Computational dvh inverse planning method based on case priori library and adaptive weight search

Abstract

The present application relates to the field of medical radiation physics and brachytherapy, and discloses a DVH reverse planning method based on case priori library and adaptive weight search, comprising the following steps: S1, case priori library construction and initial plan generation: extracting key features and corresponding "high-quality seed arrangement mode" from the past clinically recognized high-quality iodine-125 particle implantation treatment cases to construct the case priori library, after extracting the key features of the new case, S2, DVH error driven weight adaptive outer ring, S3, short-range inner ring search of maintaining existing optimizer, the present application is composed of three parts of case priori library construction and initial plan generation, DVH error driven weight adaptive outer ring and short-range inner ring search of maintaining existing optimizer, can effectively improve the planning efficiency, enhance the planning stability and reproducibility, improve the planning accuracy, have the cross-case migration ability, strong compatibility and low implementation cost.

Description

Technical Field

[0001] This invention relates to the field of medical radiation physics and brachytherapy planning technology, specifically to a method for calculating DVH backward planning based on a case prior library and adaptive weight search. Background Technology

[0002] In the field of radiotherapy, precisely delivering radiation doses to maximize tumor cell killing while minimizing radiation damage to surrounding organs at risk (OARs) is a core objective for improving treatment efficacy and patient quality of life. Dose-volume histograms (DVHs), as a key tool for quantitatively assessing dose distribution, condense complex three-dimensional dose data and organ volume information into intuitive statistical charts, clearly presenting the radiation dose levels received by tissues of different volumes. This provides an objective basis for evaluating the quality of treatment plans and has become a core component of radiotherapy planning systems (RTPS). DVH-based inverse planning methods, by pre-defining the dose-volume targets for the target area and organs at risk, use mathematical optimization algorithms to solve for the optimal beam parameters (such as beam intensity and residence time). Compared to traditional forward planning that relies on physician experience, this significantly improves the accuracy and efficiency of planning and has been widely applied in advanced treatment technologies such as intensity-modulated radiotherapy (IMRT) and volumetric modulated radiotherapy (VMAT).

[0003] With the continuous evolution of radiotherapy technology, the clinical demand for personalized and precise treatment plans continues to increase. The traditional DVH backward planning method has gradually exposed many limitations and is difficult to fully meet the needs of complex clinical scenarios.

[0004] In low-dose-rate (LDR) permanent implantation therapy, dose planning software (TPS) is the core tool connecting clinical treatment needs with actual implantation procedures. Its planning quality and efficiency directly affect treatment outcomes and patient safety. Iodine-125 (I-125) particles, due to their moderate half-life and concentrated radiation dose distribution, are widely used in the treatment of solid tumors such as the prostate, lung, liver, and pancreas. Currently, mainstream TPS technology is based on the TG-43 dose calculation model and the DVH inverse optimization framework, but many problems still need to be addressed in clinical applications. 1. Weight settings rely on human experience, lacking uniformity and reproducibility; 2. The planning process is trapped in a cycle of "parameter tuning and re-optimization," resulting in low efficiency. 3. Fixed-weight strategies have poor adaptability and insufficient ability to transfer data across cases; 4. Poor quality of initial values ​​leads to slow convergence or getting stuck in local optima.

[0005] To address the shortcomings of existing technologies, this invention proposes a reverse planning method for calculating DVH based on a prior case library and adaptive weight search. Specifically, it relates to a treatment planning system (TPS) for permanent implantation of low dose rate (LDR) radioactive particles and its automatic reverse planning method. More specifically, this invention targets iodine-125 (I-125) particle implantation therapy. Within the framework of a TG-43 dose calculation model and a reverse optimization framework based on dose-volume histogram (DVH), it provides a TPS and method that utilizes a prior historical case library for initial value construction and completes automatic reverse planning through an adaptive iterative search driven by DVH error. This invention is applicable to preoperative or intraoperative planning, dose assessment, and needle tract-loading scheme output for I-125 particle implantation in solid tumors such as the prostate, lung, liver, and pancreas. It represents an iterative optimization and upgrade of existing clinically mature TPS technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a reverse planning method for calculating DVH based on a case prior library and adaptive weight search. This method solves the problem that in low-dose-rate (LDR) permanent implantation therapy with radioactive particles, dose planning software (TPS) is the core tool connecting clinical treatment needs with actual particle implantation operations. Its planning quality and efficiency directly affect treatment efficacy and patient safety. Iodine-125 (I-125) particles, due to their moderate half-life and concentrated radiation dose distribution, are widely used in the treatment of solid tumors such as the prostate, lung, liver, and pancreas. Currently, mainstream TPS technology is based on the TG-43 dose calculation model and the DVH reverse optimization framework, but many problems still need to be solved in clinical applications: 1. Weight setting relies on manual experience, lacking uniformity and reproducibility; 2. The planning process is trapped in a "parameter tuning-re-optimization" cycle, resulting in low efficiency; 3. Fixed weight strategies have poor adaptability and insufficient cross-case transferability; 4. Poor quality of initial optimization values ​​leads to slow convergence or getting trapped in local optima.

