A method for predicting dose of brachytherapy for glioma

Abstract

The present invention discloses a method for predicting the therapeutic dose of brachytherapy for gliomas, which relates to the technical field of therapeutic dose prediction. The method comprises the following steps: acquiring patient data and extracting key patient data, constructing a network CDC Transformer model, training the network CDC Transformer model, and outputting a therapeutic dose prediction result through the network CDC Transformer model. The present invention realizes automatic prediction of the therapeutic dose of brachytherapy for gliomas by constructing a network CDC Transformer model based on the Transformer model. Channel spatial attention is added before the downsampling process of the original Transformer model to automatically allocate channel weights and enhance the prediction result. The method does not rely on the personal experience of doctors and physicists, and can more accurately predict the dose distribution.

Description

Technical Field

[0001] The present invention relates to the technical field of therapeutic dose prediction, and in particular to a method for predicting the therapeutic dose of a brain glioma brachytherapy. Background Art

[0002] Brachytherapy refers to the implantation of a radioactive source near a tumor in the patient's body, using the radioactive source to irradiate the tumor, thereby killing cancer cells. Brachytherapy is essentially a form of precision radiotherapy. Radioactive particles are implanted near the tumor through an implantation gun and an implantation needle, which can effectively kill the tumor and has a lower impact on normal tissue. However, there are also quite a few problems. For example, the preoperative plan and the operation during the operation cannot be fully matched, and deviations are prone to occur. In addition, the formulation of the surgical plan is greatly related to the doctor's clinical experience and subjective factors. The formulation efficiency is low, and the results may not be the optimal solution. The position and dose calculation of radioactive particles are also the key to the success of particle implantation interventional surgery. Traditional methods make it difficult to make the dose distribution meet clinical requirements, and dose radiation cold and hot zones are prone to occur. Currently, there are few dose prediction methods for brachytherapy.

[0003] In brachytherapy, the radiation dose distribution is usually closely related to the particle position, number, and activity. The dose distribution decays dramatically with the distance from the particle. Therefore, after planning the dose distribution, it is also difficult to determine the particle position. Currently, there are not many automated prediction methods for brachytherapy dose distribution. Most plans are still based on the personal experience of doctors and physicists. The calculations to determine the particle position are also very complex. There are basically no deep learning methods for brachytherapy. Therefore, the present invention proposes a dose prediction method for brain glioma brachytherapy to solve the problems existing in the prior art. Summary of the Invention

[0004] In response to the above problems, the purpose of the present invention is to propose a method for predicting the dose distribution of brachytherapy for brain gliomas, so as to solve the problem that there are not many automated prediction methods for the dose distribution of brachytherapy in the prior art. Most of them are still based on the personal experience of doctors and physicists for planning, and the calculation to determine the particle position is also very complicated. There is basically no deep learning method for brachytherapy.

[0005] To achieve the purpose of the present invention, the present invention is implemented by the following technical solution: a method for predicting the dose of brain glioma brachytherapy, comprising the following steps:

[0006] Step 1: Collect medical data of glioma patients from the hospital's medical record system, pre-process the collected medical data, and then extract the patient's key data;

[0007] Step 2: Add channel spatial attention before the downsampling process of the Transformer model to build a network CDC Transformer model based on the Transformer model;

[0008] Step 3: Using the key data of different glioma patients extracted in Step 1 as the dataset, the network CDC Transformer model constructed in Step 2 is trained;

[0009] Step 4: The patient's CT image, dose information, and the doctor's outline of the brain contour and PTV area are used as input to the network CDC Transformer model. The network CDC Transformer model outputs the patient's 3D dose image to predict the dose of brachytherapy for gliomas.

[0010] Further improvements are as follows: in step 1, the specific steps for preprocessing medical data are: first, remove erroneous records in the data set by using data cleaning, then convert each data into a unified format by using data standardization, and then expand the data set by using data enhancement technology.

