Lung cancer IMRT multi-prescription dosage prediction method and device based on beam path distance map

Abstract

The invention provides a lung cancer IMRT multi-prescription dosage prediction method and device based on a beam path distance map, and is applied to the technical field of data processing. The method is developed around lung cancer IMRT multi-prescription dosage prediction, CT images, organ outlines and other data of a plurality of patients are collected firstly, and data set division, resampling, normalization and data enhancement preprocessing are carried out; a signed distance map is generated based on the beam path, a cascade model CasU-Net-BDM containing GD-Net and RD-Net is constructed, and training is completed by taking MAE as a loss function and adopting an Adam optimizer and the like; evaluating the performance of the model through indexes such as voxel MAE, HI and CI and a visual means; and finally, inputting patient related data, outputting an accurate multi-prescription dose distribution prediction result by the trained model, and being suitable for various IMRT and SIB prescription scenes.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for predicting multiple prescription doses for lung cancer IMRT based on beam path distance maps. Background Technology

[0002] Lung cancer is one of the malignant tumors with high incidence and mortality rates worldwide. Intensity-modulated radiotherapy (IMRT) is a commonly used technique in lung cancer radiotherapy. Its core is to reduce the radiation dose to surrounding normal tissues while ensuring target area coverage through precise dose distribution design.

[0003] In current lung cancer IMRT planning, the determination of dose distribution relies on physicists manually adjusting parameters, which has problems such as low efficiency and poor adaptability to multi-prescription scenarios. Existing dose prediction technologies mostly lack effective utilization of beam path spatial characteristics, resulting in insufficient prediction accuracy and difficulty in meeting the needs of multi-prescription scenarios (including conventional IMRT and SIB simultaneous dose delivery).

[0004] Beam information directly impacts dose distribution results, and rapid dose attenuation outside the beam path is crucial for protecting normal tissues (such as lung tissue, preventing radiation pneumonitis). Current technologies largely rely solely on anatomical information such as CT images and organ contours, or only incorporate target volume (PTV) distance maps, failing to fully integrate spatial distance characteristics of the beam path. This results in models struggling to accurately capture dose drop patterns near beam boundaries, leading to significant dose prediction errors in areas outside the beam path. Furthermore, data ambiguity in multiple overlapping target areas within the SIB program also affects the learning performance of dose prediction models.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of this application, a method for predicting multi-prescription dose for IMRT in lung cancer based on beam path distance maps is provided, comprising: acquiring CT images, organ contours, beam directions, and planned dose distribution data of several lung cancer IMRT multi-prescription patients, covering various IMRT and SIB prescription schemes, wherein the organ contours include PTV and OARs of the esophagus, heart, lung, and spinal cord; dividing the data into training set, validation set, and test set according to a preset ratio, and performing resampling, CT value and dose value normalization, and data augmentation processing such as flipping, rotation, and translation on all data; generating a signed distance map based on the beam path, and using it, along with CT images, beam directions, and anatomical structure images as multi-channel inputs, to construct a hierarchical network consisting of GD-Net and RD-Net. A convolutional neural network, CascU-Net-BDM, was developed. Using MAE as the loss function, the model was trained in a Python 3.10 + PyTorch 2.0 environment with Kaiming initialization, Adam optimizer, and cosine annealing. The performance of the CascU-Net-BDM model was comprehensively evaluated using key dosimetric parameters such as voxel MAE, HI, CI, PTV, and OARs, combined with dose distribution visualization and DVH curve comparison. Based on the trained CascU-Net-BDM model, CT images, organ contours, beam directions, and signed distance maps of beam paths from lung cancer patients were input to generate multi-prescription dose distribution prediction results for lung cancer IMRT.

[0008] Another aspect of this application discloses a lung cancer IMRT multi-prescription dose prediction device based on beam path distance maps, comprising: a data acquisition module for acquiring CT images, organ contours, beam directions, and planned dose distribution data of several lung cancer IMRT multi-prescription patients, covering various IMRT and SIB prescription schemes, wherein the organ contours include PTV and OARs of the esophagus, heart, lung, and spinal cord; a data preprocessing module for dividing the data into training, validation, and test sets according to a preset ratio, and performing resampling, CT value and dose value normalization, and data augmentation processing such as flipping, rotation, and translation on all data; and a model building module for generating a signed distance map based on the beam path, and using it, along with CT images, beam directions, and anatomical structure images as multi-channel inputs to construct a cascaded rollup composed of GD-Net and RD-Net. The CascU-Net-BDM integrative neural network consists of four modules: a model training module, a model training module, and a model evaluation module. The CascU-Net-BDM model is trained using MAE as the loss function, employing Kaiming initialization, Adam optimizer, and cosine annealing strategy within a Python 3.10 + PyTorch 2.0 environment. A model evaluation module is used to comprehensively evaluate the performance of the CascU-Net-BDM model by combining key dosimetric parameters such as voxel MAE, HI, CI, PTV, and OARs with dose distribution visualization and DVH curve comparison. A dose prediction module is used to generate multi-prescription dose distribution prediction results for lung cancer IMRT based on the trained CascU-Net-BDM model, inputting CT images, organ contours, beam directions, and signed distance map data of beam paths from lung cancer patients.

[0009] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for predicting multiple prescription doses for lung cancer IMRT based on beam path distance maps by executing the executable instructions.

[0010] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described method for predicting multiple prescription doses for lung cancer IMRT based on beam path distance maps.

[0011] This application provides a method and device for predicting multiple prescription doses in IMRT for lung cancer based on beam path distance maps. The method involves collecting CT images, organ contours, and beam directions from lung cancer patients, performing a 7:1:2 dataset partitioning, resampling, normalization, and data augmentation preprocessing. A signed distance map is generated based on the beam path, and this map, along with multiple types of images, forms a multi-channel input. A CascU-Net-BDM model, cascaded from GD-Net and RD-Net, is constructed. The model is trained using weighted MAE as the loss function, combined with Kaiming initialization, Adam optimizer, and cosine annealing strategy. Evaluation is achieved through voxel MAE, HI, CI, and other indicators, as well as visualization techniques. The method outputs predicted multiple prescription dose distributions based on patient data, adapting to IMRT and SIB scenarios.

[0012] This innovative approach introduces a signed distance map of the beam path, accurately capturing dose drop patterns at beam boundaries with a voxel MAE as low as 1.19±0.41Gy, significantly outperforming traditional models in prediction accuracy. It is compatible with multiple prescriptions and SIB regimens ranging from 45-65Gy, resolving ambiguities related to target overlap through PTV cropping and merging. A single model meets the needs of all scenarios, demonstrating strong generalization capabilities. The cascaded model architecture and optimization strategies work synergistically, balancing global feature extraction with local accuracy optimization. Key dosimetric parameters meet clinical standards, significantly reducing the workload of manual adjustments for physicists. A multi-dimensional evaluation system combining quantitative indicators and visual validation ensures the clinical suitability of the prediction results, providing an efficient and reliable reference for radiotherapy planning.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 This document illustrates a flowchart of a method for predicting multiple prescription doses for lung cancer IMRT based on beam path distance maps, provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of a lung cancer IMRT multiprescription dose prediction device based on beam path distance map according to an embodiment of this application. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] The following is combined Figure 1This application describes a method for predicting multiple prescription doses for IMRT in lung cancer based on beam path distance maps, according to exemplary embodiments thereof. It should be noted that the following application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application are applicable to any suitable scenario.

