A BNCT multi-field irradiation plan optimization method, system, device and medium

Abstract

The application belongs to the technical field of radiotherapy, and specifically discloses a BNCT multi-field irradiation plan optimization method, system, device and medium. According to the application, the total input condition characteristics are first constructed by fusing anatomical structures and beam directions, and then the three-dimensional biological equivalent dose rate distribution of each direction is output through a three-dimensional dose rate component prediction neural network and post-processing, thereby replacing time-consuming Monte Carlo simulation, greatly improving the calculation efficiency while ensuring the accuracy; further, a multi-direction combination is enumerated and a target dose value search sequence is set, a variable irradiation time is taken as a variable, a dose deviation, uniformity and under-dose penalty term considering the target dose value are introduced to construct an optimization model and solve the model, thereby overcoming the defects of single-direction irradiation, such as insufficient deep dose and poor uniformity; finally, an optimal scheme meeting the coverage rate is output according to the total dose distribution, thereby realizing fast and accurate BNCT multi-field irradiation plan optimization.

Description

Technical Field

[0001] This application belongs to the field of radiotherapy technology, specifically relating to a method, system, equipment, and medium for optimizing BNCT multifield irradiation plans. Background Technology

[0002] Boron neutron capture therapy (BNCT) is a binary radiotherapy technique that relies on the combined effects of boron drug enrichment and neutron irradiation. Its dose is contributed by multiple physical processes, including the boron capture reaction, nitrogen reaction, hydrogen backflushing, and accompanying gamma component. The dose distribution is closely related to the patient's three-dimensional tissue structure, tumor location, boron concentration distribution, incident direction, beam aperture, and tissue depth. Because the transport, slowing, and capture processes of neutrons in human tissues are complex, BNCT dose calculations typically rely on precise simulations using the Monte Carlo method.

[0003] While Monte Carlo methods offer high physical accuracy, in scenarios involving multiple candidate irradiation directions, multi-field combination screening, and individualized irradiation time optimization, repeated three-dimensional dose calculations are required for each direction and each scheme, leading to high computational costs and making it difficult to meet the needs of rapid candidate scheme comparison and plan-aided design. Especially when it is necessary to simultaneously compare multiple fixed directions, different field numbers, and different prescription constraint levels, traditional directional complete Monte Carlo calculations significantly increase plan generation time.

[0004] Existing radiation dose prediction technologies primarily focus on the total dose distribution of photon and proton therapy, neglecting the different dose components generated by boron trapping, nitrogen reaction, hydrogen backflush, and gamma rays in BNCT. Because these dose components differ in their physical origins, spatial distribution characteristics, and numerical scales, it is difficult to handle joint prediction of multiple dose components using a single network and a unified loss function. Furthermore, some existing BNCT dose prediction methods combine Monte Carlo simulations with neural networks. While this approach reduces the computational burden of high-precision Monte Carlo simulations, it still requires pre-performing Monte Carlo calculations for each patient and each candidate irradiation direction. Therefore, in multi-directional dose prediction and multi-field combination screening, the computation time remains long and the data processing flow is complex.

[0005] Furthermore, in practical individualized treatment planning, higher prescribed doses may be infeasible due to skin or normal tissue dose limitations. Judging feasibility solely under a single fixed prescription could lead to the direct rejection of some directional combinations with good dosimetric potential, and it would be difficult to record the highest practically feasible prescription achievable under current constraints. Additionally, unidirectional irradiation can result in lower doses to deeper tumors and poor overall tumor homogeneity. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application aims to provide a method, system, device, and medium for optimizing BNCT multi-field irradiation plans. The goal is to solve the following problems: Traditional BNCT dose distribution often uses neural networks with high-precision or low-precision Monte Carlo simulations as input to output unidirectional irradiation schemes, which often suffers from insufficient dose to deep tumors, poor uniformity of overall tumor dose distribution, and long computation time.

[0007] The first aspect of this application relates to a BNCT multi-field irradiation planning optimization method, comprising: step S10, constructing a total input condition feature including target object structural input features and beam irradiation direction input features based on three-dimensional voxel data of the target object for multiple preset candidate irradiation directions; step S20, inputting the total input condition feature of each candidate irradiation direction into multiple three-dimensional dose rate component prediction neural networks to obtain the three-dimensional physical dose rate component distribution under each candidate irradiation direction, thereby obtaining the three-dimensional bioequivalent dose rate distribution of each candidate irradiation direction; step S30, enumerating multiple direction combinations from multiple preset candidate irradiation directions, and setting a target dose value search sequence for each direction combination; step S40, searching according to the target dose value. For each target dose value in the sequence, the preset optimization model is solved one by one in combination with the three-dimensional bioequivalent dose rate distribution until the solution of the preset optimization model makes the three-dimensional total bioequivalent dose distribution of the target area meet the target dose coverage requirement, so as to output the feasible combination in each direction combination and the corresponding irradiation time and target dose value; wherein, the preset optimization model is: using the irradiation time of each irradiation direction in the direction combination as the optimization variable, an objective function is established including the dose deviation term, dose uniformity term and underdose penalty term of the target area, and irradiation time constraint and dose constraint are set; the dose deviation term, dose uniformity term and underdose penalty term are all dose evaluation function terms constructed by taking into account the influence of the current target dose value and the three-dimensional bioequivalent dose rate distribution on the dose distribution characteristics of the target area.

[0008] In one embodiment, step S10 includes: preprocessing the three-dimensional voxel data of the computed tomography scan of the target object to obtain the three-dimensional tissue classification code and the three-dimensional target object mask as the target object structural input features; generating a three-dimensional beam depth distribution, a three-dimensional beam radius distribution and a three-dimensional aperture mask per voxel according to the beam injection direction and beam central axis of each candidate irradiation direction as the beam irradiation direction input features; and constructing a multi-channel three-dimensional input tensor according to the target object structural input features and the beam irradiation direction input features as the total input condition features.

[0009] In one embodiment, the multiple three-dimensional dose rate component prediction neural networks include: a three-dimensional boron physical dose rate component prediction model, a three-dimensional nitrogen physical dose rate component prediction model, a three-dimensional hydrogen physical dose rate component prediction model, and a three-dimensional photon physical dose rate component prediction model; each physical dose rate component prediction model is implemented through a three-dimensional U-shaped neural network structure with residual modules and attention jump connections; step S20 includes: inputting the total input condition features of each candidate irradiation direction into the three-dimensional boron physical dose rate component prediction model, the three-dimensional nitrogen physical dose rate component prediction model, the three-dimensional hydrogen physical dose rate component prediction model, and the three-dimensional photon physical dose rate component prediction model, respectively, to obtain the boron physical dose rate component distribution, nitrogen physical dose rate component distribution, hydrogen physical dose rate component distribution, and photon physical dose rate component distribution under each candidate irradiation direction; based on the multiple physical dose rate components under each candidate irradiation direction, the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction is obtained by adding the biological effect coefficients voxel by voxel.

[0010] In one embodiment, step S30 includes: setting a maximum number of irradiation fields, enumerating all direction combinations that satisfy the maximum number of irradiation fields limit from multiple preset candidate irradiation directions; setting a target dose value search sequence for each direction combination, the target dose value search sequence containing multiple target dose values ​​arranged in descending order.

[0011] In one embodiment, step S40 includes: selecting the maximum value in the target dose value search sequence under the current direction combination as the current target dose value; combining the three-dimensional biological equivalent dose rate distribution, using a sequential least squares programming algorithm to solve a preset optimization model under the current target dose value condition; calculating the three-dimensional total biological equivalent dose distribution based on the irradiation time of each direction obtained from the solution; when the current three-dimensional total biological equivalent dose distribution meets the target dose coverage requirement, determining the current direction combination as a feasible combination, and outputting the irradiation time of the feasible combination; when the current three-dimensional total biological equivalent dose distribution does not meet the target dose coverage requirement, taking the next target dose value in the search sequence as the current target dose value according to the decreasing order, and executing the step of solving the preset optimization model; if all target dose values ​​in the search sequence do not meet the target dose coverage requirement, determining the current direction combination as an infeasible direction combination.

