Training method and system of dose distribution prediction model fusing anatomy and beam information

Abstract

The present disclosure provides a training method and system of a dose distribution prediction model fusing anatomy and beam information, wherein the training method of the dose distribution prediction model comprises: obtaining a plurality of groups of sample training data corresponding to different sample structures; wherein any one group of sample training data comprises: training sample fusion information and corresponding training sample dose distribution information under the sample fusion information; the sample fusion information is determined based on sample anatomy structure information and sample beam configuration information corresponding to an esophageal cancer patient; a first preset model is trained based on each group of sample fusion information and corresponding sample dose distribution information to obtain a dose distribution prediction model for predicting predicted dose distribution information corresponding to a target structure under target beam configuration information. The dose distribution prediction model trained by the present disclosure can quickly realize dose prediction in the pre-planning stage, provide guidance for beam selection, and improve the efficiency and consistency of radiotherapy plan preparation.

Description

Technical Field

[0001] This disclosure relates to the field of medical data processing technology, and in particular to a training method and system for a dose distribution prediction model that integrates anatomical and beam information. Background Technology

[0002] Esophageal cancer is a common malignant tumor of the chest, and intensity-modulated radiotherapy (IMRT) is one of its core treatment methods. Developing radiotherapy plans for esophageal cancer faces many challenges: the target area is long and irregular in shape, and is anatomically close to critical organs such as the lungs, heart, and spinal cord. Clinically, complex prescription strategies such as simultaneous dose escalation and boosting are often employed. Furthermore, to balance target coverage and organ protection, the number and angle of beams in the radiotherapy plan are individually designed based on the patient's unique anatomical characteristics, resulting in highly heterogeneous beam configurations.

[0003] Existing deep learning-based methods for predicting three-dimensional dose distribution in radiotherapy are mostly anatomically driven models. These models implicitly assume that the beam configurations of the training and test sets are homogeneous, which can lead to insufficient prediction stability under heterogeneous beam configurations in esophageal cancer radiotherapy. In particular, the prediction error increases significantly in low- and medium-dose regions (such as the lungs and heart) dominated by beam direction and superposition effects.

[0004] Traditional dose prediction and planning methods suffer from inefficiency and poor consistency: physicists must repeatedly adjust beam configurations and optimize parameters through trial and error, making plan development time-consuming; plans developed by physicists with different experience levels exhibit significant dose differences, affecting the uniformity of radiotherapy quality. Existing dose prediction methods that integrate beam information often rely on optimized data such as beam weights and MLC (multi-leaf collimator) sequences, or involve complex modeling processes and insufficient representation of beam geometry, failing to provide guidance for beam selection during the pre-planning stage and limiting their clinical applicability. Summary of the Invention

[0005] The technical problem to be solved by this disclosure is to overcome the shortcomings of the prior art in that it cannot provide guidance for beam selection in the pre-planning stage, and to provide a method for training and predicting dose distribution prediction models that integrate anatomical and beam information.

[0006] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0007] According to a first aspect of this disclosure, a training method for a dose distribution prediction model that integrates anatomical and beam information is provided, the training method comprising:

[0008] Obtain several sets of sample training data corresponding to different sample structures;

[0009] Wherein, any set of the sample training data includes: training sample fusion information and training sample dose distribution information corresponding to the sample fusion information; the sample fusion information is determined based on the sample anatomical structure information and sample beam configuration information corresponding to esophageal cancer patients.

[0010] A first preset model is trained based on the sample fusion information and the corresponding sample dose distribution information of each group to obtain the dose distribution prediction model used to predict the predicted dose distribution information of the target structure under the target beam configuration information.

[0011] Optionally, the step of obtaining the sample fusion information includes:

[0012] Based on the anatomical information of the sample, the anatomical features of the sample structure are obtained.

[0013] The sample anatomical features are multi-channel anatomical features, with each channel corresponding to a mask for an organ at risk or a planned target area mask or body contour mask for embedding a prescription dose.

[0014] Based on the sample beam configuration information, obtain the sample beam geometric features;

[0015] The sample beam geometry features include a normalized beam coverage map and a beam overlap mask.

[0016] Spatial registration and resampling are performed on the anatomical features of the sample, the normalized beam coverage map, and the beam overlap mask;

[0017] The sample anatomical features, normalized beam coverage map, and beam overlap mask of the resampled sample are stitched together to obtain the sample fusion information.

[0018] Optionally, the sample anatomical information includes: a planned target area mask embedding prescription dosage information, several organs at risk masks, and a body contour mask;

[0019] The sample beam configuration information includes: beam quantity information and / or beam angle information;

[0020] The step of obtaining the sample anatomical structure features based on the sample anatomical structure information includes:

[0021] The prescription dose information is assigned to the planned target area mask;

[0022] The planned target area mask with embedded prescription dosage information and several organs at risk masks are respectively mapped to different independent channels;

[0023] The independent channels corresponding to all the planned target area masks, organs at risk masks, and body contour masks that embed prescription dosage information form the anatomical features of the sample.

[0024] And / or,

[0025] The step of obtaining the sample beam geometric features based on the beam configuration information includes:

[0026] Based on the beam quantity information and the beam angle information, ray tracing is performed from the radiation source along the central axis of each beam to determine the intersection relationship between the ray and the patient's voxel space;

[0027] The number of times each voxel intersects with the beam rays is counted to obtain a beam count voxel map;

[0028] The beam count voxel map is normalized according to the total number of beams to generate a normalized beam coverage map, so as to obtain the proportion of each voxel in the patient voxel space that is traversed by the beam based on the normalized beam coverage map.

[0029] A preset beam overlap threshold is obtained, and beam overlap masks under different thresholds are generated based on the beam count voxel map; the beam overlap threshold is the minimum number of overlaps for each voxel in the patient voxel space.

