Dose prediction method, apparatus, and electronic device

Abstract

The present disclosure provides a dose prediction method, comprising: generating an initial vector based on initial prescription information and patient structure data; generating an initialization parameter based on the initial vector; generating an initial dose distribution according to the initialization parameter; and adjusting the initial dose distribution based on a plurality of agents to generate a target dose distribution, wherein a single agent is responsible for at least partial adjustment in the initial dose distribution. The present disclosure also provides an apparatus and an electronic device for performing the method.

Description

Technical Field

[0001] This disclosure relates to the field of medical radiation, and more specifically, to a dose prediction method, apparatus, and electronic device. Background Technology

[0002] In current cancer radiotherapy planning, dose optimization methods primarily rely on rule-driven planning systems and manual adjustments based on expert experience. These traditional methods face significant drawbacks. First, they are labor-intensive, requiring clinicians to spend considerable time and effort manually adjusting dose distribution to ensure a balance between target coverage and organ protection at risk. This manual operation is not only inefficient but also prone to fluctuations in outcome quality, limited by individual experience levels.

[0003] In recent years, the application of deep learning technology in dose prediction has gradually increased. However, these methods are mostly limited to supervised learning frameworks, unable to dynamically adjust dose distribution, and lack confidence assessment of the output results. Furthermore, while existing reinforcement learning techniques are beginning to be introduced into radiotherapy planning optimization, most methods employ a single-agent architecture, making it difficult to solve the collaborative optimization problem of dose distribution for multiple structures, conflicts, and targets. When dealing with complex anatomical structures, single agents often fail to fully consider the individual needs of different regions, leading to less than ideal optimization results, especially when multiple target organs need to be considered simultaneously. Summary of the Invention

[0004] In view of this, the present disclosure provides a dose prediction method, apparatus and electronic device.

[0005] One aspect of this disclosure provides a dose prediction method, comprising: generating an initial vector based on initial prescription information and patient structural data; generating initialization parameters based on the initial vector; generating an initial dose distribution based on the initialization parameters; and adjusting the initial dose distribution based on multiple agents to generate a target dose distribution, wherein a single agent is responsible for at least a portion of the adjustment of the initial dose distribution.

[0006] According to embodiments of this disclosure, the agent is a reinforcement learning-based agent, adjusting the initial dose distribution includes: performing the following operations until a stopping condition is met: acquiring a dose distribution input, wherein the dose distribution input is the dose distribution result generated by the previous adjustment or the initial dose distribution; adjusting the dose distribution input based on multiple agents to generate an intermediate dose distribution result corresponding to the adjustment; and in response to meeting the stopping condition, using the intermediate dose distribution result corresponding to the last adjustment as the target dose distribution.

[0007] According to embodiments of this disclosure, the stopping condition includes at least one of the following: in N consecutive adjustments, the reward increase of each agent in reinforcement learning is less than a preset reward threshold, where N is a positive integer greater than or equal to 5; the number of adjustments is greater than or equal to a preset number threshold; and the intermediate dose distribution result obtained after adjustment meets a preset clinical constraint.

[0008] According to an embodiment of this disclosure, the process of generating an intermediate dose distribution result in one step includes: obtaining a dose distribution input; at least one first agent among a plurality of agents adjusting the dose distribution input to obtain a first dose distribution; at least one second agent among a plurality of agents adjusting the dose distribution input based on the first dose distribution to obtain a second dose distribution; and obtaining an intermediate dose distribution result generated by the adjustment based on the first dose distribution and the second dose distribution; wherein the first agent and the second agent are different.

[0009] According to embodiments of this disclosure, the process of generating intermediate dose distribution results in one step includes:

[0010] Each agent generates its own corresponding third dose distribution based on the dose distribution input. The third dose distribution represents the adjustment result of the agent to the dose distribution input. Based on the degree of influence of each third dose distribution on the patient's structural data, each third dose distribution is adjusted. Based on the adjusted third dose distributions, an intermediate dose distribution result is generated. The degree of influence at least represents the dose level of the third dose to the target organ in the patient's structural data. The target organs targeted by different agents are at least partially non-overlapping.

[0011] According to embodiments of this disclosure, the intermediate dose distribution result includes a target distribution value fluctuation range. The process of generating the intermediate dose distribution result once includes: each agent generating a fourth dose distribution multiple times based on the dose distribution input; generating a target distribution value fluctuation range based on each fourth dose distribution; wherein, the process of generating the fourth dose distribution once includes one of the following operations: randomly obtaining the action value corresponding to reinforcement learning from a preset first probability distribution, and generating the fourth dose distribution based on the action value; randomly obtaining a preset weight distribution from a preset second probability distribution, adjusting the weights of each network in the agent according to the preset weight distribution, and generating the fourth dose distribution based on the adjusted agent; randomly selecting a target neuron connection in the agent, adjusting the agent according to the target neuron connection, and generating the fourth dose distribution based on the adjusted agent, wherein the target neuron connection does not participate in the generation process of the fourth dose distribution.

[0012] According to embodiments of this disclosure, the method further includes: generating feature information based on patient structural data, the feature information at least characterizing the spatial relationships between various target regions in the patient, and / or the medical characteristics of each target region, the target region including at least one of the patient's organs at risk, the target area targeted by the dose distribution, the dose gradient transition area, and the patient's normal tissue areas; determining the target computational region targeted by each agent based on the feature information, the target computational region containing at least one target organ; adjusting the initial dose distribution, including: each agent adjusting the dose distribution of the target computational region targeted by each agent in the initial dose distribution to generate a target dose distribution.

[0013] According to embodiments of this disclosure, structural features are generated based on patient structural data, including: generating feature information based on the spatial overlap between the target area and the organs at risk, and / or the distribution location of the organs at risk, and / or the number of organs at risk, and / or the difficulty of predicting the dose distribution of each target area.

[0014] Another aspect of this disclosure provides a dose prediction device, comprising: a first generation module for generating an initial vector based on initial prescription information and patient structural data; a second generation module for generating initialization parameters based on the initial vector; a third generation module for generating an initial dose distribution according to the initialization parameters; and a fourth generation module for adjusting the initial dose distribution based on multiple agents to generate a target dose distribution, wherein a single agent is responsible for at least a portion of the adjustment of the initial dose distribution.

[0015] Another aspect of this disclosure provides an electronic device including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the dose prediction method of any of the foregoing embodiments.

