Adaptive radiotherapy dose optimization method and device, and storage medium

Abstract

The invention relates to a self-adaptive radiotherapy dose optimization method and device and a storage medium, which are used for optimizing dose distribution in radiotherapy. Through combination of deep reinforcement learning (DRL) and a heuristic algorithm (a particle swarm optimization algorithm), the method maximizes the dose of a tumor area, and protects surrounding normal tissues at the same time. Firstly, an anatomical structure and a tumor area of a patient are extracted through medical image data, and initial dose distribution is generated; then, the deep reinforcement learning agent adjusts dose distribution according to the dynamic anatomical change of the patient, a heuristic algorithm is further optimized, and a local optimal solution is avoided; real-time feedback in the treatment process is used for adjusting dose distribution, and accuracy and timeliness of a treatment scheme are ensured. Finally, the generated optimal dose distribution scheme is provided for a clinician to refer, dose maximization of the tumor area is ensured, and normal tissue is protected. According to the method, the precision and effect of radiotherapy are improved, and a personalized treatment scheme is provided.

Description

Technical Field

[0001] The present invention relates to radiotherapy technology, and in particular to an adaptive radiotherapy dose optimization method, device and storage medium. The method is applied in the process of radiotherapy, and aims to improve the tumor treatment effect by optimizing the radiotherapy dose distribution, while protecting normal tissue to the maximum extent. In particular, in adaptive radiotherapy, it can respond to changes in the patient's anatomical structure in real time, and improve the accuracy and safety of treatment. Background Art

[0002] Radiation therapy is one of the important means of treating cancer. It uses high-energy rays to accurately irradiate the tumor area to kill tumor cells and inhibit tumor growth. In order to ensure the treatment effect and reduce radiation damage to normal tissues, the dose must be accurately optimized during radiotherapy. The goal of dose optimization is to maximize the concentration of radiation dose in the tumor area and control the dose of surrounding normal tissues within a safe range. Existing technologies include

[0003] Existing methods for optimizing radiation therapy doses mainly rely on traditional optimization algorithms, including linear programming, nonlinear programming, and gradient descent methods. These methods have defects such as low computational efficiency and local optimal solution problems when dealing with complex dose distributions. In addition, although heuristic algorithms (including genetic algorithms and particle swarm optimization) can effectively avoid local optimal problems, they are difficult to meet the real-time requirements in clinical treatment due to large consumption of computing resources, slow convergence speed, and complex parameter adjustment. In adaptive radiation therapy, existing technologies cannot adapt to patient anatomical changes in a timely manner and cannot quickly update dose distribution during treatment.

[0004] Existing technologies face many challenges in radiotherapy, including low computational efficiency, inability to guarantee a global optimal solution, and difficulty adapting to dynamic changes in the patient's anatomical structure. With the development of adaptive radiotherapy, the need to adjust the treatment plan in real time has become more urgent. In order to solve these problems, a hybrid optimization algorithm that can combine multiple optimization strategies is urgently needed to improve computational efficiency, respond to patient anatomical changes in real time during treatment, and provide a globally optimal dose distribution, thereby improving treatment efficacy and reducing side effects. Summary of the invention

[0005] The present invention relates to an adaptive radiation therapy dose optimization method, device and storage medium, aiming to solve the problems of low computational efficiency, difficulty in ensuring the global optimal solution and poor adaptability to anatomical structure changes in existing radiation therapy dose optimization technologies. The present invention provides an efficient and accurate dose planning method that can combine multiple optimization strategies to achieve real-time and dynamic adaptive adjustment while ensuring the treatment effect.

[0006] To achieve the above object, the present invention proposes the following technical solution: an adaptive radiation therapy dose optimization method, the method comprising:

[0007] S1, collect multimodal medical images of patients;

[0008] S2. Use deep learning algorithms to fuse multimodal images and extract anatomical features of the tumor area and surrounding normal tissues to provide basic information for subsequent dose optimization;

[0009] S3, using linear programming to generate preliminary dose distribution D0 based on the key features of the patient's anatomy and tumor location;

[0010] S4. Adopt hybrid optimization strategy to optimize the preliminary dose distribution;

[0011] S5. Finally, the optimized dose distribution is generated and the treatment plan is output for clinicians' reference to ensure that the dose in the tumor area is maximized while protecting healthy tissues.