[0017] In one implementation, Figure 1 A schematic flowchart of a lung cancer IMRT multiprescription dose prediction method based on beam path distance map according to an embodiment of this application is shown.

[0018] S101 acquires CT images, organ contours, beam direction, and planned dose distribution data of several lung cancer IMRT multi-prescription patients, covering a variety of IMRT and SIB prescription regimens.

[0019] In one implementation, complete radiotherapy-related data were acquired from 190 lung cancer patients with multiple IMRT prescriptions. The data were derived from cases treated within a preset time period and had been approved by the ethics committee. The data covered various standard IMRT prescription regimens and concurrent bolus radiotherapy (SIB) regimens, with patient prescription doses ranging from 45 to 65 Gy, and fractionated doses of 1.5 to 3 Gy. Each patient had 1 to 3 planning target volumes (PTVs), meeting the needs of dose prediction in multiple scenarios. Data was acquired using a Philips 16-slice large-aperture CT scanner with a slice thickness of 5 mm and a single-slice pixel matrix of 512×512, to accurately represent the patient's thoracic anatomy and provide an anatomical basis for subsequent organ contour delineation, beam path planning, and dose distribution calculation.

[0020] The organ contour data is obtained, which includes two core structures, as detailed below: PTV: Determined based on the patient's tumor location, size, and clinical treatment needs. It must cover the tumor tissue and potential infiltrative areas. Some patients have multiple overlapping PTVs (such as standard dose PTV and boosted dose PTV).

[0021] Organs at Risk (OARs): These explicitly include the esophagus, heart, left lung, right lung, both lungs, and spinal cord. All are converted into independent binary mask files, with the internal voxel labels set to 1 and the external labels set to 0, to facilitate model recognition of anatomical boundaries.

[0022] Acquire beam direction data and optimize and determine it in EclipseTPS. Use a 6MV photon beam with 5 to 9 coplanar directions and an angle range of 0° to 360°. The beam direction should avoid areas with dense critical OARs to balance tumor irradiation and normal tissue protection.

[0023] Obtain planned dose distribution data, which is the actual clinical dose distribution data and serves as the ground truth for model training. This data is used for comparison and evaluation of subsequent model prediction results to ensure that the predicted doses meet clinical treatment standards.

[0024] For SIB plans with overlapping PTVs, a PTV trimming and merging process is adopted: the voxels of the boosted PTVs within the regular dose PTVs are trimmed and discarded, and normalized and merged with 65 Gy as the standard dose. Each PTV prescription dose label is assigned by a weighted summation method to ensure the consistency and effectiveness of data in multi-PTV scenarios and to provide a unified data format for the model to learn the multi-prescription dose mapping relationship.

[0025] S102 divides the data into training, validation and test sets according to a preset ratio, and performs data enhancement processing such as resampling, normalization of CT values ​​and dose values, flipping, rotating and translating on all data.

[0026] In one implementation, the data is randomly divided in a 7:1:2 ratio. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and overfitting monitoring during training, and the test set is used for final evaluation of the model's generalization ability. Taking 190 patient data as an example, after the division, there are 133 cases in the training set, 19 cases in the validation set, and 38 cases in the test set. The division process ensures that the prescription schemes in each set are evenly distributed, avoiding data bias that could affect model performance.

[0027] All data were resampled using trilinear interpolation to unify 3D data such as CT images, organ contours, and dose distributions to a grid size of 128×128×128. During resampling, consistent voxel spacing was maintained to ensure the relative positions of anatomical structures remained unchanged, providing the model with uniform input data and reducing feature learning interference caused by size differences.

[0028] CT values ​​are clipped and normalized within the range of [-1.024, 1.5]. CT values ​​reflect differences in human tissue density. To reduce the influence of extreme values ​​and noise, CT values ​​are clipped and normalized. The CT values ​​of all voxels are restricted to the range of [-1.024, 1.5]. Voxel values ​​exceeding this range are forcibly set as boundary values ​​and then mapped to the [0, 1] interval through a linear transformation. This preserves key tissue density information and accelerates the convergence speed during model training.

[0029] Using 65 Gy as the standard dose, the planned dose distribution for all patients was normalized. A linear normalization formula was used to map the dose values ​​of different prescribed doses (45 Gy to 65 Gy) to the range of [0,1], with the formula: Normalized dose value = Original dose value / 65. This process allows the model to focus on learning the mapping relationship between the spatial characteristics of the dose distribution and anatomical structures without needing to adapt to different dose scales.

[0030] Data augmentation was performed, including random flipping of the x and z axes, random rotation around the z axis from 0° to 320°, and random translations of up to 20 pixels. Various geometric transformations were used to expand the training data and improve the model's generalization ability, as follows: Random Flipping: The data was randomly flipped along either the x or z axis with a 50% probability in each direction, simulating the impact of patient positional differences on dose distribution; Random Rotation: Random rotations around the z axis of 0°, 40°, 80°, 120°, 160°, 200°, 240°, 280°, and 320° with a step size of 40°, covering the range of common beam field angles in clinical practice; Random Translation: Random translations of up to 20 pixels were performed in the x, y, and z directions. The translation range reflects the actual scenario of slight patient positional deviations during clinical radiotherapy, enhancing the model's robustness to positional changes.

[0031] For the SIB program, PTV trimming and merging were employed. For multiple overlapping PTV structures in vivo, the added PTV voxels within the standard-dose PTV were trimmed and discarded, using 65 Gy as the standard dose. Multiple PTVs are combined and assigned a prescription dose label, where... , , , It is a cropped binary mask (with internal voxel labels of 1 and external labels of 0). The cropping logic discards the pvoxels of the pvoxels within the standard dose pvoxels to resolve the data ambiguity problem of overlapping target regions. The cropped pvoxels are weighted and summed using the ratio of each pvoxel prescription dose to the standard dose as a weighting coefficient, and finally the merged pvoxel data is obtained. This data can be directly used as model input to achieve unified characterization and prediction of multiple prescription doses of 45-65 Gy.

[0032] For multiple overlapping PTV structures in vivo, the core processing logic is to trim and discard the voxels of the increased-dose PTV within the regular-dose PTV. That is, for overlapping PTVs receiving different prescription doses, the complete structure of the high-dose prescription PTV is retained first, and the parts of the low-dose prescription PTV that overlap with the high-dose prescription PTV are removed. This avoids model learning bias caused by ambiguous voxel attribution and ensures that each voxel corresponds to only a single prescription dose label.

[0033] After cropping, each PTV (including PTV45, PTV50, PTV60, and PTV65) is in binary mask format, with the internal voxel labels uniformly set to 1 and the external voxel labels set to 0. This format clearly defines the spatial range of each PTV, facilitating rapid target region boundary identification by the model and providing clear anatomical localization basis for subsequent dose mapping relationship learning.

[0034] Using 65 Gy as the standard dose, multiple cropped target volume televisions (PTVs) are merged into unified target area data using a weighted summation formula, and prescription dose labels are simultaneously assigned. The weighting coefficient is determined by the ratio of the prescription dose of each PTV to the standard dose (65 Gy), which preserves the dose difference characteristics of different PTVs while achieving normalized integration of multi-prescription dose information. The merged PTVs data can be directly used as model input, enabling the model to learn the distribution patterns of multiple prescription doses in a single target area representation without the need for separate modeling for different prescription regimens, significantly improving the model's generalization ability in multi-scenario dose prediction.