[0030] Optionally, after obtaining the dose distribution prediction model for predicting the target dose distribution information corresponding to the target structure under the target beam configuration information, the training method further includes:

[0031] Acquire fusion information of several sets of validation samples and corresponding validation dose distribution information;

[0032] The validation sample fusion information is input into the dose distribution prediction model to obtain the predicted dose distribution information corresponding to the validation sample fusion information.

[0033] Based on the verified dose distribution information and the predicted dose distribution information, difference information is obtained;

[0034] If the difference information is less than a first threshold, the dose distribution prediction model is output.

[0035] In response to the difference information being greater than or equal to the first threshold, the mean square error is obtained based on the verified dose distribution information and the predicted dose distribution information. The weighted sum of the mean square errors is used as a loss function to optimize the dose distribution prediction model until the difference information is less than the first threshold, and the optimized dose distribution prediction model is output.

[0036] And / or,

[0037] The sample structure includes at least one of the following: the target area structure corresponding to esophageal cancer patients, heart structure, spinal cord structure, lung structure, and body contour.

[0038] And / or,

[0039] The first preset model includes a neural network model based on AS-NeSt (a three-dimensional asymmetric distributed attention residual network for voxel-level dose prediction).

[0040] According to a second aspect of this disclosure, a dose distribution prediction method is provided, the prediction method comprising:

[0041] Obtain target sample fusion information of the target structure to be predicted, and input the target sample fusion information into the dose distribution prediction model trained by the training method of the dose distribution prediction model as described in the first aspect of this disclosure;

[0042] Output the target dose distribution information corresponding to the target sample fusion information.

[0043] According to a third aspect of this disclosure, a training system for a dose distribution prediction model that integrates anatomical and beam information is provided, the training system comprising:

[0044] The data acquisition module is used to acquire several sets of sample training data corresponding to different sample structures.

[0045] Wherein, any set of the sample training data includes: training sample fusion information and training sample dose distribution information corresponding to the sample fusion information; the sample fusion information is determined based on the sample anatomical structure information and sample beam configuration information corresponding to esophageal cancer patients.

[0046] The model training module is used to train a first preset model based on the fusion information of each group of samples and the corresponding sample dose distribution information, so as to obtain the dose distribution prediction model for predicting the predicted dose distribution information of the target structure under the target beam configuration information.

[0047] Optionally, the data acquisition module is further configured to:

[0048] Based on the anatomical information of the sample, the anatomical features of the sample structure are obtained.

[0049] The sample anatomical features are multi-channel anatomical features, with each channel corresponding to a mask for endangered organs or a planned target area mask or a body contour mask that embeds prescription dosage information.

[0050] Based on the sample beam configuration information, obtain the sample beam geometric features;

[0051] The sample beam geometry features include a normalized beam coverage map and a beam overlap mask.

[0052] Spatial registration and resampling are performed on the anatomical features of the sample, the normalized beam coverage map, and the beam overlap mask;

[0053] The sample anatomical features, normalized beam coverage map, and beam overlap mask of the resampled sample are stitched together to obtain the sample fusion information.

[0054] Optionally, the sample anatomical information includes: a planned target area mask embedding prescription dosage information, several organs at risk masks, and a body contour mask;

[0055] The sample beam configuration information includes: beam quantity information and / or beam angle information;

[0056] The data acquisition module is also used for:

[0057] The prescription dose information is assigned to the planned target area mask;

[0058] The planned target area mask, several organs at risk mask and body contour mask with embedded prescription dosage information are respectively mapped to different independent channels;

[0059] The independent channels corresponding to all the planned target area masks, organs at risk masks, and body contour masks that embed prescription dosage information form the anatomical features of the sample.

[0060] And / or,

[0061] The data acquisition module is also used for:

[0062] Based on the beam quantity information and the beam angle information, ray tracing is performed from the radiation source along the central axis of each beam to determine the intersection relationship between the ray and the patient's voxel space;

[0063] The number of times each voxel intersects with the beam rays is counted to obtain a beam count voxel map;

[0064] The beam count voxel map is normalized according to the total number of beams to generate a normalized beam coverage map, so as to obtain the proportion of each voxel in the patient voxel space that is traversed by the beam based on the normalized beam coverage map.

[0065] A preset beam overlap threshold is obtained, and beam overlap masks under different thresholds are generated based on the beam count voxel map; the beam overlap threshold is the minimum number of overlaps for each voxel in the patient voxel space.

[0066] Optionally, the training system further includes: a model validation module, which is used after obtaining the dose distribution prediction model for predicting the target dose distribution information corresponding to the target structure under the target beam configuration information:

[0067] Acquire fusion information of several sets of validation samples and corresponding validation dose distribution information;

[0068] The validation sample fusion information is input into the dose distribution prediction model to obtain the predicted dose distribution information corresponding to the validation sample fusion information.

[0069] Based on the verified dose distribution information and the predicted dose distribution information, difference information is obtained;

[0070] If the difference information is less than a first threshold, the dose distribution prediction model is output.

[0071] In response to the difference information being greater than or equal to the first threshold, the mean square error is obtained based on the verified dose distribution information and the predicted dose distribution information. The weighted sum of the mean square errors is used as a loss function to optimize the dose distribution prediction model until the difference information is less than the first threshold, and the optimized dose distribution prediction model is output.

[0072] And / or,

[0073] The sample structure includes at least one of the following: the target area structure corresponding to esophageal cancer patients, heart structure, spinal cord structure, lung structure, and body contour.

[0074] And / or,

[0075] The first preset model includes a neural network model based on AS-NeSt.

[0076] According to a fourth aspect of this disclosure, a dose distribution prediction system is provided, the prediction system comprising:

[0077] The target sample fusion information module is used to obtain target sample fusion information of the target structure to be predicted, and input the target sample fusion information into the dose distribution prediction model trained by the training system of the dose distribution prediction model as described in the third aspect of this disclosure.