[0012] Further, the multimodal images include computed tomography (CT), magnetic resonance imaging (MRI), and cone beam computed tomography (CBCT) images;

[0013] The process of extracting the anatomical features of the tumor area and surrounding normal tissues is as follows:

[0014] S21, perform mean filtering and denoising processing on the original image to eliminate the noise generated during the scanning process and improve the quality of the original image. The specific formula is:

[0015]

[0016] Where I(i) represents the value of the i-th pixel, n is the number of pixels in the filter, and I d represents the denoised image;

[0017] S22, segmenting the tumor area from the normal tissue area, and extracting the boundary and contour of the tumor, thereby determining the tumor area; the segmentation is to segment the tumor area by comparing the pixel value with the threshold T tumor and normal tissue T normal , the specific formula is:

[0018]

[0019] Among them, T tumor is the binary image of the tumor area, I d (x,y) is the image I d The pixel value at position (x, y), 1 represents the tumor area, and 0 represents the background;

[0020] The method for extracting the boundary and contour of the tumor is to use the Canny operator to extract the boundary and contour of the tumor, and the Canny operator extracts the edge by calculating the gradient and performing non-maximum suppression;

[0021] S23, expanding the tumor area of ​​the image through a dilation operation, and removing background noise of the image through an erosion operation;

[0022] The specific formula of the expansion operation is: Among them, ⊕ represents the expansion operation, and B is the structural element;

[0023] The specific formula of the corrosion operation is: in represents the corrosion operation, and B is the structural element;

[0024] S24, tumor region morphological feature extraction, including extracting volume and shape features of the tumor region,

[0025] The volume of the tumor area is obtained by calculating the number of pixels in the tumor area and the resolution of the image. The specific formula is:

[0026]

[0027] Among them, V tumor is the tumor volume, T tumor (x, y) is the pixel of the tumor area, ΔV is the volume of each pixel;

[0028] The shape characteristics of the tumor region include the tumor boundary curvature and aspect ratio,

[0029] The boundary curvature can be obtained by calculating the change rate of the edge of the tumor area, and the formula is:

[0030]

[0031] in is the curvature, x and y are the coordinates on the boundary of the tumor region;

[0032] S25, extracting normal tissue density features, the specific formula for calculating the normal tissue density features is:

[0033]

[0034] Among them, ρ normal is the average density of the normal tissue area, N is the number of pixels in the area, T normal is the gray value of the corresponding pixel in the normal tissue area.

[0035] Furthermore, the fusion of the multimodal images is to fuse the CT, MRI, and PET images using image registration and data fusion methods;

[0036] The goal of the image registration is to maximize MI, and the specific formula of MI is:

[0037]

[0038] Where p(I1(i), I2(j)) is the joint probability distribution of images I1 and I2, and p(I1(i)) and p(I2(j)) are the marginal probability distributions of images I1 and I2, respectively;

[0039] The data fusion is to weighted average the pixel values ​​of different modal images through weighted fusion method. The specific operation formula is:

[0040] I fused (x, y)=w1I1(x, y)+w2I2(x, y)+…+w n I n (x, y)

[0041] Among them, I fused (x, y) is the fused image, I1, I2, ..., I n are images of different modalities, w1, w2, …, w n is the corresponding weight.

[0042] Furthermore, the linear programming is to maximize the dose in the tumor area and minimize the dose in the normal tissue area to generate a preliminary dose distribution D0, which is expressed as follows:

[0043]

[0044] Among them, D i and D j are the dose distributions in the tumor and normal tissue regions, respectively, f1 and f2 are the objective functions of tumor maximization and normal tissue minimization, and λ is the weight coefficient of normal tissue;

[0045] The minimum dose constraint of the tumor area is:

[0046] D i ≥D min for all i∈Tumor

[0047] The maximum dose constraint for normal tissue area is:

[0048] D j ≤D max for all j∈Normal

[0049] The sum of the dose to the tumor area and the dose to the normal tissue area does not exceed the set dose threshold.

