Dual scene fusion head and neck cancer radiotherapy 3D dose prediction method and system

Abstract

The present application relates to the field of radiotherapy 3D dose prediction, in order to solve the problem of low accuracy and poor real-time performance of existing dose prediction, a double-scene fusion head and neck cancer radiotherapy 3D dose prediction method and system are provided. The double-scene fusion head and neck cancer radiotherapy 3D dose prediction method comprises: acquiring multi-source original data of a head and neck cancer patient and forming a multi-channel tensor; using a pre-trained 3D dose prediction model to extract different scale anatomical features of the head and neck of the multi-channel tensor, fusing the different scale anatomical features and upsampling to obtain the head and neck cancer radiotherapy 3D dose prediction result of the head and neck cancer patient; the loss function of training the 3D dose prediction model is composed of the weighted sum of the global dose error, the DVH index error, the multi-scale Dice loss and the low dose area accuracy loss. It can improve the accuracy of the head and neck cancer radiotherapy 3D dose prediction result and meet the real-time demand.

Description

Technical Field

[0001] This invention relates to the field of 3D dose prediction in radiotherapy, and more particularly to a dual-scene fusion method and system for 3D dose prediction in head and neck cancer radiotherapy. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Radiation therapy is a core treatment for head and neck cancers. Its core objective is to deliver a precise dose to the planning target volume (PTV) while maximizing the protection of surrounding organs at risk (OARs, such as the parotid gland, brainstem, and spinal cord). Traditional radiotherapy planning relies on the experience of the planner for iterative adjustments, which is subjective and inefficient.

[0004] To address the aforementioned issues, existing technologies have provided dose prediction models. However, to improve accuracy, dose prediction models often employ complex convolutional blocks or transposed convolutional upsampling, leading to a surge in the number of parameters and computational load, making it difficult to meet the needs of real-time clinical planning. Furthermore, the skip connections in the cascaded architecture of existing dose prediction models often use simple splicing, which easily introduces redundant information and lacks multi-scale intermediate supervision. This results in an imbalance in the model's feature learning at different dose scales, reducing the accuracy of 3D dose prediction results for head and neck cancer radiotherapy and failing to meet real-time requirements. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a dual-scene fusion method and system for predicting 3D dose in head and neck cancer radiotherapy, which can improve the accuracy of 3D dose prediction results for head and neck cancer radiotherapy and meet real-time requirements.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of the present invention provides a dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy.

[0008] In one or more embodiments, a dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy is provided, including:

[0009] We acquire multi-source raw data from head and neck cancer patients and construct a multi-channel tensor, which includes clinical CT imaging data, anatomical structure masks of organ at risk areas, and planned target area localization masks.

[0010] The 3D dose prediction model is pre-trained to extract multi-channel tensor anatomical features of the head and neck at different scales. The anatomical features at different scales are fused and upsampled to convert the 3D dose into a single-channel prediction tensor output, thus obtaining the 3D dose prediction results for head and neck cancer radiotherapy in patients with head and neck cancer.

[0011] The loss function for training the 3D dose prediction model consists of a weighted sum of four factors: global dose error, DVH (Dose and Volume Histogram) index error, multi-scale Dice loss, and low-dose region accuracy loss. The low-dose region accuracy loss consists of three factors: low-dose region voxel error, low-dose gradient consistency, and clinical threshold achievement rate.

[0012] As one implementation method, the weight of accuracy loss in low-dose areas is dynamically adjusted based on the number of training rounds.

[0013] As one implementation method, the low-dose region voxel error sub-item is: the average error between the dose prediction value of the voxel by the 3D dose prediction model and the actual dose value.

[0014] As one implementation, the low-dose gradient consistency sub-item is set as follows: :

[0015] ;

[0016] In the formula: P is the set of voxels in the gradient region defined by the boundary of the organ at risk. For 3D gradient operators, cos( ) represents the gradient direction cosine similarity; The dose value predicted by the 3D dose prediction model. This is the actual dose value; The total number of voxels contained in the voxel set P corresponding to the gradient region that defines the boundary of the organ at risk, i.e., the total number of all voxels in set P.

[0017] As one implementation method, the clinical threshold achievement rate sub-item is: :

[0018] ;

[0019] in, Number of organs at risk; The loss function; The first dose predicted by the 3D dose prediction model The percentage of the volume in the organ at risk of infection where the dose is ≤ the corresponding clinical threshold. The first under the true clinical gold standard The percentage of the volume in the organ at risk that is ≤ the corresponding clinical threshold.