[0012] In one embodiment, the target dose coverage requirement is that 80% of the volumetric dose of the target target area determined based on the current three-dimensional total biological equivalent dose distribution is not lower than the current target dose value.

[0013] In one embodiment, after step S40, the method further includes: prioritizing all feasible options based on feasible target dose values, target area dose uniformity, dose to critical sites, and total irradiation time, and outputting the highest priority feasible option as the recommended irradiation option; obtaining the corresponding reference three-dimensional total biological equivalent dose distribution calculated by Monte Carlo based on the direction combination and irradiation time of the recommended irradiation option; and reviewing the recommended irradiation option based on the reference three-dimensional total biological equivalent dose distribution and the three-dimensional total biological equivalent dose distribution.

[0014] The second aspect of this application relates to a BNCT multi-field irradiation planning optimization system, comprising: a feature construction module, used to construct a total input condition feature including target object structural input features and beam irradiation direction input features for multiple preset candidate irradiation directions, combined with three-dimensional voxel data of the target object; a dose rate component prediction module, used to input the total input condition feature of each candidate irradiation direction into a three-dimensional dose rate component prediction neural network to obtain multiple three-dimensional physical dose rate component distributions under each candidate irradiation direction, so as to obtain the three-dimensional bioequivalent dose rate distribution of each candidate irradiation direction; a combination enumeration and search module, used to enumerate multiple direction combinations from multiple preset candidate irradiation directions, and set a target dose value search sequence for each direction combination; and an irradiation time optimization and feasibility verification module, used to solve the preset optimization model one by one according to each target dose value of the target dose value search sequence, and in combination with the three-dimensional bioequivalent dose rate distribution, calculate the three-dimensional total bioequivalent dose distribution according to the irradiation time of each direction obtained by the solution, until the solution of the preset optimization model makes the three-dimensional total bioequivalent dose distribution of the target target area meet the target dose coverage requirement, so as to output the feasible combination in each direction combination and the corresponding irradiation time and target dose value.

[0015] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0016] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0017] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0018] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0019] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application first constructs a total input condition feature that includes the structural input features of the target object and the beam irradiation direction input features for multiple preset candidate irradiation directions, combined with the three-dimensional voxel data of the target object. This enables the subsequent neural network to simultaneously perceive anatomical structure and irradiation geometry information, laying the foundation for accurate prediction.

[0020] Next, the total input condition features of each candidate direction are input into the three-dimensional dose rate component prediction neural network, which directly outputs the distribution of multiple three-dimensional physical dose rate components under each direction and thus obtains the three-dimensional biological equivalent dose rate distribution. This method replaces the traditional time-consuming Monte Carlo simulation or neural network input that relies on low-resolution Monte Carlo results. Therefore, it not only greatly shortens the dose calculation time, but also avoids the prediction error caused by low input resolution, thus improving the efficiency and accuracy of dose calculation.

[0021] Then, multiple direction combinations are enumerated from multiple preset candidate directions, and a target dose value search sequence is set for each combination. For each target dose value, the preset optimization model is solved one by one in combination with the obtained three-dimensional bioequivalent dose rate distribution. The results are then verified for the coverage of the target dose value in the target area. The model uses the irradiation time of each direction as a variable and constructs a dose deviation term, a dose uniformity term, and an underdose penalty term that take into account the current target dose value and the three-dimensional bioequivalent dose rate distribution as objective functions. Irradiation time constraints and dose constraints are set, so that feasible solutions for different direction combinations under different target doses can be systematically searched, overcoming the inherent defects of difficulty in increasing the dose of deep tumors and uneven dose distribution when irradiating in a single direction.

[0022] Finally, based on the priority ranking of each feasible combination, the recommended irradiation scheme was determined by irradiation direction, irradiation time, and target dose value. This scheme was then verified against the Monte Carlo reference three-dimensional total bioequivalent dose distribution to ensure the clinical acceptability of the final plan.

[0023] In summary, this technical solution introduces an end-to-end three-dimensional dose rate component prediction network to replace traditional simulation, and combines multi-directional enumeration and constrained optimization solutions. While significantly improving the calculation speed, it effectively solves the problems of insufficient dose and poor uniformity of deep tumors caused by unidirectional irradiation, and achieves efficient, accurate and comprehensive BNCT multi-field irradiation plan optimization. Attached Figure Description

[0024] Figure 1This is a flowchart illustrating the BNCT multi-field irradiation planning optimization method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the nine beam irradiation directions provided in the embodiments of this application; Figure 3 This is a box plot of the mean absolute error between the predicted bioequivalent dose rate distribution and the reference bioequivalent dose rate distribution of different tissues of the target object in 144 test results within the test set provided in this application embodiment; Figure 4 A box plot showing the mean absolute percentage error between the predicted bioequivalent dose rate distribution and the reference bioequivalent dose rate distribution of different tissues of the target object in 144 test results within the test set provided in this application embodiment; Figure 5 DVH diagram of the optimal solution for the target object under different numbers of irradiation fields provided in the embodiments of this application; Figure 6 This is a detailed flowchart of the BNCT multi-field irradiation planning optimization method provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0027] In this application, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0028] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0029] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0030] In the subsequent description of this application, there are some similar terms, but they belong to different concepts. These are listed here for explanation: Three-dimensional physical dose rate component distribution: Multiple three-dimensional distributions output by the neural network, each corresponding to the dose rate of a physical process (boron capture, nitrogen reaction, hydrogen recoil, gamma rays), without biological equivalent weighting, and in units of Gy / min.

[0031] Three-dimensional biological equivalent dose rate distribution: The single three-dimensional dose rate distribution obtained by multiplying each of the above components by its respective biological effect coefficient and then summing them, with the unit being / Gy(w) / min, reflects the comprehensive biological effect of a single irradiation direction.

[0032] Three-dimensional total bioequivalent dose distribution: The three-dimensional cumulative dose distribution is obtained by multiplying the three-dimensional bioequivalent dose rate distribution in each direction by the corresponding irradiation time and then superimposing them. The unit is Gy(w), which is the dose result of the final scheme.

[0033] Reference three-dimensional total biological equivalent dose distribution: The three-dimensional total biological equivalent dose distribution (such as Monte Carlo calculation results, historical protocols) serves as a control benchmark and is used to verify or compare the actual output total biological equivalent dose distribution.

[0034] In light of the background technology, this application proposes an embodiment of a BNCT multi-field irradiation planning optimization method. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating the BNCT multi-field irradiation planning optimization method provided in the embodiments of this application.

[0035] In this embodiment, the method includes: Step S10: For multiple preset candidate irradiation directions, construct a total input condition feature that includes the target object structure input feature and the beam irradiation direction input feature by combining the three-dimensional voxel data of the target object.

[0036] It should be noted that the preset candidate irradiation direction refers to a set of multiple possible beam inrush angles, which are typically determined based on the location of the target area, the geometry of the surrounding normal tissue, and the mechanical degrees of freedom of the neutron beam delivery system. The three-dimensional voxel data of the target object refers to the meshed representation of the human anatomical structure reconstructed from medical images (such as CT and MRI), with each voxel containing tissue type (such as tumor, skin, brain tissue, etc.) and density information.

[0037] It should be noted that the target object structure input features are used to describe the anatomical geometry information of the target object, such as the tissue category encoding of voxels, spatial coordinates, relative positions, etc.; while the beam irradiation direction input features characterize parameters such as the incident direction vector of the beam, the size of the beam aperture, the distance of each voxel from the beam axis, and the distance of each voxel from the beam source.

