Radiotherapy dose dynamic prediction and plan optimization method and system based on artificial intelligence

Abstract

The invention discloses a radiotherapy dose dynamic prediction and plan optimization method and system based on artificial intelligence. The method comprises the following steps: S1, acquiring initial radiotherapy plan data of a patient; s2, simulating to generate N virtual treatment CT images representing different anatomical states possibly occurring in the treatment process, wherein N is an integer greater than 1; s3, quickly predicting a three-dimensional dose deviation distribution diagram; s4, generating a group of risk indexes; and S5, outputting the optimized radiotherapy plan parameters. According to the method, in the treatment plan design stage, the three-dimensional dose deviation possibly generated by the current plan due to anatomical change in the future fractional treatment can be quickly and accurately predicted, the influence of the three-dimensional dose deviation on key clinical indexes is quantitatively evaluated, the optimization process is actively guided to generate a treatment plan with high robustness to uncertainty, and the treatment plan design efficiency is improved. And the treatment accuracy and safety are improved from the source.

Description

Technical Field

[0001] This invention relates to the field of medical physics and artificial intelligence, specifically to a method and system for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence. Background Technology

[0002] Radiation therapy is one of the three core methods of cancer treatment, with approximately 70% of cancer patients requiring it during their treatment. Its core objective is to kill tumor cells using high-energy radiation while protecting surrounding normal tissues and vital organs as much as possible. Intensity-modulated radiotherapy (IMRT) and its advanced technique, volumetric modulated radiotherapy (VMAT), can modulate the intensity distribution of the beam to create a dose distribution highly conformal to the three-dimensional shape of the tumor target area, and has become a mainstream precision radiotherapy technique. A typical IMRT / VMAT treatment workflow includes: simulation localization (acquiring CT images), target area and OAR (Organizational Angiography) delineation, treatment plan design, plan validation, and fractionated treatment execution.

[0003] Treatment planning is a crucial step in the radiotherapy process. Physicists, using a treatment planning system, set a series of clinical goals based on simulated CT images of the patient (e.g., 95% of the target volume receiving at least 95% of the prescribed dose, a maximum spinal dose below 45 Gy, etc.), and calculate the required beam parameters using an inverse optimization algorithm, thus generating an "optimal" static plan under baseline conditions. However, this static plan faces significant challenges during fractionated treatments lasting several weeks. The patient's anatomy changes due to tumor shrinkage, weight fluctuations, varying bladder / rectal filling levels, respiratory movements, and daily positioning errors. This leads to inconsistencies between the actual anatomical state during treatment and the state at the time of planning, potentially causing significant dose deviations in the pre-calculated "optimal" static plan during actual execution, jeopardizing treatment efficacy and patient safety. To address this issue, existing technologies mainly offer the following solutions: Image-guided radiotherapy and replanning: This is the current standard clinical approach for managing anatomical changes. Before each treatment, images of the patient are acquired using equipment such as cone-beam CT and registered with the planning CT scan to correct for positioning errors. When significant anatomical changes are found that could lead to unacceptable dosimetric results, clinicians may decide to interrupt the current plan, perform a new CT scan, and develop a new plan. However, this approach is reactive; replanning is costly, time-consuming, and heavily reliant on the physician's experience and judgment, lacking objective and quantifiable decision-making criteria.

[0004] Dose reconstruction and accumulation: Some research methods attempt to perform deformation registration between pre-treatment CBCT images and planning CT scans in fractionated therapy, then calculate the original planned dose on the CBCT images, and finally map the fractionated doses back to the planning CT coordinate system using a deformation field for superposition to obtain the actual cumulative dose. This method can achieve post-treatment total dose deviation assessment, but the calculation process is cumbersome and time-consuming, and it cannot be used for prospective prediction and proactive intervention during the treatment planning stage.

[0005] Probabilistic robust optimization: This approach predefines a series of possible simple error scenarios during plan optimization, and the optimizer simultaneously optimizes the dose distribution under all scenarios to obtain a plan insensitive to uncertainty. However, this method typically only considers simple geometric error models, making it difficult to simulate real, complex, and patient-specific organ deformations. Furthermore, it is computationally intensive, slow in optimization, and its clinical application is limited.

[0006] In summary, existing technologies have the following main shortcomings: They are lagging and reactive, with mainstream approaches operating on a "problem-finding and then solving" model, failing to anticipate and avoid dose deviation risks before treatment begins; they lack forward-looking prediction, as current methods cannot provide clinicians with intuitive, quantitative assessments of future dose deviation risks during the planning and design phase; they are insufficient in modeling complex anatomical changes, as traditional robust optimization methods struggle to effectively simulate real, non-rigid organ deformations, resulting in limited optimization effectiveness; and they rely on experience for decision-making, lacking objective, quantitative indicators based on dose deviation risk to support whether replanning is necessary, potentially leading to decision delays or inconsistencies. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, one of the objectives of this invention is to provide a method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence. This method aims to quickly and accurately predict the three-dimensional dose deviation that may occur in future fractionation treatments due to anatomical changes during the treatment planning stage, quantitatively assess its impact on key clinical indicators, and proactively guide the optimization process to generate a treatment plan that is highly robust to uncertainty, thereby improving the accuracy and safety of treatment from the source.

[0008] The second objective of this invention is to provide a method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence. This method aims to quickly and accurately predict the three-dimensional dose deviation that may occur in future fractionation treatments due to anatomical changes during the treatment planning stage, quantitatively assess its impact on key clinical indicators, and proactively guide the optimization process to generate a treatment plan that is highly robust to uncertainty, thereby improving the accuracy and safety of treatment from the source.