[0020] In one implementation, the 3D dose prediction model includes an encoder, an extended skip connection block, and a decoder, wherein the encoder is used to extract anatomical features of the head and neck at different scales from multi-channel tensors;

[0021] The extended skip connection block and decoder include multi-scale convolutional blocks, channel attention blocks, and cascaded channel attention blocks and spatial attention blocks;

[0022] The decoder is used to fuse anatomical features at different scales and upsample them to convert 3D dose into a single-channel prediction tensor output.

[0023] In one implementation, the decoder employs a 3D macrokernel grouping attention gate to fuse anatomical features at different scales. The 3D macrokernel grouping attention gate is used to divide the encoder features and decoder upsampled features into three groups according to the organ-at-risk zone-planned target zone grouping rule: the planned target zone exclusive group, the organ-at-risk zone exclusive group, and the organ-at-risk zone-planned target zone boundary group.

[0024] A second aspect of the present invention provides a dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy.

[0025] In one or more embodiments, a dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy includes:

[0026] The data acquisition module is used to acquire multi-source raw data from head and neck cancer patients and construct a multi-channel tensor, which includes clinical CT image data, anatomical structure mask of organ at risk area and planned target area localization mask.

[0027] The dose prediction module is used to extract multi-channel tensor anatomical features of the head and neck at different scales using a pre-trained 3D dose prediction model. It then fuses the anatomical features at different scales and upsamples them to convert the 3D dose into a single-channel prediction tensor output, thus obtaining the 3D dose prediction results for head and neck cancer radiotherapy in patients with head and neck cancer.

[0028] The loss function for training the 3D dose prediction model consists of a weighted sum of four factors: global dose error, DVH index error, multi-scale Dice loss, and low-dose region accuracy loss. The low-dose region accuracy loss consists of three factors: low-dose region voxel error, low-dose gradient consistency, and clinical threshold achievement rate.

[0029] A third aspect of the present invention provides a computer-readable storage medium.

[0030] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the dual-scene fusion method for predicting 3D doses for head and neck cancer radiotherapy as described above.

[0031] A fourth aspect of the present invention provides an electronic device.

[0032] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the dual-scene fusion method for predicting 3D dose for head and neck cancer radiotherapy as described above.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This invention utilizes a 3D dose prediction model to extract multi-channel tensors of anatomical features at different scales in the head and neck region. The fusion of these features yields 3D dose prediction results for head and neck cancer radiotherapy in patients with head and neck cancer. Specifically, the training of the 3D dose prediction model achieves dual-scene fusion of training and performance evaluation scenarios for 3D dose prediction in head and neck cancer radiotherapy. In the performance evaluation scenario, the training loss is determined by a weighted sum of four factors: global dose error, DVH index error, multi-scale Dice loss, and low-dose region accuracy loss. In the model training scenario, the low-dose region accuracy loss considers voxel error, low-dose gradient consistency, and clinical threshold achievement rate, guiding the model to focus on learning the numerical accuracy, gradient trend, and clinical compliance of the low-dose region. This ensures that the training direction aligns with the needs of organ-at-risk area protection, accurately capturing the spatial adjacency between organ-at-risk areas and the planned target area in the complex anatomical structure of the head and neck. This assists physicians in confirming the anatomical boundaries of organ-at-risk areas, avoiding diagnostic bias caused by structural ambiguity, and achieving the goal of automated clinical radiotherapy planning and high-precision dose distribution generation. Attached Figure Description

[0035] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0036] Figure 1 This is a flowchart of the dual-scene fusion method for 3D dose prediction in head and neck cancer radiotherapy according to an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the 3D dose prediction model according to an embodiment of the present invention;

[0038] Figure 3 This is a comparison result between the dose prediction map and the actual dose map of an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the structure of the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy according to an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, 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 invention pertains.

[0043] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0044] Figure 1 A schematic diagram of the principle of the dual-scene fusion 3D dose prediction method for head and neck cancer radiotherapy according to an embodiment of the present invention is provided. Figure 1 The dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy in this embodiment may include the following steps S101 to S102.

[0045] The specific implementation process of steps S101 to S102 is as follows:

[0046] Step S101: Obtain multi-source raw data from head and neck cancer patients and construct a multi-channel tensor, which includes clinical CT image data, anatomical structure mask of organ at risk area, and planned target area localization mask.

[0047] In practice, clinical CT imaging data is acquired through CT imaging equipment.

[0048] The process of obtaining the anatomical structure mask of organ at risk areas is as follows: The acquisition of the anatomical structure mask of organ at risk areas (OARs) is based on head and neck medical images: First, CT volumetric images (DICOM format) of the patient's head and neck are acquired, which need to completely cover the radiotherapy target area and surrounding target OARs (such as spinal cord, parotid gland and other organs) to ensure the spatial integrity of anatomical structure information; then, the contours of each OAR in the CT images are segmented. In clinical practice, the radiation oncologist can manually delineate the organ boundaries layer by layer based on the images, or combine the segmentation model adapted to the head and neck anatomy to achieve semi-automatic annotation, and clarify the 3D spatial range of each type of OAR.

