Executable resident time plan generation method, system and application thereof

Abstract

The present application relates to the technical field of brachytherapy plan automatic optimization, and particularly relates to a method and system for generating a dwell time plan and application thereof, the method comprising converting unconstrained internal variables into non-negative dwell time through a differentiable non-negative mapping; a differentiable dose calculation operator accumulates a three-dimensional dose distribution according to a linear source dose calculation model, and analytically calculates a dwell time gradient when back-propagating; a differentiable clinical objective function is constructed according to the three-dimensional dose distribution; according to the violation degree of each clinical objective relative to a clinical threshold, a rule feedback controller automatically updates objective weights, drives a gradient optimizer to iteratively update the dwell time, and forms a closed-loop feedback process; and after a termination condition is met, a dwell time plan that can be directly executed by a afterloading treatment device is output. The present application is applicable to high-dose rate afterloading treatment of cervical cancer, and can also be extended to other brachytherapy scenarios using fixed applicator dwell positions.

Description

Technical Field

[0001] This invention relates to the field of automatic optimization technology for brachytherapy planning, and in particular to an executable residence time plan generation method, system, and application thereof. Background Technology

[0002] High-dose-rate afterloading therapy involves delivering a high-activity, sealed radiation source into the patient's body via an applicator or catheter, where it remains at multiple predetermined residence sites for varying durations to create a conformal dose distribution. The dose rate contribution of each residence site to the calculated point is determined using the TG-43 line source model based on air kerma intensity, dose rate constant, geometric functions, radial dose functions, and two-dimensional anisotropy functions. Residence time is a control variable in the actual execution of the afterloading device. Clinical planning is typically evaluated using dose-volume histograms such as HR-CTV V100, D90, and D2cc for the bladder and rectum. Image-guided afterloading therapy for cervical cancer requires applicator placement, image acquisition, target delineation, treatment planning, and irradiation to be completed within the same treatment timeframe. The planning time directly impacts patient waiting time and treatment efficiency.

[0003] Existing treatment planning methods have the following main shortcomings: manual or semi-automatic forward planning relies on the physicist's experience, requiring repeated observation of dose-volume histograms and gradual modification of residence time, resulting in strong operator dependence and significant differences in results between different operators or with different initial values; although inverse planning algorithms in commercial treatment planning systems (such as IPSA, HIPO, etc.) can directly optimize residence time, they still require manual pre-setting and repeated adjustment of dose target weights for each case, and fixed weights cannot automatically reflect the degree of competition between target coverage and organ-at-risk protection in different cases; deep learning-based three-dimensional dose prediction methods output intermediate dose distributions rather than residence times that can be directly executed by afterloading devices, still requiring additional optimization steps to convert predicted doses into residence times, and the conversion process may introduce errors. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an executable residence time plan generation method, system, and its application, solving the technical problems of the traditional TG-43 dose calculation program being disconnected from the gradient optimization framework, the inability of fixed target weights to adapt to different cases, and the inability of automatic plans to directly output executable residence times. Specifically: Traditional TG-43 dose calculation procedures only provide forward dose results and cannot directly backpropagate errors in target coverage and dose-volume targets of organs at risk to the residence time variable, resulting in a disconnect between physical dose calculation and gradient optimization framework. Fixed target weights in commercial treatment planning systems cannot automatically reflect the degree of competition between target coverage and organ protection in different cases. Physicists need to manually adjust them repeatedly for each patient, which leads to strong operator dependence and poor consistency. Existing automated planning methods cannot simultaneously meet the three requirements of physical dose calculation in the optimization loop, direct output of the executable dwell time of the afterloading device, and automatic correction according to clinical rules. Furthermore, a single optimization result cannot fully express the technical problems of diverse clinical preferences.

[0005] To address the aforementioned technical problems, this invention provides an executable residence time plan generation method, which includes the following steps: The residence position of the fixed radioactive source is set, and an unconstrained internal variable is set for each residence position. The unconstrained internal variable is converted into a non-negative residence time through a differentiable non-negative mapping. The coordinates of the dwell position, the source axis direction, the non-negative dwell time, and the radiation source dose calculation parameters are input into the differentiable dose calculation operator. The differentiable dose calculation operator performs forward dose accumulation on each calculation point according to the line source dose calculation model to obtain the three-dimensional dose distribution, and analyzes the dwell time gradient using the linear relationship between dose and dwell time during reverse propagation. Based on the three-dimensional dose distribution, a differentiable clinical objective function is constructed, which includes a target coverage loss term, an organ at risk dose excess loss term, and a residence time regularization term. The degree of violation of each clinical objective relative to the clinical threshold is calculated based on the current three-dimensional dose distribution, and the degree of violation is input into the rule feedback controller to automatically update the target weight corresponding to each clinical objective; The gradient optimizer is driven by the updated target weights, and the unconstrained internal variables are iteratively updated using the residence time gradient, forming a closed-loop feedback process of dose calculation, clinical evaluation, weight update and residence time update. Once the termination conditions are met, the system outputs a stay time plan that can be directly executed by the afterloading therapy device.

[0006] This invention connects a differential dose calculation operator, a clinical objective function, a rule feedback controller, and a gradient optimizer into a complete closed loop, forming a physical link with a computable gradient between the residence time variable and the three-dimensional dose distribution. The error of the clinical objective function can be directly backpropagated to each residence time variable. At the same time, the rule feedback controller automatically adjusts the target weights according to real-time clinical indicators, so that the entire optimization process can be completed without manual intervention.

[0007] Furthermore, the differentiable nonnegative mapping adopts softplus The function, the softplus The function expression is: in, For the first The unconstrained internal variables corresponding to each residence location; The non-negative dwell time after conversion; Alternatively, the differentiable nonnegative mapping may be any one of an exponential mapping, a square mapping, or a parameterized mapping with upper and lower bounds.

[0008] Furthermore, the forward computation of the differentiable dose calculation operator is performed in parallel, with each calculation point as a parallel task unit. For each calculation point, all dwell positions are traversed and the dose contribution of each dwell position is accumulated to obtain the total dose at that calculation point, expressed as: in, The total dose contributed by the cumulative dose per unit time; Number of locations to be occupied; For the first Each stationary location calculates the points within a unit of stationary time. The dose contribution generated; When the number of computation points is large, the computation points are input in blocks to control storage usage; The backpropagation process calculates the residence time gradient using the analytical derivative relationship between dose and residence time. This analytical derivative relationship is as follows: No. The dose at the calculation point for the first... The partial derivative of the dwell time at the first dwelling location is equal to the first dwelling location. The first residence location for the first The dose contribution value per unit time at each calculation point.