[0035] The core objective of this processing method is to resolve the data ambiguity caused by overlapping multiple PTVs in the SIB program. By unifying the data format and dose labeling rules, the model can accurately capture the mapping relationship between different prescription doses and target spatial locations. Simultaneously, the normalization processing based on the 65 Gy standard dose forms a logical closed loop with subsequent dose value normalization (mapping all dose values ​​to the [0,1] range), reducing dose scale interference during model training and improving training efficiency and prediction accuracy.

[0036] S103 generates a signed distance map based on the beam path, and uses it, along with CT images, beam direction, and anatomical structure images as multi-channel inputs to construct a cascaded convolutional neural network CascU-Net-BDM composed of GD-Net and RD-Net.

[0037] In one implementation, a beam mask is generated based on the gantry angle and voxels on the PTV surface. The mask labels are adjusted according to the inverse square law to characterize the cumulative irradiation intensity of the voxels. Based on the gantry angle and voxel coordinates on the PTV surface in the clinical radiotherapy plan, a beam mask simulating the beam irradiation range is generated, covering the spatial area the beam may traverse. The voxel labels within the mask are adjusted using the inverse square law; the label value is inversely proportional to the square of the distance from the voxel to the beam source point. That is, voxels closer to the source point have larger label values, and those farther from the source point gradually decrease to 0. This accurately characterizes the cumulative irradiation intensity of each voxel, providing the model with preliminary spatial information on beam dose attenuation.

[0038] The distance from a voxel to a point on the beam mask surface is calculated based on the Euclidean distance transformation, using the following formula: Where S is the set of points on the object's surface, and (i,j,k) are the coordinates of the surface points. , , These represent the voxel spacing in the three directions. This formula quantifies the spatial relationship between voxels and the beam boundary, providing fundamental data for subsequent range map construction.

[0039] First, coordinate alignment is performed to ensure that the target voxel coordinates (x, y, z) and the beam mask surface point coordinates (i, j, k) are in the same grid system, based on the unified voxel spacing after resampling. , , To avoid distance calculation errors caused by inconsistent coordinate systems, a point-by-point distance calculation is performed. For each target voxel (x, y, z), all points (i, j, k) in the beam mask surface point set S are traversed, and the Euclidean distance is calculated using the formula. During the calculation, the voxel coordinate difference is converted into the actual spatial distance difference through an equivalent method, taking into account the anisotropy of the voxel spacing. Next, all the calculated distance values ​​are compared, and the minimum value is selected as the final distance D(x, y, z) from the target voxel to the beam mask surface, ensuring that the result reflects the shortest path distance from the voxel to the beam boundary.

[0040] The calculated distance value D(x,y,z) is the core foundation for constructing the signed distance map. Subsequently, by distinguishing whether the voxel is inside (∈M) or outside (∈ / M) the beam mask, the distance value is assigned a positive or negative sign (negative for inside, positive for outside, and 0 for the boundary), thus forming a signed distance map. After integrating this distance map with CT images, beam direction, and anatomical structure images into a multi-channel input, it can help the CascU-Net-BDM model accurately capture the rapid dose drop characteristics near the beam boundary, especially improving the dose prediction accuracy in the region outside the beam path, and adapting to the low-dose protection requirements for normal tissues (such as the lungs and esophagus) in multi-prescription dose prediction for lung cancer IMRT.

[0041] By distinguishing the signs of the distance values ​​inside and outside the mask, a signed distance map is constructed, and the formula is as follows: Where (x,y,z) are the target voxel coordinates to be labeled in three-dimensional space, consistent with the grid coordinate system of the beam mask and CT image; M is the beam mask region, i.e. the spatial range that the beam may illuminate, defined by the binary label of the beam mask (1 for voxels inside the mask and 0 for those outside); D(x,y,z) is the shortest distance from the target voxel to the surface of the beam mask, calculated by Euclidean distance transformation, and is a non-negative value; SMD(x,y,z) is the final value of the target voxel in the signed distance map, which distinguishes the positional relationship between the voxel and the beam mask by the sign, and the distance value quantifies the spatial distance.

[0042] The construction of the signed distance map (SMD) is a core step in quantifying beam path spatial information into features that the model can recognize. By symbolically distinguishing the positional relationship and distance between voxels and the beam mask, it provides accurate beam boundary perception basis for the CascU-Net-BDM model, which is elaborated as follows: The construction needs to be based on the generated beam mask (generated according to the gantry angle and PTV surface voxels, with labels adjusted by the inverse square law to characterize the cumulative irradiation intensity) and voxel-beam mask surface distance data (calculated through Euclidean distance transformation). The core objective is to clearly present the spatial positional relationship of each voxel relative to the beam boundary through symbol differentiation and distance quantization, so as to help the model accurately capture the rapid dose drop pattern at the beam boundary and improve the accuracy of multi-prescription dose prediction.

[0043] For each target voxel (x, y, z), determine whether it belongs to the inside ((x, y, z) ∈ M) or outside ((x, y, z) ∈ M) of the beam mask based on its binarized label. Voxels at the mask boundary are determined to belong to both inside and outside, and are subsequently assigned a value of 0. Assign a sign to the distance value D(x, y, z) based on the position determination result—voxels inside the mask are assigned a value of 0. The outer voxels are assigned the value D(x,y,z), while the boundary voxels are directly assigned the value 0, thus achieving a symbolic representation of positional relationships. After traversing all voxels and completing the symbolic assignment, a complete signed distance map is formed. The value of each voxel in the map contains both the distance information (absolute value) to the beam boundary and the position information (positive or negative sign), forming a beam path spatial feature matrix in three-dimensional space.

[0044] In the signed distance graph, negative values ​​indicate that the voxel is inside the beam mask, with larger absolute values ​​indicating a greater distance from the beam boundary; positive values ​​indicate that the voxel is outside the beam mask, with larger absolute values ​​indicating a greater distance from the beam boundary; a value of 0 corresponds to the boundary voxel of the beam mask, clearly defining the boundary line of the beam illumination range. This feature representation allows the model to quickly identify the location of the beam boundary and the relative distance between the voxel and the beam.

[0045] This signed distance map, integrated with CT images (tissue density information), beam direction information, and anatomical structure images (PTV and OARs masks) as a multi-channel input, provides the CascU-Net-BDM model with comprehensive spatial-anatomical-beam features. Especially in multi-prescription IMRT scenarios for lung cancer, the model can accurately perceive the dose attenuation pattern at the beam boundary under different prescription doses based on this map, significantly improving the low-dose prediction accuracy for normal tissues outside the beam boundary (such as the lungs and esophagus), thus adapting to the dual requirements of tumor irradiation adequacy and normal tissue protection in multi-prescription dose distributions.

[0046] The signed distance map is integrated with CT images, beam direction information, and anatomical structure images into a multi-channel input. The anatomical structure images include one PTV mask, six OAR masks, and one beam mask. The constructed signed distance map is then integrated with CT images, beam direction information, and anatomical structure images into the model's multi-channel input data. The anatomical structure images include one PTV mask, six OAR masks (esophagus, heart, left lung, right lung, both lungs, and spinal cord), and one beam mask. Each mask is a binary image (voxels inside the structure are 1, and those outside are 0). This multi-channel input integrates anatomical structure, beam information, and spatial distance features, providing the model with comprehensive feature input to learn dose distribution mapping relationships.

[0047] A cascaded convolutional neural network, CascU-Net-BDM, consisting of GD-Net and RD-Net, is constructed. GD-Net extracts global features and outputs a coarse dose distribution, while RD-Net combines the original input to optimize and output a high-precision dose distribution. The model consists of two cascaded U-Net sub-models, GlobalDoseNet (GD-Net) and RefineDoseNet (RD-Net), which are connected sequentially. The output of GD-Net and the original multi-channel input are used together as the input of RD-Net, forming a cascaded architecture of "coarse prediction-fine optimization".

