Radiotherapy plan making method based on deep learning and related equipment

Abstract

The invention discloses a radiotherapy plan making method based on deep learning and related equipment, and relates to the technical field of radiotherapy, and the method comprises the steps: generating machine parameters needed by a therapy machine based on a machine parameter prediction model according to initial dose distribution; according to the generated machine parameters, obtaining first dose distribution based on a rapid dose calculation model of transfer learning; automatically adjusting machine parameters according to the difference between the first dose distribution and the clinical prescription dose distribution and an organ-endangering exposure dose limit value; and based on the rapid dose calculation model, in combination with the adjusted machine parameters, obtaining final dose distribution, and comprehensively considering multiple factors such as the integrating degree of the target dose distribution and the clinical prescription dose, the dose control endangering organs, the physical feasibility of the machine parameters and the like, so as to generate an efficient, accurate, safe and feasible radiotherapy plan.

Description

Technical Field

[0001] The present invention relates to the technical field of radiotherapy, and in particular to a radiotherapy plan making method and related equipment based on deep learning. Background Art

[0002] Particle therapy (proton, carbon ion) has significant physical property advantages compared with traditional photon radiotherapy. Its energy release rapidly decreases after reaching the tumor site, which can maximize the avoidance of damage to healthy tissues. This property makes it particularly suitable for the treatment of tumors near important organs. To achieve a precise treatment dose distribution, most current particle therapy systems adopt a pencil beam scanning treatment method. In the pursuit of better treatment effects, some scanning systems use multi-leaf collimators to trim the edge dose of the scanning beam spot in order to optimize the dose distribution and improve the treatment accuracy.

[0003] Although particle therapy has advantages in physical properties, there are still many problems to be solved in the actual radiotherapy plan making: for the plan making of complex tumor patients, the existing plan making schemes generally take a long time, which not only affects the timeliness of patients receiving treatment, but also reduces the utilization efficiency of medical resources to a certain extent. At the same time, the dose distribution presented by the treatment plan is not ideal enough to accurately meet the dual requirements of killing tumor cells and protecting healthy tissues, resulting in limited improvement of the treatment effect. Especially for treatment systems with beam limiting elements, the existing schemes face the dilemma of complex plan making. The cumbersome making process further prolongs the time for plan formulation, and at the same time, the dose distribution of its plan still cannot reach an ideal state, and the plan quality needs to be improved urgently. These problems restrict the wide application of particle therapy technology in clinical practice and the further optimization of the treatment effect, and a more efficient and accurate radiotherapy plan making scheme is needed to improve the current situation. Summary of the Invention

[0004] The purpose of the present invention is to provide a radiotherapy plan making method and related equipment based on deep learning, comprehensively considering multiple factors such as the matching degree between the target area dose distribution and the clinical prescription dose, the dose control of critical organs, and the physical feasibility of machine parameters, so as to generate an efficient, accurate, safe and feasible radiotherapy plan.

[0005] The purpose of the present invention is achieved by the following technical solutions:

[0006] In a first aspect, a radiotherapy plan making method based on deep learning, the method includes:

[0007] Based on the preliminary dose distribution and the machine parameter prediction model, generate the machine parameters required by the treatment machine;

[0008] Based on the generated machine parameters, a fast dose calculation model based on transfer learning is used to obtain a first dose distribution;

[0009] According to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk, the machine parameters are automatically adjusted;

[0010] Based on the fast dose calculation model and combined with the adjusted machine parameters, the final dose distribution is obtained.

[0011] Preferably, the method for obtaining the preliminary dose distribution includes:

[0012] Based on an automatic segmentation network model, according to the simulated CT of the patient, the target area and the organs at risk are segmented to obtain the contoured site information;

[0013] According to the simulated CT of the patient and the contoured site information, a preliminary dose distribution is generated using a dose prediction network model.

[0014] Preferably, the method further includes:

[0015] Evaluating historical samples and selecting training samples for each model according to the evaluation results; including:

[0016] Obtaining scores for multiple indicators of each sample; the indicators include the dose uniformity calculated by the treatment plan, the target coverage, the target conformity, and the dose control degree of healthy organs;

[0017] According to the scores of each indicator in the sample, the comprehensive score of the sample is obtained;

[0018] According to the comprehensive score, the training samples for each model are selected.

[0019] Preferably, the step of segmenting the target area and the organs at risk according to the simulated CT of the patient based on the automatic segmentation network to obtain the contoured site information includes:

[0020] Training an automatic segmentation network model for the target area and the organs at risk through training samples; the automatic segmentation network is a U-net network;

[0021] Through the trained U-net network, combined with the simulated CT of the patient, the target area and the organs at risk are segmented to obtain a preliminary segmentation result;

[0022] The preliminary segmentation result is confirmed and corrected; the final contoured site information is obtained.

[0023] Preferably, after the preliminary segmentation result is confirmed and corrected, the following steps are further included:

[0024] Performing a difference analysis between the corrected contoured site information and the original automatic segmentation result to generate a corrected region mask;

[0025] Based on the corrected region mask, update the parameters of the automatic segmentation network model through a dynamic weight transfer learning algorithm.

[0026] Preferably, the dose prediction network model is a generative adversarial network structure, including a discriminator and a generator, which generates a preliminary dose distribution according to the simulated CT of the patient and the delineated region, including:

[0027] The generator receives the simulated CT and the delineated region information and generates a preliminary dose distribution;

[0028] The discriminator evaluates the overall dose rationality and the target - organ at risk boundary gradient;

[0029] Introduce dosimetric prior knowledge into the loss function for training the dose prediction network, including DVH curve matching loss, dose gradient alignment loss, and organ dose limit constraint terms.

[0030] Preferably, the machine parameters include gantry angle, treatment couch angle, scan point positions under each angle combination, and the particle beam intensity required for each scan point.

[0031] Preferably, according to the preliminary dose distribution, generate the machine parameters required by the treatment machine based on the machine parameter prediction model, including:

[0032] Build a machine parameter prediction model using the network architecture of reinforcement learning, optimize the machine parameters into a Markov decision process, and the state includes the preliminary dose distribution and the current machine parameters;

[0033] The reward function combines the dose target achievement degree and parameter feasibility, and during the training process, punish the machine parameters that violate the physical constraints through the feasibility loss function;

[0034] Simulate dose calculation uncertainty by injecting Monte Carlo noise during training.

[0035] Preferably, the feasibility loss function is obtained by the following formula:

[0036]

[0037] L feasible is the feasibility loss value, is the weight coefficient calculated according to the correlation between the machine parameters and the organ importance; MA h is the actual value of the h - th machine parameter; MA h,max is the maximum value allowed for the h - th machine parameter.

[0038] Preferably, the fast dose calculation model is a pre-trained Transformer architecture model; obtaining the first dose distribution based on the generated machine parameters and the fast dose calculation model based on transfer learning includes:

[0039] Serializing the machine parameters and adding positional embeddings, and inputting them into the dose calculation network of the Transformer architecture; the dose calculation network generates a dose distribution corresponding to the machine parameters through the multi-head self-attention mechanism and the causal self-attention mechanism;

[0040] On the basis of the pre-trained model, fine-tuning is performed through a dynamic weight freezing strategy.

[0041] Preferably, automatically adjusting the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk includes:

[0042] Constructing a multi-objective loss function through the target dose gap, organ limit violation and machine parameter feasibility;

[0043] Iteratively optimizing the parameters through gradient backpropagation or a hybrid strategy of gradient backpropagation combined with an evolutionary algorithm until the preset maximum number of iterations is reached or the change in the multi-objective loss function value is less than the preset threshold, and outputting a set of machine parameters that meet the clinical dose requirements.

[0044] Preferably, the multi-objective loss function is obtained by the following formula:

[0045]

[0046] where L mu is the multi-objective loss function; D pred is the predicted target dose distribution, D 处方 is the clinical prescription dose; e represents different organs at risk; is the maximum dose of the e-th organ at risk predicted by the model; is the clinically specified maximum tolerance dose of the e-th organ at risk; L MA is the loss term related to the physical feasibility of the machine parameters; q1, q2, and q3 are dynamic weights.

[0047] In a second aspect, the present application provides a radiotherapy plan making device based on deep learning, and the device includes:

[0048] A machine parameter generation module, configured to generate machine parameters required for a treatment machine according to a preliminary dose distribution and based on a machine parameter prediction model;

[0049] A first dose distribution acquisition module, configured to obtain a first dose distribution according to the generated machine parameters and based on a fast dose calculation model based on transfer learning;

[0050] A machine parameter adjustment module, configured to automatically adjust machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk;

[0051] A final dose distribution acquisition module, configured to obtain a final dose distribution based on the fast dose calculation model and in combination with the adjusted machine parameters.

[0052] In a third aspect, the present application provides a radiotherapy system, which includes:

[0053] A particle accelerator, configured to generate a high-energy particle beam;

[0054] The aforementioned radiotherapy plan making device, configured to calculate the dose of the radiotherapy plan.

[0055] In a fourth aspect, the present application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method described above in the present application or the functions of the device described in the present application are implemented.

[0056] In a fifth aspect, the present application provides a computer-readable storage medium, which stores computer instructions. When a computer reads the computer instructions, the computer executes the steps of the method described above in the present application.