[0009] Furthermore, the target coverage loss term adopts an underdose squared penalty form. A penalty is incurred when the dose quantile corresponding to the target coverage volume ratio within the target area is lower than the prescribed dose. The penalty value is the square of the difference between the prescribed dose and the dose quantile, i.e.: in, Loss of target coverage; This is the prescription dosage; These are dose percentile values; The penalty for exceeding the dose limit for organs at risk adopts the form of a squared penalty for exceeding the limit. A penalty is generated when the highest irradiated volume dose of the organ at risk exceeds the corresponding dose limit. The penalty value is the square of the excess amount, that is: in, Indicates organs at risk Dose over-limit loss; To endanger organs The highest 2cm 3 dose; For corresponding dosage restrictions; The dwell time regularization term includes the total dwell time regularization term. Smoothing terms with adjacent residence points The adjacent dwell point smoothing term sums the squares of the dwell time differences between adjacent dwell positions within the same applicator channel, i.e.: in, This refers to the set of adjacent dwelling positions within the same channel; and The dwell time of adjacent dwell positions within the same applicator channel; The total residence time regularization term It is used to suppress unnecessary total irradiation time while meeting clinical goals.

[0010] Furthermore, the rule feedback controller adopts a discrete clinical rule feedback method, specifically as follows: Calculate the degree of normalization violation of target coverage, bladder dose, and rectal dose relative to the minimum clinical standard, respectively; Compare the degree of violation of each normalization and determine the dominant clinical objective with the greatest degree of violation; Perform a multiplicative update on the target weights corresponding to the dominant clinical target, multiplying the target weights by an update coefficient greater than 1, while keeping the weights of the remaining target items unchanged. The minimum clinical criteria and preferred program objectives constitute two levels of clinical thresholds; When the degree of violation of all clinical objectives is zero, or when the preset maximum number of feedback rounds is reached, the minimum clinical criteria are determined to be met and the optimization termination condition is triggered. The preferred plan target level is used to determine the coverage level constraint during candidate plan screening.

[0011] Furthermore, the rule feedback controller adopts a continuous adaptive feedback method, specifically as follows: For each clinical target—target coverage, bladder dose, and rectal dose—candidate weights are calculated using a monotonically increasing function based on the degree of violation. in, As a monotonically increasing function, the greater the degree of violation, the higher the candidate weights generated. To update the round number The target weight; The target weights are updated using an exponential smoothing formula that drives the coefficients. The exponential smoothing formula is as follows: in, is the momentum coefficient, and its value ranges from greater than or equal to 0 to less than 1; In each round of feedback, the target weights for target coverage, bladder dose, and rectal dose can all change simultaneously.

[0012] Furthermore, the method also includes: In the optimization trajectory, candidate plans are saved according to iteration intervals or clinical indicator improvement events. Each candidate plan record includes a residence time vector, values ​​of each clinical target indicator, and weights of each target. Multiple candidate plans are simultaneously output from the candidate plan record according to the following rules: The target area priority candidate plan selects the candidate plan with the largest target area coverage volume ratio; The bladder protection candidate plan is selected from the candidate plans that meet the target coverage level constraints and have the lowest dose to the highest irradiated volume of the bladder. The candidate program for rectal protection is selected from those that meet the target coverage level constraints, and the one with the lowest maximum irradiated volume dose in the rectum is chosen.

[0013] Furthermore, before outputting the stay time plan, a deliverability check is performed on the stay time plan, the deliverability check including: Verify that the dwell time of all dwell locations is non-negative, the dwell time of a single dwell location does not exceed the set upper limit, the total dwell time does not exceed the device's allowed upper limit, and the dwell location index is consistent with the applicator channel definition; The dwell time plan is output in DICOM RTPLAN format or a dwell time file format compatible with afterloading therapy devices.

[0014] The present invention also provides a system for implementing the above-described executable residence time schedule generation method, the system comprising: The dwell time parameterization module is used to fix the dwell position of the radioactive source and convert the unconstrained internal variables corresponding to each dwell position into non-negative dwell time through a differentiable non-negative mapping. The differential dose calculation module is connected to the residence time parameterization module and is used to perform forward dose accumulation on each calculation point according to the line source dose calculation model to obtain a three-dimensional dose distribution, and to establish the gradient relationship between the three-dimensional dose distribution and the non-negative residence time. The clinical target calculation module is connected to the differential dose calculation module and is used to calculate the target coverage loss, organ-at-risk dose over-limit loss and residence time regularization term according to the three-dimensional dose distribution, and to obtain the total loss value by weighting and summing according to the weight of each target. The clinical rule feedback module is connected to the clinical goal calculation module and is used to automatically update the weight of each clinical goal according to the degree of violation of each clinical goal relative to the clinical threshold, and output the updated goal weight to the clinical goal calculation module. The gradient optimization module is data-connected to the differential dose calculation module and the clinical target calculation module, and is used to receive the residence time gradient and perform iterative updates on the unconstrained internal variables according to the total loss value. The candidate plan management module is connected to the gradient optimization module and is used to save multiple candidate plans in the optimization trajectory and filter and output them from the candidate plans according to the clinical priority rules of target area priority, bladder protection and rectal protection. The plan output module is connected to the candidate plan management module for performing deliverability checks on the screened residence time plans and outputting residence time plan files that can be directly executed by the afterloading therapy device.

[0015] The executable residence time plan generation method proposed in this invention can be applied to the adaptive inverse planning design of high dose rate brachytherapy for cervical cancer, and can also be extended to brachytherapy scenarios such as prostate, breast, skin, head and neck, or other brachytherapy scenarios using fixed applicator / catheter residence positions.

[0016] By employing the above technical solution, the present invention provides an executable method, system, and application for generating residence time plans, which has at least the following beneficial effects: 1. This invention realizes the TG-43 line source dose calculation as a differentiable operator that can participate in automatic differentiation, and uses the linear relationship between dose and residence time to analytically calculate the gradient, so that the error of the clinical objective function can be directly backpropagated to the residence time variable, fundamentally eliminating the computational bottleneck of relying on black box trial or finite difference in traditional schemes; combined with the forward dose calculation mechanism of the graphics processor with the calculation point as the parallel task unit, the automatic optimization of a single case can be completed in seconds, meeting the time requirements of the same-day adaptive planning of image-guided brachytherapy for cervical cancer.

