Radiotherapy plan automatic optimization design method based on machine learning

Abstract

The invention discloses a radiotherapy plan automatic optimization design method based on machine learning, and relates to the technical field of radiotherapy, and the method comprises the steps: firstly collecting and preprocessing radiotherapy core data to generate a training set, training feature extraction, dose prediction and optimization adjustment of a model; processing to-be-processed data through a feature extraction model, screening three optimal similar cases in combination with cosine similarity, and extracting effective features; based on the feature vectors and the effective features, outputting an initial dose plan, constraint parameters and a deposition matrix by a dose prediction model; and finally, constructing a target function, and iteratively optimizing the radiation field parameters and the photon flux by adopting an improved ant colony algorithm and a conjugate gradient method. According to the method, clinical experience and individual differences are fused, the radiotherapy plan is efficiently and accurately generated, and the quality stability and clinical suitability of the plan are improved.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, and more specifically to a method for automatically optimizing radiotherapy plans based on machine learning. Background Technology

[0002] Designing a conventional radiotherapy treatment plan is a time-consuming and labor-intensive process, requiring physicists to continuously adjust and optimize field parameters to find the optimal plan. Furthermore, the experience differences among plan designers, the time invested in plan design, and the clinical indicators of the medical institution are all closely related to this process.

[0003] Existing automated radiotherapy planning methods suffer from problems such as excessive iterations, poor optimization targeting, and difficulty in accurately adapting to individual case differences. Some methods adjust optimization conditions globally, which can easily affect the dose distribution to non-target organs, leading to unstable planning quality. With the development of radiotherapy technology, clinical requirements for the accuracy, personalization, and efficiency of radiotherapy planning are increasing. There is an urgent need for a method that integrates machine learning technology to achieve efficient, accurate, and personalized automated optimization of radiotherapy plans. Summary of the Invention

[0004] In view of this, the present invention provides an automatic optimization design method for radiotherapy plans based on machine learning, in order to solve the problems existing in the background technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A machine learning-based method for automated optimization design of radiotherapy plans, comprising: S1: Collect core radiotherapy data and preprocess it to output the training dataset; S2, Based on the training dataset, construct and train the feature extraction model, the dose prediction model, and the optimization and adjustment model; S3, use the feature extraction model to process the current data to be processed, output a feature vector, and use the feature vector and the training dataset to determine similar cases based on cosine similarity, and obtain effective features through the similar cases; S4. Based on the feature vector and effective features, the dose prediction model is used to output the initial dose plan, constraint parameters and deposition matrix; S5. Combine the initial dose plan, constraint parameters, and deposition matrix to construct the objective function, and iteratively optimize it through the optimization adjustment model to output the optimized radiotherapy plan.

[0006] Preferably, S3 specifically includes: S3.1, using the trained feature extraction model Process the current data to be processed. , to obtain the feature vector ;calculate With training dataset Feature vectors of all cases Cosine similarity of the included angle The three cases with the highest similarity were selected. S3.2 Extract clinically validated effective features directly related to radiotherapy plan generation from the complete historical data of three similar cases, including clinical constraint parameters, plan adaptation parameters, and optimization experience parameters; S3.3, organize the above parameters of the three similar cases into a structured set. ,in Indicates the first i A complete set of auxiliary parameters for similar cases.

[0007] Preferably, the clinical constraint parameters include the PTV dose range. OAR dose threshold CI threshold HI threshold The planned adaptation parameters include the number of firing fields. Optimization range of shooting angle Key coefficients of the sedimentation matrix The optimization empirical parameters include the weighting coefficients of PTV and OAR in the objective function. .

[0008] Preferably, S4 specifically includes: S4.1, Receive the feature vector With structured sets ,Will The clinical constraint parameters and plan adaptation parameters of similar cases were standardized and compared with those of similar cases. spliced ​​as As a dose prediction model Input; S4.2, Invoke the dose prediction model ,based on Output initial dose plan : Prescription dosage Based on tumor type and Prescription dosage for similar cases generate; ; in, The baseline prescription dose is for the corresponding tumor type. The prescription dose for the i-th similar case, Weights for tumor characteristics in the current case; Shooting parameters ,based on Midfield range and angle optimization range The number of shooting fields that generate the current data to be processed. , field angle and initial intensity diagram ; ; in, The baseline firing angle corresponding to the current data to be processed. Let be the shooting angle of the i-th similar case. Weights are applied to similar case experiences; S4.3, from The selected constraint parameters were clinically validated and optimized using general clinical standards to form the constraint parameters. Including PTV dose range OAR dose threshold CI threshold With HI threshold ; S4.4, based on The field parameters in the data are calculated using the pencil beam algorithm to determine the dose contribution per unit intensity for each beam to all voxels. Constructing a sedimentation matrix ; ; ; in, This is the ray attenuation correction factor. Voxel tissue attenuation coefficient, For the first m The first beam passed through the first j path length of individual elements The scattering radius of the beam within the voxel is denoted as . Let be the scattered dose distribution function. The total number of voxels. Total number of beams .

