Real-time dose monitoring and adjustment method and system for tumor radiotherapy

Abstract

The present invention relates to the field of deep learning technology, and discloses a real-time dose monitoring and adjustment method and system for tumor radiotherapy, wherein the method comprises: collecting real-time dose data of a treatment area to obtain surface dose distribution data and pH value change data, and identifying respiratory motion features of a patient to generate a respiratory motion feature set; performing four-dimensional space-time division to construct a dynamic dose distribution network; performing feature extraction and integration to obtain a fused distribution feature set; inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data; dynamically adjusting the original control parameters of a radiotherapy device to obtain target control parameters. The method integrates multi-source heterogeneous data, constructs an accurate dose distribution prediction model, and realizes intelligent adjustment of radiotherapy device parameters.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a real-time dose monitoring and adjustment method and system for tumor radiotherapy. Background Art

[0002] Tumor radiotherapy is one of the important means of cancer treatment. Its treatment effect and patient prognosis directly depend on accurate dose distribution. However, traditional radiotherapy methods often have difficulty in real-time monitoring and adjusting dose distribution, resulting in uncertainty in treatment effects and potential risks of side effects. With the development of medical technology, real-time dose monitoring and adjustment have become the key to improving the accuracy of radiotherapy.

[0003] Most existing dose monitoring methods rely on offline measurements or simplified models, which cannot accurately reflect the real-time dose distribution changes during treatment. In addition, factors such as the patient's respiratory movement and organ deformation can also cause deviations between the actual dose distribution and the planned dose. These problems seriously restrict the accuracy and individualization of radiotherapy, and there is an urgent need for a method that can accurately monitor dose distribution in real time and make dynamic adjustments. Summary of the invention

[0004] The present invention provides a real-time dose monitoring and adjustment method and system for tumor radiotherapy, which are used to integrate multi-source heterogeneous data, build an accurate dose distribution prediction model, and realize intelligent adjustment of radiotherapy equipment parameters.

[0005] In a first aspect, the present invention provides a real-time dose monitoring and adjustment method for tumor radiotherapy, the real-time dose monitoring and adjustment method for tumor radiotherapy comprising:

[0006] Collect real-time dose data of the treatment area to obtain surface dose distribution data and pH value change data, and identify the patient's respiratory motion characteristics to generate a respiratory motion feature set;

[0007] Performing four-dimensional space-time division on the treatment area, constructing a dynamic network structure with space-time blocks as nodes, and obtaining a dynamic dose distribution network;

[0008] According to the dynamic dose distribution network, feature extraction and integration are performed on the surface dose distribution data and the pH value change data to obtain a fused distribution feature set;

[0009] Inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data;

[0010] According to the real-time dose distribution prediction data, the original control parameters of the radiotherapy equipment are dynamically adjusted to obtain the target control parameters of the radiotherapy equipment.

[0011] In a second aspect, the present invention provides a real-time dose monitoring and adjustment system for tumor radiotherapy, the real-time dose monitoring and adjustment system for tumor radiotherapy comprising:

[0012] The acquisition module is used to collect real-time dose data of the treatment area, obtain surface dose distribution data and pH value change data, and identify the patient's respiratory motion characteristics to generate a respiratory motion feature set;

[0013] A construction module is used to divide the treatment area into four-dimensional space and time, construct a dynamic network structure with space and time blocks as nodes, and obtain a dynamic dose distribution network;

[0014] An extraction module, used for performing feature extraction and integration on the surface dose distribution data and the pH value change data according to the dynamic dose distribution network to obtain a fused distribution feature set;

[0015] A prediction module, used for inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data;

[0016] The adjustment module is used to dynamically adjust the original control parameters of the radiotherapy equipment according to the real-time dose distribution prediction data to obtain the target control parameters of the radiotherapy equipment.

[0017] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned real-time dose monitoring and adjustment method for tumor radiotherapy.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned real-time dose monitoring and adjustment method for tumor radiotherapy.

[0019] In the technical solution provided by the present invention, by collecting surface dose distribution data, pH value change data and respiratory motion characteristics, the effective integration of multi-source heterogeneous data is realized, and more comprehensive treatment area information is provided. The four-dimensional spatiotemporal division and dynamic network structure are adopted to better capture the spatiotemporal change characteristics of the treatment area, and improve the accuracy and spatiotemporal resolution of dose distribution prediction. Combining the hybrid architecture of 3DCNN, GCN and GRU, the space, graph structure and timing information are fully utilized, and the accuracy and generalization ability of dose distribution prediction are significantly improved. The reinforcement learning algorithm based on the double Q network realizes the intelligent real-time adjustment of radiotherapy equipment parameters, and effectively copes with the uncertainty factors in the treatment process. Real-time monitoring of dose and physiological parameters is realized through external sensors, avoiding the potential risks of traditional invasive methods, and improving the patient's comfort and treatment compliance. Taking into account the patient's respiratory motion characteristics, more accurate individualized dose distribution prediction and adjustment are achieved, which helps to improve the treatment effect and reduce side effects. The multi-level feature extraction and optimization strategy is adopted to improve the calculation efficiency of the algorithm while ensuring the prediction accuracy, and meet the needs of real-time monitoring and adjustment.

[0020] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of an embodiment of a real-time dose monitoring and adjustment method for tumor radiotherapy according to an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of an embodiment of a real-time dose monitoring and adjustment system for tumor radiotherapy according to an embodiment of the present invention;

[0024] Figure 3 FIG. 1 is a schematic diagram of an embodiment of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products, or device ends.

[0027] To facilitate understanding of this embodiment, a real-time dose monitoring and adjustment method for tumor radiotherapy disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0028] 101. Collect real-time dose data of the treatment area to obtain surface dose distribution data and pH value change data, and identify the patient's respiratory motion characteristics to generate a respiratory motion feature set;

[0029] It is understandable that the execution subject of the present invention may be a real-time dose monitoring and adjustment system for tumor radiotherapy, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0030] Specifically, a high-precision fiber array detector is arranged in the treatment area, and the original light signal data from the treatment area is collected through the detector, which contains information about the surface dose distribution of the treatment area. The original light signal data is Fourier transformed, converted from the time domain to the frequency domain, and the frequency domain feature data is obtained to capture the spatial distribution characteristics of the surface dose. Based on the frequency domain feature data, the signal is inverted or restored to generate accurate surface dose distribution data. At the same time, a pH-sensitive optical sensor is arranged in the treatment area to collect the original pH spectrum data. The original pH spectrum data is subjected to principal component analysis to obtain pH value change data. The patient's respiratory movement is monitored in real time. By segmenting the patient's body surface image, the region of interest related to breathing is identified, a respiratory region mask is generated, and the body parts that need to be monitored are defined. The displacement field is calculated between consecutive image frames, and the changes in the respiratory area between different time points are analyzed to obtain the vector field of the respiratory movement, which reflects the displacement and movement characteristics of the patient's body surface area during breathing. Wavelet transform is performed based on the respiratory motion vector field to capture the changes in different frequency components in the respiratory motion, extract the time-frequency features, reflect the periodicity and irregularity of breathing, and finally combine the wavelet transform results to generate a complete set of respiratory motion features.