[0048] The GD-Net module design is as follows: The encoder consists of 5 modules. The first module comprises two 3×3×3 regular convolutional blocks (stride 1). The following four modules all use 3×3×3 convolutions (stride 2) for downsampling, and then extract features through 3×3×3 convolutions (stride 1). Instance normalization (IN) and ReLU activation functions are applied after each convolutional block. After each downsampling, the number of channels doubles (from 16 to 256), and the spatial size is halved (from 128×128×128 to 16×16×16).

[0049] Decoder: Trilinear interpolation upsampling is used to restore the spatial size. Each decoding module is upsampled first and then 3×3×3 convolution is performed. The first three decoding modules contain two 3×3×3 convolutions (stride 1) after upsampling, and the last decoding module contains only one 3×3×3 convolution and one 1×1×1 convolution. The feature maps of the corresponding layers of the encoder are passed to the decoder through skip connections to supplement high-resolution detail information and finally output a coarse dose distribution.

[0050] The RD-Net module has a similar structure to GD-Net, but the number of channels is doubled (initially 32, eventually increased to 512) to enhance feature learning capabilities. The input includes a coarse dose distribution output by GD-Net and the original multi-channel input data. By fusing the coarse prediction results and the original features, the dose distribution is finely optimized, and the final output is a high-precision dose distribution with a single channel size of 128×128×128.

[0051] S104, using MAE as the loss function, employs Kaiming initialization, Adam optimizer, and cosine annealing strategy to train the CascU-Net-BDM model in a Python 3.10 + PyTorch 2.0 environment.

[0052] In one implementation, a weighted L1 loss function incorporating GD-Net and RD-Net is designed based on the mean absolute error between the predicted dose and the actual dose. The formula is as follows: ,in, This represents the dose value predicted by GD-Net for the i-th voxel. This represents the dose value predicted by RD-Net for the i-th voxel; Let represent the actual dose value of the i-th voxel, and N represent the total number of dose voxels. Considering the crucial role of RD-Net in high-precision dose optimization, weighting coefficients α=0.5 and β=1 are set. By giving RD-Net higher weights, the model training focuses more on high-precision dose optimization. The overall function is based on the mean absolute error (MAE). By calculating the error between the predicted values ​​of the two sub-models and the true values ​​and then weighting and summing them, the model prediction bias is quantified, guiding the model iterative optimization and ultimately adapting to the accuracy requirements of multi-prescription dose prediction in lung cancer IMRT.

[0053] Calculate the absolute error between the predicted dose and the actual dose for each voxel i using GD-Net and RD-Net respectively. and This captures the dose prediction bias of individual voxels. The sum of the absolute errors for all voxels, divided by the total number of voxels N, yields the global average loss of the GD-Net. The global average loss of RD-Net This reflects the overall dose prediction bias. By using preset weighting coefficients α and β to weight and sum the global average losses of the two sub-models, the overall loss L is obtained, achieving a balance between coarse prediction and fine optimization losses.

[0054] As a fine-tuning module, RD-Net's output of high-precision dose distribution directly determines the model's final prediction performance. Therefore, β=1 is set to make its loss account for a higher proportion of the overall loss. GD-Net only provides basic coarse predictions, so α=0.5 is set to ensure its basic role while avoiding its loss from excessively affecting the model's optimization of high-precision doses.

[0055] This function is adapted to the multi-prescription dose prediction scenario of lung cancer IMRT. By focusing on high-precision dose optimization, the model can accurately capture the distribution pattern of different prescription doses (45Gy to 65Gy), especially the dose drop characteristics near the beam boundary and organs at risk (OARs). This ensures that the prediction results meet the dose coverage requirements of the tumor target volume (PTV) while protecting normal tissues from excessive radiation, which is in line with clinical dosimetry standards.

[0056] The Kaiming initialization method was used to initialize the weights of all convolutional layers in CascU-Net-BDM. This method is compatible with the ReLU activation function and can avoid the vanishing or exploding gradient problem in the early stages of training, laying the foundation for rapid model convergence. The batch size was set to 2 to match the GPU memory capacity (two Nvidia RTX 3090s with 24GB of VRAM), balancing computational efficiency and memory usage while ensuring training stability. The maximum number of iterations was set to 80,000, with each epoch containing 68 iterations, for a total of 1176 epochs of training, ensuring the model fully learns the dose distribution mapping patterns in the data.

[0057] The optimization strategy design is based on the complex characteristics of multi-prescription dose prediction in lung cancer IMRT—it is necessary to accurately learn the mapping relationship between different prescription doses and anatomical structures and beam paths. The goal is to balance rapid exploration in the early stage of model training with stable convergence in the later stage, avoid gradient oscillation or convergence stagnation, and improve the model's prediction accuracy for dose distribution in multiple scenarios.

[0058] The Adam optimizer and cosine annealing learning rate strategy are detailed below. They integrate momentum gradient descent and adaptive learning rate adjustment mechanisms, accelerating gradient updates through momentum terms and dynamically adjusting the learning rate for different parameters. This adapts to the complex gradient changes in 3D dose prediction tasks, significantly improving training efficiency. The initial learning rate is set to 3×10. 4. This value balances the initial parameter update speed with the risk of gradient explosion, and is suitable for the feature learning requirements of the cascaded model architecture.

[0059] The cosine annealing learning rate strategy is as follows: the learning rate gradually decays according to a cosine function as the training progresses (per epoch), simulating the training rhythm of "high learning rate exploration → low learning rate fine-tuning." Initially, the parameter space is rapidly traversed, and later parameters are precisely optimized to avoid the model oscillating around the optimal solution. The learning rate decays to a minimum of 1×10-1. Training is terminated at 7 o'clock to ensure that the model converges fully and to avoid overfitting due to overtraining.

[0060] A small server equipped with two Nvidia RTX 3090 GPUs with 24GB of video memory can significantly improve training speed through multi-GPU parallel computing. The large 24GB of video memory can meet the high memory requirements of 128×128×128 3D data, avoiding training interruptions or accuracy loss caused by batch processing of data. It needs to meet the parallel computing requirements of multi-channel input data (CT images + anatomical structure images + beam directions + signed distance maps) and ensure efficient processing of batch data (batch size set to 2).

[0061] The Python 3.10 programming language, paired with the PyTorch 2.0 deep learning framework, leverages PyTorch's automatic differentiation mechanism to efficiently perform gradient calculations and parameter updates, adapting well to the training logic of cascaded convolutional neural networks and providing excellent support for GPU acceleration. The framework version is highly compatible with hardware drivers and model dependency libraries (such as those for calculating Euclidean distance and loss functions), ensuring stable operation of core processes such as beam path distance map generation and weighted L1 loss calculation.

[0062] Preprocessed multi-channel input data is batched (batch size = 2) and input into the model. Forward propagation yields a coarse dose distribution from GD-Net and a high-precision dose distribution from RD-Net. Errors are calculated using a weighted L1 loss function, and model weights are updated via backpropagation and the Adam optimizer. A cosine annealing strategy is used to dynamically adjust the learning rate, iterating repeatedly until the maximum number of iterations (80,000) is reached or the learning rate reaches its minimum. The maximum number of iterations is 80,000, with each epoch containing 68 iterations, for a total of 1176 epochs of training. A single complete training session takes approximately 14 hours, ensuring the model fully learns the core patterns of multi-prescription dose distributions while balancing training efficiency and generalization ability.