[0009] Preferably, S5 specifically includes: S5.1, Receive initial dose plan Constraint parameters and sedimentary matrix Analyze the shooting field parameters , constrain parameter quantity Transform it into the algorithm optimization boundary; S5.2, based on the number of constraint parameters Clinical constraints, fusion deposition matrix Dose contribution logic, construct objective function ; S5.3, with the initial dose plan To optimize the starting point, the objective function of S5.2 is used as the core evaluation criterion. An improved ant colony algorithm is initiated to optimize the field intensity, blade position, and field angle, and the conjugate gradient method is integrated to optimize the photon flux. After each round of parameter adjustment, the actual dose of the voxel is calculated based on the deposition matrix logic of S5.1, and the current dose distribution is derived. When the current dose distribution meets the core clinical indicators, the iteration stops, the optimal parameter combination is output, and the optimized radiotherapy plan is obtained.

[0010] Preferably, S5.2 specifically includes: ; ; ; in, For PTV weighting, For OAR weights, For PTV voxel weights, For OAR voxel weights, The value represents the dosage; 1 indicates that the dosage exceeds the limit, and 0 indicates that it is compliant. For the first m Photon flux of a beam For PTV voxels, for i Number of OAR voxels, For the first i Clinical dose thresholds for each OAR.

[0011] Preferably, the improved ant colony algorithm specifically includes: Pheromones update mechanism: ; in, To update the pheromone concentration, The volatility coefficient is... For the number of ants, The pheromone increment for the nth ant. This is a score correction factor. This is the score for the current optimal plan.

[0012] Path selection probability: ; in, Pheromones as a heuristic factor As the expected heuristic factor, As a heuristic value, The dose loss cost for parameter adjustment, where U is the set of unoptimized parameters. Direction guiding operator; CG algorithm photon flux optimization: ;in, Let k be the photon flux in the kth iteration. To optimize step size, The gradient of the objective function; Optimize the strength of the jungler (based on Iterative adjustment), multi-leaf collimator blade position , field angle The pheromone update mechanism guides ants to search for the optimal parameter combination, and the conjugate gradient method is used to optimize photon flux. Improve dose distribution convergence efficiency; in each iteration, based on With current photon flux Calculate the actual dose of voxels .

[0013] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an automatic optimization design method for radiotherapy plans based on machine learning, which effectively solves the problems of time-consuming and labor-intensive conventional radiotherapy plan design, insufficient targeting of existing automatic planning methods, difficulty in adapting to individual case differences, and unstable plan quality. By accurately matching similar cases to extract clinically effective features, and combining machine learning models to quickly generate initial plans, and then using a hybrid optimization strategy of improved ant colony algorithm and conjugate gradient method, the method ensures accurate delivery of dose to the tumor target volume (PTV) while strictly controlling the over-limit of dose to organs at risk (OAR), significantly shortening the plan design time, and taking into account the personalization, efficiency and clinical reliability of the plan, providing a high-quality and highly adaptable optimization solution for radiotherapy. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1 A flowchart of the method steps provided by the present invention; Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention discloses a machine learning-based automatic optimization design method for radiotherapy plans, such as... Figure 1 As shown, it includes: S1: Collect core radiotherapy data and preprocess it to output the training dataset; S2, based on the training dataset, constructs and trains a feature extraction model, a dose prediction model, and an optimization and adjustment model; S3 uses a feature extraction model to process the current data to be processed, outputs a feature vector, and determines similar cases by comparing the feature vector with the training dataset based on cosine similarity, and obtains effective features through similar cases. S4, based on eigenvectors and effective features, uses a dose prediction model to output the initial dose plan, constraint parameters, and deposition matrix; S5 combines the initial dose plan, constraint parameters, and deposition matrix to construct the objective function. Through iterative optimization of the optimization adjustment model, the optimized radiotherapy plan is output.

