Glioma radiation dose prediction system based on attention-enhanced convolutional neural network

Abstract

The invention relates to the technical field of radiotherapy dose prediction, in particular to a glioma radiation dose prediction system based on an attention-enhanced convolutional neural network, which comprises a data collection module, a data preprocessing module, a convolutional neural network module, an attention mechanism module and an output layer. According to the invention, by introducing the attention mechanism, the important feature region can be effectively highlighted, so that the prediction accuracy of the radiation dose of the glioma patient is improved, and the individuation and precision of the treatment scheme are ensured; a dose distribution result can be quickly generated through automatic dose prediction, time and energy investment of doctors in radiotherapy plan design is reduced, working efficiency is improved, accuracy and efficiency of glioma radiotherapy are remarkably improved, clinical decisions are optimized, medical cost is reduced, and support is provided for intelligentization and automation of radiotherapy. Through the research of the project, the application of the deep learning technology in the field of medical image analysis is promoted, and the overall level of radiotherapy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radiotherapy dose prediction, and in particular to a glioma radiation dose prediction system based on an attention enhanced convolutional neural network. Background Art

[0002] Radiotherapy, as an important means of treating malignant tumors such as glioma, has been widely used in recent years. With the continuous advancement of radiation technology, especially the development of intensity-modulated radiation therapy (IMRT) and volumetric modulated radiation therapy (VMAT), the accuracy and effect of radiotherapy have been significantly improved. However, the planning process of radiotherapy still faces many challenges, mainly including the following aspects:

[0003] Insufficient data samples:

[0004] In the automated planning of radiotherapy, deep learning methods usually require a large amount of labeled data for training; however, case data of glioma patients is relatively scarce, resulting in small data samples and high data requirements during model training, which limits the generalization ability of the model.

[0005] Complexity of dose prediction:

[0006] The dose distribution of radiotherapy is affected by many factors, including the patient's anatomical structure, tumor location and size, etc. Traditional dose prediction methods often rely on experience and manual calculations, which are inefficient and prone to errors.

[0007] Barriers to application in clinical practice:

[0008] Although there are many automated plan design methods based on deep learning, in actual clinical applications, due to the poor generalization ability of the model, the effect is not ideal and it is difficult to meet clinical needs;

[0009] Workload of doctors and physiotherapists:

[0010] The design process of radiotherapy plan is usually time-consuming and labor-intensive. Doctors and physicists need to spend a lot of time on target delineation and dose distribution assessment, which affects work efficiency.

[0011] Technology integration requirements:

[0012] With the continuous development of radiotherapy technology, how to effectively integrate advanced artificial intelligence technology with the existing radiotherapy process has become an important issue to improve the efficiency and accuracy of radiotherapy.

[0013] Therefore, to address the above problems, a glioma radiation dose prediction system based on attention enhanced convolutional neural network was proposed. Summary of the invention

[0014] The purpose of the present invention is to provide a glioma radiation dose prediction system based on attention enhanced convolutional neural network to solve the problems raised in the above background technology.

[0015] To achieve the above object, the present invention provides the following technical solutions:

[0016] Glioma radiation dose prediction system based on attention-enhanced convolutional neural network, including:

[0017] Data collection module: used to collect radiotherapy data of glioma patients, including CT imaging data, dose distribution data, clinical information of patients and treatment plan information;

[0018] Data preprocessing module: receives radiotherapy data, extracts CT image data, annotates, cleans and enhances CT image data, and generates training data sets;

[0019] Convolutional neural network module: includes multiple convolutional layers and pooling layers; uses multiple convolutional layers to extract feature maps from input image data; in the convolutional layer, sets convolution kernels of different sizes, including 3×3, 5×5, and 7×7; each convolutional layer is followed by a pooling layer, which reduces the dimension of the feature map through the maximum pooling operation, reduces the amount of calculation, and retains important features at the same time;

[0020] Attention mechanism module: optimizes the feature map and generates a feature map adjusted by self-attention, channel attention and spatial attention. The attention mechanism module includes:

[0021] Self-attention layer: calculates the autocorrelation of the feature map and generates an attention weight matrix to highlight important feature areas and suppress unimportant areas;

[0022] Channel attention layer: Calculate the channel weights through global average pooling and global maximum pooling, adjust the importance of each channel, and use the Sigmoid activation function to generate channel weights;

[0023] Spatial attention layer: Generates spatial weights according to the importance of spatial positions, combines channel and spatial attention information, optimizes feature representation, and uses convolution operations to generate spatial weight maps;

[0024] Output layer: used to generate radiation dose prediction results for glioma patients. The output layer is implemented in the following ways:

[0025] The feature map optimized by the attention mechanism module is input into the fully connected layer for dose prediction;

[0026] The activation function is used to output the predicted dose distribution, and the output format is the dose distribution map and the corresponding DVH data.

[0027] As a preferred solution, the data collection module obtains three-dimensional CT images of glioma patients through high-resolution CT scanning equipment with a resolution of not less than 1mm³; records the radiation dose distribution information actually applied to the patient during radiotherapy, and the radiation dose distribution information includes a dose volume histogram; the patient's clinical information and treatment plan information include the patient's basic information, medical records, and treatment plans.

[0028] As a preferred solution, the data preprocessing module marks the target areas and organs at risk in the CT images, removes incomplete and erroneous data, uses a data cleaning algorithm to detect outliers, and uses image enhancement technology to enhance the CT images. The image enhancement technology includes random rotation, random cropping, color jittering, and noise addition. The rotation angle range of random rotation is -30° to 30°, the cropping ratio of random cropping is 0.8 to 1.0, and the brightness, contrast, and saturation of color jittering vary within a range of ±20%.

[0029] As a preferred solution, the convolutional neural network module includes at least five convolutional layers and three pooling layers. The arrangement structure of the convolutional layers and pooling layers in the convolutional neural network module is: convolutional layer-activation layer-pooling layer-convolutional layer-activation layer-pooling layer-convolutional layer-activation layer-fully connected layer.

[0030] As a preferred solution, the attention mechanism module identifies glioma target areas and organs at risk by combining channel and spatial attention information, and the attention weight matrix is ​​normalized by the Softmax function to ensure that the sum of the weights is 1.

[0031] As a preferred solution, the output layer converts the feature map optimized by the attention mechanism module into a one-dimensional feature vector, and inputs the flattened feature vector into the fully connected layer. The fully connected layer linearly transforms the input feature vector through the weight matrix and adds a bias term, which is expressed as: Z=W×X+b, where Z is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias term; the result of the linear transformation Z is processed by an activation function to introduce nonlinearity. The activation function uses the ReLU activation function, and the formula is: A=ReLU(Z)=max(0,Z); the output A of the activation function is input to the output layer to generate the radiation dose prediction results for glioma patients.

[0032] As a preferred solution, the output layer outputs a dose distribution map based on the radiation dose prediction results for glioma patients, calculates the dose volume histogram based on the predicted dose distribution map, counts the volumes at these doses by setting dose thresholds, calculates the corresponding volume percentages, and presents the calculated DVH data in a graphical manner.

[0033] As a preferred solution, a model training algorithm is also included, and the model training algorithm includes the following steps:

[0034] Step 1: Initialize the weight parameters of the convolutional neural network and set the hyperparameters, including learning rate, batch size, and number of training rounds;

[0035] Step 2: Randomly extract a batch of samples from the training data set and perform data preprocessing, including normalization and enhancement;

[0036] Step 3: Input the preprocessed data into the convolutional neural network and perform forward propagation calculation to obtain the predicted dose distribution;

[0037] Step 4: Calculate the loss between the predicted result and the actual dose, using the mean square error as the loss function;

[0038] Step 5: Calculate the gradient of the loss function to the network parameters through the back propagation algorithm, update the network parameters, and use the Adam optimizer to update the parameters;

[0039] Step 6: Repeat steps 2 to 5 until the set number of training rounds is reached or the loss converges.