[0017] 2. This invention automatically updates target weights based on the real-time violation of target coverage, bladder dose, and rectal dose relative to clinical thresholds by setting a rule feedback controller, replacing the physicist's repeated manual adjustment of weights for each patient; in the validation of high-dose-rate loading plans for 299 cervical cancer patients, the discrete feedback mode increased the simultaneous pass rate of all minimum clinical standards from 10.4% in manual clinical planning to 50.5%, significantly improving planning consistency and clinical target achievement rate.

[0018] 3. This invention sets unconstrained internal variables for each dwell position and converts them into non-negative dwell time through differentiable non-negative mapping. This ensures that the dwell time always meets the physical executable constraints of the afterloading device throughout the optimization process. No additional constraint projection or dwell time inverse calculation steps are required. It directly outputs the dwell time plan that the afterloading treatment device can execute, eliminating the conversion error between the predicted dose and the final deliverable dose in the dose prediction scheme.

[0019] 4. This invention introduces an exponential smoothing weight update formula for the driving force coefficient through a continuous adaptive feedback controller, which allows the weights of multiple clinical targets to be adjusted simultaneously in each round of feedback. This effectively suppresses abrupt changes and oscillations in target weights among competing targets, and enables the optimized trajectory to move continuously along the trade-off region between target coverage and organ protection. The continuous adaptive strategy can generate candidate plans with different priorities in about 21 seconds.

[0020] 5. This invention saves a candidate plan library in the optimized trajectory and automatically filters and outputs multiple candidate plans according to clinical priority rules such as target area priority, bladder protection and rectal protection. This enables physicians to make rapid and interpretable clinical choices between target area coverage and organ protection based on the patient's anatomical characteristics, previous radiation dose and organ status, and solves the technical problem that a single automatic plan cannot fully express diverse clinical preferences. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart for an executable method for generating a residency time plan; Figure 2 Representative validation cases for comparing TG-43-CU and TPS dose calculations; Figure 3 HR-CTV V100, rectal D2cc, and bladder D2cc distribution for TPS plan and AutoBTPlan generation plan; Figure 4 Representative dose distributions of clinical manual plans and AutoBTPlan generated plans in 5 anonymous clinical cases; Figure 5 The distribution of TPS and AutoBTPlan generation plans in the target-organ-at-risk tradeoff space. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0023] Example 1: This embodiment provides an executable residence time plan generation method, applicable to adaptive inverse plan design for high-dose-rate afterloading therapy for cervical cancer. For example... Figure 1 As shown, this method takes fixed dwell position and optimizable dwell time as its core, and connects microdose calculation, clinical target evaluation, rule feedback weight update and gradient optimizer into a closed loop to directly generate a dwell time plan that can be executed by afterloading therapy device.

[0024] Step S1: Obtain treatment plan input data.

[0025] Read the patient's three-dimensional medical images, the three-dimensional contours of the high-risk clinical target area HR-CTV, the three-dimensional contours of the bladder and rectum, the three-dimensional spatial coordinates of each applicator channel, the spatial coordinates and local source axis direction of each candidate residence position, the radiation source type and corresponding TG-43 parameters, the prescribed dose for this treatment, and the clinical dose targets and dose limits for each structure from the treatment planning system or independent data interface.

[0026] This embodiment unifies the contours of each structure into a three-dimensional dose grid by voxelizing them, generating three-dimensional voxel masks for HR-CTV, bladder, and rectum. The voxel masks are used to extract the dose values ​​of the corresponding voxels from the complete three-dimensional dose grid according to the structural index, for subsequent clinical target calculation.

[0027] In this embodiment, the single prescription dose is set to 6 Gy. The minimum acceptable coverage standard for HR-CTV is set at V100 not less than 85%, and the preferred planning target is set at V100 not less than 90%. The minimum acceptable limit for D2cc in the bladder and rectum is set at 80% of the prescription dose, and the preferred limit is set at 70% of the prescription dose.

[0028] Step S2: Initialize the residence time optimization variable. Fix the residence locations of all radioactive sources, and only use the residence time at each residence location as the variable to be optimized. For the... Each residence location is assigned an unconstrained internal variable. and through softplus The actual dwell time obtained by nonnegative mapping is: in, For the first The length of stay at each designated location.

[0029] Through the above mapping, the dwell time obtained during the optimization process is always not less than zero, thereby avoiding the generation of physically unexecutable negative dwell times. Each internal variable can be initialized to the same value, or it can be initialized according to the donor geometry, target area envelopment relationship, or a preset dwell time distribution.

[0030] Step S3: Calculate the three-dimensional dose using the differentiable dose calculation module. Input the dwell position coordinates, local source axis direction, dwell time, and the TG-43 radiation source parameter table into the differentiable dose calculation module. For any dose calculation point... The dose contribution per unit time at each residence location is calculated and accumulated, i.e.: in, The total dose contributed by the cumulative dose per unit time; Number of locations to be occupied; For the first Each stationary location calculates the points within a unit of stationary time. The dose contribution produced.

[0031] Dose contribution per unit time The dose calculation model based on the TG-43 line source includes air kerma intensity, dose rate constant, line source geometry function, radial dose function, and two-dimensional anisotropy function. The expression is as follows: in, The intensity of air kerma; It is the dose rate constant; For line source geometric functions; Radial dose function; It is a two-dimensional anisotropic function; , For reference purposes, in this embodiment... , ; , The source coordinates are .

[0032] In this embodiment, the input tensor is first established: for each computation point Save 3D coordinates For each stationing location Save coordinates Source axis and length of stay Load source-specific TG-43 lookup table and physical constants.

[0033] Then let axial component radial component Thus we obtain and .for For values ​​close to zero, a preset lower limit or source truncation is used to avoid singular values.

[0034] The radial dose function is obtained through one-dimensional interpolation, and the two-dimensional anisotropy function is obtained through two-dimensional interpolation in the distance and angle directions. When the query point exceeds the range of the parameter table, boundary truncation or preset boundary value processing is adopted.

[0035] Forward computation uses computation points as parallel task units. Each computation point independently traverses all dwell positions and accumulates the dose contribution, making full use of the parallel computing capabilities of the graphics processor. When the number of computation points is large, the computation points are input in blocks to control storage usage.

