A swarm intelligence-driven radiotherapy planning optimization system

Abstract

This invention discloses a swarm intelligence-driven radiotherapy planning optimization system, which relates to the field of radiotherapy system optimization technology. The invention employs a swarm intelligence-driven multi-agent collaborative optimization framework, mapping each structure to an independent agent. A communication graph is constructed using geometric and dose features to achieve precise interaction. Combining self-organizing coordination and conflict fusion mechanisms, dose conflicts are dynamically handled. Fuzzy logic decision-making is introduced to adapt to clinical needs, ensuring compatibility with existing TPS systems and improving dose optimization accuracy. Conflict regions are finely allocated to reduce the risk of underdose in the target area or overdose in the oral radiation area (OAR). The system improves optimization efficiency; the intelligent communication graph reduces redundant interactions; self-organizing collaboration accelerates convergence; and the system is easy to implement and robust.

Description

Technical Field

[0001] This invention relates to the field of radiation system optimization technology, and in particular to a swarm intelligence-driven radiation therapy planning optimization system. Background Technology

[0002] Most existing radiotherapy planning optimization models are based on aggregated centralized optimization methods and optimization solvers, but they have the following drawbacks: optimization depends on given constraint weights, and there is a lack of flexible coordination mechanisms when objectives conflict; multi-objective conflicts cannot be explained or actively coordinated; optimization interpretability is poor, optimization paths are difficult to trace, and there is no way to provide interaction with doctors and support personalized adjustments; the algorithm is constrained and it is difficult to support the integration of future related technologies (such as dynamic planning, inputting multi-source data).

[0003] Current AI technologies applied in radiotherapy, such as deep learning and interactive classification systems, often prioritize predictive performance but fail to adequately consider multi-objective collaboration, the flexibility of the optimization process, and interpretability. Therefore, it is necessary to explore a swarm intelligence-like and multi-agent collaborative optimization mechanism to achieve efficient data optimization for radiotherapy planning in a distributed, adaptive, and interpretable manner. Summary of the Invention

[0004] The purpose of this invention is to provide a swarm intelligence-driven radiotherapy planning optimization system to improve the above-mentioned technical problems.

[0005] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A swarm intelligence-driven radiotherapy planning optimization system includes:

[0007] The data acquisition module is used to acquire DICOM-RT data through the data and structure layer; the DICOM-RT data includes structural data and initial radiotherapy plans;

[0008] The initialization module is used to initialize the structured data through a multi-agent optimization layer, mapping each piece of structured data to an Agent, and determining multiple Agents and their corresponding initial values ​​for state variables. Each Agent includes a corresponding optimization sub-objective, optimization objective function, and decision strategy. The state variables include dose gain value, satisfaction index, conflict point location set, negotiation willingness value, and dose parameters. The dose parameters include dose budget and conflict area dose value.

[0009] The graph structure construction module is used to determine the adjacency relationships and dose correlations between each Agent; based on each adjacency relationship and each dose correlation, the corresponding communication graph structure is constructed; the adjacency relationship includes the geometric overlap between structural data and the dose influence range;

[0010] The optimization and fusion module is used to perform information exchange, self-organizing coordination optimization, and conflict fusion on each Agent based on the communication graph structure and the state variables, optimization sub-objectives, optimization objective functions, and decision strategies of each Agent, so as to obtain dose distribution optimization data.

[0011] The conversion module is used to convert the dose distribution optimization data into a format to obtain TPS executable parameters; the TPS executable parameters are then input into the TPS system; the TPS executable parameters include beam optimization parameters and treatment equipment optimization parameters.

[0012] In the above scheme, this invention adopts a multi-agent collaborative optimization framework driven by swarm intelligence, mapping each structure to an independent agent, and constructing a communication graph through geometric and dose features to achieve precise interaction; combined with self-organizing coordination and conflict fusion mechanisms, it dynamically handles dose conflicts, introduces fuzzy logic decision-making to adapt to clinical needs, is compatible with existing TPS systems, improves the accuracy of dose optimization, and performs fine allocation of conflict areas to reduce the risk of underdose in the target area or overdose in the oral radiation area; it improves optimization efficiency, the intelligent communication graph reduces redundant interactions, self-organizing collaboration accelerates convergence, and it is easy to implement and has strong robustness.

[0013] Furthermore, the processing procedure of the graph structure construction module is as follows:

[0014] Voxel coordinate data of structural data are extracted from image data; based on the voxel coordinate data, the geometric center, overlap ratio and minimum boundary distance of the corresponding structural data are calculated.

[0015] Based on the geometric center, overlap ratio and minimum boundary distance, calculate the overlapping area / overlap range between each structural data, and determine the geometric overlap and dose influence range between each structural data.

[0016] Calculate the dose-deposition matrix; based on the dose-deposition matrix and the coordinate data of each voxel, according to the formula:

[0017] ;

[0018] ;

[0019] ;

[0020] Calculate the dose correlation between structural data; where, Indicates weight, Indicates the first The structure data to the first The dose item of the structural data, Indicates the first The structure data to the first The dose item of the structural data, Indicates the first The first structural data and the first The dose correlation between structural data, Indicates hyperparameters, Indicates the first The first structural data and the first Voxel distance between voxel coordinates corresponding to each structural data point. This represents an exponential function with base e. Indicates the first The structure data to the first Linear dose coupling operator for structured data, Denotes the norm F, Indicates the first Mask of structured data The absolute value, Represents the dose-deposition matrix. Indicates the first Voxel selection matrix for structured data. Represents the selection matrix. Represents a constant. The dimension representing the dose deposition matrix;

[0021] A communication graph construction strategy is selected; based on dose correlation, geometric overlap, and dose influence range, the communication graph structure is constructed using the selected communication graph construction strategy.

[0022] In the above scheme, this invention provides an objective basis for judging the spatial relationship between structures by extracting voxel coordinates to calculate quantitative parameters such as geometric center and overlap ratio, avoiding subjective misjudgments of geometric overlap and association, and ensuring the accuracy of subsequent association analysis; it quantifies the intensity of dose interaction between structures, rather than relying solely on spatial distance; at the same time, it supports selecting the communication graph construction strategy according to system complexity and hardware capabilities, and constructs the communication graph based on core indicators such as dose correlation and geometric overlap, which can accurately screen the agents that need to interact, avoid redundant connections of irrelevant agents, and significantly reduce the computational complexity of subsequent multi-agent collaborative optimization.

[0023] Furthermore, the construction of the communication graph structure includes:

[0024] Choose a communication graph construction strategy; the communication graph construction strategy includes the Voronoi graph space partitioning strategy and the K-nearest neighbor (KNN) sparse graph strategy;

[0025] Each Agent is used as a graph node to construct the initial communication graph structure;

[0026] Based on dose correlation, geometric overlap, and dose influence range, the initial communication graph structure is optimized by the selected communication graph construction strategy, and the edges and their weights are determined to obtain the communication graph structure.

[0027] In the above scheme, the present invention flexibly selects Voronoi graph space partitioning or K-nearest neighbor sparse graph strategy to adapt to different optimization scenarios and improve strategy adaptability; it constructs an initial communication graph with each Agent as a node to ensure that all Agents participating in the optimization are included in the topological framework and to ensure basic integrity; then it optimizes the initial graph based on core indicators such as dose correlation, geometric overlap, and dose influence range and determines the edge weights, accurately selects Agents with actual correlations and establishes connections, eliminates redundant edges of irrelevant Agents to reduce the computational load of subsequent multi-Agent collaboration, and can reflect the correlation strength through edge weights, laying the foundation for efficient information interaction between Agents.

