Proton intensity modulated therapy radiation field direction automatic planning method, system and device based on deep learning and multi-objective optimization, and storage medium

Abstract

The invention provides a proton intensity modulated therapy radiation field direction automatic planning method, system and device based on deep learning and multi-objective optimization, and a storage medium, and the method comprises the steps: inputting a patient medical image, a tumor target region, an endangered organ structure and a treatment plan radiation field direction into a pre-trained 3DU-Net model, and obtaining a three-dimensional radiation field direction probability distribution diagram; performing Fibonacci spherical sampling to generate a candidate radiation field direction, calculating a path length score and an organ-at-risk avoidance score in each direction, and dynamically weighting and balancing the two scores; a proton range error and a patient positioning error are introduced to simulate dose distribution, target area coverage and organ-endangering dose change under the error are evaluated, and a robustness score is calculated; performing collision detection on the treatment rack, and rejecting directions which are not reachable or have collision risks; and selecting a direction to encrypt and sample near the direction, and optimizing a final radiation field direction according to comprehensive scoring results. According to the method, the radiation field direction can be automatically optimized, target dose coverage and organ-endangering protection are improved, and the robustness of a treatment plan to errors is enhanced.

Description

Technical Field

[0001] This invention relates to the technical field of radiotherapy, and in particular to a method, system, device, and storage medium for automatic planning of proton intensity-modulated therapy (IMRT) field direction based on deep learning and multi-objective optimization. Background Technology

[0002] Limitations of traditional photon radiotherapy: The photon beam dose decays exponentially, resulting in higher doses to normal tissues behind the tumor (such as the brainstem and spinal cord). Advantages of proton therapy: Utilizing the Bragg peak characteristic, a rapid dose drop can be achieved at the end of the target area. Intensity-modulated proton therapy (IMPT), as the current mainstream proton therapy technology, achieves unprecedented dose conformity by actively scanning a pencil beam and precisely controlling the intensity of each beam spot. However, this high degree of flexibility makes the choice of radiation field direction particularly critical, as its quality directly affects the robustness of dose distribution, the complexity of treatment planning, and the final clinical efficacy. The selection of proton therapy radiation field direction requires a balance between tumor coverage, protection of normal tissue, equipment limitations, anatomical and density heterogeneity, and patient specificity. It typically requires multi-objective optimization using a treatment planning system (TPS) and dynamic adjustments based on clinical experience to ultimately achieve individualized precision treatment.

[0003] The current technical status and bottlenecks of proton therapy radiation field planning are mainly reflected in the following aspects: (1) High dependence on manual labor and high uncertainty in the plan: The selection of the clinical firing field angle is almost entirely done manually. Different planners have subjective differences in their judgment of the merits of the angle, which leads to uncertainty in the quality of the plan. Complex cases often require planners to repeatedly try and adjust the firing field settings, which is cumbersome and inefficient.

[0004] (2) Lack of automatic angle selection support in TPS: Currently, mainstream treatment planning systems have limited support for optimizing the field angle. Ideally, the field direction selection and dose optimization should be performed simultaneously, but in practice, the field incident direction is usually preset manually first, and then the dose intensity is optimized by TPS. This separate process makes it difficult to achieve the global optimum for the field angle, limiting the space for further optimization of the plan.

[0005] (3) Low utilization of non-coplanar angles and difficulty in integrating collision avoidance and robustness: Increasing the non-coplanar incident angle has potential advantages in improving dose distribution. However, the search combination of non-coplanar fields is extremely large, and it is very difficult to exhaustively select them manually. Randomly using non-coplanar angles may also lead to the risk of collisions between the gantry and the bed or patient, but the existing TPS cannot detect collisions in real time during planning and optimization, and it is necessary to rely on human experience to avoid them in advance or adjust them afterward. Therefore, in clinical practice, except for difficult cases in certain special anatomical locations, planners rarely use non-coplanar fields, and the potential advantages of non-coplanar schemes have not been fully explored.

[0006] (4) Lack of quantitative evaluation indicators for ray path quality: Currently, there is no unified quantitative indicator system to evaluate the quality of candidate ray field paths. For each possible incident direction, there is a lack of clearly defined evaluation functions for "path quality" factors such as the length of the ray path, the type and volume of tissue traversed, and whether it passes through sensitive organs, leading planners to rely on subjective judgment based on experience. This lack of objective scoring standards makes it difficult for automatic algorithms to comprehensively consider factors such as path length and tissue avoidance in selecting the best path.

[0007] (5) Insufficient robustness of plans in complex anatomy or muscular organ situations: In cases of tumors with respiratory motion, such as the lungs and liver, or in situations where the density of tissues surrounding the tumor is highly heterogeneous, the field plan needs to be robust to anatomical changes and range uncertainties. However, in the current clinical planning process, field angle selection and robustness optimization are carried out separately. Usually, the field direction is determined first, and then dose error tolerance is considered in the optimization. This post-implementation robustness assessment means that it is impossible to judge the overall quality and robustness of a certain field configuration in advance during the planning stage. Even if the angle combination can bring an ideal nominal dose distribution, it may be very sensitive to position errors or density changes and lack stability under delivery errors, and vice versa. Existing TPS cannot organically integrate collision detection and robustness analysis into an integrated angle selection optimization process, making it particularly difficult to quickly generate safe and robust field plans in complex anatomical scenarios.

