A radiotherapy dose distribution prediction method based on beam level pseudo field prior guidance

Abstract

This invention provides a method for predicting radiotherapy dose distribution based on beam-level pseudo-field prior guidance. The method includes: acquiring a dataset; dividing the preprocessed dataset into training, validation, and test sets; parsing the preprocessed radiotherapy planning parameters; generating a cumulative opening response map of beam-level control points; determining the effective incident propagation path range of the beam and generating a surface hit weight map; calculating the propagation path length of the beam within the body and generating a depth attenuation weight map; generating beam-level pseudo-field prior tensors for all beams; constructing a three-dimensional dose distribution prediction model; training the model using the training set; inputting the validation set into the trained model for optimization; and inputting the test set into the finally optimized model to obtain the three-dimensional radiotherapy dose distribution prediction results for the test set. This invention solves the technical problems of existing radiotherapy dose prediction methods, such as coarse expression of beam-related information, insufficient utilization of planning parameters, and difficulty in finely characterizing the differences in interaction between beams.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent prediction of radiotherapy dose distribution, and more specifically, to a method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-field. Background Technology

[0002] One of the core components of automated radiotherapy planning is dose distribution prediction. Its main purpose is to minimize the radiation dose to organs at risk (OARs) while ensuring that the planned target volume (PTV) meets the prescribed dose requirements, thereby providing a reference for subsequent inverse optimization, planning parameter initialization, and planning quality assessment.

[0003] In existing technologies, the solutions related to predicting radiotherapy dose distribution mainly fall into two categories:

[0004] The first category is knowledge-based planning methods. These methods typically utilize existing high-quality treatment plans and analyze the geometric features of the patient's anatomy, such as overlap volume histograms, target distance histograms, and some dose statistics, to estimate the dose volume histogram (DVH) or planning constraints for new patients. This type of method primarily focuses on transferring planning experience but struggles to directly recover the voxel-level three-dimensional dose distribution.

[0005] The second category comprises deep learning-based voxel-level dose prediction methods. These methods typically take planned computed tomography (CT) images, a planned target volume mask, and an organ-at-risk mask as input, and utilize convolutional neural networks, encoder-decoder networks, generative adversarial networks, or Transformer networks to directly output a three-dimensional voxel-level dose distribution. Some methods further incorporate distance maps, boundary maps, beam pattern maps, field coverage maps, or dose-related loss functions to enhance the model's ability to represent target coverage and organ-at-risk protection.

[0006] Building upon the aforementioned deep learning methods, some technical approaches have attempted to incorporate beam-related information to enhance the model's ability to express the directionality of dose deposition. For example, beam direction, field coverage, or beam path information can be used as auxiliary inputs, fed into the prediction model along with CT images and structural masks. However, most existing methods still use beam-related information in the form of total field level, merge level, or coarse geometry, and their ability to characterize differences in control point opening variations, surface incidence relationships, and in vivo propagation differences between different beams remains relatively limited.

[0007] While existing radiotherapy dose prediction methods have been able to generate reference dose distributions to some extent, they still have the following shortcomings:

[0008] (1) Most existing beam-related information uses total field coverage map, single direction map or combined beam representation, which fails to further distinguish the differences of different beams in the control point opening response, incident path and propagation process in the body, resulting in insufficient refinement of the expression of actual dose deposition directionality;

[0009] (2) Existing methods do not make full use of parameters known in the radiotherapy planning file, such as gantry angle, collimator angle, treatment bed angle, multi-leaf collimator control point opening and monitoring unit weight, and do not fully convert these planning parameters into patient-specific auxiliary information suitable for three-dimensional dose prediction;

[0010] (3) Although some methods introduce beam direction or field geometry information, their expression is relatively coarse and it is difficult to simultaneously reverse the beam coverage shape, control point cumulative opening effect and in vivo propagation attenuation relationship. Therefore, the ability to model dose distribution under complex anatomical conditions is still limited.

[0011] (4) Although the predicted dose distribution output by the existing method can be used for planning assistance, there is still a lack of a technical solution that can more fully combine planning geometry information and form a three-dimensional beam-level auxiliary prior, given that the planning parameters are known.

[0012] Therefore, it is necessary to propose a new technical solution to explicitly transform the beam delivery-related information in the radiotherapy planning parameters into beam-level pseudo-field priors and apply them to dose distribution prediction, so as to improve the directional expression ability of the reference dose distribution and the individual patient adaptation ability. Summary of the Invention

[0013] The present invention aims to solve the technical problems of existing radiotherapy dose prediction methods, such as the coarse expression of beam-related information, insufficient utilization of planning parameters, and difficulty in accurately characterizing the differences in interaction between beams.

[0014] Therefore, the purpose of this invention is to propose a method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-field.

[0015] To achieve the above objectives, the present invention provides a method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields. This method includes: Step S1: acquiring a dataset; wherein the dataset includes: CT images, PTV structure, OARs structure, clinical reference dose distribution, and corresponding radiotherapy planning parameters for each patient; Step S2: preprocessing the dataset; Step S3: dividing the preprocessed dataset into a training set, a validation set, and a test set according to a preset ratio; Step S4: analyzing the preprocessed radiotherapy planning parameters based on the b-th beam; Step S5: generating a cumulative opening response map of beam-level control points based on the analysis results of the radiotherapy planning parameters of the b-th beam; wherein the cumulative opening response map of beam-level control points is used to characterize the b-th beam. Step S6: Based on the patient's surface mask and the beam's incident direction, determine the effective incident propagation path range of the beam entering the patient's body from the outside, and then generate a surface hit weight map based on the effective incident propagation path range; Step S7: Based on the surface hit weight map and the beam's incident direction, calculate the in vivo propagation path length of the beam from the surface to the target voxel after entering the patient's body, and then generate a depth attenuation weight map based on the in vivo propagation path length; Step S8: Combine the beam-level control point cumulative aperture response map, the surface hit weight map, and the depth attenuation weight map to generate the beam-level pseudo-field prior tensor corresponding to all beams; Step S9: Construct a three-dimensional dose distribution prediction model based on the beam-level pseudo-field prior guidance; wherein, the three-dimensional dose The distribution prediction model includes: an anatomical information backbone branch, a beam-level pseudo-field prior coding branch, an encoder, a decoder, and a multi-scale fusion module. The encoder includes a first encoder, a second encoder, and a third encoder connected in sequence. The first encoder includes a first-level wavelet feature enhancement module and a first-level beam prior-guided fusion module connected in series. The second encoder includes a second-level wavelet feature enhancement module and a second-level beam prior-guided fusion module connected in series. The third encoder is a third-level wavelet feature enhancement module. The decoder includes a first decoder and a second decoder connected in series. The first decoder is a first-level wavelet hybrid upsampling module. The second decoder is a second-level wavelet hybrid upsampling module. The multi-scale fusion module includes a first-level multi-scale fusion module and a second-level multi-scale fusion module connected in series. The anatomical information backbone branch takes patient CT images, PTV masks, and OAR masks as input, and after initial convolution mapping, feeds them into the first-level wavelet feature enhancement module of the encoder. The beam-level pseudo-field prior coding branch takes the beam-level pseudo-field prior tensor as input, and extracts prior features of different scales through the prior encoder. and prior features and the prior features and prior features The high-frequency detail features output by the first-stage beam prior-guided fusion module and the second-stage beam prior-guided fusion module are respectively injected into the encoder. The high-frequency detail features output by the first-stage beam prior-guided fusion module are directly input to the first-stage wavelet hybrid upsampling module via a high-frequency skip connection. The high-frequency detail features output by the second-stage beam prior-guided fusion module are directly input to the second-stage wavelet hybrid upsampling module via a high-frequency skip connection. The multi-scale fusion module employs an adaptive mixing mechanism in its second-stage multi-scale fusion module, adaptively fusing the output features of the second encoder and the third encoder, and then inputting the fused features to the second decoder. The first-stage multi-scale fusion module employs an adaptive mixing mechanism... The output features of the first encoder and the output features of the second decoder are adaptively fused, and then the fused features are input into the first decoder; the output features of the first decoder are directly fed into the output convolutional layer to generate the three-dimensional radiotherapy dose distribution prediction results corresponding to the training set, validation set, and test set; the output features of the first decoder serve as the supervision object of the first loss function; the output features of the second decoder serve as the supervision object of the second loss function; the three-dimensional radiotherapy dose distribution prediction results serve as the supervision object of the main loss function; Step S10: Based on the first loss function, the second loss function, and the main loss function, and using the training set, the three-dimensional dose distribution prediction model is trained; Step S11: The validation set is input into the trained three-dimensional dose distribution prediction model for optimization; Step S12: The test set is input into the finally optimized three-dimensional dose distribution prediction model to obtain the three-dimensional radiotherapy dose distribution prediction results corresponding to the test set.