[0038] It should be noted that the total input conditional features are a multidimensional tensor obtained by fusing the two types of features mentioned above, which serves as the input to the subsequent neural network. This step can be achieved by using image processing software (such as the SimpleITK library in Python) to read DICOM data and generate a feature matrix by combining the beam configuration parameters. Alternatives include using point cloud representations or graph neural network input formats, but three-dimensional voxel tensors are more conducive to convolution operations.

[0039] Understandably, the effect of this step is to uniformly encode the patient's individualized anatomical information and beam geometry information, enabling the neural network to perceive spatial location and orientation dependence, thereby laying the foundation for subsequent rapid dose prediction.

[0040] Step S20: Input the total input condition features of each candidate irradiation direction into multiple three-dimensional dose rate component prediction neural networks to obtain the three-dimensional physical dose rate component distribution under each candidate irradiation direction, so as to obtain the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction.

[0041] It should be noted that the three-dimensional dose rate component prediction neural network is a trained deep learning model. Its input is the total input conditional features generated in step S10, and its output is a single three-dimensional physical dose rate component distribution.

[0042] It is understood that the three-dimensional physical dose rate components here include, but are not limited to: the boron dose rate component of alpha particles and lithium nuclei produced by the boron capture reaction, the nitrogen dose rate component of protons produced by the nitrogen reaction, the hydrogen dose rate component of protons produced by hydrogen recoil, and the γ dose rate component consisting of accompanying γ rays and secondary γ rays. Each component corresponds to a three-dimensional tensor, the value of which represents the energy deposited in the voxel per unit time (unit: Gy / min).

[0043] It should be noted that each neural network is responsible for predicting a single dose rate component. The four neural networks have the same structure but independent weights. A typical implementation can be based on 3D U-Net or its variants (such as V-Net), with training data derived from Monte Carlo simulations. Alternative options include using a Transformer architecture or a diffusion model, but 3D convolutional networks are more mature for medical imaging tasks. After obtaining each physical dose rate component, the three-dimensional biological equivalent dose rate distribution can be obtained by multiplying it by the corresponding weighting factor (such as the complex biological effect CBE of boron trapping reaction) and summing the results.

[0044] Understandably, this distribution reflects the equivalent dose rate that comprehensively considers the biological effects of each component. The benefits of this step are: it eliminates the need to rerun Monte Carlo simulations for each candidate direction; the dose rate distribution for all directions can be quickly obtained with just a few forward propagations of the neural network, significantly reducing computation time; and by decoupling the component predictions, it avoids the problem of a single network struggling to fit multi-scale components.

[0045] Step S30: Enumerate multiple direction combinations from multiple preset candidate irradiation directions, and set a target dose value search sequence for each direction combination.

[0046] It should be noted that enumerating multiple direction combinations from multiple preset candidate illumination directions means combining and arranging all candidate directions. For example, selecting k directions from N directions to form a combination (k=1, 2, ..., N) to form a series of multi-field schemes to be evaluated.

[0047] It should be noted that the target dose value search sequence is a set of pre-set prescription dose values ​​for each directional combination, such as values ​​taken at equal intervals from low to high, or several candidate values ​​set based on clinical experience (such as 10 Gy, 12 Gy, 15 Gy, etc.). This sequence is used to explore the feasibility of the directional combination at different prescription levels.

[0048] It should be noted that enumeration can be implemented through programming loops, such as using Python's itertools.combinations function. Alternatives include using heuristic search (such as genetic algorithms) to reduce the number of enumerations. For the small set of candidate directions involved in this embodiment, the full enumeration method is simple to implement, yields stable results, and avoids the problem that heuristic search may miss the optimal combination.

[0049] Understandably, the effect of this step is to: systematically cover all possible combinations of irradiation field numbers and directions, avoiding the omission of potentially high-quality combinations; and by setting a search sequence, allow for the evaluation of multiple prescription levels under the same combination, rather than just determining the feasibility of a single fixed prescription, thereby discovering combinations that are limited under high prescriptions but perform well under medium prescriptions. It also avoids selecting irradiation directions as part of the optimization algorithm, thus preventing the formation of a convex objective function. Step S40: Search for each target dose value in the target dose value search sequence, and solve the preset optimization model one by one in combination with the three-dimensional biological equivalent dose rate distribution until the solution of the preset optimization model makes the three-dimensional total biological equivalent dose distribution of the target target area meet the target dose coverage requirement, so as to output the feasible combination in each direction combination and the corresponding irradiation time and target dose value.

[0050] It should be noted that the preset optimization model is a mathematical programming model used to solve for the optimal irradiation time allocation under a given combination of directions and target dose values. The optimization variable is the irradiation time in each direction. Constraints typically include: the bioequivalent dose to the target area is not less than the target dose value (or the prescribed dose is reached within a certain volume), the total irradiation time does not exceed the drug metabolism time, and the bioequivalent dose to organs at risk (such as skin and brainstem) does not exceed the tolerance limit. The objective function can be to maximize the minimum tumor dose, minimize the integral dose to normal tissues, or minimize the total irradiation time, etc.

[0051] It should be noted that this optimization model can be expressed as a linear programming or quadratic programming problem, and can be solved using the simplex method, interior point method, or gradient descent method. For example, if the dose rate distribution is known and is linearly superimposed (because the dose contribution of the neutron beam is approximately proportional to the irradiation time), then the constraints become linear inequalities, which can be solved efficiently using linear programming.

[0052] It is understandable that "until the solution of the preset optimization model makes the three-dimensional total bioequivalent dose distribution of the target area meet the target dose coverage requirement" means that for the current target dose value, there exists a set of non-negative irradiation times that satisfy all constraints, and the three-dimensional total bioequivalent dose distribution of the target area under this solution meets the target dose coverage requirement. The benefits of this step are: automatically determining the maximum feasible prescription dose achievable for each directional combination under given constraints, avoiding manual trial and error; and obtaining specific irradiation time allocations, providing direct parameters for subsequent treatment execution.

[0053] Specifically, the preset optimization model proposed in this application is as follows: taking the irradiation time of each irradiation direction in the directional combination as the optimization variable, an objective function including the dose deviation term, dose uniformity term and underdose penalty term of the target area is established, and irradiation time constraints and dose constraints are set; the dose deviation term, dose uniformity term and underdose penalty term are all dose evaluation function terms constructed by taking into account the influence of the current target dose value and the three-dimensional biological equivalent dose rate distribution on the dose distribution characteristics of the target area.

[0054] It should be noted that the target area refers to the image information of the target object. Specifically, it is the image data formed by marking the area to be covered by the prescription dose on the tomographic image of the target object after acquiring it through medical imaging equipment. This image information uses pixels or voxels as the basic unit to form a mask image, label image, or contour image that marks the scope of the target area.

[0055] In other words, the target area in this application is an image object in the image space, which is digital image data stored in computer memory, rather than a physically existing tissue entity on the patient's anatomy. This definition allows the target area to be directly used as a data carrier for image processing, neural network calculation, and optimization model evaluation, which is naturally compatible with the planning optimization method based on three-dimensional voxel data and neural networks proposed in this application in terms of data structure.

[0056] It should be noted that irradiation time refers to the actual duration for which the neutron beam is turned on and acts on the target object in each specific irradiation direction. Since the bioequivalent dose rate at each voxel in the target object in each irradiation direction is given by the aforementioned three-dimensional dose rate component prediction neural network, and the bioequivalent dose is linearly related to the irradiation time, the bioequivalent dose contributed by that direction to the target area and surrounding normal tissues can be controlled by adjusting the irradiation time in each direction.