[0028] Furthermore, when selecting the Voronoi graph space partitioning strategy, the optimized initial communication graph structure includes:

[0029] Using the geometric center of each graph node as a seed point, Voronoi subdivision is performed in 3D space to obtain Voronoi units exclusive to each Agent. If the structures corresponding to two Agents have geometric overlap or strong dose correlation, and the corresponding Voronoi units are not directly adjacent, then an edge is established between the two Agents. Based on the minimum boundary distance and dose correlation of each graph node, the edge weight between each pair of different graph nodes is determined by weighted calculation. The edge weight is applied to the corresponding edge to complete the optimization of the initial communication graph structure.

[0030] When selecting the K-Nearest Neighbors (KNN) sparse graph strategy, the optimized initial communication graph structure includes:

[0031] Based on each geometric center, calculate the Euclidean distance between each pair of different Agents to generate a distance matrix; for each Agent, initially select the K Agents with the smallest distances to establish edges; if the... If an Agent has high geometric overlap or strong dose correlation with other structural data, its K value is increased to establish edges with more relevant Agents; based on the minimum boundary distance and dose correlation of each graph node, the edge weight between each pair of different graph nodes is determined by weighted calculation; the edge weight is applied to the corresponding edge to complete the optimization of the initial communication graph structure.

[0032] In the above schemes, the Voronoi diagram spatial partitioning strategy divides the space into dedicated units using the geometric center of each agent as the seed point. It also breaks through unit boundary limitations, adding connections for agents with geometric overlap or strong dose correlations to avoid missing key associations. The K-nearest neighbor sparse graph strategy first selects the initial K associated agents based on their geometric center distance, and then dynamically increases the K value for agents with high geometric overlap or strong dose correlations, balancing connection simplicity with the integrity of key associations. Both strategies calculate edge weights based on minimum boundary distance and dose correlation, accurately quantifying the association strength between agents. The final optimized communication graph can eliminate redundant connections of irrelevant agents to reduce the computational load of subsequent collaborative optimization, while ensuring efficient connectivity for agents with actual spatial or dose correlations.

[0033] Furthermore, the processing procedure of the optimized fusion module is as follows:

[0034] Based on the graph communication structure, determine the set of neighboring agents corresponding to each agent;

[0035] The local search algorithm is used to optimize the structural regions corresponding to each Agent in the current iteration, optimize the state variables of each Agent, adjust the dose parameters and corresponding optimization objective functions of each Agent, and obtain the corresponding optimized state variables.

[0036] The optimized state variables of each Agent are packaged and used as neighbor state vectors; each neighbor state vector is broadcast to the corresponding neighbor Agent through a graph communication structure.

[0037] Based on the neighboring state vectors received by each Agent, the optimized state variables of each Agent are further optimized using a self-organizing coordination optimization algorithm and a conflict fusion mechanism to obtain the Agent after secondary optimization in the current iteration.

[0038] Based on the Agent after secondary optimization, calculate the energy value of the current iteration; based on the energy value of the current iteration, determine whether the convergence state of the current iteration is converged; if so, end the iteration and obtain the dose distribution optimization data; otherwise, enter the next iteration until the satisfaction index of all Agents exceeds the preset satisfaction threshold or the total number of conflict points is less than the structural volume ratio or the maximum number of iterations is reached, and the current iteration number is incremented by 1.

[0039] In the above scheme, the present invention first accurately locates the neighboring related objects of each Agent based on the graph communication structure, and then optimizes the state variables and dose parameters of each Agent in a targeted manner through a local search algorithm, which fits the exclusive optimization objectives of different structures and avoids the one-sidedness of global optimization. Subsequently, efficient information sharing among Agents is realized through state vector broadcasting, providing data support for collaborative decision-making. Then, a second optimization is carried out by combining self-organizing coordination optimization and conflict fusion mechanism, which effectively resolves the dose conflict that may be generated by local optimization and improves the coordination of global optimization.

[0040] Furthermore, adjusting the dosage parameters and corresponding optimization objective functions of each Agent includes:

[0041] The local search algorithm is used to optimize the structural regions corresponding to each agent in the current iteration, thereby optimizing the dose budget of each agent and obtaining the updated dose budget for the current iteration.

[0042] When the OAR satisfaction index of a neighboring agent of a certain agent is higher than the OAR satisfaction index threshold and the negotiation willingness value is lower than the OAR negotiation willingness threshold, then according to the formula:

[0043] ;

[0044] Reduce the current iteration count in the overlapping region of this Agent. PTV dose gain value ;in, Indicates the current iteration number Next The negotiation willingness value corresponding to each Agent. Indicates the adjustment rate;

[0045] If an Agent receives dosage negotiation requests from multiple neighboring Agents, then according to the formula:

[0046] ;

[0047] The optimization objective function of the Agent in the current iteration is updated to obtain the updated optimization objective function. ;in, This represents the optimization objective function before the Agent update. Indicates the penalty coefficient. Indicates the current iteration number Next The satisfaction index for each agent. , These represent the current iteration number, respectively. Next A set of Agents and their neighboring Agents The first in The set of conflict point locations corresponding to each Agent Represents the intersection-union ratio function. This represents the summation function;

[0048] Determine whether the satisfaction index of each agent is higher than the satisfaction index threshold; if so, reduce the weight of the agent whose satisfaction index is higher than the threshold.

[0049] In the above scheme, this invention optimizes the dose budget of each agent in a targeted manner through a local search algorithm, which makes the dose allocation more in line with the needs of different structures, avoids the drawbacks of "one-size-fits-all" adjustment, prioritizes the protection of OARs from excessive dose, and balances the dose requirements of the target area and organs at risk. If an agent receives multiple dose negotiation requests, by optimizing the objective function and the penalty term based on the intersection-union ratio of conflict points, the overlap of dose hotspots can be effectively suppressed, reducing dose conflicts between multiple agents. At the same time, the weight of high-satisfaction agents is dynamically reduced according to the satisfaction index, which can "make way" for structures that have not yet met the standards, prevent dominant agents from over-consuming dose resources, and ensure a more balanced global optimization.

[0050] Furthermore, the secondary optimization of the optimized state variables of each Agent using the self-organizing coordination optimization algorithm and conflict fusion mechanism includes:

[0051] The conflict fusion mechanism is used to detect conflicts in each agent, update the conflict point positions of each agent, and calculate the corresponding dose conflict tensor.

[0052] The edge dose of each Agent in a region with a conflict is updated using a self-organizing coordination optimization algorithm. The dose budget of each Agent is updated using a conflict fusion mechanism to obtain the updated edge dose and the updated dose budget for the current iteration.

[0053] The dose recommendation value is determined by using the conflict fusion mechanism, which yields the dose sacrifice in the current iteration;

[0054] Based on the updated edge dose, updated dose budget, and dose sacrifice, the optimized state variables of each Agent are updated to obtain the Agent after secondary optimization in the current iteration.