[0016] Another aspect of this disclosure provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform a dose prediction method according to any of the foregoing embodiments.

[0017] Another aspect of this disclosure provides a computer program product, including a computer program / instructions, characterized in that the computer program / instructions, when executed by a processor, implement the operation of the dose prediction method of any of the foregoing embodiments.

[0018] According to embodiments of this disclosure, the dose prediction scheme provided has at least one of the following beneficial effects: First, by introducing a multi-agent architecture, the system can adaptively divide tasks based on the anatomical characteristics of different patients, ensuring that each agent focuses on a specific target computational region. This approach significantly improves the fine-grained control of dose distribution, especially in handling complex structures and multi-target conflicts, enabling more efficient dose optimization. This disclosure employs a feature-information-driven agent partitioning and adjustment mechanism, giving the system good medical interpretability and engineering operability. Feature information includes spatial overlap, the location and number of organs at risk, and the difficulty of dose distribution prediction, helping the system to have greater flexibility and accuracy when handling high-risk areas. Furthermore, by modeling uncertainty and outputting the confidence interval of the dose map, the credibility of clinical decisions is improved. Doctors can identify high-risk areas and manually intervene based on uncertainty cues, reducing clinical risks and enhancing the acceptance of AI systems in clinical applications. Attached Figure Description

[0019] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments of this disclosure with reference to the accompanying drawings, in which:

[0020] Figure 1 A flowchart illustrating a dose prediction method according to an embodiment of the present disclosure is shown schematically;

[0021] Figure 2 A flowchart illustrating the adjustment of the initial dose distribution in a dose prediction method according to an embodiment of the present disclosure is shown schematically.

[0022] Figure 3 A flowchart illustrating the generation of intermediate dose distribution results in a dose prediction method according to an embodiment of the present disclosure is shown in the schematic diagram.

[0023] Figure 4 This schematically illustrates another flowchart of the dose prediction method according to an embodiment of the present disclosure, which generates intermediate dose distribution results in one step.

[0024] Figure 5 This schematically illustrates another flowchart of the dose prediction method according to an embodiment of the present disclosure, which generates intermediate dose distribution results in one step.

[0025] Figure 6 Another flowchart of a dose prediction method according to an embodiment of the present disclosure is illustrated schematically;

[0026] Figure 7 A block diagram of a dose prediction device according to an embodiment of the present disclosure is schematically shown; and

[0027] Figure 8A block diagram of an electronic device suitable for implementing the methods described above, according to embodiments of the present disclosure, is illustrated schematically. Detailed Implementation

[0028] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0031] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0032] In the embodiments disclosed herein, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.

[0033] Embodiments of this disclosure provide a dose prediction method, comprising: generating an initial vector based on initial prescription information and patient structural data; generating initialization parameters based on the initial vector; generating an initial dose distribution based on the initialization parameters; and adjusting the initial dose distribution based on multiple agents to generate a target dose distribution, wherein a single agent is responsible for at least a portion of the adjustment of the initial dose distribution.

[0034] Figure 1 A flowchart illustrating a dose prediction method according to an embodiment of the present disclosure is shown schematically.

[0035] like Figure 1 As shown, the dose prediction method may include at least operations S110 to S140.

[0036] In operation S110, an initial vector is generated based on the initial prescription information and patient structure data.

[0037] Initial prescribing information can be the constraints of a radiotherapy plan formulated by clinicians based on the patient's condition and treatment goals. It typically includes minimum target volume dose coverage requirements, maximum dose limits for organs at risk, and requirements for treatment fractions and total dose. For example, initial prescribing information may include a target volume D95 (dose received at 95% of the target volume) of 70 Gy, and a maximum dose to organs at risk such as the spinal cord not exceeding 45 Gy.

[0038] Patient structural data can be anatomical information about a patient obtained through medical imaging techniques. It typically includes computed tomography (CT) images, outlines of the target area and organs at risk, voxel masks, and related spatial geometry. For example, patient structural data may include CT images of a lung cancer patient, where the target area is the tumor region, the outlines of organs at risk including the lungs, esophagus, and heart, and voxel representations of these structures.

[0039] An initial vector is a unified mathematical representation of the input state for reinforcement learning environments, formed by feature extraction and encoding of initial prescription information and patient structural data. Initial vectors typically include the spatial coordinates of the target area and organs at risk, overlap, dose distribution prediction difficulty, and historical dose-volume histogram (DVH) features. For example, an initial vector can be a high-dimensional feature vector containing the three-dimensional coordinates of the target center point, the spatial overlap between the target area and the heart (expressed as a percentage of voxel overlap), and DVH predictions based on historical cases.

[0040] In operation S120, initialization parameters are generated based on the initial vector.

[0041] Initialization parameters, which can be a set of parameters used to initialize the reinforcement learning system, can be generated based on an initialization vector by referencing a historical case database or a pre-trained model. These parameters include the generative model weights for the initial dose distribution, the initial weights of the reinforcement learning policy network, or control parameters of the heuristic algorithm. For example, initialization parameters may include a dose distribution template similar to the current patient's structure extracted from a historical lung cancer case database, or reference values ​​based on the weights of a pre-trained convolutional neural network model used to generate the initial dose distribution.

[0042] In operation S130, an initial dose distribution is generated based on the initialization parameters.

[0043] The initial dose distribution, which can be a preliminary dose distribution map generated based on initialization parameters using a pre-trained model, heuristic algorithm, or historical case template, serves as the starting point for multi-agent reinforcement learning optimization. The initial dose distribution is typically represented in the form of a three-dimensional voxel grid, with each voxel corresponding to a dose value, initially meeting the prescription requirements but not yet reaching optimality. For example, the initial dose distribution could be a three-dimensional dose map where the dose values ​​of the target area voxels initially meet the requirement of D95=70Gy, but the dose distribution for organs at risk, such as the lungs, still needs further optimization to reduce side effects.

[0044] In operation S140, based on multiple agents, the initial dose distribution is adjusted to generate the target dose distribution, wherein each agent is responsible for at least a portion of the adjustment of the initial dose distribution.

[0045] An agent is an independent decision-making unit built based on reinforcement learning algorithms. Each agent learns and optimizes strategies through interaction with the environment, and is responsible for adjusting the dose values ​​of specific regions or targets in the initial dose distribution. Agents typically employ deep reinforcement learning models, such as Dual-Action Deep Q-Networks (D3QN), to achieve efficient action selection and policy optimization.

