Method, apparatus and device for predicting three-dimensional radiotherapy dose distribution

Abstract

The application provides a three-dimensional radiotherapy dose distribution prediction method, device and equipment. The target area and the organ at risk drawn in the clinic are converted into an anatomical structure hyperedge, so that multiple feature markers belonging to the same clinical structure can exchange group-level information, thereby explicitly describing the high-order relationship inside the same clinical structure. Meanwhile, the adaptive feature relationship, the local space hypergraph and the anatomical structure hypergraph can reserve the patient-specific feature similarity relationship, take into account the three-dimensional space continuity and the clinical structure semantics, perform relationship propagation in two deep semantic scales, the higher resolution scale reserves more spatial details, the lower resolution scale provides more context information in a larger range, so as to take into account the dose gradient boundary and the global dose field structure, and finally realize accurate prediction of the three-dimensional radiotherapy dose distribution.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and equipment for predicting three-dimensional radiotherapy dose distribution. Background Technology

[0002] The goal of radiotherapy planning is to ensure adequate prescription dose coverage of the planned target area while minimizing unnecessary radiation to surrounding organs at risk. Current clinical radiotherapy planning typically requires medical physicists to repeatedly adjust and optimize targets, constraints, and their weights based on the patient's anatomy and clinical prescription. After obtaining the dose distribution, they check target coverage and organ-at-risk doses, and then modify planning parameters based on the results. This process is time-consuming, relies heavily on the planner's experience, and can lead to discrepancies in results between different planners.

[0003] Patient-specific 3D dose prediction can generate a reference dose distribution based on the patient's computed tomography (CT) images and clinically delineated structures before treatment planning optimization, thus providing dose guidance for subsequent planning optimization. Existing automated dose prediction methods have gradually evolved from knowledge-based planning methods based on historical cases and artificial geometric features to deep learning-based anatomical structure-to-3D dose distribution mapping methods. Techniques such as 3D U-shaped networks, generative adversarial networks, diffusion generative models, geometrically perceptual input, and clinical target loss have all been used to learn the correspondence between CT images and target areas, organ-at-risk masks, and voxel-level dose distributions.

[0004] However, existing deep learning dose prediction methods typically use the target area and the mask of organs at risk as ordinary input channels, and indirectly learn the relationship between anatomical structures through local convolution, attention mechanisms or generation processes, without explicitly representing the higher-order relationships between multiple voxels or feature markers belonging to the same clinical structure.

[0005] Clinical dose distribution is not only influenced by local anatomical morphology, but also needs to reflect the overall balance between target area prescription coverage and organ at risk protection. Steep dose gradients are typically required in the regions adjacent to the target area and organs at risk. Without explicit high-order structural relationship modeling, features belonging to the same target area or the same organ at risk may be inconsistently updated during network refinement, and the clinically important dose transition near the target area-organ at risk boundary may be over-smoothed.

[0006] Therefore, it is necessary to propose a method that can jointly utilize patient-specific feature relationships, three-dimensional local spatial relationships, and clinical anatomical structure relationships to predict three-dimensional dose distribution from coarse to fine at different semantic scales. This would address the problems of insufficient modeling of high-order structural relationships between the target area and organs at risk, inconsistent feature updates within the same clinical structure, and low dose prediction accuracy caused by the smoothing of dose gradients in the target area-organ at risk neighborhood. Summary of the Invention

[0007] This application proposes a three-dimensional radiotherapy dose distribution prediction method, device, and equipment, which can solve one of the problems existing in the background art.

[0008] To achieve the above objectives, this application adopts the following technical solution:

[0009] Firstly, a three-dimensional radiotherapy dose distribution prediction method is provided, including:

[0010] Obtain patient prediction data, which includes: computed tomography (CT) volume data, target area mask, and organ at risk mask;

[0011] The prediction data is input into a three-dimensional dose prediction network for processing to obtain the three-dimensional radiotherapy dose distribution prediction results for the patient.

[0012] The three-dimensional dose prediction network includes:

[0013] The first-stage coarse dose field estimation network is used to generate coarse dose predictions and first-stage decoding features using the prediction data as input; the first-stage coarse dose field estimation network adopts a first three-dimensional encoder-decoder structure;

[0014] The second-stage structure-aware refinement network is used to generate the final refined dose prediction by taking the decoded features from the first stage and the spliced ​​features of the prediction data as input. The second-stage structure-aware refinement network adopts a second 3D encoder-decoder structure. On the encoding side, the second 3D encoder-decoder structure processes several deep semantic scale feature labels as follows: constructing adaptive feature proximity relationships based on the similarity between the feature labels; constructing a local spatial hypergraph based on the 3D local adjacent positions of the feature labels; constructing a target area / endangered organ anatomical structure hypergraph of the feature labels based on the target area / endangered organ mask; fusing the adaptive feature proximity relationships, the local spatial hypergraph, and the target area / endangered organ anatomical structure hypergraph at the propagation matrix level to obtain a fused propagation matrix; and updating the feature labels using the fused propagation matrix through a gated residual method.