[0063] The optimization strategy and environment configuration are precisely adapted to the multi-prescription dose prediction scenario of lung cancer IMRT: the adaptive characteristics of the Adam optimizer can cope with the gradient differences of different prescription doses (45Gy to 65Gy), and the cosine annealing strategy avoids the model from overfitting to specific prescription scenarios; the large-memory GPU and efficient framework ensure the rapid processing of multi-channel, large-size 3D data, enabling the model to stably learn key clinical features such as beam boundary dose drop and OAR protection, which meets the requirements of clinical dosimetry standards for prediction accuracy.

[0064] S105 comprehensively evaluates the performance of the CascU-Net-BDM model by using key dosimetric parameters of voxel MAE, HI, CI, PTV, and OARs, combined with dose distribution visualization and DVH curve comparison.

[0065] In one implementation, based on the target of evaluating the accuracy of model dose prediction, voxel MAE, PTV homogeneity index HI, and conformity index CI are selected as core evaluation indicators to comprehensively quantify the consistency between predicted dose and manually planned clinical dose. The HI formula is as follows: ,in, , These are doses covering 2% and 98% of the PTV volume, respectively. For the prescribed dose, the closer the HI value is to 0, the more uniform the dose distribution within the PTV. The CI formula is: ,in, For the PTV volume covered by the prescribed dose, For PTV size, The CI value is the volume covered by the prescription dose. The closer the CI value is to 1, the better the conformity of the prescription dose to the PTV. Where N is the number of voxels in the PTV or OARs region. The predicted dose for the i-th voxel. For the manually planned dose of the i-th voxel, the smaller the value, the more accurate the overall dose prediction.

[0066] Voxel MAE focuses on "overall error control," HI focuses on "target area homogeneity," and CI focuses on "dose-target conformity." These three indicators approach the problem from different dimensions, avoiding the limitations of a single indicator. A comprehensive assessment using all three indicators is necessary—if the voxel MAE is low, HI is close to 0, and CI is close to 1, it indicates that the model not only has a small overall prediction error but also ensures uniform PTV dose coverage and precise conformity, meeting the core requirements of clinical radiotherapy for target area treatment efficacy.

[0067] Regardless of whether the patient's prescribed dose is 45Gy, 50Gy, 60Gy, or 65Gy, all three indicators can be uniformly quantified and assessed without adjusting the assessment criteria for different prescriptions, making it suitable for multiple scenarios including SIB and conventional IMRT. HI and CI directly correspond to the core clinical requirements of radiotherapy planning (target volume dose homogeneity and conformity), while voxel MAE reflects the model's dose prediction accuracy for OARs. The combination of these three indicators can comprehensively verify the model's performance under the dual objectives of "tumor irradiation adequacy" and "normal tissue protection," ensuring that the prediction results have clinical application value.

[0068] Focusing on the core indicators of voxel MAE, HI, and CI, this study supplements key dosimetric parameters of PTV and OARs to form a comprehensive multi-dimensional evaluation system covering both the target area and normal tissues, further refining the model performance evaluation logic. The details are as follows: The design of supplementary indicators is based on the clinical integrity requirements of IMRT multiprescription dose prediction for lung cancer. The core indicators focus on the overall error and basic characteristics of the target area, while the supplementary indicators focus on the adequacy of target area dose coverage and the risk of radiation to normal tissues. The core objective is to refine the clinical suitability of the assessment model prediction from the two dimensions of "target area treatment effect" and "normal tissue protection" to ensure that the assessment is comprehensive and more in line with the actual clinical needs of radiotherapy.

[0069] The supplementary dosimetric parameters for PTV are as follows: D99% (dose covering 99% of PTV volume), D98% (dose covering 98% of PTV volume), D95% (dose covering 95% of PTV volume), Dmax (maximum dose within PTV), and Dmean (average dose within PTV). D99%, D98%, and D95% are used to determine the adequacy of target area dose coverage, ensuring that the vast majority of target area voxels receive sufficient therapeutic dose. Dmax is used to mitigate the risk of damage to normal tissues caused by excessively high local doses in the target area. Dmean is used to assess whether the overall dose level of the target area meets the prescription requirements, jointly verifying the rationality of the target area dose distribution.

[0070] The supplementary dosimetric parameters for OARs are as follows: Dmax (maximum dose in organs at risk), Dmean (mean dose in organs at risk), V40Gy (volume percentage of those receiving 40Gy or more), V30Gy (volume percentage of those receiving 30Gy or more), V20Gy (volume percentage of those receiving 20Gy or more), and V5Gy (volume percentage of those receiving 5Gy or more).

[0071] Dmax and Dmean quantify the overall radiation dose level of organs at risk, avoiding serious complications caused by excessive dose to critical organs; volumetric dose parameters such as V40Gy, V30Gy, V20Gy, and V5Gy assess the radiation range of organs for different dose thresholds, especially suitable for the low-dose protection needs of OARs such as lung, esophagus, and spinal cord in lung cancer radiotherapy, and comprehensively assess the radiation risk to normal tissues.

[0072] The target area synergy is as follows: the core indicators HI (homogeneity) and CI (conformity) are combined with supplementary indicators D95%, Dmean, and Dmax to assess both the "distribution pattern" of the dose in the target area and verify the "coverage level" of the dose, ensuring that the target area is both homogeneous and conformal, with sufficient dose and no local overdose. The overall assessment synergy is as follows: the core indicator Voxel MAE reflects global error, while supplementary indicators focus on the detailed dose characteristics of PTV and OARs, forming a three-layer assessment logic of "global error + target area details + normal tissue risk," avoiding the limitation of a single indicator failing to cover key clinical dimensions.

[0073] Regardless of the prescribed dose of 45Gy, 50Gy, 60Gy, or 65Gy, the supplementary indicators can be compared with the corresponding prescribed dose to accurately assess target coverage and OAR protection under different dosage regimens without adjusting the assessment criteria. The supplementary indicators directly correspond to the core review points of radiotherapy clinical plans (e.g., PTV D95% must be ≥ 95% of the prescribed dose, bilateral lung V20Gy ≤ 30%), and the assessment results can be directly used to determine whether the model's predictions meet clinical treatment guidelines, providing crucial quantitative evidence for the clinical translation of the model.

[0074] By combining quantitative indicators with visual verification, a dual evaluation logic of "numerical quantification + visual intuition" is formed to improve the credibility of the evaluation, as detailed below: Dose distribution visualization comparison: The three-dimensional dose distribution predicted by the model is compared with the clinically manually planned dose distribution in the transverse, coronal and sagittal planes to intuitively observe the dose distribution pattern, the dose drop trend at the beam boundary and the dose matching of key areas (such as the edge of PTV and the area adjacent to OARs).

[0075] Dose-volume histogram (DVH) curve consistency analysis: The predicted dose-volume DVH curves of PTV and each OAR are plotted with the manually planned DVH curves. By comparing the curve shape and the degree of overlap of key dose points (such as the volume or dose value corresponding to V20Gy and Dmean), the consistency of the dose-volume relationship is quantitatively evaluated. The higher the degree of overlap of the curves, the closer the dose-volume characteristics predicted by the model are to clinical reality.