[0009] To achieve one of the objectives of this invention, the following solution is adopted: An AI-based method for dynamic prediction and planning optimization of radiotherapy dose includes the following steps: S1. Obtain the patient's initial radiotherapy plan data, which includes at least the planned CT images, structure set, initial static planning parameters, and the planned dose distribution calculated on the planned CT images; S2. Based on the planned CT image and the structure set, simulate and generate N virtual treatment CT images representing different anatomical states that may occur during the treatment process, where N is an integer greater than 1; S3. For each of the virtual treatment CT images, a pre-trained deep learning model is used to quickly predict the three-dimensional dose deviation distribution map generated when the radiotherapy plan corresponding to the initial static planning parameters is executed in the virtual anatomical state. S4. Based on the N three-dimensional dose deviation distribution maps and the planned dose distribution, quantitatively assess the dose deviation risk faced by the initial radiotherapy plan in future fractionated treatments, and generate a set of risk indicators. S5. With improving the risk indicators as one of the optimization objectives, iteratively optimize the initial static planning parameters until the preset convergence condition is met, and output the optimized radiotherapy planning parameters.

[0010] Further, step S2 includes: The planned CT images and the structure set are input into a trained generative adversarial network model; The generative adversarial network model outputs N three-dimensional deformation vector fields; The planned CT image is spatially transformed using the three-dimensional deformation vector field to generate the N virtual treatment CT images.

[0011] Further, step S2 includes: Based on a pre-established statistical shape model for key organs, parameters are sampled in the shape space to generate N sets of new organ shape parameters.

[0012] The sampled organ shapes are embedded into the planned CT image, replacing the original organ outlines, and the N virtual treatment CT images are generated through image fusion.

[0013] Further, the pre-trained deep learning model is a 3D U-Net network; step S3 includes: Calculate the difference image between the virtual treatment CT image and the planned CT image; The difference image is concatenated with the encoded initial static plan parameters to form a combined feature map; The combined feature map is input into the 3D U-Net network, and the 3D U-Net network outputs the three-dimensional dose deviation distribution map.

[0014] Furthermore, the 3D U-Net network is trained through the following steps: Acquire training data from historical patients, including planned CT images, CBCT images acquired during treatment, corresponding radiotherapy planning parameters, and actual dose distribution recalculated on the CBCT images; Based on the deformation registration algorithm, the deformation vector field from the planned CT image to the CBCT image is calculated, and the actual dose distribution is mapped back to the planned CT coordinate system to obtain the actual dose distribution after deformation. The difference between the actual dose distribution after deformation and the original planned dose distribution is calculated and used as the training true value; The 3D U-Net network is trained using the difference images between the planned CT and CBCT and the corresponding radiotherapy planning parameters as inputs, and the training ground truth as a supervision signal.

[0015] Further, step S4 includes: For the scene corresponding to the i-th virtual treatment CT image, the predicted three-dimensional dose deviation distribution map is added to the planned dose distribution to obtain the predicted dose distribution for that scene; On the predicted dose distribution, a predefined clinical dose-volume histogram index is calculated; For each of the N scenarios, the worst value of the clinical dose-volume histogram index is extracted across all scenarios, and the worst value constitutes the risk index.

[0016] Furthermore, the risk indicators include at least: the lowest D95 dose value of the planned target area in all simulated scenarios, and / or the highest Vx dose volume percentage value of the organ at risk in all simulated scenarios.

[0017] Further, in step S5, an extended optimization objective function is adopted, which is expressed as: Where F(D_plan) is the traditional optimization objective function based on the planned dose distribution, C({Risk_Indicator}) is the robustness penalty term based on the risk indicator, and α and β are weighting coefficients.

[0018] Furthermore, the robustness penalty term C({Risk_Indicator}) includes: The penalty term C1 = -min(D95_i) used to improve the target dose in the worst scenario is optimized to maximize the minimum value of the target D95 index in all simulation scenarios. Alternatively, the penalty term C2 = max(V50_i) is used to control the dose of organs at risk in the worst-case scenario, and its optimization objective is to minimize the maximum value of the V50 index of organs at risk in all simulation scenarios.

[0019] To achieve the second objective of this invention, the following solution is adopted: An AI-based system for dynamic prediction and planning optimization of radiotherapy dose includes: The data input module is used to acquire the patient's initial radiotherapy plan data, which includes at least the planned CT images, structure set, initial static planning parameters, and the planned dose distribution calculated on the planned CT images; The anatomical change simulation module is used to simulate and generate N virtual treatment CT images representing different anatomical states that may occur during treatment, based on the planned CT images and the structure set, where N is an integer greater than 1; The AI ​​dose deviation prediction module is used to quickly predict the three-dimensional dose deviation distribution map generated when the radiotherapy plan corresponding to the initial static planning parameters is executed in the virtual anatomical state for each virtual treatment CT image using a pre-trained deep learning model. The dose deviation assessment and quantification module is used to quantitatively assess the dose deviation risk faced by the initial radiotherapy plan in future fractionated treatments based on N three-dimensional dose deviation distribution maps and the planned dose distribution, and generate a set of risk indicators. The AI-guided planning optimization module is used to iteratively optimize the initial static planning parameters with the improvement of the risk indicators as one of the optimization objectives, until the preset convergence conditions are met, and output the optimized radiotherapy planning parameters.