[0036] The differentiable dose calculation module is constructed as part of the automatic differential calculation graph, and the dose calculation results maintain a gradient correlation with the residence time variable. The backpropagation uses the analytical derivative relationship between dose and residence time to calculate the residence time gradient, which is: No. The dose at the calculation point for the first calculation point The partial derivative of the dwell time at the first dwelling location is equal to the first dwelling location. The first residence location for the first The dose contribution value per unit time at each calculation point.

[0037] Specifically, due to dose contribution Regarding the length of stay The linear, residence-time gradient can be expressed analytically as follows: The differential dose computation operator receives the upstream gradient. And calculate: In the embodiment with a fixed dwell position, only the dwell time gradient is returned; in the extended embodiment, the derivatives of coordinate transformation, radial interpolation, and anisotropic interpolation can be calculated analytically, and the line source geometry function can be derived using analytical derivatives or scalar finite differences, thereby further optimizing the dwell position or source axis.

[0038] Step S4: Calculate the target dose. Extract the dose values ​​of all voxels within the HR-CTV from the three-dimensional dose distribution, and calculate the coverage index of the HR-CTV based on dose value sorting or a differentiable dose volume approximation method. When the target coverage volume ratio of the HR-CTV is set to... At that time, calculate the corresponding dose quantile value. Target coverage loss The underdose squared penalty is adopted, that is: in, This is the prescription dosage.

[0039] In this embodiment, when the target coverage volume ratio within the target area is... Corresponding dose quantile Reaching or exceeding the prescribed dosage At that time, target coverage loss The value is zero; when the target coverage volume ratio within the target area is zero. Corresponding dose quantile Less than the prescription dose At that time, target coverage loss It increases with increasing dose gap.

[0040] Step S5: Calculate the dose target for organs at risk. Extract the dose values ​​of all voxels in the bladder and rectum from the three-dimensional dose distribution, and calculate the doses for the corresponding structures. Or its differentiable approximation, This indicates the highest dose to a certain organ at risk is 2cm. 3 Volume-specific dose. For organs at risk. Its dosage over-limit loss The overlimit square penalty is adopted, namely: in, To endanger organs The highest 2cm 3 dose; For corresponding dosage restrictions.

[0041] In this embodiment, the dose at which the organ at risk of exposure is the highest irradiated volume. Reaching or exceeding the corresponding dose limit A penalty is incurred when the dose exceeds the limit, and the penalty value is the square of the excess dose; no penalty is incurred when the maximum irradiated volume dose to the organ at risk is below the limit.

[0042] When precise dosage cannot be directly determined In the implementation of differentiation, the high-dose voxel mean, tail conditional mean, or differentiable quantile can be used as approximations of the dose. Differentiable agent target.

[0043] Step S6: Calculate the residence time regularization term. To avoid unnecessary drastic changes in residence time between adjacent residence positions within the same applicator channel, a smoothing term for adjacent residence points is set. ,Right now: in, This represents the set of adjacent dwelling positions within the same channel.

[0044] You can also set a regular expression for total dwell time. It is used to suppress unnecessary total irradiation time while meeting clinical goals.

[0045] Step S7: Construct the initial optimization objective function. Construct the total loss function based on the target area, bladder, rectum, and residence time regularization terms. ,Right now: in, Loss of target coverage; Bladder dose loss; This is due to rectal dose loss; Target coverage weight; Bladder dose weighting; Rectal dose weighting; Weights for the dwell time smoothing term; The weight of the regularization term for total dwell time.

[0046] At the start of optimization, the target weights for the target area, bladder, and rectum can be set to the same initial value, or different initial values ​​can be set according to clinical priorities.

[0047] Step S8: Perform gradient descent optimization. Use the gradient descent optimizer to optimize the unconstrained internal variables. Perform iterative updates. In each iteration, execute the following steps in sequence: based on the current internal variables... softplus The function mapping calculates the nonnegative residence time; the differentiable dose calculation module is called to calculate the three-dimensional dose distribution; dose indices for HR-CTV, bladder, and rectum are extracted; the total loss function value is calculated; the total loss is backpropagated across the residence time variable, and the gradient is analytically calculated using the linear relationship between dose and residence time, i.e.: in, For the first The first residence location for the first The dose contribution value per unit time at each calculation point; the internal variables are updated according to the gradient; and a validity check is performed on the updated residence time.

[0048] After a preset number of internal gradient updates, the transient treatment plan for the current stage is obtained.

[0049] Step S9: Perform discrete clinical rule feedback. Calculate the normalized violation degree of each clinical goal relative to the minimum clinical standard based on the current transient treatment plan, i.e.: in, , and These represent the degree of normalization violation for target coverage, bladder dose, and rectal dose, respectively. The minimum acceptable target area coverage ratio; This represents the planned target area coverage volume ratio. , The maximum 2cm for bladder and rectal doses, respectively. 3 dose; , These are the corresponding dose limits for bladder and rectal doses, respectively.

[0050] System Comparison , and Identify the dominant clinical objective item with the greatest degree of violation. The target weights corresponding to the dominant clinical objectives. Perform a multiplicative update while keeping the weights of the remaining target terms unchanged, i.e.: in, To update the round number The target weight; To update the coefficients.

[0051] After the weights are updated, the next round of residence time gradient optimization is performed, thus forming a closed-loop feedback process of dose calculation, clinical evaluation, weight adjustment, and residence time update.

[0052] Step S10: Determine the termination condition for optimization. When the normalization violation of all clinical objectives is zero, that is, when the target coverage, bladder dose, and rectal dose all meet the minimum clinical standards, the termination condition is determined to be met; or when the preset maximum number of feedback rounds is reached, termination is also triggered; optimization can also be terminated when the maximum number of gradient iterations is reached or when the change of the loss function is less than the preset threshold for several consecutive rounds.

[0053] Minimum clinical criteria and preferred program objectives constitute two levels of clinical thresholds. The preferred program objective level is used to determine the coverage level constraints when screening candidate programs.

[0054] Step S11: Generate and output an executable plan. Perform a deliverability safety check on the final dwell time plan, including: verifying that the dwell time at all dwell locations is non-negative, the dwell time at a single dwell location does not exceed the set upper limit, the total dwell time does not exceed the device's allowed upper limit, and the dwell location index is consistent with the applicator channel definition.