[0040] As a preferred solution, it also includes a model evaluation module, which divides the training data set into a training set, a validation set and a test set, preprocesses the test set, and inputs the preprocessed test set into the convolutional neural network module and the attention mechanism module for prediction to obtain the dose distribution map and the corresponding DVH data; uses the mean square error, mean absolute error and R² determination coefficient to calculate the performance indicators of the convolutional neural network module and the attention mechanism module, calculates the difference between the predicted DVH and the true DVH, uses the DVH scoring indicator for evaluation, and puts forward suggestions for model improvement based on the evaluation results.

[0041] It can be seen from the technical solution provided by the present invention that the glioma radiation dose prediction system based on attention enhanced convolutional neural network provided by the present invention has the following beneficial effects: by introducing the attention mechanism, it can effectively highlight important feature areas, thereby improving the prediction accuracy of the radiation dose of glioma patients and ensuring the personalization and precision of the treatment plan; automated dose prediction can quickly generate dose distribution results, reduce the time and energy investment of doctors in radiotherapy plan design, and improve work efficiency. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0043] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with specific implementation methods.

[0044] The embodiment of the present invention provides a glioma radiation dose prediction system based on attention-enhanced convolutional neural network, including a data collection module, a data preprocessing module, a convolutional neural network module, an attention mechanism module and an output layer;

[0045] The data collection module is used to collect radiotherapy data of glioma patients, which includes CT image data, dose distribution data, clinical information of patients and treatment plan information;

[0046] Furthermore, the data collection module acquires three-dimensional CT images of glioma patients through high-resolution CT scanning equipment, with a resolution of not less than 1 mm³; records the radiation dose distribution information actually applied to the patient during radiotherapy, and the radiation dose distribution information includes a dose volume histogram; the patient's clinical information and treatment plan information include the patient's basic information, medical records, and treatment plans;

[0047] The data preprocessing module receives radiotherapy data, extracts CT image data, annotates, cleans, and enhances the CT image data, and generates a training data set;

[0048] Furthermore, the data preprocessing module annotates the target areas and organs at risk in the CT images, removes incomplete and erroneous data, uses a data cleaning algorithm to detect outliers, and uses image enhancement technology to enhance the CT images. The image enhancement technology includes random rotation, random cropping, color jittering, and noise addition. The rotation angle range of random rotation is -30° to 30°, the cropping ratio of random cropping is 0.8 to 1.0, and the brightness, contrast, and saturation of color jittering vary within a range of ±20%.

[0049] The convolutional neural network module includes multiple convolutional layers and pooling layers. Multiple convolutional layers are used to extract feature maps from input image data. In the convolutional layers, convolution kernels of different sizes are set, including 3×3, 5×5, and 7×7. Each convolutional layer is followed by a pooling layer, which reduces the dimension of the feature map through the maximum pooling operation, reduces the amount of calculation, and retains important features.

[0050] The convolutional neural network module includes at least five convolutional layers and three pooling layers. The arrangement structure of the convolutional layers and the pooling layers in the convolutional neural network module is: convolutional layer-activation layer-pooling layer-convolutional layer-activation layer-pooling layer-convolutional layer-activation layer-fully connected layer;

[0051] The attention mechanism module optimizes the feature map and generates a feature map adjusted by self-attention, channel attention and spatial attention. The attention mechanism module identifies the glioma target area and organs at risk by combining the channel and spatial attention information. The attention weight matrix is ​​normalized by the Softmax function to ensure that the sum of the weights is 1. The attention mechanism module includes:

[0052] Self-attention layer: calculates the autocorrelation of the feature map and generates an attention weight matrix to highlight important feature areas and suppress unimportant areas;

[0053] Channel attention layer: Calculate the channel weights through global average pooling and global maximum pooling, adjust the importance of each channel, and use the Sigmoid activation function to generate channel weights;

[0054] Spatial attention layer: Generates spatial weights according to the importance of spatial positions, combines channel and spatial attention information, optimizes feature representation, and uses convolution operations to generate spatial weight maps;