[0011] Further improvement is that in step 1, the specific steps of extracting the patient's key data are: first determine the key fields of the key data, then extract the determined key fields from the preprocessed data set, and then integrate the extracted data into a structured format to obtain key data including CT images, ptv, particle position, dose and external.

[0012] A further improvement is that in step 2, before the downsampling process, the CBAM module is first used to perform channel attention processing on the model to enhance the model's attention to important features, and then the CBAM module is used to perform spatial attention processing on the model.

[0013] A further improvement is that in step 2, during the downsampling process of the Transformer model, a distance attention module is used to perform distance convolution calculation, and a vector aggregation module is added to the Transformer model.

[0014] Further improvements are as follows: in step 3, when training the network CDCTransformer model, the model training batch size is set to 2, the initial learning rate is set to 3e-6, the Adam optimizer is used, the cosine annealing algorithm is used to ensure the training results, and the total number of training iterations is set to 100,000.

[0015] A further improvement is that in step three, the dose of each voxel point is extracted as input during the training of the network CDC Transformer model, so that the network CDC Transformer model can learn the relationship between the dose of each voxel and the particle position.

[0016] A further improvement is that in step 3, the loss function during the training of the network CDCTransformer model is the MSE function, which is expressed by the following formula:

[0017]

[0018] Where i represents each voxel, n represents the total number of voxels, and P i Represents the predicted value of each voxel, G i Indicates the true dose value of each voxel;

[0019] The model training evaluation indicators use the DICE coefficient and MSE function. The DICE coefficient is expressed as follows:

[0020]

[0021] Among them, X is the actual dose result, and Y is the predicted dose result.

[0022] The beneficial effects of the present invention are as follows: the present invention realizes the automatic prediction of the dose of brain glioma brachytherapy by constructing a network CDCTransformer model based on the Transformer model, and adds channel spatial attention before the downsampling process of the original Transformer model to automatically allocate channel weights and enhance the prediction results. It does not need to rely on the personal experience of doctors and physicists, and can more accurately predict the dose distribution, help doctors analyze the dose prediction results more intuitively, and assist doctors in formulating radiotherapy plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 1 is a flow chart of a method for predicting the dose of brachytherapy for gliomas according to the present invention;

[0025] Figure 2 is a visualization diagram of the extracted particle positions in an embodiment of the present invention;

[0026] Figure 3 Schematic diagram of the channel space attention module structure in an embodiment of the present invention;

[0027] Figure 4 Schematic diagram of the distance attention module structure in an embodiment of the present invention;

[0028] Figure 5 is a schematic diagram of the distance calculation process in an embodiment of the present invention;

[0029] Figure 6 Schematic diagram of the structure of a vector aggregation module in an embodiment of the present invention;

[0030] Figure 7 is a comparison chart of the actual and predicted doses for Patient I in an embodiment of the present invention;

[0031] Figure 8 This is a comparison chart of the actual and predicted doses for Patient II in an embodiment of the present invention;

[0032] Figure 9 is a DVH map of patient I in an embodiment of the present invention;

[0033] Figure 10 is the DVH map of patient II in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0035] See also Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 、 Figure 9 、 Figure 10 This embodiment provides a method for predicting the dose of brachytherapy for glioma, comprising the following steps:

[0036] Step 1: Obtain patient data and extract key patient data

[0037] Medical data from 50 patients with glioma were collected from the hospital's medical record system. The collected medical data was preprocessed by first removing erroneous records from the dataset using data cleaning, then converting the data into a unified format using data standardization to facilitate subsequent analysis, and then using data augmentation techniques to expand the dataset to improve the generalization ability of subsequent models.