[0020] Compared to existing technologies, the beneficial effects of this invention are as follows: This invention achieves forward-looking and predictive optimization. By simulating and generating virtual treatment CT images representing different anatomical states that may occur during treatment during the treatment planning stage, and rapidly predicting three-dimensional dose deviation distribution maps, this invention transforms the traditional open-loop optimization mode based on a single-phase static image into a closed-loop mode that can proactively predict, assess, and adapt to uncertainties during treatment. This transformation allows dose deviation risks to be foreseen and avoided before treatment begins, improving the accuracy and safety of treatment from the source. This invention elevates clinical decision-making from experience-driven to data-driven. By quantifying and generating a set of risk indicators, this invention transforms abstract judgments dependent on physician subjective experience into objective and quantifiable dose deviation risk data. This provides clinicians with clear, consistent, and objective quantitative evidence to judge the quality of the current plan, assess future risks, and decide whether a replanning is necessary, significantly reducing decision-making uncertainty and delays. This invention significantly improves the robustness of treatment plans. This invention uses improving risk indicators as one of its optimization objectives for iterative optimization, its core objective being… This invention directly targets the worst-case performance of the treatment plan across all simulated scenarios. The final optimized plan not only meets the requirements under baseline anatomical conditions but, more importantly, demonstrates stronger stability in key clinical dosimetry indicators when facing various complex anatomical changes in the simulation. This ensures the achievement of dosimetry goals throughout the treatment cycle. This invention significantly improves the efficiency of plan optimization. It employs a pre-trained deep learning model to rapidly predict dose deviations, overcoming the computational bottleneck of traditional physical dosimetry algorithms. This model can complete predictions for a single scenario within seconds, making it possible to integrate the analysis of dozens of complex scenarios in the optimization loop. The entire multi-scenario robust optimization process can be completed within a clinically acceptable timeframe. This invention achieves personalized, precise, and robust optimization. Both the simulation and prediction are based on the patient's initial radiotherapy plan data, including their unique planned CT images and anatomical structure set. This means that the simulated anatomical changes and subsequent optimizations are highly personalized processes for specific patients. Compared to traditional robust optimization based on general error models, the optimization results of this method are more targeted and closer to the patient's actual condition. In summary, this invention creatively combines a rapid AI prediction model with a treatment planning optimization loop, enabling proactive management and avoidance of dosage deviations caused by complex and patient-specific anatomical variations. Compared to traditional robust optimization techniques that can only handle simple errors and are computationally slow, this invention achieves breakthroughs in computational efficiency and simulation realism using artificial intelligence, making the optimization process more intelligent and closer to clinical practice. The resulting treatment plan exhibits excellent robustness and safety. Attached Figure Description

[0021] Figure 1 This is a flowchart of the AI-based dynamic prediction and planning optimization method for radiotherapy dose in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the training and application of the AI ​​dose deviation prediction model in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the specific steps of the AI-based dynamic prediction and planning optimization method for radiotherapy dose in an embodiment of the present invention. Figure 4 This is a block diagram of the AI-based dynamic prediction and planning optimization system for radiotherapy dose in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0023] Example 1 This invention provides an AI-based method for dynamic prediction and planning optimization of radiotherapy dose, constructing a dynamic, closed-loop radiotherapy planning optimization method based on artificial intelligence. This method transforms treatment planning from a traditional "open-loop" model based on a single-phase static image into a "closed-loop" model capable of predicting, evaluating, and proactively adapting to uncertainties during treatment. In the planning stage, this AI-based method simulates anatomical changes during treatment using generative AI, rapidly assesses the dosimetric consequences of these changes using predictive AI, and uses this assessment result as a key signal guiding the optimizer's search direction, ultimately outputting a robust plan insensitive to anatomical changes.

[0024] like Figure 1-3 As shown, this embodiment of the invention provides a method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence, including the following steps: S1. Obtain the patient's initial radiotherapy plan data, which includes at least the planned CT images, structure set, initial static planning parameters, and the planned dose distribution calculated on the planned CT images.

[0025] Initialization and data input: First, the data input interface module receives the initial data packet from the traditional treatment planning system (TPS). This data packet contains the following key elements: Planning CT Image: A three-dimensional digital image matrix that represents the anatomical position of the patient during simulation and serves as the baseline geometric model for dose calculation.

[0026] Structure Set: A set of three-dimensional contours outlined on a planning CT scan, including one or more planning target volumes (PTVs) and multiple organs at risk (OARs).

[0027] Initial static plan parameters include, but are not limited to, beam energy, field angle, subfield shape and weight, multileaf grating (MLC) sequence, and machine hop count (MU). These parameters define an initial IMRT or VMAT plan that meets clinical requirements.

[0028] Planned Dose Distribution (D_plan): A three-dimensional dose matrix calculated by TPS based on the planned CT and initial planning parameters.

[0029] The function of this step is to complete the standardized reading and verification of data, preparing it for subsequent processing.

[0030] S2. Based on the planned CT image and the structure set, simulate and generate N virtual treatment CT images representing different anatomical states that may occur during the treatment process, where N is an integer greater than 1.

[0031] Modeling treatment uncertainty: This is the first step in achieving prediction from "static" to "dynamic". The task of this step is to generate a series of images that can represent different patient anatomical states that may occur in future staged treatments, namely "virtual treatment CT".

[0032] There are two preferred ways to implement this technology: Method 1: Deformation field prediction model based on generative adversarial networks: Model input: planned CT images, structure set (especially organs closely related to deformation, such as bladder and rectum).

[0033] Model Architecture: A Conditional Generative Adversarial Network (GAN) is employed. The generator (G) is typically a convolutional neural network with a U-Net structure, which learns to generate a three-dimensional deformation vector field (DVF) conditioned on the aforementioned input. Each vector in this DVF indicates the direction and distance that the corresponding voxel on the planned CT scan should move when anatomical changes occur.

[0034] Model training: Training was conducted using a large amount of paired historical patient data. The training data included planning CT scans and CBCT scans acquired at different treatment fractions. A deformation registration algorithm was used to calculate the true DVF from the planning CT scan to each treatment CBCT scan as a supervisory signal. The discriminator (D) then attempted to distinguish between the generator-generated "virtual DVF" and the true "treatment DVF".

[0035] Application reasoning: The trained generator G receives the planned CT scan and structure set of a new patient and can quickly generate multiple different and reasonable DVFs (denoted as {DVF_i}, i=1, 2, ..., N). Each DVF_i is applied to the planned CT scan through a SpatialTransformer Layer to obtain the corresponding virtual treatment CT_i. By controlling the random noise vector input to the generator, a wide variety of samples can be generated.