[0015] In one possible design of the first aspect, the first three-dimensional encoder-decoder structure adopts a three-dimensional U-shaped network encoder-decoder structure, including: four successive downsampling steps, a bottleneck layer, corresponding successive upsampling steps, and skip connections between the encoder and the decoder. Each encoder step includes two sets of three-dimensional convolutions, instance normalization, and ReLU activation operations. The decoder performs upsampling using trilinear interpolation and extracts features through three-dimensional convolution. The upsampled features are concatenated with the skip connection features of the corresponding encoder and then input into the subsequent convolutional block.

[0016] In one possible design of the first aspect, the second 3D encoder-decoder structure adopts a 3D U-shaped network encoder-decoder structure, including: four successive downsampling steps, a bottleneck layer, corresponding successive upsampling steps, and skip connections between the encoder and the decoder. Each encoder step contains two sets of 3D convolutions, instance normalization, and ReLU activation operations. The decoder performs upsampling using trilinear interpolation and extracts features through 3D convolution. The upsampled features are concatenated with the skip connection features of the corresponding encoder and then input into subsequent convolutional blocks. The second 3D encoder-decoder structure outputs 3D feature maps at different deep semantic scales at the fourth encoder and the bottleneck layer, respectively. The 3D feature maps serve as the input to the relation enhancement and refinement module of the second 3D encoder-decoder structure. The relation enhancement and refinement module is used to flatten the 3D feature maps into a feature label matrix and perform subsequent processing to obtain the updated feature labels.

[0017] In one possible design of the first aspect, the relationship enhancement and refinement module is specifically used to: for the current feature label, calculate its feature distances with other feature labels, and through... Search for the nearest one Each feature label is used to form an adaptive relationship with the current feature label and its selected nearest neighbors, and the first correlation matrix of the adaptive feature proximity relationship is filled with weights calculated from the feature distance.

[0018] In one possible design approach of the first aspect, the relationship enhancement and refinement module is specifically used for: at scale On the three-dimensional feature marker volume, local windows are slid with a step size. Each effective window constitutes a hyperedge and simultaneously connects all feature markers within the window, thus obtaining the second correlation matrix of the local spatial hypergraph.

[0019] In one possible design approach of the first aspect, the relation enhancement and refinement module is specifically used to: binarize the input target / organ-at-risk mask, and then downsample the target / organ-at-risk mask to a specific scale using three-dimensional adaptive max pooling. If a pooling unit overlaps with the target area / endangered organ, the corresponding feature marker is considered to belong to the target area / endangered organ, and the downsampled target area / endangered organ mask is rearranged into the third correlation matrix of the anatomical structure hypermap.

[0020] In one possible design approach of the first aspect, the relationship enhancement and refinement module is specifically used to: convert the first correlation matrix, the second correlation matrix, and the third correlation matrix into a first propagation matrix, a second propagation matrix, and a third propagation matrix, respectively; and to scale... The feature labels are subjected to global average pooling to obtain channel descriptors; then the channel descriptors are input into a multilayer perceptron to generate the unnormalized scores of the adaptive feature proximity relationship, the local spatial hypergraph, and the target / endangered organ anatomical structure hypergraph; the fusion weights of the first propagation matrix, the second propagation matrix, and the third propagation matrix are obtained through the Softmax function; the first propagation matrix, the second propagation matrix, and the third propagation matrix are fused using the fusion weights to obtain the fused propagation matrix.

[0021] In one possible design approach of the first aspect, the first-stage coarse dose field estimation network is supervised by structure-weighted voxel loss; the second-stage structure-aware refinement network is jointly supervised by structure-weighted voxel loss and dose-volume histogram tail proxy loss.

[0022] In a second aspect, a three-dimensional radiotherapy dose distribution prediction device is provided, comprising one or more functional units or modules for performing the method as described in any possible implementation of the first aspect.

[0023] Thirdly, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the method as described in any possible implementation of the first aspect.