[0031] 102. Divide the treatment area into four-dimensional space and time, construct a dynamic network structure with space and time blocks as nodes, and obtain a dynamic dose distribution network;

[0032] Specifically, the treatment area is processed into a three-dimensional spatial grid, and the entire treatment area is divided into multiple small spatial units to obtain spatial grid data. The spatial grid data is expanded in the time dimension, and the time dimension is introduced on the basis of the original three-dimensional spatial grid to form a four-dimensional spatiotemporal data set containing time and space information. The four-dimensional data set can record the dynamic change characteristics of these doses over time. Block division is performed based on the four-dimensional spatiotemporal data set. The four-dimensional data set is divided into multiple spatiotemporal blocks, each block represents a small spatial position in the treatment area and its changes in a specific time period. After the division is completed, multiple spatiotemporal blocks form a block set, and each spatiotemporal block is regarded as a node in the network, thereby constructing an initial network structure. The nodes in the initial network structure are analyzed for spatial adjacency relationships to obtain the relative position relationship of each node in space, determine which nodes are adjacent to each other in space, and generate spatial correlation data based on this, reflecting the possible mutual influence between adjacent grids in physical space. At the same time, the nodes are analyzed for time series correlation to determine the change trend of the same node in different time periods and the temporal correlation between different nodes. Through time series analysis, we can obtain the association information of nodes in the time dimension, generate time-related data, and reflect the law of dose variation over time and the time interaction between different nodes. Based on the spatial association data and time association data, we can calculate the comprehensive association strength between each node in the network. According to the comprehensive association strength, we can optimize the initial network structure, adjust and strengthen important associations, reduce unnecessary connections, and form the final dynamic dose distribution network, which can reflect the distribution of dose in the treatment area and its trend over time in real time.

[0033] 103. According to the dynamic dose distribution network, the surface dose distribution data and pH value change data are feature extracted and integrated to obtain a fused distribution feature set;

[0034] Specifically, the network structure feature extraction of the dynamic dose distribution network is performed, and the complex network structure information is converted into an operable numerical form to generate a network feature vector. Based on the network feature vector, the surface dose distribution data is weighted by spatial attention. Through the spatial attention mechanism, the spatial regions with high importance in the network are focused on, and higher weights are given to these regions, so that the weighted surface dose distribution data can more accurately reflect the dose changes in key positions. Multi-level spatial feature extraction is performed on the weighted surface dose distribution data, and the spatial information in the data is gradually mined in depth through multiple feature extraction layers to obtain a surface dose feature set with high expression ability. At the same time, based on the network feature vector, the pH value change data is weighted by temporal attention, emphasizing the more important part of the data in the time dimension. By giving higher weights to key time points, the weighted pH value change data can more accurately reflect the dose change trend in time. Time series feature extraction is performed on the weighted pH value change data to capture the key patterns of pH value changes over time from the time dimension to obtain the pH value feature set. The surface dose feature set and the pH value feature set are feature concatenated, and the data features from two different sources are spliced ​​and integrated to form an initial fused feature set. The initial fused feature set is subjected to feature dimension reduction and denoising. Through feature dimension reduction, redundant information is removed and the computational complexity is reduced; through denoising, the inevitable noise in data acquisition is effectively filtered out to obtain a compressed feature set. Based on the compressed feature set, a feature graph is constructed to intuitively represent the association and distribution between each feature. In order to optimize these features, a non-local attention mechanism is used for feature optimization. The non-local attention mechanism captures the potential association between long-distance features by paying global attention to the entire feature space, effectively improving the expressiveness and generalization of the features, and finally obtaining a fused distribution feature set that comprehensively reflects the key information in the spatial dose distribution, the temporal pH value changes, and the network structure.

[0035] 104. Inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data;

[0036] Specifically, the fused distribution feature set is reconstructed in three dimensions to generate feature volume data, which reflects the spatial dose distribution of the treatment area and the changing trend of pH value. The feature volume data is registered by 3D spatial transformation network to make the data more aligned in space, reduce errors, and obtain the aligned feature volume. The aligned feature volume is input into the three-dimensional convolutional neural network module for processing. The three-dimensional convolutional neural network module contains 5 3D convolutional layers, and each convolutional layer is connected to 3D batch normalization and LeakyReLU activation function to ensure the stability of feature data after convolution operation and the introduction of nonlinear features. At the same time, the 3D maximum pooling layer is introduced after the 3rd and 5th convolution layers. The maximum pooling can downsample the features and extract important information in the spatial distribution. Through multi-level convolution operations, a spatial feature map is obtained to integrate spatial information at different levels. The spatial feature map is reshaped into a graph structure so that it can be input into the graph convolution layer for graph feature extraction. In the graph convolution operation, the Chebyshev polynomial approximation method is used for graph convolution calculation, which can effectively reduce the computational complexity while capturing the association between nodes. After each graph convolution operation, batch normalization and ReLU activation function are applied to ensure the stability and nonlinear expression ability of the features. After three graph convolution operations, the final graph feature representation is extracted to fully capture the local and global associations in the spatial data. Global average pooling and global maximum pooling are performed on the graph feature representation to compress and aggregate these features, and then the pooling results are spliced ​​to generate a graph-level feature vector. The respiratory motion feature set is input into the multi-layer perceptron for feature transformation to obtain the respiratory feature vector. The multi-layer perceptron can effectively extract important patterns in the respiratory features and use them in combination with the dose distribution data. The graph-level feature vector is spliced ​​with the respiratory feature vector to obtain a fused feature vector. The fused feature vector is expanded by time step and input into the bidirectional GRU (gated recurrent unit) network for learning of temporal features. The bidirectional GRU network contains two layers of GRU units, each with 128 hidden units, which can effectively capture the forward and backward information in the time series data. Through this bidirectional structure, key information is extracted from historical and future dose distribution trends to improve the accuracy of prediction. The self-attention mechanism is used to analyze the temporal feature sequence. By calculating the correlation between different time steps, more important time steps are given higher weights to obtain weighted temporal features. The weighted temporal features are input into a three-dimensional deconvolution network to predict the dose distribution. The three-dimensional deconvolution network contains four 3D deconvolution layers, each of which is followed by three-dimensional instance normalization and ReLU activation function. The features are upsampled through deconvolution operation and restored to the same spatial resolution as the input data. In the deconvolution layer, the Tanh activation function is used to ensure that the output dose prediction value is within an appropriate range. Generate real-time dose distribution prediction data.

[0037] Among them, the weighted time series features are reshaped and converted into input tensors suitable for three-dimensional deconvolution operations to generate initial input features. Based on the initial input features, the first 3D deconvolution operation is performed. The convolution kernel size used in this operation is 3x3x3, the step size is 2, and the padding is 1. With these parameters, the deconvolution operation can upsample the input features, expand the size of the feature map, and obtain the first layer of feature maps. Three-dimensional instance normalization and ReLU activation function are applied to the first layer of feature maps to obtain the first layer of activation features. Three-dimensional instance normalization effectively adjusts the distribution of each channel in the feature map and eliminates the imbalance caused by feature changes between different samples. The ReLU activation function introduces nonlinearity to help the model better capture complex features. The first layer of activation features is input to the second 3D deconvolution layer for processing. The convolution kernel size used in this layer is 3x3x3, the step size is 2, and the padding is 1, the purpose of which is to further upsample the spatial resolution of the feature map. After the second deconvolution operation, the second layer feature map is generated, and 3D instance normalization and ReLU activation function are applied to it again to obtain the second layer activation features. The second layer activation features are residually connected. The second layer activation features are fused with the first layer activation features. In order to match the dimensions of the two, the first layer activation features are reduced in dimension through 1x1x1 convolution. The first layer features after dimension reduction are added to the second layer activation features to form the fused features after residual connection. Residual connection can alleviate the gradient vanishing problem in deep networks, while retaining the useful information of the previous layer and improving the expressiveness of features. The fused features are input into the third 3D deconvolution layer. The convolution kernel size of this layer is also 3x3x3, the stride is 1, and the padding is 1. The feature map is refined through convolution operations with smaller strides to generate a more refined third layer feature map. 3D instance normalization and ReLU activation function are applied to the feature map to obtain the third layer activation features. The third layer activation features are processed by spatial attention mechanism. By calculating the spatial weight, the important areas in the feature map are identified and weighted. After weighted processing, the obtained spatial enhancement features can more effectively highlight the key dose distribution information in space. The spatial enhancement features are input into the fourth 3D deconvolution layer for processing. The convolution kernel size used in this layer is still 3x3x3, the step size is 1, and the padding is 1. Through this layer of operation, the feature map of the dose distribution is obtained. On this feature map, the Tanh activation function is applied to generate the dose distribution prediction result. The Tanh activation function can ensure that the output value is in the range of -1 to 1, providing a smoother output for subsequent processing. The dose distribution prediction results are smoothed by three-dimensional median filtering. Median filtering can effectively remove noise in the prediction results and ensure that the dose distribution is smoother and continuous. On this basis, a thresholding operation is performed to map the predicted value to the preset dose range to obtain real-time dose distribution prediction data that meets actual treatment needs.