[0055] After the check is passed, the plan is output in DICOM RTPLAN format or a residence time file format compatible with the afterloading therapy device, for direct execution by the afterloading therapy device. The output may also include three-dimensional dose distribution, HR-CTV, DVH of the bladder and rectum, HR-CTV-V100, D2cc of the bladder and rectum, total residence time, residence time smoothness index, plan data file that can be read by the treatment planning system or afterloading therapy device, dose-volume histogram index of organs at risk, and optimization audit log, thereby obtaining an automatically optimized plan that can be directly executed by the afterloading therapy device.

[0056] This embodiment generates an afterloading therapy dwell time sequence by binding the final dwell time to the corresponding dwell location. The final dwell time is adapted to the device's precision, for example, by rounding it according to the time resolution allowed by the afterloading therapy machine. After rounding, the dose is recalculated and clinical indicators are evaluated. When rounding causes clinical indicators to exceed the allowable range, local fine-tuning can be performed on the dwell time.

[0057] The specific clinical thresholds, weight update coefficients, number of optimization rounds, learning rate, and termination conditions in each of the above steps can be adjusted according to the treatment site, prescription regimen, clinical guidelines, and equipment performance, without affecting the basic technical concept of this method to automatically optimize residence time through microdose calculation and rule feedback.

[0058] Example 2: The data acquisition, residence time initialization, dose calculation, clinical target calculation, and residence time gradient update methods in this embodiment are the same as in Embodiment 1. The difference is that the clinical target weight is updated using a continuous adaptive momentum smoothing method, and multiple candidate plans are saved in the optimization trajectory and output according to clinical priority.

[0059] For each target item in the target area, bladder, and rectum, candidate weights are calculated based on their clinical violation severity. ,Right now: in, It is a monotonically increasing function; the greater the degree of violation, the higher the candidate weights generated.

[0060] To avoid optimization trajectory oscillations caused by frequent switching between different clinical objectives, an exponential smoothing formula with momentum coefficients is used to update the weights of each objective, i.e.: in, is the momentum coefficient, and its value ranges from 0 to less than 1.

[0061] In each feedback round, the target weights of target coverage, bladder dose, and rectal dose can change simultaneously. The residence time optimization process can continuously move along the trade-off region between target coverage and protection of organs at risk, effectively suppressing abrupt changes and oscillations in target weights among competing targets.

[0062] During the optimization process, candidate plans are saved according to iteration intervals or clinical indicator improvement events. Each candidate plan record includes a residence time vector, HR-CTV-V100, bladder D2cc, rectal D2cc, total residence time, weights of each target, and timestamps. Multiple candidate plans are then selected from the saved candidate plan records and output simultaneously according to the following rules: Target volume priority candidate plan: Select the plan with the largest HR-CTV V100 from the candidate set, that is, the candidate plan with the largest target volume coverage ratio; when V100 is the same or the difference is less than the preset tolerance, the lower rectal D2cc, bladder D2cc or total residence time are used as secondary sorting conditions in turn.

[0063] Bladder protection candidate plans: First, select a set of candidate plans that meet the preferred coverage target; if the set is empty, select a set that meets the minimum coverage criterion; then select the candidate plan with the lowest bladder D2cc from the selected set. That is, select the candidate plan with the smallest maximum bladder irradiated volume dose from the candidate plans that meet the target coverage level constraints.

[0064] Rectal protection candidate plans: Using the same coverage level screening rules as bladder protection candidate plans, the candidate plan with the lowest rectal D2cc is selected from the candidate set that meets the coverage level constraints. That is, the candidate plan with the smallest maximum irradiated volume dose in the rectum is selected from the candidate plans that meet the target area coverage level constraints.

[0065] The three candidate plans mentioned above are output simultaneously, allowing physicians or physicists to select the final treatment plan based on the patient's anatomical characteristics, previous radiation dose, cumulative dose from fractionated treatments, and clinical priority. The continuous adaptive feedback strategy can generate candidate plans with different priorities within approximately 21 seconds, supporting rapid clinical decision-making in same-day adaptive treatment workflows. In other implementations, candidate plans may also include the plan with the shortest total treatment time, the plan with the lowest target hotspot, the plan with the smoothest dwell time, or the plan with the highest overall score.

[0066] Example 3: This embodiment provides an executable residence time plan generation system. The system maps each processing step of the method described in Embodiments 1 and 2 to a functional module with a clear data interface, thus forming a complete closed-loop automatic optimization system.

[0067] The data acquisition module is used to read patient images, structural contours, applicator geometry, residence position coordinates and local source axis direction, TG-43 parameter table, prescription dose and clinical dose target from the treatment planning system or independent data interface, organize the above data into the data format required by subsequent modules and pass it down.

[0068] The structure processing module is connected to the data acquisition module, receives structural contour data, performs voxelization processing on HR-CTV, bladder and rectum to generate a three-dimensional voxel mask, and outputs it to the differential dose calculation module and the clinical target calculation module for extracting dose values ​​by structural index.

[0069] The residence time parameterization module receives residence location information, sets unconstrained internal variables for each residence location, converts them into non-negative residence time through a differentiable non-negative mapping, and outputs the residence time tensor to the differentiable dose calculation module.

[0070] The microdose calculation module is connected to the residence time parameterization module. It receives residence position coordinates, source axis direction, non-negative residence time and TG-43 parameter table. Based on the TG-43 line source model, it performs forward dose accumulation in parallel to obtain the three-dimensional dose distribution and establishes the gradient relationship between the three-dimensional dose distribution and the non-negative residence time. The three-dimensional dose tensor is output to the clinical target calculation module, and the residence time gradient is output to the gradient optimization module.

[0071] The clinical target calculation module is connected to the differential dose calculation module. It receives the three-dimensional dose distribution, extracts the dose values ​​of each structure voxel, calculates the target coverage loss, the dose over-limit loss of organs at risk, and the residence time regularization term. It then calculates the total loss value by weighting and summing according to the weight of each target and outputs it to the gradient optimization module and the clinical rule feedback module.

[0072] The clinical rule feedback module is connected to the clinical goal calculation module. It receives the values ​​of each clinical goal indicator, calculates the degree of violation of each goal relative to the clinical threshold, automatically updates the weight of each goal in a discrete feedback or continuous adaptive feedback manner, and outputs the updated goal weights to the clinical goal calculation module to drive the next round of total loss value calculation.