[0057] Compared with the prior art, the beneficial effects of the present invention at least include: using a deep learning model to quickly generate machine parameters and dose distributions, greatly shortening the treatment plan making time; compared with the existing treatment plan making solutions, reducing the time-consuming for making plans for complex tumor patients, enabling patients to receive treatment more promptly, and improving the utilization efficiency of medical resources. At the same time, the fast dose calculation model is based on transfer learning, without the need to perform a large amount of calculations from scratch, accelerating the dose calculation speed and optimizing the entire treatment plan formulation process. By iteratively optimizing the machine parameters, the final dose distribution is more consistent with the clinical prescription dose distribution, better meeting the dual requirements of killing tumor cells and protecting healthy tissues. In particle therapy, the energy can be more precisely concentrated on the tumor site, reducing damage to surrounding healthy tissues. Especially for the treatment of tumors near important organs, the treatment accuracy can be significantly improved, avoiding treatment deviations caused by unreasonable dose distributions; aiming at the problems of complex plan making and unsatisfactory dose distributions in existing treatment systems with limited beam elements, by constructing a multi-objective loss function to comprehensively consider the dose gap in the target area, organ limit violations, and the feasibility of machine parameters, the machine parameters are comprehensively optimized, so that the treatment plan not only meets the clinical dose requirements but also ensures the rationality and feasibility of the plan, improving the plan quality, and contributing to the wide application of particle therapy technology in clinical practice and the further optimization of treatment effects. Description of the Drawings

[0058] Figure 1 It is a schematic diagram of a radiotherapy plan making method based on deep learning according to an embodiment of the present invention;

[0059] Figure 2 It is a schematic diagram of a method for obtaining a preliminary dose distribution according to an embodiment of the present invention;

[0060] Figure 3 It is a schematic diagram of a method for selecting model training samples according to an embodiment of the present invention;

[0061] Figure 4 It is a schematic diagram of a radiotherapy plan making device based on deep learning according to an embodiment of the present invention. Detailed implementation manners

[0062] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this invention will be more complete and comprehensive, and will fully convey the concept of the example embodiments to those skilled in the art. Like reference numerals in the figures denote the same or similar structures, and thus their repetitive description will be omitted.

[0063] The words expressing positions and directions described in the present invention are all illustrated by taking the accompanying drawings as examples, but can be changed according to needs, and all the changes made are included in the protection scope of the present invention.

[0064] Refer to Figure 1 : The embodiments of the present application provide a radiotherapy plan making method based on deep learning, and the method includes:

[0065] Based on the preliminary dose distribution and a machine parameter prediction model, generate the machine parameters required by the treatment machine;

[0066] Based on the generated machine parameters and a fast dose calculation model based on transfer learning, obtain a first dose distribution;

[0067] Automatically adjust the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk;

[0068] Based on the fast dose calculation model, combine with the adjusted machine parameters to obtain the final dose distribution.

[0069] In a possible implementation manner, the machine parameters include gantry angle, treatment couch angle, scanning point positions under each angle combination, and the particle beam intensity required for each scanning point, etc.

[0070] The working principle of the above technical solution is as follows: First, based on the preliminary dose distribution, a pre-constructed machine parameter prediction model is used to generate the machine parameters required for the treatment machine. This prediction model is trained based on a large amount of historical data and deep learning algorithms. It can understand the complex mapping relationship between the dose distribution and the machine parameters, and thus output the corresponding machine parameters according to the input preliminary dose distribution, such as gantry angle, treatment couch angle, scanning point positions under various angle combinations, and the required particle beam intensity at each scanning point, etc.

[0071] After obtaining the generated machine parameters, they are input into the fast dose calculation model based on transfer learning. Transfer learning enables this model to utilize the knowledge learned in other related fields or a large number of similar tasks to quickly and relatively accurately calculate the corresponding first dose distribution under these machine parameters. It skips the long process of learning the complex physical process from scratch and instead performs fast adaptation calculations based on the existing knowledge base.

[0072] The first dose distribution is compared with the clinical prescription dose distribution to calculate the gap between the two. At the same time, considering the key factor of the dose limit for the organs at risk, through specific algorithms and logics, the machine parameters are automatically adjusted. The aim is to make the adjusted machine parameters enable the dose distribution to be closer to the clinical prescription dose distribution and ensure that the dose received by the organs at risk does not exceed the limit.

[0073] The adjusted machine parameters are input into the fast dose calculation model again, and the model recalculates based on these parameters to obtain the final dose distribution; this final dose distribution is obtained after parameter optimization and adjustment and is more in line with the requirements of clinical treatment.

[0074] The effects of the above technical solution are as follows: By adopting a fast dose calculation model based on transfer learning, the time required for dose calculation is greatly reduced. Compared with the traditional method of calculating the dose from scratch, transfer learning can quickly utilize existing knowledge for calculation, shortening the time of the entire radiotherapy plan making process, improving the efficiency of treatment planning, and enabling patients to enter the treatment stage faster. By continuously comparing the gap between the first dose distribution and the clinical prescription dose distribution and automatically adjusting the machine parameters in combination with the dose limits of the organs at risk, the finally obtained dose distribution can more accurately meet the clinical treatment requirements. It not only ensures that the treatment dose for the tumor site is sufficient but also effectively controls the dose received by the organs at risk, improving the safety and effectiveness of radiotherapy and enhancing the treatment quality. The machine parameter prediction model can generate corresponding machine parameters according to different preliminary dose distributions and can automatically adjust the parameters according to the dose distribution gap subsequently. This adaptive feature enables this method to be applicable to the individualized radiotherapy needs of different patients. Regardless of the differences in the tumor location, shape, and physical condition of the patients, a relatively appropriate radiotherapy plan can be generated. The degree of automation of the whole process is relatively high. From the generation of machine parameters to parameter adjustment and then to the acquisition of the final dose distribution, many steps are automatically completed by the model and algorithm, reducing the workload of manual intervention and complex calculations by radiotherapy planners, lowering the labor cost, and at the same time reducing the errors that may be caused by human factors.

[0075] In some embodiments, the method for obtaining the preliminary dose distribution includes:

[0076] Based on an automatic segmentation network model, the target area and the organs at risk are segmented according to the simulated CT of the patient to obtain the information of the delineated parts;

[0077] According to the simulated CT of the patient and the information of the delineated parts, a preliminary dose distribution is generated by using a dose prediction network model.

[0078] The working principle of the above technical solution is as follows:

[0079] The automatic segmentation network model takes the simulated CT image of the patient as the input. This model is trained with a large amount of labeled CT image data and learns the characteristic patterns of different tissues, organs, and target areas during the training process. When a new simulated CT image of the patient is input, the model automatically identifies and segments the target area and the organs at risk in the image based on these learned characteristic patterns, thereby outputting the information of the delineated parts and accurately defining the tumor area and the range of the organs at risk that need to be protected.

[0080] The dose prediction network model takes the simulated CT images of the patient and the contour information obtained in the previous step as inputs. This model is also trained with a large amount of historical case data, which includes CT image features of different patients, corresponding target area and organ-at-risk information, and actual dose distribution. By learning the complex relationships in these data, the model can understand the connection between different anatomical structures (reflected by the simulated CT and contour information) and the appropriate initial dose distribution. Based on this learned connection, the dose prediction network model generates an initial dose distribution according to the specific situation of the current patient, providing an initial basis for the subsequent dose optimization process.

[0081] The effects of the above technical solution are as follows: The automatic segmentation network model can capture subtle feature differences in CT images using deep learning algorithms. Compared with traditional manual segmentation methods, it greatly improves the accuracy of target area and organ-at-risk segmentation. More accurate segmentation means that subsequent dose calculations and treatment planning can better fit the actual anatomical structure of the patient, reducing the accidental irradiation of normal tissues and improving the treatment effect. The processes of automatic segmentation and initial dose distribution generation are both automatically completed by the model, greatly reducing the time for manual contour drawing and dose calculation. The automatic segmentation network model and the dose prediction network model follow unified standards and algorithms during training, and have consistency in the processing methods for different patients. This avoids the inconsistency of segmentation and dose calculation results caused by differences in personal experience and operation habits of different radiotherapy planners, ensuring the stability and reliability of radiotherapy plan formulation.

[0082] In some embodiments, the method further includes:

[0083] Evaluating historical samples and obtaining training samples for each model according to the evaluation results; including:

[0084] Obtaining multiple index scores for each sample; the indexes include the dose uniformity of the treatment plan calculation, target area coverage, target area conformity, and the control degree of the irradiated dose of healthy organs;

[0085] Obtaining the comprehensive score of the sample according to the scores of each index in the sample; the sample comprehensive score is obtained by weighted averaging each index, and the weights of each index are adjusted according to the radiotherapy type; for example, radical radiotherapy focuses on target area coverage and organ protection; palliative radiotherapy focuses on organ protection and patient tolerance;

[0086] Selecting the training samples of the model according to the comprehensive score.

[0087] For example: Calculate the dose uniformity score SDU of the treatment plan calculation dose uniformity using the dose uniformity index,

[0088] SDU = 1 - (J2 - J98) / J50; J2, J98, and J50 represent the doses received by 2%, 98%, and 50% of the target volume, respectively.

[0089] The target coverage score is the ratio of the volume of the target actually receiving the prescribed dose to the total volume of the target.

[0090] The target conformity score can be measured using the conformity index.