[0049] After segmentation, the contour regions of each OAR are converted into binary masks (voxel values ​​"1" correspond to OAR regions, and "0" correspond to non-regions). Finally, clinicians review the matching degree between the mask and the actual anatomical structure to ensure its spatial positioning accuracy. This mask serves as the basis for the subsequent positioning of the voxel set of the "OAR boundary ±3mm gradient region," providing a spatial range benchmark for the calculation of sub-items such as low-dose gradient consistency.

[0050] The process of obtaining the target area localization mask is as follows: The target area localization mask is obtained based on the clinical target volume (CTV) and combined with the control requirements of radiotherapy position variation: First, CT images of the patient's head and neck are acquired. The radiation oncologist manually delineates the clinical target area of ​​the tumor and potential invaded tissues on the images according to the tumor pathology type and invasion range. Then, in accordance with the clinical quality control standards for head and neck radiotherapy, the boundaries of the CTV are expanded in three dimensions to determine the spatial contour of the target area.

[0051] The PTV contour is then converted into a binary mask (a voxel value of "1" corresponds to the PTV region and "0" corresponds to the non-region). Finally, the spatial range of the mask is reviewed by the radiation physicist and clinician to ensure that it fully covers the possible location variations of the tumor while avoiding excessive expansion that would increase the dose to normal tissue. This mask will serve as the spatial positioning benchmark for the "high-dose coverage target area" in subsequent dose prediction.

[0052] Step S102: Use a pre-trained 3D dose prediction model to extract multi-channel tensor anatomical features of the head and neck at different scales, fuse the anatomical features at different scales and upsample them, and convert the 3D dose into a single-channel prediction tensor output to obtain the 3D dose prediction results for head and neck cancer radiotherapy in patients with head and neck cancer.

[0053] Figure 3 Figure (a) in the diagram is the actual dose diagram. Figure 3Figure (b) in the figure is a predicted dose diagram. It can be seen that the prediction results of the dual-scene fusion 3D dose prediction method for head and neck cancer radiotherapy in this embodiment of the invention are close to the actual dose, which can guide doctors to make clinical decisions on the 3D dose of head and neck cancer radiotherapy.

[0054] In embodiments of the present invention, such as Figure 2 As shown, the 3D dose prediction model includes an encoder, an extended skip connection block, and a decoder. The encoder is used to extract anatomical features of the head and neck at different scales from the multi-channel tensor.

[0055] The extended skip connection block and decoder consist of multi-scale convolutional blocks, channel attention blocks, and cascaded channel attention blocks and spatial attention blocks;

[0056] The decoder is used to fuse and upsample anatomical features at different scales, converting 3D dose into a single-channel prediction tensor output.

[0057] The role of multi-scale convolutional blocks (MSCBs) is to simultaneously capture the spatial features of small-volume OARs (such as the optic nerve, with a small receptive field) and large-volume planning target areas (PTVs) / parotid glands (with large receptive fields) in the head and neck by using convolutional kernels of different sizes such as 3×3, 5×5, and 7×7. This outputs a feature map that integrates multi-scale information, solving the problem of traditional single-scale convolutions "missing OAR boundary details or losing global PTV relationships". This provides a complete anatomical-dose feature basis for accurate prediction of low-dose areas (around OARs).

[0058] The role of the Channel Attention Block (CAB) is to weight the feature channels after multi-scale convolution (such as assigning high weights to the "OARs low-dose region associated channel" and filtering the "irrelevant normal tissue channel"), output a channel-level focused feature map, solve the problem of redundancy in head and neck feature channels, strengthen the supervision signal in the low-dose region (avoiding low-dose features being masked by the high-dose PTV channel), and allow the decoder to prioritize learning features related to OARs protection.

[0059] The purpose of the cascaded channel + spatial attention block (CAB+SAB) is to further weight the spatial location of the feature map on the basis of channel weighting (such as assigning high weight to the "OAR boundary ±3mm low dose area"), output a feature map with spatial + channel dual focus, accurately locate the voxel position in the low dose area, improve the prediction accuracy of voxel error and gradient consistency in the low dose area, and avoid local dose abrupt changes in OARs.

[0060] In the decoder, a 3D macrokernel grouping attention gate is used to fuse anatomical features at different scales. The 3D macrokernel grouping attention gate is used to divide the encoder features and decoder upsampled features into three groups according to the organ at risk area-planned target area grouping rule: the planned target area exclusive group, the organ at risk area exclusive group, and the organ at risk area-planned target area boundary group.