[0024] Beneficial effects:

[0025] Based on the above technical solution, the clinically delineated target area and organs at risk are converted into anatomical structure hyperedges, enabling multiple feature markers belonging to the same clinical structure to exchange information at the group level, thereby explicitly describing high-order relationships within the same clinical structure. Simultaneously, adaptive feature relationships, local spatial hypergraphs, and anatomical structure hypergraphs are used to preserve patient-specific feature similarities while maintaining 3D spatial continuity and clinical structural semantics. Relationship propagation is performed at two deep semantic scales, with the higher-resolution scale preserving more spatial details and the lower-resolution scale providing a wider range of contextual information, thus balancing dose gradient boundaries and global dose field structure. Through learning-based relationship fusion, the relative weights of the three relationships are dynamically determined based on current patient characteristics, and the early training process of the stable model is updated through zero-initialization gated residuals. Simultaneously, coarse dose prediction and final refined dose prediction are supervised, and the average dose and near-maximum dose of the high and low dose tails of the target area and organs at risk are constrained by the dose-volume histogram tail proxy loss, improving the consistency of structural-level dose prediction. Ultimately, accurate prediction of 3D radiotherapy dose distribution is achieved. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is an overall framework diagram provided in the embodiments of this application;

[0028] Figure 2 This is a schematic diagram of the dual-scale three-view hypergraph refinement module provided in an embodiment of this application;

[0029] Figure 3 This is a comparison chart of the three-dimensional dose prediction results of various methods on the OpenKBP dataset provided in the embodiments of this application;

[0030] Figure 4 This is a comparison chart of the DVH curves predicted by various methods on the OpenKBP dataset provided in this application embodiment and the clinical reference DVH curves;

[0031] Figure 5 This is a comparison chart of the three-dimensional dose prediction results of various methods on the RLCD dataset provided in the embodiments of this application;

[0032] Figure 6 This is a comparison chart of the DVH curves predicted by various methods and the clinical reference DVH curves on the RLCD dataset provided in this application embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0036] This embodiment belongs to the fields of medical image processing, radiotherapy planning assistance, and artificial intelligence technology. Specifically, it relates to a method for predicting the three-dimensional radiotherapy dose distribution of a patient by using the patient's computed tomography CT body data, target area mask, and organ at risk mask, through a two-stage three-dimensional encoder-decoder network and dual-scale three-view hypergraph relationship propagation.

[0037] This embodiment employs a two-stage three-dimensional dose prediction framework, progressing from coarse to fine. The first stage generates coarse dose prediction and decoding features based on patient CT body data, target area masks, and organs at risk masks. The second stage concatenates the decoded features from the first stage with the original anatomical input and, at two deep semantic scales, jointly utilizes adaptive feature relationships, local spatial hypergraphs, and anatomical structure hypergraphs to refine the coarse dose features using structure-aware techniques, generating the final three-dimensional dose prediction result.

[0038] Specifically, this embodiment includes the following steps:

[0039] Step S1: Acquire patient CT body data, target area mask, and organ at risk mask; normalize the CT body data and reference dose; and combine the CT images and masks of different target areas and organs at risk along the channel dimension into a multi-channel input tensor.

[0040] Step S2: Input the multi-channel input tensor into the first-stage coarse dose field estimation network. The first stage adopts a three-dimensional encoder-decoder structure, generating coarse dose predictions and the final decoded features of the first stage through multi-level encoding, downsampling, skip connections, and step-by-step decoding.

[0041] Step S3: Adjust the number of channels of the last-stage decoding feature in the first stage using three-dimensional convolution, and concatenate the adjusted decoding feature with the original multi-channel input tensor along the channel dimension to obtain the input tensor of the second stage.

[0042] Step S4: Input the second-stage input tensor into the structure-aware thinning network. The second stage adopts a 3D encoder-decoder structure, and sets up three-view hypergraph thinning modules at the first deep scale output by the fourth-level encoder and the second deep scale output by the bottleneck layer.

[0043] Step S5: At each hypergraph refinement scale, the 3D feature map is converted into a feature label matrix, and adaptive feature relations, local spatial hypergraphs, and anatomical structure hypergraphs are constructed respectively. The adaptive feature relations are determined based on the feature distances between current feature labels; the local spatial hypergraph is determined based on 3D local windows; and the anatomical structure hypergraph is determined based on the downsampled target region and the mask of organs at risk.

[0044] Step S6: Calculate the propagation matrix based on the association matrices corresponding to the three relationships. Generate fusion weights for the three relationships using the learning-based fusion module, perform weighted fusion on the propagation matrix, and update the feature labels using a gated residual method. The updated feature labels are then restored to a 3D feature map and input into the subsequent encoding or decoding layer of the second stage.

[0045] Step S7: During training, both the first-stage coarse dose prediction and the second-stage final dose prediction are supervised simultaneously. The first stage uses structure-weighted voxel loss for auxiliary supervision; the second stage uses structure-weighted voxel loss and dose-volume histogram tail proxy loss for joint supervision.