[0016] Preferably, in step S2, the preprocessed radiotherapy planning parameters include: beam geometry parameters and control point related parameters; the beam geometry parameters and control point related parameters include: gantry angle, collimator angle, treatment bed angle, source axis distance, multi-leaf collimator control point opening information, and monitoring unit weight; the multi-leaf collimator control point opening information includes: multi-leaf collimator opening profile.

[0017] Preferably, step S4 specifically includes: Step S4.1: Based on the gantry angle, collimator angle, treatment bed angle, and source axis distance of the b-th beam, determine the incident direction and projection relationship of the b-th beam in the patient's three-dimensional coordinate system to establish a geometric mapping relationship between the b-th beam and the patient's voxel space; Step S4.2: For the b-th beam, read the multi-leaf collimator opening profile and corresponding monitoring unit weights under all control points of the b-th beam to form a control point parameter sequence for the b-th beam; wherein, the control point parameter sequence is used to reflect the dynamic changes of the opening shape and monitoring unit weights of the b-th beam throughout the entire delivery process.

[0018] Preferably, step S5 specifically includes: step S5.1: based on the control point parameter sequence of the b-th beam, according to the opening profile and beam incident direction corresponding to each control point, project the opening region of the corresponding control point onto the patient voxel space to characterize the potential coverage range of the corresponding control point to different spatial locations within the patient; step S5.2: weighted summation of the projection results of all control points of the same beam according to their monitoring unit weights to generate the cumulative opening response map of the beam-level control points.

[0019] Preferably, step S6 specifically includes: Step S6.1: Determine the possible incident boundary position of the beam based on the patient's body surface mask or the patient's outer contour, and determine the effective incident propagation path range of the beam entering the patient's body from the outside, so as to establish the correspondence between the body surface and the beam; Step S6.2: Determine whether the b-th beam can hit the path of the target voxel along its incident direction, and generate the body surface hit weight map based on the determination result; When the determination result is yes, assign the corresponding body surface hit weight to the voxel position located within the effective incident propagation path range of the beam; When the determination result is no, reduce or set the corresponding weight to zero for the voxel position not located within the effective incident propagation path range of the beam.

[0020] Preferably, step S7 specifically includes: step S7.1: for voxel position p, calculate the length of the in vivo propagation path of the b-th beam from the incident point on the body surface to the voxel position p, so as to characterize the propagation depth of the beam in the patient's body;

[0021] Step S7.2: Based on the preset attenuation model, the lengths of the propagation paths within the body are converted into corresponding depth attenuation weights to generate a depth attenuation weight map; wherein, the preset attenuation model adopts an exponential attenuation form, a percentage depth dose curve mapping method, an empirical piecewise function method, or other models that can characterize the propagation attenuation trend within the beam fluid.

[0022] Preferably, step S8 specifically includes:

[0023] Step S8.1: Generate the prior value of the beam-level pseudo-field at voxel position p for the b-th beam:

[0024] ;

[0025] In the formula, Indicates the cumulative open response of the control point; Indicates the surface hit weight; Indicates the depth decay weight;

[0026] Step S8.2: Perform steps S4 to S7 and step 8.1 for all beams respectively;

[0027] Step S8.3: Generate the cumulative beam channel, average propagation depth channel, and coverage statistics channel;

[0028] Step S8.4: Based on the cumulative beam channel, average propagation depth channel, coverage statistics channel, and the beam-level pseudo-field prior value of each beam at voxel position p, generate the beam-level pseudo-field prior tensor corresponding to all beams to enhance the overall beam morphology expression capability.

[0029] ;

[0030] In the formula, B represents the beam-level pseudo-field prior tensor for all beams; B represents the number of beams. Prior value of the beam level pseudo-field at voxel position p for the B-th beam; Indicates the cumulative beam current channel; Indicates the average propagation depth channel; This indicates coverage of statistical channels.

[0031] Preferably, in step S9, the corresponding structures of the encoder and decoder are implemented through a wavelet domain encoding-decoding structure.

[0032] Preferably, in step S10, the total loss function for training the three-dimensional dose distribution prediction model is expressed as:

[0033] ;

[0034] In the formula, Represents the total loss function; ; and Let these represent the first loss function and the second loss function, respectively. , and These represent the weighting coefficients of the corresponding loss terms.

[0035] Preferably, the main loss function The voxel-level mean absolute error loss function is used; where the main loss function is... Represented as:

[0036] ;

[0037] In the formula, This represents the predicted dose value of the main output branch at the i-th voxel position; represents the reference dose value at the i-th voxel position; N represents the total number of voxels involved in the loss calculation.

[0038] Compared with existing technologies, the radiotherapy dose distribution prediction method based on beam-level pseudo-field prior guidance provided by this invention has the following beneficial effects:

[0039] (1) It can construct patient-specific beam-level pseudo-field priors based on radiotherapy planning parameters, and more detailed characterize the differences of different beams in direction, opening response and in vivo propagation process;

[0040] (2) It can effectively introduce the beam-level pseudo-field prior into the three-dimensional dose distribution prediction backbone through independent prior coding branches and hierarchical guided fusion, thereby improving the ability to express the overall dose pattern and spatial coverage relationship;

[0041] (3) It can incorporate the control point opening response, surface incidence relationship and in vivo propagation attenuation trend of different beams into the dose prediction process, so that the prediction model can make full use of the patient's individualized planning geometry information and beam space information when generating the three-dimensional reference dose distribution.

[0042] (4) The output three-dimensional reference dose distribution can be further used for automatic plan optimization, inverse plan parameter initialization, plan quality assessment or candidate plan screening.

[0043] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0044] Figure 1 A schematic diagram of a three-dimensional dose distribution prediction model based on beam-level pseudo-field prior guidance, according to an embodiment of the present invention, is shown.

[0045] Figure 2 A schematic diagram of the low-frequency feature refinement unit structure in the feature enhancement module of an embodiment of the present invention is shown;

[0046] Figure 3 A schematic diagram of the reconstructed feature refinement unit structure in the feature enhancement module of an embodiment of the present invention is shown;

[0047] Figure 4A schematic diagram of the three-dimensional dose distribution prediction results for a head and neck cancer patient according to an embodiment of the present invention is shown.