[0007] This invention provides the following technical solution: a method for calculating DVH backward planning based on a case prior library and adaptive weight search, comprising the following steps: S1. Case Priority Library Construction and Initial Plan Generation: Key features and corresponding "high-quality seed placement patterns" are extracted from previously clinically recognized high-quality iodine-125 particle implantation treatment cases to construct a case priority library. After extracting key features of new cases, similar historical cases are retrieved through a multi-dimensional feature matching algorithm, and their seed distribution is transferred to the target area of ​​new cases through registration or morphological mapping to generate the initial plan P0. Key features of the cases include target volume, morphological complexity, spatial distance from vital organs, prescription dosage level, and commonly used needle path direction; The "high-quality seed arrangement pattern" includes seed spatial distribution, density variation patterns, and hot spot or cold spot treatment methods; S2, DVH error-driven weighted adaptive outer loop: After calling the existing optimizer to complete one round of optimization, the DVH index of the current plan is read and compared with the clinical threshold. The weights of optimization targets such as target coverage, organ at risk protection, and hotspot inhibition are adaptively adjusted according to the error direction and magnitude. The DVH indicators include V100 and D90 of the target area PTV and Dmax and D2cc of the organ at risk OARs, etc. S3. Maintain the short-range inner loop search of the existing optimizer: Set parameters such as the number of optimization iterations, search step size, and convergence criteria, perform short-range fast optimization under the adjusted fixed weights, feed the results back to the weight adaptive outer loop, and iterate until the DVH index reaches the clinical goal. The existing optimizers include IPSA, FSA, simulated annealing algorithm, and genetic algorithm.

[0008] Preferred technical solution 1: The calculation method for the morphological complexity of the target area is as follows: the target area is three-dimensionally modeled to obtain contour data, the ratio of surface area to volume, the irregularity coefficient, and the concavity / convexity parameter are calculated, and the three are multiplied by preset weight coefficients and then summed to obtain a quantitative value; wherein the irregularity coefficient is calculated by the degree of difference in surface area between the target area and a sphere of the same volume, and the concavity / convexity parameter is obtained by analyzing the curvature change of the contour points of the target area and statistically analyzing the number and depth of the concave areas.

[0009] Preferred technical solution 2: The spatial distance between the target area and the vital organ is determined by: calculating the straight-line distance between the geometric center of the target area and the vital organ and the shortest distance between each point on their contours; multiplying the straight-line distance by a first preset weighting coefficient and the shortest distance by a second preset weighting coefficient, and then summing the results; the sum of the two weighting coefficients is 1.

[0010] Preferred technical solution 3: The implementation method of the multi-dimensional feature matching algorithm is as follows: set weight coefficients for each case feature, calculate the similarity of numerical features using Euclidean distance, assign similarity values ​​to categorical features according to the matching situation, calculate the comprehensive similarity by weighted summation, and select the top N (N ranges from 5 to 10) historical cases as similar cases.

[0011] This solution can effectively improve the accuracy and relevance of similar case retrieval, ensure the high quality and adaptability of the initial plan, and enhance cross-case migration capabilities and planning stability.

[0012] Preferred technical solution four: The registration or morphological mapping process adopts a combination of rigid registration and flexible registration: first, the target area is roughly aligned by using the geometric center and main axis, then fine adjustments are made based on the contour feature points and internal structure, and finally the seed distribution of similar cases is mapped to the target area of ​​the new case according to the registration relationship.

[0013] This solution can quickly achieve initial target alignment, improve registration efficiency, accurately match target morphological details, ensure registration accuracy, ensure accurate migration and clinical adaptability of seed distribution, reduce subsequent optimization costs, and improve overall planning efficiency.

[0014] Preferred technical solution five: Adaptive weight adjustment strategy includes: when the target coverage index is below the lower threshold, the target weight is increased by Wnew=Wold×(1+k1×E) (k1 value is 0.2-0.5); when the organ at risk index exceeds the upper threshold, the protection weight is increased by Woarnew=Woarold×(1+k2×Eoar) (k2 value is 0.1-0.6) and a weight locking factor is set; when the index is close to meeting the standard, the weight update is weakened; when there is a hotspot, the inhibition weight is increased by Whotnew=Whotold×(1+k4×S×R) (k4 value is 0.2-0.4).

[0015] Preferred technical solution six: The number of optimization iterations M for short-range inner loop search is 50-200, the search step size is dynamically set according to the weight adjustment range, and the convergence criterion is that the change in the DVH index is less than 1% or the preset number of iterations is reached.

[0016] The preferred technical solution seven: The extraction process of the "high-quality seed arrangement pattern" is as follows: analyze the clinically validated high-quality treatment plan, obtain the spatial distribution coordinates of the seeds, statistically analyze the seed density distribution in the target area, analyze the location and range of hot spots and cold spots and summarize the treatment methods, and record the distance settings between the seeds and the target area boundary, important organs and the seed activity selection strategy in different areas.

[0017] This solution can more accurately capture the core deployment logic of high-quality plans, ensure the quality of initial plans and standardized extraction processes, and improve the stability and reproducibility of cross-case migration.