[0036] Method 2: Parametric sampling based on statistical shape models: Model Construction: Based on a large sample database, a statistical shape model (SSM) is established for the surface contours of key organs (such as the prostate and bladder). Principal component analysis (PCA) is used to extract the main patterns (eigenvectors) of shape changes and their variances (eigenvalues).

[0037] Parameter sampling: Within the shape space defined by the PCA model, random sampling is performed according to the variance of the features to generate a series of new organ shape parameters.

[0038] Image synthesis: The sampled new organ shape is embedded into the planned CT using mesh deformation technology to replace the original organ outline, and a virtual treatment CT image that looks realistic is generated through image fusion algorithm.

[0039] The output of this step is a set of N (e.g., N=20) virtual therapeutic CT images {CT_virtual_i}, which together constitute a quantitative and visual description of the treatment uncertainty.

[0040] S3. For each of the virtual treatment CT images, a pre-trained deep learning model is used to quickly predict the three-dimensional dose deviation distribution map generated when the radiotherapy plan corresponding to the initial static planning parameters is executed in the virtual anatomical state.

[0041] Rapid prediction of dose-response effects: This is the core technology of this invention. The goal of this step is to bypass time-consuming physical dose calculations and directly and quickly predict the deviation between the dose distribution and the original planned dose D_plan when the initial plan is executed on each virtual treatment CT_i.

[0042] Input data preparation: For each virtual scene i, calculate the difference image between its virtual treatment CT_i and planning CT, i.e., ΔCT_i = CT_virtual_i - Planning_CT. This difference image highlights the density and geometric variations of anatomical structures. Simultaneously, initial planning parameters (such as field direction) are encoded into feature maps that match the CT image size.

[0043] Prediction Model Architecture: 3D U-Net is preferably used as the main architecture of the prediction network. The U-Net encoder (shrinking path) extracts multi-level features from the input ΔCT_i and planning parameter feature maps; the decoder (expanding path) upsamples and fuses these features to progressively reconstruct a high-resolution 3D dose deviation map. Skip connections ensure that low-level details are not lost.

[0044] Model Input and Output: The model input consists of concatenated ΔCT_i images and a planning parameter feature map. The output is a three-dimensional matrix, namely the predicted dose deviation map ΔD_predicted_i. The value of each voxel in this map represents the dose change (in Gy) at that location due to anatomical variations. Positive values ​​indicate increased dose, and negative values ​​indicate decreased dose.

[0045] Model training (offline process): Training data: A large-scale historical dataset is required. Each sample contains: a planned CT scan, a CBCT scan of a specific treatment, and the actual dose D_actual recalculated on the CBCT scan at that treatment using an accurate algorithm (such as Monte Carlo).

[0046] Truth preparation: Map the DVF obtained by deformation registration of D_actual back to the planning CT coordinate system to obtain D_actual_deformed. Then the true dose deviation map is: ΔD_truth = D_actual_deformed - D_plan.

[0047] Loss function: During training, the difference between the predicted value ΔD_predicted and the true value ΔD_truth is minimized. Mean squared error (MSE) is usually used or an improved loss function that incorporates dose gradient information (such as gradient loss weighted MSE) to better preserve the marginal characteristics of the dose distribution.

[0048] The advantage of this step lies in its extremely high inference speed. Once the model is trained, it can make predictions for a new virtual scene in seconds or even sub-seconds, making it possible to integrate a large amount of scene analysis into the optimization loop.

[0049] S4. Based on the N three-dimensional dose deviation distribution maps and the planned dose distribution, quantitatively assess the dose deviation risk faced by the initial radiotherapy plan in future fractionated treatments, and generate a set of risk indicators.

[0050] Risk Quantitative Assessment: This module is responsible for converting the physical quantities (dose deviation) predicted by AI into risk indicators that have direct guiding significance for clinical decision-making.

[0051] Predicted dose calculation: For each virtual scene i, the predicted dose deviation ΔD_predicted_i is added to the original planned dose D_plan to obtain the predicted dose distribution in that virtual anatomical state: D_predicted_i = D_plan + ΔD_predicted_i.

[0052] Key metric extraction: For each D_predicted_i, calculate the dose-volume histogram (DVH) metric of greatest clinical interest. For example: For the target volume (PTV): D98 (dose received by 98% of the target volume), D95, and Dmean. These metrics measure target coverage and dose uniformity.

[0053] For organs at risk (OARs): Dmax (maximum dose), Dmean, Vx (volume percentage of those receiving doses of x Gy or higher, such as V50 for the rectum and V40 for the bladder).

[0054] Risk aggregation: Iterate through all N virtual scenarios, find the worst value of each key indicator across all scenarios, and form a set of "robust risk indicators". For example: Target coverage risk: Risk_PTV_D95 = min( D95_i ​​) over i=1 to N (Minimum D95 value of PTV in all scenarios) Risk of exceeding organ limits: Risk_Rectum_V50 = max( V50_i ) over i=1 to N (the maximum value of rectal V50 in all scenarios) This set of {Risk_Indicator} quantifies the worst-case scenario that the current static plan may face in future treatments.

[0055] S5. With improving the risk indicators as one of the optimization objectives, iteratively optimize the initial static planning parameters until the preset convergence condition is met, and output the optimized radiotherapy planning parameters.

[0056] Closed-loop optimization execution: This step is the "brain" of the closed-loop system, using the aforementioned risk indicators to redefine "what constitutes a good plan".

[0057] Extended Optimization Objective Function: The traditional inverse optimization objective function F only considers the dose distribution D_plan on the planned CT scan. This invention extends it to a multi-objective robust optimization function F_robust: Where F(D_plan) is the original objective function, ensuring that the optimized plan maintains high quality under the baseline dissection conditions.