[0076] Statistical processing was performed on all evaluation results. The Shapiro-Wilk test was used to determine the normality of the data. For normally distributed data, a two-tailed paired t-test was used, and for non-normally distributed data, a Wilcoxon signed-rank test was used to comprehensively assess the dose prediction performance of the CascU-Net-BDM model. Statistical analysis was conducted on all evaluation indicators to ensure the scientific rigor and soundness of the evaluation conclusions. First, the Shapiro-Wilk test was used to determine the normality of each evaluation data point; the closer the test statistic W is to 1, the better the normality of the data. If the data conformed to a normal distribution, a two-tailed paired t-test was used to compare the mean difference between the predicted results and the manual planning results; if the data did not conform to a normal distribution, a Wilcoxon signed-rank test was used for non-parametric comparison.

[0077] All tests used a significance level of α=0.05 as the judgment criterion. When the p-value was ≥0.05, it indicated that there was no statistically significant difference between the predicted results and the manual planning results, and the model's dose prediction performance met the clinical suitability requirements. When the p-value was <0.05, it was necessary to consider the actual difference in the indicator and comprehensively judge whether the model's deviation on that indicator affected clinical application. Through the above statistical analysis, the stability and clinical applicability of the dose prediction performance of the CascU-Net-BDM model were comprehensively determined.

[0078] S106, based on the trained CascU-Net-BDM model, takes CT images, organ contours, beam directions and signed distance map data of beam paths from lung cancer patients as input, and generates prediction results for the multi-prescription dose distribution of lung cancer IMRT.

[0079] In one implementation, a trained CascU-Net-BDM model is obtained based on the lung cancer IMRT multiprescription dose prediction target, and the data type and specifications of the input data are clearly defined. The input data includes CT images of the target lung cancer patient, organ contours, beam directions, and signed distance maps of the beam path. Specifically, the CT images must be chest images acquired by a Philips 16-slice large-aperture CT scanner, with a slice thickness of 5mm and a single-slice pixel matrix of 512×512; the organ contours include PTV and OARs of the esophagus, heart, left lung, right lung, both lungs, and spinal cord, all in binary mask format (internal voxel label 1, external 0); the beam direction has 5 to 9 coplanar directions (angles from 0° to 360°); the signed distance map needs to be generated through Euclidean distance transformation of the beam mask, clearly distinguishing between voxels inside and outside the mask.

[0080] The input data undergoes validity validation to verify CT image resolution, organ contour integrity, beam direction rationality, and signed distance map standardization, generating data validation results. Multi-dimensional data validation rules are established to comprehensively validate the input data, ensuring it meets model input requirements, as detailed below: For CT image resolution verification, it is verified whether the image slice thickness and pixel matrix meet the preset specifications and whether there are any resolution abnormalities, artifacts or missing data.

[0081] For organ contour integrity verification, check whether the PTV and each OAR contour are completely drawn, without boundary breaks or missing areas, and ensure that the contours correspond accurately to the anatomical structures in the CT images.

[0082] For the rationality verification of beam direction, it was confirmed that the number of beam directions is within the range of 5 to 9, the angle distribution avoids the dense area of ​​OARs, and there are no repeated or unreasonable field directions.

[0083] For the standardization verification of signed distance maps, this section verifies whether the distance map correctly distinguishes between the inside and outside of the mask (negative inside, positive outside, 0 at the boundary) and whether the distance calculation conforms to the requirements of the Euclidean distance transformation algorithm. After verification, data verification results are generated, clearly identifying qualified and unqualified items and their reasons.

[0084] The prediction process is designed based on a cascaded architecture logic of "coarse prediction - fine optimization". The core objective is to first obtain the basic dose distribution through global feature extraction, and then refine and optimize it in combination with the original input to ensure that the prediction results not only conform to the global anatomy-beam feature rules, but also accurately capture local dose details. At the same time, compliant results are screened through clinical validation rules to adapt to the treatment needs of multi-prescription scenarios of lung cancer IMRT.

[0085] The validated CT images, anatomical structure images (1 PTV mask + 6 OARs masks + 1 beam mask), beam direction, and signed distance maps are integrated in a multi-channel format to form an input data format that the model can recognize, integrating three core features: anatomy, beam direction, and spatial distance. The input data must first undergo validity validation to ensure that the CT image resolution is compliant, organ outlines are complete, beam direction is reasonable, and signed distance maps are standardized, avoiding invalid data from affecting prediction accuracy.

[0086] The GD-Net global feature extraction and coarse prediction operations are as follows: Through progressive downsampling using five modules, the number of channels increases from 16 to 256, and the spatial size decreases from 128×128×128 to 16×16×16. Global anatomical-beam correlation features are extracted using 3×3×3 convolution, instance normalization, and ReLU activation. The decoder uses trilinear interpolation upsampling to restore the spatial size and combines skip connections to supplement high-resolution details, ultimately outputting a coarse dose distribution of 128×128×128 size, laying the foundation for the global dose distribution framework.

[0087] The RD-Net refinement and high-precision output are as follows: It receives the coarse dose distribution output from GD-Net and the original multi-channel input data, achieving deep fusion of the coarse prediction results and the original features. The number of channels is twice that of GD-Net (increased from 32 to 512), enhancing local feature learning capabilities. Dose distribution details are optimized through refined convolution operations, ultimately outputting a single-channel, voxel-level high-precision dose distribution (128×128×128). Validation rules are set based on clinical dosimetry standards, clearly defining PTV dose coverage requirements and OAR dose limitation thresholds to ensure that the prediction results comply with clinical radiotherapy treatment guidelines. The PTV dose coverage requirements are as follows: D95% ≥ 95% of the prescription dose, ensuring that more than 95% of the target area voxels receive sufficient therapeutic dose; Dmean is within ±5% of the prescription dose, ensuring that the overall dose level of the target area meets the standard, and avoiding insufficient or excessive dose.

[0088] The dose limits for OARs are as follows: Lungs: V20Gy ≤ 30%, Dmean ≤ 15Gy, to reduce the risk of radiation pneumonitis. Heart: V30Gy ≤ 40%, Dmean ≤ 25Gy, to protect cardiac function from excessive radiation. Esophagus: Dmax ≤ 60Gy, Dmean ≤ 35Gy, to reduce esophageal injury complications. Spinal cord: Dmax ≤ 45Gy, strictly limiting the maximum dose to highly sensitive organs to avoid severe radiation damage.

[0089] GD-Net's coarse prediction provides a global framework for RD-Net, while RD-Net's fine optimization focuses on local detail correction. Together, they achieve "globally accurate + locally precise" dose distribution prediction, adapting to the distribution differences of multiple prescription doses from 45Gy to 65Gy.

[0090] The model output results are validated for compliance using clinical validation rules. Results that meet the PTV dose coverage and OARs dose limits are selected, and data that exceeds the limits are removed. This ensures that the prediction results are not only numerically accurate, but also feasible for direct application to clinical radiotherapy planning.

[0091] Regardless of whether the patient's prescription dose is a single dose (e.g., 50Gy, 60Gy) or a mixed dose (e.g., 45-60Gy, 50-65Gy), the process can generate the corresponding dose distribution through a unified cascaded optimization logic. The validation rules can be dynamically adapted to the threshold by the prescription dose (e.g., the D95% threshold is adjusted synchronously with the prescription dose), without needing to modify the process for different prescriptions.

[0092] The prediction process balances the comprehensiveness of feature extraction with the precision of optimization. The validation rules directly correspond to the core review standards of clinical radiotherapy plans, enabling prediction results to be quickly integrated with clinical decisions, reducing the workload of physicists in manual adjustments, and improving the efficiency of multi-prescription radiotherapy plan design.