[0018] Preferred technical solution eight: applicable to preoperative or intraoperative planning, dosage assessment and needle-path loading protocol output for iodine-125 particle implantation in solid tumors of the prostate, lung, liver and pancreas.

[0019] Preferred technical solution nine: A storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1-9.

[0020] Compared with the prior art, the present invention provides a method for calculating DVH backward planning based on a case prior library and adaptive weight search, which has the following beneficial effects: The present invention consists of three parts: case prior library construction and initial plan generation, DVH error-driven adaptive weight outer loop, and short-range inner loop search that maintains the existing optimizer. 1. Effectively improves planning efficiency: Through the construction of a prior case library and the generation of initial plans, the initial plan P0 for new cases is generated based on the seed arrangement pattern of historical high-quality cases, so that the optimization process starts from the region close to the high-quality plan, significantly reducing the random search time; at the same time, the DVH error-driven weighted adaptive outer loop replaces manual parameter tuning, avoiding the "parameter tuning-re-optimization" cycle. Combined with the rapid feedback of short-range inner loop search, the time for treatment plan formulation is greatly shortened, which can meet the needs of clinical practice, especially for rapid intraoperative planning.

[0021] 2. Enhanced planning stability and reproducibility: The weight adjustment in this invention is achieved through an automated strategy, unaffected by differences in human experience. For the same case, regardless of who performs the operation, a consistent weight adjustment process and optimization results can be obtained, improving the reproducibility of the treatment plan. At the same time, the initial plan based on prior case information is of higher quality. Combined with an adaptive weight iteration mechanism, it can effectively avoid the optimization process from getting stuck in local optima, ensuring that different cases can obtain stable and reliable treatment plans, significantly enhancing the stability of the planning.

[0022] 3. Improved planning accuracy: When there is insufficient target coverage or excessive risk to organs, the weighted adaptive outer loop will automatically adjust the corresponding weights according to the DVH error, so that the optimization process can accurately focus on the indicators that have not been met and optimize them in a targeted manner. At the same time, the "high-quality seed distribution mode" in the case prior library is clinically validated and can provide a more reasonable initial reference for seed distribution for new cases, further improving the accuracy of treatment planning, ensuring that the target area receives sufficient dose coverage, and protecting organs at risk to the greatest extent.

[0023] 4. Cross-case transfer capability: The adaptive weighting mechanism of this invention can automatically adjust the weighting parameters according to the characteristics and DVH index performance of different cases, without the need to perform a lot of weighting adjustments for each new case, and has a strong cross-case transfer capability. At the same time, the case prior library covers a variety of high-quality cases with different characteristics, which can provide suitable initial plan references for new cases of different types and characteristics, further expanding the scope of application of this invention.

[0024] 5. Strong compatibility and low implementation cost: Without changing the existing TG-43 dose calculation and DVH inverse optimization framework, this invention can upgrade the existing TPS by simply adding a case prior library module and a weight adaptive outer loop module. It does not require replacing the existing mature optimizer and has good compatibility with existing technologies. At the same time, the construction of the case prior library can be based on the hospital's previous clinical case data, without the need for additional investment in data collection. The implementation logic of the weight adaptive outer loop and short-range inner loop search is simple, easy to program and integrate, and has a low implementation cost. Attached Figure Description

[0025] Figure 1 This is a flowchart of the S1 case prior library construction and initial plan generation process based on the DVH calculation backward planning method proposed in this invention using a case prior library and adaptive weight search. Figure 2 The flowchart of the S2DVH error-driven weighted adaptive outer loop based on the DVH calculation backward programming method based on case prior library and adaptive weight search proposed in this invention is shown. Figure 3 This is a flowchart of the short-range inner loop search of the S3-preserving optimizer based on the DVH backward programming method proposed in this invention, which is based on a case prior library and adaptive weight search. Detailed Implementation

[0026] Please see Figure 1-3 , Example 1: A backward programming method for calculating DVH based on a case prior library and adaptive weight search, including the following steps: S1. Case Priority Library Construction and Initial Plan Generation: Key features and corresponding "high-quality seed placement patterns" are extracted from previously clinically recognized high-quality iodine-125 particle implantation treatment cases to construct a case priority library. After extracting key features of new cases, similar historical cases are retrieved through a multi-dimensional feature matching algorithm, and their seed distribution is transferred to the target area of ​​new cases through registration or morphological mapping to generate the initial plan P0. Key features of the case include target volume, morphological complexity, spatial distance from vital organs, prescribed dose level, and commonly used needle path direction; "High-quality seed arrangement pattern" includes seed spatial distribution, density variation pattern, hot spot or cold spot treatment method, etc. S2, DVH error-driven weighted adaptive outer loop: After calling the existing optimizer to complete one round of optimization, the DVH index of the current plan is read and compared with the clinical threshold. The weights of optimization targets such as target coverage, organ at risk protection, and hotspot inhibition are adaptively adjusted according to the error direction and magnitude. DVH indicators include V100 and D90 of the target volume PTV and Dmax and D2cc of the organ at risk OARs. S3. Maintain the short-range inner loop search of the existing optimizer: Set parameters such as the number of optimization iterations, search step size, and convergence criteria, perform short-range fast optimization under the adjusted fixed weights, feed the results back to the weight adaptive outer loop, and iterate until the DVH index reaches the clinical goal. Existing optimizers include IPSA, FSA, simulated annealing algorithm, and genetic algorithm.