[0008] Therefore, how to construct a unified scoring mechanism that simultaneously addresses structural avoidance (avoidance of organs at risk), dose protection (target coverage and normal tissue dose control), and path quality for objective evaluation and selection of radiation field directions; how to couple deep learning prediction models with multi-objective optimization algorithms to achieve automatic angle selection of non-coplanar radiation fields within minutes, significantly improving planning efficiency and reducing manual intervention; and how to integrate robustness assessment and gantry collision detection into the angle selection optimization process to ensure the dose robustness of the generated plan and the safety of gantry movement from the source, thereby improving the clinical feasibility and safety reliability of the radiation field plan—solving these issues will provide an intelligent, efficient, and reliable technical foundation for proton therapy radiation field planning. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes an automatic planning method, system, device, and storage medium for proton intensity-modulated therapy (IMRT) field direction based on deep learning and multi-objective optimization. This invention generates the optimal field direction through deep learning and multi-objective dynamic optimization algorithms, minimizing path length, reducing the volume of normal tissue exposed to radiation, lowering the integrated dose, and maximizing organ at risk (OAR) avoidance. This ensures the field path avoids high-risk and easily changeable organs, supports robust planning for non-coplanar fields, and improves clinical feasibility by combining gantry collision detection and error simulation.

[0010] In a first aspect, the present invention provides an automatic planning method for the firing field direction of proton intensity-modulated therapy based on deep learning and multi-objective optimization, the method comprising: The system acquires the patient's 3D CT medical image data, the patient's tumor target area, the structure of organs at risk, and the direction of the treatment plan's radiation field. It then inputs these data into a pre-trained 3DU-Net deep neural network model for analysis, predicts the probability of each candidate radiation field direction becoming the optimal radiation field, and generates a 3D radiation field direction probability distribution map. Based on the Fibonacci sphere sampling method, multiple candidate shooting directions are uniformly generated on a unit sphere, and the path length score and the organ-at-risk avoidance score are calculated for each candidate shooting direction. According to the predetermined dynamic weighting strategy, the path length score and the endangered organ avoidance score are weighted and integrated to dynamically adjust the trade-off between endangered organ priority and path length. Dose simulations were performed for each candidate firing field direction by introducing proton range error and patient positioning error, to evaluate the changes in target area dose coverage and organ-at-risk dose under error conditions, and to calculate the robustness score for each candidate firing field direction. The collision detection model of the treatment gantry is used to detect the collision risk of each candidate firing direction and eliminate the candidate firing directions that are unreachable or have a collision risk. Based on the three-dimensional firing direction probability distribution map, candidate firing directions within a set probability range are selected, and local dense sampling is performed in their neighboring areas. The final firing direction is then selected based on the integrated scoring results.

[0011] Furthermore, the 3DU-Net deep neural network model takes CT medical image data, tumor target area, organ structure at risk, and treatment plan field direction as multi-channel inputs, and outputs a three-dimensional field direction probability distribution map covering all candidate field directions, with each field direction corresponding to the probability value of becoming the optimal field.

[0012] Furthermore, by sampling with Fibonacci spheres, a set of candidate field directions is uniformly generated on a unit sphere, which eliminates unreachable directions that are beyond the rotation range of the treatment gantry or the angle limit of the treatment bed.

[0013] Furthermore, the path length score is obtained by calculating and normalizing the distance from the center of the tumor target area to the patient's body surface in the direction of the candidate firing field; the shorter the distance, the higher the path length score.

[0014] Furthermore, the endangered organ avoidance score is obtained by calculating the minimum distance between the ray path of the candidate radiation field direction and each endangered organ, and by performing a function transformation on the minimum distance based on a preset safe distance threshold. The larger the minimum distance, the higher the endangered organ avoidance score.

[0015] Furthermore, the dynamic weighting strategy adaptively adjusts the weights of the path length score and the endangered organ avoidance score based on the priority of the endangered organ and the path length threshold. Specifically, the weight of the endangered organ avoidance score is increased for high-risk endangered organs, and the weight of the path length score is increased when the path length of the candidate firing field direction exceeds a preset threshold.

[0016] Furthermore, the dose simulation of the proton range error is performed by adjusting the proton range within a preset deviation range to generate the corresponding dose distribution and calculate the target area dose coverage; the dose simulation of the positioning error is performed by translating the patient position within a preset distance range to assess the change in the dose to organs at risk, and the robustness score of each candidate field direction is determined based on the change in the target area dose coverage and the dose to organs at risk.

[0017] Secondly, this invention provides an automatic proton intensity-modulated therapy (IMRT) field direction planning system based on deep learning and multi-objective optimization, the system comprising: Data acquisition module: used to acquire and preprocess the input data required for proton therapy planning, including the patient's 3D CT medical image data, tumor target area, organs at risk structures, and treatment plan field direction; The CNN prediction module is used to input the patient's 3D CT medical image data, tumor target area and organs at risk structure, and treatment plan field direction into the trained 3DU-Net deep learning model for analysis, and output a 3D field direction probability distribution map composed of the probability values ​​of each field direction. The multi-objective optimization module is used to filter directions within a set probability range based on the three-dimensional shooting direction probability distribution map and generate a set of candidate shooting directions by sampling multiple candidate shooting directions in their neighborhood using Fibonacci spheres. For each candidate shooting direction, a path length score and an organ avoidance score are calculated, and the scores are weighted and integrated using a dynamic weighting strategy to determine the optimal shooting direction. The collision detection module is used to calculate the minimum distance between the gantry and the patient under each candidate firing direction based on the relative position relationship between the treatment gantry and the patient, and compare the minimum distance with a preset safety threshold to detect collision risk and eliminate candidate firing directions with collision risk. The robustness analysis module is used to simulate proton range error and patient positioning error for each candidate firing field direction, evaluate the changes in target dose coverage and organ-at-risk dose under the error conditions, and calculate the robustness score for each candidate firing field direction. Output Result Module: Used to output the final determined proton therapy field direction planning scheme.