[0078] The target information output module is used to output the target dose distribution information corresponding to the target sample fusion information.

[0079] According to a fifth aspect of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and for running on the processor, wherein the processor executes the computer program to implement a training method for a dose distribution prediction model according to a first aspect of this disclosure, and / or a dose distribution prediction method according to a second aspect of this disclosure.

[0080] According to a sixth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a training method for a dose distribution prediction model according to a first aspect of this disclosure, and / or a dose distribution prediction method according to a second aspect of this disclosure.

[0081] According to a seventh aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements a training method for a dose distribution prediction model as described in a first aspect of this disclosure, and / or a dose distribution prediction method as described in a second aspect of this disclosure.

[0082] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of this disclosure.

[0083] The positive and progressive effects of this disclosure are as follows:

[0084] The training method for the dose distribution prediction model provided in this disclosure introduces beam angle information available in the pre-planning stage into dose prediction, constructs voxel-level beam geometry features through ray tracing and fuses them with anatomical structure features, and trains a high-precision, high-robust dose prediction model. Based on this dose model, dose prediction in the pre-planning stage is realized, providing guidance for beam selection and improving the efficiency and consistency of radiotherapy planning. Attached Figure Description

[0085] Figure 1 A flowchart illustrating the training method for the dose distribution prediction model provided in Embodiment 1 of this disclosure;

[0086] Figure 2 This is a schematic diagram of the process for obtaining sample fusion information provided in Embodiment 1 of this disclosure;

[0087] Figure 3 This is a schematic diagram of the process for obtaining the geometric features of a sample beam according to Embodiment 1 of this disclosure;

[0088] Figure 4 The difference graph of the isodose volume DSC (Dice similarity coefficient) curves of the AADP model and the ADP model in the prediction results on the conventional test set provided in Embodiment 1 of this disclosure;

[0089] Figure 5A plot showing the difference in isodose volume DSC curves between the AADP model and the ADP model in a rare beam configuration test set, as provided in Embodiment 1 of this disclosure.

[0090] Figure 6 A flowchart illustrating the verification of model accuracy provided in Embodiment 1 of this disclosure;

[0091] Figure 7 This is a schematic diagram of the structure of the neural network model provided in Embodiment 1 of this disclosure;

[0092] Figure 8 Beam geometry feature diagram of a representative five-field case provided in Embodiment 1 of this disclosure;

[0093] Figure 9 This is a flowchart illustrating the dose distribution prediction method provided in Embodiment 2 of this disclosure;

[0094] Figure 10 This is a schematic diagram of the structure of the training system for the dose distribution prediction model provided in Embodiment 3 of this disclosure;

[0095] Figure 11 This is a schematic diagram of the dose distribution prediction system provided in Embodiment 4 of this disclosure;

[0096] Figure 12 This is a schematic diagram of the structure of the electronic device provided in Embodiment 5 of this disclosure. Detailed Implementation

[0097] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0098] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the claims or the context of the embodiments, and should not be construed as an unnecessary limitation. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0099] In this embodiment of the disclosure, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information comply with relevant laws and regulations and do not violate public order and good morals.

[0100] Example 1

[0101] like Figure 1As shown, this embodiment provides a training method for a dose distribution prediction model. The training method includes:

[0102] S11: Obtain several sets of sample training data corresponding to different sample structures;

[0103] Any set of sample training data includes: training sample fusion information and corresponding training sample dose distribution information under the sample fusion information; the sample fusion information is determined based on the sample anatomical structure information and sample beam configuration information corresponding to esophageal cancer patients;

[0104] S12: Train the first preset model based on the fusion information of each group of samples and the corresponding sample dose distribution information to obtain a dose distribution prediction model for predicting the predicted dose distribution information of the target structure under the target beam configuration information.

[0105] By incorporating beam angle information available in the pre-planning stage into dose prediction, constructing voxel-level beam geometry features through ray tracing and fusing them with anatomical features, a high-precision and robust dose prediction model is trained to achieve dose prediction in the pre-planning stage, providing guidance for beam selection and improving the efficiency and consistency of radiotherapy planning.

[0106] Among them, such as Figure 2 As shown, step S11 includes:

[0107] S111: Based on the sample's anatomical structure information, obtain the sample's anatomical structure features.

[0108] Among them, the anatomical features of the sample are multi-channel anatomical features, with each channel corresponding to a mask of an organ at risk or a planned target area mask or a body contour mask that embeds prescription dosage information.

[0109] S112: Based on the sample beam configuration information, obtain the sample beam geometric features; wherein, the sample beam geometric features include the normalized beam coverage map and the beam overlap mask.

[0110] S113: Spatial registration and resampling of sample anatomical features, normalized beam coverage map and beam overlap mask;

[0111] S114: The sample anatomical features, normalized beam coverage map, and beam overlap mask of the resampled samples are stitched together to obtain sample fusion information.

[0112] By spatially registering and resampling anatomical features, normalized beam coverage maps, and beam overlap masks, the voxel resolution of each feature is made consistent. Then, channel splicing is performed to obtain input features, ensuring spatial consistency of the model input.

[0113] The sample anatomical structure information in this embodiment includes: a planned target area mask with embedded prescription dosage information, several organ-at-risk masks, and a body contour mask;

[0114] The sample beam configuration information includes: beam quantity information and / or beam angle information;

[0115] Step S111 includes:

[0116] Incorporate prescription dosage information into the planned target area mask;

[0117] The planned target area mask, several organs at risk mask and body contour mask with embedded prescription dosage information are respectively mapped to different independent channels;

[0118] The independent channels corresponding to all the planned target area masks, organ at risk masks, and body contour masks that embed prescription dosage information form the anatomical features of the sample.