[0057] Understandably, using irradiation time in each direction as a decision variable can directly reflect controllable parameters at the treatment execution level, possessing clear physical significance and clinical feasibility. This model can be executed by the optimization calculation module within the treatment planning system. This module includes a processor, memory, and a computer program stored in the memory. When the processor executes the program, it reads the three-dimensional bioequivalent dose rate distribution in each direction output by the neural network, constructs and solves the optimization model.

[0058] It should be noted that the objective function is composed of three dose evaluation function terms: dose deviation, dose uniformity, and underdose penalty. These three terms are all evaluation function terms constructed to take into account the influence of the current target dose value and the three-dimensional bioequivalent dose rate distribution on the dose distribution characteristics of the target area, and evaluate the quality of the target area dose distribution from different perspectives.

[0059] It should be noted that the dose deviation term is used to evaluate the overall deviation between the actual biological equivalent dose of each voxel within the target area and the current target dose value. This term encourages the optimization results to make the overall dose of the target area as close as possible to the target dose value, avoiding both insufficient overall dose and unnecessary damage to normal tissues caused by excessive irradiation.

[0060] It should be noted that the dose uniformity term is used to evaluate the consistency of dose distribution within the target area, and the generation of inhibitor hotspots and colds. This term promotes the uniformity of dose distribution among voxels within the target area, thereby ensuring that the target area receives sufficient and consistent bioequivalent dose coverage.

[0061] It should be noted that the underdose penalty is used to impose an additional penalty on voxels within the target area that have not reached the current target dose value. Unlike the dose deviation term, the underdose penalty term emphasizes the rigid requirement of prescription dose coverage, that is, it applies a non-linear amplification penalty to underdose voxels to ensure the coverage of the prescription dose in a clinical sense and avoid local dose insufficiency in the target area that affects the treatment effect.

[0062] It is understandable that the above three items are combined into the overall objective function through weighted coefficients. Users can adjust the weights of each item according to clinical preferences, for example, increasing the weight of the underdosage penalty item when emphasizing prescription coverage, and increasing the weight of the uniformity item when emphasizing dose distribution uniformity. Without departing from the core concept of this application, a hierarchical optimization strategy can also be adopted, optimizing coverage first and then uniformity, or the three items can be combined into a single weighted comprehensive evaluation index.

[0063] Understandably, the preset optimization model has two types of constraints: irradiation time constraints and dose constraints. Irradiation time constraints are used to ensure that the optimized irradiation time is within the physically feasible range; dose constraints are used to protect organs at risk and ensure the lower limit of the target dose; both can be set according to the actual situation.

[0064] At this point, given the direction combination and the current target dose value, the objective function and constraints are submitted to the numerical optimization solver for solution. The solver finds a set of irradiation time allocations that satisfy the optimal objective function and all constraints under the constraints; this is called the solution of the preset optimization model. The irradiation time for this direction combination at the current target dose value is then output.

[0065] It should be noted that, based on the irradiation time of each directional combination obtained in step S40, the three-dimensional total biological equivalent dose distribution corresponding to each combination can be calculated. The calculation method is as follows: multiply the biological equivalent dose rate distribution of each direction by its irradiation time and then sum them up to obtain the cumulative dose of each voxel.

[0066] Then, dose-volume histogram (DVH) analysis is used to assess the dose coverage of the tumor target volume and the achievement of dose limits for organs at risk. Dose coverage can be evaluated using various metrics, such as V100% (the percentage of tumor volume receiving the prescribed dose) and V95% (the percentage of volume receiving 95% of the prescribed dose); volume-dose metrics can also be used, such as D80 (the dose value corresponding to 80% target volume coverage) and D95 ​​(the dose value corresponding to 95% target volume coverage). In clinical practice, target dose coverage usually refers to the percentage of tumor volume that achieves at least the target dose value, commonly using standards such as V100% ≥ 95% or D80 meeting the prescribed dose. Furthermore, dose limits for organs at risk, such as the maximum skin dose, must be assessed to ensure they meet tolerability requirements. The final output recommended irradiation regimen includes the selected directional combination, irradiation time for each direction, and the target dose values ​​used. This output can be directly used for the final validation and execution of the treatment planning system (TPS).

[0067] Understandably, the effects and benefits of this step are: it realizes a fully automated process from multi-directional combination enumeration to optimal solution selection, significantly reducing manual intervention and the number of calculation iterations; at the same time, it ensures that the recommended solutions meet clinical constraints in terms of physical dosimetry and biological equivalent doses, improving the efficiency and reliability of BNCT individualized plan development.

[0068] In this embodiment, by first constructing the total input condition features that integrate anatomical structures and beam direction, and then using a three-dimensional dose rate component prediction neural network and post-processing to output the three-dimensional biological equivalent dose rate distribution in each direction, the time-consuming Monte Carlo simulation is replaced, which significantly improves computational efficiency while ensuring accuracy. Furthermore, by enumerating multi-directional combinations and setting a target dose value search sequence, and using irradiation time as a variable, and introducing dose deviation, uniformity, and underdose penalty terms that take into account the target dose value, an optimization model is constructed and solved, overcoming the defects of insufficient deep dose and poor uniformity of single-directional irradiation. Finally, the optimal scheme is output based on the total dose distribution, realizing fast and accurate optimization of BNCT multi-field irradiation plans.

[0069] Based on the above embodiments, this application provides a feasible implementation method.

[0070] In this embodiment, step S10 includes: preprocessing the three-dimensional voxel data of the computed tomography scan of the target object to obtain the three-dimensional tissue classification code and the three-dimensional target object mask of the target object as the structural input features of the target object; generating a three-dimensional beam depth distribution, a three-dimensional beam radius distribution and a three-dimensional aperture mask per voxel according to the beam injection direction and beam central axis of each candidate irradiation direction as the beam irradiation direction input features; and constructing a multi-channel three-dimensional input tensor according to the target object structural input features and the beam irradiation direction input features as the total input condition features.

[0071] It should be noted that this embodiment first preprocesses the computed tomography scan data of the target object to obtain a three-dimensional tissue classification code and a three-dimensional target object mask.

[0072] It is understood that the three-dimensional voxel data of a target object's computed tomography scan refers to a digital three-dimensional representation of the target object's anatomical structure acquired through a CT scanning device. Each voxel stores a CT value (Henness unit), reflecting the attenuation characteristics of the corresponding tissue to X-rays. Preprocessing involves a series of image processing operations on the raw CT data to convert it into a standardized format suitable for subsequent neural network input. Specific preprocessing operations may include: converting CT values ​​to electron density or tissue type; resampling CT images to standardize voxel sizes; cropping or padding images to a fixed size; and adjusting window width and level to enhance soft tissue contrast. Preprocessing can be performed by an image processing module in a computer, which can be written based on an open-source library.

[0073] It should be noted that 3D tissue classification coding refers to assigning an integer label to each voxel to identify the tissue type to which that voxel belongs (e.g., 0 represents air, 1 represents soft tissue, etc.). This coding can be automatically generated from CT data through threshold segmentation, machine learning classifiers, or deep learning semantic segmentation networks.

[0074] In this embodiment, the three-dimensional tissue classification coding includes five categories: air, soft tissue, bone tissue, skin, and target area. It should be noted that a 3D target object mask is a binary image used to distinguish between voxels inside the target object and voxels in the surrounding air. For example, voxels belonging to the target object (i.e., the patient's body) are marked as 1, while non-target areas such as the surrounding air are marked as 0.

[0075] Understandably, the effect of this step is to convert the original continuous CT values ​​into discrete semantic information, enabling the neural network to directly perceive the differences in the effects of different tissues on neutron transport and dose deposition. At the same time, the target object mask provides spatial range limitation for subsequent beam feature calculations.