[0061] The multi-scale convolutional block (MSCB) in this embodiment of the invention is a 3D multi-scale convolutional block (MSCB) used to replace ordinary convolutional blocks: the MSCB uses three sets of 3×3 3D depth convolutions with different dilatancy rates (d=1, 2, 3), which effectively cover the receptive fields of 3×3, 5×5, and 7×7 convolutional kernels, taking into account both small-volume OARS and large-volume PTV feature capture, and providing an accurate multi-scale feature foundation for LDRI supervision;

[0062] Based on the core design of the original EMCAD 2D architecture, the Multi-Scale Convolutional Block (MSCB) and Channel-Spatial Attention Module (CAB+SAB) are extended to 3D versions, replacing the traditional convolutional blocks in the traditional DOSE-PYFER cascaded encoder architecture. MSCB uses multiple sets of 3D convolutions with different dilation rates to capture anatomical features of the head and neck at different scales. CAB highlights the channel information related to OARs and PTV through feature weight allocation. SAB uses large-kernel convolution to focus on the dose gradient region around the tumor, constructing an enhanced encoder feature extraction unit to solve the problem of insufficient feature discriminativeness in traditional encoders.

[0063] The 3D Channel Attention (CAB) and 3D Spatial Attention (SAB) are combined: CAB uses 3D average pooling + 3D max pooling + 1×1×1 convolution to highlight the feature weights of channels related to OARs and PTV (e.g., the weight of the parotid gland channel is increased by 20%-30%); SAB uses a 7×7×7 large kernel 3D convolution (validated on the OpenKBP dataset, compared with 3×3×3 and 5×5×5 kernels, the 7×7×7 kernel can optimally cover the 5-10mm low-dose gradient region at the junction of head and neck OARs and PTV, while avoiding computational redundancy caused by excessive kernel size), focusing on the dose gradient region around the tumor;

[0064] The final output consists of several scale features (e.g., 5 scale features with dimensions of 128×128×128, 64×64×64, 32×32×32, 16×16×16, and 8×8×8 respectively), with the number of feature channels set to 16×2. (i-1) The distribution balances feature discriminative power and computational efficiency. Here, i = 1~5.

[0065] 3D Multiscale Convolutional Block (MSCB) Formula:

[0066] ;

[0067] In the formula: X is the 3D input feature tensor, which is also the residual connection term, ensuring stable gradient propagation; is 3D depthwise convolution; ChannelShuffle is channel shuffling operation; BN is 3D batch normalization. It is a convolution.

[0068] 3D Channel Attention Block (CAB) Formula:

[0069] ;

[0070] In the formula: σ is the Sigmoid activation function, AvgPool3D is 3D adaptive averaging; MaxPool3D is 3D max pooling; ⊙ is element-wise multiplication, which strengthens key channel features through weight allocation; This is the activation function.

[0071] 3D Spatial Attention Block (SAB) Formula:

[0072] ;

[0073] In the formula: It is the average along the channel dimension; Max pooling along the channel dimension; 7×7×7 convolution captures a wide range of spatial relationships and focuses on low-dose gradient regions; ⊙ represents element-wise multiplication; σ is the Sigmoid activation function; BN is 3D batch normalization.

[0074] Feature scale output formula:

[0075] ;

[0076] In the formula: For the output features at the i-th scale, the dimension decreases with the downsampling level, while the number of channels increases with the level, balancing details and semantic information; .

[0077] In the feature fusion stage of the decoder (corresponding to skip1~skip4 layers of the encoder), the traditional feature stitching method is replaced by a 3D large kernel grouped attention gate (LGAG) to solve the problem of redundant information interference. LGAG divides the encoder features and decoder upsampled features into 3 groups according to the "OARs-PTV grouping rule" (based on the PTV mask, voxels belonging only to PTV are grouped into the PTV exclusive group; those belonging only to OARS are grouped into the OARS exclusive group; voxels less than 3mm from the PTV or OARS boundary are grouped into the boundary group). The spatial relationship within each group is captured by 3×3×3 grouped convolution, generating attention weights (the boundary group weight is increased by 1.5 times, focusing on strengthening low-dose region features). Effective features are selected and fused according to the weights to improve the feature fusion accuracy of the low-dose region around OARs, providing high-quality fused features for subsequent LDRI supervision. After optimizing the feature fusion accuracy through LGAG in this step, it is necessary to further solve the contradiction of 'high computational cost and weak detail reconstruction' in the decoder upsampling stage. Therefore, a 3D efficient upconvolutional block (EUCB) replacement design is implemented.