[0038] 105. According to the real-time dose distribution prediction data, the original control parameters of the radiotherapy equipment are dynamically adjusted to obtain the target control parameters of the radiotherapy equipment.

[0039] Specifically, the real-time dose distribution prediction data and the original control parameters of the radiotherapy equipment are state-encoded, and the two types of data are integrated into a unified representation to generate the current state vector, which contains the specific situation of the dose distribution in the treatment area and the current operating parameters of the equipment. The action space is constructed based on the current state vector. The action space represents the possible adjustment options of all adjustable parameters of the radiotherapy equipment in the current state. In order to facilitate calculation, the adjustable parameters of the radiotherapy equipment are discretized into multiple adjustment options to form an action set. Each action in the action set represents a specific adjustment to a certain parameter of the equipment, such as the change of the intensity of the radiation dose, the irradiation angle or the radiation time. Each action in the action set is evaluated by an online Q network to calculate the value of each action in the current state, which is specifically expressed as a Q value. The Q value represents the impact of the radiotherapy equipment on the future dose distribution prediction and treatment effect after selecting a certain action. Through evaluation, an action value sequence is obtained, which contains the Q values ​​of all possible actions and reflects the expected effect of each adjustment option. Based on the action value sequence, the ε-greedy strategy is used to select the current optimal action. The ε-greedy strategy is a common method in reinforcement learning. It balances random exploration and greedy selection to ensure that there is still a certain probability of exploring new action choices based on certain information. At the current stage, the optimal action selected by the ε-greedy strategy represents the operation plan that is most likely to bring ideal results to the equipment adjustment under the existing knowledge. The optimal action is applied to the target Q network to calculate the target Q value. The target Q value can reflect the impact of the optimal action on the dose distribution prediction over a longer time range and further optimize the control parameters of the equipment. Based on the target action value and the previously calculated action value sequence, the temporal difference error is calculated. The temporal difference error measures the difference between the current Q network prediction value and the target Q network value. The error is used to adjust the parameters of the online Q network through the back propagation mechanism. By minimizing the loss function constructed by the error, the parameters of the online Q network are gradually optimized, improving its ability to predict the effect of radiotherapy equipment adjustment. In order to ensure the stability of the model and the continuous improvement of learning, the parameters of the optimized online Q network are copied to the target Q network every preset update cycle. The update of the target Q network can improve the accuracy and stability of future action selection and prevent the network from falling into local optimality or parameter drift. Based on the updated target Q network, the target control parameters of the radiotherapy equipment are output.

[0040] In the embodiment of the present invention, by collecting surface dose distribution data, pH value change data and respiratory motion characteristics, the effective integration of multi-source heterogeneous data is achieved, and more comprehensive treatment area information is provided. The four-dimensional spatiotemporal division and dynamic network structure can better capture the spatiotemporal change characteristics of the treatment area, and improve the accuracy and spatiotemporal resolution of dose distribution prediction. Combining the hybrid architecture of 3DCNN, GCN and GRU, the space, graph structure and timing information are fully utilized, and the accuracy and generalization ability of dose distribution prediction are significantly improved. The reinforcement learning algorithm based on the double Q network realizes the intelligent real-time adjustment of radiotherapy equipment parameters, and effectively copes with the uncertainty factors in the treatment process. The real-time monitoring of dose and physiological parameters is realized through external sensors, which avoids the potential risks of traditional invasive methods and improves the patient's comfort and treatment compliance. Taking into account the patient's respiratory motion characteristics, more accurate individualized dose distribution prediction and adjustment are achieved, which helps to improve the treatment effect and reduce side effects. The multi-level feature extraction and optimization strategy is adopted to improve the calculation efficiency of the algorithm while ensuring the prediction accuracy, and meet the needs of real-time monitoring and adjustment.

[0041] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0042] Arrange a high-precision fiber array detector in the treatment area to collect raw light signal data, and perform Fourier transform on the raw light signal data to obtain frequency domain feature data;

[0043] Perform signal restoration based on frequency domain feature data to obtain surface dose distribution data;

[0044] Arranging a pH-sensitive optical sensor in the treatment area, collecting raw pH spectrum data, and performing principal component analysis on the raw pH spectrum data to obtain pH value change data;

[0045] Perform body surface image segmentation on the patient, identify the breathing-related region of interest, and obtain the breathing region mask;

[0046] Calculate the displacement field of the respiratory region mask between consecutive frames to obtain the respiratory motion vector field;

[0047] Wavelet transform and time-frequency feature extraction are performed based on the respiratory motion vector field to obtain a respiratory motion feature set.

[0048] Specifically, multiple physical quantities are monitored and analyzed in the treatment area. High-precision fiber array detectors are arranged in the treatment area to collect raw light signal data in real time. Light signals reflecting surface dose distribution are captured. The raw light signal data is Fourier transformed to convert it into frequency domain feature data. The Fourier transform formula is:

[0049]

[0050] Among them, X(f) represents the frequency domain feature data, x(t) is the original time domain signal, f is the frequency, t is the time variable, and j is the imaginary unit. Through Fourier transform, the frequency components in the original signal are decomposed, making it easier to identify and remove noise interference, and obtain clear frequency domain feature data, which effectively describes the distribution of dose in the treatment area. Based on the frequency domain feature data, signal restoration is performed to obtain more intuitive surface dose distribution data. The process of signal restoration uses inverse Fourier transform, and its formula is:

[0051]

[0052] In this way, the frequency domain feature data is restored to the time domain signal to obtain the dose distribution on the surface of the treatment area, reflecting the intensity distribution of the radiation dose at different positions in the treatment area. A pH-sensitive optical sensor is arranged in the treatment area to capture specific spectral data to reflect the pH value changes in the treatment area and obtain the original pH spectral data. Spectral data usually contains information in multiple dimensions. The original pH spectral data is subjected to dimensionality reduction processing. Principal component analysis projects high-dimensional data into low-dimensional space, retains the most important change patterns, and reduces the dimension of the data. The formula for principal component analysis is:

[0053] Z = XW;

[0054] Among them, Z represents the principal component after projection to low-dimensional space, X is the matrix of original pH spectral data, and W is the projection matrix containing the principal component. Through principal component analysis, key features reflecting pH changes are extracted, and these features are used to monitor changes in the cell microenvironment during treatment. At the same time, during tumor radiotherapy, the patient's respiratory movement will cause changes in the position of the treatment area. The patient's body surface image is segmented to identify the regions of interest related to breathing, which are defined as respiratory area masks. Image segmentation uses an algorithm based on convolutional neural networks to identify the areas on the patient's body surface that move with breathing changes and obtain the precise outline of the respiratory area. The displacement field between consecutive frames of the respiratory area mask is calculated to obtain the respiratory motion vector field. The displacement field is calculated using the optical flow method, which calculates the displacement vector by analyzing the pixel changes between adjacent image frames. Assuming that between two consecutive frames, the displacement of a point is represented by a vector v(x, y, t), where x and y are pixel coordinates and t is time, the basic formula of the optical flow method is:

[0055] I(x+v x ,y+v y ,t+1)=I(x,y,t);

[0056] Where I(x,y,t) represents the image intensity at position (x,y) at time t, v x ,v y is the displacement component in the x,y direction. The vector field of respiratory motion in the treatment area is calculated by the formula. Wavelet transform is performed based on the respiratory motion vector field. Wavelet transform is a time-frequency analysis method that can simultaneously capture the local frequency and time characteristics of the signal. By performing wavelet transform on the respiratory motion vector field, multi-scale motion patterns related to breathing are extracted to obtain a more accurate set of respiratory features. The formula for wavelet transform is:

[0057]

[0058] Among them, W(a,b) represents the wavelet coefficient, a and b are the scale and translation parameters respectively, ψ(t) is the wavelet function, and f(t) is the time function of the respiratory motion vector field. By adjusting the scale parameter a, the wavelet transform captures the characteristics of different time scales in the respiratory process and extracts a complete set of respiratory motion features.