[0091] For each healthy organ, calculate the ratio of the actual dose received to the tolerance dose; the irradiated dose control score for healthy organs is the reciprocal of the average of the ratios of the actual doses received to the tolerance doses for all healthy organs.

[0092] The irradiated dose control score for healthy organs can also be obtained in the following way:

[0093] For the score of a single organ,

[0094]

[0095] JS v is the single score of the irradiated dose control for the v-th healthy organ, which is used to measure the performance of the irradiated dose control of this organ. The higher the score, the better the irradiated dose control of this organ. JD_actual, v represents the actual irradiation dose received by the v-th healthy organ; in the actual radiotherapy process, the radiation dose value received by this organ is obtained through measurement or calculation; JD_constraint, v is the dose constraint value of the v-th healthy organ, which is preset according to medical standards, clinical experience, or treatment plans, and represents the reasonable maximum dose that this organ can receive in this treatment. The overall formula is 1 minus the ratio of the excess dose to the constraint dose; when there is an excess dose, the larger the proportion of the excess part to the constraint dose, the lower the single score, indicating that the irradiated dose control of this organ is worse.

[0096] Perform a weighted sum of the scores of each healthy organ to obtain the irradiated dose control score for healthy organs. Here, it can be stipulated that the weights of the spinal cord and the lens are 2 times, or 2.5 times, etc. of those of other organs; the reason for setting the weights like this is that the spinal cord and the lens are relatively more sensitive in the human body, and excessive irradiation may cause serious health problems. Therefore, when evaluating the irradiated dose control of healthy organs, higher weights are given to them to reflect the key attention to these key organs.

[0097] In a possible implementation, the training samples of the selection model according to the comprehensive score include:

[0098] Samples with a comprehensive score threshold greater than the preset score threshold are used as training samples.

[0099] In another possible implementation, the selection of the training samples of the model according to the comprehensive score includes:

[0100] The samples are divided into three grades according to the comprehensive score; the first grade is excellent, the second grade is medium, and the third grade is poor;

[0101] The first grade is directly used as the core of the training set;

[0102] For the second grade, samples covering different failure modes are selected by sampling according to a preset ratio.

[0103] The use of the first grade as the core of the training set includes:

[0104] Samples are drawn from the first grade according to the occurrence frequency of each anatomical part; that is, by counting the occurrence frequencies of each anatomical part (such as head and neck, chest and abdomen, pelvic cavity) in a large number of past clinical cases, samples are drawn from the first grade of different anatomical parts according to this frequency ratio; ensuring that the proportion of samples of each anatomical part in the core of the training set is close to the actual clinical distribution;

[0105] The samples of the first grade are stratified according to the dose gradient; the same proportion of samples is drawn from each dose gradient layer;

[0106] Different weights are assigned to the samples through time decay weighting.

[0107] By counting the occurrence frequencies of each anatomical part (such as head and neck, chest and abdomen, pelvic cavity, etc.) in a large number of past clinical cases, when selecting samples of the first grade, samples are drawn from the excellent samples of different anatomical parts according to this frequency ratio, ensuring that the proportion of samples of each anatomical part in the core of the training set is close to the actual clinical distribution, so that the model can learn the best treatment plan features under different anatomical parts.

[0108] Based on the calculation of relevant indicators such as dose uniformity according to the treatment plan in the samples, the samples of the first grade are stratified according to the dose gradient. For example, samples with high dose uniformity and a large dose change range are divided into the high dose gradient layer, and samples with high dose uniformity but a small dose change range are divided into the low dose gradient layer. The same proportion (50% for both the high gradient area and the low gradient area) of samples is drawn from each dose gradient layer, ensuring that the samples with different dose gradient characteristics in the core of the training set are covered;

[0109] Considering the development of medical technology and the timeliness of data, weights are assigned to the samples of the first grade, and the more recent the sample, the higher the weight.

[0110] For the second - level samples, the sampling ratio can be preset between 30% and 50%. During sampling, analyze the performance of the sample in terms of indicators such as the dose uniformity calculated in the treatment plan, target coverage, target conformity, and the control degree of the irradiated dose of healthy organs, and select samples with different failure modes. For example, select some samples with "high target coverage but low target conformity", and some samples with "low control degree of the irradiated dose of healthy organs but high dose uniformity calculated in the treatment plan", etc., so that the model can learn the treatment plan adjustment strategies in the case of deviations in different indicators and enhance the model's adaptability to complex situations.

[0111] For the third - level samples, they can be retained after strict manual review at a very low ratio of no more than 5%. Only select samples with typical teaching significance, such as samples in extreme cases like "severe under - irradiation of the target area but full score for the control degree of the irradiated dose of healthy organs". These samples can be used as a supplement to model training to warn the model to avoid similar serious errors and improve the model's robustness. However, due to their overall poor quality, their proportion in the training set is strictly controlled. Or they can be directly not used.

[0112] Integrate the first - level, second - level (sampled part), and third - level (review - retained part) samples after the above - mentioned processing to construct a complete model training sample set. At the same time, add detailed metadata tags to each sample, recording information such as the source of the sample (such as the affiliated hospital, case number), collection time, specific scores of each indicator, and sample level, etc., to facilitate the subsequent management, maintenance, and optimization of the training sample set. For example, when the model training effect is not good, the problem samples can be quickly located based on the metadata, the reasons can be analyzed, and targeted adjustments can be made.

[0113] The effects of the above - mentioned technical solutions are as follows:

[0114] By scoring multiple key indicators (dose uniformity calculated in the treatment plan, target coverage, target conformity, and control degree of the irradiated dose of healthy organs) and obtaining a comprehensive score based on this, high - quality samples can be screened out for training the model. These high - quality samples can provide accurate and effective information for the model, help the model learn the ideal treatment plan characteristics, thereby improving the model's ability to predict and generate accurate treatment plans, making the final radiotherapy dose distribution more in line with clinical needs and enhancing the treatment effect.

[0115] Adjust the weights of each indicator according to the radiotherapy type (such as radical radiotherapy, palliative radiotherapy) to calculate the comprehensive score, so that the model training can be optimized for different radiotherapy purposes. The model can learn the key points and critical factors in different radiotherapy scenarios and can better adapt to various radiotherapy types in practical applications, providing more appropriate initial dose distributions and treatment plans for patients with different treatment needs.

[0116] When selecting training samples, the frequency of occurrence of each anatomical part is fully considered, and samples of different anatomical parts are extracted from the first-level samples according to the actual clinical distribution. This ensures that samples of various anatomical parts are covered in the training set, enabling the model to learn the radiotherapy characteristics of different body parts, improving the generalization ability of the model when facing different patient anatomical structures, and enhancing the adaptability of the model to diverse clinical cases.

[0117] The first-level samples are stratified and sampled according to the dose gradient to ensure that the core of the training set contains samples with different dose gradient characteristics. This enables the model to learn the treatment plan mode under different dose conditions, and when dealing with the actual patient dose distribution problem, it can more comprehensively consider the dose factor and improve the rationality and accuracy of dose prediction and planning.

[0118] Assigning time decay weights to samples allows the model to learn more recent valid data as medical technology develops and data is updated. The model can adapt to changes brought about by medical technology advances in a timely manner, avoid model performance lags caused by the use of outdated data, and ensure that the model always maintains good performance and applicability in the evolving medical environment.

[0119] For the second-level samples, sampling is performed according to the preset ratio and samples covering different failure modes are selected to expose the model to situations where various indicators are biased. During the training process, the model can learn adjustment strategies for different failure modes. When encountering complex or undesirable situations in actual clinical practice, the model can make better reasonable responses, thereby improving the robustness of the model and its ability to deal with complex problems.

[0120] The third-level samples are strictly screened, and only samples with typical teaching significance (such as extreme error situations) are retained, and their proportion in the training set is controlled. These samples can be used as warning cases for model training, allowing the model to learn how to avoid serious errors, further improve the robustness of the model, and ensure the safety and reliability of the model in practical applications. Alternatively, the third-level samples can be directly omitted to ensure the safety and reliability of the model in practical applications.

[0121] Add detailed metadata tags to each sample, record information such as sample source, collection time, various indicator scores and sample level, which helps to effectively manage and maintain the training sample set. When the model training effect is not good, it can quickly locate the problem samples based on metadata, analyze the causes and perform targeted optimization, thereby improving the efficiency of model training and optimization.

[0122] In some embodiments, the automatic segmentation network model construction and training process includes:

[0123] Slice the simulated CT images of the patient and input the slice sequence of each plane into the encoder sub-module composed of multiple convolutional layers and pooling layers;

[0124] Fuse the feature maps output by different plane sub-modules; Output the overall spatial envelope feature map containing all target regions and organs at risk;

[0125] Based on dynamic label encoding, construct a hierarchical label weight matrix;

[0126]

[0127] Among them, D organ is the distance field from the pixel to the organ surface; W i,j,k is the label weight value at the three-dimensional spatial coordinate (i, j, k); β is a hyperparameter, D organ (i, j, k) is the distance field from the pixel at the coordinate (i, j, k) to the surface of the target organ;

[0128] Use the anatomical constraint loss function to train the model. The anatomical constraint loss function is:

[0129]

[0130] L is the entire anatomical constraint loss; L Dice is the Dice loss; α is a balance parameter, 0 < α < 1; M is the total number of segmentation regions participating in the calculation; m, n are indices; IoU((S m , S n ) is the intersection over union; τ mn is the preset threshold.