[0055] In the above scheme, the conflict fusion mechanism accurately detects the conflict points between agents and calculates the dose conflict tensor, providing a quantitative basis for conflict handling; the self-organizing coordination optimization algorithm updates the dose at the edge of the conflict region and adjusts the dose budget in combination with the conflict fusion mechanism, which can resolve local dose contradictions in a targeted manner and avoid the one-sidedness of single agent optimization; the dose sacrifice degree is determined by the conflict fusion mechanism, which can balance the dose requirements of each agent while protecting key structures (such as OAR), and avoid excessive sacrifice of the optimization goal of a certain structure.

[0056] Further, the calculation of the corresponding dose conflict tensor includes:

[0057] Based on the optimized state variables of each Agent and the received neighboring state vectors, that is, by using the set of conflict point locations corresponding to the optimized state variables of each Agent and the received neighboring state vectors, the voxel regions with spatial overlap or dose conflict are determined.

[0058] Calculate the number of voxels in voxel regions with spatial overlap or dose conflict; determine whether the number of each voxel exceeds the minimum voxel threshold; if so, determine that the two Agents corresponding to the voxel region have significant conflict, that is, there is a conflict region and enter the coordination list.

[0059] According to the formula:

[0060] ;

[0061] Calculate the dose conflict tensor of each voxel in a voxel region for agents with significant pairwise conflicts. ;in, , These represent the expected PTV dose and the maximum permissible OAR dose for a given voxel, respectively. , These represent indicator functions for the PTV and OAR regions, respectively, indicating whether a voxel is a PTV or an OAR. Represents voxels, Indicates the current dose of a voxel.

[0062] In the above scheme, the dose conflict tensor calculation process has significant benefits, specifically: it accurately locates voxel regions with spatial overlap or dose conflict based on the optimized state and neighboring state vectors of each agent, ensuring the targeting of conflict detection; it filters out significant conflict regions by using a voxel number threshold and includes them in the coordination list, avoiding minor conflicts from interfering with optimization priorities and focusing on key contradictions that need to be addressed first; and it uses a formula combined with the expected dose of PTV, the maximum allowable dose of OAR, and the indicator function of the voxel's region to quantitatively calculate the dose conflict tensor of each conflicting voxel, intuitively reflecting the degree of dose conflict between different structures (PTV and OAR) on the same voxel, providing a precise quantitative basis for subsequent self-organized coordination optimization and conflict fusion, ensuring that conflict handling is more targeted, and effectively improving the efficiency and accuracy of dose conflict resolution in radiotherapy planning.

[0063] Furthermore, the step of updating the edge dose for agents in pairs with conflicting regions using a self-organizing coordination optimization algorithm includes:

[0064] Information exchange is performed between Agents that have conflicting regions; that is, the two Agents exchange corresponding shared information. The shared information includes the local dose gradient of the current iteration, the target dose requirement, the optimization status, and historical dose transfer records.

[0065] The dose transfer weights are updated based on the dose conflict tensor of each voxel in the conflict region.

[0066] Based on the updated dose budget of each dose transfer weight, each local dose gradient, and the conflict fusion mechanism in the previous iteration, the edge dose distribution of the corresponding conflict region is adjusted to obtain the updated edge dose.

[0067] In the above scheme, the present invention eliminates information barriers and ensures information symmetry in collaborative optimization by exchanging local dose gradients and target requirements with agents in conflict areas. The dose transfer weights are updated based on the dose conflict tensor, so that the weights accurately match the degree of voxel conflict (e.g., high-conflict voxels correspond to high weights), ensuring that dose adjustment resources are prioritized for areas with prominent conflicts. The edge dose distribution is then adjusted by combining the transfer weights, local dose gradients, and the dose budget from the previous iteration. This not only specifically alleviates the dose conflict in conflict areas but also connects with historical optimization results to avoid repeated adjustments. Ultimately, the updated edge dose distribution is more in line with clinical dose requirements, significantly improving the accuracy of conflict resolution and the efficiency of collaborative optimization.

[0068] Furthermore, the method of updating the dose budget for each agent using the conflict fusion mechanism includes:

[0069] Determine if there is an intersection between the dose budgets of the two agents corresponding to the conflict area; if so, determine if the current dose in each conflict area belongs to the intersection of the dose budgets of the two agents corresponding to the conflict area; otherwise, use the target offset mechanism to adjust the tolerance boundary of the dose budget of the two agents corresponding to the conflict area, expand or compress the budget within the allowable range, and complete the update of the dose budget of each agent.

[0070] If the current dose in each conflict zone is the intersection of the dose budgets of the two agents corresponding to the conflict zone, then the budget allocation factor is calculated based on historical sacrifice records, target importance, and remaining budget; based on the budget allocation factor, the current dose in each conflict zone is updated to obtain the updated dose;

[0071] If the current dose in each conflict region belongs to the intersection of the dose budgets of the two agents corresponding to the conflict region, then determine whether the updated dose satisfies the corresponding optimization sub-objective; if so, reduce the dose budget of the two agents corresponding to the conflict region based on the edge dose distribution of the previous iteration; otherwise, increase the dose budget of the two agents corresponding to the conflict region based on the edge dose distribution of the previous iteration; and complete the update of the dose budget for each agent.

[0072] In the above scheme, this invention first determines whether there is an overlap in the agent dose budgets in the conflict area, and then flexibly adopts different processing strategies: when there is no overlap, a target offset mechanism is used to adjust the budget tolerance boundary to avoid optimization deadlock caused by hard constraints; when there is overlap, the allocation factor is calculated by combining historical sacrifice records, target importance, and remaining budget, which can fairly allocate budget resources and take into account structural priority and historical contribution. Then, the budget is dynamically adjusted according to whether the updated dose meets the optimization sub-objective. If the target is met, the budget is reduced to avoid resource waste; if the target is not met, the budget is increased to promote optimization, and finally, the dose requirements of PTV and OAR in the conflict area are accurately balanced. Attached Figure Description

[0073] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is a system structure diagram in an embodiment of the present invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0076] Please see Figure 1 This embodiment provides a swarm intelligence-driven radiotherapy planning optimization system, which includes:

[0077] The data acquisition module is used to acquire DICOM-RT data through the data and structure layer. The DICOM-RT data includes structural data and an initial radiotherapy plan. The structural data is drawn from the image data and includes the contour data of the planned target volume (PTV) and organs at risk (OAR). The image data includes CT images and PET-CT images. The initial radiotherapy plan is based on the structural data and includes beam parameters, dose distribution data, and treatment equipment parameters. The beam parameters include beam energy, dose rate, and beam direction. The treatment equipment information includes the MLC (Multi-Leaf Collimator) blade position sequence and a control point table, which includes gantry angles and collimator angles.

[0078] The initialization module is used to initialize the structural data through a multi-agent optimization layer, mapping each structural data to an Agent and determining multiple Agents and their corresponding initial values ​​for state variables. Each Agent includes a corresponding optimization sub-objective, optimization objective function, and decision strategy. The state variables include dose gain value, satisfaction index, conflict point location set, negotiation willingness value, and dose parameters. The dose parameters include dose budget and conflict area dose value. In this embodiment, the initial values ​​of dose gain value, satisfaction index, conflict point location set, and negotiation willingness value are all set to 0. The initial values ​​of the dose parameters are determined by the initial radiotherapy plan. For example, the dose budget for PTV is usually a high dose budget, which can be set to [68, 74] Gy, and the dose budget for OAR is usually a low dose budget, which can be set to [0, 50] Gy.