[0059] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0060] Perform three-dimensional spatial gridding on the treatment area to obtain spatial grid data, and then expand the spatial grid data in time dimension to obtain a four-dimensional space-time data set;

[0061] Based on the four-dimensional space-time data set, block division is performed to obtain a space-time block set, and each block in the space-time block set is set as a network node to construct an initial network structure;

[0062] Perform spatial adjacency analysis on the nodes in the initial network structure to obtain spatial correlation data, and perform time series correlation analysis on the nodes in the initial network structure to obtain time correlation data;

[0063] Based on the spatial correlation data and temporal correlation data, the comprehensive correlation strength between nodes is calculated, and the initial network structure is optimized according to the comprehensive correlation strength to obtain a dynamic dose distribution network.

[0064] Specifically, the treatment area is gridded in three dimensions, and the entire treatment area is divided into many small cubic units, each of which represents a spatial position in the area and contains corresponding dose data. Assuming that the size of the treatment area is L×W×H, it is divided into N x ×N y ×N z A cubic grid, each grid unit size is Δx, Δy, Δz, the corresponding spatial grid data is expressed as:

[0065] S(x,y,z)={S i,j,k |i=1,…,N x ; j = 1,…,N y ; k = 1,…,N z};

[0066] Among them, S i,j,k Represents the dose data in the (i, j, k) grid cell. The spatial grid data is expanded in the time dimension to generate a four-dimensional spatiotemporal data set. The data of each grid cell is sampled at each time step t, and time t is added to the data set as the fourth dimension. By continuously collecting dose data that changes over time, a four-dimensional data set containing spatiotemporal information is obtained, which is formally expressed as:

[0067] T(x,y,z,t)={T i,j,k,t |t=1,…,T};

[0068] Among them, T i,j,k,t Represents the dose data in the (i, j, k) grid cell at time t. The four-dimensional data set reflects the dose distribution in the treatment area that changes with time and space. Based on the obtained four-dimensional spatiotemporal data set, the spatiotemporal block division is performed. The entire four-dimensional data set is divided into multiple spatiotemporal blocks, each of which represents the dose information in a certain time period and spatial region. Assume that each block covers M in space. x ×M y ×M z grid cells and covers M in time t time steps, the space-time block is represented as:

[0069] B(x,y,z,t)={B p,q,r,s |p=1,…,P; q=1,…,Q; r=1,…,R; s=1,…,S};

[0070] Among them, B p,q,r,s Represents the dose distribution data in time period s and spatial block (p, q, r). These blocks are regarded as nodes in the network, and the initial spatiotemporal network structure is constructed, in which each node represents a specific spatiotemporal block. In order to analyze the relationship between nodes, the nodes are analyzed for spatial adjacency. Spatial adjacency refers to which nodes are adjacent or close to each other in physical space. Assume that node B i and B j They correspond to two space-time blocks, and their spatial coordinates are x i =(x i ,y i ,z i ) and x j =(x j ,y j ,zj ), then their spatial distance is expressed as:

[0071]

[0072] By calculating the distance d between nodes ij , get the spatial adjacency matrix A between nodes s , where if d ij If it is less than a preset threshold ∈, it is considered that there is a spatial association between the two nodes. That is:

[0073]

[0074] Time series correlation analysis is performed to determine the association between nodes in the time dimension. Time correlation is achieved by calculating the correlation coefficient of the dose data between nodes over time. Assume that node B i and B j The dose distributions in T time steps are {T i (t)} and {T j (t)}, then their time correlation coefficient r ij It is expressed as:

[0075]

[0076] in, and Node B i and B j The average dose value. By calculating the time correlation coefficient, the time correlation matrix A is obtained t , where A t (i,j) represents node B i and B j The temporal correlation strength between them. Based on the spatial adjacency matrix A s and the time correlation matrix A t , calculate the comprehensive correlation strength between each node. Comprehensive correlation strength S ij is a weighted combination of spatial and temporal associations, and is given by:

[0077] S ij =αA s (i,j)+βA t (i,j);

[0078] Among them, α and β are the weight coefficients of spatial association and temporal association, which are used to adjust the influence of the two on the comprehensive association strength. Through the formula, the preliminary node comprehensive association matrix S is obtained. The initial network structure is optimized based on the comprehensive association strength. The connections between nodes are weighted and adjusted according to the comprehensive association strength, the connections with weak association strength are removed, and the important associations are retained and strengthened to obtain a more compact and efficient dynamic dose distribution network.

[0079] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0080] Extracting the network structure features of the dynamic dose distribution network to obtain the network feature vector, and performing spatial attention weighting on the surface dose distribution data based on the network feature vector to obtain the weighted surface dose distribution data;

[0081] Perform multi-level spatial feature extraction on the weighted surface dose distribution data to obtain a surface dose feature set;

[0082] Based on the network feature vector, the pH value change data is weighted by time attention to obtain the weighted pH value change data, and the time series feature extraction is performed on the weighted pH value change data to obtain the pH value feature set;

[0083] The surface dose feature set and the pH value feature set are feature concatenated to obtain an initial fused feature set, and the initial fused feature set is subjected to feature dimension reduction and denoising to obtain a compressed feature set;

[0084] A feature map is constructed based on the compressed feature set, and a non-local attention mechanism is applied to optimize the features to obtain a fused distributed feature set.

[0085] Specifically, the network structure features of the dynamic dose distribution network are extracted. The graph convolution network is applied to process the network structure. The graph convolution network performs convolution operations on nodes through the adjacency matrix and feature matrix to capture the global structural information between nodes. Assuming that the adjacency matrix is ​​A, the calculation of the graph convolution is expressed as:

[0086] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) );

[0087] Among them, H (l) is the input feature matrix of the lth layer, W (l)is the weight matrix of the layer, D is the node degree matrix, A is the adjacency matrix, and σ represents the activation function. After multiple layers of graph convolution operations, the characteristic vector of the network is obtained, which effectively describes the position and importance of each node in the dynamic dose distribution network. Based on the extracted network characteristic vector, the surface dose distribution data is spatially weighted. Different weights are assigned to the surface dose data according to the characteristics of different nodes in the network, so as to focus on the areas with high importance in the network and optimize the distribution and adjustment of dose. Assume that the network characteristic vector is v i , the surface dose distribution data is S i , then the weighted dose data is expressed as:

[0088] S′ i =v i ·S i ;

[0089] Through weighting, the dose data of important areas are amplified, while the dose data of secondary areas are weakened accordingly, generating weighted surface dose distribution data. Multi-level spatial feature extraction is performed on the weighted surface dose distribution data. By using a multi-layer convolutional neural network, the spatial features of the surface dose data are extracted layer by layer. The convolution operation generates a surface dose feature set by extracting local and global spatial patterns. It contains spatial information at different levels and helps to accurately capture the complex changes in dose distribution. At the same time, based on the network feature vector, the pH change data is weighted by time attention. Focusing on specific time points, the impact of pH changes at certain critical moments on the treatment effect is identified. Assume that the network feature vector is v t , pH data is P t , then the weighted pH value data is:

[0090] P′ t =v t ·P t ;

[0091] Through the weighting process, the time periods with significant changes during the treatment process are analyzed. The weighted pH data are subjected to time series feature extraction, and the dynamic patterns in time are extracted using a long short-term memory network to finally generate a pH feature set. The surface dose feature set and the pH feature set are subjected to feature cascade operation. Data from different sources are combined into a comprehensive feature set so that both spatial dose distribution and temporal pH changes can be considered in subsequent processing. Assume that the surface dose feature set is F s , pH value feature set is F p , then the initial fusion feature set is expressed as:

[0092] F fusion =[F s ; F p ];

[0093] Perform feature dimension reduction and denoising on the initial fused feature set. Use principal component analysis or autoencoder to effectively reduce the dimension of the data and remove redundant features to obtain a compressed feature set. Construct a feature graph based on the compressed feature set. A feature graph is a way of using features as nodes and constructing a graph structure through the relationships between nodes, which can better capture the potential correlation between features. Based on the feature graph, a non-local attention mechanism is applied for feature optimization. The non-local attention mechanism calculates the feature correlation in a global range so that each node in the feature graph can perceive the information of distant nodes, breaking the limitation that traditional convolution operations can only focus on local areas. The formula for non-local attention is:

[0094]

[0095] Among them, Y i is the optimized feature, X i and X j are the features of nodes i and j respectively, f(X i ,X j ) is a function that calculates the correlation between two nodes, g(X j ) is the feature X j The transformation of , C(i) is the coefficient used for normalization. Through the non-local attention mechanism, the correlation between global features is better captured, and finally a fusion distribution feature set is generated.

