and "in spot" indicates the location of the dot matrix.

[0061] Step S3: Set all weights of the lattice positions to preset values, and calculate the reference dose matrix of the single-field proton robust plan dose using matrad.

[0062] Specifically, based on the planning information obtained in step S1, the corresponding reference dose is obtained by recalculating the dose by setting the weight values ​​of all point positions to a uniform value. This reference dose can be regarded as a combined dose after normalizing all pen weights to assist the network in comparative learning. Optionally, in this embodiment, the corresponding reference dose matrix is ​​calculated by setting all weight values ​​to 1.

[0063] Step S4: Set all weights of the dot matrix positions to the planned weights in the planned information to obtain the planned weight matrix.

[0064] Specifically, based on the dot mask position matrix obtained in step S2 and the pen beam order in the planning information, the weight value corresponding to each dot position is found and assigned to the voxel to obtain the planning weight matrix corresponding to the robust dose of the proton single-field optimization for the patient. This matrix is ​​the result of the model's label data supervised optimization model.

[0065] Step S5: Obtain robust planning data for the patient at different preset angles. The robust planning data includes a 3D CT image matrix, a dose matrix for a single-field proton robust plan, a dot matrix mask matrix, a reference dose matrix, and a planning weight matrix. Rotate all data in the matrices counterclockwise by a preset angle so that they are 0° relative to the proton beam incident direction. Determine the boundary position of the dose range based on the rotated dose matrix of the single-field proton robust plan. Crop the robust planning data by uniformly expanding it outward to a uniform dimension based on the boundary position of the dose range to obtain a training dataset.

[0066] The dose distribution in the data sample is cropped and centered; the cropped data dimension covers the range of dose diffusion.

[0067] Specifically, the robust planning data of patients at preset angles obtained from steps S1-S4 are all rotated to 0° of the incident direction to ensure consistency of data at all angles and enhance model learning. After rotation, the matrices of all data samples are uniformly cropped to achieve a unified dimensional size and complete the dataset production. The dose distribution is centered in the cropped data samples, and the cropped data dimensions cover the dose diffusion range.

[0068] In this embodiment, the uniform dimension size takes into account three factors: First, it must include the three-dimensional dose distribution range of all data samples to ensure that the model can fully learn the corresponding features of dose and weight; second, it must meet the dimension requirements of the skip connection after four downsamplings of the network; and finally, on this basis, the dimension should be as small as possible to reduce the proportion of invalid background information.

[0069] Specifically, in this embodiment, the dimensions are determined to be (192, 64, 64), meaning the cross-sectional dimension perpendicular to the irradiation direction is 128mm × 128mm, and the depth along the irradiation direction is 384mm. The dimensions can be adjusted according to patient type and actual needs. During trimming, the maximum boundary position around the planned dose is first found in the cross-section perpendicular to the irradiation direction. Then, the trimming dimension is expanded uniformly outward from the maximum boundary position to ensure the effective dose distribution area is centered in the matrix, thus ensuring the uniformity of all data and assisting the network in focusing on a unified area. It should be noted that the trimming starting point can be selected along the irradiation direction, but it must be ensured that the air region at the front end of the irradiated body structure is included. This helps the network learn the dose at the irradiation front end and the air conditions; therefore, it is not recommended to start trimming directly from the location of the incident voxel.

[0070] In one specific embodiment, training data of 240 cases is extracted from medical images and radiotherapy plans for each patient site using the above method. This training dataset will be used to train the established 3DA attention Unet network to obtain a proton robust plan weight generation model. The number of data samples can be increased or decreased, but it is not recommended to have less than 180 training data samples. The number of data samples also needs to be allocated to a validation set; too few data samples will affect the model's learning performance.

[0071] The 3D Attention Unet network used in this embodiment integrates a soft attention mechanism into the network framework based on the 3D Unet network. Attention weights are derived through gradient calculation and feedback learning, helping the network automatically focus on significant regional features while effectively suppressing irrelevant features. This integration is not only simple and direct, requiring only minimal additional model parameters, but also significantly improves model accuracy.

[0072] like Figure 2As shown, each input information enters the encoder structure of the model for feature extraction. In this embodiment, four downsampling operations are performed in the encoder, each consisting of a double convolution and a max-pooling layer with a kernel size of 3, a stride of 1, and the ReLU function as the activation function. After encoding, the feature map is then transmitted to the decoder structure, involving three operations: upsampling, skip connections, and attention gates. Upsampling uses a transposed convolution with a stride of 2 to halve the number of channels in the high-level feature map and double its dimension. Skip connections integrate the low-level feature map in the encoder with the upsampled feature map in the decoder, enabling the model to learn combinations of low-level and high-level features. The attention gate mechanism calculates attention coefficients for the shallow features of the encoder and the deep features of the decoder at the corresponding downsampling levels before skip connections, further enhancing the fusion of low-level and high-level features. This gating operation suppresses redundant shallow features of the encoder layer by layer while automatically focusing on salient feature regions.