[0093] The dose distribution results output by the model were validated for compliance, verifying whether the key indicators of PTV (D95% and Dmean) and OARs (V20Gy and Dmax) met clinical requirements. Based on the established validation rules, the dose distribution results output by the model were validated for compliance, specifically for the key indicators of PTV: calculating and validating whether D95% and Dmean met the dose coverage requirements, ensuring dose uniformity (HI close to 0) and conformity (CI close to 1).

[0094] For OARs key indicator verification: Calculate and verify whether V20Gy (both lungs), Dmax (esophagus, spinal cord), and other dosimetric parameters (V40Gy, V30Gy, Dmean, etc.) meet the limit thresholds. If any indicator exceeds the limit, mark it as a compliance anomaly, specifying the abnormal indicator and the extent of the exceedance; if all indicators meet the requirements, it is deemed compliant.

[0095] Integrating multi-dimensional verification results and generating structured final prediction output, the core is to ensure that the output results are highly accurate, clinically compliant, and interpretable. Specifically, the logic for generating the final result is based on a closed-loop process of "data verification + model prediction + compliance verification". The core objective is to integrate key information in the prediction process and output structured results that include both accurate dosage data and clinical suitability assessment, so as to meet the decision-making needs of physicists and clinicians in the multi-prescription scenario of lung cancer IMRT.

[0096] The voxel-level three-dimensional dose distribution is shown below. The voxel-level dose data, with a grid size of 128×128×128, is consistent with the model input data specifications, accurately presenting the dose distribution in the three-dimensional space of the patient's chest, covering the spatial distribution characteristics of multiple prescription doses from 45Gy to 65Gy. The high-precision dose distribution generated after coarse prediction by GD-Net and fine optimization by RD-Net has passed the clinical validation rules screening, ensuring that beam boundary dose drop, PTV dose coverage, and OAR dose limitations meet the requirements.

[0097] The key dosimetric parameter set includes PTV-related parameters and OARs-related parameters. PTV-related parameters include D99%, D98%, D95%, Dmax, Dmean, CI, and HI, comprehensively quantifying the adequacy, homogeneity, and conformity of target area dose coverage, directly corresponding to core review indicators for clinical radiotherapy planning. OARs-related parameters include Dmax, Dmean, V40Gy, V30Gy, V20Gy, and V5Gy, systematically assessing the radiation risk to normal tissues such as the esophagus, heart, lungs, and spinal cord, providing quantitative evidence for the effectiveness of normal tissue protection.

[0098] The clinical suitability assessment report should clearly indicate whether the predicted results meet clinical dosimetric standards. If any indicators exceed the limits (e.g., bilateral lung V20Gy > 30%), the abnormal indicators and the extent of the exceedance should be listed in detail. The report should summarize the validity verification results of the input data, explaining the compliance of CT images, organ contours, beam direction, and signed distance maps, providing support for the reliability of the results. For results with non-compliance, targeted adjustment suggestions should be provided (e.g., optimizing beam direction, adjusting dose constraints, etc.) to assist the physicist in revising subsequent plans.

[0099] The output format is uniformly clinically interpretable and compatible with radiotherapy planning systems (such as Eclipse TPS), facilitating direct access and comparative analysis by physicists and reducing data format conversion costs. Regardless of whether the patient's prescription is a single dose (e.g., 50Gy, 60Gy) or a mixed dose (e.g., 45-60Gy, 50-65Gy), the output results correspond to the key parameters and distribution characteristics of the prescription dose, eliminating the need to adjust the output structure for different prescriptions.

[0100] Key dosimetric parameters and clinical suitability assessments are directly aligned with radiotherapy plan review standards, allowing physicists to quickly determine whether predicted results meet treatment needs without recalculating core indicators, thus improving plan design efficiency. It is compatible with various prescription scenarios, including conventional IMRT and SIB concurrent bolus radiotherapy, and a single model can output prediction results for all scenarios, avoiding the need to generate separate reports for different prescription regimens and reducing the complexity of clinical applications.

[0101] In one implementation, such as Figure 2 As shown, this application also provides a lung cancer IMRT multi-prescription dose prediction device based on beam path distance map, comprising: The data acquisition module 201 is used to acquire CT images, organ contours, beam directions and planned dose distribution data of several lung cancer IMRT multi-prescription patients, covering a variety of IMRT and SIB prescription schemes, including PTV and OARs of esophagus, heart, lung and spinal cord. The data preprocessing module 202 is used to divide the data into training set, validation set and test set according to a preset ratio, and to perform resampling, normalization of CT value and dose value, and data enhancement processing such as flipping, rotation and translation on all data. Model building module 203 is used to generate a signed distance map based on the beam path, and use it, along with CT images, beam direction and anatomical structure images as multi-channel inputs to build a cascaded convolutional neural network CascU-Net-BDM composed of GD-Net and RD-Net; Model training module 204 is used to train the CascU-Net-BDM model in a Python 3.10 + PyTorch 2.0 environment, using MAE as the loss function and employing Kaiming initialization, Adam optimizer, and cosine annealing strategy. The model evaluation module 205 is used to comprehensively evaluate the performance of the CascU-Net-BDM model by combining key dosimetric parameters of voxel MAE, HI, CI and PTV, OARs with dose distribution visualization and DVH curve comparison. The dose prediction module 206 is used to generate multi-prescription dose distribution prediction results for lung cancer IMRT based on the trained CascU-Net-BDM model, inputting CT images, organ contours, beam directions and signed distance map data of lung cancer patients.

[0102] The computer-readable storage medium provided in the above embodiments of this application and the lung cancer IMRT multiprescription dose prediction method based on beam path distance map provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0103] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the lung cancer IMRT multi-prescription dose prediction method, system, electronic device, and readable storage medium based on beam path distance maps are basically similar to the embodiments of the lung cancer IMRT multi-prescription dose prediction method based on beam path distance maps described above, and therefore are described relatively simply. Relevant parts can be referred to in the descriptions of the embodiments of the lung cancer IMRT multi-prescription dose prediction method based on beam path distance maps described above.

Claims

1. A method for predicting multi-prescription dose in IMRT for lung cancer based on beam path distance map, characterized in that, include: We acquired CT images, organ contours, beam directions, and planned dose distribution data for several lung cancer patients with multiple IMRT prescriptions, covering various IMRT and SIB prescription protocols. The organ contours included PTV and OARs for the esophagus, heart, lungs, and spinal cord. The data is divided into training set, validation set and test set according to a preset ratio. All data are resampled, CT value and dose value are normalized, and data enhancement processing such as flipping, rotation and translation is performed. A signed distance map is generated based on the beam path. This map, along with CT images, beam direction, and anatomical structure images, is used as multi-channel inputs to construct a cascaded convolutional neural network, CascU-Net-BDM, composed of GD-Net and RD-Net. Using MAE as the loss function, and employing Kaiming initialization, Adam optimizer, and cosine annealing strategy, the CascU-Net-BDM model was trained in a Python 3.10 + PyTorch 2.0 environment. The performance of the CascU-Net-BDM model was comprehensively evaluated by using key dosimetric parameters of voxel MAE, HI, CI, PTV, and OARs, combined with dose distribution visualization and DVH curve comparison. Based on the trained CascU-Net-BDM model, inputting CT images, organ contours, beam directions, and signed distance map data of beam paths from lung cancer patients, the system generates prediction results for the multi-prescription dose distribution of lung cancer IMRT.