[0046] Adjusting the initial dose distribution refers to the process by which multiple agents dynamically adjust the dose values ​​in a voxel grid based on the current dose distribution and environmental feedback through an iterative reinforcement learning process, in order to meet clinical prescription requirements and optimize the quality of the dose distribution. The adjustment process involves agents selecting actions (such as increasing or decreasing the dose value of a voxel) according to their respective target regions (such as the target area or specific organs at risk), evaluating the adjustment effect through a reward function, and ultimately forming a globally optimized target dose distribution.

[0047] This disclosure employs a collaborative optimization mechanism using multiple agents based on reinforcement learning, overcoming the limitations of traditional manual parameter tuning methods that suffer from "local modification, global uncontrollability." Traditional manual planning is constrained by the human decision-making ability's limited perception of multi-variable linkage effects, easily falling into a mindset of "the current dose being acceptable is the endpoint," thus neglecting potential optimization space in the future. In contrast, reinforcement learning agents do not aim solely at immediate improvement, but rather adjust strategies based on long-term cumulative rewards, ensuring that each dose modification is embedded in global optimization judgments for future sequences.

[0048] Furthermore, the multi-agent architecture allows each agent to focus on dose adjustment for specific structures or regions, simulating the collaborative work of multiple experts. This strengthens the modeling of coordination between conflicting structures, effectively avoiding the problem of single agents getting trapped in local optima, and improving the global consistency and robustness of dose distribution across spatial and structural dimensions. Therefore, the dose prediction method provided in this disclosure, while ensuring clinical dose constraints, can generate higher-quality, more coordinated, and more stable target dose maps, possessing stronger individual adaptability and clinical applicability.

[0049] Figure 2 A flowchart illustrating the adjustment of the initial dose distribution in a dose prediction method according to an embodiment of the present disclosure is shown.

[0050] According to embodiments of this disclosure, the intelligent agent is a reinforcement learning-based intelligent agent, such as... Figure 2 As shown, based on the aforementioned embodiments, S140 may include repeatedly executing operations S210 to S220 until the stop condition is met, and operation S230.

[0051] In operation S210, obtain the dose distribution input, which is the dose distribution result generated by the previous adjustment or the initial dose distribution.

[0052] The dose distribution input can be considered as the dose map data representing the current state in a reinforcement learning loop. The dose distribution input can be the initial dose distribution generated by the system initially, or an intermediate dose map generated by multiple agents in the previous adjustment operation. This distribution data is expressed as a three-dimensional voxel matrix, with each voxel containing a dose value, allowing each agent to select its next action based on the target structural region it is responsible for. For example, in the first round, the dose distribution input is a dose map generated based on initialization parameters, while in the second round it is an intermediate dose map obtained after the agents' collaborative adjustment in the first round.

[0053] In operation S220, the dose distribution input is adjusted based on multiple agents to generate the intermediate dose distribution result corresponding to the adjustment.

[0054] Each reinforcement learning agent fine-tunes the dose values ​​of the current dose map locally or across different regions based on its assigned computational region (such as a specific target area or organ at risk). Each agent makes an action selection based on its internal reinforcement learning model (such as D3QN) for the current state, such as increasing, decreasing, or maintaining the dose in a certain region. The outputs of multiple agents can be fused sequentially or in parallel to synthesize a new intermediate dose distribution. The fusion process considers the weighting of each region's impact on the clinical goal; for example, regions with higher target importance are given greater weight during fusion. This intermediate dose distribution is used to subsequently determine whether the stopping condition has been met.

[0055] In operation S230, in response to the satisfaction of the stop condition, the intermediate dose distribution result corresponding to the last adjustment is used as the target dose distribution.

[0056] The system determines whether the current optimization is complete based on the stopping criteria set during the reinforcement learning optimization process. If the stopping criteria are met, the current intermediate dose map is used as the final output target dose map for clinical use or for subsequent validation.

[0057] According to embodiments of this disclosure, the stopping condition includes at least one of the following: in N consecutive adjustments, the reward increase of each agent in reinforcement learning is less than a preset reward threshold, where N is a positive integer greater than or equal to 5; the number of adjustments is greater than or equal to a preset number threshold; and the intermediate dose distribution result obtained after adjustment meets a preset clinical constraint.

[0058] Specifically, the stopping condition can be: if the reward value increase obtained by each agent is less than a certain set threshold (e.g., 0.01) in N consecutive rounds (e.g., N=10), it indicates that the system has entered the learning saturation state, and at this time it is determined to be converged.

[0059] Specifically, the stopping condition can be: if the number of iterations reaches the maximum round limit (e.g., 100 rounds), to prevent infinite loops or model overfitting, the system can also force the output of the current dose map as the result.

[0060] Specifically, the stopping condition can be: if all dose-volume histogram (DVH) indicators of the current intermediate dose map meet the prescription limits (e.g., the target area D95 reaches 70 Gy, and the organ at risk Dmax does not exceed the limit value), it can also be regarded as reaching the optimization endpoint.

[0061] Through the aforementioned optimization mechanism, this disclosure achieves a closed-loop design of the entire process of "state-action-reward-update" in the reinforcement learning system, ensuring the traceability and progressiveness of the optimization results. Compared with the traditional one-time adjustment method based on heuristic rules, this scheme introduces multiple rounds of dynamic updates, enabling each dose modification action to generate a coherent response based on historical context, supporting the exploration of long-term optimal solutions.

[0062] Figure 3 A flowchart illustrating the generation of intermediate dose distribution results in a dose prediction method according to an embodiment of the present disclosure is shown.

[0063] like Figure 3 As shown, based on the aforementioned embodiments, S220 may include operations S310 to S340.

[0064] In operation S310, obtain dose distribution input.

[0065] In operation S320, at least one of the multiple agents, the first agent, adjusts the dose distribution input to obtain a first dose distribution.

[0066] In this step, the first agent adjusts the dose value of the dose distribution input based on structural features within its control range (such as a target area or an organ at risk). This agent, based on a reinforcement learning policy model, selects actions such as increasing or decreasing the dose in a specific voxel region, thereby generating the adjusted first dose distribution. The adjustment process is based on the "environmental state" of the dose distribution input; the agent outputs corresponding actions according to the state value and the policy network, and updates the dose value of the controlled area. For example, if the first agent controls a portion of the tumor target area, it may preferentially increase the central dose when the target area D95 is insufficient, thus forming the first dose distribution.