[0073] The gradient optimization module is connected to the differentiable dose calculation module and the clinical goal calculation module. It receives the residence time gradient and the total loss value, performs iterative updates on the unconstrained internal variables based on the gradient, and outputs the updated internal variables to the residence time parameterization module, completing one iteration cycle. The gradient optimization module can use stochastic gradient descent, Adam, AdamW, L-BFGS, or other optimization algorithms that can update variables based on gradients.

[0074] The candidate plan management module is connected to the gradient optimization module. Multiple candidate plans are saved in the optimization trajectory according to the iteration interval or the improvement event of clinical indicators. The plan is then selected and output according to the clinical priority rules of target area priority, bladder protection and rectal protection. The selection results are then passed to the plan output module.

[0075] The plan output module is connected to the candidate plan management module to perform a deliverability check on the filtered dwell time plan, verifying the non-negativity of dwell time, the upper limit of single point, the upper limit of total time, and the consistency of dwell location index. After the check passes, it outputs a dwell time file in DICOM RTPLAN format or compatible with afterloading therapy equipment for direct execution by afterloading therapy equipment.

[0076] The modules in the above system are interconnected through data and feedback flows, forming a complete closed-loop automatic optimization system that can be deployed on independent computing workstations, hospital servers, treatment planning system plugins, or remote computing nodes. The microdose calculation module can be executed by CPU, GPU, dedicated accelerator, or a combination thereof to fully utilize parallel computing capabilities to achieve second-level single-case optimization.

[0077] It is understood that the nonnegative mapping of this invention is not limited to the softplus function, but can also employ exponential mapping. Square mapping Projection truncation or parameterized mapping with upper and lower bounds can all guarantee non-negative dwell time and achieve the same physical executable constraint effect.

[0078] It is understandable that the differentiable D2cc proxy target is not limited to the exact sorting method. Differentiable approximation methods such as high-dose voxel mean, tail conditional mean, SoftSort, or high-dose voxel weighted sum can also be used, which are also applicable in implementation scenarios where the exact D2cc cannot be directly differentiated.

[0079] It is understood that this invention is not limited to high-dose-rate brachytherapy for cervical cancer, but can also be extended to brachytherapy for the prostate, breast, skin, head and neck, or other sites where a fixed applicator or catheter is used. Only the corresponding clinical threshold, prescription dose, and dose limit for organs at risk need to be adjusted according to the treatment site. This does not affect the basic technical concept of this invention, which achieves automatic optimization of residence time through microdose calculation and rule-guided closed-loop feedback.

[0080] Verification example: 1. Validation of TG-43-CU dose calculation Table 1 summarizes the consistency of TG-43-CU and TPS dose calculations when using clinical residence location and residence time. Compared with TPS, the mean absolute errors (MAE) of dose at HR-CTV, rectum, and bladder were 0.698±0.341 Gy, 0.017±0.007 Gy, and 0.007±0.002 Gy, respectively.

[0081] Table 1. Consistency of TG-43-CU and TPS dose calculation results in clinical planning Figure 2 Representative validation cases comparing TG-43-CU and TPS dose calculations are provided. Each row in the figure represents one clinical case, sequentially showing the TPS dose, the calculated TG-43-CU dose, the dose difference map, the horizontal dose profile, and the vertical dose profile. The dose map shows the high-dose region with doses not less than 1 Gy. The black dashed outline represents the 6 Gy prescription isodose line. The dose map overlay shows the structural outlines, with HR-CTV in red, rectum in cyan, and bladder in orange. In the dose profile diagram, the TPS dose is represented by a solid line, and the TG-43-CU dose is represented by a dotted line. Figure 2 It can be seen that the TG-43-CU can reproduce the spatial dose distribution of the prescription dose area well, and the 6 Gy isodose lines of the two calculation results are in high agreement. In most of the high-dose areas assessed, the dose profiles of TPS and TG-43-CU almost overlap, with the main difference occurring near the location of the radiation source, where the peak doses of the two calculation results differ. Apart from these local source-near areas, the dose distributions of the two are generally well consistent.

[0082] 2. Evaluation of AutoBTPlan The single-case optimization time for the discrete clinical rule feedback strategy was 3.60±2.63s, while the single-case optimization time for the continuous adaptive feedback strategy was 20.77±4.21s.

[0083] Table 2 summarizes the dosimetric performance of the TPS plan and the AutoBTPlan generation plan. The HR-CTVV100 of the clinical artificial plan was 0.87±0.08, and the rectal and bladder D2cc were 4.90±0.93 Gy and 5.28±1.05 Gy, respectively.

[0084] The discrete clinical rule feedback strategy generated plans with lower organ-at-risk (OAT) doses compared to the TPS plan, specifically reducing rectal D2cc to 4.61 ± 0.57 Gy and bladder D2cc to 4.70 ± 0.50 Gy. However, this reduction in OAT dose was accompanied by a decrease in HR-CTV coverage, with V100 decreasing to 0.83 ± 0.05. These results indicate that the discrete clinical rule feedback strategy tends to rapidly meet minimum clinical criteria, but its target coverage is relatively conservative. The continuous adaptive feedback strategy, while maintaining HR-CTV coverage close to the TPS plan, can generate candidate plans with different clinical priorities and clinical interpretability. The target-priority candidate plan achieved the highest HR-CTV V100 (0.88 ± 0.07), higher than the TPS plan, but increased rectal D2cc to 5.23 ± 0.74 Gy. The bladder-protection candidate plan reduced bladder D2cc to 4.90 ± 0.65 Gy while maintaining an HR-CTV V100 of 0.86 ± 0.05. Similarly, the rectal protection candidate plan reduced rectal D2cc to 4.84±0.70 Gy, with an HR-CTVV100 of 0.86±0.05. Compared to discrete clinical rule-based feedback strategies, continuous adaptive feedback strategies better maintain target coverage and allow for plan selection based on clinical priorities of the target area, bladder, or rectum.

[0085] Table 2 Dosimetric Comparison of TPS Plan and AutoBTPlan Generation Plan Figure 3 This study further illustrates the distribution characteristics of target coverage and organ-at-risk dose under different AutoBTPlan strategies in the patient cohort. Boxes represent interquartile ranges, orange lines represent the median, black diamonds represent the mean, and semi-transparent scatter dots represent individual patient cases. Horizontal reference lines represent the minimum clinical criteria (dashed lines) and preferred planning targets (dotted lines). For HR-CTV, the corresponding V100 reference values ​​are 0.85 and 0.90, respectively; for rectum and bladder, the corresponding D2cc reference values ​​are 4.8 Gy and 4.2 Gy, respectively.