[0076] Ultimately, the 3D tissue classification code and the 3D target object mask together constitute the target object structural input features.

[0077] It should be noted that, for each candidate irradiation direction, based on the spatial coordinates of the three-dimensional voxel grid, and combined with the beam inrush direction, beam central axis and beam radius corresponding to the candidate irradiation direction, a three-dimensional beam depth distribution, a three-dimensional beam radius distribution and a three-dimensional beam aperture mask are generated on a voxel-by-voxel basis.

[0078] Among them, the three-dimensional beam depth distribution is used to characterize the projection distance of each voxel relative to the preset reference plane along the beam injection direction; the three-dimensional beam radius distribution is used to characterize the vertical distance of each voxel to the beam center axis; and the three-dimensional beam aperture mask is used to characterize whether each voxel is located within the preset beam radius on the projection plane perpendicular to the beam injection direction.

[0079] Ultimately, the three-dimensional beam depth distribution, the three-dimensional beam radius distribution, and the three-dimensional beam aperture mask together constitute the input characteristics of the beam illumination direction.

[0080] Therefore, each candidate illumination direction corresponds to a multi-channel three-dimensional input tensor containing target object structural information and beam illumination direction information. In a specific embodiment, the candidate illumination directions include nine fixed directions: top, left, right, front, rear, left front oblique, right front oblique, left rear oblique, and right rear oblique, as shown below. Figure 2 As shown in (a) and (b) above. The number and orientation of candidate irradiation directions are not limited to the examples above and can be set according to the structure of the boron neutron capture therapy (BNCT) device, the target placement conditions, and the beam inrush restrictions.

[0081] It should be noted that Ming uses the three-dimensional tissue structure information of the target object and the beam direction geometry information corresponding to each candidate irradiation direction to complete the dose rate prediction. There is no need to perform low particle number, coarse mesh or other forms of Monte Carlo dose calculation for the target object to be predicted and the candidate irradiation direction in advance, which reduces the computation time and computational resources required to generate multi-directional dose data.

[0082] It should be noted that the candidate irradiation direction and the spatial geometric relationship of the beam are characterized by the three-dimensional beam depth distribution, the three-dimensional beam radius distribution and the three-dimensional beam aperture mask. This enables different candidate irradiation directions to be converted into network input features with consistent dimensions, thereby generating three-dimensional dose rate distributions corresponding to multiple candidate irradiation directions in batches through a unified dose prediction process.

[0083] In this embodiment, the three-dimensional dose rate component prediction neural network includes: a three-dimensional boron physical dose rate component prediction model, a three-dimensional nitrogen physical dose rate component prediction model, a three-dimensional hydrogen physical dose rate component prediction model, and a three-dimensional photon physical dose rate component prediction model. It should be noted that the prediction models for each physical dose rate component are implemented using a three-dimensional U-shaped neural network structure with residual modules and attention jump connections; specifically, each single-component prediction model can adopt a three-dimensional encoder-decoder structure containing residual modules and attention jump connections.

[0084] Therefore, step S20 includes: inputting the total input condition features of each candidate irradiation direction into the three-dimensional boron physical dose rate component prediction model, the three-dimensional nitrogen physical dose rate component prediction model, the three-dimensional hydrogen physical dose rate component prediction model, and the three-dimensional photon physical dose rate component prediction model, respectively, to obtain the three-dimensional boron physical dose rate component under each candidate irradiation direction. B Distribution, three-dimensional nitrogen physical dose rate components N Distribution, three-dimensional hydrogen physical dose rate components H Distribution and three-dimensional physical dose rate components of light γ Distribution: Based on the multiple physical dose rate components under each candidate irradiation direction, and combined with the biological effect coefficients, the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction is obtained by adding them voxel by voxel.

[0085] Understandably, for each candidate irradiation direction, the four three-dimensional physical dose rate component distributions are weighted and summed on a voxel-by-voxel basis according to the tissue to which the voxel belongs or the corresponding composite biological effect factor (CBE) and relative biological effect factor (RBE) to obtain the three-dimensional biological equivalent dose rate distribution corresponding to the candidate irradiation direction.

[0086] For any voxel i , No. j Candidate irradiation directions in voxels i Bioequivalent dose rate at: j bio ( i ) =CBE ( i )× B,j ( i ) +RBE H × H,j ( i ) +RBE N × N,j ( i )+ RBE γ × γ,j ( i (1).

[0087] In the formula, j Indicates the candidate irradiation direction. i This represents a voxel in a three-dimensional voxel mesh. CBE ( iThe appropriate value can be taken based on the tissue type to which the voxel belongs. RBE H , RBE N and RBE γ These represent the relative biological effect factors corresponding to the hydrogen, nitrogen, and photon dose components, respectively.

[0088] During the training phase, the four physical dose rate components can be normalized to reduce the impact of differences in numerical scale between different dose rate components on the training process. The loss function can be a weighted smooth L1 loss function, with different weights set for the target area, skin area, bone tissue area, and high-dose area; at the same time, a three-dimensional target object mask is used to apply hard constraints to the air area outside the target object, so that the predicted dose rate of the voxels outside the target object is zero or close to zero.

[0089] A total of 144 test results were predicted for 16 target objects in the test set under 9 different illumination directions. Figure 3 Box plots show the mean absolute error (MAE) between the predicted and reference results for the target object in different tissue regions across 144 test results in the test set. Figure 4 A box plot of the mean absolute percentage error (MAPE) is shown. For those containing... N For each evaluation voxel region, MAE and MAPE can be represented as follows: (2); (3).

[0090] In the formula, and They represent the first i The predicted bioequivalent dose rate values ​​of individual elements and the reference values ​​of bioequivalent dose rates calculated by Monte Carlo (MC) are used. N This indicates the number of voxels in the region.

[0091] In addition, there is the MAPE10 index. It indicates that MAPE is calculated only on voxels within the corresponding tissue region where the reference value is not less than 10% of the maximum reference value in that region, and is used to reduce the amplification effect of low-dose voxels on relative error.

[0092] Depend on Figure 3 , 4 The results show that the average MAPE of the tumor target area in the 144 test results was 1.35%, indicating that the model has high prediction accuracy in the high-dose voxel area of ​​the target area. The average overall skin MAPE was 3.34%, while the MAPE10 was 1.35%, indicating that low-dose skin voxels amplify the percentage error, but the error is significantly lower in the medium-to-high dose areas that are of greater clinical concern.

[0093] Furthermore, in an exemplary result, the mean MAE for the tumor target area was 0.10 Gy(w) / min, and the MAPE was 1.35%; the mean MAE for the skin area was 0.02 Gy(w) / min, the standard MAPE was 3.34%, and the MAPE10 was 1.35%. This result indicates that low-dose voxels in the skin area amplify the standard MAPE, while the predicted results for high-dose areas in the skin show good agreement with the reference results. The above values ​​are only used to illustrate the feasibility of the prediction process and do not constitute a limitation on the prediction accuracy or dose units of this application.

[0094] It should be noted that this application predicts the three-dimensional physical dose rate component distributions of boron, nitrogen, hydrogen, and photons separately, and obtains the three-dimensional biological equivalent dose rate distribution through the corresponding biological effect coefficients. This preserves the interpretability of different dose sources and their biological weighting processes, facilitating component-level error localization, dose contribution analysis, and dosimetric evaluation. Preferably, employing four independent single-component prediction models can reduce the impact of numerical scale differences between different dose components on the joint training process and improve the accuracy of smaller dose components. In this embodiment, step S30 includes: setting a maximum number of irradiation fields, enumerating all direction combinations that meet the maximum number of irradiation fields limit from multiple preset candidate irradiation directions; setting a target dose value search sequence for each direction combination, the target dose value search sequence containing multiple target dose values ​​arranged in descending order.