[0131] The working principle of the above technical solution is:

[0132] In the encoder part, slice the simulated CT images of the patient in the sagittal plane, coronal plane, and transverse plane; For the slice sequence of each plane, input it into the encoder sub-module composed of multiple convolutional layers and pooling layers respectively. For example, each sub-module contains 2 - 3 consecutive 3×3 convolutional layers for extracting local features within the plane, followed by a 2×2 max pooling layer for downsampling to reduce the data dimension and expand the receptive field.

[0133] Fuse the feature maps output by different planar sub-modules; calculate the attention weights between the planar feature maps, specifically by calculating the dot product between the feature maps or using a more complex attention mechanism (such as the self-attention mechanism) to obtain the weight coefficients of each plane at different positions; multiply these weight coefficients by the corresponding planar feature maps and then sum them to achieve the fusion of multi-planar features, so as to perform a more comprehensive spatial rough localization of the target area and organs at risk in the simulated CT images. Finally, output an overall spatial envelope feature map containing all target areas and organs at risk as one of the inputs to the decoder.

[0134] Based on dynamic label encoding, construct a hierarchical label weight matrix according to the distance field from the pixel to the organ surface;

[0135]

[0136] where D organ is the distance field from the pixel to the organ surface; W i,j,k is the label weight value at the three-dimensional spatial coordinate (i, j, k) position; this weight is used to adjust the contribution degree of pixels at different positions to the segmentation result during the fine segmentation process, highlighting key areas, such as the boundary part of the organ; β is a hyperparameter that controls the steepness of the weight function. The larger the β value, the more drastic the change of the weight function, and the greater the discrimination of pixels at different distances from the organ surface; conversely, the smaller the β value, the smoother the change of the weight function, and the relatively less obvious the discrimination of pixels at different positions. D organ (i, j, k): is the distance field from the pixel at the coordinate (i, j, k) to the surface of the target organ; the distance field reflects the spatial distance relationship between this pixel point and the organ surface, and determines the importance weight of this pixel in the segmentation process through this distance; the pixels closer to the organ surface are usually more important in the segmentation task, and their corresponding weight values will also be larger;

[0137] During the upsampling and convolution operations of the decoder, apply the label weight matrix W i,j,k to pixel-level calculations; for each upsampled feature map, obtain the corresponding weight value according to its pixel coordinates. When calculating the loss function, combine this weight with the predicted value and the true label of the pixel to make the network pay more attention to the segmentation accuracy of key areas such as near the organ surface.

[0138] Train the automatic segmentation network through the anatomical constraint loss function, and the anatomical constraint loss function is:

[0139]

[0140] L is the value of the entire anatomical constraint loss function, which is used to measure the degree of difference between the segmentation result predicted by the model and the true anatomical structure. By minimizing this loss function, the segmentation result of the model can be made more consistent with the anatomically correct structure; α is a balancing parameter with a value range between 0 and 1, which is used to balance the contributions of the two parts in the loss function, that is, α determines the relative importance of the Dice loss L Dice and the subsequent IoU-based constraint term. When α is close to 1, the Dice loss plays a dominant role in the entire loss function; when α is close to 0, the subsequent IoU-based constraint term has a greater impact. L Dice is the Dice loss function, which is used to measure the degree of overlap between the segmented region predicted by the model and the true segmented region; the closer the Dice coefficient is to 1, the higher the degree of overlap between the predicted result and the true result, and the better the segmentation effect; M is the total number of segmented regions involved in the calculation; for example, when segmenting multiple organs, M is the number of organs; m and n are indices used to traverse the indices of different segmented regions, m ranges from 1 to M - 1, and n is the index of the other region corresponding to m; by comparing the relationship between different regions S m and S n , the rationality of the segmentation result is constrained; IoU((S m , S n ) is the intersection over union, which is used to calculate the ratio of the intersection to the union of two segmented regions S m and S n , measuring the degree of overlap between these two regions; τ mn is a preset threshold used to determine whether the degree of overlap between regions S m and S n conforms to the anatomical structure; when IoU((S m , S n ) < τ mn , max(0, IoU((S m , S n ) - τ mn ) 2 will produce a non-zero value, which contributes to the loss function, prompting the model to adjust the segmentation result so that the overlap relationship between regions is more in line with the anatomical reality.

[0141] During the training process of the U-Net, through the backpropagation algorithm, the parameters of the network are updated according to the gradient calculated by this loss function to minimize the loss value and improve the segmentation accuracy of the model.

[0142] Preprocess the simulated CT image data of the training set, including normalization processing to adjust the pixel value range to between [0, 1] to accelerate the convergence speed of the model; at the same time, perform data augmentation operations such as random rotation, translation, and flipping to expand the data volume and improve the generalization ability of the model.

[0143] Accurately label the true labels of the target area and organs at risk for each simulated CT image as the supervision information during the training process.

[0144] The effects of the above technical solution are as follows: The simulated CT images are sliced along the sagittal plane, coronal plane, and transverse plane and respectively input into the encoder sub-module composed of convolutional layers and pooling layers. This multi-plane processing method can comprehensively capture the local features of the image from different perspectives. For example, a continuous 3×3 convolutional layer can effectively extract the features within the plane, and a 2×2 max pooling layer reduces the data dimension and expands the receptive field, enabling the model to be efficient in the feature extraction stage and laying a foundation for subsequent accurate segmentation.

[0145] By calculating the attention weights between the feature maps of each plane, the feature maps output by different plane sub-modules are fused to achieve multi-plane feature fusion. This operation can perform a more comprehensive spatial rough positioning of the target area and organs at risk, output an overall spatial envelope feature map containing all target areas, enabling the model to comprehensively understand the anatomical structure in the image from multiple dimensions, and improving the accuracy and comprehensiveness of segmentation.

[0146] Based on dynamic label encoding, a hierarchical label weight matrix is constructed, and the pixel weights at different positions are determined according to the distance field from the pixel to the organ surface. During the segmentation process, the pixel weights in key areas such as near the organ surface are larger, and the model pays more attention to these areas, which helps to more accurately outline the organ boundary and improve the accuracy of the segmentation result in key parts. Especially for organs with complex boundaries, the segmentation accuracy can be significantly improved.

[0147] The model is trained using an anatomical constraint loss function. This loss function comprehensively considers the Dice loss and the anatomical structure constraint term based on the intersection over union (IoU). By adjusting the weights of the two through a balance parameter, when the model optimizes the segmentation result, it not only pays attention to the overlap degree between the segmented area and the true area (Dice loss), but also ensures that the overlap relationship between different segmented areas conforms to the anatomical reality (constraint based on IoU), effectively improving the fit between the segmentation result and the actual anatomical structure.

[0148] The simulated CT image data in the training set is preprocessed by normalization, and the pixel value range is adjusted to [0,1], which can accelerate the model convergence speed and reduce the training time. At the same time, data augmentation operations such as random rotation, translation, and flipping are used to expand the data volume, enabling the model to be exposed to more diverse data, improving its adaptability to different situations, enhancing the generalization ability, avoiding overfitting, and enabling the model to maintain good segmentation performance for the CT images of various patients in practical applications.

[0149] Accurately label the true labels of the target area and organs at risk for each simulated CT image as training supervision information, providing a clear learning goal for model training, enabling the model to continuously adjust parameters during training to minimize the difference between the prediction result and the true label, thereby improving the segmentation accuracy and ensuring that the model learns an accurate segmentation pattern that meets clinical requirements.

[0150] In some embodiments, after confirming and correcting the preliminary segmentation result, the following steps are further included:

[0151] Perform a difference analysis between the corrected delineation site information and the original automatic segmentation result to generate a correction region mask; after obtaining the preliminary segmentation result of the U-Net, the doctor can use a pen to make manual adjustments through the interactive interface; the system extracts the topologically closed region manually adjusted by the doctor through handwriting trajectory coordinate parsing; combine the prior knowledge of organ anatomy to perform morphological dilation processing on the discrete handwriting points, for example, perform multiple dilation operations using a suitable structural element (such as a circular or square structural element) to generate a three-dimensional correction influence domain with anatomical rationality, that is, the correction region mask;

[0152] Based on the correction region mask, update the automatic segmentation network parameters through the dynamic weight transfer learning algorithm, specifically including:

[0153] Construct a correction loss function:

[0154]

[0155] where M edit is the correction region mask, S pred is the original predicted label; S edit is the corrected label; γ is a dynamic balance factor, and its value changes with the distance field between the correction region and the organ boundary;

[0156] Freeze the network encoder layer and only backpropagate the gradient to the decoder layer to achieve efficient local parameter update.

[0157] The generation steps of the correction region mask include:

[0158] Extract the topologically closed region manually adjusted by the doctor through handwriting trajectory coordinate parsing;

[0159] Combine the prior knowledge of organ anatomy to perform morphological dilation processing on the discrete handwriting points to generate a three-dimensional correction influence domain with anatomical rationality.

[0160] The triggering conditions of the dynamic weight transfer learning algorithm include:

[0161] The change in the Dice coefficient between the corrected region and the original segmentation, ΔDice > θdice (a preset change threshold, e.g., 0.05), or the average Hausdorff distance of the corrected boundary, ΔD > θdist (a preset distance threshold, e.g., 2 mm);

[0162] The thresholds θdice and θdist are dynamically adjusted according to the clinical importance of the target organ. Among them, the tumor target area is 40%-60% smaller than the critical organ. For example, for the tumor target area, θdist = 1 mm, and for large organs such as the liver, θdist = 3 mm.