[0086] Compared to the TPS plan, the discrete clinical rule feedback strategy showed a significantly narrower distribution across all three clinical objectives: the interquartile range of HR-CTVV100 decreased from 0.125 to 0.030, the interquartile range of rectal D2cc decreased from 1.01 Gy to 0.39 Gy, and the interquartile range of bladder D2cc decreased from 1.08 Gy to 0.42 Gy. This narrowing of the distribution range indicates that the discrete clinical rule feedback strategy generates a more consistent plan, but HR-CTVV100 is also more concentrated near the minimum acceptable coverage criteria.

[0087] Continuous adaptive feedback strategies exhibited a broader and more clinically interpretable distribution pattern. Target-priority candidate plans shifted HR-CTVV100 to higher coverage levels; bladder-protection and rectal-protection candidate plans, on the other hand, reduced D2cc of the corresponding organs at risk while maintaining target coverage near acceptable levels. As expected, target-priority candidate plans were typically accompanied by increased doses to organs at risk, reflecting the competition between target coverage and organ-at-risk protection.

[0088] Figure 4 Representative spatial dose distributions from both manual clinical planning and AutoBTPlan-generated plans are presented. Each row in the figure represents one case, and the columns are, in order, the TPS plan, the discrete clinical rule feedback plan, the target area priority candidate plan, the bladder protection candidate plan, and the rectal protection candidate plan. The dose map shows the high-dose region with doses not less than 1 Gy. The black dashed outline represents the 6 Gy prescription isodose line. The dose map overlay shows the structural outlines, with HR-CTV in red, rectum in cyan, and bladder in orange.

[0089] These cases illustrate the anatomical dose distribution characteristics corresponding to DVH results. Target-priority candidate plans typically expand the 6 Gy prescription isodose line around the HR-CTV, thereby improving target coverage, but at the same time extending the dose range more towards adjacent organs at risk. Conversely, bladder-protection and rectal-protection candidate plans reduce the dose extension towards the corresponding priority organs at risk; discrete clinical rule feedback plans result in a more conservative dose distribution with lower doses to organs at risk, but also correspondingly reduced target coverage.

[0090] 3. Plan acceptability and clinical standard pass rate Table 3 summarizes the pass rates for the minimum clinical criteria for each protocol. The discrete clinical rule feedback strategy achieved the highest simultaneous pass rate for all minimum clinical criteria, increasing the proportion of plans meeting all minimum clinical criteria from 10.4% for manual clinical plans to 50.5%. This improvement was primarily driven by enhanced organ-at-risk protection capabilities. The pass rate for rectal D2cc increased from 46.2% to 62.9%, and the pass rate for bladder D2cc increased from 32.1% to 61.2%.

[0091] Target-priority candidate plans improved HR-CTV coverage, with 77.4% of cases meeting the minimum V100 criteria and 41.1% achieving the preferred V100 target. However, the simultaneous pass rate for all minimum clinical criteria for these candidate plans was only 5.1%, primarily reflecting more frequent organ-at-risk dose exceedances. Bladder-protection and rectal-protection candidate plans partially restored the pass rate for organ-at-risk dose limitations, with simultaneous pass rates for all minimum clinical criteria of 27.7% and 26.7%, respectively. These results indicate that discrete clinical rule feedback strategies are most effective in meeting minimum clinical criteria, while continuous adaptive feedback strategies can provide candidate plans with different clinical priorities and clinical interpretability.

[0092] Table 3. Pass Rate of Minimum Clinical Standards and Optimized Program Goals Figure 5 The distribution of the TPS plan and AutoBTPlan generation plan in the target-organ-at-risk tradeoff space is shown. Each point in the figure represents a clinical case, with HR-CTV V100 on the x-axis and rectal or bladder D2cc on the y-axis. Reference lines represent the minimum acceptable HR-CTV coverage and the dose limit for organ-at-risk D2cc, respectively.

[0093] The discrete clinical rule feedback strategy shifts many cases towards lower rectal and bladder D2cc regions, but simultaneously clusters near the lowest target coverage boundary, consistent with its emphasis on meeting minimum clinical criteria. In contrast, target-priority candidate plans shift towards higher HR-CTV V100 regions, while bladder-protective and rectal-protective candidate plans are distributed in the intermediate regions, improving protection of the corresponding organs at risk. This distribution illustrates that the continuous adaptive feedback strategy can generate candidate plans with diverse clinical characteristics, rather than a single dominant approach.

[0094] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0096] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for generating an executable residence time plan, characterized in that, The method includes the following steps: The residence position of the fixed radioactive source is set, and an unconstrained internal variable is set for each residence position. The unconstrained internal variable is converted into a non-negative residence time through a differentiable non-negative mapping. The coordinates of the dwell position, the source axis direction, the non-negative dwell time, and the radiation source dose calculation parameters are input into the differentiable dose calculation operator. The differentiable dose calculation operator performs forward dose accumulation on each calculation point according to the line source dose calculation model to obtain the three-dimensional dose distribution, and analyzes the dwell time gradient using the linear relationship between dose and dwell time during reverse propagation. Based on the three-dimensional dose distribution, a differentiable clinical objective function is constructed, which includes a target coverage loss term, an organ at risk dose excess loss term, and a residence time regularization term. The degree of violation of each clinical objective relative to the clinical threshold is calculated based on the current three-dimensional dose distribution, and the degree of violation is input into the rule feedback controller to automatically update the target weight corresponding to each clinical objective; The gradient optimizer is driven by the updated target weights, and the unconstrained internal variables are iteratively updated using the residence time gradient, forming a closed-loop feedback process of dose calculation, clinical evaluation, weight update and residence time update. Once the termination conditions are met, the system outputs a stay time plan that can be directly executed by the afterloading therapy device.

2. The executable residence time plan generation method according to claim 1, characterized in that, The differentiable nonnegativity mapping adopts softplus The function, the softplus The function expression is: in, For the first The unconstrained internal variables corresponding to each residence location; The non-negative dwell time after conversion; Alternatively, the differentiable nonnegative mapping may be any one of an exponential mapping, a square mapping, or a parameterized mapping with upper and lower bounds.