[0038] The key patient data is then extracted from the preprocessed dataset. The specific steps are as follows: first, determine the key fields of the key data, then extract the determined key fields from the preprocessed dataset, and then integrate the extracted data into a structured format, such as a CSV file or a database table, to obtain key data including CT images, PTV (planned target volume), particle position, dose (dose distribution), and external (external information);

[0039] Given the high impact of particle number and position on dose distribution, this embodiment extracts particle position and number information from the patient's medical data as a channel input. The extracted particle position visualization diagram is shown in Figure 2. Figure 2 ;

[0040] Step 2: Build a network CDC Transformer model

[0041] Before the downsampling process of the Transformer model, channel spatial attention is added to build a network CDC (Change Detection and Classification) Transformer model based on the Transformer model to automatically assign channel weights and enhance the prediction results. The channel spatial attention module consists of a channel attention module and a spatial attention module. For its structure diagram, see Figure 3 The channel attention module of the channel spatial attention module can assign weights by itself and assign larger training weights to important channels. In this way, the model can determine the important input channels by itself and adjust the weights to improve the training effect. The spatial attention module can better capture spatial information. Due to the complex structure of the brain, improving the ability to capture spatial information can make the model learn better.

[0042] Before downsampling, this embodiment uses the CBAM module (Convolutional Block Attention Module) to perform channel attention processing on the model to enhance the model's focus on important features. Then, the CBAM module is used to perform spatial attention processing on the model, which helps improve the performance of the model.

[0043] In this embodiment, during the downsampling process of the Transformer model, a distance attention module is used to perform distance convolution calculation (i.e., each value on the original voxel is multiplied by an inverse square coefficient to transform the relationship between the distance between each voxel and the particle voxel). A vector aggregation module is added to the model to integrate different features into a feature vector to represent the probability of the existence of the feature and the spatial structure. The distance attention module integrates a formula for the inverse distance to assist in the input of particle positions. This module only integrates one inverse relationship, and the remaining parameters are learned by the model during training, so that the model can accurately learn distance-related dose distribution information.

[0044] The distance attention module structure can be found in Figure 4 , where the distance calculate process is as follows Figure 5 That is, since the dose distribution of particle radiation is inversely proportional to the distance between particles, this embodiment embeds a formula for the inverse relationship. The obtained output feature is then convolved with the original distance_map to obtain the distance feature, which is then combined with the features of the previous part of the model to form a new feature and input into the next layer of the network. That is, the initial dose is a dose value at each voxel point, which is converted into a new voxel value through the inverse formula calculation, and then other steps are performed. In this way, the model can learn the relationship between the particle position and the distance between each position;

[0045] Vector aggregation module structure diagram see Figure 6 The vector aggregation module can list the eigenvalues ​​of each direction of the bottom-level feature map as vectors for calculation. The modulus of the vector represents the probability of the feature's existence, and the angle of the vector represents the spatial structure of the feature. This allows the model to learn spatial features more clearly. The model can not only predict the dose distribution, but also the position of the particles, which can reduce the cost of particle position calculation for doctors.

[0046] Step 3: Train the network CDCTransformer model

[0047] The key data of 50 patients with glioma extracted in step 1 were used as the dataset to train the network CDCTransformer model constructed in step 2. During the training process, the dose of each voxel point was extracted as input for the network CDCTransformer model to learn the relationship between the dose of each voxel and the particle position. When training the prediction model, the parameters were set as follows: the model training batch size was set to 2, the initial learning rate was 3e-6, the Adam optimizer was used, the cosine annealing algorithm was used to ensure the training results, the total number of training iterations was set to 100,000, and the NVIDIA GeForce 3080 TI GPU and PyTorch 1.70 framework were used for training;

[0048] In this embodiment, the loss function during the training of the network CDCTransformer model is the MSE function, which is expressed by the following formula:

[0049]

[0050] Where i represents each voxel, n represents the total number of voxels, and P i Represents the predicted value of each voxel, G i Indicates the true dose value of each voxel;

[0051] The model training evaluation indicators use the DICE coefficient and MSE function. The DICE coefficient is expressed as follows:

[0052]

[0053] Among them, X is the actual dose result, and Y is the predicted dose result

[0054] Step 4: Output the treatment dose prediction results through the network CDCTransformer model

[0055] The network CDC Transformer model has three input channels. In this embodiment, the model uses the patient's CT image, dose information, and the doctor's outline of the brain contour and PTV area as inputs. The model outputs the patient's 3D dose image to predict the dose for brachytherapy of gliomas. The 3D dose image is then superimposed on the original CT image to obtain a dose distribution map, which helps doctors analyze the dose prediction results more intuitively.