[0096] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0097] Perform three-dimensional reconstruction on the fused distribution feature set to obtain feature volume data, and perform 3D spatial transformation network registration on the feature volume data to obtain aligned feature volumes;

[0098] The aligned feature volume is input into the 3D convolutional neural network module, which contains 5 3D convolutional layers, each followed by 3D batch normalization and LeakyReLU activation function, and a 3D maximum pooling layer is added after the 3rd and 5th layers to obtain the spatial feature map;

[0099] The spatial feature map is reshaped into a graph structure and input into the graph convolution layer for graph feature extraction to obtain graph feature representation. The graph convolution layer contains 3 graph convolution operations, each of which is implemented using Chebyshev polynomial approximation, and graph batch normalization and ReLU activation function are applied after each graph convolution operation;

[0100] Perform global average pooling and global maximum pooling on the graph feature representation, and concatenate the pooling results to obtain the graph-level feature vector;

[0101] The respiratory motion feature set is input into a multi-layer perceptron for feature transformation to obtain a respiratory feature vector, and the graph-level feature vector and the respiratory feature vector are concatenated to obtain a fused feature vector;

[0102] The fused feature vector is expanded by time step and input into the bidirectional GRU network to obtain a temporal feature sequence. The bidirectional GRU network contains 2 layers of bidirectional GRU units, with 128 hidden units in each layer.

[0103] Perform self-attention mechanism analysis on the time series feature sequence, calculate the correlation between different time steps, and obtain weighted time series features;

[0104] The weighted temporal features are input into the three-dimensional deconvolution network for dose distribution prediction to obtain real-time dose distribution prediction data. The three-dimensional deconvolution network contains 4 3D deconvolution layers, each layer is followed by three-dimensional instance normalization and ReLU activation function, and the last layer uses Tanh activation function.

[0105] Specifically, the fused distribution feature set represents the overall information of the spatial dose distribution and related physiological data of the tumor area. The fused distribution feature set is reconstructed in three dimensions, and the one-dimensional or two-dimensional features are converted into a three-dimensional tensor with depth information to generate a feature volume data that can reflect spatial changes. The feature volume data T(x, y, z) contains the dose distribution at each spatial position in the treatment area. The feature volume data is aligned with the 3D spatial transformation network to ensure that the data at different time points during the treatment process are spatially aligned. The volume data at different time points are aligned to eliminate body position deviation caused by breathing or other factors, and to ensure the continuity and consistency of the analysis. Assume that the coordinates of the feature volume at two time points are (x t ,y t ,z t ) and (x t+1 ,y t+1 ,z t+1 ), the registration process is performed through the transformation matrix M:

[0106]

[0107] Through the registration operation, the aligned feature volume data is obtained to ensure that subsequent operations are based on consistent spatial positions. The aligned feature volume data is input into the 3D convolutional neural network module for processing. The 3D convolutional neural network module contains 5 3D convolution layers, each layer uses a 3×3×3 convolution kernel to extract spatial information in the feature volume through stereo convolution operations. The convolution calculation formula is:

[0108] O i,j,k =∑ m,n,p X i+m,j+n,k+p ·W m,n,p +b;

[0109] Among them, O i,j,k is the convolution output, X i,j,k is the input feature volume, W m,n,p is the convolution kernel weight, and b is the bias term. Through 5 layers of convolution operations, the network extracts the spatial features of the treatment area layer by layer. After each layer of convolution, 3D batch normalization and LeakyReLU activation function are performed to ensure the stability and nonlinear expression ability of the network. 3D maximum pooling layers are added to the 3rd and 5th layers to reduce the dimension of the features, reduce the computational complexity while retaining important information. The spatial feature map is reshaped into a graph structure. The graph structure can effectively represent the relationship between different positions in the spatial area. Each position in the spatial feature map is regarded as a node in the graph, and the relationship between adjacent nodes is represented by the edges connecting them. The graph structure is input into the graph convolution layer for graph feature extraction. The graph convolution layer uses Chebyshev polynomial approximation to calculate the feature aggregation of nodes on the graph, and the formula is:

[0110]

[0111] Among them, H (l) is the node feature matrix of the lth layer, is based on the graph Laplacian matrix The Chebyshev polynomial of α k is the weight coefficient, and σ is the activation function. After three graph convolution operations, a high-level graph feature representation is extracted from the graph structure. In order to compress the features, global average pooling and global maximum pooling are performed on the graph feature representation. Average pooling extracts the overall information of the features, while maximum pooling captures the most important local features. By splicing the results of the two, a graph-level feature vector is obtained. At the same time, the patient's respiratory motion characteristics are analyzed to capture the displacement of the treatment area caused by breathing. The respiratory motion feature set is transformed by a multi-layer perceptron to obtain a respiratory feature vector. The multi-layer perceptron maps the input features to a high-dimensional space through linear transformation and nonlinear activation functions to generate features that can represent the respiratory motion pattern. The respiratory feature vector is spliced ​​with the graph-level feature vector to form a fused feature vector. The fused feature vector is expanded by time step and input into the bidirectional GRU network for temporal feature extraction. The bidirectional GRU (gated recurrent unit) network can capture the trend of feature changes over time, and its calculation formula is:

[0112]

[0113] Among them, h t is the hidden state at the current moment, z t It is the update gate. is a candidate hidden state. The bidirectional GRU can process both forward and backward time information and capture more complex time dependencies. The network consists of 2 layers, each with 128 hidden units, and outputs a time series feature sequence. The time series feature sequence is analyzed by the self-attention mechanism, and a weight is assigned to each time step by calculating the correlation between different time steps. The calculation formula is:

[0114]

[0115] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key. Through the self-attention mechanism, we identify which time steps are most important for dose distribution prediction and obtain weighted time series features. The weighted time series features are input into the three-dimensional deconvolution network for dose distribution prediction. The three-dimensional deconvolution network contains 4 3D deconvolution layers, each of which is followed by three-dimensional instance normalization and ReLU activation function. The deconvolution operation is used to upsample the features and restore them to the same spatial resolution as the input. The calculation formula is:

[0116] O i,j,k =∑ m,n,p X i+m,j+n,k+p ·W m,n,p +b;

[0117] In the last layer of deconvolution operation, the Tanh activation function is used to ensure that the output value is within an appropriate range. Through the deconvolution operation, real-time dose distribution prediction data is generated to reflect the dose distribution of the tumor area at a specific time point. In order to smooth the prediction results, the dose distribution prediction results are processed by three-dimensional median filtering. The median filter can remove noise and ensure the smoothness and continuity of the dose distribution. The predicted value is mapped to the preset dose range through the thresholding operation to ensure that the output dose meets the treatment requirements.