[0055] The output layer is used to generate radiation dose prediction results for glioma patients. The output layer is implemented in the following way:

[0056] The feature map optimized by the attention mechanism module is input into the fully connected layer for dose prediction;

[0057] The activation function is used to output the predicted dose distribution, and the output format is the dose distribution map and the corresponding DVH data;

[0058] Further, the output layer converts the feature map optimized by the attention mechanism module into a one-dimensional feature vector, and inputs the flattened feature vector into the fully connected layer. The fully connected layer linearly transforms the input feature vector through the weight matrix and adds a bias term, which is expressed as: Z=W×X+b, where Z is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias term; the result of the linear transformation Z is processed by an activation function to introduce nonlinearity. The activation function uses the ReLU activation function, and the formula is: A=ReLU(Z)=max(0,Z); the output A of the activation function is input to the output layer to generate the radiation dose prediction result for glioma patients; the output layer outputs a dose distribution map based on the radiation dose prediction result for glioma patients, calculates the dose volume histogram based on the predicted dose distribution map, counts the volume at these doses by setting the dose threshold, and calculates the corresponding volume percentage, and presents the calculated DVH data in a graphical manner.

[0059] In this embodiment, a model training algorithm is also included, and the model training algorithm includes the following steps:

[0060] Step 1: Initialize the weight parameters of the convolutional neural network and set the hyperparameters, including learning rate, batch size, and number of training rounds;

[0061] Step 2: Randomly extract a batch of samples from the training data set and perform data preprocessing, including normalization and enhancement;

[0062] Step 3: Input the preprocessed data into the convolutional neural network and perform forward propagation calculation to obtain the predicted dose distribution;

[0063] Step 4: Calculate the loss between the predicted result and the actual dose, using the mean square error as the loss function;

[0064] Step 5: Calculate the gradient of the loss function to the network parameters through the back propagation algorithm, update the network parameters, and use the Adam optimizer to update the parameters;

[0065] Step 6: Repeat steps 2 to 5 until the set number of training rounds is reached or the loss converges.

[0066] In this embodiment, a model evaluation module is also included. The model evaluation module divides the training data set into a training set, a validation set, and a test set, preprocesses the test set, and inputs the preprocessed test set into the convolutional neural network module and the attention mechanism module for prediction, and obtains the dose distribution map and the corresponding DVH data; the performance indicators of the convolutional neural network module and the attention mechanism module are calculated using the mean square error, the mean absolute error, and the R² determination coefficient, the difference between the predicted DVH and the true DVH is calculated, and the DVH scoring indicator is used for evaluation, and suggestions for model improvement are put forward according to the evaluation results;

[0067] The embodiments of the present invention are described in further detail below:

[0068] This embodiment will describe in detail the various modules and operation steps of the glioma radiation dose prediction system based on attention-enhanced convolutional neural network, including data collection, data preprocessing, convolutional neural network module, attention mechanism module, output layer, model training algorithm and model evaluation module;

[0069] 1. Data Collection Module

[0070] Function: Collect radiotherapy data of glioma patients, including CT images, dose distribution, clinical information and treatment plans;

[0071] Steps:

[0072] High-resolution CT scan:

[0073] Use high-resolution CT scanning equipment to scan glioma patients and obtain three-dimensional CT images, ensuring that the resolution is no less than 1mm³;

[0074] Recording radiation dose distribution:

[0075] During radiotherapy, the actual radiation dose distribution information applied to the patient is recorded, including the dose volume histogram (DVH). This information is used for subsequent dose prediction model training.