[0067] In operation S330, at least one second agent among the multiple agents adjusts the dose distribution input based on the first dose distribution to obtain a second dose distribution.

[0068] In this step, the second agent receives the first dose distribution and the original dose distribution inputs, and determines its action strategy based on these two. Unlike the direct adjustment by the first agent, the second agent can compensate for or fine-tune its strategy according to the changing trend of the first dose distribution. For example, if the first agent increases the dose to the target area, the second agent may be responsible for regulating the dose to organs at risk that are spatially adjacent to the target area to prevent them from exceeding safety limits. This sequential mechanism ensures the sequential dependencies between agents and the causal consistency of the adjustment logic, which helps improve the overall stability of the system.

[0069] In operation S340, based on the first dose distribution and the second dose distribution, the intermediate dose distribution result generated in this adjustment is obtained. The first agent and the second agent are different.

[0070] The first and second agents differ in several ways. First, this constraint ensures that each agent in the system has independent responsibilities and structural control, preventing dose oscillations or strategy convergence failures caused by repeated adjustments by the same agent. Second, the differentiated agent roles allow the system to simulate clinical regulatory approaches such as "target-first, avoidance-later" and "primary-first, secondary-later" strategies, achieving sequential modeling of phased optimization goals. Through a serial control structure, the system can introduce a posterior compensation mechanism from the second agent based on the adjustments made by the first agent, strengthening local risk control capabilities and global dose coordination.

[0071] Through the aforementioned serial agent collaboration mechanism, this embodiment effectively avoids conflicts, overlaps, or mutual cancellations that may result from parallel operations of multiple agents, enhancing the causality and coordination among decision-making units within the system. The first and second agents have clearly defined responsibilities in their structure and sequentially connected actions, making the system behavior closer to human planning, while possessing a higher-dimensional ability for coordinated adjustment at the strategy level.

[0072] Figure 4 Another flowchart illustrating the generation of intermediate dose distribution results in a dose prediction method according to an embodiment of the present disclosure is shown.

[0073] like Figure 4 As shown, based on the aforementioned embodiments, S220 may include operations S410 to S430.

[0074] In operation S410, each agent generates its own corresponding third dose distribution based on the dose distribution input. The third dose distribution represents the adjustment result of the agent in response to the dose distribution input.

[0075] In this step, multiple reinforcement learning agents receive the same dose distribution input in parallel and independently perform policy reasoning and dose adjustment within their respective target computation regions. Each agent modifies the dose field of its control structure based on its internal policy network and generates a "third dose distribution" in the complete dose map. This distribution reflects the optimal dose adjustment behavior that agent should take in the current state. For example, the agent responsible for regulating the lung region will try to reduce high-dose areas and limit the volume exceeding the limit within its region; while the agent responsible for the target region may focus on increasing dose coverage.

[0076] In conjunction with the foregoing embodiments, the third dose distribution generated by at least some of the second agents among the plurality of agents is generated based on the third dose distribution generated by at least some of the first agents among the plurality of agents. For details, please refer to the foregoing embodiments, which will not be repeated here.

[0077] In operation S420, each third dose distribution is adjusted based on the degree of influence of each third dose distribution on the patient's structural data.

[0078] The degree of influence at least characterizes the dosage of the third dose to the target organ in the patient's structural data. Specifically, the degree of influence reflects the sensitivity, contribution, and optimization importance of the third dose distribution generated by the agent to the dose response of key anatomical structures (such as the target area or organs at risk). The degree of influence can be calculated based on the following dimensions:

[0079] The rate of change of this dose distribution with respect to key DVH parameters such as mean dose, maximum dose, or D95 in the target organ. And / or,

[0080] The degree of improvement in the dose gradient within the region, such as reducing dose spillover or increasing edge dose concentration. And / or,

[0081] The degree of dose interference to adjacent structures, i.e. whether it causes unwanted collateral damage.

[0082] For example, if a third dose distribution generated by an agent causes the maximum dose in the cardiac region it is responsible for to decrease from 46 Gy to 44 Gy, while the clinical limit is 45 Gy, then the agent's adjustment behavior has a high "influence" weight.

[0083] In this context, different agents target non-overlapping organs. This "non-overlapping" does not require that the structures involved by each agent are completely independent, but rather that their task objectives have at least some structural differences. Even if multiple agents have overlapping control over the boundary regions of the same organ, they can be considered not completely overlapping as long as their focus areas, target computation sub-regions, dose adjustment strategies, or reward objectives differ significantly. For example, one agent might be responsible for dose protection of the upper lobes of the lung, while another agent focuses on volume control of the lower lobes. Although both belong to the "lung" structure, their target areas and dose strategies differ.

[0084] In operation S430, intermediate dose distribution results are generated based on the adjusted third dose distributions.

[0085] The system uses the third dose distributions generated by each agent after adjustment as candidate results for fusion processing to generate a unified intermediate dose distribution map. The fusion process adopts a structure-guided weighting approach, which does not simply average or crop the dose maps, but assigns fusion weights based on the actual impact of each dose map on the patient's key structures.

[0086] Specifically, for each voxel, the system first determines whether the voxel is within a target structural region controlled by a certain agent, such as whether it is located in the target area, high-risk organ, or transition zone. If so, the optimization contribution of the agent to the region is given priority, that is, whether the dose adjustment produced by the agent effectively improves target coverage or reduces the risk of overdose to organs at risk.

[0087] Based on this, the system dynamically assigns fusion priority weights to different agents in different regions. For example, if an agent successfully reduces the dose from 47 Gy to 44 Gy in the heart region (while the maximum dose requirement for the heart is 45 Gy), its output in the heart region will be given higher weight. If another agent increases the D95 value in the target area edge region, its output dose in that region will be prioritized.

[0088] Through the weighted strategy guided by the above structure, the system achieves local optimal output integration across the entire structure range, so that the final generated intermediate dose map not only has numerical continuity and coordination, but also demonstrates optimal control and protection effects on key structures.

[0089] Figure 5 Another flowchart illustrating the generation of intermediate dose distribution results in a dose prediction method according to an embodiment of the present disclosure is shown.

[0090] According to embodiments of this disclosure, intermediate dose distribution results may include the target distribution value fluctuation range, such as... Figure 5 As shown, based on the aforementioned embodiments, S220 may include operations S510 to S520.