[0118] In a specific embodiment, the step of executing inputs the weighted time series features into a three-dimensional deconvolution network for dose distribution prediction to obtain real-time dose distribution prediction data. The three-dimensional deconvolution network includes four 3D deconvolution layers, each layer is followed by three-dimensional instance normalization and ReLU activation function, and the process of using the Tanh activation function in the last layer can specifically include the following steps:

[0119] Reshape the weighted time series features, convert the weighted time series features into input tensors suitable for three-dimensional deconvolution operations, and obtain initial input features;

[0120] Based on the initial input features, perform the first 3D deconvolution operation with a convolution kernel size of 3x3x3, a step size of 2, and a padding of 1 to obtain the first layer feature map. Apply 3D instance normalization and ReLU activation function to the first layer feature map to obtain the first layer activation features.

[0121] The first layer activation features are input into the second 3D deconvolution layer with a convolution kernel size of 3x3x3, a step size of 2, and a padding of 1 to obtain the second layer feature map. Three-dimensional instance normalization and ReLU activation function are applied to the second layer feature map to obtain the second layer activation features.

[0122] Perform residual connection on the second layer activation features, add the second layer activation features to the first layer activation features after 1x1x1 convolution dimension reduction, and obtain fusion features;

[0123] The fused features are input into the third 3D deconvolution layer with a convolution kernel size of 3x3x3, a step size of 1, and a padding of 1 to obtain the third-layer feature map. Three-dimensional instance normalization and ReLU activation function are applied to the third-layer feature map to obtain the third-layer activation features.

[0124] The spatial attention mechanism is used to process the activation features of the third layer, and the spatial weights are calculated and weighted to obtain spatial enhanced features;

[0125] The spatial enhancement features are input into the fourth 3D deconvolution layer with a convolution kernel size of 3x3x3, a step size of 1, and a padding of 1 to obtain the final feature map. The Tanh activation function is applied to the final feature map to obtain the dose distribution prediction result.

[0126] The dose distribution prediction results are smoothed by three-dimensional median filtering, and the predicted values ​​are mapped to the preset dose range through thresholding operation to obtain real-time dose distribution prediction data.

[0127] Specifically, the weighted time series features are reshaped to convert them into input tensors suitable for three-dimensional deconvolution operations. The one-dimensional or two-dimensional time series features are converted into a three-dimensional spatial representation, and the generated initial input features T(x, y, z) contain the dose distribution data for each time step. Based on this initial input feature, the first 3D deconvolution operation is performed. Through the upsampling process, the spatial resolution of the input tensor is gradually restored to simulate the dose distribution in the treatment area. A configuration with a convolution kernel size of 3×3×3, a step size of 2, and a padding of 1 is selected to ensure that local features in the three-dimensional space can be captured and the input tensor can be appropriately expanded. The calculation formula for deconvolution is:

[0128] O i,j,k =∑ m,n,p X i+m,j+n,k+p ·W m,n,p +b;

[0129] Among them, O i,j,k is the value of the output tensor at position i, j, k, X i,j,k is the input feature volume, W m,n,pis the weight of the convolution kernel, and b is the bias term. Through this operation, the first layer feature map is obtained. In order to normalize the feature distribution, 3D instance normalization and ReLU activation function are applied to the first layer feature map. Instance normalization can eliminate the distribution differences between different input samples and ensure the stability of model training, while ReLU activation function can introduce nonlinear features to the network, improve the expressiveness of the model, and generate the first layer activation features. The first layer activation features are input to the second 3D deconvolution layer, and the same convolution kernel size and step size configuration are used to continue upsampling the features. The second deconvolution operation further expands the size of the feature map to gradually approach the spatial resolution of the original input data. After the deconvolution operation, 3D instance normalization and ReLU activation function are applied to the output feature map to generate the second layer activation features. Residual connection is performed on the second layer activation features. The information of the first layer activation features is fused with the second layer activation features to prevent information loss during multiple convolutions. To achieve residual connection, the first layer activation features are reduced in dimension by 1×1×1 convolution operation, and then added to the second layer activation features to obtain fused features. This structure allows the network to retain early key information at a deeper level and improve the expressiveness of features. The fused features are input into the third 3D deconvolution layer, where the convolution kernel size is also 3×3×3, the step size is 1, and the padding is 1. This operation does not change the spatial resolution of the feature map, but extracts finer-grained features through convolution. After the deconvolution operation, three-dimensional instance normalization and ReLU activation function are applied to the third layer feature map to generate the third layer activation features. In order to improve the spatial sensitivity of the model, the spatial attention mechanism is applied to the third layer activation features. Different weights are assigned to different regions according to the importance of spatial position, so that the model can pay more attention to the spatial regions that are important for dose distribution prediction. Assuming that the activation feature of the third layer is F(x,y,z), the spatial attention weight A(x,y,z) is calculated by the following formula:

[0130] A(x,y,z)=softmax(W2·ReLU(W1·F(x,y,z)));

[0131] Among them, W1 and W2 are trainable weight matrices. The original features F(x,y,z) are weighted by the calculated spatial weight A(x,y,z) to obtain spatial enhancement features. The spatial enhancement features are input into the fourth 3D deconvolution layer for final processing, with a convolution kernel size of 3×3×3, a step size of 1, and a padding of 1. The goal of the deconvolution operation is to generate a predicted dose distribution feature map, which reflects the spatial dose distribution of the treatment area. In order to ensure that the output value is within a reasonable range, the Tanh activation function is applied to the final feature map, which limits the output value to the range of [-1, 1] to generate the dose distribution prediction result. The dose distribution prediction results are processed to improve the smoothness and accuracy of the prediction. The prediction results are smoothed by three-dimensional median filtering. Median filtering can effectively remove noise in the prediction results and ensure that the generated dose distribution image is more continuous and smooth in space. The predicted value is mapped to the preset dose range through the thresholding operation to ensure that the final output dose distribution meets the requirements of radiotherapy. For example, if the standard range of radiation dose is [0, 1], the negative values ​​in the prediction results are set to 0 through the thresholding operation, and the values ​​exceeding 1 are limited to 1, so as to obtain the final real-time dose distribution prediction data.

[0132] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0133] Perform state encoding on the real-time dose distribution prediction data and the original control parameters of the radiotherapy equipment to obtain the current state vector;

[0134] Constructing an action space based on the current state vector, discretizing the adjustable parameters of the radiotherapy equipment into multiple adjustment options, and obtaining an action set;

[0135] Apply the online Q network to evaluate each action in the action set, calculate the Q value of each action, and obtain the action value sequence;

[0136] Based on the action value sequence, the ε-greedy strategy is used to select the optimal action, obtain the current optimal adjustment plan, and apply the current optimal adjustment plan to the target Q network, calculate the target Q value, and obtain the target action value;

[0137] The temporal difference error is calculated based on the target action value and the action value sequence to obtain the loss function value, and the loss function value is used to update the parameters of the online Q network to obtain the optimized online Q network;

[0138] At every preset update cycle, the parameters of the optimized online Q network are copied to the target Q network to obtain an updated target Q network, and the target control parameters of the radiotherapy equipment are output based on the updated target Q network.