[0018] Thirdly, the present invention provides an automatic proton intensity-modulated therapy (IMRT) field direction planning device based on deep learning and multi-objective optimization, comprising: The proton therapy head, mounted on a rotatable treatment gantry, is used to emit a proton beam to irradiate the patient's tumor target area; It also includes the aforementioned automatic proton intensity-modulated therapy (IMRT) field direction planning system based on deep learning and multi-objective optimization, which is used to automatically plan the proton therapy field direction according to the patient's medical imaging data, tumor target area, and information on organs at risk.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a processor, cause the processor to perform the steps of the above-described method for automatic planning of proton intensity-modulated therapy firing field direction based on deep learning and multi-objective optimization.

[0020] Compared with the prior art, the present invention has the following beneficial effects: Firstly, this invention presents an automatic proton field direction planning method based on deep learning and multi-objective optimization, enabling intelligent selection and optimization of field angles. Its advantages include: utilizing a pre-trained 3DU-Net model to automatically predict the probability distribution of the optimal field direction from patient images and organ information, significantly reducing reliance on manual angle selection and improving planning efficiency. Simultaneously, it combines multi-objective scoring such as path length and organ-at-risk avoidance, and adaptively balances tumor dose coverage and organ-at-risk protection through a dynamic weighting strategy, ensuring that the selected field effectively covers the tumor while minimizing the need to avoid sensitive organs, thus improving planning quality. The method also incorporates collision detection and robustness assessment, eliminating gantry-inaccessible angles and considering the impact of range and positioning errors, ensuring that the final selected field direction is clinically feasible, safe, reliable, and insensitive to errors. Overall, this automatic planning method reduces reliance on human experience and can quickly generate high-quality field configurations comparable to manual plans, achieving highly efficient automation of proton therapy field selection.

[0021] Secondly, the 3DU-Net deep neural network of this invention uses CT images, tumor target area, structures of organs at risk, and treatment plan field directions as multi-channel inputs, outputting a three-dimensional field direction probability map covering all candidate directions. This multi-channel 3DU-Net model can simultaneously consider the spatial anatomical relationship between the tumor and surrounding organs, automatically learn expert angle selection experience from medical images, and thus accurately predict the probability of each direction becoming the optimal field. By outputting a comprehensive field probability distribution, the system can intuitively obtain an evaluation heatmap of the advantages and disadvantages of each angle, facilitating the identification of high-probability preferred direction areas and reducing blind searching. This prediction module significantly improves the intelligence and efficiency of angle selection, enabling subsequent optimization to focus on high-probability directions, thereby improving the accuracy and rationality of field selection.

[0022] Third, this invention generates a large number of candidate firing directions uniformly on a unit sphere through Fibonacci spherical sampling, and eliminates inaccessible directions that exceed the gantry rotation range or bed angle limitations. Fibonacci spherical sampling can very uniformly cover all directions on the sphere, avoiding biases and blind spots in direction selection and increasing the probability of finding the globally optimal firing field. Uniform sampling also ensures that the algorithm fairly evaluates all possible angles and does not miss potential high-quality directions. In addition, pre-eliminating physically inaccessible or collision-risk directions reduces the search space, avoids wasting computational resources on infeasible solutions, and improves planning efficiency and the feasibility of the results.

[0023] Fourth, this invention calculates a "path length score" for each candidate radiation field direction, which is the normalized distance from the center of the tumor target area along that direction to the patient's body surface. The shorter the distance, the higher the score. Through this scoring index, the system tends to select radiation field directions with shorter entry paths into human tissue. A shorter range means that the proton beam passes through less tissue in the body, resulting in less energy loss and scattering, which helps to accurately deposit the Bragg peak on the tumor. Therefore, a short-path radiation field can reduce the dose received by normal tissue, reduce the risk of side effects, and reduce the impact of proton range uncertainty on dose distribution. After normalization, the path length score is easy to integrate with other scores, making the angle selection optimization more scientific and reasonable, ensuring tumor coverage while optimizing the entry path.

[0024] Fifth, a "risk organ avoidance score" is calculated for each candidate direction. This score is derived by determining the minimum distance between the radiation field path and each risk organ, and then converting this distance based on a preset safety distance threshold. A larger minimum distance results in a higher score, prompting the system to prioritize radiation field directions that are far from critical organs. If the radiation path maintains a sufficiently large distance from the risk organ (exceeding the safety threshold), the direction receives a high score due to its minimal impact on the organ; conversely, if the path is close to a sensitive organ, a significant deduction is made. In this way, the potential irradiation risk to critical organs from each candidate radiation field is quantified, ensuring that the final selected radiation field direction avoids risk organs as much as possible, achieving better protection for normal tissues. By setting a safety distance threshold for function conversion, the score is ensured to be sensitive to changes in organ spacing and has clinical significance, making the optimization process more in line with radiotherapy safety requirements.

[0025] Sixth, this invention employs a dynamic weighting strategy, adaptively adjusting the weights of path length scores and organ avoidance scores based on the priority of organs at risk and a threshold for the length of the firing field path. Dynamic weighting empowers the algorithm to flexibly prioritize optimization based on specific clinical circumstances. When certain organs at risk are critical and high-risk, the avoidance score weight is automatically increased to prioritize the protection of sensitive organs; conversely, when candidate firing field paths are too long (exceeding the threshold), the path length score weight is increased to avoid selecting directions that penetrate too much tissue. This adaptive adjustment ensures a balance between optimization objectives for different patients and under different firing paths, overcoming the suboptimal results that may result from fixed weights. Overall, this strategy improves the adaptability of firing field selection to individual anatomy and risk distribution, ensuring both tumor dosage and critical organ safety in the planning process, resulting in optimization outcomes that better meet clinical expectations.