[0119] like Figure 3 As shown, step S112 includes:

[0120] S1121: Based on the beam number information and beam angle information, perform ray tracing from the radiation source along the central axis of each beam to determine the intersection relationship between the ray and the patient's voxel space;

[0121] S1122: Count the number of times each voxel is intersected by the beam rays to obtain the beam count voxel map;

[0122] Beam count voxel maps visually reflect the degree of overlap of beam irradiation on different parts of the patient's body, and are the basis for analyzing target coverage uniformity, normal organ irradiation dose distribution, and beam redundancy.

[0123] S1123: Normalize the beam count voxel map according to the total number of beams to generate a normalized beam coverage map, so as to obtain the proportion of each voxel in the patient voxel space that is traversed by the beam based on the normalized beam coverage map.

[0124] Dividing the beam count of each voxel by the total number of beams (i.e., the total number of beams) yields a normalized value, typically ranging from 0 to 1. This normalized beam coverage map can be considered a beam coverage density map, representing the proportion of each voxel that is traversed by all beams. Normalization eliminates the dimensional effects caused by different total number of beams, allowing for cross-sectional comparisons of coverage between different plans or different patients.

[0125] S1124: Obtain a preset beam overlap threshold, and generate beam overlap masks under different thresholds based on the beam count voxel map; the beam overlap threshold is the minimum number of overlaps for each voxel in the patient voxel space.

[0126] In this process, a preset beam overlap threshold is set (e.g., ≥1, ≥2, ≥3 beams intersect), and then a binary beam overlap mask under different thresholds is generated based on the beam count voxel map.

[0127] By using a binarized beam overlap mask, the degree of beam overlap between the target area and organs at risk can be intuitively quantified, supporting the establishment of clinical constraints on the minimum number of beam coverages and improving the controllability and robustness of radiotherapy planning.

[0128] As shown in Table 1 and Figure 4 As shown in Table 1, the prediction errors of various dosimetric parameters of the AADP model (the method model described in this disclosure, i.e., the dose prediction model that integrates patient anatomical structure and beam geometry information) and the ADP model (the dose prediction model based solely on patient anatomical structure) on the conventional test set are compared. The average prediction error of the AADP model on the conventional test set decreased from 2.88% for the ADP model to 2.02%. Figure 4 This shows the difference curve of the Dice similarity coefficient (DSC) of the isodose volume between the AADP model and the ADP model on the standard test set. Figure 4 The x-axis represents the isodose volume, and the y-axis represents the difference between the AADP model prediction result and the Dice similarity coefficient of the ADP model prediction result. The AADP model's DSC (a statistical indicator used to measure the similarity between two sets or samples, here used to measure the similarity between the predicted dose and the clinically true dose) is consistently higher than the ADP model across the isodose volume range, increasing from an average DSC value of 0.9 to 0.93. Further details are shown in Table 2. Figure 5 As shown in Table 2, the prediction errors of the AADP model and the ADP model for various dosimetric indices are compared on the rare beam configuration test set. The average prediction error of the AADP model for rare beam configurations is reduced to 3.08%, which is far better than the 4.52% of the ADP model. The average DSC value is improved from 0.79 to 0.87, which solves the robustness problem of the model driven by only anatomical structure under heterogeneous beam configurations. The model trained in this disclosure (AADP model) significantly reduces the dose prediction error of the target area and organs at risk of conventional models (such as the ADP model) that do not consider the geometric information of the radiation field. In particular, the prediction accuracy is significantly improved in low and medium dose areas such as the lungs and heart.

[0129] Furthermore, the beam geometry features used in this disclosure are generated based solely on the number and angle of beams available in the pre-planning stage, without requiring optimized data such as beam weights, MLC (multi-leaf collimator) sequences, or machine hop counts. This enables dose prediction before radiotherapy planning optimization, providing direct guidance for physicists in beam selection and filling the gap in dose prediction in the pre-planning stage of existing technologies.

[0130] Table 1

[0131]

[0132] Table 2

[0133]

[0134] like Figure 6 As shown, after obtaining the dose distribution prediction model for predicting the target dose distribution information corresponding to the target structure under the target beam configuration information, the training method in this embodiment further includes:

[0135] S61: Obtain fusion information of several sets of validation samples and corresponding validation dose distribution information;

[0136] S62: Input the validation sample fusion information into the dose distribution prediction model to obtain the predicted dose distribution information corresponding to the validation sample fusion information;

[0137] S63: Obtain the difference information based on the verified dose distribution information and the predicted dose distribution information; if the difference information is less than the first threshold, output the dose distribution prediction model; if the difference information is greater than or equal to the first threshold, obtain the mean square error based on the verified dose distribution information and the predicted dose distribution information, use the weighted sum of the mean square errors as the loss function to optimize the dose distribution prediction model until the difference information is less than the first threshold, and output the optimized dose distribution prediction model.

[0138] By using threshold judgment in the validation process, we ensure that the final output model meets the clinically acceptable accuracy standards. We use weighted mean square error as the loss function, which allows clinical priorities (such as emphasizing the weight of target dose accuracy) to be incorporated into the model training process, making the model more in line with actual treatment needs.

[0139] The sample structures in this embodiment include at least one of the following: the planned target area structure corresponding to esophageal cancer patients, heart structure, spinal cord structure, lung structure, and body contour.

[0140] like Figure 7 As shown, the first preset model includes a neural network model based on AS-NeSt.

[0141] The model uses an AS-NeSt-based neural network as its foundation. This model is based on 3D U-Net (a deep learning architecture designed specifically for 3D medical image segmentation) and incorporates the segmentation attention mechanism of 3D ResNeSt blocks (a neural network module for 3D deep learning). It includes 4 encoding stages and 2 decoding stages, and retains spatial context information through skip connections. The 6-7+N channel input features are input into the model to output a single-channel 3D dose distribution prediction map.