[0046] In step S8, during the inference process, the dose prediction generated in the second stage is used as the final output. The negative values ​​in the prediction results and the voxel values ​​outside the effective dose region are set to zero, and the normalized dose is restored to a three-dimensional dose distribution in gray (Gy).

[0047] Specifically:

[0048] 1. Symbol Explanation and Background Introduction

[0049] Given the patient's CT body data and clinical delineation structure, the task of this embodiment is to predict the corresponding three-dimensional radiotherapy dose distribution for the patient.

[0050] Each case is represented as a multi-channel input tensor: .in, Represents the real number field; , and These represent the number of voxels in the depth, height, and width directions of the 3D volume data, respectively. This represents the total number of channels for clinically delineated structures. The first channel of the input tensor is the normalized CT image, and the rest... The channels are for the target area and the mask for organs at risk.

[0051] Let the number of target areas be The number of organs at risk is ,but: The normalized reference dose is denoted as: The effective dose mask used to define the scope of monitoring, post-processing, and evaluation is denoted as: The dose prediction network learns the following mapping: ...(1)

[0052] in, The parameter set is A three-dimensional dose prediction network; This includes convolution parameters, channel projection parameters, relational fusion parameters, and gating parameters in the network; This represents the predicted normalized three-dimensional dose distribution. Effective dose mask. It is not used as network input, but only for loss calculation, prediction result post-processing, and evaluation metric calculation. In the OpenKBP embodiment, the number of target regions... endangering the number of organs ,therefore Each case's input tensor contains a total of 11 channels.

[0053] 2. Methods

[0054] 2.1 Overall Framework

[0055] like Figure 1 As shown, this embodiment adopts a cascaded two-stage three-dimensional dose prediction structure, including a first-stage coarse dose field estimation network and a second-stage structure-aware refinement network.

[0056] The first stage is denoted as Its input is the patient's multi-channel input tensor. The first stage employs an encoder-decoder structure similar to a 3D U-shaped network, including four downsampling operations, a bottleneck layer, corresponding progressive upsampling, and skip connections between the encoder and decoder. Each encoder stage contains two sets of 3D convolutions, instance normalization, and ReLU activation operations. The decoder performs upsampling using trilinear interpolation and... Features are extracted using 3D convolution; the upsampled features are concatenated with the skip connection features of the corresponding encoder and then input into subsequent convolutional blocks.

[0057] Phase 1 Output Crude Dose Prediction and the final level of decoding features .in, Used to provide auxiliary crude dose monitoring Used to construct the input for the second stage. 3D Convolution Projection Adjustment The number of channels is calculated, and the adjusted features are compared with the original input tensor. splicing along the channel: .in, express 3D convolutional projection; This indicates that feature splicing is performed along the channel dimension; This represents the input tensor for the second stage.

[0058] The second stage is referred to as ,by As input. The second stage also adopts a four-stage three-dimensional encoder-decoder structure, with the output of the fourth-stage encoder serving as the input. Scale and bottleneck layer output Each scale is set with a hyper-view refinement module.

[0059] The second-stage decoder upsamples the enhanced deep features step by step and fuses encoded features at different scales through skip connections. Finally, it refines the dose prediction by outputting the dose prediction head. During the training process, It accepts auxiliary supervision of structure-weighted voxel loss. It accepts joint supervision from structure-weighted voxel loss and dose-volume histogram tail proxy loss. The inference process employs... This serves as the final dose prediction result.

[0060] 2.2 Dual-Scale Three-View Hypermap Refinement

[0061] Figure 2 The dual-scale three-view hypermap refinement module shown is not an independent dose prediction network, but rather... Figure 1 The second-stage network shown The relationship (enhanced) refinement unit in the middle.

[0062] Second stage input First, it undergoes multi-stage encoder processing, with the output at the fourth-stage encoder. Scale features, output at the bottleneck layer Scale features. Features at both scales are input into the hypermap refinement module of the three-view drawing, which has the same structure and independent parameters. For scale... The input feature map is denoted as: .in, Indicates the number of cases in the batch; Representing scale The number of feature channels below; , and These represent the depth, height, and width of the feature map at this scale, respectively.

[0063] Flatten the 3D feature map into a feature label matrix: .in, Only indicating scale The number of feature markers below does not represent the number of clinical structures.

[0064] Subsequently, adaptive feature proximity relationships, local spatial hypermaps, and anatomical structure hypermaps are constructed based on feature similarity, 3D local neighborhood, and target / endangered organ masks, respectively. After message propagation of each of the three relationships, they are weighted and fused using learned fusion weights, and then added to the original feature labels using a gated residual method. The updated feature labels are then reconstructed into 3D feature maps and output to... Figure 1 The subsequent layers of the second-stage network.