[0027] Example 2: The difference between this example and Example 1 is that the calculation method for the complexity of the target area morphology is as follows: the target area is three-dimensionally modeled to obtain contour data, the ratio of surface area to volume, the irregularity coefficient, and the concavity / convexity parameter are calculated, and the three are multiplied by a preset weight coefficient and then summed to obtain a quantitative value; the irregularity coefficient is calculated by the degree of difference in surface area between the target area and a sphere of the same volume, and the concavity / convexity parameter is obtained by analyzing the curvature change of the target area contour points and statistically analyzing the number and depth of the concave areas.

[0028] Example 3: The difference between this example and Example 1 is that the spatial distance between the target area and the vital organ is determined as follows: the straight-line distance between the geometric center of the target area and the vital organ and the shortest distance between each point on their contours are calculated. The straight-line distance is multiplied by a first preset weighting coefficient, and the shortest distance is multiplied by a second preset weighting coefficient and then summed. The sum of the two weighting coefficients is 1.

[0029] Example 4: The difference between this example and Example 1 is that the multi-dimensional feature matching algorithm is implemented as follows: weight coefficients are set for each case feature, numerical features are calculated using Euclidean distance, categorical features are assigned similarity values ​​according to matching conditions, weighted summation is used to obtain the comprehensive similarity, and the top N (N ranges from 5 to 10) historical cases are selected as similar cases.

[0030] It can effectively improve the accuracy and relevance of similar case retrieval, ensure the high quality and adaptability of the initial plan, and enhance the ability to migrate across cases and the stability of the plan.

[0031] Example 5: The difference between this example and Example 1 is that the registration or morphological mapping process adopts a combination of rigid registration and flexible registration: first, the target area is roughly aligned by using the geometric center and the main axis, then fine adjustments are made based on the contour feature points and internal structure, and finally the seed distribution of similar cases is mapped to the target area of ​​the new case according to the registration relationship.

[0032] It can quickly achieve initial alignment of the target area, improve registration efficiency, accurately match target morphological details, ensure registration accuracy, ensure accurate migration and clinical adaptability of seed distribution, reduce subsequent optimization costs, and improve overall planning efficiency.

[0033] Example 6: The difference between this example and Example 1 is that the adaptive weight adjustment strategy includes: when the target coverage index is below the lower threshold, the target weight is increased by Wnew=Wold×(1+k1×E) (k1 takes a value of 0.2-0.5); when the organ at risk index exceeds the upper threshold, the protection weight is increased by Woarnew=Woarold×(1+k2×Eoar) (k2 takes a value of 0.1-0.6) and a weight locking factor is set; when the index is close to meeting the target, the weight update is weakened; when there is a hotspot, the inhibition weight is increased by Whotnew=Whotold×(1+k4×S×R) (k4 takes a value of 0.2-0.4).

[0034] Example 7: The difference between this example and Example 1 is that the number of optimization iterations M for the short-range inner loop search is 50-200, the search step size is dynamically set according to the weight adjustment range, and the convergence criterion is that the change in the DVH index is less than 1% or the preset number of iterations is reached.

[0035] Example 8: The difference between this example and Example 1 is that the extraction process of the "high-quality seed arrangement pattern" is as follows: analyze the clinically validated high-quality treatment plan, obtain the spatial distribution coordinates of the seeds, statistically analyze the seed density distribution in the target area, analyze the location and range of hot spots and cold spots and summarize the treatment methods, and record the distance settings between the seeds and the target area boundary, important organs and the seed activity selection strategy in different areas.

[0036] To more accurately capture the core deployment logic of high-quality plans, ensure the quality of initial plans and standardized extraction processes, and improve the stability and reproducibility of cross-case migration.

[0037] Example 9: The difference between this example and Example 1 is that it is applicable to the preoperative or intraoperative planning, dosage assessment and needle-path loading protocol output for iodine-125 particle implantation in solid tumors of the prostate, lung, liver and pancreas.

[0038] Example 10: The difference between this example and Example 1 is that a storage medium stores a computer program thereon, and when the computer program is executed by a processor, it implements the steps of any one of claims 1-9.

[0039] In this embodiment, dose planning software (TPS) is the core tool connecting clinical treatment needs with actual particle implantation in low-dose-rate (LDR) permanent implantation therapy. Its planning quality and efficiency directly affect treatment efficacy and patient safety. Iodine-125 (I-125) particles are widely used in the treatment of solid tumors such as prostate, lung, liver, and pancreas due to their moderate half-life and concentrated radiation dose distribution. Currently, mainstream TPS technology is based on the TG-43 dose calculation model and the DVH inverse optimization framework, but many problems still need to be solved in clinical applications. 1. Weight settings rely on human experience, lacking uniformity and reproducibility; 2. The planning process is trapped in a cycle of "parameter tuning and re-optimization," resulting in low efficiency. 3. Fixed-weight strategies have poor adaptability and insufficient ability to transfer data across cases; 4. Poor quality of initial values ​​leads to slow convergence or getting stuck in local optima.