[0163] Among them,

[0164]

[0165] Dice og is the Dice of the original segmentation; ΔDice is the coefficient change;

[0166]

[0167] ΔD is the average Hausdorff distance of the organ boundary before and after correction, which is used to quantify the change degree of the organ boundary position after correcting the segmentation result; N is the number of points selected on the boundary of the original segmentation area on; is a point on the boundary of the original segmentation area; is a point on the boundary of the corrected segmentation area; is the boundary of the original segmentation area, that is, the edge of the target area or critical organ obtained by the original segmentation; is the boundary of the corrected segmentation area; means that for each point p on the boundary of the original segmentation area, calculate its distance to all points on the boundary of the corrected segmentation area and take the minimum value; means summing up the minimum distances corresponding to all points p on the boundary of the original segmentation area.

[0168] The model update also includes:

[0169] (a) When it is detected that the correction causes an anatomical structure conflict, directly trigger the update by ignoring the threshold condition;

[0170] (b) For the correction of non-dose-sensitive regions, batch update is performed after accumulating at least M similar corrections, where M ≥ 2 and is determined by the dose gradient value of this region; for example, the gradient norm fb of the correction loss L exceeds 2 standard deviations of the historical mean, it is determined as a critical correction and forced to update.

[0171] Freeze the weights of the encoder layer of the U-Net network and only backpropagate the gradients to the decoder layer; according to the gradients calculated by the modified loss function L fb Update the parameters of the decoder layer to achieve efficient update of local parameters, enabling the model to better adapt to the modified segmentation requirements.

[0172] The effects of the above technical solutions are as follows:

[0173] Extract the topologically closed region manually adjusted by the physician through handwriting trajectory coordinate analysis, and perform morphological dilation processing in combination with organ anatomy prior knowledge to generate a modified region mask. This method can accurately define the scope of the physician's manual correction and ensure the anatomical rationality of the modified region, providing an accurate target region for subsequent model updates, thereby optimizing the segmentation results more pertinently.

[0174] Construct a modified loss function based on the modified region mask, freeze the network encoder layer, and only backpropagate the gradients to the decoder layer, reducing the number of parameters to be updated, lowering the computational complexity, and achieving efficient update of local parameters. This not only avoids the instability caused by large-scale adjustment of the entire model's parameters but also enables the model to quickly adapt to the modified segmentation requirements, saving training time and resources.

[0175] Set dynamic thresholds to trigger the dynamic weight transfer learning algorithm based on indicators such as the change in the Dice coefficient between the modified region and the original segmentation and the average Hausdorff distance of the modified boundary. Adjust the thresholds according to the clinical importance of the target organ, such as setting a smaller threshold for the tumor target area and a larger threshold for large organs, so that the model update better fits the actual needs of different organs. Trigger the update when anatomical structure conflicts or a certain number of similar corrections are accumulated in non-dose-sensitive regions to ensure that the model can promptly adapt to important corrections and improve the accuracy and reliability of the segmentation results.

[0176] By performing differential analysis on the information of the modified delineation site and the original automatic segmentation result to generate a modified region mask and then updating the parameters of the automatic segmentation network based on this, the model can learn the experience and knowledge of the physician's manual correction. This helps improve the accuracy of subsequent automatic segmentation, making the model's segmentation results more in line with clinical actual needs, reducing missegmentation situations, and providing more accurate target area and organ-at-risk segmentation data for radiotherapy treatment planning.

[0177] The physician can use a pen for manual adjustment through the interactive interface, and the system can parse the handwriting and generate a modified region mask. The whole process is easy to operate and conforms to the physician's clinical operation habits. At the same time, the model update mechanism takes into account clinical actual situations, such as the correction processing of anatomical structure conflicts and non-dose-sensitive regions, enhancing the practicality and convenience of this technology in clinical applications and improving the physician's work efficiency.

[0178] The correction of non-dose-sensitive regions adopts the cumulative batch update method. On the premise of ensuring the model performance, it makes full use of multiple correction data of the same type to improve the data utilization efficiency. It is forced to update by judging key corrections (such as the gradient norm of the correction loss exceeding 2 standard deviations of the historical mean), ensuring that important correction information can be incorporated into the model in a timely manner and further optimizing the model performance.

[0179] In some embodiments, the dose prediction network is a generative adversarial network structure, including a discriminator and a generator. According to the simulated CT of the patient and the delineated region information, a preliminary dose distribution is generated using the dose prediction network model, which includes:

[0180] The generator receives the simulated CT and the delineated region information and generates a preliminary dose distribution; a preliminary dose distribution can be generated through multi-organ attention gating.

[0181] The discriminator evaluates the overall dose rationality and the gradient of the target region-critical organ boundary.

[0182] Dosimetric prior knowledge is introduced into the loss function for training the dose prediction network, and the loss function includes the DVH curve matching loss, the dose gradient alignment loss, and the organ dose limit constraint term.

[0183] In an implementable way, the DVH curve matching loss L DVH can be implemented through a differential DVH calculation layer; in deep learning, this constraint term can be allowed to participate in the backpropagation process, so as to update the model parameters and make the model optimize in the direction of reducing L DVH

[0184]

[0185] Among them, JL is the set of dose levels (for example, JL = {0.95D 处方 , D max , 0.5D 处方}); V pred (jl) is the volume of the target region in the predicted dose distribution that receives a dose ≥ jl; histogram statistics + Straight-Through Estimator (STE) is used to transfer the gradient; D 处方 represents the prescription dose, that is, the dose that is specified to be given to the target region in the treatment plan; D max represents the maximum dose; V real (jl) is the volume of the target region in the true dose distribution that receives a dose greater than or equal to jl. V target ​is the total volume of the target region. For each dose level, calculate the difference in the proportion of the target region volume where the received dose is greater than or equal to jl in the predicted dose distribution and the true dose distribution, take the L2 norm of it (i.e., the Euclidean norm, which is the square of the difference here), and then sum the results for all dose levels.

[0186] It can be applied to situations where high computational accuracy is required, the data volume is relatively sufficient, and computational resources permit. When dealing with relatively regular dose distribution data with a not particularly huge number of voxels, by accurately calculating the proportion of the target region volume to measure the matching degree of the DVH curve, it can more accurately reflect the difference between the predicted and true dose distributions, thus providing more accurate guidance for model optimization.

[0187] Approximate the volume percentage of the dose distribution through a differentiable sorting algorithm (such as SoftSort), thereby constructing a differentiable DVH loss:

[0188] In another achievable way, the DVH curve can be approximated by randomly sampling dose points;

[0189] Randomly sample G voxels from the dose distribution and calculate their dose values;

[0190] Statistically calculate the proportion of the sampling points where the dose ≥ jl, approximating the volume proportion V(jl) / V target

[0191] Calculate the difference between the predicted and true sampling proportions to obtain the DVH curve matching loss L DVH ,

[0192]

[0193] where JL is the set of dose levels (for example, JL = {0.95D 处方 , D max , 0.5D 处方}); G pred (jl) is the number of sampling points where the dose is greater than or equal to jl in the predicted dose distribution; G real (jl) is the number of sampling points where the dose is greater than or equal to jl in the true dose distribution; G is the total number of voxels randomly sampled from the dose distribution; this formula is more suitable for use in scenarios with a large data volume, or when computational resources are limited and real-time requirements are high. By approximately calculating the relevant proportions of the dose distribution through random sampling, the computational amount can be reduced to a certain extent. Although it is an approximate calculation, in many cases, it can also well reflect the characteristics of the DVH curve and meet the requirements for loss calculation in model training.

[0194] In a possible implementation, the dose gradient alignment loss mainly constrains the dose gradient in the boundary region between the target area and the organs at risk (OARs). By calculating the sum of the absolute values of the differences between the predicted dose gradient and the true dose gradient in the boundary region, it prompts the change of the predicted dose at the boundary to conform to the real situation, avoiding unreasonable dose mutations in the boundary region.

[0195] Among them, the interface between the target area and the OARs is extracted by performing dilation and erosion operations on the delineated mask. This interface is the key area for focusing on the dose gradient because in radiotherapy, while ensuring that the target area receives sufficient dose, it is necessary to minimize the irradiation of the surrounding organs at risk. Therefore, controlling the dose change at the boundary is very important.

[0196] In some embodiments, the organ dose limit constraint term includes:

[0197] Calculate the dose limit violation for each organ at risk separately to obtain the organ dose limit violation penalty (organ dose limit constraint term):

[0198]

[0199] Or

[0200]

[0201] is the maximum dose of the e-th organ at risk predicted by the model; D 约束，e is the dose upper limit of the e-th organ specified clinically; λ e is the weight of the organ; when the predicted dose exceeds the dose constraint value D of the organ 约束，e calculate the penalty for the exceeded part, and if it does not exceed the limit, the penalty is 0.

[0202] Among them, is applicable to the situation where the penalty requirement for dose over-limit is relatively gentle. For example, when the clinic hopes to control the dose over-limit situation to a certain extent but does not want to give too heavy a penalty for minor over-limit, this formula can be used. For example, for some relatively less sensitive organs at risk, minor dose over-limit may not immediately have a serious impact on the patient's health. At this time, using the formula in the form of the first power can remind the model to pay attention to the dose over-limit problem while giving the model a certain adjustment space, avoiding difficulties in model training or generating unreasonable treatment plans due to excessive penalty.