[0078] The decoder replaces the traditional transposed convolution upsampling with 3D efficient upconvolutional blocks (EUCB), improving the accuracy of low-dose region detail reconstruction while reducing computational cost. The EUCB process involves: first doubling the feature size through trilinear interpolation (e.g., 16×16×16 → 32×32×32), then enhancing low-dose region details (e.g., OAR boundary gradients) with 3×3×3 depthwise convolutions, and finally adjusting the number of channels (to match the number of feature channels in the corresponding skip layer) using 1×1×1 convolutions. The number of parameters is only 1 / 16 of that of the traditional transposed convolution (e.g., reducing the number of parameters of the transposed convolution from 256×512 to 16×32), reducing decoder computation by 40% and further compressing single-example prediction time.

[0079] A multi-stage output supervision structure is constructed, setting dose prediction output heads at key stages of the decoder to generate prediction results at different scales. A combined loss function is used for model pre-training: global dose error loss ensures overall prediction accuracy, DVH index loss aligns with clinical evaluation criteria, and multi-scale Dice loss enhances low-dose region learning. During pre-training, relevant parameters for feature extraction are fixed, and only the parameters of the supervised output layer are optimized to accelerate model convergence and lay a foundation for accuracy.

[0080] Using preprocessed real clinical data (including CT images, 7 types of OARs masks, PTV masks, and corresponding real clinical dose maps) as the training set, end-to-end formal training was carried out based on a multi-scale pre-trained model. During training, the AdamW optimizer was used to dynamically adjust the learning rate and weight decay parameters of the dose prediction model according to the training iterations to suppress overfitting. At the same time, the "global-local-clinical" three-layer combined loss with LDRI was used to enhance the learning of low-dose areas. After training, the optimal weight file was saved.

[0081] The process of constructing the training sample set for training the 3D dose prediction model is as follows:

[0082] CT images of patients with head and neck cancer, masks of 7 types of OARs (such as brainstem and spinal cord) and PTV masks were acquired to form 9-channel 3D data, which were uniformly adjusted to 128×128×128 size; CT values ​​were normalized, OARs / PTV masks were binarized, and dose values ​​were subsequently normalized; OpenKBP public dataset (e.g., 340 cases) was used.

[0083] The loss function for training the 3D dose prediction model consists of a weighted sum of four factors: global dose error, DVH index error, multi-scale Dice loss, and low-dose region accuracy loss. The low-dose region accuracy loss consists of three factors: low-dose region voxel error, low-dose gradient consistency, and clinical threshold achievement rate.

[0084] The loss function for training the 3D dose prediction model is:

[0085] ;

[0086] In the formula: ~ The loss weights were determined through ablation experiments on the OpenKBP validation set, and the optimal values ​​were: =0.3、 =0.2、 =0.2、 =0.3, prioritizing monitoring intensity in low-dose areas. This refers to the global dose error. This refers to the error in the DVH index. For multi-scale Dice loss; This results in a loss of accuracy in low-dose areas.

[0087] Global dose error Using L1 loss, the mean absolute error between the predicted dose and the actual dose for all elements is calculated to ensure the accuracy of the overall dose distribution.

[0088] DVH index error Calculate the mean absolute error of 5 key DVH indicators to conform to routine clinical evaluation standards;

[0089] The five key DVH metrics are as follows:

[0090] 95% volumetric dose received in the PTV (D95): The dose received in 95% of the planned target volume (PTV).

[0091] 100% volumetric dose received in the planned target volume (PTV) (D100): The dose received in 100% of the planned target volume (PTV);

[0092] 5% volume dose (D5) of OARs: The highest dose received in 5% of the volume of an organ at risk (OAR, such as the spinal cord or parotid gland).

[0093] 50% volume dose received by OARs (D50): The dose received by 50% of the volume of an organ at risk (OAR).

[0094] Maximum dose to OARs (Dmax): The highest dose value (i.e., peak dose) received by all voxels in an organ at risk (OAR).

[0095] Multiscale Dice loss : Calculate the Dice coefficient for prediction maps at different scales (focusing on PTV and OARs regions), and then take the average value to enhance the consistency of multi-scale features;

[0096] The accuracy loss in the low-dose region is decomposed into three differentiable terms. Targeted optimization of the 5-20 Gy low-dose region surrounding OARs is then performed synergistically with 3D feature enhancement.

[0097] Low-dose region voxel error sub-item ( ):

[0098] ;

[0099] In the formula: V is the set of voxels within the "OARs mask perimeter range (e.g., 3mm) + dose (e.g., 5-20Gy)" (generated by the intersection of the OARs mask and the dose range mask), and N is the total number of voxels in V. Only the voxel error in the low-dose region is calculated to avoid global errors masking local problems. The model's dose prediction for voxel v. : The actual dose value of voxel v (clinically measured or gold standard dose).