[0058] C({Risk_Indicator}) is a newly added robustness penalty term. Its specific form depends on the clinical objective. For example, to improve the target dose in the worst-case scenario, C1 can be set to -min(D95_i), and the optimization objective is to maximize min(D95_i), which is equivalent to minimizing C1. To control the risk of OARs, C2 can be set to max(V50_i), and the optimization objective is to minimize C2.

[0059] α and β are weighting coefficients used to balance the trade-off between the quality and robustness of the baseline program.

[0060] Optimize the loop: The optimizer (e.g., gradient descent, stochastic parallel gradient descent) calculates the corresponding F_robust based on the current planning parameters. Then, the optimizer fine-tunes the planning parameters (e.g., subfield weights) and requests the system to recalculate the predicted dose and risk indicators for all virtual scenarios under the new parameters (i.e., rapid prediction of repeated dose consequences and risk quantification assessment, but typically does not require rerunning treatment uncertainty modeling). Through iteration, the optimizer finds the set of planning parameters that minimizes F_robust. The loop terminates when the objective function converges or reaches the maximum number of iterations.

[0061] Final output: After the optimization cycle is complete, the final optimized plan parameters and their dose distribution on the planning CT scan are output. This new plan not only meets clinical requirements on baseline images, but more importantly, it demonstrates greater stability in the face of various simulated anatomical variations. Clinicians can simultaneously review the dose distribution of the plan on the planning CT scan, as well as robustness analysis reports (such as the degree of improvement of each risk indicator compared to the initial plan), thereby making more confident clinical decisions.

[0062] Further, step S2 includes: The planned CT images and the structure set are input into a trained generative adversarial network model; The generative adversarial network model outputs N three-dimensional deformation vector fields; The planned CT image is spatially transformed using the three-dimensional deformation vector field to generate the N virtual treatment CT images.

[0063] Further, step S2 includes: Based on a pre-established statistical shape model for key organs, parameters are sampled in the shape space to generate N sets of new organ shape parameters. The sampled organ shapes are embedded into the planned CT image, replacing the original organ outlines, and the N virtual treatment CT images are generated through image fusion.

[0064] Further, the pre-trained deep learning model is a 3D U-Net network; step S3 includes: Calculate the difference image between the virtual treatment CT image and the planned CT image; The difference image is concatenated with the encoded initial static plan parameters to form a combined feature map; The combined feature map is input into the 3D U-Net network, and the 3D U-Net network outputs the three-dimensional dose deviation distribution map.

[0065] Furthermore, the 3D U-Net network is trained through the following steps: Acquire training data from historical patients, including planned CT images, CBCT images acquired during treatment, corresponding radiotherapy planning parameters, and actual dose distribution recalculated on the CBCT images; Based on the deformation registration algorithm, the deformation vector field from the planned CT image to the CBCT image is calculated, and the actual dose distribution is mapped back to the planned CT coordinate system to obtain the actual dose distribution after deformation. The difference between the actual dose distribution after deformation and the original planned dose distribution is calculated and used as the training true value; The 3D U-Net network is trained using the difference images between the planned CT and CBCT and the corresponding radiotherapy planning parameters as inputs, and the training ground truth as a supervision signal.

[0066] Further, step S4 includes: For the scene corresponding to the i-th virtual treatment CT image, the predicted three-dimensional dose deviation distribution map is added to the planned dose distribution to obtain the predicted dose distribution for that scene; On the predicted dose distribution, a predefined clinical dose-volume histogram index is calculated; For each of the N scenarios, the worst value of the clinical dose-volume histogram index is extracted across all scenarios, and the worst value constitutes the risk index.

[0067] Furthermore, the risk indicators include at least: the lowest D95 dose value of the planned target area in all simulated scenarios, and / or the highest Vx dose volume percentage value of the organ at risk in all simulated scenarios.

[0068] Further, in step S5, an extended optimization objective function is adopted, which is expressed as: Where F(D_plan) is the traditional optimization objective function based on the planned dose distribution, C({Risk_Indicator}) is the robustness penalty term based on the risk indicator, and α and β are weighting coefficients.

[0069] Furthermore, the robustness penalty term C({Risk_Indicator}) includes: The penalty term C1 = -min(D95_i) used to improve the target dose in the worst scenario is optimized to maximize the minimum value of the target D95 index in all simulation scenarios. Alternatively, the penalty term C2 = max(V50_i) is used to control the dose of organs at risk in the worst-case scenario, and its optimization objective is to minimize the maximum value of the V50 index of organs at risk in all simulation scenarios.

[0070] Experimental example: Take VMAT treatment for prostate cancer as an example: 1. Data Preparation and Model Training (Offline): Data from 500 prostate cancer patients who had completed treatment were collected. Each case included planning CT scans, 10-20 treatment CBCT scans, VMAT planning parameters, and doses recalculated using the Monte Carlo algorithm on each treatment CBCT scan. After rigorous deformation registration and data processing, a 3D U-Net model was trained. This model learns the effects of bladder and rectal volume and shape changes on dose distribution.

[0071] 2. New patient planning design (online): Step 1: The doctor sketches structures such as the prostate, seminal vesicles, bladder, and rectum. The physicist designs an initial VMAT plan (static plan) that meets clinical objectives.

[0072] Step 2: The anatomical change simulation module generates 15 sets of virtual treatment CT based on the statistical changes in bladder and rectal volume in the prostate cancer population, covering various possible combinations from extreme bladder emptying to fullness and rectal expansion from emptiness to expansion.

[0073] Step 3: Input 15 sets of image pairs (planning CT, virtual therapeutic CT) and initial planning parameters into the trained AI model. The model predicts 15 dose deviation maps ΔD_i in parallel within 10 seconds.