[0203] It can be applied to scenarios with relatively strict requirements for dose over-limit punishment. For those very sensitive organs at risk, such as the spinal cord and lens, even a small dose over-limit may cause serious irreversible damage to the patient. Therefore, the model needs to more strictly control the dose of these organs. Using a quadratic formula can make the model highly concerned about the dose over-limit situation of these organs during the training process. Once over-limit occurs, it will receive a heavier punishment, thus prompting the model to generate a treatment plan that better meets the clinical safety requirements.

[0204] The effect of the above technical solution is that the generator receives the simulated CT and the delineated site information, and generates a preliminary dose distribution through the multi-organ attention gating. It enables the model to generate a dose distribution targeted according to the characteristics and importance of different organs, improving the fit between the preliminary dose distribution and the patient's actual anatomical structure and treatment needs, providing a more reliable basis for subsequent optimization, and helping to improve the accuracy of dose planning in radiotherapy.

[0205] The discriminator evaluates the overall dose rationality and the target-organ-at-risk boundary gradient. Through this evaluation mechanism, the model can judge whether the generated dose distribution is reasonable from both the overall and key region levels, effectively avoiding the generation of dose distributions that do not meet clinical requirements, further ensuring the accuracy and safety of dose prediction, and ensuring that the target area can be effectively treated while maximizing the protection of organs at risk during radiotherapy.

[0206] Dosimetric prior knowledge is introduced into the loss function, including the DVH curve matching loss, the dose gradient alignment loss, and the organ dose limit constraint term. These prior knowledge provide a clear optimization direction for model training, enabling the model to continuously adjust parameters during the training process and generate results that better conform to the actual dosimetric characteristics of the clinic. For example, the DVH curve matching loss prompts the generated dose distribution to have a closer target volume ratio to the real situation at different dose levels; the dose gradient alignment loss ensures a reasonable dose change in the boundary region between the target area and the organ at risk; the organ dose limit constraint term prevents the organ at risk from receiving excessive irradiation, comprehensively improving the performance of the model and the quality of the generated dose distribution.

[0207] Two methods for calculating the DVH curve matching loss are provided. The method implemented through the differential DVH calculation layer is applicable to situations with high calculation accuracy requirements, relatively sufficient data volume, and available calculation resources. It can accurately reflect the difference between the predicted and real dose distributions and provide precise guidance for model optimization. The method of randomly sampling dose points to approximate the DVH curve is suitable for scenarios with a large data volume, limited calculation resources, or high real-time requirements. While reducing the calculation amount, it can also better reflect the DVH curve characteristics, meet the loss calculation requirements of model training in different application scenarios, and improve the flexibility and applicability of model training.

[0208] The dose gradient alignment loss constrains the dose gradient in the boundary region between the target area and the organs at risk. By calculating the sum of the absolute values of the differences between the predicted dose gradient and the true dose gradient in the boundary region, it promotes the change of the predicted dose at the boundary to conform to the actual situation. This is crucial for radiotherapy, effectively avoiding unreasonable dose mutations in the boundary region, ensuring sufficient dose to the target area while minimizing unnecessary irradiation to the organs at risk, and improving the safety and effectiveness of radiotherapy.

[0209] The organ dose limit constraint term provides two different penalty calculation methods according to the characteristics and clinical requirements of different organs at risk. The first - order form is applicable to the situation where the penalty requirement for dose over - limit is relatively gentle, giving the model a certain adjustment space to avoid excessive penalty affecting the rationality of model training and treatment plan generation; the second - order form is applicable to scenarios with strict penalty requirements for dose over - limit, such as sensitive organs at risk like the spinal cord and lens, which can prompt the model to highly focus on and strictly control the dose of these organs, ensuring the generation of treatment plans that meet clinical safety requirements and satisfying the diverse dose control needs of different organs at risk in radiotherapy.

[0210] In a possible implementation manner, generating the machine parameters required by the treatment machine based on the preliminary dose distribution and the machine parameter prediction model includes:

[0211] Building a machine parameter prediction model using the network architecture of reinforcement learning; optimizing the machine parameters into a Markov decision process, where the state includes the preliminary dose distribution and the current machine parameters;

[0212] The reward function synthesizes the dose target achievement degree and parameter feasibility, and during the training process, the machine parameters that violate physical constraints are punished by the feasibility loss function;

[0213] During training, the uncertainty of dose calculation is simulated by injecting Monte Carlo noise; Gaussian noise is added to the input dose during training to simulate dose calculation errors; uniform noise is added to the action output to simulate equipment execution errors.

[0214] In a possible implementation manner, the feasibility loss function is obtained by the following formula:

[0215]

[0216] L feasible is the feasibility loss value, is the weight coefficient calculated according to the association between the machine parameters and the organ importance; MA h is the actual value of the h - th machine parameter; MA h,max is the maximum value allowed for the h - th machine parameter.

[0217] The working principle and effects of the above technical solution are as follows: A machine parameter prediction model is built using the network architecture of reinforcement learning. Reinforcement learning enables an agent to interact with the environment and learn the optimal policy based on the reward signal feedback from the environment. Here, the machine parameter optimization problem is transformed into a Markov decision process, and the state space includes the initial dose distribution and the current machine parameters. This means that the model will, based on the current dose distribution and existing machine parameters, explore by continuous trial and error to learn how to select appropriate machine parameters to achieve the best treatment effect.

[0218] The reward function comprehensively considers the dose target achievement degree and parameter feasibility. The dose target achievement degree measures the closeness of the dose distribution corresponding to the generated machine parameters to the expected dose target, and the closer to the target, the higher the reward. Parameter feasibility ensures that the generated machine parameters are executable in the actual treatment device and comply with physical constraints. During the training process, for machine parameters that violate physical constraints, they are punished through the feasibility loss function. For example, if the machine parameters exceed the physical limit range of the device, such as the gantry angle exceeding the rotatable range, they will be determined to be infeasible, and their scores in the learning process will be reduced through the feasibility loss function, prompting the model to avoid generating such infeasible parameters. On the basis of ensuring parameter feasibility, it better protects critical organs and improves the safety of radiotherapy.

[0219] During training, to make the model more robust and adaptable, Monte Carlo noise is injected to simulate dose calculation uncertainty. On the one hand, Gaussian noise is added to the input dose to simulate the possible errors in the dose calculation process. Since the actual dose calculation is affected by various factors and there is a certain degree of uncertainty, adding Gaussian noise allows the model to learn how to generate appropriate machine parameters under different dose calculation error conditions. On the other hand, uniform noise is added to the action output (i.e., the generated machine parameters) to simulate the possible errors during the device execution process, enabling the model to consider the uncertainty factors in the actual operation of the device.

[0220] The feasibility loss value depends on the weight coefficient calculated based on the association between the machine parameters and the organ importance, the actual value of the machine parameters, and the maximum value allowed for this machine parameter; the weight coefficient reflects the importance degree of different organs for the machine parameter limitations. For example, for the machine parameter limitations that endanger important organs, the weight coefficient will be set higher. Through this formula, when the machine parameters are closer to or exceed the allowed maximum value, the feasibility loss value becomes larger, thereby imposing a greater penalty on the model and guiding the model to generate machine parameters that are physically feasible and meet the organ protection requirements.

[0221] In a possible implementation, the fast dose calculation model is a pre-trained Transformer architecture model; obtaining the first dose distribution based on the generated machine parameters and the fast dose calculation model based on transfer learning includes:

[0222] Serialize the machine parameters and add positional embeddings, and input them into the dose calculation network of the Transformer architecture; the dose calculation network generates a dose distribution corresponding to the machine parameters through the multi-head self-attention mechanism and the causal self-attention mechanism;

[0223] On the basis of the pre-trained model, fine-tune it through the dynamic weight freezing strategy.

[0224] The working principle of the above technical solution is as follows:

[0225] Select a Transformer architecture model pre-trained on a large amount of dose calculation-related data, load its parameters into the self-built model, and complete the construction of the transfer learning foundation;

[0226] Sort out and encode the generated machine parameters, fuse auxiliary information such as the patient's anatomical structure, and perform input data processing;

[0227] Arrange the gantry angle, treatment couch angle, scanning point coordinates, and beam intensity in a sequence X ∈ R T×d , where T is the total number of irradiation fields and d is the parameter dimension;

[0228] Add learnable positional encoding; enable the model to perceive the irradiation order;

[0229] Build a model architecture with Transformer as the main body, including a multi-head self-attention mechanism, a causal self-attention mechanism, and positional embeddings; split the input into multiple heads, each head independently learns the parameter correlations of different dimensions, splice the outputs of each head, and obtain the final attention feature through a linear transformation; apply a lower triangular mask matrix in the decoder to ensure that the output at a certain position only depends on the previous step information; the loss function includes the L1-L2 mixed loss of the voxel-level dose difference;

[0230] Input the processed input information into the trained model, and generate the first dose distribution through the processing and analysis of the model components.