2. The method as described in claim 1, characterized in that, The data is divided into training, validation, and test sets according to a preset ratio. All data undergoes resampling, CT and dose value normalization, and data enhancement processing including flipping, rotation, and translation. The data is randomly divided into training set, validation set and test set according to the preset ratio of 7:1:

2. All data were resampled, and a uniform grid size of 128×128×128 was applied. The CT values ​​were clipped and normalized, with a normalization range of [-1.024, 1.5]. Based on the standard dose of 65 Gy, the dose values ​​were normalized, and the normalization range was [0,1]. The data is augmented by random flipping of the x and z axes, random rotation around the z axis from 0° to 320°, and random translation of up to 20 pixels. For the SIB program, PTV trimming and merging were employed. For multiple overlapping PTV structures in vivo, the added PTV voxels within the standard-dose PTV were trimmed and discarded, using 65 Gy as the standard dose. Multiple PTVs are combined and assigned a prescription dose label, where... , , and The structural endogenous voxel labels are all 1.

3. The method as described in claim 1, characterized in that, A signed distance map is generated based on the beam path. This map, along with CT images, beam direction, and anatomical structure images, is used as multi-channel input to construct a cascaded convolutional neural network, CascU-Net-BDM, consisting of GD-Net and RD-Net. Based on the rack angle and the voxels on the PTV surface, a beam mask is generated, and the mask label is adjusted according to the inverse square law to characterize the cumulative intensity of voxel irradiation. The distance from a voxel to a point on the beam mask surface is calculated based on the Euclidean distance transformation, using the following formula: Where S is the set of points on the object's surface, and (i,j,k) are the coordinates of the surface points. , , These represent the voxel spacing in the three directions, respectively. By distinguishing the signs of the distance values ​​inside and outside the mask, a signed distance map is constructed, and the formula is as follows: ； The signed distance map is integrated with CT images, beam direction, and anatomical structure images into a multi-channel input. The anatomical structure images include one PTV mask, six OARs masks, and one beam mask. A cascaded convolutional neural network, CascU-Net-BDM, consisting of GD-Net and RD-Net, is constructed. GD-Net extracts global features and outputs a coarse dose distribution, while RD-Net combines the original input to optimize and output a high-precision dose distribution.

4. The method as described in claim 1, characterized in that, Using MAE as the loss function, and employing Kaiming initialization, Adam optimizer, and cosine annealing strategy, the CascU-Net-BDM model was trained in a Python 3.10 + PyTorch 2.0 environment, including: Based on the mean absolute error between the predicted dose and the actual dose, a weighted L1 loss function incorporating GD-Net and RD-Net is designed, as shown in the formula: ,in, This represents the dose value predicted by GD-Net for the i-th voxel. This represents the dose value predicted by RD-Net for the i-th voxel; represents the actual dose value of the i-th voxel, and N represents the total number of dose voxels; The Kaiming initialization method is used to initialize the model weights, with a batch size of 2 and a maximum number of iterations of 80,000. The Adam optimizer was selected to accelerate model convergence, and the initial learning rate was set to... The cosine annealing strategy is used to gradually decay the learning rate in each epoch. ; The CascU-Net-BDM model was trained using a Python 3.10 + PyTorch 2.0 environment and on an Nvidia RTX 3090 GPU server equipped with two 24GB VRAM chips.

5. The method as described in claim 1, characterized in that, The performance of the CascU-Net-BDM model was comprehensively evaluated using key dosimetric parameters of voxel MAE, HI, CI, PTV, and OARs, combined with dose distribution visualization and DVH curve comparison. This evaluation included: Based on the objective of evaluating the accuracy of model dose prediction, voxel MAE, PTV homogeneity index HI, and conformity index CI are selected as core evaluation indicators. The formula for HI is: ,in, , These are doses covering 2% and 98% of the PTV volume, respectively. This is the prescription dosage; the CI formula is... ,in, For the PTV volume covered by the prescribed dose, For PTV size, For the volume covered by the prescribed dose; Where N is the number of voxels in the PTV or OARs region. These are the predicted dose and manually planned dose for voxel i, respectively; Supplement the dosimetric parameters of PTV (D99%, D98%, D95%, Dmax, Dmean) and OARs (Dmax, Dmean, V40Gy, V30Gy, V20Gy, V5Gy) to construct a multi-dimensional evaluation system. By combining dose distribution visualization comparison and dose-volume histogram curve consistency analysis, a comprehensive evaluation logic of quantitative indicators and visualization verification is formed. Statistical analysis was performed on the evaluation results. The normality of the data was determined by the Shapiro-Wilk test. For normally distributed data, a two-tailed paired t-test was used, and for non-normally distributed data, the Wilcoxon signed-rank test was used to comprehensively evaluate the dose prediction performance of the CascU-Net-BDM model.

6. The method as described in claim 5, characterized in that, Based on the trained CascU-Net-BDM model, inputting CT images, organ contours, beam directions, and signed distance maps of beam paths from lung cancer patients, the model generates multi-prescription dose distribution prediction results for lung cancer IMRT, including: Based on the target of IMRT multiprescription dose prediction for lung cancer, the trained CascU-Net-BDM model is obtained, and the CT images, organ contours, beam directions and signed distance map data of the target lung cancer patient are input. The input data is validated to confirm the resolution of CT images, the integrity of organ contours, the rationality of beam direction, and the standardization of signed distance maps, and the data validation results are generated. The validated data is input into the CascU-Net-BDM model. Global features are extracted by GD-Net to output a coarse dose distribution. RD-Net combines the original input to optimize and generate a high-precision dose distribution. Based on clinical dosimetry standards, result verification rules were established to clarify PTV dose coverage requirements and OARs dose limitation thresholds. The model output dose distribution results were verified for compliance, and the key indicators of PTV (D95%) and Dmean, as well as OARs (V20Gy) and Dmax, were verified to meet clinical requirements. By integrating data validation results, model-predicted dose distribution, and compliance validation conclusions, a final result for predicting the multi-prescription dose distribution of lung cancer IMRT is generated, which includes voxel-level three-dimensional dose distribution, key dosimetric parameters, and clinical suitability assessment.

7. A lung cancer IMRT multi-prescription dose prediction device based on beam path distance map, characterized in that, The system includes: The data acquisition module is used to acquire CT images, organ contours, beam directions and planned dose distribution data of several lung cancer IMRT multi-prescription patients, covering a variety of IMRT and SIB prescription schemes, including organ contours including PTV and OARs of esophagus, heart, lung and spinal cord; The data preprocessing module is used to divide the data into training set, validation set and test set according to a preset ratio, and to perform data enhancement processing such as resampling, normalization of CT value and dose value, flipping, rotation and translation on all data. The model building module is used to generate a signed distance map based on the beam path, and use it, along with CT images, beam direction and anatomical structure images as multi-channel inputs, to build a cascaded convolutional neural network CascU-Net-BDM composed of GD-Net and RD-Net. The model training module is used to train the CascU-Net-BDM model in a Python 3.10 + PyTorch 2.0 environment, using MAE as the loss function and employing Kaiming initialization, Adam optimizer, and cosine annealing strategy. The model evaluation module is used to comprehensively evaluate the performance of the CascU-Net-BDM model by combining key dosimetric parameters of voxel MAE, HI, CI, PTV, and OARs with dose distribution visualization and DVH curve comparison. The dose prediction module is used to generate multi-prescription dose distribution prediction results for lung cancer IMRT based on the trained CascU-Net-BDM model, inputting CT images, organ contours, beam directions and signed distance map data of the beam path of lung cancer patients.

8. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the lung cancer IMRT multiprescription dose prediction method based on beam path distance map according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the lung cancer IMRT multiprescription dose prediction method based on beam path distance map as described in any one of claims 1 to 6.

Images