[0091] During operation of S510, each agent generates a fourth dose distribution multiple times based on the dose distribution input.

[0092] In this step, to estimate the uncertainty in the dose prediction process, each reinforcement learning agent, given the same dose distribution input, uses different perturbation mechanisms to repeatedly perturb its internal decision-making process. Each perturbation generates a complete dose map, i.e., the "fourth dose distribution." These distributions constitute a sample set of uncertainties in dose prediction, used for subsequent statistical inference of the target dose's fluctuation range. For example, an agent targeting lung structures can generate 20 dose maps based on the same input state, with each image changing due to internal policy perturbations, thus characterizing the output variability.

[0093] The process of generating the fourth dose distribution in one step includes one of the following operations:

[0094] The agent randomly selects action values ​​corresponding to reinforcement learning from a pre-defined first probability distribution, and generates a fourth dose distribution based on these action values. Specifically, given the current state input, the agent no longer adopts greedy actions or maximum Q-value policies, but instead samples actions from a pre-defined policy distribution (such as a Softmax policy or a Boltzmann distribution). The sampled actions are used to adjust the dose values ​​(such as increasing, decreasing, or maintaining), thereby outputting a dose map. Through multiple random samplings, the volatility of the policy space can be covered, reflecting the output differences caused by policy uncertainty.

[0095] A preset weight distribution is randomly obtained from a preset second probability distribution. The weights of each network in the agent are adjusted according to this preset weight distribution, and a fourth dose distribution is generated based on the adjusted agent. Specifically, this simulates the sources of uncertainty at the model structure level. The specific approach is to sample network parameters (such as the weight matrix in a neural network) from a Bayesian prior distribution or a known Gaussian perturbation, keeping the input constant, and apply these parameters to the policy network. A new dose map is then generated based on the perturbated network. This method can be used to evaluate the impact of changes in model structure on the output dose map, and falls under the category of structural uncertainty estimation.

[0096] A target neuron connection is randomly selected within the agent. Based on this target neuron connection, the agent is adjusted, and a fourth dose distribution is generated. The target neuron connection does not participate in the generation of the fourth dose distribution. Specifically, the uncertainty of neural network activation paths is modeled by randomly selecting some neuron connections (such as connections in intermediate hidden layers or output layers) and temporarily masking the activation paths of these connections. This allows the agent to use different sub-network structures for reasoning in this decision-making process. Each perturbation outputs a new dose map, thus constructing a set of dose distributions for the uncertainty of activation paths.

[0097] In operation S520, a target distribution value fluctuation range is generated based on each fourth dose distribution.

[0098] The system treats multiple fourth dose distributions generated under the same input state as a dose value sample set and statistically analyzes their variation range at each voxel location. Specifically, the system calculates the mean and variance for all sample maps at the same voxel location, thereby forming the expected dose value and its confidence interval. For example, if the dose value distribution of a certain voxel in 20 dose maps has a mean of 70 Gy and a standard deviation of 1.5 Gy, the system can calculate that the dose fluctuation range of that voxel at a 95% confidence level is 67 Gy to 73 Gy, and use this as part of the target dose fluctuation range map. The final output includes: a central predicted dose map + a dose confidence interval map for each voxel, which helps clinicians identify areas with high uncertainty, thereby enabling manual intervention or risk reassessment.

[0099] In traditional systems, due to the limited number of training data samples and the high individual variability in tumor types and structural layouts, especially for rare structural layouts (such as tumors close to the heart), models often struggle to output sufficiently representative and stable prediction results. If the system only outputs a single "dose map prediction value," clinicians will be unable to determine its reliability, thus posing a risk of "blindly trusting AI." This disclosed embodiment introduces a mechanism of generating a "fourth dose distribution" through multiple perturbations, introducing uncertainty into the dose prediction process and statistically inferring the mean and fluctuation range of the dose for each voxel. Ultimately, the system output includes a central dose prediction plus a replica floating map, clearly indicating which regions have stable and reliable predictions and which regions have large fluctuations or high uncertainty. Doctors can identify potentially risky locations in the system's predictions by viewing high-uncertainty areas. In cases of complex structures such as tumors near high-risk organs, uncertainty bandwidth indicators can guide the treatment team to adjust segmentation strategies or adopt more conservative approaches, improving the credibility, transparency, and clinical acceptability of the AI ​​radiotherapy system.

[0100] Figure 6 Another flowchart of a dose prediction method according to an embodiment of the present disclosure is illustrated schematically.

[0101] like Figure 6 As shown, based on the aforementioned embodiments, the dose prediction method may include operations S610 to S620.

[0102] In operation S610, feature information is generated based on patient structural data. The feature information at least characterizes the spatial relationship between various target regions in the patient and / or the medical characteristics of each target region. The target regions include at least one of the patient's various organs at risk, the target area targeted by the dose distribution, the dose gradient transition area, and the patient's normal tissue areas.

[0103] In this step, the system extracts information from the patient's CT images, MRI data, or other anatomical images to describe the spatial relationships and medical characteristics between various target regions. This information includes not only the geometric shape of the target region (such as volume, boundary curvature, and spatial location) but also its functional attributes in treatment (such as tissue type, sensitivity, and known dose limits).

[0104] The system identifies vulnerable organ regions in patients, such as the heart, spinal cord, and esophagus, which require special attention during radiotherapy. By recognizing the spatial location, proximity, and historical dose limitations of these regions, the system determines their priority and influencing factors within the feature information.

[0105] The dose distribution targets the primary tumor treatment area, typically including the Gross Tumor Volume (GTV), Clinical Target Volume (CTV), and Planning Target Volume (PTV). The system performs detailed delineation of the target area and uses feature information to annotate its spatial relationship with surrounding Organisms of Interest (OARs) to optimize dose coverage.

[0106] The dose gradient transition zone refers to the dose buffer band between the target area and normal tissue, typically a 1-2 cm wide transition area. The system identifies the rate of dose change in this region using characteristic information to ensure that the dose gradient is appropriately controlled in the treatment plan.

[0107] The patient's normal tissue areas can include background areas such as lung parenchyma, soft tissue, and bone. The system uses feature information to describe the extent of their impact throughout the dose distribution to avoid over-irradiation.

[0108] In operation S620, based on feature information, the target computing region targeted by each agent is determined, and the target computing region contains at least one target organ.