[0142] The implementation principle of the training method for the dose distribution prediction model in this embodiment is explained below with specific implementation details:

[0143] First, sample acquisition and preprocessing were carried out, specifically including: collecting clinical data from 751 patients with esophageal cancer undergoing intensity-modulated radiotherapy, and screening out cases containing complete CT images, structural files, beam configuration information (4-9 fields, personalized beam angle design), and clinically approved three-dimensional dose distributions. These cases were divided into a model development set (576 cases, 500 for training and 76 for validation), an independent test set (100 cases), a rare beam configuration set (33 cases), and a clinical utility evaluation set (42 cases).

[0144] The anatomical structures (including the planned target volume (PTV), organs at risk (OARs), and body contour) of all cases were encoded: three-dimensional segmentation masks were generated for PTV, heart, spinal cord, lung, and body contour as independent channels; each patient contained 1-2 PTVs, and the prescribed dose (41.4Gy / 50.4Gy / 50.4Gy & 60.2Gy) was assigned corresponding voxel values ​​in the PTV channels. The PTV, OARs, and body contour formed a total of 5-6 channels of anatomical structure features.

[0145] Secondly, the geometric features of the generated beam specifically include: First, let... Representing the patient's three-dimensional voxel space, This represents any one of the voxels. For each instance containing... Strip beam ( The IMRT (Intensity Modulated Radiotherapy) protocol involves tracing the beam along the central axis of each beam, starting from the linear accelerator radiation source. During this process, only the intersection of the beam with the patient's body contour is determined, without considering modulation information such as MU (Monitoring Unit, used to control the amount of X-rays or electron beams exposed during treatment), beam weighting, or MLC (Multi-Leaf Collimator).

[0146] To characterize the coverage of a single beam over a patient's anatomical structures at the voxel level, the first... The indicator function corresponding to the strip beam is:

[0147]

[0148] Based on the above definition, a three-dimensional beam counting volume can be further constructed:

[0149] in, Indicated in voxels The number of beam paths covering a given location in space can be visually represented by voxel-level beam counting, which can intuitively reflect the spatial distribution trend of beams within the patient's body.

[0150] Considering the differences in the number of beams among different patients, directly using Variations in the number of beams may introduce model training bias, such as causing the model to adopt the erroneous paradigm that more beam fields equate to higher doses. Therefore, this disclosure further normalizes the beam counting volume:

[0151] This yields a normalized beam coverage map with values ​​ranging from [0, 1].

[0152] This normalization process makes the beam coverage features comparable under different beam configurations and helps the model stably learn beam spatial information across patients and protocols. (Five representative patients) For example Figure 8 As shown in (b); where, Figure 8 Representative five-field IMRT (Intensity Modulated Radiotherapy) cases: beam geometry representation based on ray tracing technology: (a) spatial distribution of beam paths within anatomical structures; (b) normalized beam coverage map (the beam coverage area is colored from light to dark according to the beam density of ⅕-1); (c-g) thresholded beam overlap mask (corresponding to the intersection areas of ≥1 to ≥5 beams respectively).

[0153] Using only continuous beam coverage values ​​may be insufficient to fully characterize the overlapping structural features of beams within the body. To enhance the model's ability to express beam overlap patterns, this disclosure further constructs a set of thresholded beam overlap masks based on beam counting volume:

[0154]

[0155] in, This indicates a pre-defined beam overlap level; different

[0156] The mask describes the distribution characteristics of beam intersection density at a voxel in a hierarchical manner: smaller... The overall range covered by the anti-mapping beam, while the larger This highlights spatial regions where multiple beams highly overlap. This hierarchical representation helps the model distinguish between low-beam-density regions and high-beam-intersection regions, thereby more effectively learning the potential correlation between beam geometry and dose distribution patterns.

[0157] During the model input construction phase, the normalized beam coverage map is... (1 channel) and its corresponding multi-level beam overlap mask (like Figure 8(cg) (N channels) are treated as independent channels and spliced ​​together with the patient's anatomical structure channels (including PTV, OARs and body contour mask) (5-6 channels) in the channel dimension to form the final multi-channel input of the network (6-7+N channels).

[0158] After obtaining the anatomical structure features and channel beam geometry features, feature fusion is required. This includes: spatial registration of the 5-6 channel anatomical structure features and the N+1 channel beam geometry features, and uniform resampling to voxel resolution. The two channels are concatenated to obtain 6-7+N channels of model input features; simultaneously, the clinical 3D dose distribution annotation map is resampled to... , used as a label for model training.

[0159] After obtaining the fused features, the model is further trained. The model training process specifically includes:

[0160] The model uses an AS-NeSt-based neural network as its foundation. This model is based on 3D U-Net (a deep learning architecture designed for 3D medical image segmentation) and incorporates the segmentation attention mechanism of 3D ResNeSt blocks (a neural network module for 3D deep learning). It includes 4 encoding stages and 2 decoding stages, and preserves spatial context information through skip connections.

[0161] Inputting 6-7+N channels into the feature model outputs a single-channel 3D dose distribution prediction map. Feature fusion and model schematic diagrams are shown below. Figure 5 As shown;

[0162] The loss function is a weighted sum of the mean squared errors of global voxels and PTV voxels:

[0163]

[0164] Among them, D C For clinically labeled dosage, D P Here, N represents the global voxel count, and the target voxel sets are respectively... The corresponding number of voxels are respectively and If there is only one target region, then M=0. The Adam optimizer (an adaptive learning rate optimization algorithm used in deep learning) is employed, with the learning rate set to... The batch size was 2, and the training was conducted for 400 rounds on two NVIDIA A100 GPUs (a type of graphics processor). Training was stopped when the loss function value on the validation set tended to stabilize, resulting in a three-dimensional dose distribution prediction model for esophageal cancer radiotherapy.