[0026] Seventh, this invention introduces proton range error and patient positioning error into each candidate radiation field direction for dose simulation, evaluates target area dose coverage and organ-at-risk dose changes under error conditions, calculates robustness scores, and incorporates uncertainties such as range deviation and positioning deviation into the radiation field optimization evaluation, ensuring that the selected radiation field maintains good dosimetric effects even under actual clinical error scenarios. By simulating the impact of proton range changes on dose distribution and the impact of patient position deviation on organ-at-risk dose, the system can identify directions that are less sensitive to errors and have more stable dose coverage, and assign them high robustness scores. This mechanism ensures that the finally selected radiation field direction has stronger robustness in actual treatment: it avoids underdose in the target area caused by range error and prevents unexpectedly excessive dose to organs-at-risk under small positioning errors, improving the safety margin of the plan and the reliability of the treatment effect.

[0027] In summary, this invention provides an automated and high-precision method for proton therapy radiation field direction planning. It generates the optimal radiation field direction through deep learning and multi-objective dynamic optimization algorithms, achieving the following objectives: minimizing path length, reducing the volume of normal tissue exposed to radiation, and lowering the integrated dose; maximizing OAR avoidance: ensuring the radiation field path avoids high-risk organs (such as the spinal cord and optic nerve) and easily changeable organs (such as the intestines and stomach); and supporting robust planning for non-coplanar fields: combining gantry collision detection and error simulation to improve clinical feasibility. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of an automatic proton intensity-modulated therapy (IMRT) field direction planning method based on deep learning and multi-objective optimization, according to an embodiment of the present invention. Detailed Implementation

[0029] The following examples illustrate the implementation of the present invention in detail, but they do not constitute a limitation on the invention and are merely illustrative. Furthermore, the advantages of the present invention will become clearer and easier to understand by explaining them.

[0030] Example 1: This example provides an automatic proton intensity-modulated therapy (IMRT) field direction planning system based on deep learning and multi-objective optimization. The system includes: a data acquisition module, a CNN prediction module, a multi-objective optimization module, a robustness analysis module, a collision detection module, and an output result module. The structure and function of each module are as follows: Data Acquisition Module: This module acquires and preprocesses the input data required for proton therapy planning, including the patient's medical images (such as 3D CT scans), tumor target volume, structures of organs at risk, and the direction of the treatment plan's radiation field. This module standardizes and processes the data before providing it to subsequent modules.

[0031] The CNN prediction module takes the patient's 3D CT medical image data, tumor target area, organs at risk, and treatment plan radiation field directions and inputs them into a pre-trained 3DU-Net deep learning model for analysis. It outputs a 3D radiation field direction probability distribution map composed of the probability values ​​of each radiation field direction. This module takes the patient's images and target / organ at risk information as input and uses a pre-trained offline deep learning model to quickly estimate the impact of each radiation field direction on tumor dose coverage and normal tissue dose distribution. The 3D radiation field direction probability distribution map output by the CNN prediction module provides a basis for subsequent path scoring and optimization.

[0032] The multi-objective optimization module filters directions within a defined probability range based on a 3D radiation field direction probability distribution map. It then uses Fibonacci sphere sampling in the neighborhood to generate a candidate radiation field direction set. For each candidate direction, it calculates a path length score and an organ-at-risk avoidance score, and uses a dynamic weighting strategy to weight and integrate these scores to determine the optimal radiation field direction. During the iteration process, the multi-objective optimization module integrates the scores provided by the path scoring module and dose prediction information, evaluating, selecting, cross-referencing, and mutating radiation field combinations, evolving generation by generation to generate a set of candidate optimal solutions. Finally, this module outputs a set of optimized radiation field directions that achieve a balance across all objectives, which can be further filtered by the robustness analysis module or selected by clinical users.

[0033] Robustness Analysis Module: This module simulates proton range errors and patient positioning errors for each candidate field direction, assesses changes in target dose coverage and organ-at-risk dose under these error conditions, and calculates a robustness score for each candidate field direction. This module performs uncertainty robustness assessment on candidate field plans generated by the multi-objective optimization module. Based on uncertainties specific to proton therapy (e.g., patient positioning errors, proton range errors), this module simulates and evaluates the dose distribution changes of each field direction under offset or error scenarios. The robustness analysis module can introduce several representative uncertainty scenarios (e.g., positional offsets within ± millimeters in each direction, or changes in proton range within ± several percentage points) to reassess whether target dose coverage and organ-at-risk dose meet requirements. For field direction combinations with poor robustness (e.g., target coverage falling below prescription requirements in certain adverse scenarios), this module can filter them out or notify the optimization module to adjust weights and re-optimize. Through the robustness analysis module, it ensures that the output field plans still have good dosimetric performance within the actual clinical error range.

[0034] Collision Detection Module: This module simulates proton range errors and patient positioning errors for each candidate radiation field direction, assesses changes in target dose coverage and organ-at-risk dose under these error conditions, and calculates the robustness score for each candidate radiation field direction. This module verifies the physical feasibility of the selected radiation field direction to avoid collisions between the treatment gantry and the patient or other equipment. The module stores geometric models of the treatment equipment (such as rotating gantry, treatment bed, treatment head, etc.) and the patient's shape. For each candidate radiation field direction, the collision detection module calculates the spatial relative positions of the treatment head, collimator, bed, and patient body based on the corresponding gantry and bed angles, detecting any geometric overlap or distances below a safe threshold. If a radiation field direction has a potential collision risk, it is marked as unusable or an adjustment suggestion is provided (e.g., slightly shifting the patient bed). The collision detection module can filter out obviously infeasible directions during the direction sampling stage and can also validate the final solution during the optimization stage to ensure the recommended combination of radiation field directions is electrically safe and reliable during clinical implementation.