[0079] Specifically, structural data forms the basis for Agent initialization. Each structure is mapped to an Agent, such as a PTV or an OAR. This Agent carries the dose optimization objectives for the corresponding structure (e.g., dose coverage requirements for the PTV, dose limitation requirements for the OAR), and performs local optimization and collaborative negotiation based on these objectives. Optimization sub-objectives include improving dose distribution uniformity within the target area and reducing the maximum spinal cord dose. The optimization objective function includes the D function of the PTV. 95 (The dose received by 95% of the target volume) is not less than 95% of the prescribed dose, and the D of a certain OAR max The maximum dose accepted is less than 20 Gy. Decision-making strategies include local genetic algorithms and Q-learning algorithms.

[0080] The dose gain value is used to determine the satisfaction index, conflict point location set, and negotiation willingness value during the optimization iteration process.

[0081] This invention maps each structure (such as PTV and OAR) to an independent Agent. Each Agent carries the exclusive optimization goal for the corresponding structure (such as dose coverage for PTV and dose limitation for OAR). The dose parameters are adjusted in a targeted manner through a local search algorithm, avoiding the drawbacks of the "one-size-fits-all" approach in traditional global optimization, so that the dose distribution is more in line with the clinical needs of different structures.

[0082] The graph structure construction module is used to determine the adjacency relationships and dose correlation between each Agent; based on each adjacency relationship and each dose correlation, the corresponding communication graph structure is constructed; the adjacency relationship includes the geometric overlap between structural data and the dose influence range.

[0083] The processing procedure of the graph structure construction module is as follows:

[0084] S3-1. Extract voxel coordinate data of structural data based on image data; calculate the geometric center, overlap ratio and minimum boundary distance of the corresponding structural data based on each voxel coordinate data.

[0085] Geometric center, overlap ratio, and minimum boundary distance are the core quantitative parameters for determining adjacency relationships. For geometric overlap, the overlap ratio is a direct quantitative indicator. By calculating the percentage of overlapping voxels between two structures (such as PTV and OAR), it directly reflects the degree of spatial overlap; the higher the overlap ratio, the more significant the geometric overlap. The minimum boundary distance is used to assist in judging the state of geometric overlap. If the distance is 0, it indicates that the structures have geometrically overlapped; if the distance is greater than 0, there is no direct overlap. The geometric center, through the relative position between centers (such as whether they are within each other's geometric range), can preliminarily determine whether there is a potential overlap trend in the structures. For example, structures with a high degree of overlap in center coordinates are more likely to have geometric overlap.

[0086] S3-2. Based on the geometric center, overlap ratio and minimum boundary distance, calculate the overlapping area / overlap range between each structural data, and determine the geometric overlap and dose influence range between each structural data.

[0087] The dose influence range refers to the spatial extent in which a dose change corresponding to one structural data point affects other structural data points. Its definition requires consideration of geometric relationships and dose physics characteristics. Therefore, the geometric center, overlap ratio, and minimum boundary distance are key parameters for determining the dose influence range. Specifically, the geometric center is determined by calculating the distance between the centers of different structures, which can initially divide the area into a "close-range influence zone" (structures with close center distances have more significant dose interactions) and a "far-range low-influence zone," providing a spatial benchmark for dose correlation calculations (such as exponential decay models).

[0088] Therefore, the overlapping area / overlap range between structural data and the combined overlap and dose influence range between structural data are calculated by intersecting three-dimensional geometric bodies; since the method of obtaining the overlap range of intersecting geometric bodies by intersecting three-dimensional geometric bodies is an existing technology, it will not be described in detail.

[0089] S3-3, Calculate the dose deposition matrix; based on the dose deposition matrix and the coordinate data of each voxel, calculate the dose correlation between structural data;

[0090] The formula for dose correlation is:

[0091] ;

[0092] ;

[0093] ;

[0094] Calculate the dose correlation between structural data; where, Indicates weight, Indicates the first The structure data to the first The dose item of the structural data, Indicates the first The structure data to the first The dose item of the structural data, Indicates the first The first structural data and the first The dose correlation between structural data, Indicates hyperparameters, Indicates the first The first structural data and the first Voxel distance between voxel coordinates corresponding to each structural data point. This represents an exponential function with base e. Indicates the first The structure data to the first Linear dose coupling operator for structured data, Denotes the norm F, Indicates the first Mask of structured data The absolute value, Represents the dose-deposition matrix. Indicates the first Voxel selection matrix for structured data. Indicates the first Selection matrix for structured data Represents a constant. This represents the dimension of the dose deposition matrix.

[0095] in, When beamlet weight When the value is 1, the dose deposition matrix is ​​obtained. Select matrix For the first The set of valid beamlet variables for the nth structured data, for those belonging to the nth structured data. The beamlet for the illumination region corresponding to each structure data is set to 1, and the others are set to 0.

[0096] S3-4. Select a communication graph construction strategy; based on dose correlation, geometric overlap, and dose influence range, construct the communication graph structure using the selected communication graph construction strategy; the communication graph construction strategy includes Voronoi graph space partitioning strategy and K-nearest neighbor (KNN) sparse graph strategy;

[0097] S3-4 includes:

[0098] S3-4-1. Select a communication graph construction strategy based on actual usage requirements and scenarios;

[0099] S3-4-2. Construct the initial communication graph structure by treating each Agent as a graph node;

[0100] S3-4-3. Based on dose correlation, geometric overlap and dose influence range, the initial communication graph structure is optimized by the selected communication graph construction strategy, the edges and their weights are determined, and the communication graph structure is obtained.

[0101] When the Voronoi graph space partitioning strategy is selected, S3-4-3 includes:

[0102] Using the geometric center of each graph node as a seed point, Voronoi subdivision is performed in 3D space to obtain Voronoi units (spatial regions) exclusive to each Agent. If the structures corresponding to two Agents have geometric overlap (overlap ratio > 0) or strong dose correlation (dose correlation degree > threshold), and the corresponding Voronoi units are not directly adjacent, then an edge is established between the two Agents (to ensure that Agents in conflict regions can communicate).

[0103] Based on the minimum boundary distance and dose correlation of each graph node, the edge weights between different graph nodes are determined by weighted calculation; the edge weights are then applied to the corresponding edges to complete the optimization of the initial communication graph structure.

[0104] When the K-Nearest Neighbors (KNN) sparse graph strategy is selected, S3-4-3 includes:

[0105] Based on each geometric center, calculate the Euclidean distance between each pair of different Agents to generate a distance matrix; for each Agent, initially select the K Agents with the smallest distances to establish edges; if the... If an Agent has high geometric overlap (overlap ratio > 20%) or strong dose correlation (dose correlation degree > 0.8) with other structural data, its K value is increased to ensure that it establishes edges with more relevant Agents (to avoid missing critical conflicts).

[0106] Based on the minimum boundary distance and dose correlation of each graph node, the edge weights between different graph nodes are determined by weighted calculation; the edge weights are then applied to the corresponding edges to complete the optimization of the initial communication graph structure.

[0107] This invention constructs a communication graph based on geometric features (overlap ratio, minimum boundary distance) and dose correlation, allowing only agents with spatial or dose correlation to establish connections (such as Voronoi partitioning or KNN strategy), avoiding redundant interactions between irrelevant agents, significantly reducing computational complexity and improving optimization efficiency.

[0108] The optimization and fusion module is used to perform information exchange, self-organizing coordination optimization, and conflict fusion on each Agent based on the communication graph structure and the state variables, optimization sub-objectives, optimization objective functions, and decision strategies of each Agent, so as to obtain dose distribution optimization data.