[0018] In one specific embodiment, the feature extraction model M feat A CNN-Transformer hybrid network is used for the dose prediction model. M dose An improved U-Net is adopted, which combines the encoder ResNet block with the decoder attention mechanism; In one specific embodiment, S3 specifically includes: S3.1, using the trained feature extraction model Process the current data to be processed. , to obtain the feature vector ;calculate With training dataset Feature vectors of all cases Cosine similarity of the included angle The three cases with the highest similarity were selected. S3.2 Extract clinically validated effective features directly related to radiotherapy plan generation from the complete historical data of three similar cases, including clinical constraint parameters, plan adaptation parameters, and optimization experience parameters; S3.3, organize the above parameters of the three similar cases into a structured set. ,in Indicates the first i A complete set of auxiliary parameters for similar cases.

[0019] In one specific embodiment, the clinical constraint parameters include the PTV dose range. OAR dose threshold CI threshold HI threshold The planned adaptation parameters include the number of firing fields. Optimization range of shooting angle Key coefficients of the sedimentation matrix The optimization empirical parameters include the weighting coefficients of PTV and OAR in the objective function. .

[0020] In one specific embodiment, S4 specifically includes: S4.1, Receive feature vector With structured sets ,Will The clinical constraint parameters and plan adaptation parameters of similar cases were standardized and compared with those of similar cases. spliced ​​as As a dose prediction model Input; S4.2, Invoke the dose prediction model ,based on Output initial dose plan : Prescription dosage Based on tumor type and Prescription dosage for similar cases generate; ; in, The baseline prescription dose is for the corresponding tumor type. The prescription dose for the i-th similar case, Weights for tumor characteristics in the current case; Shooting parameters ,based on Midfield range and angle optimization range The number of shooting fields that generate the current data to be processed. , field angle and initial intensity diagram ; ; in, The baseline firing angle corresponding to the current data to be processed. Let be the shooting angle of the i-th similar case. Weights are applied to similar case experiences; Number of shooting fields currently awaiting processing from The number of clinically effective shooting fields extracted from three similar cases is denoted as [missing information]. Calculate its mean and reasonable range: ; Case feature extraction model Output 256-dimensional feature vector In the process, anatomical feature dimensions that are strongly correlated with the number of shot fields are selected, and after standardization, a weighted anatomical complexity correction factor is calculated: ; in, The average PTV volume for cases with the same tumor type; For feature weights; For PTV size, For the overlap rate between OAR and PTV, The irregularity of the target area morphology; Based on the average number of field counts in similar cases, an initial value was calculated using a weighted factor adjusted for anatomical complexity, and then... Integer rounding: ; ; in, For similar cases, experience weighting, Weights for the anatomical features of the current case; It is a rounding function; This indicates taking the intersection of intervals, and the range of values ​​for the number of shooting fields. .

[0021] Generating the initial intensity map includes: The input dose prediction model extracts dose features through ResNet blocks, then uses a decoder to reconstruct the spatial dose distribution, and finally outputs the original intensity distribution tensor as the output layer mapped to the field intensity distribution. : Among them, tensor dimension ,in This represents the maximum total number of beams. The size of the voxel of the field of view; The original intensity distribution tensor Prune the data based on the number of shooting fields currently pending processing, retaining the top-ranked data. The intensity tensor is obtained by taking the intensity values ​​of each pencil beam. : ; .

[0022] For the matched intensity tensor After performing two layers of clinical constraint filtering and dose-intensity normalization, a structured initial intensity map is finally obtained. .

[0023] S4.3, from The selected constraint parameters were clinically validated and optimized using general clinical standards to form the constraint parameters. Including PTV dose range OAR dose threshold CI threshold With HI threshold Specifically, only constraint parameters from the final radiotherapy plans of the top-3 similar cases that have been clinically validated were extracted to ensure the parameters have actual clinical effectiveness. These parameters were extracted from historical data of the three similar cases. Industry-standard clinical practices were also retrieved. If the clinically validated efficacy (similar cases) is greater than the general clinical standard, and the parameters of the similar cases exceed the upper limit of the general standard, then the general standard shall be used as the boundary; if they are lower than the lower limit, the parameters of the similar cases shall be retained.

[0024] For the PTV dose range, the parameters of three similar cases are weighted and averaged (the higher the similarity, the greater the weight), and then the intersection with the general standard is taken. For the OAR dose threshold, the parameters of similar cases are matched with the general standard according to organ type, and the threshold that has been verified by similar cases and meets the general standard is preferred. The CI and HI thresholds are the average of the effective thresholds of similar cases. If the average exceeds the range of the general standard, the general standard shall prevail.