[0035] Output Result Module: This module outputs the final determined firing field direction planning scheme and related information. It takes the optimal firing field direction set, filtered through multi-objective optimization and robustness analysis and verified through collision detection, as the planning result output of this invention. The output result module can save this information as a treatment plan file for further use by the treatment planning system, or directly provide it to clinicians for review and adjustment.

[0036] The modules of the aforementioned system can be implemented using a combination of computer programs and hardware. For example, this system can be deployed on a radiotherapy planning computer, with the corresponding software units of each module executed by a processor to achieve the above functions; alternatively, hardware acceleration such as FPGA / ASIC can be used to accelerate specific modules (e.g., deep learning inference computation). The modules interact through data and signals to form the overall workflow: the data acquisition module provides preprocessed patient data to the direction sampling module and the CNN prediction module; the CNN prediction module and the path scoring module collaboratively calculate the evaluation index for each candidate direction; the multi-objective optimization module selects the optimal solution based on this, and the selection is verified by the robustness analysis module and the collision detection module; finally, the output result module generates a complete automatic planning scheme for the radiation field direction. This system can efficiently integrate deep learning prediction and optimization algorithms to achieve intelligent automatic planning of proton therapy radiation field directions, improving planning efficiency and scheme quality.

[0037] Example 2: This example provides an automatic planning method for the firing field direction of proton intensity-modulated therapy based on deep learning and multi-objective optimization. The method can be executed by the system described in Example 1 above, realizing automatic selection and optimization of the firing field direction, such as... Figure 1 As shown, the method includes: The system acquires the patient's 3D CT medical image data, the patient's tumor target area, the structure of organs at risk, and the direction of the treatment plan's radiation field. This data is then input into a pre-trained 3DU-Net deep neural network model for analysis. The model predicts the probability that each candidate radiation field direction will become the optimal radiation field and generates a 3D radiation field direction probability distribution map. The training dataset used by the pre-trained 3DU-Net deep neural network model consists of several clinical proton therapy plans, which include radiation field direction labels and dose distributions.

[0038] Based on the Fibonacci sphere sampling method, multiple candidate shooting directions are uniformly generated on a unit sphere, and the path length score and the organ-at-risk avoidance score are calculated for each candidate shooting direction. According to the predetermined dynamic weighting strategy, the path length score and the endangered organ avoidance score are weighted and integrated to dynamically adjust the trade-off between endangered organ priority and path length. Dose simulations were performed for each candidate firing field direction by introducing proton range error and patient positioning error, to evaluate the changes in target area dose coverage and organ-at-risk dose under error conditions, and to calculate the robustness score for each candidate firing field direction. The collision detection model of the treatment gantry is used to detect the collision risk of each candidate firing direction and eliminate the candidate firing directions that are unreachable or have a collision risk. Based on the three-dimensional shooting direction probability distribution map, candidate shooting directions within a set probability range (such as a high probability value range) are selected, and local encrypted sampling is performed in their neighboring areas. The final shooting direction is then selected based on the integrated scoring results.

[0039] In the above technical solution, the 3DU-Net deep neural network model takes CT medical image data, tumor target area, organ structure at risk and treatment plan field direction as multi-channel input, and outputs a three-dimensional field direction probability distribution map covering all candidate field directions, with each field direction corresponding to the probability value of becoming the optimal field.

[0040] In the above technical solution, a set of candidate field directions is uniformly generated on a unit sphere by sampling Fibonacci spheres, which consists of multiple candidate field directions, and unreachable directions that exceed the rotation range of the treatment gantry or the angle limit of the treatment bed are excluded.

[0041] In the above technical solution, the path length score is obtained by calculating the distance from the center of the tumor target area to the patient's body surface in the direction of the candidate firing field and normalizing the distance. The shorter the distance, the higher the path length score.

[0042] In the above technical solution, the endangered organ avoidance score is obtained by calculating the minimum distance between the ray path of the candidate radiation field direction and each endangered organ, and by performing a function transformation on the minimum distance based on a preset safe distance threshold. The larger the minimum distance, the higher the endangered organ avoidance score.

[0043] In the above technical solution, the dynamic weighting strategy adaptively adjusts the weights of the path length score and the endangered organ avoidance score based on the priority of the endangered organ and the path length threshold. Specifically, the weight of the endangered organ avoidance score is increased for high-risk endangered organs, and the weight of the path length score is increased when the path length of the candidate firing field direction exceeds a preset threshold.

[0044] In the above technical solution, the dose simulation of the proton range error is achieved by adjusting the proton range within a preset deviation range to generate the corresponding dose distribution and calculate the target area dose coverage; the dose simulation of the positioning error is achieved by translating the patient position within a preset distance range to assess the change in the dose to organs at risk, and the robustness score of each candidate field direction is determined based on the target area dose coverage and the change in the dose to organs at risk.

[0045] As a preferred embodiment, the automatic planning method for proton intensity-modulated therapy (IMRT) field direction based on deep learning and multi-objective optimization includes the following steps: S1: CNN candidate direction prediction 1) Input data: Patient's 3D CT image (size 512×512×N, slice thickness 3mm); segmented target area (PTV) and OAR (binary label); patient position and treatment bed coordinate system information.

[0046] 2) Output data: The three-dimensional probability map P(d)∈[0,1] represents the probability that each firing direction is the optimal solution; the set of high-probability region directions The number is 20%-40% of the original candidate directions.

[0047] 3) Network architecture: The 3DU-Net architecture is adopted, with the input layer consisting of multi-channel data (CT values ​​+ CTV / OAR). The output layer generates an orientation probability map through a spherical convolutional layer.

[0048] Loss function: Weighted cross-entropy loss In the formula, L is the loss function, and the smaller the value, the more accurate the prediction; d j ' represents the j-th candidate direction; ω(d j ') represents the direction weight; P(d) j ') represents the probability of the direction predicted by the network; y j For the true label of direction, if d j 'Is it the optimal clinical direction?' j =1, otherwise y j =0.