[0048] Figure 5 A schematic diagram illustrating the three-dimensional dose distribution prediction results for cervical cancer patients according to an embodiment of the present invention is shown. Detailed Implementation

[0049] To better understand the above-mentioned objects, features, and advantages of the present invention, such as Figures 1 to 5 As shown in the accompanying drawings and specific embodiments, the present invention will be further described in detail below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0051] The technical solution of the present invention will be illustrated below with a specific embodiment.

[0052] To construct a beam-level pseudo-field prior based on radiotherapy planning parameters and apply it to dose distribution prediction, this specific embodiment requires corresponding patient image data, structural information, reference dose information, and radiotherapy planning parameters. This specific embodiment is implemented on a publicly available head and neck cancer intensity-modulated radiotherapy (IMRT) dataset and an in-hospital cervical cancer IMRT dataset. The publicly available head and neck cancer IMRT dataset originally included data from 380 patients; after data integrity screening and preprocessing, 372 patients were ultimately included for model training and evaluation. The in-hospital cervical cancer IMRT dataset included data from 200 patients. Each patient's data includes at least: CT images, PTV structure, OAR structure, reference dose distribution, and corresponding radiotherapy planning parameters. The publicly available head and neck cancer dataset includes PTV70, PTV63, and PTV56, and OARs (Organizational Image Arrays) for brainstem, spinal cord, left parotid gland, right parotid gland, larynx, esophagus, and mandible. The in-hospital cervical cancer dataset includes OARs for bladder, rectum, spinal cord, left femoral head, and right femoral head. For different patients, due to potential differences in voxel spacing and slice count in the original images, standardized preprocessing is required before proceeding to subsequent processing steps.

[0053] The specific embodiment of the radiotherapy dose distribution prediction method based on beam-level pseudo-field prior guidance is implemented through the following steps:

[0054] (1) Step 1: Patient data acquisition and standardized preprocessing.

[0055] We acquire CT images, target area structure information, organ at risk structure information, clinical reference dose distribution, and radiotherapy planning parameters for each patient, and perform spatial alignment and standardization on the above data to form the basic input required for subsequent prior construction and dose prediction.

[0056] Step 1 specifically includes:

[0057] Step 1.1: Image and structural information preprocessing;

[0058] Three-dimensional reconstruction is performed on the CT images of each patient, and resampling is carried out according to a uniform voxel spacing. The images are then cropped or zero-padded to a preset fixed size for the three-dimensional volume data. In a preferred embodiment, the fixed size is preferably 128×128×128. The CT grayscale values ​​are truncated and normalized to ensure that the image intensity distribution among different patients meets uniform input requirements. The PTV and various OARs structures drawn by the doctor are extracted from the structure set file and rasterized into three-dimensional binary masks corresponding to the CT images, ensuring a one-to-one spatial correspondence between each structure mask and the CT image.

[0059] Step 1.2: Reference dose distribution preprocessing;

[0060] The dose file in the clinical treatment plan is used as a reference dose label. This file undergoes coordinate alignment, resampling, and size standardization to ensure that the reference dose distribution shares the same three-dimensional voxel mesh with the CT images and structural masks. This guarantees a one-to-one correspondence between images, structures, and doses in the same spatial coordinate system. After this processing, spatially aligned CT images, structural masks, and reference dose distributions are obtained, providing a unified data foundation for subsequent beam-level pseudo-field prior construction.

[0061] Step 1.3: Extraction of radiotherapy planning parameters;

[0062] The radiotherapy planning parameters are derived from the original radiotherapy planning file RTPLAN, and the beam geometry parameters and control point related parameters are obtained by parsing the RTPLAN. Preferably, these parameters include gantry angle, collimator angle, treatment bed angle, source axis distance, multi-leaf collimator control point opening information, and monitoring unit weight.

[0063] Step 1.4: Dataset partitioning;

[0064] Following steps 1.1 to 1.3, the publicly available head and neck cancer intensity-modulated radiotherapy (IMRT) dataset and the in-hospital cervical cancer IMRT dataset were processed separately. After data integrity screening and preprocessing, the publicly available head and neck cancer IMRT dataset included 372 patients for model training and evaluation, with 260 patients used as the training set, 56 as the validation set, and 56 as the test set. The in-hospital cervical cancer IMRT dataset included 200 patients, with 140 used as the training set, 30 as the validation set, and 30 as the test set. This partitioning method ensures that model training, parameter tuning, and performance evaluation are performed on independent subsets of data, thereby guaranteeing the objectivity and reliability of the model evaluation results.

[0065] (2) Step 2: Analyze the radiotherapy plan parameters according to the beam.

[0066] Step 2 specifically includes:

[0067] Step 2.1: Determine the spatial direction of the beam;

[0068] For the b-th beam, the incident direction and projection relationship of the beam in the patient's three-dimensional coordinate system are determined based on its gantry angle, collimator angle, treatment bed angle, and source axis distance, so as to establish the geometric mapping relationship between the beam and the patient's voxel space.

[0069] Step 2.2: Extract the control point parameter sequence;

[0070] For the b-th beam, the multi-leaf collimator opening profiles and corresponding monitoring unit weights at all control points are further read to form the control point parameter sequence for that beam. This control point parameter sequence reflects the dynamic changes in the opening shape and monitoring unit weights of the beam throughout the delivery process and is used to subsequently construct the cumulative opening response map of the control points.

[0071] (3) Step 3: Construct the cumulative opening response diagram of the control points.

[0072] Based on the control point parameter sequence obtained in step 2, the opening profile of each control point is projected onto the patient voxel space and accumulated according to the corresponding monitoring unit weights to generate a beam-level control point cumulative opening response map. This cumulative opening response map characterizes the cumulative coverage intensity of the beam at different spatial locations within the patient's body throughout the entire delivery process.

[0073] Step 3 specifically includes:

[0074] Step 3.1: Control point opening projection;

[0075] For each control point, the control point opening region is mapped to the patient voxel space according to its corresponding opening profile and beam incident direction, in order to characterize the potential coverage range of the control point at different spatial locations within the patient.

[0076] Step 3.2: Weighted accumulation of monitoring units;

[0077] The projection results of all control points for the same beam are weighted and summed according to their monitoring unit weights to obtain the cumulative aperture response distribution of the b-th beam in the voxel space throughout the delivery process. This response distribution reflects the relative coverage intensity and cumulative aperture effect of the beam in space.

[0078] (4) Step 4: Construct the body surface hit relationship.

[0079] Based on the patient's body surface mask and the direction of beam incidence, the effective propagation relationship of the beam from outside the body into the patient's body is determined, and a body surface hit weight map is generated.

[0080] Step 4 specifically includes:

[0081] Step 4.1: Determine the body surface area;

[0082] The possible incident boundary position of the beam is determined based on the patient's outer contour or body surface mask, thereby establishing the correspondence between the body surface and the beam.

[0083] Step 4.2: Determine the effective hit relationship;

[0084] Determine whether the b-th beam can hit the path of the target voxel along its incident direction. For voxel positions within the effective incident propagation path of the beam, assign a corresponding surface hit weight; for voxel positions outside the effective propagation path, reduce or set their corresponding weight to zero.

[0085] (5) Step 5: Construct a depth attenuation map of in vivo propagation.