[0025] S4.4, based on The field parameters in the data are calculated using the pencil beam algorithm to determine the dose contribution per unit intensity for each beam to all voxels. Constructing a sedimentation matrix ; ; ; in, This is the ray attenuation correction factor. Voxel tissue attenuation coefficient, For the first m The first beam passed through the first j Individual path length (shooting angle) Determining the incident direction of the pencil beam directly determines the first m The first beam passed through the first j (path length of individual elements) The scattering radius of the beam within the voxel is denoted as . Let be the scattered dose distribution function. The total number of voxels. Total number of beams .

[0026] In one specific embodiment, S5 specifically includes: S5.1, Receive initial dose plan Constraint parameters and sedimentary matrix Analyze the shooting field parameters , constrain parameter quantity Transform it into the algorithm optimization boundary; S5.2, based on the number of constraint parameters Clinical constraints, fusion deposition matrix Dose contribution logic, construct objective function ; S5.3, with the initial dose plan To optimize the starting point, the objective function of S5.2 is used as the core evaluation criterion. An improved ant colony algorithm is initiated to optimize the field intensity, blade position, and field angle, and the conjugate gradient method is integrated to optimize the photon flux. After each round of parameter adjustment, the actual dose of the voxel is calculated based on the deposition matrix logic of S5.1, and the current dose distribution is derived. When the current dose distribution meets the core clinical indicators, the iteration stops, the optimal parameter combination is output, and the optimized radiotherapy plan is obtained.

[0027] In one specific embodiment, S5.2 specifically includes: ; ; ; in, For PTV weighting, For OAR weights, For PTV voxel weights, For OAR voxel weights, The value represents the dosage; 1 indicates that the dosage exceeds the limit, and 0 indicates that it is compliant. For the first m Photon flux of a beam For PTV voxels, for i Number of OAR voxels, For the first i Clinical dose thresholds for each OAR.

[0028] In one specific embodiment, the improved ant colony algorithm includes: Pheromones update mechanism: ; in, To update the pheromone concentration, The volatility coefficient is... For the number of ants, The pheromone increment for the nth ant. This is a score correction factor. This is the score for the current optimal plan.

[0029] Path selection probability: ; in, Pheromones as a heuristic factor As the expected heuristic factor, As a heuristic value, The dose loss cost for parameter adjustment, where U is the set of unoptimized parameters. Direction guiding operator; CG algorithm photon flux optimization: ;in, Let k be the photon flux in the kth iteration. To optimize step size, The gradient of the objective function; Optimize the strength of the jungler (based on Iterative adjustment), multi-leaf collimator blade position , field angle The pheromone update mechanism guides ants to search for the optimal parameter combination, and the conjugate gradient method is used to optimize photon flux. Improve dose distribution convergence efficiency; in each iteration, based on With current photon flux Calculate the actual dose of voxels Output optimized radiotherapy plan ,in To optimize the post-shooter strength map, To accurately position the blades, To fine-tune the firing angle; To meet Constrained dose distribution The corresponding DVH curve; the optimization results will be presented as follows: N field The units are structurally integrated and ultimately packaged into an optimization plan. P opt 。

[0030] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The methods disclosed in the embodiments are described simply because they correspond to the methods disclosed in the embodiments; relevant parts can be found in the method section.

[0031] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for automatically optimizing radiotherapy plans based on machine learning, characterized in that, include: S1: Collect core radiotherapy data and preprocess it to output the training dataset; S2, Based on the training dataset, construct and train the feature extraction model, the dose prediction model, and the optimization and adjustment model; S3, use the feature extraction model to process the current data to be processed, output a feature vector, and use the feature vector and the training dataset to determine similar cases based on cosine similarity, and obtain effective features through the similar cases; S4. Based on the feature vector and effective features, the dose prediction model is used to output the initial dose plan, constraint parameters and deposition matrix; S5. Combine the initial dose plan, constraint parameters, and deposition matrix to construct the objective function, and iteratively optimize it through the optimization adjustment model to output the optimized radiotherapy plan.

2. The automatic optimization design method for radiotherapy planning based on machine learning according to claim 1, characterized in that, S3 specifically includes: S3.1, using the trained feature extraction model Process the current data to be processed. , to obtain the feature vector ;calculate With training dataset Feature vectors of all cases Cosine similarity of the included angle The three cases with the highest similarity were selected. S3.2 Extract clinically validated effective features directly related to radiotherapy plan generation from the complete historical data of three similar cases, including clinical constraint parameters, plan adaptation parameters, and optimization experience parameters; S3.3, organize the above parameters of the three similar cases into a structured set. ,in Indicates the first i A complete set of auxiliary parameters for similar cases.