[0100] Low dose gradient consistency sub-item ( ):

[0101] ;

[0102] In the formula: P is the set of voxels in the gradient region within the boundary setting range of OARs (e.g., ±3 mm) (voxels with absolute dose gradient values ​​> 5 Gy / mm calculated by the 3D Sobel operator). For 3D gradient operators, cos( The gradient direction cosine similarity is used to ensure that the gradient in the low-dose region is consistent with the true value, thus avoiding local dose abrupt changes in OARs. The dose value predicted by the model. It is the ground truth dose value. The total number of voxels contained in the voxel set P corresponding to the gradient region within the boundary range (e.g., ±3mm) of OARs.

[0103] Clinical threshold achievement rate sub-item ( ):

[0104] ;

[0105] In the formula: K is the number of OARs (e.g., 7), and the SmoothL1 loss avoids outlier interference and directly aligns with clinical OAR protection needs. The percentage of volumetric OARs with doses ≤ the corresponding clinical threshold in the k-th class predicted by the model. Under the true clinical gold standard, the volume percentage of OARs in class k with doses ≤ the corresponding clinical threshold.

[0106] Specifically, the weight of accuracy loss in low-dose areas is dynamically adjusted based on the number of training rounds.

[0107] For example, when the number of training rounds is less than or equal to the threshold of the first training round (e.g., 50), the weight of the accuracy loss in the low-dose region can be set to 0.2, which can be automatically adjusted according to the actual situation. At this stage, the model needs to prioritize learning the basic association between the head and neck anatomy and the dose distribution (e.g., 3D-MSCB capturing multi-scale PTV / OAR features, LGAG optimizing feature fusion). If the LDRI weight is too high (e.g., set directly to 0.3), it will cause the model to be "unbalanced", over-focusing on the low-dose region and ignoring the rationality of the overall dose distribution (e.g., insufficient high-dose coverage of PTV). Therefore, a lower weight is used to allow the model to first establish a good global feature foundation. The weight value can be flexibly changed according to the differences in the dataset, the adjustment of the radiotherapy plan, and the emphasis of clinical needs, but it must meet the two principles of unchanged stage logic and no deviation from the goal.

[0108] When the number of training epochs is greater than the first training epoch threshold (e.g., 50) and less than or equal to the second training epoch threshold (e.g., 150), the weight for accuracy loss in the low-dose region is 0.3, which can be automatically adjusted according to the actual situation. At this point, the model has mastered the basic feature associations and needs to move to the core stage of "accuracy optimization in the low-dose region" (the pain point is insufficient supervision in the low-dose region). Increasing the LDRI weight can strengthen the supervision signals of sub-items such as gradient consistency and voxel error, guiding the model to focus on correcting the prediction bias in the 5-20Gy region at the boundary of OARs. At the same time, the weight of 0.3 does not exceed the global loss ratio. =0.3), to avoid sacrificing overall accuracy; the weight value can be flexibly changed according to differences in the dataset, adjustments to the radiotherapy plan, and clinical needs, but it must meet the two principles of unchanged stage logic and no deviation from the goal.

[0109] When the number of training epochs exceeds the second training epoch threshold but is less than or equal to the cutoff training epoch threshold (e.g., 200), the weight for accuracy loss in the low-dose region is 0.25, which can be automatically adjusted according to the actual situation. This stage is the "fine-tuning stabilization period." Reducing the weight is to avoid the model's generalization ability decreasing due to overfitting to local features in the low-dose region (e.g., weakened adaptability to new patient data). At the same time, it is dynamically adjusted in conjunction with the validation set LDRI score (e.g., if there is no decrease for 5 consecutive epochs, the learning rate is decayed) to balance accuracy and stability. The weight value can be flexibly changed according to differences in the dataset, adjustments to the radiotherapy plan, and clinical needs, but it must meet the two principles of unchanged stage logic and no deviation from the goal.

[0110] The first 50 epochs: =0.2, prioritize the 3D-MSCB and LGAG modules to capture multi-scale features in the low-dose region, and avoid excessive LDRI causing the model to be unbalanced;

[0111] 51-150 epochs: =0.3, using the gradient signal of LDRI, guide the 3D-SAB module to focus on the gradient region of OARs boundary, while optimizing the reconstruction accuracy of EUCB upsampling for details in low-dose regions;

[0112] 151-200 epochs: =0.25, dynamically adjusted based on the validation set LDRI score (if the validation set LDRI score does not decrease for 5 consecutive epochs, the learning rate is triggered to decay to 1 / 10 of the original).