[0086] Based on the surface hit relationship and the beam incident direction, the propagation path length of the beam from the body surface into the patient's body to the target voxel is calculated, and a depth attenuation weight map is generated accordingly.

[0087] Step 5 specifically includes:

[0088] Step 5.1: Calculate the length of the path within the body;

[0089] For a voxel location p, calculate the length of the in vivo propagation path of the b-th beam from the incident point on the body surface to that voxel location to characterize the propagation depth of the beam within the patient's body.

[0090] Step 5.2: Generation of decay weights;

[0091] Based on the preset attenuation model, the propagation path length obtained in step 5.1 is converted into a depth attenuation weight.

[0092] In this specific embodiment, the attenuation model preferably adopts an exponential attenuation form; in other embodiments, a percentage depth dose curve mapping method, an empirical piecewise function method, or other models that can characterize the attenuation trend of propagation within the beam fluid can also be used.

[0093] (6) Step 6: Generate the beam-level pseudo-field prior.

[0094] The control point cumulative opening response map obtained in step 3, the body surface hit weight map obtained in step 4, and the depth attenuation weight map obtained in step 5 are combined to generate the beam-level pseudo-field prior map of the corresponding beam.

[0095] Step 6 specifically includes:

[0096] Step 6.1: Single-beam prior generation;

[0097] For any voxel position p, the prior value of the pseudo-field of the beam level for the b-th beam is determined by the cumulative opening response of the control point, the surface hit weight, and the depth attenuation weight. This is expressed as:

[0098] ;

[0099] in, This represents the prior value of the pseudo-field of the beam order at voxel position p for the b-th beam. This indicates the cumulative open response of the control point. Indicates the surface hit weight. This represents the depth decay weight.

[0100] Step 6.2: Generation of multi-beam prior tensors;

[0101] Steps 2 through 5 and step 6.1 are performed on all beams respectively to obtain the multi-channel beam-level pseudo-field prior tensor. Furthermore, while retaining the independent channels of each beam, additional cumulative beam channels, average propagation depth channels, or coverage statistics channels can be generated to enhance the overall beam morphology representation capability.

[0102] ;

[0103] Where B represents the number of beams, Indicates the cumulative beam channel. Indicates the average propagation depth channel. This indicates coverage of statistical channels.

[0104] (7) Step 7: Construct a three-dimensional dose distribution prediction model based on the prior guidance of the beam-level pseudo-field.

[0105] After obtaining the prior knowledge of the beam-level pseudo-field, a neural network model for three-dimensional dose distribution prediction is constructed. In this specific embodiment, the model employs... Figure 1 This is implemented using the wavelet domain coding-decoding structure shown. For ease of description, Figure 1 The feature enhancement module is labeled WFE Block, the beam prior guided fusion module is labeled BPGF Block, the beam prior response modulation module is labeled BPRM, and the upsampling reconstruction module is labeled WHU Block. These labels are only used to illustrate one specific implementation and do not limit the invention to using only this network structure.

[0106] Step 7 specifically includes:

[0107] Step 7.1: Overall structure of the model;

[0108] The neural network model includes an anatomical information backbone branch, a beam-level pseudo-field prior coding branch, an encoder, a decoder, and a multi-scale fusion module. The "beam-level pseudo-field prior coding branch" corresponds to... Figure 1 The PriorEncoder on the left extracts prior features at different scales through convolution and progressive downsampling. and The branches; WFE Block, BPGF Block and WHU Block serve as specific modules within the encoder or decoder, respectively.

[0109] The anatomical information backbone branch takes patient CT images, PTV masks, and OAR masks as inputs, and feeds them into the encoder backbone after initial convolutional mapping. The beam-level pseudo-field prior encoding branch takes the beam-level pseudo-field prior tensor as input, and extracts prior features of different scales through the prior encoder. and The multi-scale prior features are injected into the beam prior guidance fusion modules of the first two stages of the encoder. The reason for injecting multi-scale prior features into the first two stages of the encoder is that shallow and medium-scale features are more suitable for combining information on beam coverage morphology, control point opening response and in-cell propagation attenuation trend, thereby enhancing the model's ability to express beam directionality and coverage relationship.

[0110] The overall structure of the model is as follows: Figure 1As shown in part (a), the encoder includes a three-stage wavelet feature enhancement module (WFE Block), with the first two stages followed by a beam prior guided fusion module (BPGF Block); the decoder includes a two-stage wavelet hybrid upsampling module (WHU Block), which restores the spatial resolution step by step through high-frequency skip connections and adaptive mixing mechanisms, and finally generates a three-dimensional dose distribution through the output convolutional layer.

[0111] It should be noted that steps 7.2 to 7.4 are only used to illustrate a preferred implementation of a wavelet domain coding-decoding structure; in other embodiments, the feature extraction, prior fusion and decoding reconstruction processes can also be implemented using three-dimensional convolution, residual networks, attention networks or convolution-attention hybrid networks.

[0112] Step 7.2: Wavelet Feature Enhancement Module (WFE Block);

[0113] The WFE Block, as Figure 1 As shown in part (d), after performing discrete wavelet transform on the input features, the low-frequency subband is obtained. and high-frequency subband Among them, the low-frequency subband is fed into the low-frequency feature refinement block LFRB (e.g., Figure 2 As shown), this is used to enhance the overall structural information and low-frequency feature representation; the enhanced low-frequency features are then reconstructed using Inverse Discrete Wavelet Transform (IDWT) and high-frequency information. The features reconstructed by the inverse wavelet are further fed into the reconstructed feature refinement block RFRB, as shown. Figure 3 As shown, the reconstructed spatial features are supplemented and refined to obtain the final module output. .

[0114] Among them, the LFRB module features low-frequency subband characteristics. As input, local spatial information is first extracted through depthwise convolution, then the features are further integrated through the feature mapping unit FMBConv, and finally, residual enhancement output is formed by element-wise addition with the input features. .

[0115] The process can be represented as follows:

[0116] ;

[0117] in, This represents low-frequency enhanced features after depthwise convolution and feature mapping; This indicates element-wise addition.

[0118] The RFRB features after inverse wavelet transform The input is processed by parallel feature extraction via lightweight convolutional branches and residual block branches. The multi-channel features are then concatenated, and the output is obtained through convolutional mapping and residual connections. .

[0119] The process can be represented as follows:

[0120] ;

[0121] ;

[0122] in, These represent reconstructed features from different branches; Concat represents channel concatenation; Conv represents convolutional mapping; This indicates element-wise addition.

[0123] The RFRB is mainly used to supplement and refine the features after inverse wavelet reconstruction, so as to enhance the ability to restore local details and the stability of the reconstructed features.

[0124] Step 7.3: Beam Prior Guidance Fusion Module (BPGF);

[0125] BPGF Block, such as Figure 1 As shown in part (b), the encoder input features As input, a Discrete Wavelet Transform (DWT) is first performed to obtain low-frequency and high-frequency sub-band features. Then, the sub-band features obtained from the wavelet decomposition are convolved to generate the features to be modulated. On the other hand, the low-frequency subband features, the high-frequency subband features after removing the highest frequency subband HHH, and the corresponding scale of encoded prior features are combined. The beam prior response modulation module (BPRM) is fed together with the current. Figure 1 As shown in part (c), a modulation map (Map) is generated. Further, the modulation map is used to generate the features to be modulated. Element-by-element modulation is performed to obtain modulation characteristics. The module output is obtained by fusing the retained high-frequency features with the data. Among them, the high-frequency features are retained. Simultaneously, the signal is transmitted to the WHU Block of the decoder via high-frequency jump connections to enhance the detail recovery capability during the decoding stage and improve the accuracy of dose prediction.