[0165] As shown in Table 3 (Table 3 compares the efficiency and dose consistency of radiotherapy planning process under AADP model-assisted and non-assisted conditions), the assisted planning based on the model disclosed in this publication can shorten the planning time for junior physicists by more than 65% and reduce the number of planning modification iterations by 59.68%. At the same time, it reduces the dose difference between physicists with different experience levels, with the dose difference among junior physicists reduced by 80.58%, significantly improving the uniformity of radiotherapy planning and clinical work efficiency. The beam geometry features of this publication are generated based on ray tracing, directly characterizing the penetration path and superposition pattern of the beam in the patient voxel space, which conforms to the physical laws of radiotherapy dose deposition. Compared with purely data-driven feature fusion methods, it has stronger physical interpretability, and the model prediction results are more easily accepted by clinical physicists.

[0166] Table 3

[0167]

[0168] The training method for the dose distribution prediction model provided in this disclosure introduces beam angle information available in the pre-planning stage into dose prediction, constructs voxel-level beam geometry features through ray tracing and fuses them with anatomical structure features, and trains a high-precision, high-robust dose prediction model. Based on this dose model, dose prediction in the pre-planning stage is realized, providing guidance for beam selection and improving the efficiency and consistency of radiotherapy planning.

[0169] Example 2

[0170] like Figure 9 As shown, this embodiment provides a dose distribution prediction method, which includes:

[0171] S91: Obtain the target sample fusion information of the target structure to be predicted, and input the target sample fusion information into the dose distribution prediction model trained by the training method of the dose distribution prediction model;

[0172] S92: Outputs the target dose distribution information corresponding to the target sample fusion information.

[0173] The implementation principle of the dose distribution prediction method in this embodiment is explained below with specific implementation methods:

[0174] The data to be predicted includes: obtaining radiotherapy pre-planning data for a certain esophageal cancer patient, including PTV, heart, spinal cord, lung, and body contour mask obtained from chest CT image segmentation, prescription dose of 50.4 Gy, and the pre-planned 6-field beam angles (20°, 60°, 100°, 260°, 300°, 340°).

[0175] Feature processing specifically includes: following the steps in Example 1, performing ray tracing-based geometric feature acquisition on the 6 field beam angles to generate a normalized beam coverage map with 1 channel and a 6-level beam overlap mask (6 channels), for a total of 7 channels of beam geometric features; encoding the mask and prescription dose to generate 5 channels of anatomical structure features; and spatially registering and resampling the two to stitch them together into 12 channels of input features.

[0176] Dose prediction specifically includes: inputting 12-channel input features into the dose prediction model trained in Example 1, with a model inference time of approximately 6.74ms, and outputting a three-dimensional dose distribution map of the patient. This distribution map can clearly show the dose coverage of the PTV and the radiation dose distribution of the lungs, heart, and spinal cord, providing guidance for physicists to optimize beam configuration.

[0177] The dose distribution prediction method provided in this disclosure introduces beam angle information available in the pre-planning stage into dose prediction, constructs voxel-level beam geometry features through ray tracing and fuses them with anatomical structure features, and trains a high-precision, high-robust dose prediction model. Based on this dose model, dose prediction in the pre-planning stage is realized, providing guidance for beam selection and improving the efficiency and consistency of radiotherapy planning.

[0178] Example 3

[0179] like Figure 10 As shown, a training system for a dose distribution prediction model is provided. The training system includes:

[0180] Data acquisition module 101 is used to acquire several sets of sample training data corresponding to different sample structures;

[0181] Any set of sample training data includes: training sample fusion information and corresponding training sample dose distribution information under the sample fusion information; the sample fusion information is determined based on the sample anatomical structure information and sample beam configuration information corresponding to esophageal cancer patients;

[0182] The model training module 102 is used to train a first preset model based on the fusion information of each group of samples and the corresponding sample dose distribution information, so as to obtain a dose distribution prediction model for predicting the predicted dose distribution information of the target structure under the target beam configuration information.

[0183] In this embodiment, the data acquisition module 101 is also used for:

[0184] Based on the anatomical information of the sample, the anatomical features of the sample structure are obtained.

[0185] Among them, the anatomical features of the sample are multi-channel anatomical features, with each channel corresponding to a mask of an organ at risk or a planned target area mask or a body contour mask that embeds prescription dosage information.

[0186] Based on the sample beam configuration information, obtain the geometric features of the sample beam;

[0187] Among them, the sample beam geometry features include the normalized beam coverage map and the beam overlap mask;

[0188] Spatial registration and resampling were performed on the anatomical features of the samples, the normalized beam coverage map, and the beam overlap mask.

[0189] The sample anatomical features, normalized beam coverage map, and beam overlap mask of the resampled samples are stitched together to obtain sample fusion information.

[0190] The sample anatomical structure information in this embodiment includes: a planned target area mask with embedded prescription dosage information, several organ-at-risk masks, and a body contour mask;

[0191] The sample beam configuration information includes: beam quantity information and / or beam angle information;

[0192] The data acquisition module 101 is also used for:

[0193] The prescription dose information is assigned to the planned target area mask;

[0194] The planned target area mask, several organs at risk mask and body contour mask with embedded prescription dosage information are respectively mapped to different independent channels;

[0195] The independent channels corresponding to all the planned target area masks, organs at risk masks, and body contour masks that embed prescription dosage information form the anatomical features of the sample.