[0113] After training is complete, a standard-sized dose map of 128×128×128 is output, which can be directly imported into a clinical imaging diagnostic system without additional format conversion.

[0114] In this embodiment of the invention, the dual-scenario fusion uses LDRI (Low-Dose Area Precision Index) as the core link to connect the "model training scenario" and the "performance evaluation scenario," forming a technical closed loop of "targeted learning - precise verification." The first scenario is the LDRI-driven training scenario. Addressing the problem that traditional global loss is insufficient for supervising the low-dose area around OARs (Organs at Risk), LDRI is decomposed into three differentiable loss sub-items: voxel error, gradient consistency, and clinical threshold achievement rate. These are integrated into a three-layer combined loss function with dynamically adjusted weights. Combined with the 3D EMCAD module, the model is guided to focus on learning the numerical accuracy, gradient trend, and clinical compliance of the low-dose area, ensuring that the training direction is aligned with the protection needs of OARs.

[0115] The second scenario is an LDRI-guided assessment scenario, which abandons the traditional single global dose error index and uses the LDRI quantitative index corresponding to the training target as the core, combined with key clinical indicators such as the DVH curve, to accurately test the predictive effect in the low-dose area and avoid the disconnect between training and assessment.

[0116] like Figure 4 As shown, the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy provided in this embodiment of the invention can be implemented in software. The dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy includes the following software modules: data acquisition module 401 and dose prediction module 402.

[0117] The following is an introduction to the functions of each software module in the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy:

[0118] The data acquisition module 401 is used to acquire multi-source raw data of patients with head and neck cancer and form a multi-channel tensor, which includes clinical CT image data, anatomical structure mask of organ at risk area and planned target area localization mask.

[0119] The dose prediction module 402 is used to extract the anatomical features of the head and neck at different scales using a pre-trained 3D dose prediction model, fuse the anatomical features at different scales and upsample them, and convert the 3D dose into a single-channel prediction tensor output, thereby obtaining the 3D dose prediction result for head and neck cancer radiotherapy in patients with head and neck cancer.

[0120] The loss function for training the 3D dose prediction model consists of a weighted sum of four factors: global dose error, DVH index error, multi-scale Dice loss, and low-dose region accuracy loss. The low-dose region accuracy loss consists of three factors: low-dose region voxel error, low-dose gradient consistency, and clinical threshold achievement rate.

[0121] It should be noted that each module in the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy in this embodiment corresponds one-to-one with each step in the dual-scene fusion 3D dose prediction method for head and neck cancer radiotherapy in the above embodiment, and their specific implementation processes are the same, so they will not be repeated here.

[0122] The structure of the electronic device according to an embodiment of the present invention will be described in detail below. Figure 5 This is a schematic diagram of the composition structure of an electronic device provided in an embodiment of the present invention. It can be understood that... Figure 5 The diagram shows only an exemplary structure of the electronic device, not the entire structure. Some or all of the structures shown may be implemented as needed.

[0123] The electronic device provided in this embodiment of the invention includes: at least one processor 501, a memory 502, a user interface 503, and at least one network interface 504. The various components in the dual-scene fusion head and neck cancer radiotherapy 3D dose prediction system are coupled together via a bus system 505. It can be understood that the bus system 505 is used to realize the connection and communication between these components. In addition to a data bus, the bus system 505 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 5 The general designated all buses as Bus System 505.

[0124] The user interface 503 may include a monitor, keyboard, mouse, trackball, click wheel, buttons, touchpad, or touch screen.

[0125] It is understood that memory 502 can be volatile memory or non-volatile memory, or both. In this embodiment of the invention, memory 502 is capable of storing data to support the operation of the terminal. Examples of this data include any computer programs used to operate on the terminal, such as operating systems and applications. The operating system includes various system programs, such as framework layers, core library layers, driver layers, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications.

[0126] In some embodiments, the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy provided in this invention can be implemented using a combination of hardware and software. As an example, the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy provided in this invention can be a processor in the form of a hardware-decoded processor, programmed to execute the dual-scene fusion 3D dose prediction method for head and neck cancer radiotherapy provided in this invention. For example, the processor in the form of a hardware-decoded processor can employ one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0127] As an example, processor 501 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0128] As an example of the hardware implementation of the dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy provided in this embodiment of the invention, the device provided in this embodiment of the invention can be directly executed by a processor 501 in the form of a hardware decoding processor. For example, it can be executed by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components to implement the dual-scene fusion 3D dose prediction method for head and neck cancer radiotherapy provided in this embodiment of the invention.