[0109] Specifically, for example, based on the spatial overlap of target regions in the patient, the target computation region targeted by each agent is determined. By calculating the voxel overlap rate, contact area, and relative positional relationship between different target regions, the complexity of the region is judged, and the responsibilities of the agents are assigned accordingly. For example, if the target area has a high degree of overlap with the heart, the system will prioritize assigning a dedicated agent for fine-tuning the dosage.

[0110] Specifically, for example, based on the distribution location of each endangered organ and / or the number of endangered organs, the target computing area targeted by each of the aforementioned agents is determined. By identifying the spatial density and number of OARs, the system automatically adjusts the distribution strategy of the agents to ensure that high-risk areas are handled by agents with multi-target collaborative capabilities.

[0111] For example, based on the difficulty of dose distribution prediction in each target region, the target computation region targeted by each agent is determined. The system analyzes the prediction error and dose fluctuation rate of different regions based on historical data, and configures agent resources to cope with high-difficulty regions, ensuring the stability and accuracy of the optimization process.

[0112] S140 may include operation S630.

[0113] In operation S630, each agent adjusts the initial dose distribution, the dose distribution of the target computation region targeted by each agent, and generates the target dose distribution.

[0114] By introducing a feature-driven agent partitioning and adjustment mechanism, this disclosure realizes a technical shift from static task configuration to dynamic structure-aware partitioning, enabling the system to adaptively determine the area and strategy parameters that each agent should control based on the anatomical characteristics of each patient, thereby making more efficient use of computing resources and avoiding unnecessary agent redundancy.

[0115] Meanwhile, the use of feature information makes the segmentation process medically interpretable and engineering computable, facilitating doctors' understanding of the AI ​​segmentation logic and enhancing the clinical credibility of the subsequently generated dose maps. This is particularly true for cases with high structural heterogeneity and ambiguous treatment boundaries, such as areas with multiple coexisting organs in the head and neck, mediastinum, or pelvis, significantly improving optimization effectiveness and risk control capabilities.

[0116] According to embodiments of this disclosure, generating feature information based on patient structural data may include: generating feature information based on the spatial overlap between the target area and the organs at risk, the distribution location of the organs at risk, and / or the number of organs at risk, and / or the difficulty of predicting the dose distribution of each target area.

[0117] Specifically, for spatial overlap calculation, the system first extracts the three-dimensional structural information of the target area and organs at risk from the patient's CT or MRI images. By calculating the voxel overlap, contact area, and relative positional relationship between the target area and each organ at risk, the system generates spatial overlap indices. These indices are used to assess the risk and complexity of conflict during dose adjustment.

[0118] Specifically, the system identifies the location and number of organs at risk: using anatomical tags or image segmentation techniques, it identifies organs at risk within the patient and records their spatial coordinates. The system further calculates the number and density of organs at risk per unit volume to determine the density of high-risk areas, guiding the partitioning and resource allocation of intelligent agents.

[0119] Specifically, the dose distribution prediction difficulty assessment involves the system calculating the dose prediction difficulty for each target region based on historical data and model performance analysis. This includes evaluating the model's prediction error, dose variability, and gradient stability on similar structures. The difficulty assessment results are used to dynamically adjust the agent's learning strategy and optimize resource allocation to ensure dose control accuracy in high-difficulty regions.

[0120] Through the aforementioned feature information generation mechanism, this disclosure achieves a refined description and quantitative analysis of the patient's anatomical structure, enabling the intelligent agent segmentation and dosage optimization strategies to possess good medical interpretability and engineering operability. The system can not only dynamically adapt to individual differences among patients but also achieve precise dosage control in complex structural layouts, significantly improving the quality and safety of treatment planning.

[0121] Furthermore, the use of feature information enables the system to have greater flexibility and accuracy when dealing with high-risk areas. Especially when the target area and the organs at risk are highly overlapping or structurally dense, the system can optimize dose distribution and reduce clinical risks through refined feature analysis and agent configuration.

[0122] Figure 7 A block diagram of a dose prediction device according to an embodiment of the present disclosure is shown schematically.

[0123] like Figure 7 As shown, the dose prediction device 700 may include a first generation module 710, a second generation module 720, a third generation module 730, and a fourth generation module 740.

[0124] The first generation module 710 is used to generate an initial vector based on the initial prescription information and patient structure data. In some embodiments, the first generation module 710 can be used to perform operation S110 in the dose prediction method described above, which will not be elaborated here.

[0125] The second generation module 720 is used to generate initialization parameters based on the initial vector. In some embodiments, the second generation module 720 can be used to perform operation S120 in the dose prediction method described above, which will not be elaborated here.

[0126] The third generation module 720 is used to generate an initial dose distribution based on initialization parameters. In some embodiments, the third generation module 720 can be used to perform operation S130 in the dose prediction method described above, which will not be elaborated here.

[0127] The fourth generation module 720 is used to adjust the initial dose distribution based on multiple agents to generate a target dose distribution, wherein a single agent is responsible for at least a portion of the adjustment of the initial dose distribution. In some embodiments, the fourth generation module 720 can be used to perform operation S140 in the dose prediction method described above, which will not be elaborated here.

[0128] According to embodiments of this disclosure, the fourth generation module may include a first acquisition module, a first adjustment module, and a first determination module.

[0129] The first acquisition module is used to acquire the dose distribution input, which is the dose distribution result generated by the previous adjustment or the initial dose distribution. In some embodiments, the first acquisition module can be used to perform operation S210 in the dose prediction method described above, which will not be elaborated here.

[0130] The first adjustment module is used to adjust the dose distribution input based on multiple agents and generate the intermediate dose distribution result corresponding to the adjustment. In some embodiments, the first adjustment module can be used to perform operation S220 in the dose prediction method described above, which will not be elaborated here.

[0131] The first determining module is used to, in response to the satisfaction of the stopping condition, take the intermediate dose distribution result corresponding to the last adjustment as the target dose distribution. In some embodiments, the first determining module may be used to perform operation S230 in the dose prediction method described above, which will not be elaborated here.

[0132] According to embodiments of this disclosure, the first adjustment module may include a second acquisition module, a second adjustment module, a third adjustment module, and a first integration module.

[0133] The second acquisition module is used to acquire the dose distribution input. In some embodiments, the second acquisition module can be used to perform operation S310 in the dose prediction method described above, which will not be elaborated here.

[0134] The second adjustment module is used to adjust at least one of the multiple agents (first agent) in response to the dose distribution input to obtain a first dose distribution. In some embodiments, the second adjustment module can be used to perform operation S320 in the dose prediction method described above, which will not be elaborated here.