[0074] Step 4: The evaluation module calculations revealed that in three scenarios involving bladder emptying and rectal dilation, the D95_predicted value of the prostate target area was more than 5% lower than the planned value; in five scenarios, the V50_predicted value of the rectum exceeded the planned limit.

[0075] Step 5: Start the optimization module and set the optimization goal as follows: while ensuring that the target area D95 on the planned CT is greater than 95%, maximize the smallest prostate D95_predicted among the 15 scenarios and minimize the largest rectal V50_predicted among the 15 scenarios.

[0076] Step Six: After approximately 50 iterations, convergence was optimized. The final plan showed a slight decrease in prostate D95 on the planned CT scan from 96% to 95.5%, but its predicted D95 ​​in the worst-case scenario improved from 90% to 93%; the V50 of the rectum in the worst-case scenario decreased from 55% to 52%. This plan was selected as the final treatment plan due to its greater stability in the face of anatomical changes.

[0077] Example 2 like Figure 2-4 As shown, embodiments of the present invention also provide an artificial intelligence-based system for dynamic prediction and planning optimization of radiotherapy dose, including: The data input module is used to acquire the patient's initial radiotherapy plan data, which includes at least the planned CT images, structure set, initial static planning parameters, and the planned dose distribution calculated on the planned CT images.

[0078] System initialization and data input: After the system starts up, the data input interface module first receives an initial data packet from the traditional treatment planning system (TPS). This data packet contains the following key elements: Planning CT Image: A three-dimensional digital image matrix that represents the anatomical position of the patient during simulation and serves as the baseline geometric model for dose calculation.

[0079] Structure Set: A set of three-dimensional contours outlined on a planning CT scan, including one or more planning target volumes (PTVs) and multiple organs at risk (OARs).

[0080] Initial static plan parameters include, but are not limited to, beam energy, field angle, subfield shape and weight, multileaf grating (MLC) sequence, and machine hop count (MU). These parameters define an initial IMRT or VMAT plan that meets clinical requirements.

[0081] Planned Dose Distribution (D_plan): A three-dimensional dose matrix calculated by TPS based on the planned CT and initial planning parameters.

[0082] This module is designed to perform standardized data reading and verification, preparing the data for subsequent processing.

[0083] The anatomical change simulation module is used to simulate and generate N virtual treatment CT images representing different anatomical states that may occur during treatment, based on the planned CT images and the structure set, where N is an integer greater than 1.

[0084] Modeling treatment uncertainty: This is the first step towards achieving prediction from "static" to "dynamic". The module's task is to generate a series of images that can represent different patient anatomical states that may occur in future staged treatments, namely "Virtual Treatment CT".

[0085] There are two preferred ways to implement this technology: Method 1: Deformation field prediction model based on generative adversarial networks Model input: planned CT images, structure set (especially organs closely related to deformation, such as bladder and rectum).

[0086] Model Architecture: A Conditional Generative Adversarial Network (GAN) is employed. The generator (G) is typically a convolutional neural network with a U-Net structure, which learns to generate a three-dimensional deformation vector field (DVF) conditioned on the aforementioned input. Each vector in this DVF indicates the direction and distance that the corresponding voxel on the planned CT scan should move when anatomical changes occur.

[0087] Model training: Training was conducted using a large amount of paired historical patient data. The training data included planning CT scans and CBCT scans acquired at different treatment fractions. A deformation registration algorithm was used to calculate the true DVF from the planning CT scan to each treatment CBCT scan as a supervisory signal. The discriminator (D) then attempted to distinguish between the generator-generated "virtual DVF" and the true "treatment DVF".

[0088] Application reasoning: The trained generator G receives the planned CT scan and structure set of a new patient and can quickly generate multiple different and reasonable DVFs (denoted as {DVF_i}, i=1, 2, ..., N). Each DVF_i is applied to the planned CT scan through a SpatialTransformer Layer to obtain the corresponding virtual treatment CT_i. By controlling the random noise vector input to the generator, a wide variety of samples can be generated.

[0089] Method 2: Parametric Sampling Based on Statistical Shape Model Model Construction: Based on a large sample database, a statistical shape model (SSM) is established for the surface contours of key organs (such as the prostate and bladder). Principal component analysis (PCA) is used to extract the main patterns (eigenvectors) of shape changes and their variances (eigenvalues).

[0090] Parameter sampling: Within the shape space defined by the PCA model, random sampling is performed according to the variance of the features to generate a series of new organ shape parameters.

[0091] Image synthesis: The sampled new organ shape is embedded into the planned CT using mesh deformation technology to replace the original organ outline, and a virtual treatment CT image that looks realistic is generated through image fusion algorithm.

[0092] The output of this module is a set of N (e.g., N=20) virtual therapeutic CT images {CT_virtual_i}, which together constitute a quantitative and visual description of the treatment uncertainty.

[0093] The AI ​​dose deviation prediction module is used to quickly predict the three-dimensional dose deviation distribution map generated when the radiotherapy plan corresponding to the initial static planning parameters is executed in the virtual anatomical state for each virtual treatment CT image using a pre-trained deep learning model.

[0094] Rapid prediction of dose-response effects: This is the core technology of this invention. The goal of this module is to bypass time-consuming physical dose calculations and directly and quickly predict the deviation between the dose distribution and the original planned dose D_plan when the initial plan is executed on each virtual treatment CT_i.

[0095] Input data preparation: For each virtual scene i, calculate the difference image between its virtual treatment CT_i and planning CT, i.e., ΔCT_i = CT_virtual_i - Planning_CT. This difference image highlights the density and geometric variations of anatomical structures. Simultaneously, initial planning parameters (such as field direction) are encoded into feature maps that match the CT image size.