[0040] To address the shortcomings of the existing technologies, this invention proposes a method for calculating DVH backward planning based on a case prior library and adaptive weight search.

[0041] In summary, in practical implementation, the technical solution of this invention consists of three parts: case prior library construction and initial plan generation, DVH error-driven weighted adaptive outer loop, and short-range inner loop search maintaining the existing optimizer. For case prior library construction: the system extracts two types of key information from previously completed and clinically approved high-quality iodine-125 particle implantation treatment cases to construct the case prior library. The first type of information consists of key features of the case, including target volume, target morphological complexity, spatial distance between the target and vital organs, prescription dose level, and commonly used needle path direction. These features directly affect the logic and effectiveness of particle implantation treatment planning. The second type of information is the corresponding "high-quality seed arrangement pattern", which includes seed spatial distribution, density variation patterns, and common hot or cold spot treatment methods. This pattern is a clinically validated, efficient, and safe seed arrangement scheme that can provide a reliable reference for new cases.

[0042] New case feature extraction: When a new case is input, the system uses the same standards and methods as the case feature extraction in the prior case library to calculate the key features of the new case, ensuring the consistency and comparability of the features.

[0043] Similar Case Retrieval: Based on the key features of the new case, a multi-dimensional feature matching algorithm is used to retrieve the most similar set of historical cases from the prior case database. By comprehensively considering the similarity of various key features, historical cases that are closest to the treatment needs and anatomical structures of the new case are selected, providing a high-quality reference for the initial plan generation.

[0044] Initial plan generation: The seed distribution in the "high-quality seed layout pattern" of similar cases retrieved is transferred to the target area of ​​the new case through registration or morphological mapping techniques to generate a priori initial plan P0. This initial plan P0 is not the final treatment plan, but it enables the subsequent optimization process to start from the "region close to a good plan", significantly reducing the random search time of the optimization algorithm and improving optimization efficiency; For the DVH error-driven adaptive weight outer loop, this invention adds a "adaptive weight outer loop" to the traditional DVH inverse optimization framework to achieve automatic weight adjustment, replacing manual parameter tuning. The specific process is as follows: Optimized execution: Invoke an existing mature optimizer to perform one round of optimization calculations to obtain the current optimization plan.

[0045] DVH metrics reading: Read the DVH metrics of the current optimized plan, including V100 (the percentage of the target volume receiving at least 100% of the prescribed dose) and D90 (the dose received at 90% of the target volume) of the target volume, and Dmax (the maximum dose received by the organ at risk) and D2cc (the maximum dose received by 2cc of the organ at risk) of each organ at risk. These metrics are the core basis for evaluating the quality of the treatment plan.

[0046] Clinical threshold comparison: The current DVH index is compared with the preset clinical threshold to determine the error direction (exceeding or meeting the target) and magnitude (the degree of deviation from the threshold) of each DVH index. The clinical thresholds are determined based on clinical treatment guidelines, expert consensus, and a large amount of clinical practice data to ensure the safety and effectiveness of treatment.

[0047] Adaptive weight adjustment: The weights of each optimization objective are adaptively adjusted based on the direction and magnitude of the error. If insufficient target coverage is detected (e.g., PTV-V100 is below the clinical threshold), the target-related weights are automatically increased, making the optimization algorithm more focused on ensuring target dose coverage. If an organ at risk (OAR) is found to have exceeded the limit (e.g., the Dmax of the OAR is higher than the clinical threshold), the protection weight of the OAR will be automatically increased, and a weight locking factor will be set to make it more "tough" in the next round of optimization, giving priority to ensuring the safety of the organ at risk. If each indicator is close to meeting the target (with a small deviation from the clinical threshold), the weight update magnitude will be automatically reduced to avoid over-adjustment that could cause plan fluctuations.

[0048] Through the above mechanism, the outer ring replaces "manual parameter adjustment" with a set of repeatable and standardized automatic strategies to ensure the scientific nature and consistency of weight adjustment.

[0049] For short-range inner-loop search while maintaining the existing optimizer, after each round of outer-loop weight update, the system calls existing mature optimization modules (such as IPSA, FSA, simulated annealing algorithm, genetic algorithm, etc.) to perform short-range fast optimization, as follows: Optimize parameter settings: Set optimization parameters for the short-range inner loop search, including the number of iterations, search step size, and convergence criteria. Parameter settings should balance optimization efficiency and effectiveness to ensure that evaluable optimization results are obtained in a short time.

[0050] Short-range optimization execution: Under fixed weights, short-range rapid optimization is performed according to the set optimization parameters. The inner loop does not aim to reach the global optimum in one go, but rather quickly generates a "new, evaluable DVH result" to provide data support for the next round of weight adjustment of the outer loop.