[0065] Adaptive feature relationships are established based on content similarity between current feature labels. For each feature label, the feature distance between it and other labels in the same case is calculated and then analyzed via Top-... Search for the nearest one Each label is assigned a label. An adaptive relation is constructed using the current label and its selected nearest neighbors, and the relation matrix is ​​populated with weights calculated from the feature distances. Among them, subscript This represents an adaptive feature relationship. In each forward propagation, nearest neighbors are reselected based on the current features. The gradient propagates through the weights of the selected relationships, and the nearest neighbor index is used for discrete selection. This embodiment sets... Therefore, this view primarily provides paired patient-specific feature relationships; higher-order group relationships are mainly provided by the local spatial hypermap and the anatomical hypermap.

[0066] Local spatial hypergraphs are used to preserve local spatial continuity in 3D volumetric data. At scale... Slide on the three-dimensional marker with a step size of 1 A local window is defined, where each valid window forms a hyperedge, connecting all 27 feature labels within the window. The incidence matrix of the local spatial hypergraph is denoted as: Among them, subscript Indicates local spatial relationships; Representing scale The number of hyperedges in the lower local space. When the input scale is... hour: When the input scale is hour: This view enables adjacent spatial locations to exchange information and stabilizes adaptive feature relationships.

[0067] The anatomical hypermap transforms clinically delineated target areas and organs at risk masks into explicit relational priors. Let all clinical structural masks be: A threshold is applied to the input structure mask. Binarization is performed, and then the structural mask is downsampled to the scale using 3D adaptive max pooling. If a pooling unit overlaps with a clinical structure, the corresponding feature label is considered to belong to that structure. Max pooling can better preserve the membership relationships at the structural boundaries. The downsampled structural mask is rearranged into an anatomical structure association matrix: Among them, subscript This represents the clinical anatomical relationships. Each target region or organ at risk is defined by an anatomical hyperedge, connecting all feature markers belonging to that clinical structure. Therefore, the target region and organ at risk masks in this embodiment serve two purposes simultaneously: first, as input channels for the dose prediction network; and second, to construct structural-level hyperedges, providing clinical anatomical priors for hypergraph relationship propagation.

[0068] 2.3 Hypergraph Propagation

[0069] For any hypergraph incidence matrix Its propagation matrix is ​​calculated according to the following formula: ...(2)

[0070] in, This represents the association matrix of the current relational view; Represents the node degree matrix; Represents the hypermarginality matrix; Represents the diagonal hyperside weight matrix; superscript This indicates the matrix transpose.

[0071] This embodiment will Setting it as an identity matrix allows the three relational views to be distinguished primarily by their respective topological connection structures. Substituting the three relational matrices into formula (2) yields: ... (3)

[0072] in, , and Representing scale The adaptive feature propagation matrix, local spatial propagation matrix, and anatomical structure propagation matrix are used. The adaptive correlation matrix is ​​recalculated from the current features in each forward propagation, while the local spatial and anatomical correlation matrices are determined by the sliding window geometry and structural mask, respectively.

[0073] 2.4 Relationship Fusion and Gated Residual Update

[0074] The three relational views are merged at the propagation matrix level.

[0075] First, regarding scale The feature labels are subjected to global average pooling to obtain channel descriptors; then the channel descriptors are input into a multilayer perceptron including two fully connected layers and a ReLU activation function to generate unnormalized scores for three relation views; finally, the fusion weights of the three relations are obtained through the Softmax function. , and The three fusion weights satisfy: The fusion propagation matrix is ​​as follows: ... (4)

[0076] in, , and These represent the fusion weights for adaptive feature relationships, local spatial relationships, and anatomical structure relationships, respectively. Message propagation is performed using the fusion propagation matrix, and feature labels are updated using a gated residual approach. ... (5)

[0077] in, Feature tags indicating enhanced relationships; Represents the learnable channel linear projection matrix; This indicates that batch normalization is performed along the channel. Represents the ReLU activation function; This represents the learnable scalar gating parameters. Initializing to 0 makes the hypergraph refinement module approximately an identity mapping at the start of training, thereby reducing the unstable impact of relation propagation on early network optimization. Updated feature labels. The feature map is restored to a 3D feature map and then input into the subsequent encoding layer, bottleneck layer, or decoding layer in the second stage.

[0078] 2.5 Training Objectives

[0079] During training, both the first-stage coarse dose prediction and the second-stage final dose prediction were supervised. These are denoted as follows: and Before calculating the loss, a ReLU operation is performed on both prediction results to ensure that the predicted dose is non-negative.