[0109] Specifically, information exchange, self-organizing coordination optimization, and conflict fusion among the agents are essentially iterative optimizations of each agent. In each iteration, information exchange, self-organizing coordination optimization, and conflict fusion are performed on each agent until the optimization sub-objective is achieved, the optimization objective function satisfies the iteration conditions, or the maximum number of iterations is reached. For example, the dose distribution optimization data from the last iteration is used as the iterative optimization result.

[0110] In addition, in each iteration, the dose gain value is used to determine the improvement of the corresponding optimization objective function by the dose change of each Agent. The satisfaction index is used to determine the degree of satisfaction of each Agent with the dose optimization result of the current iteration, reflecting optimization pressure; the lower the satisfaction index, the higher the willingness to negotiate. The conflict point location set is used to determine the voxel region where each Agent has dose conflicts with its neighboring Agents. The willingness to negotiate value is used to determine whether each Agent is willing to relinquish a portion of its dose budget. The formulas corresponding to the dose gain value, satisfaction index, and willingness to negotiate value are as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] in, , Let represent the objective function of the previous iteration and the objective function of the current iteration, respectively. Indicates the first The dose gain value for each agent, Indicates the first Satisfaction metrics for each agent. , They represent the first The optimization objective function of each agent and the initial value of the optimization objective function. Represents the maximum value function. Indicates the first The negotiation willingness value of each agent.

[0115] Therefore, the processing procedure of the optimization fusion module is as follows:

[0116] S4-1. Based on the graph communication structure, determine the set of neighboring agents corresponding to each agent. ;

[0117] S4-2. Optimize the structural regions corresponding to each Agent in the current iteration using a local search algorithm, optimize the state variables of each Agent, adjust the dose parameters and corresponding optimization objective functions of each Agent, and obtain the corresponding optimized state variables.

[0118] S4-2 includes:

[0119] S4-2-1. Optimize the structural regions corresponding to each Agent in the current iteration using a local search algorithm, optimize the dose budget of each Agent, and obtain the updated dose budget for the current iteration.

[0120] Specifically, when the neighboring agent of an agent corresponding to a certain PTV is the agent corresponding to OAR, and the OAR satisfaction index is higher than the OAR satisfaction index threshold and the OAR negotiation willingness value is lower than the OAR negotiation willingness threshold, then according to the formula:

[0121] ;

[0122] Reduce the current iteration count in the overlapping region of this Agent. PTV dose gain value ;in, Indicates the current iteration number Next The negotiation willingness value corresponding to each Agent. This indicates the adjustment rate.

[0123] S4-2-2. If an Agent receives dosage negotiation requests from multiple neighboring Agents, then according to the formula:

[0124] ;

[0125] The optimization objective function of the Agent in the current iteration is updated by introducing a penalty term. This is beneficial for suppressing dose hotspot overlap and obtaining the updated optimization objective function. .in, This represents the optimization objective function before the Agent update. Indicates the penalty coefficient. Indicates the current iteration number Next The satisfaction index for each agent. , These represent the current iteration number, respectively. Next A set of Agents and their neighboring Agents The first in The set of conflict point locations corresponding to each Agent Represents the intersection-union ratio function. This represents the summation function;

[0126] Among them, dose request negotiation means that the dose at the current location has not reached the target value. Conflict concentration means that a large number of voxel doses in a certain agent have not reached the target value and the gap is large. In this case, it is necessary to increase the penalty for the area, that is, to increase the weight in S4-2-3.

[0127] S4-2-3. Determine whether the satisfaction index of each Agent is higher than the satisfaction index threshold; if so, reduce the weight of the Agent whose satisfaction index is higher than the threshold (the weight of the optimization objective function corresponding to the Agent) to make way for unmet structural data.

[0128] S4-3. Pack the optimized state variables of each agent and use them as the neighbor state vector. The graph communication structure connects the state vectors of each neighboring state. The broadcast is sent to the corresponding neighboring Agents, enabling each Agent to share information with its neighboring Agents;

[0129] S4-4. Based on the neighboring state vectors received by each Agent, the optimized state variables of each Agent are further optimized using a self-organizing coordination optimization algorithm and a conflict fusion mechanism to obtain the Agent after secondary optimization in the current iteration.

[0130] The self-organizing coordination optimization algorithm adopts an edge negotiation strategy: guiding dose transfer and sharing target weights in conflict areas to form a non-centralized coordination scheme. It aims to solve the problem of dose allocation conflict when multiple target structures (such as PTV and OAR) overlap or are adjacent in space, especially in the overlapping boundary area, to avoid one target unilaterally sacrificing or dominating.

[0131] In each iteration, the self-organizing coordination optimization algorithm and the conflict fusion mechanism alternate. The state exchange of the self-organizing coordination optimization algorithm provides input to the conflict fusion mechanism, and the result of the conflict fusion mechanism, in turn, modifies the strategy of the self-organizing coordination optimization algorithm, forming a closed loop of "conflict processing triggered within the protocol framework, and conflict processing results feeding back into protocol adjustment".

[0132] Therefore, S4-4 includes:

[0133] S4-4-1. Use the conflict fusion mechanism to perform conflict detection on each Agent, update the conflict point position corresponding to each Agent, and calculate the corresponding dose conflict tensor.

[0134] S4-4-1 includes:

[0135] Based on the optimized state variables of each agent and the received neighboring state vectors, i.e., using the set of conflict point locations corresponding to the optimized state variables of each agent and the received neighboring state vectors, the voxel regions with spatial overlap or dose conflict are determined; generally, the structural data / Agent with dose conflict and another structural data / Agent are PTV and OAR, respectively.

[0136] Calculate the number of voxels in spatially overlapping or dose-conflicting voxel regions; determine whether the number of each voxel exceeds the minimum voxel threshold; if so, determine that the two Agents corresponding to the voxel region have significant conflict and enter the coordination list; where the minimum voxel threshold is 20 voxels.

[0137] The formula for calculating the dose conflict tensor of each voxel in a voxel region of an Agent with significant pairwise conflicts is as follows:

[0138] ;

[0139] in, , These represent the expected PTV dose and the maximum permissible OAR dose for a given voxel, respectively. , They represent, Represents voxels, Indicates the current dose of a voxel. The dose conflict tensor of a voxel.

[0140] S4-4-2. Use the self-organizing coordination optimization algorithm to update the edge dose for Agents with conflict regions between each other, and use the conflict fusion mechanism to update the dose budget of each Agent to obtain the updated edge dose and the updated dose budget in the current iteration.

[0141] In S4-4-2, edge dose and dose budget are updated simultaneously. Edge dose adjustment by the self-organizing coordination optimization algorithm will affect the budget allocation in the conflict fusion mechanism. For example, if the edge dose of an agent's PTV is reduced too much, the conflict domain budget may be appropriately relaxed. Conversely, the budget result of the conflict fusion mechanism will also constrain the adjustment range of the protocol. For example, if the budget limit is 45 Gy, the agent's OAR will reject dose recommendations in the protocol that exceed this value.