3. The executable residence time plan generation method according to claim 1, characterized in that, The forward computation of the differentiable dose calculation operator is performed in parallel, with each calculation point as a parallel task unit. For each calculation point, all dwell positions are traversed and the dose contribution of each dwell position is accumulated to obtain the total dose at that calculation point, expressed as: in, The total dose contributed by the cumulative dose per unit time; Number of locations to be occupied; For the first Each stationary location calculates the points within a unit of stationary time. The dose contribution generated; When the number of computation points is large, the computation points are input in blocks to control storage usage; The backpropagation process calculates the residence time gradient using the analytical derivative relationship between dose and residence time. This analytical derivative relationship is as follows: No. The dose at the calculation point for the first... The partial derivative of the dwell time at the first dwelling location is equal to the first dwelling location. The first residence location for the first The dose contribution value per unit time at each calculation point.

4. The executable residence time plan generation method according to claim 1, characterized in that, The target coverage loss term adopts an underdose squared penalty form. A penalty is incurred when the dose quantile corresponding to the target coverage volume ratio within the target area is lower than the prescribed dose. The penalty value is the square of the difference between the prescribed dose and the dose quantile, i.e.: in, Loss of target coverage; This is the prescription dosage; These are dose percentile values; The penalty for exceeding the dose limit for organs at risk adopts the form of a squared penalty for exceeding the limit. A penalty is generated when the highest irradiated volume dose of the organ at risk exceeds the corresponding dose limit. The penalty value is the square of the excess amount, that is: in, Indicates organs at risk Dose over-limit loss; To endanger organs The highest 2cm 3 dose; For corresponding dosage restrictions; The dwell time regularization term includes the total dwell time regularization term. Smoothing terms with adjacent residence points The adjacent dwell point smoothing term sums the squares of the dwell time differences between adjacent dwell positions within the same applicator channel, i.e.: in, This refers to the set of adjacent dwelling positions within the same channel; and The dwell time of adjacent dwell positions within the same applicator channel; The total residence time regularization term It is used to suppress unnecessary total irradiation time while meeting clinical goals.

5. The executable residence time plan generation method according to claim 1, characterized in that, The rule feedback controller adopts a discrete clinical rule feedback method, specifically: Calculate the degree of normalization violation of target coverage, bladder dose, and rectal dose relative to the minimum clinical standard, respectively; Compare the degree of violation of each normalization and determine the dominant clinical objective with the greatest degree of violation; Perform a multiplicative update on the target weights corresponding to the dominant clinical target, multiplying the target weights by an update coefficient greater than 1, while keeping the weights of the remaining target items unchanged. The minimum clinical criteria and preferred program objectives constitute two levels of clinical thresholds; When the degree of violation of all clinical objectives is zero, or when the preset maximum number of feedback rounds is reached, the minimum clinical criteria are determined to be met and the optimization termination condition is triggered. The preferred plan target level is used to determine the coverage level constraint during candidate plan screening.

6. The executable residence time plan generation method according to claim 1, characterized in that, The rule feedback controller adopts a continuous adaptive feedback method, specifically: For each clinical target—target coverage, bladder dose, and rectal dose—candidate weights are calculated using a monotonically increasing function based on the degree of violation. ,Right now: in, As a monotonically increasing function, the greater the degree of violation, the higher the candidate weights generated. To update the round number The target weight; The target weights are updated using an exponential smoothing formula that drives the coefficients. The exponential smoothing formula is as follows: in, is the momentum coefficient, and its value ranges from greater than or equal to 0 to less than 1; In each round of feedback, the target weights for target coverage, bladder dose, and rectal dose can all change simultaneously.

7. The executable residence time plan generation method according to claim 1, characterized in that, Also includes: In the optimization trajectory, candidate plans are saved according to iteration intervals or clinical indicator improvement events. Each candidate plan record includes a residence time vector, values ​​of each clinical target indicator, and weights of each target. Multiple candidate plans are simultaneously output from the candidate plan record according to the following rules: The target area priority candidate plan selects the candidate plan with the largest target area coverage volume ratio; The bladder protection candidate plan is selected from the candidate plans that meet the target coverage level constraints and have the lowest dose to the highest irradiated volume of the bladder. The candidate program for rectal protection is selected from those that meet the target coverage level constraints, and the one with the lowest maximum irradiated volume dose in the rectum is chosen.

8. The executable residence time plan generation method according to claim 1, characterized in that, Before outputting the stay time plan, a deliverability check is performed on the stay time plan, the deliverability check including: Verify that the dwell time of all dwell locations is non-negative, the dwell time of a single dwell location does not exceed the set upper limit, the total dwell time does not exceed the device's allowed upper limit, and the dwell location index is consistent with the applicator channel definition; The dwell time plan is output in DICOM RTPLAN format or a dwell time file format compatible with afterloading therapy devices.

9. A system for implementing the executable residence time plan generation method according to any one of claims 1-8, characterized in that, The system includes: The dwell time parameterization module is used to fix the dwell position of the radioactive source and convert the unconstrained internal variables corresponding to each dwell position into non-negative dwell time through a differentiable non-negative mapping. The differential dose calculation module is connected to the residence time parameterization module and is used to perform forward dose accumulation on each calculation point according to the line source dose calculation model to obtain a three-dimensional dose distribution, and to establish the gradient relationship between the three-dimensional dose distribution and the non-negative residence time. The clinical target calculation module is connected to the differential dose calculation module and is used to calculate the target coverage loss, organ-at-risk dose over-limit loss and residence time regularization term according to the three-dimensional dose distribution, and to obtain the total loss value by weighting and summing according to the weight of each target. The clinical rule feedback module is connected to the clinical goal calculation module and is used to automatically update the weight of each clinical goal according to the degree of violation of each clinical goal relative to the clinical threshold, and output the updated goal weight to the clinical goal calculation module. The gradient optimization module is data-connected to the differential dose calculation module and the clinical target calculation module, and is used to receive the residence time gradient and perform iterative updates on the unconstrained internal variables according to the total loss value. The candidate plan management module is connected to the gradient optimization module and is used to save multiple candidate plans in the optimization trajectory and filter and output them from the candidate plans according to the clinical priority rules of target area priority, bladder protection and rectal protection. The plan output module is connected to the candidate plan management module for performing deliverability checks on the screened residence time plans and outputting residence time plan files that can be directly executed by the afterloading therapy device.

10. The application of the executable residence time plan generation method as described in any one of claims 1-8 in adaptive inverse planning design for high-dose-rate afterloading therapy for cervical cancer; Alternatively, it can be extended to brachytherapy scenarios involving the prostate, breast, skin, head and neck, or other locations with fixed applicator / catheter placement.

Images