[0080] Let the union of all target area masks be... The union of all organs endangering the mask is The voxel weight map is defined as follows: ... (6)

[0081] in, This represents prime-property multiplication; and These represent the additional voxel weights for the target area and organs at risk, respectively. This embodiment sets the following: The two-stage structure-weighted voxel loss is: ... (7) ... (8)

[0082] and These represent the structure-weighted voxel losses in the first and second stages, respectively. and Indicates the corresponding loss weight; and Representing the clinical reference three-dimensional dose distribution and the... Predicted dose distribution for the stage; This represents a voxel weight map, with the same size as the dose distribution. This embodiment is configured as follows: To improve structural-level dose consistency, further refinement of final dose prediction was conducted. Calculate the tail surrogate loss of the order-based dose-volume histogram. For each target region, within the masked and effective dose regions of that target region, calculate the mean of the highest 1% voxel dose, the mean of the lowest 5% voxel dose, and the mean of the lowest 1% voxel dose, respectively, as... , and Differentiable proxies are used. For each organ at risk, the mean dose and the mean dose of the top 8 voxels are calculated. The mean dose of the top 8 voxels is used only as a near-maximum dose surrogate stable during the training phase and is not used to replace the dose calculated based on the actual voxel spacing during the evaluation phase. Let the set consist of surrogate statistics for all target areas and organs at risk. The tail proxy loss of the dose-volume histogram is: ...(9)

[0083] in, This represents a function used to calculate a surrogate statistic from the dose distribution. This indicates the total number of statistical items.

[0084] The total training loss of the model is: ... (10)

[0085] in, Indicates the current training iteration number; This represents the dose-volume histogram loss weight corresponding to the current iteration, with a maximum value of 0.1. In the first 5000 training iterations, this weight linearly increases from 0 to 0.1. Reasoning is conducted using... As the final prediction, negative values ​​and voxel values ​​outside the effective dose region in the prediction results are set to zero, and the normalized dose is converted into a three-dimensional dose distribution in Gy.

[0086] Table 1. OpenKBP Official Test Set Scores

[0087] Note: Results are in Gy; * indicates a single random seed; † indicates that the method relying on beam information lacks beam geometry input in OpenKBP.

[0088] As shown in Table 1, the HyperDose method of this embodiment achieves a dose score of 2.504 Gy and a DVH score of 1.413 Gy on the OpenKBP official test set, both lower than the comparative methods listed in the table. Compared to HD U-Net, the dose score of this embodiment is reduced from 2.685 Gy to 2.504 Gy, and the DVH score is reduced from 1.567 Gy to 1.413 Gy, indicating that this embodiment can simultaneously improve the accuracy of voxel-level dose prediction and the accuracy of dose-volume histogram prediction for the target area and organs at risk.

[0089] Table 2. OpenKBP Auxiliary Clinical Indicators

[0090] As shown in Table 2, this embodiment achieves low results in several indicators, including global maximum dose error, target area D90 error, target area maximum dose error, target area homogeneity index error, and mean dose error of organs at risk. Specifically, the global maximum dose error is 1.308 Gy, the target area D90 error is 0.978 Gy, the target area maximum dose error is 1.322 Gy, and the mean dose error of organs at risk is 1.317 Gy. This indicates that this embodiment not only improves the overall dose field prediction accuracy but also maintains good target area coverage and limits the dose received by organs at risk.

[0091] Table 3 Key and Overall Indicators of RLCD

[0092] Note: All methods report single results for random seed 2026.

[0093] Table 4 RLCD target area and organ at risk indicators

[0094] As shown in Tables 3 and 4, on the RLCD clinical dataset, this embodiment achieved a dose score of 1.653 Gy and a DVH score of 6.632 Gy, as well as a near-maximum dose error of 7.773 Gy for organs at risk and a mean dose error of 1.977 Gy for organs at risk. These results demonstrate that this embodiment still exhibits good dose prediction performance under different data sources and clinical scenarios.

[0095] Table 5 Ablation Experiments of OpenKBP Components

[0096] Note: Single random seed diagnostic ablation; the lower the value, the better.

[0097] Table 6 OpenKBP Relation Fusion Ablation Experiment

[0098] Table 7. OpenKBP Adaptive View Nearest Neighbor Ablation Experiment

[0099] Tables 5 to 7 present the ablation results for different components of this embodiment. The Dosescore and DVH score for the complete HyperDose configuration are 2.435 Gy and 1.362 Gy, respectively. After removing gated updates, adaptive feature views, local spatial views, anatomical hyperedges, dual-scale propagation, or dose-volume histogram losses, at least one evaluation index shows a decrease to varying degrees, indicating that the three types of relational views, dual-scale propagation, learning-based relational fusion, and gated residual updates together constitute the effective technical solution of this embodiment.