[0126] BPRM is used to generate modulation maps based on wavelet subband features and coded prior features. The module uses low-frequency subband features... High-frequency subband features after removing HHH and corresponding scale-based coding prior features As input, after feature concatenation and convolution mapping, a modulation map is generated through the Sigmoid activation function.

[0127] The modulation pattern can be represented as:

[0128] ;

[0129] in, Indicates the selected wavelet sub-band features; This represents the encoded prior features at the s-th scale; Indicates feature concatenation; Conv indicates convolution mapping; This represents the Sigmoid activation function.

[0130] The modulation process can be represented as follows:

[0131] ;

[0132] in, This indicates element-wise multiplication.

[0133] Subsequently, the modulation features With the features to be modulated Further integration yields module output. In this way, the beam-level pseudo-field prior can guide and constrain low-frequency or global dose morphology features in the wavelet domain.

[0134] Step 7.4: Wavelet Hybrid Upsampling Module (WHU Block);

[0135] The WHU Block, as Figure 1 As shown in section (e), the decoder input features and high-frequency jump connection features from the encoder The input is used as input. The module first splits the input features into channels. One channel combines high-frequency skip-connection features and performs inverse wavelet transform for upsampling; the other channel performs learning-based upsampling through convolutional mapping and PixelShuffle. Then, the upsampling results from the two channels are concatenated and fused to obtain the decoded output. .

[0136] This process can be represented as:

[0137] ;

[0138] ;

[0139] ;

[0140] in, and These represent the two features after channel splitting; IDWT represents inverse wavelet transform upsampling; PixelShuffle represents pixel rearrangement upsampling; the Fusion module is as follows: Figure 1 As shown in part (f), this represents the feature fusion operation.

[0141] In this way, the module can combine the advantages of wavelet reconstruction upsampling and convolutional learning upsampling, thereby improving the spatial recovery capability in the decoding stage.

[0142] The Fusion module includes a Contrast-aware Channel Attention (CCA) module. Figure 1 (as shown in section (g)). The CCA module is used to perform channel-level recalibration and local feature integration on the spliced ​​features obtained from the wavelet reconstruction upsampling branch and the convolutional learning upsampling branch, thereby enhancing the fusion effect between features from different sources.

[0143] The CCA module features splicing characteristics. As input, where The CCA module includes two parallel branches: the first branch is a local branch, which processes the input features. Perform grouped convolution operations to extract local channel-related features and obtain local fusion features. The second branch is the channel attention branch, which first processes the input features. Global Average Pooling (GAP) is performed, followed by convolution mapping, the non-linear activation function SiLU, and another convolution mapping. Finally, channel attention weights are generated using the Sigmoid activation function. .

[0144] The channel attention weight can be expressed as:

[0145] ;

[0146] in, This indicates a global average pooling operation; Represents a convolutional mapping; Represents a non-linear activation function; This represents the Sigmoid activation function.

[0147] Furthermore, channel attention weights This is applied to input features or locally fused features to achieve channel recalibration, and then fused with the local branch output to obtain the CCA module output features. .

[0148] The output feature can be written as:

[0149] ;

[0150] in, Indicates the output characteristics of local branches; This represents element-wise multiplication; This indicates element-wise addition.

[0151] In this way, the CCA module can simultaneously utilize global channel statistics and local convolutional features to adaptively reweight and fuse features from different sources, thereby improving the effectiveness of feature integration in the decoding stage.

[0152] It should be noted that the CCA module is a preferred implementation and does not limit the present invention to only this channel attention fusion method; without departing from the basic concept of the present invention, other channel attention, spatial attention or residual fusion methods can also be used to complete feature integration.

[0153] (8) Step 8: Model Training Setup. The model is trained based on a deep learning framework. In a preferred embodiment, the Adam optimizer is used to update the network parameters. The initial learning rate is set to 0.0005 and dynamically adjusted using a cosine annealing learning rate scheduling strategy. The weight decay coefficient is set to 0.00005, and the number of training epochs is set to 200. During training, the performance of the validation set can be used as the basis for model selection, and the optimal model parameters are saved.

[0154] Step 8 specifically includes:

[0155] Step 8.1: Loss function setting;

[0156] The main output loss uses the voxel-level average absolute error loss function, expressed as:

[0157] ;

[0158] in, This represents the predicted dose value of the main output branch at the i-th voxel position; represents the reference dose value at the i-th voxel position; N represents the total number of voxels involved in the loss calculation.

[0159] Furthermore, auxiliary supervision losses can be defined for the auxiliary outputs generated by the intermediate decoding layers. and Its form is consistent with the main output loss. The auxiliary loss is used to enhance the stability of the intermediate layer feature learning process and improve the training convergence effect.

[0160] The model's total loss function can be expressed as:

[0161] ;

[0162] in, Represents the total loss function; Indicates the main output loss; and These represent the losses of the two auxiliary outputs, respectively. , and These represent the weighting coefficients of the corresponding loss terms.

[0163] make 1. , , Greater than and This highlights the dominant role of the main output branch in the training process. Furthermore, the loss function can also be implemented with regional weights, weighting the planned target area, the region adjacent to the organ at risk, or the non-zero dose region, thereby enhancing the dose prediction accuracy of key regions.

[0164] Step 8.2: Setting evaluation indicators;

[0165] To evaluate the predictive performance of the model, a series of clinical assessment metrics were used, including the Conformity Index and the Heterogeneity Index. The CI was defined as follows:

[0166] ;

[0167] Where TV is the target volume and PIV is the prescribed isodose volume. The CI is the intersection of the target volume and the prescribed isodose volume. CI measures the consistency between the target volume and the prescribed isodose volume. In predictive performance evaluation, the difference in CI between the predicted dose distribution and the reference dose distribution can be further compared; the smaller the difference, the closer the prediction result is to the reference dose distribution in terms of target area conformity. Meanwhile, HI is defined as:

[0168] ;

[0169] Where Dx is the minimum dose required to cover x% of the PTV volume absorbed. Compared to CI, a lower HI is better. Based on clinical practice and expert medical advice, there are many standards used to assess the dose coverage of PTV and OARs, including D99, D95, and D1. D99 represents the minimum dose covering 99% of the target volume, D95 represents the minimum dose covering 95% of the target volume, and D1 represents the minimum dose covering 1% of the target volume. For dose assessment of organs at risk, dose endpoints such as mean dose Dmean, small volume high dose indices D0.1cc and D0.03cc can also be used. Dmean represents the mean dose within the corresponding organ at risk region, and D0.1cc and D0.03cc represent the irradiated volume of 0.1 cm³ in the corresponding organ at risk, respectively. 3 and 0.03 cm 3 The high-dose endpoints. For the aforementioned dose endpoints, the normalized prediction error between the predicted and reference values ​​can be calculated to evaluate the consistency of the predicted dose distribution across dosimetric parameters of organs at risk.

[0170] In another preferred embodiment, a voxel-level mean absolute error can be used to evaluate the overall difference between the predicted dose distribution and the reference dose distribution. This error can be calculated across all voxels, the potentially irradiated region, or a pre-defined evaluation region.

[0171] ;

[0172] Where P is the predicted dose distribution, GT is the reference dose, N is the sum of voxels in all three-dimensional space, and i is the current voxel.