[0050] Furthermore, the deep reinforcement learning optimization process in the hybrid optimization strategy uses a deep neural network to model the anatomical structure and dose distribution, and optimizes the dose allocation strategy based on reward feedback;

[0051] Heuristic algorithm optimization uses particle swarm optimization algorithm to adjust the dose distribution and ensure the global optimal solution.

[0052] Furthermore, the specific steps of the hybrid optimization strategy are:

[0053] S41. Perform deep reinforcement learning optimization on the preliminary dose distribution D0; use a deep reinforcement learning agent to optimize the dose distribution in the tumor area by continuously adjusting the dose distribution according to the dynamic changes of the patient's anatomical structure and tumor area, and perform reinforcement learning optimization based on reverse rewards;

[0054] The deep Q network (DQN) is used to train the agent, and the formula for updating the Q value is:

[0055]

[0056] Among them, S t is the current state, A t is the currently selected dose dispensing action, R t+1 is the reward after the current action, α is the learning rate, and γ is the discount factor;

[0057] S42, optimize the preliminary dose distribution D0 using a heuristic algorithm; combine the particle swarm optimization (PSO) algorithm to globally optimize the preliminary dose distribution, avoid local optimal solutions, further optimize the dose distribution, and ensure global optimality. The update formula of particle swarm optimization is:

[0058] v i (t+1)=wv i (t)+c1r1(p best -x i (t))+c2r2(g best -x i (t))

[0059] x i (t+1)=x i (t)+v i (t+1)

[0060] where v i (t) is the velocity of particle i at time t, x i (t) is the position of the particle, p best and g best are the personal optimal position and global optimal position of the particle, r1 and r2 are random numbers, and c1 and c2 are learning factors;

[0061] S43. According to the patient's feedback during the treatment process, the optimization strategy is adjusted in real time to ensure the accuracy and timeliness of the treatment plan. The feedback formula for adjustment is:

[0062] D adjusted =D current +β(R feedback -D current )

[0063] Among them, D adjusted is the dose distribution after feedback adjustment, D current is the current dose distribution, R feedback is the treatment feedback information, and β is the adjustment factor.

[0064] An adaptive radiation therapy dose optimization device, the device comprising:

[0065] Data input module: used to receive patient images and transmit them to the processing module;

[0066] Data processing module: including deep learning algorithm module and heuristic optimization algorithm module, which are used for data feature extraction, dose calculation and optimization respectively;

[0067] Optimization calculation module: Combines deep reinforcement learning and heuristic algorithms to further optimize the preliminary dose distribution and generate the optimal dose distribution plan;

[0068] Feedback module: used to provide feedback and optimize results based on real-time images and patient anatomical changes;

[0069] Output module: used to generate the final treatment plan and output it for execution by radiotherapy equipment.

[0070] An adaptive radiation therapy dose optimization device, the device comprising a processor and a memory:

[0071] The memory is used to store program code and transmit the program code to the processor;

[0072] The processor is used to execute the adaptive radiotherapy dose optimization method according to the instructions in the program code.

[0073] A computer-readable storage medium, characterized in that the computer-readable storage medium is used to store program code, and the program code is used to execute the adaptive radiotherapy dose optimization method.