[0095] It should be noted that the maximum number of irradiation fields refers to the upper limit of the number of neutron beam irradiation directions that can be used simultaneously in an irradiation scheme, denoted as N max The setting of this parameter is based on the following considerations: On the one hand, increasing the number of irradiation fields can improve the uniformity of target area dose and reduce the dose attenuation effect of deep tumors; however, on the other hand, too many fields can lead to prolonged treatment time, decreased patient positional stability, and increased complexity of the optimization problem. Therefore, by setting... N max This parameter can strike a balance between dosimetric benefits and feasibility, with a typical value ranging from 2 to 4. As a constraint on the enumeration process, it limits the upper limit of the number of fields for all combinations of directions to be evaluated.

[0096] Specifically, let the set of candidate illumination directions be... Ω={1,2,…,K} ,from Ω Enumerating candidate illumination direction combinations S Furthermore, the number of candidate irradiation directions included in combination S does not exceed the preset maximum number of irradiation fields. N max .For example, K =9, Nmax =3.

[0097] At this point, the aforementioned three-dimensional bioequivalent dose rate distribution corresponding to all candidate irradiation directions will be used as input for subsequent multi-field irradiation time optimization. For the current combination of candidate irradiation directions... S={s 1 ,s 2 ,…,s m } Define the irradiation time vector t =(t 1 ,t 2 ,…,t m ) T ,pass t j Indicates the first in the combination j The irradiation time for each candidate irradiation direction. Since the neural network has given the bioequivalent dose rate of each candidate irradiation direction across all three-dimensional voxels, for any voxel... i The total bioequivalent dose after multi-directional irradiation for: (4).

[0098] Therefore, the neural network prediction results serve as the coefficient matrix in the multi-field optimization model, connecting irradiation time with the total bioequivalent dose for each voxel. Using a neural network to generate the bioequivalent dose rate under each irradiation direction avoids the time-consuming Monte Carlo dose calculation.

[0099] Understandably, the current target dose value is set for each candidate irradiation direction combination. Let the upper limit of the target dose value be... The lower limit of the target dose value is The program searches for target dose values ​​in descending order. To balance search speed and resolution, a coarse search can be performed using a first step size. After determining the feasible and infeasible transition ranges, a fine search can then be performed using a second step size smaller than the first step size.

[0100] It is important to emphasize that the target dose value search sequence is set for each directional combination. However, in practical applications, all directional combinations can share the same set of decreasing sequences, or customized sequences can be made based on combination characteristics (such as field number and geometric symmetry). The advantage of shared sequences is that they simplify implementation, while the advantage of customized sequences is that they allow for a more refined search for high-quality combinations.

[0101] Furthermore, this application provides a specific preset optimization model.

[0102] For the current candidate irradiation direction combination and the current target dose value Based on the three-dimensional biological equivalent dose rate distribution under various irradiation directions, an objective function for optimizing irradiation time is constructed. J(t) .

[0103] In one embodiment, the objective function includes a target dose deviation term, a target dose uniformity term, and a target underdose penalty term: (5).

[0104] The above items can be defined as follows, by way of example: (6); (7); (8).

[0105] In the formula, It is the first in the target area i Total bioequivalent dose of individual pigments N Individual factors, It is the target dose value. It is the average of the total bioequivalent dose of all voxels within the target area. , and The weights are non-negative. From equation (4), it can be seen that the weights of each voxel are... All were obtained by linearly combining the directional dose rate predicted by the neural network with the irradiation time. J(t) It is the irradiation time vector t The function. This is used to ensure that the total bioequivalent dose of each voxel in the target area is close to the current target dose value. Used to reduce the dispersion of total biological equivalent dose within the target area; Used to target voxels that are below the current target dose value.

[0106] Therefore, the three-dimensional biological equivalent dose rate distribution output by the neural network directly participates in each objective function calculation, gradient update and constraint calculation through equation (4), forming the basic input for irradiation time optimization.

[0107] Specifically, the irradiation time optimization constraints include: the irradiation time for each candidate irradiation direction is non-negative; the sum of the irradiation times for all candidate irradiation directions does not exceed the preset total irradiation time limit; and each restricted voxel or corresponding dose index in the skin, normal tissue, and other hazardous sites does not exceed the preset dose limit.

[0108] In this embodiment, step S40 includes: selecting the maximum value in the target dose value search sequence under the current direction combination as the current target dose value; combining the three-dimensional biological equivalent dose rate distribution, using a sequential least squares programming algorithm to solve the preset optimization model under the current target dose value condition; calculating the three-dimensional total biological equivalent dose distribution based on the irradiation time of each direction obtained from the solution; when the current three-dimensional total biological equivalent dose distribution meets the target dose coverage requirement, determining the current direction combination as a feasible combination, and outputting the irradiation time of the feasible combination; when the current three-dimensional total biological equivalent dose distribution does not meet the target dose coverage requirement, taking the next target dose value in the search sequence as the current target dose value according to the decreasing order, and executing the step of solving the preset optimization model; if all target dose values ​​in the search sequence do not meet the target dose coverage requirement, determining the current direction combination as an infeasible direction combination.

[0109] Understandably, the core task of step S40 is to solve the preset optimization model for each directional combination enumerated in step S30, combined with the target dose value search sequence, in order to determine the highest feasible prescription dose and corresponding irradiation time for that directional combination under the current constraints.

[0110] The specific solution logic is as follows: For the current direction combination being processed, the system first selects the maximum value from the target dose value search sequence for that combination and uses it as the current target dose value. The intention of this operation is to attempt to irradiate with the highest prescribed dose. If the highest value is feasible, then the highest feasible prescription for that combination is naturally obtained, without needing to continue searching downwards.

[0111] Subsequently, the system, combining the three-dimensional biological equivalent dose rate distribution obtained in step S20, employs a sequential least squares programming algorithm to solve the preset optimization model under the current target dose value. The goal of the solution is to obtain a set of irradiation times in each direction, such that the objective function is optimal and satisfies all constraints.

[0112] If the current target dose value is feasible (i.e., the algorithm finds an irradiation time allocation that satisfies all constraints, and the three-dimensional total biological equivalent dose distribution of the target area meets the target dose coverage requirement), the system outputs this irradiation time as the irradiation time for the current direction combination and stops further searching for that direction combination. This is because the search sequence is arranged in descending order, and the current target dose value is already the highest feasible point in the sequence; continuing to search downwards is meaningless.

[0113] If the current target dose value cannot be solved (i.e., the algorithm finds an irradiation time allocation that satisfies all constraints, but the three-dimensional total biological equivalent dose distribution of the target area does not meet the target dose coverage requirement), the system selects the next target dose value in the target dose value search sequence as the current target dose value according to the descending order, and re-solves the preset optimization model. This process continues until a solution is found for the first time or the entire search sequence is traversed.

[0114] If finding the minimum value in the target dose value search sequence is still not feasible, the system determines that the current direction combination is not feasible, that is, the combination cannot meet the constraint conditions even at the lowest prescription dose, and is therefore discarded.

[0115] It should be noted that this embodiment uses the Sequential Least Squares Programming (SLSQP) algorithm to solve the irradiation time optimization problem under the current candidate irradiation direction combination and the current target dose value. During the iteration process, the SLSQP algorithm constructs a quadratic approximation of the objective function and linearizes the constraints. By continuously solving the quadratic programming subproblems to update the irradiation time, it obtains candidate irradiation time solutions that satisfy the preset convergence conditions and constraint tolerances.

[0116] Understandably, the validation step is an independence check of the solution results of the optimization model, ensuring that the final output solution actually achieves the coverage required by clinical practice.

[0117] In one feasible implementation, the target dose coverage requirement is that 80% of the volumetric dose of the target target area determined based on the current three-dimensional total bioequivalent dose distribution is not lower than the current target dose value.