[0051] Optimization result feedback: The optimization results obtained from the short-range inner loop search (including new DVH indicators and treatment plan parameters) are fed back to the weight-adaptive outer loop. The outer loop continues to adjust the weights based on the results, and this process is repeated until the DVH indicators reach the clinical target.

[0052] In practical applications, the first step is to construct a case prior library. For data collection, 1,200 high-quality cases were selected from iodine-125 particle implantation treatment cases completed and clinically followed up between 2018 and 2023 in the hospital. These cases included 400 cases of prostate cancer, 350 cases of lung cancer, 300 cases of liver cancer, and 150 cases of pancreatic cancer, ensuring the diversity and representativeness of case types.

[0053] The following data should be collected for each case: Patient basic information: age, gender, height, weight, etc.; Tumor-related information: tumor location, pathological type, stage, target volume, and three-dimensional data of target morphology; Information on important organs: three-dimensional anatomical data of important organs such as the spinal cord, kidneys, and liver, and their spatial relationship with the target area; Treatment plan data: prescription dose level (80-160Gy), commonly used needle path direction, seed spatial distribution coordinates, seed activity (0.3-0.8mCi), and DVH index data.

[0054] Regarding the extraction of key case features, especially target volume extraction: using the medical image segmentation technology of the ITK library, the three-dimensional contour of the target area is extracted from CT or MRI images, and the target volume (unit: cm³) is automatically calculated. For example, the target volume of a certain prostate tumor case is 38.5 cm³. Target area morphological complexity calculation: The three-dimensional contour data of the target area is processed, and the surface area to volume ratio (unit: cm⁻¹), irregularity coefficient (the degree of difference in surface area between the target area and a sphere of the same volume), and concavity / convexity parameter (by analyzing the curvature changes of contour points, the number and depth of concave areas are counted) are calculated in sequence. The three parameters are multiplied by weighting coefficients of 0.3, 0.4, and 0.3 respectively and then summed to obtain a quantitative value. For example, the quantitative value of morphological complexity of a lung tumor case is 0.85×0.3+1.62×0.4+0.78×0.3=1.125. Calculation of spatial distance between the target area and vital organs: Select the geometric center of the target area and vital organs, and calculate the straight-line distance between them; extract the three-dimensional contours of the two, calculate the shortest distance between each point on the contour and take the minimum value; multiply the straight-line distance by 0.6 and the minimum spatial distance by 0.4 and then sum them to obtain the comprehensive spatial distance. For example, in a case of liver tumor, the comprehensive spatial distance between the target area and the kidney is 4.2×0.6+1.8×0.4=3.24cm. Other feature extraction: Directly extract prescription dosage levels (unit: Gy) and commonly used acupuncture directions (represented by angles in a three-dimensional coordinate system, such as α=30°, β=45°, γ=60°) from case data.

[0055] For the extraction of the "high-quality seed arrangement pattern", the spatial distribution of seeds is: the three-dimensional spatial coordinates (x, y, z) of the seeds are directly extracted from the treatment plan data. Seed density variation pattern: The target area is evenly divided into 1cm×1cm×1cm cubic sub-regions. The number of seeds in each sub-region is counted, and the seed density (number of seeds or volume of sub-region) is calculated to clarify the density distribution pattern. For example, in a case of prostate tumor, the seed density in the central sub-region of the target area is 0.8 seeds or cm³, and the seed density in the peripheral sub-region is 0.5 seeds or cm³. Hot spot or cold spot treatment methods: Analyze DVH index data, define the location and range of hot spot areas (dose ≥ 120% of prescription dose) and cold spot areas (dose ≤ 90% of prescription dose), and record the corresponding treatment methods, such as reducing the number of seeds in hot spot areas by 20% and increasing the number of seeds in cold spot areas by 30%.

[0056] For the case prior database storage, the extracted key features of the cases and the corresponding "high-quality seed layout patterns" are organized in a relational database format, and a unique ID is assigned to each case to establish a relationship between the two. The database is stored in a 2TB NVMe solid-state drive to complete the construction of the case prior database, which is convenient for subsequent retrieval and retrieval.

[0057] For generating an initial plan for a new case, taking a 45-year-old male prostate cancer case as an example, the specific procedures are as follows: New case data input The CT image data of this case (slice thickness 1mm, pixel pitch 0.5mm×0.5mm) was acquired using medical image acquisition equipment. At the same time, clinical data such as patient age, tumor pathology type (adenocarcinoma), stage (T2a stage), and prescription dose (145Gy) were entered and imported into the TPS system.

[0058] New case feature extraction Using the same method as the case prior library construction, the key characteristics of new cases were calculated: target volume 36.8 cm³, target morphological complexity quantification value 1.08, comprehensive spatial distance between the target area and the spinal cord 5.3 cm, prescription dose 145 Gy, and commonly used needle tract directions α=25°, β=40°, γ=55°.