[0100] Table 6 further shows that the learning-based relationship fusion is superior to the fusion methods of fixed equal weights and preset anatomical weights, indicating that dynamically determining the importance of the three relationships based on the current patient characteristics can improve prediction accuracy. Table 7 shows that the best comprehensive result is obtained when the number of nearest neighbors k in the adaptive feature view is 1; therefore, k is set to 1 in this embodiment.

[0101] Table 8. Average comparison of single-inference and four-view flipping in OpenKBP.

[0102] Note: The time period does not include the calculation of evaluation indicators.

[0103] As shown in Table 8, the four-view flipped averaging method can reduce the Dose score from 2.549 Gy to 2.435 Gy and the DVH score from 1.392 Gy to 1.362 Gy, but the single-case inference time increases from 0.39 seconds to 1.18 seconds. In practical applications, either single-inference or four-view flipped averaging can be selected based on the required prediction accuracy and calculation speed.

[0104] Table 9 RLCD Model Size, Inference Time, and Memory Usage

[0105] Note: Reasoning time does not include the calculation of evaluation indicators.

[0106] As shown in Table 9, the model in this embodiment contains approximately 32.85 million parameters, with a single-case inference time of approximately 1.20 seconds and a peak memory usage of approximately 2.56 GB. Compared to the diffusion-based DoseDiff method, this embodiment has fewer parameters, shorter inference time, and lower memory usage, meeting the computational efficiency requirements of clinical auxiliary applications while maintaining prediction accuracy.

[0107] Figure 3 and Figure 5 Representative cases from the OpenKBP and RLCD datasets are presented respectively. The figures show CT images, clinical reference doses, predicted doses for each comparison method, and prediction errors. The prediction results in this embodiment show high consistency with clinical reference doses in the high-dose region of the target area, the region adjacent to the target area and organs at risk, and the low-dose outward extension region, and can maintain the dose gradient boundaries well.

[0108] Figure 4 and Figure 6 The dose-volume histograms from the OpenKBP and RLCD datasets are presented for comparison. The horizontal axis represents the dose, and the vertical axis represents the proportion of structures receiving at least the corresponding dose. Different curves correspond to the target area and organs at risk, respectively. The closer the predicted curve is to the clinical reference curve, the more accurate the structural dose prediction. In this embodiment, the predicted curve maintains good consistency with the clinical reference curve in the high-dose segment of the target area and the low-dose segment of organs at risk.

[0109] This embodiment has the following advantages:

[0110] First, this embodiment converts the clinically delineated target area and organs at risk into anatomical structure hyperedges, enabling multiple feature markers belonging to the same clinical structure to exchange information at the group level, thereby explicitly describing the higher-order relationships within the same clinical structure.

[0111] Second, this embodiment uses adaptive feature relationships, local spatial hypergraphs, and anatomical structure hypergraphs simultaneously, which can preserve patient-specific feature similarity relationships while taking into account three-dimensional spatial continuity and clinical structural semantics.

[0112] Third, this embodiment performs relation propagation at two deep semantic scales, where the higher resolution scale retains more spatial details and the lower resolution scale provides a wider range of contextual information, thereby taking into account both the dose gradient boundary and the global dose field structure.

[0113] Fourth, this embodiment uses learning-based relationship fusion to dynamically determine the relative weights of the three relationships based on the current patient characteristics, and updates the early training process of the stable model through zero-initialization gated residuals.

[0114] Fifth, this embodiment simultaneously monitors coarse dose prediction and final refined dose prediction, and improves the consistency of structural dose prediction by constraining the high and low dose tails of the target area and the average and near-maximum doses of organs at risk through the proxy loss of the dose-volume histogram tail.

[0115] This application also provides a three-dimensional radiotherapy dose distribution prediction device, including one or more functional units or modules, for performing the method described in any of the above embodiments.