[0142] Therefore, the step of updating the edge dose for agents in pairs with conflicting regions using the self-organizing coordination optimization algorithm includes:

[0143] For agents with conflicting regions, information exchange occurs; that is, the two agents exchange corresponding shared information. This shared information includes the local dose gradient of the current iteration, the target dose requirement, the optimization status, and historical dose concession records. The optimization status includes the number of optimization failures and the dose achievement rate. The local dose gradient is an intermediate quantity calculated during the optimization process, representing the trend of dose distribution changes within the current conflict region. The target dose requirement refers to the ideal dose or dose distribution set for each agent. Historical dose concession records indicate whether an agent has made any sacrifices and the extent of those sacrifices.

[0144] Based on the dose conflict tensor of each voxel in the conflict region, the dose transfer weights of the two agents corresponding to the dose conflict tensor are updated. The dose transfer weights determine how the dose in the conflict region is "transferred to which side." The update formula for the dose transfer weights of the two agents is as follows:

[0145] ;

[0146] ;

[0147] in, , These represent the first and second domains where conflict exists. The Agent and the The satisfaction index for each agent. , These represent the first and second domains where conflict exists. The Agent, the Dosage transfer weight for each agent , They represent the first The Agent and the In the conflict domain where the agent exists, the first... Individual elements and the first Dosage conflict tensor of individual elements.

[0148] Based on the updated dose budget derived from the dose transfer weights, local dose gradients, and the conflict fusion mechanism from the previous iteration, the edge dose distribution of the corresponding conflict region is adjusted to obtain the updated edge dose. For planning systems such as IMRT / VMAT, FMO or DAO optimization is performed to adjust subfield weights, blade position or angle priorities, etc. For dose field simulation systems, the penalty or weight terms of each agent in the objective function of that region are updated.

[0149] Among them, the updated dose budget of the conflict fusion mechanism in the previous iteration serves as a constraint for adjusting the edge dose distribution of the conflict region in the current iteration. The adjusted edge dose cannot exceed the updated dose budget in the previous iteration.

[0150] The Agent of this invention achieves dynamic coordination through state exchange (dose gain, satisfaction, and willingness to negotiate), and can complete edge dose adjustment and objective function correction (such as introducing conflict penalty terms) without central control, forming a closed loop of "local optimization-global coordination", reducing human intervention and accelerating convergence.

[0151] In conflict zones, PTV and OAR are difficult to achieve simultaneously. Therefore, a "dose budget sharing" mechanism needs to be introduced. Agents negotiate their respective tolerable dose increase / decrease ranges, allowing each agent to allocate a "tolerable dose" space. Thus, the method of updating the dose budget for each agent using the conflict fusion mechanism includes:

[0152] Determine if there is an intersection between the dose budgets of the two agents corresponding to the conflict area; if so, determine if the current dose in each conflict area belongs to the intersection of the dose budgets of the two agents corresponding to the conflict area; otherwise, use the target offset mechanism to adjust the tolerance boundary of the dose budget of the two agents corresponding to the conflict area, expanding or compressing the budget within the allowable range.

[0153] If the current dose in each conflict zone belongs to the intersection of the dose budgets of the two agents corresponding to the conflict zone, then the budget allocation factor is calculated based on the sacrifice level in the historical dose transfer record, the weight of the optimization objective function corresponding to the agent, and the remaining budget. Based on the budget allocation factor, update the current dose in each conflict zone to obtain the updated dose; calculate the budget allocation factor using the same formula as the dose transfer weight, i.e.:

[0154] .

[0155] If the current dose in each conflict zone does not belong to the intersection of the dose budgets of the two agents corresponding to the conflict zone, the original dose budget is retained.

[0156] Determine whether the updated dose satisfies the corresponding optimization sub-objective; if so, reduce the dose budget of the two agents corresponding to the conflict region based on the edge dose distribution of the previous iteration; otherwise, increase the dose budget of the two agents corresponding to the conflict region based on the edge dose distribution of the previous iteration; complete the update of the dose budget for each agent.

[0157] This invention accurately identifies overlapping or dose-conflicting regions through a conflict fusion mechanism. By combining fuzzy logic decision-making and dose budget sharing, it achieves refined dose allocation in conflict regions. By balancing the needs of PTV and OAR through the allocation factor, it reduces the risk of "target underdose" or "OAR overdose" and improves treatment safety.

[0158] S4-4-3. Use the conflict fusion mechanism to determine the dose recommendation value, that is, to obtain the dose sacrifice under the current iteration;

[0159] Dose conflicts between different structural data (such as PTV and OAR) are often difficult to handle using traditional hard constraints. This system introduces fuzzy logic rules, comprehensively considering the priority of dose constraints, the importance level of the organ / target area, and the current dose deviation, and dynamically outputs suggested dose adjustment target values ​​or local dose adjustment directions based on these inputs. Therefore, S4-4-3 includes:

[0160] Set up fuzzy logic rules and obtain the structural importance level of each Agent.

[0161] The inputs to the fuzzy logic rules are the current dose deviation, structural importance level, and dose constraint priority of each Agent, and the output is the initial dose sacrifice. The dose constraint priority is divided into low, medium, and high, the structural importance level is divided into minor, medium, and critical, and the current dose deviation is divided into underdose, appropriate, and excessive.

[0162] The specific content of the fuzzy logic rules is as follows: (1) If the dose constraint priority, structural importance level, and current dose deviation are high, critical, and excessive, respectively, then the initial dose sacrifice is "reduced more"; (2) If the dose constraint priority, structural importance level, and current dose deviation are medium, moderate, and appropriate, respectively, then the initial dose sacrifice is "maintained"; (3) If the dose constraint priority, structural importance level, and current dose deviation are low, minor, and insufficient, respectively, then the initial dose sacrifice is "slightly increased"; (4) If the dose constraint priority, structural importance level, and current dose deviation are high, critical, and insufficient, respectively, then the initial dose sacrifice is "increased more". Among them, the dose sacrifice refers to the difference between the actual dose and the target dose when the actual dose of an Agent does not reach the target dose.

[0163] Fuzzy logic rules can be set based on actual usage scenarios. For example, among end users, different applicable scenarios can configure these rules according to their own habits. Alternatively, they can be set in the simplest way, for example, increasing the weight of the corresponding Agent by increasing the priority of the dose constraint. The structural importance level refers to the importance level of the organs involved by the Agent, and this importance level is determined based on actual needs and relevant regulations.

[0164] Calculate the difference between the optimized dose in the state variables of each agent and the dose of its optimized sub-objective; determine the current dose deviation based on the dose difference; if the dose difference is within the first dose difference range, the current dose deviation is underdose; if the dose difference is within the second dose difference range, the current dose deviation is appropriate; if the dose difference is within the third dose difference range, the current dose deviation is excessive. The first, second, and third dose difference ranges need to be set according to different diseases.

[0165] Membership functions are used to map the current dose deviation, structural importance level, and dose constraint priority of each agent to obtain the corresponding fuzzy set; for example, if the dose deviation is +6Gy, then its "excessive" membership degree is 0.8 and "appropriate" is 0.2.

[0166] Fuzzy multiplication is performed on fuzzy sets to obtain the applicability of the corresponding rules; based on fuzzy logic rules, the inference results corresponding to the fuzzy sets are obtained.

[0167] Perform fuzzy aggregation (such as maximization or weighted average) on all inference results to obtain the corresponding fuzzy aggregation result;

[0168] Based on the fuzzy aggregation results, the initial dose sacrifice is mapped back to the actual numerical range using the centroid method or the centroid method to obtain the corresponding dose sacrifice. For example, if the initial dose sacrifice is "slightly reduced", then the dose sacrifice is -1.5Gy, which means that the dose of the corresponding Agent should be reduced by 1.5Gy.