[0059] Similar Case Search Feature weight settings: According to the degree of influence of each feature on the treatment plan, set the weight coefficients as follows: target volume (0.25), target morphological complexity (0.25), spatial distance between target area and important organs (0.2), prescription dose level (0.2), and commonly used needle path direction (0.1). Similarity calculation: Numerical features (such as target volume and spatial distance) are calculated using Euclidean distance, with the formula Sim=1-(|x_new-x_hist| or max(x_range)) (x_new is the feature value of a new case, x_hist is the feature value of a historical case, and x_range is the range of values ​​for that feature in the database); the direction of the needle path is often determined by calculating the difference in direction angles, and the smaller the difference, the higher the similarity. Comprehensive similarity calculation: The similarity of each feature dimension is weighted and summed according to the corresponding weight coefficient to obtain the comprehensive similarity between the new case and each historical case; Similar case selection: The top 8 cases were selected as the most similar case group by sorting them from highest to lowest overall similarity.

[0060] Seed distribution migration Rigid registration: Calculate the geometric center coordinates and principal inertial axis directions of the target areas of the new case and 8 similar cases respectively. Translate the geometric center of the target area of ​​the similar cases to the geometric center of the target area of ​​the new case, and rotate the principal inertial axis to be consistent to achieve approximate alignment of the target areas. Elastic registration: The B-spline elastic registration algorithm is adopted to select the target area contour feature points (vertices, inflection points, etc.) and internal control points. Based on grayscale and morphological information, the correspondence between feature points is established, and the position of control points is adjusted to achieve fine matching of the target area. The registration error is controlled within 0.5mm. Seed distribution mapping: The spatial distribution of seeds from 8 similar cases is mapped to the target area of ​​new cases according to the registration relationship; the number and activity of seeds are determined by weighted average method (weight is the comprehensive similarity of similar cases), and fine-tuned by referring to the density pattern of similar cases and the treatment of hot spots or cold spots, to generate the initial plan P0 (P0 target area V100=82%, D90=128Gy, spinal cord Dmax=38Gy, kidney D2cc=32Gy).

[0061] For adaptive weight iterative optimization The initial weight settings are based on clinical experience and the weight statistics of similar cases in the prior case library. The initial weight parameters are set as follows: target coverage weight W_ptv=1.0, spinal cord protection weight W_spine=1.5, kidney protection weight W_kidney=1.2, and hotspot inhibition weight W_hot=0.8.

[0062] Outer loop weight adjustment and inner loop short-range optimization cycle The first round of optimization, inner loop short-range optimization: call the IPSA optimizer, set the number of iterations M=100, the search step size 0.05, and the convergence criterion is that the change in the DVH index is less than 1%; perform short-range optimization under the initial weights to obtain the first round of optimization plan; DVH Indicator Reading: Read the DVH indicators of the optimization plan: PTV-V100=88% (clinical lower limit threshold 95%), PTV-D90=135Gy (clinical lower limit threshold 140Gy), spinal cord Dmax=42Gy (clinical upper limit threshold 40Gy), kidney D2cc=34Gy (clinical upper limit threshold 30Gy), PTV-Dmax=165Gy (hotspot threshold 150Gy); Error calculation: Calculate the errors for each indicator separately: PTV-V100 error E1≈7.37%, PTV-D90 error E2≈3.57%, spinal cord Dmax error E3=5%, kidney D2cc error E4≈13.33%, hot spot intensity S=10%, hot spot range R=6%; Weight adjustment: Adjust weights according to adaptive strategy: W_ptv_new≈1.016, W_spine_new=1.5375 (set locking factor), W_kidney_new≈1.264 (set locking factor), W_hot_new≈0.8014.

[0063] Second round of optimization Inner loop short-range optimization: Using the adjusted weights, set the number of iterations M=120, the search step size 0.04, execute short-range optimization, and obtain the second round of optimization plan; DVH index readings: PTV-V100=92%, PTV-D90=138Gy, spinal cord Dmax=40Gy, kidney D2cc=32Gy, PTV-Dmax=155Gy; Error calculation: The calculated errors are as follows: PTV-V100 error E1≈3.16%, PTV-D90 error E2≈1.43%, kidney D2cc error E4≈6.67%, hot spot intensity S≈3.33%, and hot spot range R=4%; Weight adjustment: The adjusted weights are: W_ptv_new≈1.023, W_spine_new remains at 1.5375, W_kidney_new≈1.298 (with locking factor set), W_hot_new≈0.8017.

[0064] Multiple cycles Repeat the above process of "inner loop short-range optimization - DVH index reading - error calculation - weight adjustment" until the 8th round of optimization, when all DVH indicators reach the clinical target: PTV-V100=96%, PTV-D90=142Gy, spinal cord Dmax=38Gy, kidney D2cc=29Gy, PTV-Dmax=148Gy.

[0065] Treatment plan output Once the DVH index meets the clinical target, iterative optimization stops, and the final treatment plan is output, including needle-loading scheme (needle location coordinates, orientation angle, number, and seed quantity, activity, and spacing within each needle tract), DVH chart, and three-dimensional dose distribution image, providing precise guidance for clinical particle implantation therapy.