[0231] The pre-trained model has learned on a large amount of relevant data and already has certain general feature extraction and pattern recognition capabilities. On this basis, a dynamic weight freezing strategy is adopted for fine-tuning. The dynamic weight freezing strategy means that during the fine-tuning process, according to the progress of training and the performance of the model, it is dynamically determined which layers of weights need to be fixed (frozen) and which layers of weights can continue to be updated. For example, at the beginning of training, the weights of most pre-trained layers may be frozen, and only the weights of a few layers close to the output layer are fine-tuned to quickly adapt to the characteristics of the new task data. As training progresses, more layers of weights are gradually unfrozen for joint fine-tuning, enabling the model to more accurately learn the features and parameters required for a specific task (i.e., calculating the dose distribution according to the current machine parameters) while maintaining the pre-trained knowledge.

[0232] The effects of the above technical solution are as follows: The powerful parallel computing ability and unique attention mechanism of the Transformer architecture enable it to quickly process the machine parameter sequence. Compared with traditional dose calculation methods, it greatly shortens the calculation time and can generate a dose distribution according to the input machine parameters in a short time, meeting the requirement of rapid feedback for radiotherapy plan formulation, improving the efficiency of the entire radiotherapy process, and enabling patients to enter the treatment stage faster. The multi-head self-attention mechanism and the causal self-attention mechanism work together, enabling the model to deeply explore the complex correlation relationships between machine parameters, thereby generating a more accurate dose distribution. This accurate dose prediction helps to improve the quality of radiotherapy plans, ensure that the tumor target area can receive sufficient and appropriate doses, and at the same time minimize unnecessary irradiation to surrounding critical organs, improving the treatment effect and safety of radiotherapy.

[0233] Built based on the pre-trained model, it utilizes the general knowledge learned from a large amount of relevant data, reducing the large amount of computing resources and time required for training the model from scratch. Through transfer learning, the model can quickly adapt to the radiotherapy plan making task. Even when the training data is relatively limited, it can achieve good performance based on the pre-trained foundation.

[0234] The dynamic weight freezing strategy provides a flexible fine-tuning method. It can quickly adjust the model to adapt to the key features of the new task at the beginning of training, avoiding the degradation of model performance caused by over-adjusting the pre-trained weights. And in the later stage, by gradually unfreezing more weights, the model can more comprehensively learn the detailed features of the new task, realizing fine optimization of the model performance. This strategy enables the model to learn in the best state at different training stages, improving the adaptability and accuracy of the model in the radiotherapy plan making task.

[0235] By fine-tuning on the basis of a pre-trained model and combining a dynamic weight freezing strategy, the model not only learns the features of a specific task but also retains the generalization ability obtained during the pre-training stage. This enables the model to accurately calculate a reasonable dose distribution when faced with diverse machine parameters of different patients, improving the reliability and universality of the model in actual clinical applications.

[0236] In a possible implementation, the automatically adjusting the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk includes:

[0237] Constructing a multi-objective loss function through the target dose gap, organ limit violation, and machine parameter feasibility; constructing a multi-objective loss function, that is, combining the target dose gap, organ limit violation, and machine parameter feasibility to construct a multi-objective loss function; among them, the target dose gap loss is obtained by calculating the difference between the first dose distribution and the clinical prescription dose distribution in the target area; the organ limit violation loss measures whether the dose received by the organs at risk exceeds the limit; the machine parameter feasibility loss considers whether the machine parameters meet the physical constraints. Finally, the multi-objective loss function is composed of the weighted sum of each loss term, and the weights are dynamically adjusted according to the clinical priorities;

[0238] Iteratively optimizing the parameters through a hybrid strategy of gradient backpropagation and evolutionary algorithm, that is, first initializing the parameter population, and then sequentially performing gradient backpropagation optimization on the current population parameter set, that is, calculating the gradient of the loss function and updating the parameters to obtain a new population; then performing evolutionary algorithm optimization on the new population, including fitness evaluation, selection, crossover, and mutation operations to obtain an evolved new population; iterating in this way until the preset termination condition is met, and outputting the parameter set with the highest fitness value as the treatment machine parameters that meet the clinical dose requirements.

[0239] The working principle of the above technical solution is as follows:

[0240] Generating a set of initial machine parameters as a population, and each parameter set represents a treatment plan. For example, a set of machine parameters, including gantry angle, treatment couch angle, scan point position, and particle beam intensity, is denoted as P;

[0241] For each parameter set in the current population, input it into the fast dose calculation model to obtain the corresponding dose distribution, and then calculate the multi-objective loss function; calculate the gradient of the loss function with respect to the machine parameters through automatic differentiation technology (such as the backpropagation algorithm in deep learning frameworks); according to the calculated gradient, use gradient descent or its variants (such as Adagrad, Adadelta, RMSProp, Adam, etc.) to update the parameters; the updated parameter sets form a new population P new-grad, each parameter set in the population updated by backpropagation through the gradient is input into the fast dose calculation model again to calculate the corresponding multi-objective loss function value; the reciprocal of the loss function value is used as the fitness value, and the higher the fitness value, the better the treatment plan corresponding to the parameter set. Methods such as roulette wheel selection and tournament selection are used to select some parameter sets from the population to enter the next generation. For example, using roulette wheel selection, the probability of each parameter set being selected is proportional to its fitness value, and the parameter set with a high fitness value has a greater chance of being selected; crossover operations are performed on the selected parameter sets to generate new parameter sets P new-cross . Mutation operations are performed on the parameter sets obtained after crossover, and the values of some parameters in the parameter sets are randomly changed with a certain mutation probability to simulate gene mutations in biological evolution and increase the diversity of the population; after selection, crossover, and mutation operations, a new population P optimized by the evolutionary algorithm is obtained new-ea ; the population P optimized by the evolutionary algorithm new-ea is used as the initial population for the next round of iteration, and the backpropagation through the gradient and the evolutionary algorithm optimization process are repeated. Set the iteration termination conditions, such as reaching the maximum number of iterations or the change in the multi-objective loss function value being less than a certain threshold in several consecutive iterations. When the termination conditions are met, the parameter set with the highest fitness value is selected from the final population as the treatment machine parameters that meet the clinical dose requirements for output

[0242] In another possible implementation, the automatically adjusting the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk includes:

[0243] Constructing a multi-objective loss function through the target dose gap, organ limit violation, and machine parameter feasibility;

[0244] Iteratively optimizing the parameters through backpropagation through the gradient

[0245] The working principle of the above technical solution is:

[0246] Generate an initial set of machine parameters Pinit, covering gantry angle, treatment couch angle, scanning point position, particle beam intensity, etc. This set of initial parameters can be generated randomly or determined based on clinical experience, historical data, etc., providing a starting point for subsequent iterative optimization. Calculate the gradient of the loss function: Input the initial parameter set Pinit into the fast dose calculation model to obtain the corresponding dose distribution; then calculate the multi-objective loss function; use automatic differentiation techniques (such as the well-established backpropagation algorithm in deep learning frameworks) to calculate the gradient of the loss function with respect to the machine parameters; according to the calculated gradient, select the gradient descent or its variants (such as Adagrad, Adadelta, RMSProp, Adam, etc.) to update the parameters, obtaining the updated parameter set Pnew; use the updated parameter set as the input parameters for the next round of iteration, and repeat the process of calculating the gradient of the loss function and parameter update. Set reasonable iteration termination conditions, such as reaching the preset maximum number of iterations, or the change in the multi-objective loss function value in consecutive several iterations is less than a certain threshold; once the termination condition is met, output the parameter set at this time as the treatment machine parameters that meet the clinical dose requirements. Adopt strategies such as exponential decay, cosine annealing, etc. to balance the convergence speed and optimization accuracy. At the same time, regularly save the intermediate parameters and loss function values to facilitate the analysis of the optimization process, and promptly discover and solve possible problems, such as gradient disappearance, gradient explosion, etc.

[0247] In one possible implementation, the multi-objective loss function is obtained by the following formula:

[0248]

[0249] where L mu is the multi-objective loss function; D pred is the predicted target dose distribution, D 处方 is the prescription dose specified clinically; e represents different organs at risk; is the maximum dose of the e-th organ at risk predicted by the model; is the clinically specified maximum tolerance dose of the e-th organ at risk; L MA is the loss term related to the physical feasibility of the machine parameters; q1, q2, q3 are dynamic weights; Gamma(D pred , D 处方 ) is used to calculate the difference between the predicted target dose and the prescription dose (the specific calculation logic depends on the definition of the Gamma function, usually measuring the matching degree between the dose distribution and the target dose).

[0250] The effects of the above technical solution are as follows: By calculating the difference between the predicted target dose distribution and the clinically prescribed prescription dose, the multi-objective loss function prompts the model to generate a target dose closer to the prescription requirements, which helps ensure that the tumor target can receive sufficient and accurate radiation doses, maximally kill tumor cells, improve the treatment effect of radiotherapy on tumors, and provide strong support for the patient's recovery.

[0251] Including a loss term related to the physical feasibility of machine parameters enables the model to generate machine parameters that meet the physical limitations of the treatment device while optimizing the dose distribution, ensuring the executability of the radiotherapy plan in actual operation, avoiding treatment interruptions or equipment failures caused by infeasible parameters, and improving the stability and reliability of the radiotherapy process.

[0252] The loss function that comprehensively considers multiple objectives can guide the model to learn and optimize from multiple dimensions, avoiding the situation where the model only focuses on a single objective (such as only optimizing the target dose) and ignores other important factors (such as the protection of organs at risk and the feasibility of machine parameters). By simultaneously balancing these different objectives, the model can generate a more comprehensive radiotherapy plan that better meets the actual clinical needs, improving the overall performance of the model in radiotherapy plan making and scheme formulation tasks.