[0073] To ensure the 3D Attention Unet network accurately learns the correspondence between planned dose and lattice weights, this embodiment trains the 3D Attention Unet network using a training dataset, including:

[0074] For each training data example, the CT scan image, proton robust planned dose matrix, dot matrix position mask matrix, and reference dose matrix are normalized and then stacked in the zeroth dimension to obtain the stacked result. The input data of the stacked result is fed into the 3D Attention Unet network, so that the network calculates the (192, 64, 64) dot matrix weight matrix corresponding to the planned dose matrix. The planned weight matrix in the training dataset is used as the label data to supervise the calculation results of the 3D Attention Unet network.

[0075] After obtaining the weight matrix output by the 3D Attention UNet network, the loss function can be calculated and its partial derivatives obtained based on the model's computational results and label data to obtain the model's update gradient parameters. Optionally, in this embodiment, the loss function is chosen as the Masked Mean Square Error (Masked MSE), calculated as follows:

[0076]

[0077] Where M i,j,k This indicates the position of a weight value in the label weight matrix, i.e., the position of the dot matrix. W represents the label weight. ^ This represents the lattice weights predicted by the model.

[0078] The parameters are updated according to the gradient descent direction of the updated gradient parameters until the value of the loss function reaches the minimum and the preset maximum number of training times, thus obtaining the proton uniform scanning field matrix weight prediction model.

[0079] In this embodiment of the invention, a weight prediction model for a proton uniform scanning field matrix is ​​established based on a 3D attention mechanism Unet network. The acquired patient medical images and structural information are converted into a robust proton dose for a single field. The planned field matrix and reference dose are used as input data in the training set, and the planned weights corresponding to the planned field matrix are used as label data. The training dataset is input into the proton uniform scanning field matrix weight prediction model built on the 3D Attention Unet network structure for training. The Adam optimization algorithm is used for 400 training rounds to obtain an optimized proton uniform scanning field matrix weight prediction model. The proton uniform scanning field matrix weight prediction model established in this embodiment has high universality and fast computational efficiency, greatly reducing the time and manpower required for planned weight optimization.

[0080] Example 2

[0081] A method for generating weights for a proton uniform scanning field matrix includes:

[0082] The patient's CT scan images and data of the area to be treated are obtained and input into the matRad platform to obtain the ideal dose matrix, the ideal dot matrix mask matrix, and the ideal reference dose matrix.

[0083] The ideal dose matrix, ideal lattice mask matrix, and ideal reference dose matrix are normalized and then stacked in the zeroth dimension. The stacking result is then used as input data into the proton uniform scanning field lattice weight prediction model to obtain the ideal weight matrix that achieves the ideal dose matrix.

[0084] The ideal weight matrix is ​​restored according to the dimensions and angles of the original CT scan image to obtain the restored weight matrix. According to the arrangement order of proton beam irradiation points along the incident direction from near to far in the single-field proton robust plan, the corresponding weight values ​​are extracted from the restored weight matrix to obtain a one-dimensional weight array, which is used as the target prediction weight of the proton uniform scanning field array in the single-field proton robust plan.

[0085] The weight prediction model of the proton uniform scanning field array is established by the method for establishing the weight prediction model of the proton uniform scanning field array provided in Example 1 above.

[0086] Example 3

[0087] A computer program product includes a computer program that, when executed by a processor, implements the method for establishing a weight prediction model for a proton uniform scanning field lattice provided in Embodiment 1 above, and / or the method for generating weights for a proton uniform scanning field lattice provided in Embodiment 2 above.

[0088] Example 4

[0089] A computer-readable storage medium includes a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to execute the proton uniform scanning field matrix weight prediction model establishment method provided in Embodiment 1 above, and / or the proton uniform scanning field matrix weight generation method provided in Embodiment 2 above.

[0090] Those skilled in the art will readily understand that the above description is merely 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 within the scope of protection of the present invention.