3. The automatic optimization design method for radiotherapy planning based on machine learning according to claim 2, characterized in that, The clinical constraint parameters include the PTV dose range. OAR dose threshold CI threshold HI threshold The planned adaptation parameters include the number of firing fields. Optimization range of shooting angle Key coefficients of the sedimentation matrix The optimization empirical parameters include the weighting coefficients of PTV and OAR in the objective function. .

4. The automatic optimization design method for radiotherapy planning based on machine learning according to claim 2, characterized in that, S4 specifically includes: S4.1, Receive the feature vector With structured sets ,Will The clinical constraint parameters and plan adaptation parameters of similar cases were standardized and compared with those of similar cases. spliced ​​as As a dose prediction model Input; S4.2, Invoke the dose prediction model ,based on Output initial dose plan : Prescription dosage Based on tumor type and Prescription dosage for similar cases generate; ； in, The baseline prescription dose is for the corresponding tumor type. The prescription dose for the i-th similar case, Weights for tumor characteristics in the current case; Shooting parameters ,based on Midfield range and angle optimization range The number of shooting fields that generate the current data to be processed. , field angle and initial intensity diagram ; ； in, The baseline firing angle corresponding to the current data to be processed. Let be the shooting angle of the i-th similar case. Weights are applied to similar case experiences; S4.3, from The selected constraint parameters were clinically validated and optimized using general clinical standards to form the constraint parameters. Including PTV dose range OAR dose threshold CI threshold With HI threshold ; S4.4, based on The field parameters in the data are calculated using a pencil beam algorithm to determine the dose contribution per unit intensity for each beam to all voxels. Constructing a sedimentation matrix ; ； ； in, This is the ray attenuation correction factor. Voxel tissue attenuation coefficient, For the first m The first beam passed through the first j path length of individual elements The scattering radius of the beam within the voxel is denoted as . Let be the scattered dose distribution function. The total number of voxels. Total number of beams .

5. The automatic optimization design method for radiotherapy planning based on machine learning according to claim 4, characterized in that, S5 specifically includes: S5.1, Receive initial dose plan Constraint parameters and sedimentary matrix Analyze the shooting field parameters , constrain parameter quantity Transform it into the algorithm optimization boundary; S5.2, based on the number of constraint parameters Clinical constraints, fusion deposition matrix Dose contribution logic, construct objective function ; S5.3, with the initial dose plan To optimize the starting point, the objective function of S5.2 is used as the core evaluation criterion. An improved ant colony algorithm is initiated to optimize the field intensity, blade position, and field angle, and the conjugate gradient method is integrated to optimize the photon flux. After each round of parameter adjustment, the actual dose of the voxel is calculated based on the deposition matrix logic of S5.1, and the current dose distribution is derived. When the current dose distribution meets the core clinical indicators, the iteration stops, the optimal parameter combination is output, and the optimized radiotherapy plan is obtained.

6. The automatic optimization design method for radiotherapy planning based on machine learning according to claim 5, characterized in that, Specifically, S5.2 includes: ； ； ； in, For PTV weighting, For OAR weights, For PTV voxel weights, For OAR voxel weights, The value represents the dosage; 1 indicates an excessive dose, and 0 indicates compliance. For the first m Photon flux of a beam For PTV voxels, for i Number of OAR voxels, For the first i Clinical dose thresholds for each OAR.

7. The automatic optimization design method for radiotherapy planning based on machine learning according to claim 5, characterized in that, The improved ant colony algorithm specifically includes: Pheromones update mechanism: ; in, To update the pheromone concentration, The volatility coefficient is... For the number of ants, The pheromone increment for the nth ant. This is a score correction factor. This is the score for the current optimal plan. Path selection probability: ; in, Pheromones as a heuristic factor As the expected heuristic factor, As a heuristic value, The dose loss cost for parameter adjustment, where U is the set of unoptimized parameters. Direction guiding operator; CG algorithm photon flux optimization: ;in, Let k be the photon flux in the kth iteration. To optimize step size, The gradient of the objective function; Optimize the strength of the jungler Multi-leaf collimator blade position , field angle The pheromone update mechanism guides ants to search for the optimal parameter combination, and the conjugate gradient method is used to optimize photon flux. Improve dose distribution convergence efficiency; in each iteration, based on With current photon flux Calculate the actual dose of voxels .

Images