[0139] Specifically, the real-time dose distribution prediction data and the original control parameters of the radiotherapy equipment are state-encoded to generate the current state vector. The real-time dose distribution prediction data contains the dose distribution information in the treatment area, such as the dose distribution of the tumor and surrounding healthy tissues, while the original control parameters of the radiotherapy equipment include dose intensity, radiotherapy angle, irradiation time, etc. By integrating this information into a vector representation, the current state vector S t , reflecting the current treatment scenario. Based on the state vector S t , construct an action space. The action space defines the adjustment operations that the radiotherapy device can perform at each time step, which correspond to changes in the device control parameters. In order to make the action space easier to calculate, the adjustable parameters of the radiotherapy device are discretized into multiple adjustment options. For example, for dose intensity, it is discretized into several levels; for radiotherapy angle, it is divided into several discrete angle values. Through the discretization operation, an action set A = {a1, a2, ..., a n}, where each action a i Corresponds to a specific adjustment of a parameter of the device. Each action in the action set is evaluated by an online Q network. The Q network is a deep reinforcement learning network that learns to predict the long-term reward value of choosing different actions in the current state. For each action a i , the Q network calculates its corresponding Q value Q(S t ,a i ), that is, in the current state S t Next, perform action a i The expected return after treatment. The return is an evaluation indicator of the treatment effect, such as the dose coverage of the target area, the dose control of healthy tissues, etc. By evaluating all actions, the Q network generates an action value sequence {Q(S t ,a1),Q(S t ,a2),...,Q(S t ,a n )}, that is, the Q value corresponding to each action. Based on the action value sequence, the ∈-greedy strategy is used to select the current optimal action. The ∈-greedy strategy is a strategy that balances exploration and utilization, that is, under a certain probability ∈, the system randomly selects an action for exploration to prevent falling into a local optimal solution; and under the remaining probability 1-∈, the system selects the action with the highest current Q value for utilization. Through this strategy, in the learning process, the known best strategy can be effectively utilized, and continuous exploration can be carried out to improve the strategy. The selected action is the current optimal adjustment plan a *, represents the optimal device adjustment operation selected by the system based on the current state. After determining the optimal adjustment plan, the plan is applied to the target Q network to calculate the target Q value Q′. The target Q network is a model independent of the online Q network, which is used to generate stable target values ​​to help the system train. The calculation of the target Q value follows the Bellman equation in reinforcement learning, and the formula is as follows:

[0140]

[0141] Among them, r is the immediate return after executing the action in the current state, γ is the discount factor, which represents the influence weight S of future returns on the current decision t+1 is the new state after executing the action, max a′ Q(S t+1 ,a′) is the Q value of the best action in the next state. The temporal difference error δ is calculated based on the target Q value and the action value sequence, and the formula is as follows:

[0142] δ=Q′(S t ,a * )-Q(S t ,a * );

[0143] The temporal difference error represents the difference between the Q value predicted by the current Q network and the target Q value. The system uses this error to adjust the parameters of the online Q network. The loss function L is constructed using the temporal difference error, and the parameters of the Q network are updated by the gradient descent method to minimize the loss function value L, thereby optimizing the performance of the online Q network. At every preset update cycle, the parameters of the optimized online Q network are copied to the target Q network. The stability of the target Q network is maintained, and the latest strategies of the online Q network are gradually applied to the target Q network to improve the prediction accuracy of the Q value. Through the synchronization mechanism, the system can maintain a reasonable estimate of future returns during the update process, thereby stabilizing the learning process. Based on the updated target Q network, the target control parameters of the radiotherapy equipment are output, including dose intensity, irradiation angle, irradiation time, etc.

[0144] The above describes the real-time dose monitoring and adjustment method for tumor radiotherapy in the embodiment of the present invention. The following describes the real-time dose monitoring and adjustment system for tumor radiotherapy in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a real-time dose monitoring and adjustment system for tumor radiotherapy includes:

[0145] The acquisition module 201 is used to collect real-time dose data of the treatment area, obtain surface dose distribution data and pH value change data, and identify the patient's respiratory motion characteristics to generate a respiratory motion feature set;

[0146] A construction module 202 is used to divide the treatment area into four-dimensional space and time, construct a dynamic network structure with space and time blocks as nodes, and obtain a dynamic dose distribution network;

[0147] An extraction module 203 is used to extract and integrate the surface dose distribution data and pH value change data according to the dynamic dose distribution network to obtain a fused distribution feature set;

[0148] The prediction module 204 is used to input the fusion distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data;

[0149] The adjustment module 205 is used to dynamically adjust the original control parameters of the radiotherapy equipment according to the real-time dose distribution prediction data to obtain the target control parameters of the radiotherapy equipment.

[0150] Through the collaboration of the above components, by collecting surface dose distribution data, pH value change data and respiratory motion characteristics, the effective integration of multi-source heterogeneous data is achieved, providing more comprehensive treatment area information. The four-dimensional spatiotemporal division and dynamic network structure can better capture the spatiotemporal change characteristics of the treatment area, and improve the accuracy and spatiotemporal resolution of dose distribution prediction. Combining the hybrid architecture of 3DCNN, GCN and GRU, the spatial, graph structure and time series information are fully utilized, and the accuracy and generalization ability of dose distribution prediction are significantly improved. The reinforcement learning algorithm based on the double Q network realizes the intelligent real-time adjustment of radiotherapy equipment parameters, effectively coping with the uncertainty factors in the treatment process. The real-time monitoring of dose and physiological parameters is achieved through external sensors, avoiding the potential risks of traditional invasive methods, and improving the patient's comfort and treatment compliance. Taking into account the patient's respiratory motion characteristics, more accurate individualized dose distribution prediction and adjustment are achieved, which helps to improve the treatment effect and reduce side effects. The multi-level feature extraction and optimization strategy is adopted to improve the computational efficiency of the algorithm while ensuring the prediction accuracy, meeting the needs of real-time monitoring and adjustment.

[0151] above Figure 2 The real-time dose monitoring and adjustment system for tumor radiotherapy in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The computer device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0152] Figure 33 is a schematic diagram of the structure of a computer device provided by an embodiment of the present invention. The computer device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the computer device 300. Furthermore, the processor 310 can be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the computer device 300 to implement the steps of the above-mentioned real-time dose monitoring and adjustment method for tumor radiotherapy.

[0153] The computer device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The illustrated computer device structure does not constitute a limitation on the computer device provided by the present invention, and may include more or fewer components than illustrated, or a combination of certain components, or a different arrangement of components.

[0154] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the real-time dose monitoring and adjustment method for tumor radiotherapy.

[0155] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0156] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0157] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time dose monitoring and adjustment method for tumor radiotherapy, characterized in that: The method comprises: Collect real-time dose data of the treatment area to obtain surface dose distribution data and pH value change data, and identify the patient's respiratory motion characteristics to generate a respiratory motion feature set; Performing four-dimensional space-time division on the treatment area, constructing a dynamic network structure with space-time blocks as nodes, and obtaining a dynamic dose distribution network; According to the dynamic dose distribution network, feature extraction and integration are performed on the surface dose distribution data and the pH value change data to obtain a fused distribution feature set; Inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data; According to the real-time dose distribution prediction data, the original control parameters of the radiotherapy equipment are dynamically adjusted to obtain the target control parameters of the radiotherapy equipment.

2. The real-time dose monitoring and adjustment method for tumor radiotherapy according to claim 1, characterized in that: The real-time dose data collection of the treatment area is performed to obtain the surface dose distribution data and pH value change data, and the respiratory motion feature recognition of the patient is performed to generate a respiratory motion feature set, including: Arranging a high-precision optical fiber array detector in the treatment area to collect raw optical signal data, and performing Fourier transform on the raw optical signal data to obtain frequency domain feature data; Perform signal restoration based on the frequency domain characteristic data to obtain surface dose distribution data; Arranging a pH-sensitive optical sensor on the treatment area, collecting original pH spectrum data, and performing principal component analysis on the original pH spectrum data to obtain pH value change data; Perform body surface image segmentation on the patient, identify the breathing-related region of interest, and obtain the breathing region mask; Calculating the displacement field between consecutive frames of the respiratory region mask to obtain a respiratory motion vector field; Wavelet transformation and time-frequency feature extraction are performed based on the respiratory motion vector field to obtain a respiratory motion feature set.

3. The real-time dose monitoring and adjustment method for tumor radiotherapy according to claim 1, characterized in that: The four-dimensional space-time division of the treatment area is performed to construct a dynamic network structure with space-time blocks as nodes to obtain a dynamic dose distribution network, including: Performing three-dimensional spatial gridding processing on the treatment area to obtain spatial grid data, and performing time dimension expansion on the spatial grid data to obtain a four-dimensional space-time data set; Based on the four-dimensional spatiotemporal data set, block division is performed to obtain a spatiotemporal block set, and each block in the spatiotemporal block set is set as a network node to construct an initial network structure; Performing spatial adjacency analysis on the nodes in the initial network structure to obtain spatial correlation data, and performing time series correlation analysis on the nodes in the initial network structure to obtain time correlation data; Based on the spatial correlation data and the temporal correlation data, the comprehensive correlation strength between nodes is calculated, and the initial network structure is optimized according to the comprehensive correlation strength to obtain a dynamic dose distribution network.