Claims

1. A method for establishing a weighted prediction model for a proton uniform scanning field lattice, characterized in that, include: Build the training dataset; In the training dataset, each training data example includes a dimensionally uniform three-dimensional CT image matrix, a dose matrix of a single-field proton robust plan, a position mask matrix of the plan dot matrix, and a reference dose matrix, with the corresponding label being a weight matrix; the weight matrix is ​​the true weight corresponding to each element in the plan dot matrix; The dose matrix of the single-field proton robust plan is the dose value applied in each irradiation path in the single-field proton robust plan formulated based on the three-dimensional CT image matrix; The planned lattice matrix is ​​obtained based on the position index of all irradiation points in the single-field proton robust plan; The reference dose matrix is ​​obtained by setting the weight corresponding to each element in the planning matrix to a preset value; A three-dimensional attention mechanism Unet network is established, and the three-dimensional attention mechanism Unet network is trained using the training dataset. After training, a proton uniform scanning field lattice weight prediction model is obtained. The training of the Unet network, a three-dimensional attention mechanism, using the training dataset includes: T1. For each training data example, the three-dimensional CT image matrix, the dose matrix of the single-field proton robust plan, the position mask matrix of the plan lattice, and the reference dose matrix are normalized and then stacked in the zeroth dimension. T2. Input the stacking results as input data into the three-dimensional attention mechanism Unet network to obtain the prediction lattice weights of the single-field proton robust plan dose; T3. By minimizing the difference loss between the predicted lattice weights and the corresponding labels, the three-dimensional attention mechanism Unet network is trained to obtain the proton uniform scanning field lattice weight prediction model.

2. The method as described in claim 1, characterized in that, The difference loss between the predicted dot matrix weight matrix and the corresponding label is the weighted sum of the errors between each element in the predicted dot matrix weight matrix and the corresponding element in the corresponding label, in order to enhance the model's learning of low weight values.

3. The method as described in claim 1, characterized in that, The three-dimensional CT image matrix is ​​obtained by importing the original CT scan images into the matRad platform for processing. The original CT scan image includes the air region in front of the human body structure in the direction of proton beam irradiation.

4. The method as described in claim 3, characterized in that, The dimensions of all data samples in the training dataset are uniform on the cross-section perpendicular to the proton beam irradiation direction, and the dimension along the proton beam irradiation direction is a preset value.

5. The method as described in claim 1, characterized in that, The construction of the training dataset includes: Step S1: Import the patient's original CT scan images and the corresponding structures of the radiotherapy target area and critical organs delineated by the doctor into the matRad platform to obtain a three-dimensional CT image matrix. Generate an initial single-field proton robust plan for each patient based on different preset angles, and correct it to obtain the dose matrix and plan information of the single-field proton robust plan. Step S2: Determine the dot matrix position of the irradiation point according to the beam position and energy information in the planning information, and make a mask matrix based on the dot matrix position, setting the dot matrix position to 1 and the other positions to 0; Step S3: Set all weights of the lattice positions to preset values, and calculate the reference dose matrix of the single-field proton robust plan dose using matrad. Step S4: Set all weights of the dot matrix positions to the planned weights in the planned information to obtain the planned weight matrix; Step S5: Obtain robust planning data for the patient at different preset angles. The robust planning data includes a 3D CT image matrix, a dose matrix for a single-field proton robust plan, a dot matrix mask matrix, a reference dose matrix, and a plan weight matrix. Rotate all data in the matrix counterclockwise by a preset angle so that it is 0° relative to the proton beam incident direction. Determine the boundary position of the dose range based on the rotated dose matrix of the single-field proton robust plan. Crop the robust planning data by uniformly expanding it outward to a uniform dimension based on the boundary position of the dose range to obtain a training dataset.

6. The method as described in claim 1, characterized in that, The original CT scan images contain CT images of different parts of the human body.

7. A weight prediction method for a proton uniform scanning field lattice, characterized in that, include: The patient's CT scan images and data of the area to be treated are obtained and input into the matRad platform to obtain the ideal dose matrix, the ideal dot matrix mask matrix, and the ideal reference dose matrix. The ideal dose matrix, ideal lattice mask matrix, and ideal reference dose matrix are normalized and then stacked in the zeroth dimension. The stacking result is then used as input data into the proton uniform scanning field lattice weight prediction model to obtain the ideal weight matrix that achieves the ideal dose matrix. The ideal weight matrix is ​​restored according to the dimensions and angles of the original CT scan image to obtain the restored weight matrix. According to the arrangement order of proton beam irradiation points along the incident direction from near to far in the single-field proton robust plan, the corresponding weight values ​​are extracted from the restored weight matrix to obtain a one-dimensional weight array, which is used as the target prediction weight of the proton uniform scanning field array in the single-field proton robust plan. The weight prediction model of the proton uniform scanning field array is established by the method for establishing the weight prediction model of the proton uniform scanning field array as described in any one of claims 1 to 6.

8. A computer program product, characterized in that, The system includes a computer program that, when executed by a processor, implements the method for constructing a weight prediction model for a proton uniform scanning field lattice as described in any one of claims 1 to 6, and / or the method for weight prediction of a proton uniform scanning field lattice as described in claim 7.

9. A computer-readable storage medium, characterized in that, Includes a stored computer program; when the computer program is executed by a processor, it controls the device containing the computer-readable storage medium to perform the proton uniform scanning field lattice weight prediction model establishment method as described in any one of claims 1 to 6, and / or the proton uniform scanning field lattice weight generation method as described in claim 7.

Images