[0066] Through the above-described steps, this invention enables automated and efficient planning of iodine-125 particle implantation treatment. Compared to the traditional TPS method, the planning time of this approach is reduced from an average of 75 minutes to 22 minutes, the DVH target achievement rate is increased from 76.67% to 100%, and the planning coefficient of variation is reduced from 7.8% to 2.1%, significantly improving planning efficiency, accuracy, and stability. It is suitable for preoperative or intraoperative treatment planning of solid tumors such as the prostate, lung, liver, and pancreas.

Claims

1. A backward programming method for calculating DVH based on a case prior library and adaptive weight search, characterized by: Includes the following steps: S1. Case Priority Library Construction and Initial Plan Generation: Key features and corresponding "high-quality seed placement patterns" are extracted from previously clinically recognized high-quality iodine-125 particle implantation treatment cases to construct a case priority library. Key features of new cases are extracted, and similar historical cases are retrieved through multi-dimensional feature matching algorithms. Their seed distribution is transferred to the target area of ​​new cases through registration or morphological mapping to generate the initial plan P0. Key features of the cases include target volume, morphological complexity, spatial distance from vital organs, prescription dosage level, and commonly used needle path direction; The "high-quality seed arrangement pattern" includes seed spatial distribution, density variation pattern, and hot spot or cold spot treatment method; S2, DVH error-driven weighted adaptive outer loop: After calling the optimizer to complete one round of optimization, the DVH index of the current plan is read and compared with the clinical threshold. The weights of optimization targets such as target coverage, organ at risk protection, and hotspot inhibition are adaptively adjusted according to the error direction and magnitude. The DVH indicators include V100 and D90 of the target area PTV and Dmax and D2cc of each organ at risk OAR. S3. Maintain the short-range inner loop search of the existing optimizer: Set the number of optimization iterations, search step size, and convergence criterion parameters, perform short-range fast optimization under the adjusted fixed weights, feed the results back to the weight adaptive outer loop, and iterate until the DVH index reaches the clinical goal. The existing optimizers include IPSA, FSA, simulated annealing algorithm, and genetic algorithm.

2. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 1, characterized in that: The calculation method for the morphological complexity of the target area is as follows: the target area is modeled in three dimensions to obtain contour data, the ratio of surface area to volume, the irregularity coefficient, and the concavity and convexity parameters are calculated, and the three are multiplied by preset weight coefficients and then summed to obtain the quantized value. The irregularity coefficient is calculated by the difference in surface area between the target area and a sphere of the same volume, while the concavity parameter is obtained by analyzing the curvature changes of the target area contour points and statistically analyzing the number and depth of concave areas.

3. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 2, characterized in that: The spatial distance between the target area and the vital organ is determined as follows: calculate the straight-line distance between the geometric center of the target area and the vital organ and the shortest distance between each point on their contours, multiply the straight-line distance by a first preset weighting coefficient and the shortest distance by a second preset weighting coefficient, and then sum them up. The sum of the two weighting coefficients is 1.

4. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 3, characterized in that: The implementation of the multi-dimensional feature matching algorithm is as follows: weight coefficients are set for each case feature, similarity is calculated using Euclidean distance for numerical features, similarity is assigned to categorical features according to matching conditions, and a weighted sum is obtained to obtain the comprehensive similarity. The top N (N ranges from 5 to 10) historical cases are selected as similar cases.

5. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 4, characterized in that: The registration or morphological mapping process adopts a combination of rigid registration and flexible registration: first, the target area is roughly aligned by using the geometric center and main axis, then fine adjustments are made based on the contour feature points and internal structure, and finally the seed distribution of similar cases is mapped to the target area of ​​new cases according to the registration relationship.

6. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 5, characterized in that: The adaptive weight adjustment strategy includes: when the target coverage index is below the lower threshold, the target weight is increased by Wnew=Wold×(1+k1×E) (k1 value is 0.2-0.5); when the organ at risk index exceeds the upper threshold, the protection weight is increased by Woarnew=Woarold×(1+k2×Eoar) (k2 value is 0.1-0.6) and a weight locking factor is set; when the index is close to meeting the target, the weight is reduced and updated; when there is a hotspot, the inhibition weight is increased by Whotnew=Whotold×(1+k4×S×R) (k4 value is 0.2-0.4).

7. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 8, characterized in that: The number of iterations M for the short-range inner loop search is 50-200. The search step size is dynamically set according to the weight adjustment range. The convergence criterion is that the change in the DVH index is less than 1% or the preset number of iterations is reached.

8. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to claim 7, characterized in that: The extraction process of "high-quality seed placement pattern" is as follows: analyze clinically validated high-quality treatment plans, obtain seed spatial distribution coordinates, statistically analyze the seed density distribution in the target area, analyze the location and range of hot spots and cold spots and summarize the treatment methods, and record the distance settings between seeds and target area boundaries, important organs and seed activity selection strategies in different areas.

9. The method for calculating DVH backward programming based on a case prior library and adaptive weight search according to any one of claims 1-8, characterized in that: It is suitable for preoperative or intraoperative planning, dosage assessment, and needle-path loading protocol output for iodine-125 particle implantation in solid tumors of the prostate, lung, liver, and pancreas.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-9.

Images