[0118] It should be noted that the 80% volume dose (D80) refers to the lowest dose value received by 80% of the target volume within the target area, i.e., the dose value corresponding to 80% of the volume on the target dose-volume histogram. This indicator reflects the dose level achieved by the vast majority of the target volume and is an important parameter for evaluating the quality of prescribed dose coverage.

[0119] Understandably, if the D80 of the target area is not lower than the current target dose value, the irradiation time corresponding to the current target dose value is output as the irradiation time for the current directional combination, and further searching for the current directional combination stops. If the D80 of the target area is lower than the current target dose value, it means that although the optimization model determines that there is a solution, more than 20% of the target area volume in the actual dose distribution still does not reach the prescribed dose, and the coverage does not meet the strict requirements. In this case, the next target dose value in the descending order search sequence is used as the current target dose value, and the preset optimization model is solved again until a solution is found that makes the three-dimensional total biological equivalent dose distribution of the target area meet the target dose coverage. The target dose value corresponding to this solution is the optimal feasible prescribed dose value for the directional combination under the current constraints, and the corresponding irradiation time is the optimal irradiation time allocation for the directional combination. If the solution with the minimum value in the target dose value search sequence still does not meet the target coverage requirement, the current directional combination is determined to be infeasible.

[0120] It should be noted that the V100 standard mentioned earlier, which requires 100% of the target volume to reach the prescribed dose, is too stringent. In BNCT, the distribution of boron in tumors is naturally uneven, and edges and cold spots are unavoidable. Forcing V100=100% would force the optimizer to overdose, endangering the organ. D80 only considers the dose value obtained from 80% of the target volume, allowing for a few cold spots with underdose, which is more in line with the physical reality of BNCT.

[0121] Therefore, after step S40, the method further includes: for all the obtained feasible solutions, prioritizing all feasible solutions according to the feasible target dose value, the uniformity of the target area dose, the dose of the critical site and the total irradiation time, and outputting the feasible solution with the highest priority as the recommended irradiation solution.

[0122] Understandably, after all candidate irradiation direction combinations have completed the target dose value search and irradiation time optimization, the system summarizes all feasible solutions. Recommended solutions can first be sorted according to the highest feasible target dose value within a preset range; when multiple solutions have the same highest feasible target dose value within the preset range or are within the preset tolerance range, the solution with the lower target homogeneity index (HI) is given priority; when the HIs are still similar, the dose level of the critical site, the total irradiation time, and the normalized target function value can be further compared.

[0123] The system output includes recommended candidate irradiation direction combinations, the corresponding irradiation time for each candidate irradiation direction, the highest feasible target dose value within the preset range, dosimetric indicators of the target area, dose indicators of critical sites, total irradiation time, and feasibility markers. These outputs can be used for irradiation plan design, scheme comparison, and subsequent high-precision dose verification.

[0124] In an exemplary verification, the optimization results of multi-field irradiation for 16 target objects showed that as the number of irradiation fields increased from single to double and then to triple, the number of target objects reaching higher target dose values ​​increased, and the overall HI of the target area decreased. This indicates that, under the premise of meeting the dose constraints of the critical site and the total irradiation time constraints, multi-field irradiation can improve the dose coverage and dose uniformity of the target area.

[0125] Figure 5 This paper presents the optimal dose-to-visibility (DVH) of a target object under different irradiation fields. The single-field irradiation field shows a relatively gentle decrease in the target area curve, indicating a large dispersion in dose distribution within the target area. The two-field irradiation field shifts the target area curve towards higher doses, indicating an overall increase in the target area dose level. The three-field irradiation field curve becomes even steeper, reaching an exemplary target dose value of 40 Gy(w), indicating further improvement in target area coverage and dose uniformity. Increasing the field number from single to three reduces the target object's dose-to-visibility (HVH) from 0.84 to 0.24, a reduction of 71.43%, demonstrating that the dose contribution of different incident paths to the target volume (GTV) is complementary, significantly improving the low-dose region and dose inhomogeneity issues under single-field irradiation. In this example, the recommended irradiation field includes three directions: top, right rear oblique, and rear, with corresponding irradiation times of 14.02 min, 7.92 min, and 6.52 min, respectively. The directions, times, and target dose values ​​mentioned above are merely exemplary parameters.

[0126] To verify the reliability of multi-field optimization using neural network prediction results, a reference dose verification can be performed on the recommended scheme. Therefore, after obtaining the recommended irradiation scheme, the process also includes: obtaining the reference three-dimensional total biological equivalent dose distribution for the corresponding candidate irradiation directions calculated by Monte Carlo based on the direction combination and irradiation time of the recommended irradiation scheme; and verifying the recommended irradiation scheme based on the reference three-dimensional total biological equivalent dose distribution and the three-dimensional total biological equivalent dose distribution.

[0127] It should be noted that the review here mainly refers to the comparison of relevant indicators under two distributions. For example, dosimetric indicators of the target area and the site of danger.

[0128] In one exemplary verification result, the relative error of the maximum dose at the critical site was less than 1%, and all were below the exemplary dose limits of 11 Gy(w) and 12.5 Gy(w); the relative error of the target area D80 was 0.65%, and the relative deviation of HI was 0.54%. These results indicate that using the three-dimensional bioequivalent dose rate distribution predicted by a neural network to optimize the combination of candidate irradiation directions and irradiation time can yield multi-field schemes with good consistency with the reference dose verification results. The above results are only used to illustrate the feasibility and technical effectiveness of the calculation process in this application and do not constitute a limitation on specific target dose values, dose limits, or evaluation thresholds.

[0129] It should be noted that the steps of this embodiment are the same as those in this application. Figure 6 Detailed process Figure 1 The correspondence is one-to-one, so I won't go into details.

[0130] Understandably, this application separates the discrete direction selection problem from the continuous irradiation time optimization problem by enumerating candidate irradiation direction combinations that satisfy the field number limit; then, by combining the target dose value search and sequential least squares programming method, the irradiation time is optimized for each direction combination separately, avoiding the direct solution of the mixed optimization problem containing direction selection variables and irradiation time variables, and is able to determine the highest feasible target dose value within the preset range corresponding to each direction combination.

[0131] It is understandable that this application does not judge the merits of a scheme solely based on the objective function value returned by the optimization program. Instead, after optimizing the irradiation time, it re-verifies the target dose coverage of the target area based on the complete three-dimensional total biological equivalent dose distribution, thereby reducing misjudgments of scheme feasibility caused by optimization sampling, approximate constraints, or local solution results.

[0132] Understandably, this application can comprehensively evaluate feasible schemes based on dose-volume histograms (DVH) and region-of-interest dose indices, and output target dose indices, as well as results such as dose to critical sites, total irradiation time, irradiation direction combinations, and irradiation time in each direction, providing quantitative basis for comparison and design of multi-field irradiation schemes.

[0133] In addition, this application also proposes an embodiment of a BNCT multi-field irradiation planning optimization system. The system includes: The feature construction module is used to construct a total input condition feature that includes the target object structure input feature and the beam irradiation direction input feature for multiple preset candidate irradiation directions, combined with the three-dimensional voxel data of the target object.

[0134] The dose rate component prediction module is used to input the total input condition features of each candidate irradiation direction into the three-dimensional dose rate component prediction neural network to obtain the distribution of multiple three-dimensional physical dose rate components under each candidate irradiation direction, so as to obtain the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction.

[0135] The combination enumeration and search module is used to enumerate multiple direction combinations from multiple preset candidate irradiation directions and set a target dose value search sequence for each direction combination.