[0135] The third adjustment module is used by at least one second agent among multiple agents to adjust the dose distribution input based on the first dose distribution to obtain a second dose distribution. In some embodiments, the third adjustment module can be used to perform operation S330 in the dose prediction method described above, which will not be elaborated here.

[0136] The first integration module is used to obtain the intermediate dose distribution result generated in this adjustment based on the first dose distribution and the second dose distribution. In some embodiments, the first integration module can be used to perform operation S340 in the dose prediction method described above, which will not be elaborated here.

[0137] The first intelligent agent and the second intelligent agent are different.

[0138] According to embodiments of this disclosure, the first adjustment module may include a fifth generation module, a fourth adjustment module, and a third integration module.

[0139] The fifth generation module is used by each agent to generate its own corresponding third dose distribution based on the dose distribution input. The third dose distribution represents the adjustment result of the agent in response to the dose distribution input. In some embodiments, the fifth generation module can be used to perform operation S410 in the dose prediction method described above, which will not be elaborated here.

[0140] The fourth adjustment module is used to adjust each third dose distribution based on the degree of influence of each third dose distribution on the patient's structural data. In some embodiments, the fourth adjustment module can be used to perform operation S420 in the dose prediction method described above, which will not be elaborated here.

[0141] The third integration module is used to generate intermediate dose distribution results based on the adjusted third dose distributions. In some embodiments, the third integration module can be used to perform operation S430 in the dose prediction method described above, which will not be elaborated here.

[0142] Among them, the degree of influence at least characterizes the dosage of the third dose to the target organ in the patient's structural data;

[0143] In this case, the target organs of different intelligent agents do not overlap.

[0144] According to embodiments of this disclosure, the intermediate dose distribution result includes the target distribution value fluctuation range, and the first adjustment module may include a sixth generation module, a seventh generation module, a preset 23 module, a preset 24 module, and a preset 25 module.

[0145] The sixth generation module is used by each agent to generate a fourth dose distribution multiple times based on the dose distribution input. In some embodiments, the sixth generation module can be used to perform operation S510 in the dose prediction method described above, which will not be elaborated here.

[0146] The seventh generation module is used to generate the target distribution value fluctuation range based on each fourth dose distribution. In some embodiments, the seventh generation module can be used to perform operation S520 in the dose prediction method described above, which will not be elaborated here.

[0147] The process of generating the fourth dose distribution in one step includes one of the following operations:

[0148] Randomly obtain the action values ​​corresponding to reinforcement learning from the preset first probability distribution, and generate a fourth dose distribution based on the action values;

[0149] A preset weight distribution is randomly obtained from the preset second probability distribution. The weights of each network in the agent are adjusted according to the preset weight distribution. A fourth dose distribution is generated based on the adjusted agent.

[0150] Target neuron connections are randomly selected in the agent. The agent is adjusted based on the target neuron connections, and a fourth dose distribution is generated based on the adjusted agent. The target neuron connections do not participate in the generation process of the fourth dose distribution.

[0151] According to embodiments of this disclosure, the dose prediction device may include an eighth generation module and a third determination module.

[0152] The eighth generation module is used to generate feature information based on patient structural data. The feature information at least characterizes the spatial relationship between various target regions in the patient and / or the medical characteristics of each target region. The target regions include at least one of the patient's regions of organs at risk, the target area targeted by the dose distribution, the dose gradient transition zone, and the patient's normal tissue regions. In some embodiments, the eighth generation module can be used to perform operation S610 in the dose prediction method described above, which will not be elaborated here.

[0153] The third determining module is used to determine the target computation region targeted by each agent based on the feature information, wherein the target computation region contains at least one target organ. In some embodiments, the third determining module can be used to perform operation S620 in the dose prediction method described above, which will not be elaborated here.

[0154] The fourth generation module further includes a ninth generation module, used by each agent to adjust the dose distribution of the target calculation region targeted by each agent in the initial dose distribution, and to generate a target dose distribution. In some embodiments, the ninth generation module can be used to perform operation S630 in the dose prediction method described above, which will not be elaborated here.

[0155] Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as hardware circuitry, such as a Field-Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System-on-Chip, a System-on-a-Substrate, a System-on-Package, an Application-Specific Integrated Circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0156] For example, any and more of the first generation module 710, the second generation module 720, the third generation module 730, the fourth generation module 750, and the preset 5 modules 750 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this disclosure, at least one of the first generation module 710, the second generation module 720, the third generation module 730, the fourth generation module 750, and the preset 5 modules 750 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in a suitable combination of any of them. Alternatively, at least one of the first generation module 710, the second generation module 720, the third generation module 730, the fourth generation module 750, and the preset 5 modules 750 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0157] It should be noted that the data processing system part in the embodiments of this disclosure corresponds to the data processing method part in the embodiments of this disclosure. The specific description of the data processing system part is referred to in the data processing method part, and will not be repeated here.

[0158] Figure 8 A block diagram of an electronic device suitable for implementing the methods described above, according to embodiments of the present disclosure, is illustrated schematically. Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0159] like Figure 8As shown, an electronic device 800 according to an embodiment of this disclosure includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this disclosure.

[0160] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 802 and / or RAM 803. It should be noted that the programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0161] According to embodiments of this disclosure, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0162] According to embodiments of this disclosure, the method flow according to embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by processor 801, it performs the functions defined in the system of embodiments of this disclosure. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0163] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0164] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0165] For example, according to embodiments of this disclosure, a computer-readable storage medium may include the ROM 802 and / or RAM 803 described above and / or one or more memories other than ROM 802 and RAM 803.

[0166] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this disclosure. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the control methods provided in the embodiments of this disclosure.

[0167] When the computer program is executed by the processor 801, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0168] In one embodiment, the computer program may rely on tangible storage media such as optical storage devices or magnetic storage devices. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via communication section 809, and / or installed from removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof. According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, 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).