[0096] Prediction Model Architecture: 3D U-Net is preferably used as the main architecture of the prediction network. The U-Net encoder (shrinking path) extracts multi-level features from the input ΔCT_i and planning parameter feature maps; the decoder (expanding path) upsamples and fuses these features to progressively reconstruct a high-resolution 3D dose deviation map. Skip connections ensure that low-level details are not lost.

[0097] Model Input and Output: The model input consists of concatenated ΔCT_i images and a planning parameter feature map. The output is a three-dimensional matrix, namely the predicted dose deviation map ΔD_predicted_i. The value of each voxel in this map represents the dose change (in Gy) at that location due to anatomical variations. Positive values ​​indicate increased dose, and negative values ​​indicate decreased dose.

[0098] Model training (offline process): Training data: A large-scale historical dataset is required. Each sample contains: a planned CT scan, a CBCT scan of a specific treatment, and the actual dose D_actual recalculated on the CBCT scan at that treatment using an accurate algorithm (such as Monte Carlo).

[0099] Truth preparation: Map the DVF obtained by deformation registration of D_actual back to the planning CT coordinate system to obtain D_actual_deformed. Then the true dose deviation map is: ΔD_truth = D_actual_deformed - D_plan.

[0100] Loss function: During training, the difference between the predicted value ΔD_predicted and the true value ΔD_truth is minimized. Mean squared error (MSE) is usually used or an improved loss function that incorporates dose gradient information (such as gradient loss weighted MSE) to better preserve the marginal characteristics of the dose distribution.

[0101] The advantage of this module lies in its extremely high inference speed. Once the model is trained, it can make predictions for a new virtual scene in seconds or even sub-seconds, making it possible to integrate a large amount of scene analysis into the optimization loop.

[0102] like Figure 2 The diagram illustrates the core workflow of the AI ​​dose deviation prediction module. Training Phase (Offline): Training is performed using a large amount of historical patient data (planning CT, multiple treatment CBCTs, planning parameters, and actual doses recalculated on CBCT). First, CBCTs are registered with the planning CT, and their difference image ΔCT is calculated. The deviation ΔD between the actual and planned doses is used as the ground truth. A deep learning model (such as 3D U-Net) is trained to learn the mapping relationship from ΔCT and planning parameters to ΔD. Application Phase (Online): For new patients, the difference image ΔCT_virtual between the planning CT and the simulated virtual treatment CT, along with the planning parameters, are input into the trained model, which quickly outputs the predicted dose deviation ΔD_predicted.

[0103] The dose deviation assessment and quantification module is used to quantitatively assess the dose deviation risk faced by the initial radiotherapy plan in future fractionated treatments based on N three-dimensional dose deviation distribution maps and the planned dose distribution, and generate a set of risk indicators.

[0104] Risk Quantitative Assessment: This module is responsible for converting the physical quantities (dose deviation) predicted by AI into risk indicators that have direct guiding significance for clinical decision-making.

[0105] Predicted dose calculation: For each virtual scene i, the predicted dose deviation ΔD_predicted_i is added to the original planned dose D_plan to obtain the predicted dose distribution in that virtual anatomical state: D_predicted_i = D_plan + ΔD_predicted_i.

[0106] Key metric extraction: For each D_predicted_i, calculate the dose-volume histogram (DVH) metric of greatest clinical interest. For example: For the target volume (PTV): D98 (dose received by 98% of the target volume), D95, and Dmean. These metrics measure target coverage and dose uniformity.

[0107] For organs at risk (OARs): Dmax (maximum dose), Dmean, Vx (volume percentage of those receiving doses of x Gy or higher, such as V50 for the rectum and V40 for the bladder).

[0108] Risk aggregation: Iterate through all N virtual scenarios, find the worst value of each key indicator across all scenarios, and form a set of "robust risk indicators". For example: Target coverage risk: Risk_PTV_D95 = min( D95_i ​​) over i=1 to N (Minimum D95 value of PTV in all scenarios) Risk of exceeding organ limits: Risk_Rectum_V50 = max( V50_i ) over i=1 to N (the maximum value of rectal V50 in all scenarios) This set of {Risk_Indicator} quantifies the worst-case scenario that the current static plan may face in future treatments.

[0109] The AI-guided planning optimization module is used to iteratively optimize the initial static planning parameters with the improvement of the risk indicators as one of the optimization objectives, until the preset convergence conditions are met, and output the optimized radiotherapy planning parameters.

[0110] Closed-loop optimization execution: This module is the "brain" of the closed-loop system. It uses the aforementioned risk indicators to redefine "what constitutes a good plan".

[0111] Extended Optimization Objective Function: The traditional inverse optimization objective function F only considers the dose distribution D_plan on the planned CT scan. This invention extends it to a multi-objective robust optimization function F_robust: Where F(D_plan) is the original objective function, ensuring that the optimized plan maintains high quality under the baseline dissection conditions.

[0112] C({Risk_Indicator}) is a newly added robustness penalty term. Its specific form depends on the clinical objective. For example, to improve the target dose in the worst-case scenario, C1 can be set to -min(D95_i), and the optimization objective is to maximize min(D95_i), which is equivalent to minimizing C1. To control the risk of OARs, C2 can be set to max(V50_i), and the optimization objective is to minimize C2.

[0113] α and β are weighting coefficients used to balance the trade-off between the quality and robustness of the baseline program.

[0114] Optimize the loop: The optimizer (e.g., gradient descent, stochastic parallel gradient descent) calculates the corresponding F_robust based on the current planning parameters. Then, the optimizer fine-tunes the planning parameters (e.g., subfield weights) and requests the system to recalculate the predicted dose and risk indicators for all virtual scenarios under the new parameters (i.e., rapid prediction of repeated dose consequences and risk quantification assessment, but typically does not require rerunning treatment uncertainty modeling). Through iteration, the optimizer finds the set of planning parameters that minimizes F_robust. The loop terminates when the objective function converges or reaches the maximum number of iterations.