[0173] To evaluate the difference between the predicted results and the reference dose distribution, the average prediction error (APE) can be calculated based on the above evaluation indicators. For CI and HI, the APE between the predicted and reference values ​​can be calculated using formulas. For other clinical indicators such as D95, D99, and D1, the normalized prediction error form is used for calculation. The prediction error can be expressed as APE ± SD, where SD is the standard deviation. The formula for calculating APE is as follows:

[0174] ;

[0175] ;

[0176] in, This represents the dosimetric parameters calculated from the reference dose distribution for the i-th case. This represents the corresponding index calculated from the predicted dose distribution. This indicates the prescription dosage, and n indicates the number of cases.

[0177] Furthermore, in the confirmatory embodiments, paired statistical tests can be used to compare the differences in dosimetric errors of different methods on the same test case. If the difference reaches a preset statistical threshold, it can be marked with a symbol in a table. The statistical tests are only used to illustrate the technical effects of this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0178] Step 8.3: Input the test set into the three-dimensional dose distribution prediction model based on the prior guidance of the beam-level pseudo-field to obtain the three-dimensional radiotherapy dose distribution prediction results corresponding to the test set.

[0179] This specific embodiment provides a method for constructing a beam-level pseudo-field prior based on radiotherapy planning parameters, and applies the constructed beam-level pseudo-field prior to the dose distribution prediction process, thereby improving the model's ability to express different beam directions, coverage relationships and in vivo propagation attenuation trends, and obtaining a three-dimensional reference dose distribution that is more consistent with the characteristics of clinical planning.

[0180] Furthermore, in a preferred embodiment, to verify the effectiveness of the method of the present invention, a comparative experiment was conducted on a publicly available dataset of intensity-modulated radiotherapy for head and neck cancer and a dataset of intensity-modulated radiotherapy for cervical cancer in a hospital. The verification results are shown in Tables 1, 2 and 3.

[0181] Table 1: Comparison of normalized prediction errors (NPE, mean ± standard deviation) of organ-at-risk dose endpoints on the head and neck cancer test set in the embodiments of the present invention.

[0182]

[0183] Note: Numerical values ​​represent mean ± standard deviation. Statistical tests may be performed using paired t-tests or Wilcoxon signed-rank tests to compare the error results of different methods in the same test case. The statistical test results are only used to illustrate the technical effects of this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0184] Table 2: Comparison of errors of clinical dosimetric indices CI, HI, D95, and D1 on the head and neck cancer test set in the embodiments of the present invention.

[0185]

[0186] Note: Numerical values ​​represent mean ± standard deviation. Statistical tests may be performed using paired t-tests or Wilcoxon signed-rank tests to compare the error results of different methods in the same test case. The statistical test results are only used to illustrate the technical effects of this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0187] Table 3: Comparison of clinical dosimetric index APE on the in-hospital cervical cancer test set in the embodiments of the present invention

[0188]

[0189] Note: Numerical values ​​represent mean ± standard deviation. Statistical tests may be performed using paired t-tests or Wilcoxon signed-rank tests to compare the error results of different methods in the same test case. The statistical test results are only used to illustrate the technical effects of this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0190] Figure 4 A schematic diagram illustrating the three-dimensional dose distribution prediction results for head and neck cancer patients according to an embodiment of the present invention is shown. Figure 4 As shown, in the head and neck cancer cases presented, the high-dose region morphology, dose coverage, and surrounding dose transition trend obtained by the method in this embodiment are quite close to the reference dose distribution; the range of higher error areas in the absolute error graph is relatively small, and the overall error distribution is dominated by low-value areas. These results indicate that beam-level pseudo-field priors can provide the model with auxiliary information related to beam direction, control point opening response, and in vivo propagation process, which helps improve the spatial consistency between the predicted dose distribution and the reference dose distribution.

[0191] Figure 5 A schematic diagram illustrating the three-dimensional dose distribution prediction results for cervical cancer patients according to an embodiment of the present invention is shown. Figure 5 As shown, in the cervical cancer cases presented, the spatial morphology of the high-dose area, the lateral dose coverage, and the peripheral dose attenuation trend obtained by the method in this embodiment are quite close to the reference dose distribution; the range of the higher error area in its absolute error map is relatively small, and the overall error distribution is relatively concentrated. These results indicate that introducing patient-specific beam-level pseudo-field priors into the three-dimensional dose distribution prediction model helps the model combine patient anatomical structure and planning geometry information, improving the spatial consistency of dose distribution prediction results for cervical cancer cases.

[0192] In addition to the specific implementation shown in the above embodiments, several alternative implementations exist without departing from the basic concept of the present invention, which can also achieve the inventive objective of constructing a beam-level pseudo-field prior based on radiotherapy planning parameters and applying it to dose distribution prediction. The alternative implementations are as follows:

[0193] (1) First, different specific implementation forms can be adopted in the construction of the beam-level pseudo-field prior.

[0194] For example, in constructing the cumulative control point aperture response, geometric projection of the control point aperture profile, voxel penetration statistics, beam coverage counting, or other accumulation methods based on control point aperture and delivery weights can be used. In modeling the in-cell propagation attenuation relationship, exponential attenuation models, empirical piecewise function models, percentage depth dose curve mapping models, or other attenuation models that can describe the in-cell propagation trend of the beam can be used. Therefore, the specific calculation method for the beam-level pseudo-field prior can be adjusted according to the data source, planning format, and application scenario, and is not limited to a single fixed expression.

[0195] (2) Second, different channel design methods can be adopted in the organization of the beam-level pseudo-field prior.

[0196] For example, a multi-channel organization method can be adopted in which each beam corresponds to an independent prior channel; or, on the basis of retaining the independent channels of each beam, a cumulative beam channel, an average propagation depth channel, a coverage statistics channel, or other statistical auxiliary channels can be further added.

[0197] All of the above-mentioned different forms can characterize the patient-specific spatial distribution characteristics of the beam, and all belong to the alternative implementation methods of the present invention.

[0198] (3) Third, different lightweight feature extraction structures can be used in the implementation of the prior coding branch.

[0199] For example, a convolutional encoding method consisting of 1×1×1 convolutions and 3×3×3 convolutions can be used, as well as residual convolutional structures, grouped convolutional structures, depthwise separable convolutional structures, or other encoding methods that can map the beam-level pseudo-field prior to multi-scale prior features.

[0200] The purpose of this invention can be achieved as long as prior features corresponding to different coding stages can be obtained and used for subsequent feature-guided fusion.

[0201] (4) Fourth, different fusion strategies can be adopted in terms of the injection of prior features.

[0202] For example, a layered injection approach can be used in the first two coding stages, or a single-level injection, multi-level injection, or cross-layer shared injection approach can be used. In terms of specific fusion methods, modulation graph-guided element-wise multiplication modulation, gated fusion, feature concatenation followed by convolutional mapping, channel-weighted methods, or spatial-weighted methods can be employed. Therefore, the specific form in which prior features participate in backbone network computation is not limited to the illustrated implementation.

[0203] (5) Fifth, different three-dimensional feature extraction structures can be used in the dose prediction backbone network.

[0204] In one embodiment, the wavelet domain encoder-decoder backbone shown in this invention can be used; in other embodiments, a three-dimensional convolutional encoder-decoder network, a multi-scale residual network, a hybrid convolutional and attention network, or other network structures suitable for three-dimensional dose distribution prediction can also be used.