[0129] The memory 502 in this embodiment of the invention is used to store various types of data to support the operation of a dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy, or to store data for execution. Figure 1 The program code for the method shown. Examples of this data include: any executable instructions for operation on a dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy, such as executable instructions that can be included in the executable instructions to implement the dual-scene fusion 3D dose prediction method for head and neck cancer radiotherapy according to embodiments of the present invention.

[0130] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including functions for executing... Figure 1 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the apparatus of this application.

[0131] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy, characterized in that, include: We acquire multi-source raw data from head and neck cancer patients and construct a multi-channel tensor, which includes clinical CT imaging data, anatomical structure masks of organ at risk areas, and planned target area localization masks. The 3D dose prediction model is pre-trained to extract multi-channel tensor anatomical features of the head and neck at different scales. The anatomical features at different scales are fused and upsampled to convert the 3D dose into a single-channel prediction tensor output, thus obtaining the 3D dose prediction results for head and neck cancer radiotherapy in patients with head and neck cancer. The loss function for training the 3D dose prediction model consists of a weighted sum of four factors: global dose error, DVH dose-volume histogram error, multi-scale Dice loss, and low-dose region accuracy loss. The low-dose region accuracy loss is composed of three sub-terms: low-dose region voxel error, low-dose gradient consistency, and clinical threshold achievement rate. The 3D dose prediction model includes an encoder, an extended skip connection block, and a decoder. The encoder is used to extract multi-channel tensors of anatomical features of the head and neck at different scales. The extended skip connection block and decoder consist of multi-scale convolutional blocks, channel attention blocks, and cascaded channel attention blocks and spatial attention blocks; The decoder is used to fuse anatomical features at different scales and upsample them to convert 3D dose into a single-channel prediction tensor output; In the decoder, a 3D macrokernel grouping attention gate is used to fuse anatomical features at different scales. The 3D macrokernel grouping attention gate is used to divide the encoder features and decoder upsampled features into three groups according to the organ at risk area-planned target area grouping rule: the planned target area exclusive group, the organ at risk area exclusive group, and the organ at risk area-planned target area boundary group.

2. The dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy as described in claim 1, characterized in that, The weight of accuracy loss in low-dose areas is dynamically adjusted based on the number of training rounds.

3. The dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy as described in claim 1, characterized in that, The low-dose region voxel error sub-item is: the average error between the dose prediction value and the actual dose value of the voxel by the 3D dose prediction model.

4. The dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy as described in claim 1, characterized in that, The low-dose gradient consistency sub-item is set as follows: : ； In the formula: P is the set of voxels in the gradient region defined by the boundary of the organ at risk. For 3D gradient operators, cos( ) represents the gradient direction cosine similarity; The dose value predicted by the 3D dose prediction model. This is the actual dose value; The total number of voxels contained in the voxel set P corresponding to the gradient region that defines the boundary of the organ at risk.

5. The dual-scene fusion method for predicting 3D dose in head and neck cancer radiotherapy as described in claim 1, characterized in that, The clinical threshold achievement rate sub-item is: : ； in, Number of organs at risk; The loss function; The first dose predicted by the 3D dose prediction model The percentage of the volume in the organ at risk of infection where the dose is ≤ the corresponding clinical threshold. The first under the true clinical gold standard The percentage of the volume in the organ at risk that is ≤ the corresponding clinical threshold.

6. A dual-scene fusion 3D dose prediction system for head and neck cancer radiotherapy, characterized in that, The method for predicting 3D dose for head and neck cancer radiotherapy based on dual-scene fusion as described in any one of claims 1-5 includes: The data acquisition module is used to acquire multi-source raw data from head and neck cancer patients and construct a multi-channel tensor, which includes clinical CT image data, anatomical structure mask of organ at risk area and planned target area localization mask. The dose prediction module is used to extract multi-channel tensor anatomical features of the head and neck at different scales using a pre-trained 3D dose prediction model. It then fuses the anatomical features at different scales and upsamples them to convert the 3D dose into a single-channel prediction tensor output, thus obtaining the 3D dose prediction results for head and neck cancer radiotherapy in patients with head and neck cancer. The loss function for training the 3D dose prediction model consists of a weighted sum of four factors: global dose error, DVH dose-volume histogram index error, multi-scale Dice loss, and low-dose region accuracy loss. The low-dose region accuracy loss consists of three factors: low-dose region voxel error, low-dose gradient consistency, and clinical threshold achievement rate.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the dual-scene fusion method for 3D dose prediction in head and neck cancer radiotherapy as described in any one of claims 1-5.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the dual-scene fusion method for 3D dose prediction in head and neck cancer radiotherapy as described in any one of claims 1-5.

Images