[0116] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0117] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0118] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0119] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0120] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0121] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0122] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0123] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for predicting three-dimensional radiotherapy dose distribution, characterized in that, include: Obtain patient prediction data, which includes: computed tomography (CT) volume data, target area mask, and organ at risk mask; The prediction data is input into a three-dimensional dose prediction network for processing to obtain the three-dimensional radiotherapy dose distribution prediction results for the patient. The three-dimensional dose prediction network includes: The first-stage coarse dose field estimation network is used to generate coarse dose predictions and first-stage decoding features using the prediction data as input; the first-stage coarse dose field estimation network adopts a first three-dimensional encoder-decoder structure; The second-stage structure-aware refinement network is used to generate the final refined dose prediction by taking the decoded features from the first stage and the spliced ​​features of the prediction data as input. The second-stage structure-aware refinement network adopts a second 3D encoder-decoder structure. On the encoding side, the second 3D encoder-decoder structure processes several deep semantic scale feature labels as follows: constructing adaptive feature proximity relationships based on the similarity between the feature labels; constructing a local spatial hypergraph based on the 3D local adjacent positions of the feature labels; constructing a target area / endangered organ anatomical structure hypergraph of the feature labels based on the target area / endangered organ mask; fusing the adaptive feature proximity relationships, the local spatial hypergraph, and the target area / endangered organ anatomical structure hypergraph at the propagation matrix level to obtain a fused propagation matrix; and updating the feature labels using the fused propagation matrix through a gated residual method.

2. The method as described in claim 1, characterized in that, The first 3D encoder-decoder structure adopts a 3D U-shaped network encoder-decoder structure, including: four successive downsampling, a bottleneck layer, corresponding successive upsampling, and skip connections between the encoder and the decoder. Each encoder level contains two sets of 3D convolutions, instance normalization, and ReLU activation operations. The decoder performs upsampling using trilinear interpolation and extracts features through 3D convolution. The upsampled features are concatenated with the skip connection features of the corresponding encoder and then input into the subsequent convolutional block.

3. The method as described in claim 1, characterized in that, The second 3D encoder-decoder structure adopts a 3D U-shaped network encoder-decoder structure, including: four successive downsampling steps, a bottleneck layer, corresponding successive upsampling steps, and skip connections between the encoder and decoder. Each encoder step contains two sets of 3D convolutions, instance normalization, and ReLU activation operations. The decoder performs upsampling using trilinear interpolation and extracts features through 3D convolution. The upsampled features are concatenated with the skip connection features of the corresponding encoder and then input into subsequent convolutional blocks. The second 3D encoder-decoder structure outputs 3D feature maps at different deep semantic scales at the fourth encoder and the bottleneck layer, respectively. The 3D feature maps serve as the input to the relation enhancement and refinement module of the second 3D encoder-decoder structure. The relation enhancement and refinement module is used to flatten the 3D feature maps into a feature label matrix and perform subsequent processing to obtain the updated feature labels.

4. The method as described in claim 3, characterized in that, The relationship enhancement and refinement module is specifically used for: calculating the feature distance between the current feature label and other feature labels, and then... Search for the nearest one Each feature label is used to form an adaptive relationship with the current feature label and its selected nearest neighbors, and the first correlation matrix of the adaptive feature proximity relationship is filled with weights calculated from the feature distance.

5. The method as described in claim 4, characterized in that, The relationship enhancement and refinement module is specifically used for: at the scale On the three-dimensional feature marker volume, local windows are slid with a step size. Each effective window constitutes a hyperedge and simultaneously connects all feature markers within the window, thus obtaining the second correlation matrix of the local spatial hypergraph.

6. The method as described in claim 5, characterized in that, The relationship enhancement and refinement module is specifically used to: binarize the input target / organ at risk mask, and then downsample the target / organ at risk mask to the specified scale using three-dimensional adaptive max pooling. If a pooling unit overlaps with the target area / endangered organ, the corresponding feature marker is considered to belong to the target area / endangered organ, and the downsampled target area / endangered organ mask is rearranged into the third correlation matrix of the anatomical structure hypermap.

7. The method as described in claim 6, characterized in that, The relationship enhancement and refinement module is specifically used to: convert the first correlation matrix, the second correlation matrix, and the third correlation matrix into a first propagation matrix, a second propagation matrix, and a third propagation matrix, respectively; and to scale... The feature labels are subjected to global average pooling to obtain channel descriptors; then the channel descriptors are input into a multilayer perceptron to generate the unnormalized scores of the adaptive feature proximity relationship, the local spatial hypergraph, and the target / endangered organ anatomical structure hypergraph; the fusion weights of the first propagation matrix, the second propagation matrix, and the third propagation matrix are obtained through the Softmax function; the first propagation matrix, the second propagation matrix, and the third propagation matrix are fused using the fusion weights to obtain the fused propagation matrix.

8. The method as described in claim 7, characterized in that, The first-stage coarse dose field estimation network uses structure-weighted voxel loss for auxiliary supervision; the second-stage structure-aware refinement network uses structure-weighted voxel loss and dose-volume histogram tail proxy loss for joint supervision.

9. A three-dimensional radiotherapy dose distribution prediction device, characterized in that, It includes one or more functional units or modules for performing the method of any one of claims 1-8.

10. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-8.

Images