[0074] Compared with the prior art, the advantages of the present invention are:

[0075] By combining deep learning with heuristic optimization algorithms, the present invention can improve the computational efficiency of dose planning and avoid the problems of excessive computation and local optimal solutions in traditional methods. This method can adjust the treatment plan in real time when the patient's anatomical structure changes, and can also provide a globally optimal dose distribution. In addition, through a real-time feedback mechanism, the treatment plan can be dynamically adjusted according to the patient's actual situation, providing guidance for professional medical staff. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0077] Figure 1 is a flow chart of the method of the present invention;

[0078] Figure 2 It is a schematic diagram of the hybrid optimization strategy process of the present invention;

[0079] Figure 3 It is a schematic diagram of the multimodal data fusion and feature extraction process of the present invention. DETAILED DESCRIPTION

[0080] The specific implementation of the present invention is an adaptive radiation therapy dose optimization method, device and storage medium, which aims to optimize the dose distribution during radiation therapy by combining deep reinforcement learning (DRL) and heuristic algorithm (particle swarm optimization), improve the treatment effect, and protect the patient's healthy tissue. The technical steps and processing flow involved in this implementation are as follows: Figure 1 As shown:

[0081] S1, collect multimodal medical images of patients;

[0082] The multimodal images include computed tomography (CT), magnetic resonance imaging (MRI), and cone beam computed tomography (CBCT) images;

[0083] S2. Use deep learning algorithms to fuse multimodal images and extract the anatomical features of the patient's tumor area and surrounding normal tissues to provide basic information for subsequent dose optimization;

[0084] The fusion of the multimodal images is to fuse the CT, MRI and PET images by using image registration and data fusion methods;

[0085] The goal of the image registration is to maximize MI, and the specific formula of MI is:

[0086]

[0087] Where p(I1(i), I2(j)) is the joint probability distribution of images I1 and I2, and p(I1(i)) and p(I2(j)) are the marginal probability distributions of images I1 and I2, respectively;

[0088] The data fusion is to weighted average the pixel values ​​of different modal images through weighted fusion method. The specific operation formula is:

[0089] I fused (x, y)=w1I1(x, y)+w2I2(x, y)+…+w n I n (x, y)

[0090] Among them, I fused (x, y) is the fused image, I1, I2, ..., I n are images of different modalities, w1, w2, ..., w n is the corresponding weight.

[0091] The process of extracting the anatomical features of the tumor area and surrounding normal tissues is as follows:

[0092] S21, perform mean filtering and denoising processing on the original image to eliminate the noise generated during the scanning process and improve the quality of the original image. The specific formula is:

[0093]

[0094] Where I(i) represents the value of the i-th pixel, n is the number of pixels in the filter, and I d represents the denoised image;

[0095] S22, segmenting the tumor area from the normal tissue area, and extracting the boundary and contour of the tumor, thereby determining the tumor area; the segmentation is to segment the tumor area by comparing the pixel value with the threshold T tumor and normal tissue T normal , the specific formula is:

[0096]

[0097] Among them, T tumor is the binary image of the tumor area, I d (x,y) is the image I d The pixel value at position (x, y), 1 represents the tumor area, and 0 represents the background;

[0098] The method for extracting the boundary and contour of the tumor is to use the Canny operator to extract the boundary and contour of the tumor, and the Canny operator extracts the edge by calculating the gradient and performing non-maximum suppression;

[0099] S23, expanding the tumor area of ​​the image through a dilation operation, and removing background noise of the image through an erosion operation;

[0100] The specific formula of the expansion operation is: Among them, ⊕ represents the expansion operation, and B is the structural element;

[0101] The specific formula of the corrosion operation is: in represents the corrosion operation, and B is the structural element;

[0102] S24, tumor region morphological feature extraction, including extracting volume and shape features of the tumor region,

[0103] The volume of the tumor area is obtained by calculating the number of pixels in the tumor area and the resolution of the image. The specific formula is:

[0104]

[0105] Among them, V tumor is the tumor volume, T tumor (x, y) is the pixel of the tumor area, ΔV is the volume of each pixel;

[0106] The shape characteristics of the tumor region include the tumor boundary curvature and aspect ratio,

[0107] The boundary curvature can be obtained by calculating the change rate of the edge of the tumor area, and the formula is:

[0108]

[0109] in is the curvature, x and y are the coordinates on the boundary of the tumor region;

[0110] S25, extracting normal tissue density features, the specific formula for calculating the normal tissue density features is:

[0111]

[0112] Among them, ρ normal is the average density of the normal tissue area, N is the number of pixels in the area, T normal is the gray value of the corresponding pixel in the normal tissue area.