[0169] In addition, the Agent records the "dose sacrifice" in this round; if an Agent sacrifices too much in a row, it can obtain higher priority in the next round of edge negotiation; all dose adjustment suggestions are checked by the central controller to see if they violate the mandatory dose constraints, and if they exceed the maximum allowable dose of OAR, they are rolled back.

[0170] S4-4-4. Based on the updated edge dose, updated dose budget, and dose sacrifice, update the optimized state variables of each Agent to obtain the Agent after secondary optimization in the current iteration.

[0171] S4-5. Based on the Agent after the second optimization, calculate the energy value of the current iteration; based on the energy value of the current iteration, determine whether the convergence state of the current iteration is converged; if so, end the iteration and obtain the dose distribution optimization data; otherwise, proceed to S4-6 and increment the current iteration count by 1; the initial value of the current iteration count is 0.

[0172] The formula corresponding to the energy value is:

[0173] ;

[0174] in, This represents the total number of Agents. Indicates the current iteration number The total energy of all agents. Indicates the current iteration number Next The energy weight corresponding to each Agent Indicates the current iteration number Next The energy (global cost function) corresponding to each Agent.

[0175] In addition to the current agent's dose index, the formula for energy value can be extended to support other types, such as spatial penalty terms (e.g., adding conflict penalty terms in the dose overlap region of the PTV-OAR boundary), interaction penalty terms (introducing cross-Agent terms to reflect dose influence transmission and negotiation failure penalties), evolutionary indicators, etc. (the rate of change of historical energy trajectories can be added as an inertia factor).

[0176] S4-6. Repeat S4-2 to S4-5 until the satisfaction index of all agents exceeds the preset satisfaction threshold, the proportion of conflict points in the structure is less than the preset threshold, or the maximum number of iterations is reached, to obtain the dose distribution optimization data. The structure volume ratio is calculated from the structure data; the volume ratio is determined based on actual requirements.

[0177] The conversion module is used to convert the dose distribution optimization data into a format to obtain TPS executable parameters; the TPS executable parameters are then input into the TPS system; the TPS executable parameters include beam optimization parameters and treatment equipment optimization parameters.

[0178] The dose distribution optimization data is converted into a format acceptable to the TPS system, such as converting the dose distribution optimization data into a DICOMRTPlan file, and supporting API or script-level interaction with mainstream TPS systems.

[0179] In addition, this device can extract structural data from the latest CT image data in real time to update overlapping areas, which is beneficial for determining the direction of iterative optimization in iterative optimization.

[0180] In summary, this invention employs a multi-agent collaborative optimization framework driven by swarm intelligence, mapping each structure to an independent agent. It constructs a communication graph using geometric and dose features to achieve precise interaction. Combining self-organizing coordination and conflict fusion mechanisms, it dynamically handles dose conflicts, introduces fuzzy logic decision-making to adapt to clinical needs, is compatible with existing TPS systems, improves dose optimization accuracy, and performs fine-grained allocation of conflict regions, reducing the risk of underdose in the target area or overdose in the oral radiation area (OAR). It also improves optimization efficiency, with the intelligent communication graph reducing redundant interactions, self-organizing collaboration accelerating convergence, and is easy to implement and robust.

[0181] It should be noted that, Figure 1 The execution entity of the method corresponding to the system shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not impose any limitations on this.

[0182] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0183] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A swarm intelligence driven radiotherapy plan optimization system, characterized in that, The method comprises the following steps: a data acquisition module is configured to acquire DICOM-RT data from a data and structure layer; the DICOM-RT data comprises structure data and an initial radiotherapy plan; an initialization module is configured to initialize the structure data through a multi-agent optimization layer, map each structure data to an agent, determine a plurality of agents and initial values of corresponding state variables; each agent comprises a corresponding optimization sub-goal, an optimization objective function and a decision strategy; a state variable comprises a dose gain value, a satisfaction index, a conflict point position set, a negotiation willingness value and a dose parameter; the dose parameter comprises a dose budget and a conflict region dose value; a graph structure construction module is configured to determine adjacency relationships between the agents and dose correlation degrees; a communication graph structure is constructed based on the adjacency relationships and the dose correlation degrees; the adjacency relationships comprise geometric overlaps and dose influence ranges between the structure data; an optimization fusion module is configured to exchange information, self-organize and coordinate optimization and conflict fusion among the agents based on the communication graph structure and the state variables, the optimization sub-goals, the optimization objective functions and the decision strategies of the agents, and obtain dose distribution optimization data; a conversion module is configured to convert the dose distribution optimization data into TPS executable parameters; the TPS executable parameters are input into a TPS system; the TPS executable parameters comprise beam optimization parameters and treatment device optimization parameters; a formula corresponding to the dose correlation degree is as follows: ； ； ； Calculate the dose correlation between structural data; where, Indicates weight, Indicates the first The structure data to the first The dose item of the structural data, Indicates the first The structure data to the first The dose item of the structural data, Indicates the first The first structure data and the first The dose correlation between structural data, Indicates hyperparameters, Indicates the first The first structure data and the first Voxel distance between voxel coordinates corresponding to each structural data point. This represents an exponential function with base e. Indicates the first The structure data to the first Linear dose coupling operator for structured data, Denotes the norm F, Indicates the first Mask of structured data The absolute value, Represents the dose-deposition matrix. Indicates the first Voxel selection matrix for structured data. Represents the selection matrix. Represents a constant. The dimension representing the dose deposition matrix; a processing procedure of the optimization fusion module is as follows: based on the graph communication structure, a set of adjacent agents corresponding to each agent is determined; a local search algorithm is used to optimize structure regions corresponding to each agent at a current iteration, optimize state variables of each agent, adjust dose parameters and corresponding optimization objective functions of each agent, and obtain corresponding optimized state variables; the optimized state variables of each agent are packaged and used as adjacent state vectors; the adjacent state vectors are broadcast to corresponding adjacent agents through the graph communication structure; based on the adjacent state vectors received by each agent, a self-organizing and coordinating optimization algorithm and a conflict fusion mechanism are used to perform secondary optimization on the optimized state variables of each agent, and a secondary optimized agent at the current iteration is obtained; based on the secondary optimized agent, an energy value at the current iteration is calculated; based on the energy value at the current iteration, it is determined whether a convergence state at the current iteration is convergent; if yes, the iteration is ended, and the dose distribution optimization data is obtained; otherwise, the next iteration is entered, and the current iteration number is increased by 1 until the satisfaction index of all agents exceeds a preset threshold, the total number of conflict points is less than a structure volume ratio, or a maximum iteration number is reached; the secondary optimization on the optimized state variables of each agent using the self-organizing and coordinating optimization algorithm and the conflict fusion mechanism comprises: Conflict fusion mechanism is used to detect conflicts of each Agent, update the position of conflict points corresponding to each Agent, and calculate the corresponding dose conflict tensor; the formula corresponding to the dose conflict tensor is: ； wherein, , respectively denote the PTV desired dose and the OAR maximum allowed dose for a certain voxel, , respectively denote the indicator functions for the PTV and OAR regions, i.e. marking whether a voxel is PTV or OAR, denotes a voxel, denotes the current dose for a certain voxel, denotes the dose conflict tensor for each voxel in the voxel region of Agents that significantly conflict with each other. Self-organizing coordination optimization algorithm is used to update the edge dose of the Agent in the conflict region, and the conflict fusion mechanism is used to update the dose budget of each Agent, to obtain the updated edge dose and updated dose budget under the current iteration; The conflict fusion mechanism is used to determine the dose recommendation value, i.e. to obtain the dose sacrifice degree under the current iteration; Based on the updated edge dose, updated dose budget and dose sacrifice degree, the optimized state variables of each Agent are updated to obtain the twice-optimized Agent under the current iteration.