[0136] The irradiation time optimization and feasibility verification module is used to search for each target dose value in the target dose value search sequence, and solve the preset optimization model one by one in combination with the three-dimensional bioequivalent dose rate distribution. The module calculates the three-dimensional total bioequivalent dose distribution based on the irradiation time in each direction obtained from the solution, until the solution of the preset optimization model makes the three-dimensional total bioequivalent dose distribution of the target target area meet the target dose coverage requirement, so as to output the feasible combination in each direction combination and the corresponding irradiation time and target dose value.

[0137] It should be understood that the above system is used to execute the methods in the above embodiments. The corresponding program modules in the system are similar in implementation principle and technical effect to those described in the above methods. The working process of the system can be referred to the corresponding process in the above methods, and will not be repeated here.

[0138] Based on the methods in the above embodiments, please refer to Figure 7 This application provides an electronic device that may include a processor, a communications interface, a memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions stored in the memory to execute the methods described in the above embodiments.

[0139] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0140] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0141] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0142] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0143] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0144] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0145] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0146] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for optimizing BNCT multi-field irradiation plans, characterized in that, include: Step S10: For multiple preset candidate irradiation directions, construct a total input condition feature that includes the target object structure input feature and the beam irradiation direction input feature by combining the three-dimensional voxel data of the target object; Step S20: Input the total input condition features of each candidate irradiation direction into multiple three-dimensional dose rate component prediction neural networks to obtain the three-dimensional physical dose rate component distribution under each candidate irradiation direction, so as to obtain the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction. Step S30: Enumerate multiple direction combinations from the multiple preset candidate irradiation directions, and set a target dose value search sequence for each direction combination; Step S40: Search for each target dose value in the target dose value search sequence, and solve the preset optimization model one by one in combination with the three-dimensional bioequivalent dose rate distribution until the solution of the preset optimization model makes the three-dimensional total bioequivalent dose distribution of the target target area meet the target dose coverage requirement, so as to output the feasible combination in each direction combination and the corresponding irradiation time and target dose value. The preset optimization model is as follows: taking the irradiation time of each irradiation direction in the directional combination as the optimization variable, an objective function is established including a dose deviation term, a dose uniformity term, and an underdose penalty term for the target area, and irradiation time constraints and dose constraints are set; the dose deviation term, dose uniformity term, and underdose penalty term are all dose evaluation function terms constructed by taking into account the influence of the current target dose value and the three-dimensional biological equivalent dose rate distribution on the dose distribution characteristics of the target area.

2. The BNCT multi-field irradiation planning optimization method as described in claim 1, characterized in that, Step S10 includes: Preprocessing is performed on the three-dimensional voxel data of the computed tomography scan of the target object to obtain the three-dimensional tissue classification code and the three-dimensional target object mask of the target object, which are used as structural input features of the target object. Based on the beam injection direction and beam central axis of each candidate irradiation direction, a three-dimensional beam depth distribution, a three-dimensional beam radius distribution, and a three-dimensional beam aperture mask are generated on a voxel-by-voxel basis as input features for the beam irradiation direction. A multi-channel three-dimensional input tensor is constructed based on the target object structure input features and the beam irradiation direction input features, serving as the total input condition features.

3. The BNCT multi-field irradiation planning optimization method as described in claim 1, characterized in that, Multiple three-dimensional dose rate component prediction neural networks include: three-dimensional boron physical dose rate component prediction models, three-dimensional nitrogen physical dose rate component prediction models, three-dimensional hydrogen physical dose rate component prediction models, and three-dimensional photon physical dose rate component prediction models; each physical dose rate component prediction model is implemented through a three-dimensional U-shaped neural network structure with residual modules and attention jump connections; Step S20 includes: The total input condition features of each candidate irradiation direction are respectively input into the three-dimensional boron physical dose rate component prediction model, the three-dimensional nitrogen physical dose rate component prediction model, the three-dimensional hydrogen physical dose rate component prediction model and the three-dimensional photon physical dose rate component prediction model to obtain the boron physical dose rate component distribution, nitrogen physical dose rate component distribution, hydrogen physical dose rate component distribution and photon physical dose rate component distribution under each candidate irradiation direction. Based on the multiple physical dose rate components under each candidate irradiation direction, and by adding the biological effect coefficients voxel by voxel, the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction is obtained.

4. The BNCT multi-field irradiation planning optimization method as described in claim 1, characterized in that, Step S30 includes: Set a maximum number of irradiation fields, and enumerate all direction combinations that satisfy the maximum number of irradiation fields limit from the plurality of preset candidate irradiation directions; A target dose value search sequence is set for each of the directional combinations, the target dose value search sequence containing multiple target dose values ​​arranged in descending order.

5. The BNCT multi-field irradiation planning optimization method as described in claim 1, characterized in that, Step S40 includes: Select the maximum value from the target dose value search sequence under the current direction combination as the current target dose value; Combining the three-dimensional biological equivalent dose rate distribution, a sequential least squares programming algorithm is used to solve the preset optimization model under the current target dose value; The three-dimensional total biological equivalent dose distribution is calculated based on the irradiation time in each direction obtained from the solution. When the current three-dimensional total biological equivalent dose distribution meets the target dose coverage requirement, the current directional combination is determined to be a feasible combination, and the irradiation time of the feasible combination is output. When the current three-dimensional total biological equivalent dose distribution does not meet the target dose coverage requirement, the next target dose value in the search sequence is taken as the current target dose value according to the decreasing order, and the step of solving the preset optimization model is executed. If none of the target dose values ​​in the search sequence meet the target dose coverage requirement, the current directional combination is determined to be an infeasible directional combination.

6. The BNCT multi-field irradiation planning optimization method as described in claim 5, characterized in that, The target dose coverage requirement is that 80% of the volumetric dose of the target target area determined based on the current three-dimensional total biological equivalent dose distribution is not lower than the current target dose value.

7. The BNCT multi-field irradiation planning optimization method as described in claim 1, characterized in that, Step S40 is followed by: For all the feasible solutions obtained, the feasible solutions are prioritized according to the feasible target dose value, the dose uniformity of the target area, the dose of the critical site, and the total irradiation time. The feasible solution with the highest priority is output as the recommended irradiation solution. Based on the recommended irradiation scheme and irradiation time, the corresponding reference three-dimensional total biological equivalent dose distribution was obtained by Monte Carlo calculation. The recommended irradiation regimen was reviewed based on the reference three-dimensional total biological equivalent dose distribution and the three-dimensional total biological equivalent dose distribution.

8. A BNCT multi-field irradiation planning optimization system, characterized in that, The system includes: The feature construction module is used to construct a total input condition feature, including the target object structure input feature and the beam irradiation direction input feature, based on multiple preset candidate irradiation directions and the three-dimensional voxel data of the target object. The dose rate component prediction module is used to input the total input condition features of each candidate irradiation direction into the three-dimensional dose rate component prediction neural network to obtain the distribution of multiple three-dimensional physical dose rate components under each candidate irradiation direction, so as to obtain the three-dimensional biological equivalent dose rate distribution of each candidate irradiation direction. The combined enumeration and search module is used to enumerate multiple direction combinations from the multiple preset candidate irradiation directions, and set a target dose value search sequence for each direction combination; The irradiation time optimization and feasibility verification module is used to search for each target dose value in the target dose value search sequence, and solve the preset optimization model one by one in combination with the three-dimensional bioequivalent dose rate distribution. The module calculates the three-dimensional total bioequivalent dose distribution based on the irradiation time in each direction obtained from the solution, until the solution of the preset optimization model makes the three-dimensional total bioequivalent dose distribution of the target target area meet the target dose coverage requirement, so as to output the feasible combination in each direction combination and the corresponding irradiation time and target dose value.

9. An electronic device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the BNCT multi-field irradiation planning optimization method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the BNCT multi-field irradiation planning optimization method as described in any one of claims 1 to 7.

Images