[0113] S3. Based on the key features of the patient's anatomical structure and the tumor location, a preliminary dose distribution D0 is generated using linear programming; the linear programming is to maximize the dose in the tumor area and minimize the dose in the normal tissue area to generate the preliminary dose distribution D0, which is expressed as follows:

[0114]

[0115] Among them, D i and D j are the dose distributions in the tumor and normal tissue regions, respectively, f1 and f2 are the objective functions of tumor maximization and normal tissue minimization, and λ is the weight coefficient of normal tissue;

[0116] The minimum dose constraint of the tumor area is:

[0117] D i ≥D min for all i∈Tumor

[0118] The maximum dose constraint for normal tissue area is:

[0119] D j ≥D max for all j∈Normal

[0120] The sum of the dose to the tumor area and the dose to the normal tissue area does not exceed the set dose threshold.

[0121] S4. Adopting a hybrid optimization strategy to optimize the preliminary dose distribution; the specific steps of the hybrid optimization strategy are:

[0122] S41. Perform deep reinforcement learning optimization on the preliminary dose distribution D0; use a deep reinforcement learning agent to optimize the dose distribution in the tumor area by continuously adjusting the dose distribution according to the dynamic changes of the patient's anatomical structure and tumor area, and perform reinforcement learning optimization based on reverse rewards;

[0123] The deep Q network (DQN) is used to train the agent, and the formula for updating the Q value is:

[0124]

[0125] Among them, S t is the current state, A t is the currently selected dose dispensing action, R t+1 is the reward after the current action, α is the learning rate, and γ is the discount factor;

[0126] S42, optimize the preliminary dose distribution D0 using a heuristic algorithm; combine the particle swarm optimization (PSO) algorithm to globally optimize the preliminary dose distribution, avoid local optimal solutions, further optimize the dose distribution, and ensure global optimality. The update formula of particle swarm optimization is:

[0127] v i (t+1)=wv i (t)+c1r1(p best -x i (t))+c2r2(g best -x i (t))

[0128] x i (t+1)=x i (t)+v i (t+1)

[0129] where v i (t) is the velocity of particle i at time t, x i (t) is the position of the particle, p best and g best are the personal optimal position and global optimal position of the particle, r1 and r2 are random numbers, and c1 and c2 are learning factors;

[0130] S43. According to the patient's feedback during the treatment process, the optimization strategy is adjusted in real time to ensure the accuracy and timeliness of the treatment plan. The feedback formula for adjustment is:

[0131] D adjusted =D current +β(R feedback -D current )

[0132] Among them, D adjusted is the dose distribution after feedback adjustment, D current is the current dose distribution, R feedback is the treatment feedback information, and β is the adjustment factor.

[0133] S5. Finally, the optimized dose distribution is generated and the treatment plan is output for clinicians' reference to ensure that the dose in the tumor area is maximized while protecting healthy tissues.

[0134] The deep reinforcement learning optimization process in the hybrid optimization strategy uses a deep neural network to model the anatomical structure and dose distribution, and optimizes the dose allocation strategy based on reward feedback;

[0135] The heuristic algorithm optimization adjusts the dose distribution through the particle swarm optimization algorithm to ensure the global optimal solution.

[0136] This embodiment also provides an adaptive radiation therapy dose optimization device, the device comprising:

[0137] Data input module: used to receive patient images and transmit them to the processing module;

[0138] Data processing module: including deep learning algorithm module and heuristic optimization algorithm module, which are used for data feature extraction, dose calculation and optimization respectively;

[0139] Optimization calculation module: Combines deep reinforcement learning and heuristic algorithms to further optimize the preliminary dose distribution and generate the optimal dose distribution plan;

[0140] Feedback module: used to provide feedback and optimize results based on real-time images and patient anatomical changes;

[0141] Output module: used to generate the final treatment plan and output it for execution by radiotherapy equipment.