2. The swarm intelligence driven radiotherapy plan optimization system of claim 1, wherein, The processing process of the graph structure construction module is: Based on the image data, the voxel coordinate data of the structure data is extracted; based on each voxel coordinate data, the geometric center, overlap ratio and minimum boundary distance of the corresponding structure data are calculated; Based on the geometric center, overlap ratio and minimum boundary distance, the overlapping region / range between each structure data is calculated, and the geometric overlap and dose influence range between each structure data are determined; The dose deposition matrix is calculated; Based on the dose deposition matrix and each voxel coordinate data, the dose correlation degree between the structure data is calculated; A communication graph construction strategy is selected; based on the dose correlation degree, geometric overlap and dose influence range, the communication graph structure is constructed through the selected communication graph construction strategy.

3. The swarm intelligence driven radiotherapy treatment planning optimization system of claim 2, wherein, The communication graph structure includes: A communication graph construction strategy is selected; the communication graph construction strategy includes a Voronoi graph space division strategy and a K-nearest neighbor sparse graph strategy; Each Agent is selected as a graph node to construct an initial communication graph structure; Based on the dose correlation degree, geometric overlap and dose influence range, the initial communication graph structure is optimized through the selected communication graph construction strategy to determine the edges and their edge weights, and the communication graph structure is obtained.

4. The swarm intelligence driven radiotherapy treatment planning optimization system of claim 3, wherein, When the Voronoi graph space division strategy is selected, the optimization of the initial communication graph structure includes: The geometric center corresponding to each graph node is taken as a seed point, and Voronoi partitioning is performed in three-dimensional space to obtain a Voronoi unit dedicated to each Agent; if the structures corresponding to two Agents have geometric overlap or strong dose correlation, and the corresponding Voronoi units are not directly adjacent, an edge corresponding to the two Agents is established; based on the minimum boundary distance of each graph node and the dose correlation degree, the edge weights between each two different graph nodes are determined through weighted calculation; the edge weights are applied to the corresponding edges to complete the optimization of the initial communication graph structure; When the K-nearest neighbor (KNN) sparse graph strategy is selected, the optimization of the initial communication graph structure includes: Based on each geometric center, calculate the Euclidean distance between each pair of different Agents to generate a distance matrix; for each Agent, initially select the K Agents with the smallest distances to establish edges; if the... If an Agent has high geometric overlap or strong dose correlation with other structural data, its K value is increased to establish edges with more relevant Agents; based on the minimum boundary distance and dose correlation of each graph node, the edge weight between each pair of different graph nodes is determined by weighted calculation; the edge weight is applied to the corresponding edge to complete the optimization of the initial communication graph structure.

5. The swarm intelligence driven radiotherapy treatment planning optimization system of claim 1, wherein, The adjustment of the dose parameters of each Agent and the corresponding optimization objective function includes: Each Agent corresponding to the structure region under the current iteration is optimized through a local search algorithm, and the dose budget of each Agent is optimized to obtain the updated dose budget under the current iteration; When the satisfaction index of the OAR of a neighboring Agent of a certain Agent is higher than the OAR satisfaction index threshold and the negotiation willingness value is less than the OAR negotiation willingness threshold, the formula is used to update the edge dose of the Agent: ； reducing the current iteration number in the overlap region of the agent the PTV dose gain value under the current iteration number ; wherein, represents the negotiation willingness value corresponding to the agent under the current iteration number the first agent, represents the adjustment rate;​ If a certain Agent receives multiple dose negotiation requests from neighboring Agents, the formula is used to update the edge dose of the Agent: ； The optimization objective function of the Agent in the current iteration is updated to obtain the updated optimization objective function. ;in, This represents the optimization objective function before the Agent update. Indicates the penalty coefficient. Indicates the current iteration number Next The satisfaction index for each agent. , These represent the current iteration number, respectively. Next A set of Agents and their neighboring Agents The first in The set of conflict point locations corresponding to each Agent Represents the intersection-union ratio function. This represents the summation function; Determine whether the satisfaction index of each Agent is higher than the satisfaction index threshold; if yes, reduce the weight of the Agent whose satisfaction index is higher than the satisfaction index threshold.

6. The swarm intelligence driven radiotherapy plan optimization system of claim 1, wherein, The corresponding dose conflict tensor is calculated, including: Based on the optimized state variables of each Agent and the received neighboring state vectors, i.e., using the corresponding conflict point position set in the optimized state variables of each Agent and the received neighboring state vectors, determine the voxel region of spatial overlap or dose conflict; Calculate the number of voxels in the voxel region of spatial overlap or dose conflict; determine whether each voxel number exceeds the voxel minimum threshold; if yes, determine that the two Agents corresponding to the voxel region have significant conflict, i.e., there is a conflict region and enter the coordination list; Calculate the dose conflict tensor of each voxel in the voxel region of the Agent with significant conflict between each other.

7. The swarm intelligence driven radiotherapy treatment planning optimization system of claim 1, wherein, The self-organizing coordination optimization algorithm is used to update the edge dose of the Agent with conflict region between each other, including: Information exchange is performed for the Agent with conflict region between each other, i.e., two Agents exchange corresponding shared information; the shared information includes the local dose gradient of the current iteration, the target dose requirement, the optimization state, and the historical dose sacrifice record; Based on the dose conflict tensor of each voxel in the conflict region, update the dose sacrifice weight; Based on each dose sacrifice weight, each local dose gradient, and the updated dose budget of the conflict fusion mechanism in the last iteration, adjust the edge dose distribution of the corresponding conflict region to obtain the updated edge dose.

8. The swarm intelligence driven radiotherapy plan optimization system of claim 1, wherein, The conflict fusion mechanism is used to update the dose budget of each Agent, including: Determine whether the dose budgets of the two Agents corresponding to the conflict region have an intersection; if yes, determine whether the current dose in each conflict region belongs to the intersection of the dose budgets of the two Agents corresponding to the conflict region; otherwise, use the target offset mechanism to adjust the tolerance boundary of the dose budgets of the two Agents corresponding to the conflict region, expand or compress the budget within the allowed range, and complete the update of the dose budget of each Agent; If the current dose in each conflict region belongs to the intersection of the dose budgets of the two Agents corresponding to the conflict region, calculate the budget allocation factor based on the historical sacrifice record, the target importance, and the remaining budget; Based on the budget allocation factor, update the current dose in each conflict region to obtain the updated dose. If the current dose in each conflict region belongs to the intersection of the dose budgets of the two Agents corresponding to the conflict region, it is determined whether the updated dose satisfies the corresponding optimization sub-objective; if yes, the dose budget of the two Agents corresponding to the conflict region is reduced based on the edge dose distribution of the last iteration; otherwise, the dose budget of the two Agents corresponding to the conflict region is increased based on the edge dose distribution of the last iteration; and the update of the dose budget of each Agent is completed.

Images