[0196] And / or,

[0197] The data acquisition module 101 is also used for:

[0198] Based on the beam number and beam angle information, ray tracing is performed from the radiation source along the central axis of each beam to determine the intersection relationship between the ray and the patient's voxel space;

[0199] The number of times each voxel intersects with the beam rays is counted to obtain a beam count voxel map;

[0200] The beam count voxel map is normalized according to the total number of beams to generate a normalized beam coverage map, so as to obtain the proportion of each voxel in the patient voxel space that is traversed by the beam based on the normalized beam coverage map.

[0201] Obtain a preset beam overlap threshold, and generate beam overlap masks at different thresholds based on the beam count voxel map; the beam overlap threshold is the minimum number of overlaps for each voxel in the patient voxel space.

[0202] The training system in this embodiment further includes a model validation module 103, which is used to, after obtaining a dose distribution prediction model for predicting the target dose distribution information corresponding to the target structure under the target beam configuration information:

[0203] Acquire fusion information of several sets of validation samples and corresponding validation dose distribution information;

[0204] Input the validation sample fusion information into the dose distribution prediction model to obtain the predicted dose distribution information corresponding to the validation sample fusion information;

[0205] Difference information is obtained based on verified dose distribution information and predicted dose distribution information;

[0206] If the difference information is less than the first threshold, the dose distribution prediction model is output.

[0207] If the difference information is greater than or equal to the first threshold, the mean square error is obtained based on the verified dose distribution information and the predicted dose distribution information. The weighted sum of the mean square errors is used as the loss function to optimize the dose distribution prediction model until the difference information is less than the first threshold, and the optimized dose distribution prediction model is output.

[0208] And / or,

[0209] The sample structures include at least one of the following: the planned target area structure corresponding to esophageal cancer patients, heart structure, spinal cord structure, lung structure, and body contour.

[0210] And / or,

[0211] The first preset model includes a neural network model based on AS-NeSt.

[0212] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0213] The dose distribution prediction model training system provided in this disclosure introduces beam angle information available in the pre-planning stage into dose prediction, constructs voxel-level beam geometry features through ray tracing and fuses them with anatomical structure features, and trains a high-precision, high-robust dose prediction model. Based on this dose model, dose prediction in the pre-planning stage is realized, providing guidance for beam selection and improving the efficiency and consistency of radiotherapy planning.

[0214] Example 4

[0215] like Figure 11 As shown, this embodiment provides a dose distribution prediction system, which includes:

[0216] The target sample fusion information module 201 is used to obtain the target sample fusion information of the target structure to be predicted, and input the target sample fusion information into the dose distribution prediction model trained by the training system of the dose distribution prediction model of the third aspect of this disclosure.

[0217] The target information output module 202 is used to output the target dose distribution information corresponding to the target sample fusion information.

[0218] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0219] The dose distribution prediction system provided in this disclosure introduces beam angle information available in the pre-planning stage into dose prediction, constructs voxel-level beam geometry features through ray tracing and fuses them with anatomical structure features, and trains a high-precision, high-robust dose prediction model. Based on this dose model, dose prediction in the pre-planning stage is realized, providing guidance for beam selection and improving the efficiency and consistency of radiotherapy planning.

[0220] Example 5

[0221] Figure 12 This is a schematic diagram of the structure of an electronic device according to an example embodiment of the present disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the training method of the dose distribution prediction model described in Embodiment 1 of any of the above embodiments, and / or the dose distribution prediction method described in Embodiment 2. Figure 12The electronic device 120 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0222] like Figure 12 As shown, the electronic device 120 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 120 may include, but are not limited to: at least one processor 121, at least one memory 122, and a bus 123 connecting different system components (including memory 122 and processor 121).

[0223] Bus 123 includes a data bus, an address bus, and a control bus.

[0224] The memory 122 may include volatile memory, such as random access memory (RAM) 1221 and / or cache memory 1222, and may further include read-only memory (ROM) 1223.

[0225] The memory 122 may also include a program tool 1225 (or utility) having a set (at least one) program module 1224, such program module 1224 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0226] The processor 121 executes various functional applications and data processing by running computer programs stored in the memory 122, such as the training method of the dose distribution prediction model described in any of the above embodiments 1, and / or the dose distribution prediction method described in embodiment 2.

[0227] Electronic device 120 can also communicate with one or more external devices 124 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 125. Furthermore, electronic device 120 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public network, such as the Internet) via network adapter 126. As shown, network adapter 126 communicates with other modules of electronic device 120 via bus 123. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with electronic device 120, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0228] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0229] Example 6

[0230] This disclosure also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the training method for the dose distribution prediction model described in Embodiment 1 above, and / or the dose distribution prediction method described in Embodiment 2.

[0231] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0232] Example 7

[0233] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the training method for the dose distribution prediction model described in Embodiment 1 above, and / or the dose distribution prediction method described in Embodiment 2.

[0234] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0235] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A training method for a dose distribution prediction model that integrates anatomical and beam information, characterized in that, The training method includes: Obtain several sets of sample training data corresponding to different sample structures; Wherein, any set of the sample training data includes: training sample fusion information and training sample dose distribution information corresponding to the sample fusion information; the sample fusion information is determined based on the sample anatomical structure information and sample beam configuration information corresponding to esophageal cancer patients. A first preset model is trained based on the fusion information of each set of samples and the corresponding sample dose distribution information to obtain the dose distribution prediction model used to predict the predicted dose distribution information corresponding to the target structure under the target beam configuration information; The steps for obtaining the sample fusion information include: Based on the anatomical information of the sample, the anatomical features of the sample structure are obtained. The sample anatomical features are multi-channel anatomical features, with each channel corresponding to a mask for endangered organs or a planned target area mask or body mask that embeds prescription dosage information. Based on the sample beam configuration information, obtain the sample beam geometric features; The sample beam geometry features include a normalized beam coverage map and a beam overlap mask. Spatial registration and resampling are performed on the anatomical features of the sample, the normalized beam coverage map, and the beam overlap mask; The sample anatomical features, normalized beam coverage map, and beam overlap mask after resampling are stitched together to obtain the sample fusion information; The step of obtaining the sample beam geometric features based on the beam configuration information includes: Based on the beam quantity information and the beam angle information, ray tracing is performed from the radiation source along the central axis of each beam to determine the intersection relationship between the ray and the patient's voxel space; The number of times each voxel intersects with the beam rays is counted to obtain a beam count voxel map; The beam count voxel map is normalized according to the total number of beams to generate a normalized beam coverage map, so as to obtain the proportion of each voxel in the patient voxel space that is traversed by the beam based on the normalized beam coverage map. A preset beam overlap threshold is obtained, and beam overlap masks under different thresholds are generated based on the beam count voxel map; the beam overlap threshold is the minimum number of overlaps for each voxel in the patient voxel space.