[0142] This embodiment also provides an adaptive radiation therapy dose optimization device, the device comprising a processor and a memory:

[0143] The memory is used to store program code and transmit the program code to the processor;

[0144] The processor is used to execute the adaptive radiotherapy dose optimization method according to the instructions in the program code.

[0145] This embodiment further proposes a computer-readable storage medium, characterized in that the computer-readable storage medium is used to store program codes, and the program codes are used to execute the adaptive radiation therapy dose optimization method.

[0146] The optimization process of the present invention not only relies on traditional optimization algorithms and deep learning algorithms, but also fully considers the individual differences of patients and the dynamic changes during the treatment process. Through real-time feedback and adjustment mechanisms, the dose distribution can be dynamically optimized according to the anatomical changes of the patient, ensuring that the dose in the tumor area is maximized and normal tissues are protected during the entire treatment process, achieving the best treatment effect.

[0147] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for optimizing dose of adaptive radiotherapy, the method comprising: S1, collect multimodal medical images of patients; S2. Use deep learning algorithms to fuse multimodal images and extract anatomical features of the tumor area and surrounding normal tissues to provide basic information for subsequent dose optimization; S3, using linear programming to generate preliminary dose distribution D0 based on the key features of the patient's anatomy and tumor location; S4. Adopt hybrid optimization strategy to optimize the preliminary dose distribution; ： S5. Finally, the optimized dose distribution is generated and the treatment plan is output for clinicians' reference to ensure that the dose in the tumor area is maximized while protecting healthy tissues.

2. According to claim 1, the method for optimizing the dose of adaptive radiotherapy, the multimodal images include computed tomography (CT), magnetic resonance imaging (MRI) and cone beam computed tomography (CBCT) images; The process of extracting the anatomical features of the tumor area and surrounding normal tissues is as follows: S21, perform mean filtering and denoising processing on the original image to eliminate the noise generated during the scanning process and improve the quality of the original image. The specific formula is: Where I(i) represents the value of the i-th pixel, n is the number of pixels in the filter, and I d represents the denoised image; S22, segmenting the tumor area from the normal tissue area, and extracting the boundary and contour of the tumor, thereby determining the tumor area; the segmentation is to segment the tumor area by comparing the pixel value with the threshold T tumor and normal tissue T normal , the specific formula is: in, T tumor is the binary image of the tumor area, I d (x,y) is the image I d The pixel value at position (x, y), 1 represents the tumor area, and 0 represents the background; The method for extracting the boundary and contour of the tumor is to use the Canny operator to extract the boundary and contour of the tumor, and the Canny operator extracts the edge by calculating the gradient and performing non-maximum suppression; S23, expanding the tumor area of ​​the image through a dilation operation, and removing background noise of the image through an erosion operation; The specific formula of the expansion operation is: in, represents the expansion operation, and B is the structural element; The specific formula of the corrosion operation is: in represents the corrosion operation, and B is the structural element; S24, tumor region morphological feature extraction, including extracting volume and shape features of the tumor region, The volume of the tumor area is obtained by calculating the number of pixels in the tumor area and the resolution of the image. The specific formula is: Among them, V tumor is the tumor volume, T tumor (x, y) is the pixel of the tumor area, ΔV is the volume of each pixel; The shape characteristics of the tumor region include the tumor boundary curvature and aspect ratio, The boundary curvature can be obtained by calculating the change rate of the edge of the tumor area, and the formula is: Where κ is the curvature, x and y are the coordinates on the boundary of the tumor region; S25, extracting normal tissue density features, the specific formula for calculating the normal tissue density features is: Among them, ρ normal is the average density of the normal tissue area, N is the number of pixels in the area, T normal is the gray value of the corresponding pixel in the normal tissue area.