[0253] By ensuring the treatment effect of the target area, protecting organs at risk, ensuring the feasibility of machine parameters, and flexibly adapting to clinical needs, the multi-objective loss function ultimately helps improve the overall quality of radiotherapy. It provides doctors with a more reliable radiotherapy plan, enhances the safety and effectiveness of radiotherapy treatment, and brings better treatment experiences and outcomes for patients.

[0254] The embodiment of the present application also provides a radiotherapy plan making device based on deep learning, and the device includes:

[0255] A machine parameter generation module, configured to generate machine parameters required for the treatment machine based on the preliminary dose distribution and the machine parameter prediction model;

[0256] A first dose distribution acquisition module, configured to obtain a first dose distribution based on the generated machine parameters and the fast dose calculation model based on transfer learning;

[0257] A machine parameter adjustment module, configured to automatically adjust the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk;

[0258] A final dose distribution acquisition module, configured to obtain the final dose distribution based on the fast dose calculation model and in combination with the adjusted machine parameters.

[0259] The device further includes a training sample selection module, including:

[0260] A sample index acquisition unit, configured to acquire scores of multiple indexes of each sample; the indexes include treatment plan calculation dose uniformity, target area coverage, target area conformity, and healthy organ irradiated dose control degree;

[0261] A sample scoring unit, configured to obtain a comprehensive score of a sample according to the scores of each index in the sample;

[0262] A training sample selection unit, configured to select training samples of each model according to the comprehensive score.

[0263] The working principle and effect of the above technical solution are the same as those in the method embodiment of the present application, and will not be elaborated here.

[0264] An embodiment of the present application provides a radiotherapy system, and the radiotherapy system includes:

[0265] A particle accelerator, configured to generate a high-energy particle beam;

[0266] The aforementioned radiotherapy plan making device, configured to make a radiotherapy plan.

[0267] An embodiment of the present application further provides an electronic device, and the electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of any method described in the embodiments of the present application or the functions of the device described in the embodiments of the present application are implemented.

[0268] An embodiment of the present application further provides a computer-readable storage medium, and the computer-readable storage medium is used to store a computer program. When the computer program is executed, the steps of the method in the embodiments of the present application are implemented. The specific implementation manner is the same as the implementation manner and the achieved technical effect described in the above method embodiment, and some contents will not be elaborated.

[0269] In the present application, the readable storage medium may be any tangible medium that contains or stores a program, and the program can be used by or in combination with an instruction execution system, apparatus, or device. The program product may adopt any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0270] A computer-readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries readable program code. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. The readable storage medium may also be any readable medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, RF, etc., or any suitable combination of the foregoing. The program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and also including conventional procedural programming languages such as the C language or similar programming languages. The program code may be executed entirely on the user computing device, partially on an associated device, as a stand-alone software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., by using an Internet service provider to connect through the Internet).

[0271] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention, and all such changes should fall within the protection scope of the claims of the present invention.

Claims

1. A method for making radiotherapy plans based on deep learning, characterized in that, The method includes: Based on the preliminary dose distribution and the machine parameter prediction model, generating the machine parameters required for the treatment machine; Based on the generated machine parameters and the fast dose calculation model based on transfer learning, obtaining the first dose distribution; Automatically adjusting the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk; Based on the fast dose calculation model and combined with the adjusted machine parameters, obtaining the final dose distribution.

2. A radiotherapy plan making method based on deep learning, characterized in that The method for obtaining the preliminary dose distribution includes: Based on the automatic segmentation network model, segmenting the target area and the organs at risk according to the patient's simulated CT to obtain the delineated site information; Based on the patient's simulated CT and the delineated site information, using the dose prediction network model to generate the preliminary dose distribution.

3. The radiotherapy plan making method according to claim 1 or 2, characterized in that, The method further includes: Evaluating the historical samples and selecting the training samples for each model according to the evaluation results; including: Obtaining the scores of multiple indicators for each sample; the indicators include the dose uniformity calculated by the treatment plan, the target area coverage, the target area conformity, and the dose control degree of the healthy organs; Based on the scores of each indicator in the sample, obtaining the comprehensive score of the sample; Selecting the training samples for each model according to the comprehensive score.

4. The radiotherapy plan making method according to claim 2, characterized in that, The segmenting the target area and the organs at risk according to the patient's simulated CT based on the automatic segmentation network to obtain the delineated site information; includes: Training the automatic segmentation network model for the target area and the organs at risk through the training samples; the automatic segmentation network is a U-net network; Through the trained U-net network, combined with the patient's simulated CT, segmenting the target area and the organs at risk to obtain the preliminary segmentation result; Confirming and correcting the preliminary segmentation result; obtaining the final delineated site information.

5. The radiotherapy plan making method according to claim 4, wherein After the preliminary segmentation result is confirmed and corrected, the following steps are further included: Performing difference analysis between the corrected delineated site information and the original automatic segmentation result to generate a corrected area mask; Based on the corrected area mask, updating the parameters of the automatic segmentation network model through the dynamic weight transfer learning algorithm.

6. The radiotherapy plan making method according to claim 2, wherein, The dose prediction network model is an adversarial generation network structure, including a discriminator and a generator. The generating the preliminary dose distribution using the dose prediction network model according to the patient's simulated CT and the delineated site information; includes: The generator receives the simulated CT and the delineated site information and generates the preliminary dose distribution; The discriminator evaluates the overall dose rationality and the target-organ at risk boundary gradient; Introducing dosimetric prior knowledge into the loss function for training the dose prediction network, including the DVH curve matching loss, the dose gradient alignment loss, and the organ limit constraint term.

7. The radiotherapy plan making method according to claim 1, characterized in that The machine parameters include the gantry angle, the treatment couch angle, the scanning point positions under each angle combination, and the particle beam intensity required for each scanning point.

8. The radiotherapy plan making method according to claim 7, characterized in that, The generating the machine parameters required for the treatment machine based on the preliminary dose distribution and the machine parameter prediction model; includes: Building the machine parameter prediction model using the network architecture of reinforcement learning, optimizing the machine parameters into a Markov decision process, and the state includes the preliminary dose distribution and the current machine parameters; The reward function combines the dose target achievement and parameter feasibility, and during the training process, the machine parameters that violate the physical constraints are punished by the feasibility loss function; During training, Monte Carlo noise is injected to simulate dose calculation uncertainty.

9. The radiotherapy plan making method according to claim 8, wherein The feasibility loss function is obtained through the following formula: L feasible is the feasibility loss value, is the weight coefficient calculated according to the association between machine parameters and organ importance; MA h is the actual value of the h-th machine parameter; MA h,max is the maximum value allowed for the h-th machine parameter.

10. The radiotherapy plan making method according to claim 7, characterized in that, The fast dose calculation model is a pre-trained Transformer architecture model; based on the generated machine parameters, the fast dose calculation model based on transfer learning to obtain the first dose distribution, including: Serializing the machine parameters and adding positional embeddings, and inputting them into the dose calculation network of the Transformer architecture; the dose calculation network generates a dose distribution corresponding to the machine parameters through the multi-head self-attention mechanism and the causal self-attention mechanism; Based on the pre-trained model, fine-tuning is performed through the dynamic weight freezing strategy.

11. The radiotherapy plan making method according to claim 1, wherein Automatically adjusting the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk, including: Constructing a multi-objective loss function through the target dose gap, organ limit violation, and machine parameter feasibility; Iteratively optimizing the parameters through gradient backpropagation or a hybrid strategy of gradient backpropagation combined with an evolutionary algorithm until the preset maximum number of iterations is reached or the change in the multi-objective loss function value is less than the preset threshold, and outputting a set of machine parameters that meet the clinical dose requirements.

12. The radiotherapy plan making method according to claim 11, characterized in that, The multi-objective loss function is obtained through the following formula: Among them, L mu is the multi-objective loss function; D pred is the predicted target dose distribution, and D 处方 is the prescribed dose specified clinically; e represents different organs at risk; is the maximum dose of the e-th organ at risk predicted by the model; is the clinically specified maximum tolerance dose of the e-th organ at risk; L MA is the loss term related to the physical feasibility of machine parameters; q1, q2, and q3 are dynamic weights.

13. A radiotherapy plan making device based on deep learning, characterized in that, The device includes: A machine parameter generation module, configured to generate the machine parameters required for the treatment machine based on the preliminary dose distribution and the machine parameter prediction model; A first dose distribution acquisition module, configured to obtain the first dose distribution based on the generated machine parameters and the fast dose calculation model based on transfer learning; A machine parameter adjustment module, configured to automatically adjust the machine parameters according to the gap between the first dose distribution and the clinical prescription dose distribution and the dose limit of the organs at risk; A final dose distribution acquisition module, configured to obtain the final dose distribution based on the fast dose calculation model and in combination with the adjusted machine parameters.

14. A radiotherapy system, characterized in that, The radiotherapy system includes: A particle accelerator for generating high-energy particle beams; The radiotherapy plan making device according to claim 13, for making a radiotherapy plan.

15. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the method according to any one of claims 1-12 or the functions of the device according to claim 13 are implemented.

16. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions, and when the computer reads the computer instructions, the computer executes the steps of the method according to any one of claims 1-12.

Images