[0115] Final output: After the optimization cycle is complete, the system outputs the final optimized plan parameters and their dose distribution on the planning CT scan. This new plan not only meets clinical requirements on baseline images, but more importantly, it demonstrates greater stability in the face of various simulated anatomical variations. Clinicians can simultaneously review the dose distribution of the plan on the planning CT scan, as well as the robustness analysis report provided by the system (such as the degree of improvement of each risk indicator compared to the initial plan), thereby making more confident clinical decisions.

[0116] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the patient's initial radiotherapy plan data, which includes at least the planned CT images, structure set, initial static planning parameters, and the planned dose distribution calculated on the planned CT images; S2. Based on the planned CT image and the structure set, simulate and generate N virtual treatment CT images representing different anatomical states that may occur during the treatment process, where N is an integer greater than 1; S3. For each of the virtual treatment CT images, a pre-trained deep learning model is used to quickly predict the three-dimensional dose deviation distribution map generated when the radiotherapy plan corresponding to the initial static planning parameters is executed in the virtual anatomical state. S4. Based on the N three-dimensional dose deviation distribution maps and the planned dose distribution, quantitatively assess the dose deviation risk faced by the initial radiotherapy plan in future fractionated treatments, and generate a set of risk indicators. S5. With improving the risk indicators as one of the optimization objectives, iteratively optimize the initial static planning parameters until the preset convergence condition is met, and output the optimized radiotherapy planning parameters.

2. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 1, characterized in that, Step S2 includes: The planned CT images and the structure set are input into a trained generative adversarial network model; The generative adversarial network model outputs N three-dimensional deformation vector fields; The planned CT image is spatially transformed using the three-dimensional deformation vector field to generate the N virtual treatment CT images.

3. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 1, characterized in that, Step S2 includes: Based on a pre-established statistical shape model for key organs, parameters are sampled in the shape space to generate N sets of new organ shape parameters. The sampled organ shapes are embedded into the planned CT image, replacing the original organ outlines, and the N virtual treatment CT images are generated through image fusion.

4. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 1, characterized in that, The pre-trained deep learning model is a 3D U-Net network; step S3 includes: Calculate the difference image between the virtual treatment CT image and the planned CT image; The difference image is concatenated with the encoded initial static plan parameters to form a combined feature map; The combined feature map is input into the 3D U-Net network, and the 3D U-Net network outputs the three-dimensional dose deviation distribution map.

5. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 4, characterized in that, The 3D U-Net network is trained through the following steps: Acquire training data from historical patients, including planned CT images, CBCT images acquired during treatment, corresponding radiotherapy planning parameters, and actual dose distribution recalculated on the CBCT images; Based on the deformation registration algorithm, the deformation vector field from the planned CT image to the CBCT image is calculated, and the actual dose distribution is mapped back to the planned CT coordinate system to obtain the actual dose distribution after deformation. The difference between the actual dose distribution after deformation and the original planned dose distribution is calculated and used as the training true value; The 3D U-Net network is trained using the difference images between the planned CT and CBCT and the corresponding radiotherapy planning parameters as inputs, and the training ground truth as a supervision signal.

6. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 1, characterized in that, Step S4 includes: For the scene corresponding to the i-th virtual treatment CT image, the predicted three-dimensional dose deviation distribution map is added to the planned dose distribution to obtain the predicted dose distribution for that scene; On the predicted dose distribution, a predefined clinical dose-volume histogram index is calculated; For each of the N scenarios, the worst value of the clinical dose-volume histogram index is extracted across all scenarios, and the worst value constitutes the risk index.

7. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 6, characterized in that, The risk metrics include at least: the lowest D95 dose value of the planned target area in all simulated scenarios, and / or the highest Vx dose volume percentage value of the organ at risk in all simulated scenarios.

8. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 1, characterized in that, In step S5, the extended optimization objective function is adopted, which is expressed as: Where F(D_plan) is the traditional optimization objective function based on the planned dose distribution, C({Risk_Indicator}) is the robustness penalty term based on the risk indicator, and α and β are weighting coefficients.

9. The method for dynamic prediction and planning optimization of radiotherapy dose based on artificial intelligence according to claim 8, characterized in that, The robustness penalty term C({Risk_Indicator}) includes: The penalty term C1 = -min(D95_i) used to improve the target dose in the worst scenario is optimized to maximize the minimum value of the target D95 index in all simulation scenarios. Alternatively, the penalty term C2 = max(V50_i) is used to control the dose of organs at risk in the worst-case scenario, and its optimization objective is to minimize the maximum value of the V50 index of organs at risk in all simulation scenarios.

10. A radiotherapy dose dynamic prediction and planning optimization system based on artificial intelligence, characterized in that, include: The data input module is used to acquire the patient's initial radiotherapy plan data, which includes at least the planned CT images, structure set, initial static planning parameters, and the planned dose distribution calculated on the planned CT images; The anatomical change simulation module is used to simulate and generate N virtual treatment CT images representing different anatomical states that may occur during treatment, based on the planned CT images and the structure set, where N is an integer greater than 1; The AI ​​dose deviation prediction module is used to quickly predict the three-dimensional dose deviation distribution map generated when the radiotherapy plan corresponding to the initial static planning parameters is executed in the virtual anatomical state for each virtual treatment CT image using a pre-trained deep learning model. The dose deviation assessment and quantification module is used to quantitatively assess the dose deviation risk faced by the initial radiotherapy plan in future fractionated treatments based on N three-dimensional dose deviation distribution maps and the planned dose distribution, and generate a set of risk indicators. The AI-guided planning optimization module is used to iteratively optimize the initial static planning parameters with the improvement of the risk indicators as one of the optimization objectives, until the preset convergence conditions are met, and output the optimized radiotherapy planning parameters.

Images