[0205] Any backbone network that can combine beam-level pseudo-field priors to complete the prediction of three-dimensional reference dose distribution can be considered an alternative implementation of the present invention.

[0206] (6) Sixth, different implementation methods can also be used in the decoding and feature fusion stages.

[0207] For example, during the decoding upsampling process, a hybrid approach combining inverse wavelet transform and learned upsampling can be used, or a combination of deconvolutional upsampling, interpolation upsampling, and convolutional mapping can be employed. In the fusion module, channel attention feature fusion, residual fusion, convolutional fusion after splicing, or other feature integration methods can be used. Therefore, the specific structure of the decoding and fusion stages does not constitute the sole limitation of this invention.

[0208] In summary, the core of this invention does not lie in a fixed network name or a fixed module implementation, but rather in: constructing a patient-specific beam-level pseudo-field prior based on radiotherapy planning parameters, and introducing the prior into a three-dimensional dose distribution prediction model in the form of multi-scale features, thereby enhancing the model's ability to express beam directionality, opening response differences, and in vivo propagation trends.

[0209] Under this premise, equivalent substitutions or conventional modifications made to prior construction details, feature encoding forms, injection methods, backbone structures, and decoding fusion methods can all be considered as alternative implementations in the subsequent claims of this invention.

[0210] The key technical points of this invention are summarized below, based on the specific embodiments described above: The key technical points of this invention do not lie in a fixed network name or a general attention module, but rather in a complete technical chain built around the relationship between radiotherapy planning parameters, beam delivery process, and three-dimensional dose distribution prediction. Specifically, this includes the following aspects:

[0211] (1) Constructing a priori beam-level pseudo-field based on radiotherapy planning parameters.

[0212] The core of this technique lies in analyzing the geometric parameters and control point parameters of each beam from the original radiotherapy planning file RTPLAN, and combining this with the cumulative opening response of the control points, the surface hit relationship, and the in vivo propagation depth attenuation relationship to construct a three-dimensional prior channel for each beam. This technique differs from existing methods that use beam information as a total field-level coverage map, a coarse radiation pattern, or a merged-level prior input model, and can more meticulously characterize the differences between different beams in terms of direction, spatial coverage, and propagation.

[0213] (2) Multi-channel organization of beam-level pseudo-field prior.

[0214] The core of this technique lies in organizing each beam into an independent prior channel, and further constructing cumulative beam current channels, average propagation depth channels, coverage statistics channels, or other statistical auxiliary channels as needed, thereby forming a patient-specific prior tensor suitable for three-dimensional dose distribution prediction. The key point of this technique is not a fixed number of channels, but the expression of "constructing and organizing multiple prior channels according to the beam".

[0215] (3) Multi-scale coding method for beam-level pseudo-field prior.

[0216] The core of this technique lies in using the beam-level pseudo-field prior as an independent input branch. A prior encoder extracts prior features at different scales, forming a multi-scale prior representation corresponding to the encoding stage, which is then used to guide subsequent fusion. The key to this technique is the "independent prior branch + multi-scale prior features" approach, rather than being limited to a specific combination of convolutional layers.

[0217] (4) Layered injection and guided fusion method of beam prior in the coding stage.

[0218] Its core lies in injecting prior features of different scales into the feature extraction process of the first two or more levels of the encoder. Through modulation map guidance, element-wise weighting, gated fusion, or feature splicing and mapping, the beam prior can be applied to the backbone feature extraction process, thereby enhancing the model's ability to express beam directionality, aperture response differences, and propagation trends within the body.

[0219] The preferred implementation method is as follows: during the wavelet domain feature extraction process, the wavelet subband features and the prior features of the corresponding scale are used to generate a modulation map, and then the features to be modulated are modulated element by element to complete the prior-guided fusion.

[0220] (5) The fusion mechanism of multi-scale prior features and backbone features.

[0221] The core of this approach lies in fusing the multi-scale prior features obtained from the prior encoding of the beam-level pseudo-field with the features in the dose prediction backbone network. The fusion method can be modulation map-guided, element-wise weighted, gated fusion, feature concatenation followed by mapping, channel weighting, or spatial weighting. In a preferred embodiment, this fusion process can be implemented in the wavelet subband feature space.

[0222] (6) Combine the three-dimensional dose prediction backbone to complete the generation of reference dose distribution.

[0223] Its core lies in: based on the modeling of anatomical structure information by the three-dimensional dose distribution prediction backbone network, the beam-level pseudo-field prior is introduced through independent branches and hierarchical fusion, thereby realizing the prediction of three-dimensional reference dose distribution based on individual patient planning parameters. The backbone network can adopt wavelet domain encoder-decoder, three-dimensional convolutional encoder-decoder, multi-scale residual network, or convolutional and attention hybrid network.

[0224] (7) Application of reference dose distribution.

[0225] Its core lies in further utilizing the output three-dimensional reference dose distribution for automatic plan optimization, inverse plan parameter initialization, plan quality assessment, or candidate plan screening, rather than simply as a simple image regression result. This part can be laid out as an application-layer protection point or subordinate protection content.

[0226] The following is a summary of the points to be protected by the present invention in conjunction with the above specific embodiments: The focus of the present invention is not on a fixed network name, a fixed attention module or a fixed convolution module, but on the combination of the following technical solutions: (1) Constructing a beam-level pseudo-field prior based on radiotherapy planning parameters; (2) Organizing the beam-level pseudo-field prior into a multi-channel patient-specific prior tensor; (3) Extracting multi-scale prior features through independent prior coding branches; (4) Injecting the multi-scale prior features into the coding stage of the three-dimensional dose distribution prediction backbone; (5) In a preferred embodiment, the prior-guided fusion can be implemented in the wavelet subband feature space; (6) Further using the three-dimensional reference dose distribution for automatic plan optimization, inverse plan parameter initialization, plan quality assessment or candidate plan screening.

[0227] In summary, compared with the prior art, the radiotherapy dose distribution prediction method based on beam-level pseudo-field prior guidance provided by the present invention has unparalleled technical advantages.