4. The real-time dose monitoring and adjustment method for tumor radiotherapy according to claim 1, characterized in that: The feature extraction and integration of the surface dose distribution data and the pH value change data are performed according to the dynamic dose distribution network to obtain a fused distribution feature set, including: Extracting network structure features of the dynamic dose distribution network to obtain a network feature vector, and performing spatial attention weighting on the surface dose distribution data based on the network feature vector to obtain weighted surface dose distribution data; Performing multi-level spatial feature extraction on the weighted surface dose distribution data to obtain a surface dose feature set; Performing time-attention weighting on the pH value change data based on the network feature vector to obtain weighted pH value change data, and performing time series feature extraction on the weighted pH value change data to obtain a pH value feature set; Performing feature cascading on the surface dose feature set and the pH value feature set to obtain an initial fused feature set, and performing feature dimension reduction and denoising processing on the initial fused feature set to obtain a compressed feature set; A feature map is constructed based on the compressed feature set, and a non-local attention mechanism is applied to perform feature optimization to obtain a fused distributed feature set.

5. The real-time dose monitoring and adjustment method for tumor radiotherapy according to claim 1, characterized in that: The step of inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data includes: Performing three-dimensional reconstruction on the fused distribution feature set to obtain feature volume data, and performing 3D space transformation network registration on the feature volume data to obtain an aligned feature volume; Inputting the aligned feature volume into a three-dimensional convolutional neural network module, the three-dimensional convolutional neural network module comprises five 3D convolutional layers, each layer is followed by a 3D batch normalization and a LeakyReLU activation function, and a 3D maximum pooling layer is added after the third and fifth layers to obtain a spatial feature map; Reshaping the spatial feature map into a graph structure and inputting the graph convolution layer to extract graph features to obtain a graph feature representation, wherein the graph convolution layer includes three graph convolution operations, each operation is implemented using Chebyshev polynomial approximation, and graph batch normalization and ReLU activation function are applied after each graph convolution operation; Performing global average pooling and global maximum pooling on the graph feature representation, and concatenating the pooling results to obtain a graph-level feature vector; Inputting the respiratory motion feature set into a multi-layer perceptron for feature transformation to obtain a respiratory feature vector, and concatenating the graph-level feature vector and the respiratory feature vector to obtain a fused feature vector; Expanding the fused feature vector by time step and inputting it into a bidirectional GRU network to obtain a temporal feature sequence, wherein the bidirectional GRU network comprises two layers of bidirectional GRU units, each layer having 128 hidden units; Performing a self-attention mechanism analysis on the time series feature sequence, calculating the correlation between different time steps, and obtaining a weighted time series feature; The weighted temporal features are input into a three-dimensional deconvolution network for dose distribution prediction to obtain real-time dose distribution prediction data. The three-dimensional deconvolution network includes four 3D deconvolution layers, each layer is followed by three-dimensional instance normalization and ReLU activation function, and the last layer uses Tanh activation function.

6. The real-time dose monitoring and adjustment method for tumor radiotherapy according to claim 5, characterized in that: The weighted temporal features are input into a three-dimensional deconvolution network for dose distribution prediction to obtain real-time dose distribution prediction data. The three-dimensional deconvolution network includes four 3D deconvolution layers, each layer is followed by three-dimensional instance normalization and ReLU activation function, and the last layer uses Tanh activation function, including: Reconstructing the shape of the weighted time series features, converting the weighted time series features into an input tensor suitable for a three-dimensional deconvolution operation, and obtaining initial input features; Based on the initial input features, perform a first 3D deconvolution operation with a convolution kernel size of 3×3×3, a step size of 2, and a padding of 1 to obtain a first layer feature map, and apply three-dimensional instance normalization and a ReLU activation function to the first layer feature map to obtain a first layer activation feature; The first layer of activated features is input into the second 3D deconvolution layer with a convolution kernel size of 3×3×3, a step size of 2, and a padding of 1 to obtain the second layer of feature maps. Three-dimensional instance normalization and ReLU activation function are applied to the second layer of feature maps to obtain the second layer of activated features. Perform residual connection on the second layer activation features, add the second layer activation features to the first layer activation features after dimensionality reduction by 1×1×1 convolution, and obtain fusion features; Input the fused features into the third 3D deconvolution layer, with a convolution kernel size of 3×3×3, a step size of 1, and a padding of 1, to obtain a third-layer feature map, and apply three-dimensional instance normalization and a ReLU activation function to the third-layer feature map to obtain a third-layer activation feature; Performing spatial attention mechanism processing on the activation features of the third layer, calculating spatial weights and weighting them to obtain spatial enhancement features; The spatial enhancement feature is input into the fourth 3D deconvolution layer, the convolution kernel size is 3×3×3, the step size is 1, and the padding is 1, to obtain the final feature map, and the Tanh activation function is applied to the final feature map to obtain the dose distribution prediction result; The dose distribution prediction result is smoothed by three-dimensional median filtering, and the prediction value is mapped into a preset dose range by thresholding operation to obtain real-time dose distribution prediction data.

7. The real-time dose monitoring and adjustment method for tumor radiotherapy according to claim 1, characterized in that: The method of dynamically adjusting the original control parameters of the radiotherapy device according to the real-time dose distribution prediction data to obtain the target control parameters of the radiotherapy device includes: Performing state encoding on the real-time dose distribution prediction data and the original control parameters of the radiotherapy device to obtain a current state vector; Constructing an action space based on the current state vector, discretizing the adjustable parameters of the radiotherapy device into multiple adjustment options, and obtaining an action set; Applying an online Q network to evaluate each action in the action set, calculating the Q value of each action, and obtaining an action value sequence; Based on the action value sequence, an ε-greedy strategy is used to select the optimal action to obtain a current optimal adjustment scheme, and the current optimal adjustment scheme is applied to the target Q network to calculate the target Q value to obtain the target action value; Calculating a temporal difference error based on the target action value and the action value sequence to obtain a loss function value, and using the loss function value to update parameters of the online Q network to obtain an optimized online Q network; At every preset update cycle, the parameters of the optimized online Q network are copied to the target Q network to obtain an updated target Q network, and the target control parameters of the radiotherapy device are output based on the updated target Q network.

8. A real-time dose monitoring and adjustment system for tumor radiotherapy, characterized in that: A real-time dose monitoring and adjustment method for performing tumor radiotherapy according to any one of claims 1 to 7, the system comprising: The acquisition module is used to collect real-time dose data of the treatment area, obtain surface dose distribution data and pH value change data, and identify the patient's respiratory motion characteristics to generate a respiratory motion feature set; A construction module is used to divide the treatment area into four-dimensional space and time, construct a dynamic network structure with space and time blocks as nodes, and obtain a dynamic dose distribution network; An extraction module, used for performing feature extraction and integration on the surface dose distribution data and the pH value change data according to the dynamic dose distribution network to obtain a fused distribution feature set; A prediction module, used for inputting the fused distribution feature set and the respiratory motion feature set into a preset dose distribution prediction model to perform real-time dose distribution prediction to obtain real-time dose distribution prediction data; The adjustment module is used to dynamically adjust the original control parameters of the radiotherapy equipment according to the real-time dose distribution prediction data to obtain the target control parameters of the radiotherapy equipment.

9. A computer device, characterized in that: The computer device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the computer device to execute the real-time dose monitoring and adjustment method for tumor radiotherapy according to any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the real-time dose monitoring and adjustment method for tumor radiotherapy according to any one of claims 1 to 7 is implemented.

Images