[0049] 4) Training data: Dataset: 500 clinical proton therapy plans, including field orientation labels and dose distribution. Data augmentation: rotation (±15°), translation (±10mm), noise injection (Gaussian noise σ=50HU).

[0050] S2: Multi-objective algorithm optimization 1) Candidate direction generation Fibonacci sphere sampling: generating a uniformly distributed set of candidate directions on a unit sphere. In the formula, m is the number of sampling points; the number of sampling points ϕm = 4πr 2 / ΔθΔϕ, where Δθ and Δϕ are the interval between the azimuth and polar angles, and r is the radius of the Fibonacci sphere.

[0051] Directional filtering: Exclude angles that are physically inaccessible (rack angle limitations, bed angle limitations).

[0052] 2) Path length scoring Ray path definition: from the tumor centroid c T The DJ direction extends to the body surface: In the formula, L(dj) is the path length in the dj direction, t is the step size parameter (scalar) along the ray direction, and c T Let dj be the centroid coordinates of the tumor target volume PTV, and dj be the unit vector generated by sampling from the Fibonacci sphere. VbodyThis refers to the boundary of the patient's curved skin surface. Path length and normal tissue dose: The smaller the L(dj) value, the smaller the volume of normal tissue the radiation passes through, and the lower the normal tissue integrated dose (NTID). For example, in head and neck treatment, L=8cm is better than L=15cm (40% reduction in skin dose).

[0053] Relationship with proton range: The actual range Rp of the proton beam must satisfy: Rp≥L(d j )+Target depth Otherwise, this direction must be excluded (it cannot cover the far end of the target area).

[0054] Dynamic adjustment mechanism: Obese patients: automatically increase Lmax (e.g., 40cm); Children: limit L to ≤20cm (to reduce the impact on growth and development).

[0055] Table 1. Actual CT Cases (Head and Neck Tumors) If the end of the radiation field is close to an organ at risk, it must also be ruled out.

[0056] Normalized score: In the formula, Slength(dj) is the path length score of direction dj, dj is the unit vector generated by Bonach sphere sampling, L(dj) is the shortest straight-line distance from the tumor centroid along dj to the body surface, and Lmax is the maximum body thickness of the patient.

[0057] Table 2 Clinical Decision Thresholds 3) OAR (Occupational Alternative Ranking) Minimum distance calculation: ray path P(dj) to each OARO i The minimum geometric distance dmin(dj,O) i ).

[0058] Exponential decay score: In the formula, S OAR (dj) represents the OAR avoidance score for direction dj, dsafe,i is the safe distance threshold for OAR, γ is the attenuation coefficient, and dj represents the candidate firing field direction. For the first i One organ at risk; d min (d j O i ) represents the minimum distance from the ray path to the OAR; dsafe,i represents the safe distance threshold of the OAR.

[0059] Security protection mechanism: when d min≪ d When safe: exp(⋅) → 0 → Forcefully eliminate the dangerous direction. when d min≫ d When safe: exp(⋅) → 1 → safe direction Multi-OAR Collaboration: Each OAR is scored independently and summed → directions to avoid all high-risk organs are automatically identified.

[0060] If all feasible directions cannot avoid the OAR (e.g., pancreatic cancer encasing mesenteric vessels): a pre-defined condition that must be met in subsequent dose optimization should be set to guide the predictive rule for selecting the current direction.

[0061] Hierarchical optimization method: (1) Hard constraint priority: set an absolute dose limit for extremely high-risk sequential organs or OARs that clinicians require to be avoided, and adopt a preventive geometric exclusion rule in the field direction selection stage; (2) Soft constraint trade-off: adopt a comprehensive risk score based on geometric features for intermediate-risk OARs to quantitatively assess the relative risk of different field directions. This score comprehensively considers the beam's travel length in the OAR, the path's position relative to the geometric center of the OAR, and the uncertainty brought about by organ motion and tissue heterogeneity, so as to identify the field direction with the lowest risk when it is necessary to pass through the OAR.

[0062] The OAR (Occupational Awareness Assessment) scoring formula is: The OAR risk scoring formula is: In the formula: For path length risk, this is the normalized value of the beam's travel length within the OAR. A higher value indicates a higher integrated dose that the OAR may receive. The weight is reduced for sequential organs and increased for parallel organs (such as the liver and lungs) because their damage risk is related to the irradiated volume.

[0063] f centralityFor centrality risk, the position of the beam path relative to the local geometric center of the OAR is assessed. For each candidate direction, the system identifies the local region where the beam actually intersects the OAR, and calculates the local centroid and equivalent radius of that region. Centrality risk is defined as the ratio of the distance from a representative point on the path to the local centroid to the local equivalent radius, accurately reflecting the potential risk of forming a high-dose hotspot within the local region. A higher value (closer to the center) indicates a higher risk of forming a high-dose hotspot. For sequential organs (e.g., nerves, small intestine) or any organ with a concern about hotspots, a higher weight should be given because the risk of damage is related to the maximum dose. sensitivity For sensitivity risks, additional uncertainties arising from organ motion and tissue heterogeneity are comprehensively considered. Weights are increased for motion-sensitive organs (such as the lungs and liver) or regions with complex interfaces. w1, w2, and w3 are the relative weights for path length risk, centrality risk, and sensitivity risk, respectively.

[0064] 4) Dynamic weighted comprehensive score Comprehensive scoring function: In the formula, Total(dj) is the overall score of direction dj, α is the path length weight, Slength(dj) is the path length score, and β is the OAR avoidance weight. OAR avoidance rating.