3. According to the adaptive radiation therapy dose optimization method of claim 1, the fusion of the multimodal images is to fuse the CT, MRI, and PET images by using image registration and data fusion methods; The goal of the image registration is to maximize MI, and the specific formula of MI is: Where p(I1(i), I2(j)) is the joint probability distribution of images I1 and I2, and p(I1(i)) and p(I2(j)) are the marginal probability distributions of images I1 and I2, respectively; The data fusion is to weighted average the pixel values ​​of different modal images through weighted fusion method. The specific operation formula is: I fused (x,y)=w1I1(x,y)+w2I2(x,y)+…+w n I n (x,y) in, I fused (x, y) is the fused image, I1, I2, ..., I n are images of different modalities, w1, w2, ..., w n is the corresponding weight.

4. According to the adaptive radiation therapy dose optimization method of claim 1, the linear programming is to maximize the dose of the tumor area and minimize the dose of the normal tissue area to generate a preliminary dose distribution D0, which is expressed as follows: in, D i and D j are the dose distributions in the tumor and normal tissue regions, respectively, f1 and f2 are the objective functions of tumor maximization and normal tissue minimization, and λ is the weight coefficient of normal tissue; The minimum dose constraint of the tumor area is: D i ≥D min for all i∈Tumor The maximum dose constraint for normal tissue area is: D j ≤D max for all j∈Normal The sum of the dose to the tumor area and the dose to the normal tissue area does not exceed the set dose threshold.

5. According to the adaptive radiotherapy dose optimization method of claim 1, the deep reinforcement learning optimization process in the hybrid optimization strategy is to use a deep neural network to model the anatomical structure and dose distribution, and optimize the dose allocation strategy based on reward feedback; Heuristic algorithm optimization uses particle swarm optimization algorithm to adjust the dose distribution and ensure the global optimal solution.

6. According to the adaptive radiation therapy dose optimization method of claim 5, the specific steps of the hybrid optimization strategy are: S41. Perform deep reinforcement learning optimization on the preliminary dose distribution D0; use a deep reinforcement learning agent to optimize the dose distribution in the tumor area by continuously adjusting the dose distribution according to the dynamic changes of the patient's anatomical structure and tumor area, and perform reinforcement learning optimization based on reverse rewards; The deep Q network (DQN) is used to train the agent, and the formula for updating the Q value is: Among them, S t is the current state, A t is the currently selected dose dispensing action, R t+1 is the reward after the current action, α is the learning rate, and γ is the discount factor; S42, optimize the preliminary dose distribution D0 using a heuristic algorithm; combine the particle swarm optimization (PSO) algorithm to globally optimize the preliminary dose distribution, avoid local optimal solutions, further optimize the dose distribution, and ensure global optimality. The update formula of particle swarm optimization is: υ i (t+1)=wv i (t)+c1r1(p best -x i (t))+x2r2(g best -x i (t)) x i (t+1)=x i (t)+v i (t+1) where v i (t) is the velocity of particle i at time t, x i (t) is the position of the particle, p best and g best are the personal optimal position and global optimal position of the particle, r1 and r2 are random numbers, and c1 and c2 are learning factors; S43. According to the patient's feedback during the treatment process, the optimization strategy is adjusted in real time to ensure the accuracy and timeliness of the treatment plan. The feedback formula for adjustment is: D adjusted =D current +β(R feedback -D current ) Among them, D adjusted is the dose distribution after feedback adjustment, D current is the current dose distribution, R feedback is the treatment feedback information, and β is the adjustment factor.

7. An adaptive radiation therapy dose optimization device, the device comprising: Data input module: used to receive patient images and transmit them to the processing module; Data processing module: including deep learning algorithm module and heuristic optimization algorithm module, which are used for data feature extraction, dose calculation and optimization respectively; Optimization calculation module: Combines deep reinforcement learning and heuristic algorithms to further optimize the preliminary dose distribution and generate the optimal dose distribution plan; Feedback module: used to provide feedback and optimize results based on real-time images and patient anatomical changes; Output module: used to generate the final treatment plan and output it for execution by radiotherapy equipment.

8. An adaptive radiation therapy dose optimization device, the device comprising a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the adaptive radiotherapy dose optimization method according to any one of claims 1 to 6 according to the instructions in the program code.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program codes, and the program codes are used to execute the adaptive radiation therapy dose optimization method according to any one of claims 1-6.

Images