[0169] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present disclosure can be combined and / or combined in various ways, even if such combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments of this disclosure may be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0170] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A dose prediction method, comprising: generating an initial vector based on initial prescription information and patient structure data; generating an initialization parameter based on the initial vector; the initialization parameter comprises a generated model weight of an initial dose distribution, an initial weight of a reinforcement learning policy network, or a control parameter of a heuristic algorithm; generating an initial dose distribution according to the initialization parameter; adjusting the initial dose distribution based on a plurality of agents to generate a target dose distribution, wherein a single agent is responsible for at least part of the adjustment of the initial dose distribution, and different agents are responsible for different regions of the patient; the agent is a reinforcement learning-based agent, and the adjusting the initial dose distribution comprises: performing the following operations until a stop condition is met: obtaining a dose distribution input, which is a dose distribution result generated by a previous adjustment, or the initial dose distribution; adjusting the dose distribution input based on a plurality of agents to generate an intermediate dose distribution result corresponding to this adjustment; in response to meeting the stop condition, taking the intermediate dose distribution result corresponding to the last adjustment as the target dose distribution; the generating the intermediate dose distribution result corresponding to this adjustment comprises: obtaining the dose distribution input; at least one first agent of the plurality of agents adjusts the dose distribution input to obtain a first dose distribution; at least one second agent of the plurality of agents adjusts the dose distribution input based on the first dose distribution to obtain a second dose distribution; obtaining the intermediate dose distribution result generated by this adjustment based on the first dose distribution and the second dose distribution; wherein the first agent and the second agent are different; the intermediate dose distribution result comprises an intermediate dose distribution and a target distribution value floating range, and the process of generating the intermediate dose distribution result once comprises: each agent generates a fourth dose distribution based on the dose distribution input multiple times; generating a target distribution value floating range based on each fourth dose distribution; wherein the process of generating a fourth dose distribution once comprises one of the following operations: randomly obtaining an action value corresponding to the reinforcement learning from a preset first probability distribution, and generating a fourth dose distribution according to the action value; randomly obtaining a preset weight distribution from a preset second probability distribution, adjusting the weights of each network in the agent according to the preset weight distribution, and generating a fourth dose distribution according to the adjusted agent; randomly selecting a target neuron connection in the agent, adjusting the agent according to the target neuron connection, and generating the fourth dose distribution according to the adjusted agent, wherein the target neuron connection does not participate in the generation process of the fourth dose distribution.

2. The method of claim 1, wherein, the stop condition comprises at least one of the following: in the last N adjustments, the reward of each agent in the reinforcement learning is less than a preset reward threshold, wherein N is a positive integer greater than or equal to 5; the number of adjustments is greater than or equal to a preset number threshold; the intermediate dose distribution result obtained after adjustment meets a preset clinical constraint.

3. The method of claim 1, wherein, A process of generating an intermediate dose distribution result, comprising: each of the agents generates a respective third dose distribution according to the dose distribution input, the third dose distribution representing an adjustment result of the agent for the dose distribution input; adjusting each of the third dose distributions based on an influence degree of each of the third dose distributions on the patient structure data; generating the intermediate dose distribution result based on the adjusted third dose distributions; wherein the influence degree at least represents a dose degree of the third dose on a target organ in the patient structure data; wherein the target organs for different agents at least partially do not overlap.

4. The method of claim 1, wherein, The method further comprises: generating feature information according to the patient structure data, the feature information at least representing a spatial relationship between each of target regions in the patient, and / or a medical feature of each of the target regions, the target regions including at least one of an organ-at-risk region of the patient, a target volume for the dose distribution, a dose gradient transition region, and a normal tissue region of the patient; determining a target calculation region for each of the agents according to the feature information, the target calculation region containing at least one target organ; the adjusting the initial dose distribution comprises: each of the agents adjusts a dose distribution of the target calculation region for the agent in the initial dose distribution to generate the target dose distribution.

5. The method of claim 4, wherein, The generating the feature information according to the patient structure data comprises: generating the feature information according to a spatial overlap degree between the target volume and the organ-at-risk for the dose distribution input, and / or, a distribution position of the organ-at-risk, and / or a number of the organ-at-risk, and / or, a dose distribution prediction difficulty of each of the target regions.

6. A dose prediction device, comprising: a first generation module configured to generate an initial vector based on initial prescription information and patient structure data; a second generation module configured to generate an initialization parameter based on the initial vector, the initialization parameter including a generation model weight of an initial dose distribution, an initial weight of a reinforcement learning strategy network, or a control parameter of a heuristic algorithm; a third generation module configured to generate an initial dose distribution according to the initialization parameter; and a fourth generation module configured to adjust the initial dose distribution based on a plurality of agents to generate a target dose distribution, wherein each of the agents is responsible for at least part of the adjustment of the initial dose distribution, and different agents are responsible for different regions of the patient. The agent is a reinforcement learning-based agent, and the adjusting the initial dose distribution comprises: performing the following operations until a stop condition is met: obtaining a dose distribution input, the dose distribution input being a dose distribution result generated by a previous adjustment or the initial dose distribution; adjusting the dose distribution input based on a plurality of agents to generate an intermediate dose distribution result corresponding to this adjustment; in response to the stop condition being met, taking the intermediate dose distribution result corresponding to the last adjustment as the target dose distribution; the generating the intermediate dose distribution result corresponding to this adjustment comprises: ​ obtaining the dose distribution input; at least one first agent of the plurality of agents adjusts the dose distribution input to obtain a first dose distribution; at least one second agent of the plurality of agents adjusts the dose distribution input based on the first dose distribution to obtain a second dose distribution; based on the first dose distribution and the second dose distribution, an intermediate dose distribution result generated by the current adjustment is obtained; wherein the first agent and the second agent are different; the intermediate dose distribution result includes an intermediate dose distribution and a target distribution value floating range, and a process of generating the intermediate dose distribution result once includes: each of the agents generates a fourth dose distribution multiple times based on the dose distribution input; generating a target distribution value floating range based on each fourth dose distribution; wherein the process of generating a fourth dose distribution once includes one of the following operations: randomly obtaining an action value corresponding to the reinforcement learning from a preset first probability distribution, and generating a fourth dose distribution according to the action value; randomly obtaining a preset weight distribution from a preset second probability distribution, adjusting the weights of each network in the agent according to the preset weight distribution, and generating a fourth dose distribution according to the adjusted agent; randomly selecting a target neuron connection in the agent, adjusting the agent according to the target neuron connection, and generating the fourth dose distribution according to the adjusted agent, wherein the target neuron connection does not participate in the generation process of the fourth dose distribution.

7. An electronic device comprising: one or more processors; a memory for storing one or more computer programs, characterized in that the one or more processors execute the one or more computer programs to implement the method according to any one of claims 1-5.

Images