[0065] Weight adaptive strategy: α and β are dynamically adjusted based on OAR priority, and the priority of dosimetric targets is predefined. For example, the β weight of high-risk organs such as the spinal cord is increased by 50%.

[0066] If the path length exceeds the threshold Lth, the α weight increases linearly.

[0067] Table 3 Dynamic Weight Adjustment Strategy 5) Robustness optimization Range error simulation: Calculate the potential irradiation area of ​​the field direction under nominal and perturbed ranges, generate a dose distribution with ±3% range error for each candidate direction dj, and calculate the target coverage Crobust.

[0068] Positioning error simulation: By translating the patient model (±3mm), the stability of the geometric relationship between the radiation field path and the OAR was assessed, and the risk surrogate value of OAR dose variation was estimated using geometric surrogate. . is a dimensionless score that quantifies the relative risk of a beam direction causing one or more OAR dose over - exposures under a set of error scenarios. The higher the value, the greater the risk. Usually, number (such as ) of representative setup error vectors are used, covering the vertices and the center of a ±3 mm cube in three - dimensional space. The geometric model of the OAR is translated along the vector , and the minimum distance between the beam path and the translated OAR is calculated. A decay function is used to map the distance to a risk value for a single scenario. The risk values for all error scenarios are aggregated to obtain an overall risk proxy value for that OAR. Usually, the worst - case scenario (the maximum value among all scenarios) is taken because the clinical concern is mainly about the maximum possible risk. is the clinically - set safety distance threshold. In the formula, is the candidate beam direction; K is the total number of error scenarios, with a default of 81 (3 ranges×27 setups×1 motion); Coveragek is the target coverage (V95%) for scenario k, calculated by dose grid resampling; λ is the penalty weight for the risk proxy value of OAR dose change (set to 1.0), and if the OAR is a motion - sensitive organ (such as the intestine) → λ = 2.0.

[0069] Table 4 Response to changes in organ contents Dynamic avoidance: 4D - CT - driven minimum - distance mapping → ensuring safety distance under respiratory motion; response strategy for content changes → reducing the risk of intestinal / bladder dose over - exposure by 60%.

[0070] Multi - physics collaboration: joint optimization of range / setup / motion / content errors; inter - field dose complementary design.

[0071] Clinical universality: the plan covers thoracic and abdominal tumors (lung cancer, liver cancer, pancreatic cancer).

[0072] 6) Non - coplanar beam collision detection Gantry kinematic model: Based on the treatment room coordinate system, calculate the minimum distance between the gantry angle θ, couch angle ϕ and the patient contour.

[0073] Collision marking: If the gantry - patient distance dcollision < dsafe, exclude this direction.

[0074] S3: Result integration and output 1) Coarse screening of CNN candidate directions: CNN - predicted high - probability direction set The number of candidate directions is mCNN=800 (originally 2000).

[0075] Directions with a probability below the threshold Pth=0.3 are excluded, reducing the computational load by 50%.

[0076] 2) Multi-objective algorithm optimization for refined sampling: exist Locally refined sampling of the Fibonacci sphere is performed within the neighborhood (the spacing is refined from 5° to 2°).

[0077] 3) Dynamic weight adjustment: Clustering and dimensionality reduction are used to extract representative schemes for shooting fields with similar directions, and multi-field direction results are output based on the order of clinical scores.

[0078] This invention achieves the following effects: increased efficiency, reducing planning time from hours to minutes to meet real-time clinical needs; optimized dosimetry, achieving simultaneous target coverage and normal tissue protection; and broad clinical applicability, supporting complex scenarios such as sensitive pediatric cases and tumors of the musculoskeletal system, while improving the utilization rate of non-coplanar fields.

[0079] The working principle of the above technical solution is as follows: Utilizing diverse candidate radiation fields provided by directional sampling, combined with the rapid estimation of dose effects for each direction by the CNN prediction module, the merits of radiation fields can be efficiently evaluated in a broad search space. Path scoring further quantifies the comprehensive value of each direction, initially screening out inferior directions, thereby reducing optimization complexity. Guided by the deep learning evaluation results, the multi-objective optimization module automatically searches for radiation field combinations that balance tumor control and normal tissue protection, and ensures that the selected scheme is resistant to errors through robustness analysis. Finally, collision detection verifies the feasibility of the scheme. The close cooperation of each module automates and intelligently plans the entire process.

[0080] The above technical solution has the following beneficial effects: First, by introducing a deep learning prediction model, the speed of radiation field direction evaluation is significantly improved. Compared with the traditional traversal search that relies entirely on physical dose calculation, this invention can narrow down the range of preferred directions more quickly. Second, by using a multi-objective optimization algorithm to consider multiple dosimetric objectives simultaneously, the resulting scheme is more optimized than manual experience-based adjustments, achieving a reasonable trade-off between maximizing tumor dose coverage and minimizing dose to organs at risk. Third, the combination of path scoring functions and robustness analysis ensures that this invention not only focuses on recognized planning quality indicators but also pre-considers the sensitivity of the plan to anatomical changes and physical uncertainties, guaranteeing that the recommended scheme has robust dose distribution performance in actual clinical practice. Finally, the collision detection module automatically avoids mechanical collision risks, ensuring that the optimized results can be safely implemented on the treatment equipment. Therefore, the automatic planning method and system for proton intensity-modulated therapy radiation field direction based on deep learning and multi-objective optimization provided by this invention significantly improves the intelligence level and efficiency of radiotherapy planning, and can provide high-quality and safe and reliable radiation field angle schemes for clinical use.

[0081] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described in this application.

[0082] This application also provides a computer-readable storage medium for storing a computer program. When the computer program is executed, it implements the steps of any of the methods described in this application. The specific implementation method is consistent with the implementation method and the technical effect achieved in the above method embodiments, and some contents will not be repeated.