[0076] Collect clinical information:

[0077] Collect the patient's basic information (such as age, gender), medical records (such as medical history, treatment response) and treatment plan (such as radiotherapy plan, dose distribution) to provide comprehensive background information for the model;

[0078] 2. Data preprocessing module

[0079] Function: Process the collected radiotherapy data to generate training data sets;

[0080] Steps:

[0081] Data extraction:

[0082] Extract CT image data from the collected radiotherapy data in preparation for subsequent annotation and cleaning;

[0083] Marking target areas and organs at risk:

[0084] Label the target area (tumor area) and organs at risk (such as brain stem, optic nerve, etc.) in the CT image so that the subsequent model training can be accurate;

[0085] Data cleaning:

[0086] Remove incomplete or erroneous data and use data cleaning algorithms to detect outliers to ensure data quality;

[0087] Image Enhancement:

[0088] Use image enhancement technology to process CT images to increase the diversity and robustness of the data, including:

[0089] Random rotation: The rotation angle range is -30° to 30°;

[0090] Random cropping: the cropping ratio ranges from 0.8 to 1.0;

[0091] Color jitter: Brightness, contrast, and saturation change range is ±20%;

[0092] Noise addition: Add random noise to the image to improve the robustness of the model;

[0093] 3. Convolutional Neural Network Module

[0094] Function: Extract features from input image data;

[0095] Steps:

[0096] Build a convolutional neural network model:

[0097] The convolutional neural network model includes at least five convolutional layers and three pooling layers to extract image features;

[0098] Set the convolution kernel:

[0099] Convolution kernels of different sizes (3×3, 5×5, and 7×7) are set in the convolution layer to capture features of different scales;

[0100] Pooling operation:

[0101] Each convolutional layer is followed by a pooling layer, which uses the maximum pooling operation to reduce the dimension of the feature map and the amount of calculation while retaining important features;

[0102] Convolutional neural network model arrangement:

[0103] The arrangement structure of the convolution layer and the pooling layer is:

[0104] Convolution layer-activation layer (ReLU)-pooling layer-convolution layer-activation layer (ReLU)-pooling layer-convolution layer-activation layer (ReLU)-fully connected layer;

[0105] Convolution operation formula:

[0106] The output of the convolutional layer can be expressed as follows:

[0107] in, Indicates that the convolutional layer output feature map is at coordinates The element value at is the input image, is the convolution kernel, and They represent the index variables when traversing and summing the input image and the convolution kernel, respectively, and are used to determine the positions of the elements involved in the convolution operation;

[0108] Activation function:

[0109] Use ReLU activation function for nonlinear transformation:

[0110] Among them, Z is the output of the convolutional layer;

[0111] 4. Attention Mechanism Module

[0112] Function: Optimize feature maps and highlight important feature areas;

[0113] Steps:

[0114] Self-Attention Layer:

[0115] Calculate the autocorrelation of the feature map and generate an attention weight matrix to highlight important feature areas and suppress unimportant areas; the formula for self-attention is:

[0116] in, is the query matrix, is the key matrix, represents the attention weight of the i-th position to the j-th position in the input feature map, represents the total number of positions in the feature map, Represents a transpose operation;

[0117] Channel Attention Layer:

[0118] The channel weights are calculated through global average pooling and global maximum pooling, the importance of each channel is adjusted, and the channel weights are generated using the Sigmoid activation function:

[0119] in, is the feature map after global average pooling, After the global maximum pooling, the feature map is the weight matrix;

[0120] Spatial Attention Layer:

[0121] Generate spatial weights according to the importance of spatial positions, combine channel and spatial attention information, and optimize feature representation; the calculation formula for spatial attention is:

[0122] in, is the feature map, is the spatial weight matrix;

[0123] 5. Output Layer

[0124] Function: Generate radiation dose prediction results for glioma patients;

[0125] Steps:

[0126] Feature map input:

[0127] The feature map optimized by the attention mechanism module is input into the fully connected layer for dose prediction;

[0128] Activation function:

[0129] Use activation functions (such as ReLU) to output the predicted dose distribution, and the output format is the dose distribution map and the corresponding DVH data;

[0130] Eigenvector transformation:

[0131] The optimized feature map is converted into a one-dimensional feature vector and input into the fully connected layer for linear transformation, which is expressed as:

[0132] Among them, Z is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias term;

[0133] Generate the final prediction results:

[0134] The result Z of the linear transformation is processed by the activation function to introduce nonlinearity. The activation function uses ReLU, and the formula is:

[0135] The output A of the activation function is input to the output layer to generate the radiation dose prediction results for glioma patients;

[0136] Calculate DVH data:

[0137] Calculate DVH data based on the predicted dose distribution map, count the volumes at these doses by setting dose thresholds, calculate the corresponding volume percentages, and present the DVH data in a graphical manner;

[0138] 6. Model training algorithm

[0139] Function: Train convolutional neural network model;

[0140] Steps:

[0141] Step 1. Initialize parameters:

[0142] Initialize the weight parameters of the convolutional neural network and set hyperparameters (learning rate, batch size, number of training rounds);

[0143] Step 2: Random sampling:

[0144] Randomly extract a batch of samples from the training dataset and perform data preprocessing (normalization and enhancement);

[0145] Step 3, forward propagation calculation:

[0146] The preprocessed data is input into the convolutional neural network for forward propagation calculation to obtain the predicted dose distribution;

[0147] Step 4: Loss calculation:

[0148] Calculate the loss between the predicted result and the true dose, using the mean square error as the loss function:

[0149] in, is the true value, is the predicted value, is the sample size;

[0150] Step 5: Back propagation update parameters:

[0151] The gradient of the loss function to the network parameters is calculated through the back-propagation algorithm, and the Adam optimizer is used to update the parameters:

[0152] in, is a parameter, is the learning rate, and are the first and second moment estimates, is a constant to prevent division by zero;

[0153] Step 6: Repeat the training process:

[0154] Repeat steps 2 to 5 until the set number of training rounds is reached or the loss converges;

[0155] 7. Model Evaluation Module

[0156] Function: Evaluate the performance of the model;

[0157] Steps:

[0158] Dataset division:

[0159] Divide the training data set into training set, validation set and test set;

[0160] Test set preprocessing:

[0161] Preprocess the test set and input the preprocessed test set into the convolutional neural network module and the attention mechanism module for prediction to obtain the dose distribution map and the corresponding DVH data;

[0162] Performance index calculation:

[0163] The performance indicators of the model are calculated using mean square error, mean absolute error and R² coefficient of determination:

[0164] Mean Squared Error (MSE):

[0165] Mean Absolute Error (MAE):

[0166] Coefficient of determination (R²):

[0167] in, is the mean of the true values;

[0168] DVH comparison:

[0169] Calculate the difference between the predicted DVH and the true DVH and evaluate it using the DVH score metric;

[0170] Improvement suggestions:

[0171] Based on the evaluation results, suggestions for model improvement are put forward, such as adjusting the model structure, optimizing hyperparameters, increasing training data, etc.

[0172] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A glioma radiation dose prediction system based on attention-enhanced convolutional neural network, characterized by: include: Data collection module: used to collect radiotherapy data of glioma patients, the radiotherapy data including CT image data, dose distribution data, clinical information of patients and treatment plan information; Data preprocessing module: receives radiotherapy data, extracts CT image data, annotates, cleans and enhances CT image data, and generates training data sets; Convolutional neural network module: includes multiple convolutional layers and pooling layers; uses multiple convolutional layers to extract feature maps from input image data; in the convolutional layer, sets convolution kernels of different sizes, including 3×3, 5×5, and 7×7; each convolutional layer is followed by a pooling layer, which reduces the dimension of the feature map through the maximum pooling operation, reduces the amount of calculation, and retains important features at the same time; Attention mechanism module: optimizes the feature map to generate a feature map adjusted by self-attention, channel attention and spatial attention. The attention mechanism module includes: Self-attention layer: calculates the autocorrelation of the feature map and generates an attention weight matrix to highlight important feature areas and suppress unimportant areas; Channel attention layer: Calculate the channel weights through global average pooling and global maximum pooling, adjust the importance of each channel, and use the Sigmoid activation function to generate channel weights; Spatial attention layer: Generates spatial weights according to the importance of spatial positions, combines channel and spatial attention information, optimizes feature representation, and uses convolution operations to generate spatial weight maps; Output layer: used to generate radiation dose prediction results for glioma patients. The output layer is implemented in the following ways: The feature map optimized by the attention mechanism module is input into the fully connected layer for dose prediction; The activation function is used to output the predicted dose distribution, and the output format is the dose distribution map and the corresponding DVH data.

2. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 1 is characterized in that: The data collection module acquires a three-dimensional CT image of a glioma patient through a high-resolution CT scanning device, with a resolution of not less than 1 mm³; records the radiation dose distribution information actually applied to the patient during radiotherapy, and the radiation dose distribution information includes a dose volume histogram; the patient's clinical information and treatment plan information include the patient's basic information, medical records, and treatment plans.

3. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 2 is characterized in that: The data preprocessing module marks the target area and organs at risk in the CT image, removes incomplete and erroneous data, uses a data cleaning algorithm to detect outliers, and uses image enhancement technology to enhance the CT image. The image enhancement technology includes random rotation, random cropping, color jittering, and noise addition. The rotation angle range of the random rotation is -30° to 30°, the cropping ratio of the random cropping is 0.8 to 1.0, and the brightness, contrast, and saturation of the color jittering vary in the range of ±20%.

4. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 1, characterized in that: The convolutional neural network module includes at least five convolutional layers and three pooling layers. The arrangement structure of the convolutional layers and the pooling layers in the convolutional neural network module is: convolutional layer-activation layer-pooling layer-convolutional layer-activation layer-pooling layer-convolutional layer-activation layer-fully connected layer.

5. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 1, characterized in that: The attention mechanism module identifies glioma target areas and organs at risk by combining channel and spatial attention information, and the attention weight matrix is ​​normalized by the Softmax function to ensure that the sum of the weights is 1.

6. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 1, characterized in that: The output layer converts the feature map optimized by the attention mechanism module into a one-dimensional feature vector, and inputs the flattened feature vector into the fully connected layer. The fully connected layer linearly transforms the input feature vector through the weight matrix and adds a bias term, which is expressed as: Z=W×X+b, where Z is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias term; the result Z of the linear transformation is processed by an activation function to introduce nonlinearity. The activation function uses the ReLU activation function, and the formula is: A=ReLU(Z)=max(0,Z); the output A of the activation function is input to the output layer to generate the radiation dose prediction result for glioma patients.

7. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 6, characterized in that: The output layer outputs a dose distribution map according to the radiation dose prediction results for glioma patients, calculates the dose volume histogram according to the predicted dose distribution map, counts the volumes under these doses by setting dose thresholds, calculates the corresponding volume percentages, and presents the calculated DVH data in a graphical manner.

8. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 1, characterized in that: Also included is a model training algorithm, the model training algorithm comprising the following steps: Step 1: Initialize the weight parameters of the convolutional neural network and set the hyperparameters, including learning rate, batch size, and number of training rounds; Step 2: Randomly extract a batch of samples from the training data set and perform data preprocessing, including normalization and enhancement; Step 3: Input the preprocessed data into the convolutional neural network and perform forward propagation calculation to obtain the predicted dose distribution; Step 4: Calculate the loss between the predicted result and the actual dose, using the mean square error as the loss function; Step 5: Calculate the gradient of the loss function to the network parameters through the back propagation algorithm, update the network parameters, and use the Adam optimizer to update the parameters; Step 6: Repeat steps 2 to 5 until the set number of training rounds is reached or the loss converges.

9. The glioma radiation dose prediction system based on attention-enhanced convolutional neural network according to claim 1, characterized in that: The method also includes a model evaluation module, which divides the training data set into a training set, a validation set and a test set, preprocesses the test set, inputs the preprocessed test set into the convolutional neural network module and the attention mechanism module for prediction, and obtains the dose distribution map and the corresponding DVH data; calculates the performance indicators of the convolutional neural network module and the attention mechanism module using the mean square error, the mean absolute error and the R² determination coefficient, calculates the difference between the predicted DVH and the true DVH, uses the DVH scoring indicator for evaluation, and proposes suggestions for model improvement based on the evaluation results.

Images