[0056] The output of the network CDC Transformer model in this embodiment is a three-dimensional dose distribution. Two patients (patient I and patient II) are randomly selected and the visualized dose distribution results are compared. The comparison diagram of the actual dose and predicted dose of patient I is shown in Figure 7 ( Figure 7 The left side is the actual dose, and the right side is the predicted dose). For the comparison chart of the actual and predicted doses of patient II, see Figure 8 ( Figure 8 The left side is the actual dose, and the right side is the predicted dose). The DVH diagram of patient I is shown in Figure 9 , the DVH diagram of patient II is shown in Figure 10 , where the solid line is the patient's actual DVH curve, and the dotted line is the patient's predicted DVH curve. As can be seen from the figure, the prediction error is very small.

[0057] 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, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the dose of brachytherapy for glioma, characterized in that: The following steps are involved: Step 1: Collect medical data of glioma patients from the hospital's medical record system, pre-process the collected medical data, and then extract the patient's key data; Step 2: Add channel spatial attention before the downsampling process of the Transformer model to build a network CDC Transformer model based on the Transformer model; Before downsampling, the CBAM module is used to perform channel attention processing on the model to enhance the model's attention to important features, and then the CBAM module is used to perform spatial attention processing on the model; During the downsampling process of the Transformer model, a distance attention module is used to perform distance convolution calculations. This involves multiplying each value on the original voxel by an inverse square coefficient to transform the relationship between the distance between each voxel and the particle voxel. A vector aggregation module is added to the Transformer model to integrate different features into a feature vector to represent the probability of the feature's existence and the spatial structure. The distance attention module integrates a formula for the inverse distance to assist in particle position input. This module only integrates an inverse relationship, and the remaining parameters are learned by the model during training, enabling the model to accurately learn distance-related dose distribution information. Step 3: Using the key data of different glioma patients extracted in Step 1 as the dataset, the network CDC Transformer model constructed in Step 2 is trained; Step 4: The patient's CT image, dose information, and the doctor's outline of the brain contour and PTV area are used as input to the network CDCTransformer model. The network CDCTransformer model outputs the patient's 3D dose image to predict the dose of brachytherapy for gliomas.

2. The method for predicting the dose of brachytherapy for glioma according to claim 1, characterized in that: In step 1, the specific steps of preprocessing the medical data are: first, remove erroneous records in the data set by using data cleaning, then convert each data into a unified format by using data standardization, and then expand the data set by using data enhancement technology.

3. The method for predicting the dose of brachytherapy for glioma according to claim 1, characterized in that: In step 1, the specific steps for extracting the patient's key data are: first determine the key fields of the key data, then extract the determined key fields from the preprocessed data set, and then integrate the extracted data into a structured format to obtain key data including CT images, PTV, particle position, dose and external.

4. The method for predicting the dose of brachytherapy for glioma according to claim 1, characterized in that: In step 3, when training the network CDC Transformer model, the model training batch size is set to 2, the initial learning rate is set to 3e-6, the Adam optimizer is used, the cosine annealing algorithm is used to ensure the training results, and the total number of training iterations is set to 100,000.

5. The method for predicting the dose of brachytherapy for glioma according to claim 1, characterized in that: In the step 3, the dose of each voxel point is extracted as input during the training process of the network CDC Transformer model, so that the network CDC Transformer model can learn the relationship between the dose of each voxel and the particle position.

6. The method for predicting the dose of brachytherapy for glioma according to claim 1, characterized in that: In step 3, the loss function during the network CDC Transformer model training process is the MSE function, which is expressed by the following formula: Where i represents each voxel, n represents the total number of voxels, and P i Represents the predicted value of each voxel, G i Indicates the true dose value of each voxel; The model training evaluation indicators use the DICE coefficient and MSE function. The DICE coefficient is expressed as follows: Among them, X is the actual dose result, and Y is the predicted dose result.

Images