[0083] In this application, a readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The program product can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0084] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar programming languages. The program code may be executed entirely on a user computing device, partially on an associated device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).

[0085] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention without departing from the principles and spirit of the invention, and all such changes should fall within the protection scope of the claims of the present invention.

Claims

1. A method for automatic planning of proton intensity-modulated therapy (IMRT) field direction based on deep learning and multi-objective optimization, characterized in that, The method includes: The system acquires the patient's 3D CT medical image data, the patient's tumor target area, the structure of organs at risk, and the direction of the treatment plan's radiation field. It then inputs these data into a pre-trained 3DU-Net deep neural network model for analysis, predicts the probability of each candidate radiation field direction becoming the optimal radiation field, and generates a 3D radiation field direction probability distribution map. Based on the Fibonacci sphere sampling method, multiple candidate shooting directions are uniformly generated on a unit sphere, and the path length score and the organ-at-risk avoidance score are calculated for each candidate shooting direction. According to the predetermined dynamic weighting strategy, the path length score and the endangered organ avoidance score are weighted and integrated to dynamically adjust the trade-off between endangered organ priority and path length. Dose simulations were performed for each candidate firing field direction by introducing proton range error and patient positioning error, to evaluate the changes in target area dose coverage and organ-at-risk dose under error conditions, and to calculate the robustness score for each candidate firing field direction. The collision detection model of the treatment gantry is used to detect the collision risk of each candidate firing direction and eliminate the candidate firing directions that are unreachable or have a collision risk. Based on the three-dimensional firing direction probability distribution map, candidate firing directions within a set probability range are selected, and local dense sampling is performed in their neighboring areas. The final firing direction is then selected based on the integrated scoring results.

2. The method according to claim 1, characterized in that, The 3DU-Net deep neural network model takes CT medical image data, tumor target area, organ structures at risk, and treatment plan field directions as multi-channel inputs, and outputs a three-dimensional field direction probability distribution map covering all candidate field directions, with each field direction corresponding to the probability value of becoming the optimal field.

3. The method according to claim 1, characterized in that, By sampling with Fibonacci spheres, a set of candidate field directions is uniformly generated on a unit sphere, and unreachable directions that are beyond the rotation range of the treatment gantry or the angle limit of the treatment bed are excluded.

4. The method according to claim 1, characterized in that, The path length score is obtained by calculating and normalizing the distance from the center of the tumor target area to the patient's body surface in the direction of the candidate shooting field. The shorter the distance, the higher the path length score.

5. The method according to claim 1, characterized in that, The endangered organ avoidance score is obtained by calculating the minimum distance between the ray path of the candidate radiation field direction and each endangered organ, and then performing a function transformation on the minimum distance based on a preset safe distance threshold. The larger the minimum distance, the higher the endangered organ avoidance score.

6. The method according to claim 1, characterized in that, The dynamic weighting strategy adaptively adjusts the weights of the path length score and the endangered organ avoidance score based on the priority of the endangered organ and the path length threshold. Specifically, the weight of the endangered organ avoidance score is increased for high-risk endangered organs, and the weight of the path length score is increased when the path length of the candidate firing field direction exceeds a preset threshold.

7. The method according to claim 1, characterized in that, The dose simulation of the proton range error is performed by adjusting the proton range within a preset deviation range to generate the corresponding dose distribution and calculate the target area dose coverage; the dose simulation of the positioning error is performed by translating the patient position within a preset distance range to assess the change in the dose to organs at risk, and the robustness score of each candidate field direction is determined based on the change in the target area dose coverage and the dose to organs at risk.

8. An automatic proton intensity-modulated therapy (IMRT) field direction planning system based on deep learning and multi-objective optimization, characterized in that, The system includes: Data acquisition module: used to acquire and preprocess the input data required for proton therapy planning, including the patient's 3D CT medical image data, tumor target area, organs at risk structures, and treatment plan field direction; The CNN prediction module is used to input the patient's 3D CT medical image data, tumor target area and organs at risk structure, and treatment plan field direction into the trained 3DU-Net deep learning model for analysis, and output a 3D field direction probability distribution map composed of the probability values ​​of each field direction. The multi-objective optimization module is used to filter directions within a set probability range based on the three-dimensional shooting direction probability distribution map and generate a set of candidate shooting directions by sampling multiple candidate shooting directions in their neighborhood using Fibonacci spheres. For each candidate shooting direction, a path length score and an organ avoidance score are calculated, and the scores are weighted and integrated using a dynamic weighting strategy to determine the optimal shooting direction. The collision detection module is used to calculate the minimum distance between the gantry and the patient under each candidate firing direction based on the relative position relationship between the treatment gantry and the patient, and compare the minimum distance with a preset safety threshold to detect collision risk and eliminate candidate firing directions with collision risk. The robustness analysis module is used to simulate proton range error and patient positioning error for each candidate firing field direction, evaluate the changes in target dose coverage and organ-at-risk dose under the error conditions, and calculate the robustness score for each candidate firing field direction. Output Result Module: Used to output the final determined proton therapy field direction planning scheme.

9. An automatic proton intensity-modulated therapy (IMRT) field direction planning device based on deep learning and multi-objective optimization, characterized in that, include: The proton therapy head, mounted on a rotatable treatment gantry, is used to emit a proton beam to irradiate the patient's tumor target area; It also includes the proton intensity-modulated therapy (IMRT) field direction automatic planning system based on deep learning and multi-objective optimization as described in claim 8, which is used to automatically plan the proton therapy field direction according to the patient's medical imaging data and information on the tumor target area and organs at risk.

10. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions that, when executed by a processor, cause the processor to perform the steps of the method as described in any one of claims 1-7.

Images