[0228] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-field, characterized in that, include: Step S1: Obtain the dataset; wherein the dataset includes: CT images, PTV structure, OARs structure, clinical reference dose distribution and corresponding radiotherapy plan parameters for each patient; Step S2: Preprocess the dataset; Step S3: Divide the preprocessed dataset into training set, validation set and test set according to the preset ratio; Step S4: Based on the b-th beam, analyze the preprocessed radiotherapy plan parameters; Step S5: Based on the analysis results of the radiotherapy planning parameters of the b-th beam, generate a cumulative opening response map of the beam-level control points; wherein, the cumulative opening response map of the beam-level control points is used to characterize the cumulative coverage intensity formed by the b-th beam at different spatial locations in the patient's body throughout the entire delivery process. Step S6: Based on the patient's body surface mask and the beam incident direction, determine the effective incident propagation path range of the beam entering the patient's body from outside the body, and then generate a body surface hit weight map based on the effective incident propagation path range of the beam. Step S7: Based on the surface hit weight map and the beam incident direction, calculate the length of the in vivo propagation path of the beam from the body surface into the patient's body to the target voxel, and then generate a depth attenuation weight map based on the length of the in vivo propagation path. Step S8: Combine the cumulative aperture response map of the beam-level control point, the body surface hit weight map, and the depth attenuation weight map to generate the beam-level pseudo-field prior tensor corresponding to all beams. Step S9: Construct a three-dimensional dose distribution prediction model based on beam-level pseudo-field prior guidance; wherein, the three-dimensional dose distribution prediction model includes: an anatomical information backbone branch, a beam-level pseudo-field prior coding branch, an encoder, a decoder, and a multi-scale fusion module; The encoder includes: a first encoder, a second encoder, and a third encoder connected in sequence; the first encoder includes: a first-level wavelet feature enhancement module and a first-level beam prior-guided fusion module connected to each other; the second encoder includes: a second-level wavelet feature enhancement module and a second-level beam prior-guided fusion module connected to each other; the third encoder is a third-level wavelet feature enhancement module; the decoder includes: a first decoder and a second decoder connected to each other; the first decoder is a first-level wavelet hybrid upsampling module; the second decoder is a second-level wavelet hybrid upsampling module; the multi-scale fusion module includes: a first-level multi-scale fusion module and a second-level multi-scale fusion module connected to each other; The anatomical information backbone branch takes patient CT images, PTV masks, and OAR masks as input, and after initial convolutional mapping, it is fed into the first-level wavelet feature enhancement module of the encoder; the beam-level pseudo-field prior coding branch takes the beam-level pseudo-field prior tensor as input, and extracts prior features of different scales through the prior encoder. and prior features and the prior features and prior features The first-stage beam prior-guided fusion module and the second-stage beam prior-guided fusion module were injected respectively; The encoder's high-frequency detail features output by its first-stage beam prior-guided fusion module are directly input to the first-stage wavelet hybrid upsampling module via a high-frequency jump connection; the high-frequency detail features output by its second-stage beam prior-guided fusion module are directly input to the second-stage wavelet hybrid upsampling module via a high-frequency jump connection. The multi-scale fusion module has a second-level multi-scale fusion module that uses an adaptive mixing mechanism to adaptively fuse the output features of the second encoder and the output features of the third encoder, and then inputs the fused features to the second decoder; its first-level multi-scale fusion module uses an adaptive mixing mechanism to adaptively fuse the output features of the first encoder and the output features of the second decoder, and then inputs the fused features to the first decoder. The output features of the first decoder are directly fed into the output convolutional layer to generate the three-dimensional radiotherapy dose distribution prediction results corresponding to the training set, validation set and test set through the output convolutional layer; The output features of the first decoder serve as the supervision object of the first loss function; the output features of the second decoder serve as the supervision object of the second loss function; and the three-dimensional radiotherapy dose distribution prediction result serves as the supervision object of the main loss function. Step S10: Based on the first loss function, the second loss function, and the main loss function, and using the training set, train the three-dimensional dose distribution prediction model; Step S11: Input the validation set into the trained three-dimensional dose distribution prediction model for optimization; Step S12: Input the test set into the final optimized three-dimensional dose distribution prediction model to obtain the three-dimensional radiotherapy dose distribution prediction result corresponding to the test set.

2. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 1, characterized in that, In step S2, the preprocessed radiotherapy planning parameters include: beam geometry parameters and control point related parameters; the beam geometry parameters and control point related parameters include: gantry angle, collimator angle, treatment bed angle, source axis distance, multi-leaf collimator control point opening information, and monitoring unit weight; the multi-leaf collimator control point opening information includes: multi-leaf collimator opening profile.

3. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 2, characterized in that, Step S4 specifically includes: Step S4.1: Based on the gantry angle, collimator angle, treatment bed angle, and source axis distance of the b-th beam, determine the incident direction and projection relationship of the b-th beam in the patient's three-dimensional coordinate system, so as to establish the geometric mapping relationship between the b-th beam and the patient's voxel space. Step S4.2: For the b-th beam, read the multi-leaf collimator opening profile and corresponding monitoring unit weights under all control points of the b-th beam to form the control point parameter sequence of the b-th beam; wherein, the control point parameter sequence is used to reflect the dynamic changes of the opening shape and monitoring unit weights of the b-th beam throughout the entire delivery process.

4. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 3, characterized in that, Step S5 specifically includes: Step S5.1: Based on the control point parameter sequence of the b-th beam, according to the opening profile and beam incident direction corresponding to each control point, project the opening region of the corresponding control point onto the patient voxel space to characterize the potential coverage range of the corresponding control point to different spatial locations in the patient's body. Step S5.2: The projection results of all control points of the same beam are weighted and summed according to their monitoring unit weights to generate the cumulative aperture response map of the beam-level control points.

5. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 4, characterized in that, Step S6 specifically includes: Step S6.1: Determine the possible incident boundary position of the beam based on the patient's body surface mask or the patient's outer contour, and determine the effective incident propagation path range of the beam entering the patient's body from outside the body, so as to establish the correspondence between the body surface and the beam. Step S6.2: Determine whether the b-th beam can hit the path of the target voxel along its incident direction, and generate the surface hit weight map based on the determination result; when the determination result is yes, assign the corresponding surface hit weight to the voxel position within the effective incident propagation path of the beam; when the determination result is no, reduce or set the corresponding weight to zero for the voxel position not within the effective incident propagation path of the beam.

6. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 5, characterized in that, Step S7 specifically includes: Step S7.1: For voxel location p, calculate the length of the in vivo propagation path of the b-th beam from the incident point on the body surface to the voxel location p, in order to characterize the propagation depth of the beam in the patient's body. Step S7.2: Based on the preset attenuation model, the lengths of the propagation paths within the body are converted into corresponding depth attenuation weights to generate a depth attenuation weight map; wherein, the preset attenuation model adopts an exponential attenuation form, a percentage depth dose curve mapping method, an empirical piecewise function method, or other models that can characterize the propagation attenuation trend within the beam fluid.

7. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 6, characterized in that, Step S8 specifically includes: Step S8.1: Generate the prior value of the beam-level pseudo-field at voxel position p for the b-th beam: ； In the formula, Indicates the cumulative open response of the control point; Indicates the surface hit weight; Indicates the depth decay weight; Step S8.2: Perform steps S4 to S7 and step 8.1 for all beams respectively; Step S8.3: Generate the cumulative beam channel, average propagation depth channel, and coverage statistics channel; Step S8.4: Based on the cumulative beam channel, average propagation depth channel, coverage statistics channel, and the beam-level pseudo-field prior value of each beam at voxel position p, generate the beam-level pseudo-field prior tensor corresponding to all beams to enhance the overall beam morphology expression capability. ； In the formula, B represents the beam-level pseudo-field prior tensor for all beams; B represents the number of beams. Prior value of the beam level pseudo-field at voxel position p for the B-th beam; Indicates the cumulative beam current channel; Indicates the average propagation depth channel; This indicates coverage of statistical channels.

8. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 7, characterized in that, In step S9, the corresponding structures of the encoder and decoder are implemented through a wavelet domain encoding-decoding structure.

9. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to any one of claims 1 to 8, characterized in that, In step S10, the total loss function for training the three-dimensional dose distribution prediction model is expressed as: ； In the formula, Represents the total loss function; ; and Let these represent the first loss function and the second loss function, respectively. , and These represent the weighting coefficients of the corresponding loss terms.

10. The method for predicting radiotherapy dose distribution based on prior guidance of beam-level pseudo-fields according to claim 9, characterized in that, Main loss function The voxel-level mean absolute error loss function is used; where the main loss function is... Represented as: ； In the formula, This represents the predicted dose value of the main output branch at the i-th voxel position; represents the reference dose value at the i-th voxel position; N represents the total number of voxels involved in the loss calculation.

Images