2. The training method for the dose distribution prediction model integrating anatomical and beam information according to claim 1, characterized in that, The sample anatomical information includes: a planned target area mask with embedded prescription dosage information, several organ-at-risk masks, and a body contour mask; The sample beam configuration information includes: beam quantity information and / or beam angle information; The step of obtaining the sample anatomical structure features based on the sample anatomical structure information includes: The prescription dose information is assigned to the planned target area mask; The planned target area mask, several organs at risk mask and body contour mask with embedded prescription dosage information are respectively mapped to different independent channels; The independent channels corresponding to all the planned target area masks, organ at risk masks, and body contour masks that embed prescription dosage information form the anatomical features of the sample.

3. The training method for the dose distribution prediction model integrating anatomical and beam information according to claim 1 or 2, characterized in that, After the step of obtaining the dose distribution prediction model for predicting the target dose distribution information corresponding to the target structure under the target beam configuration information, the training method further includes: Acquire fusion information of several sets of validation samples and corresponding validation dose distribution information; The validation sample fusion information is input into the dose distribution prediction model to obtain the predicted dose distribution information corresponding to the validation sample fusion information. Based on the verified dose distribution information and the predicted dose distribution information, difference information is obtained; If the difference information is less than a first threshold, the dose distribution prediction model is output. In response to the difference information being greater than or equal to the first threshold, the mean square error is obtained based on the verified dose distribution information and the predicted dose distribution information. The weighted sum of the mean square errors is used as a loss function to optimize the dose distribution prediction model until the difference information is less than the first threshold, and the optimized dose distribution prediction model is output. And / or, The sample structure includes at least one of the following: the target area structure corresponding to esophageal cancer patients, heart structure, spinal cord structure, lung structure, and body contour. And / or, The first preset model includes a neural network model based on AS-NeSt.

4. A dose distribution prediction method, characterized in that, The prediction method includes: Obtain target sample fusion information of the target structure to be predicted, and input the target sample fusion information into the dose distribution prediction model trained by the training method of the dose distribution prediction model that fuses anatomical and beam information as described in any one of claims 1-3; Output the target dose distribution information corresponding to the target sample fusion information.

5. A training system for a dose distribution prediction model that integrates anatomical and beam information, characterized in that, The training system includes: The data acquisition module is used to acquire several sets of sample training data corresponding to different sample structures. Wherein, any set of the sample training data includes: training sample fusion information and training sample dose distribution information corresponding to the sample fusion information; the sample fusion information is determined based on the sample anatomical structure information and sample beam configuration information corresponding to esophageal cancer patients. The model training module is used to train a first preset model based on the sample fusion information of each group and the corresponding sample dose distribution information, so as to obtain the dose distribution prediction model used to predict the predicted dose distribution information corresponding to the target structure under the target beam configuration information; The data acquisition module is also used for: Based on the anatomical information of the sample, the anatomical features of the sample structure are obtained. The sample anatomical features are multi-channel anatomical features, with each channel corresponding to a mask for endangered organs or a planned target area mask or a body contour mask that embeds prescription dosage information. Based on the sample beam configuration information, obtain the sample beam geometric features; The sample beam geometry features include a normalized beam coverage map and a beam overlap mask. Spatial registration and resampling are performed on the anatomical features of the sample, the normalized beam coverage map, and the beam overlap mask; The sample anatomical features, normalized beam coverage map, and beam overlap mask after resampling are stitched together to obtain the sample fusion information. The data acquisition module is also used for: Based on the beam quantity information and the beam angle information, ray tracing is performed from the radiation source along the central axis of each beam to determine the intersection relationship between the ray and the patient's voxel space; The number of times each voxel intersects with the beam rays is counted to obtain a beam count voxel map; The beam count voxel map is normalized according to the total number of beams to generate a normalized beam coverage map, so as to obtain the proportion of each voxel in the patient voxel space that is traversed by the beam based on the normalized beam coverage map. A preset beam overlap threshold is obtained, and beam overlap masks under different thresholds are generated based on the beam count voxel map; the beam overlap threshold is the minimum number of overlaps for each voxel in the patient voxel space.

6. A dose distribution prediction system, characterized in that, The prediction system includes: The target sample fusion information module is used to acquire the target sample fusion information of the target structure to be predicted, and input the target sample fusion information into the dose distribution prediction model trained by the training system of the dose distribution prediction model that fuses anatomical and beam information as described in claim 5. The target information output module is used to output the target dose distribution information corresponding to the target sample fusion information.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the training method for the dose distribution prediction model that integrates anatomical and beam information as described in any one of claims 1 to 3, and / or the dose distribution prediction method as described in claim 4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the dose distribution prediction model that integrates anatomical and beam information as described in any one of claims 1 to 3, and / or the dose distribution prediction method as described in claim 4.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method for the dose distribution prediction model that integrates anatomical and beam information as described in any one of claims 1 to 3, and / or the dose distribution prediction method as described in claim 4.

Images