A lung function adaptive radiotherapy plan generation and update method based on a generative deep model and voxel-level damage optimization

Abstract

The application provides a lung function adaptive radiotherapy plan generation and update method based on a generative deep model and a voxel-level damage optimization algorithm. The method comprises the following steps: first, collecting missing multi-modal diagnosis and treatment images and lung perfusion function images of a patient; constructing an image generation network (MAFA-Net) with missing modal perception and adaptive attention fusion mechanism; generating a lung perfusion image of a new patient by using a positioning CT image and the network, and completing segmentation of organs, tumors and perfusion regions; based on the generated image, the CT image and the segmentation result, building a voxel-level damage optimization automatic planning algorithm (BVIO-Plan) guided by biophysical information, obtaining and executing an initial function protection radiotherapy plan; acquiring online and / or offline CT images during radiotherapy; iteratively updating the lung perfusion image by using the MAFA-Net, and evaluating the dose distribution deviation; and again using the BVIO-Plan algorithm to complete adaptive adjustment of the plan. The application can quickly and efficiently realize lung function adaptive radiotherapy and improve the lung function protection effect.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, radiotherapy planning optimization, and deep learning technology, and particularly to a method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization. Background Technology

[0002] Lung cancer is the most common malignant tumor in my country. In 2022, there were approximately 1.061 million new cases and 733,000 deaths from lung cancer in China, ranking first among malignant tumors in both incidence and mortality. Radiotherapy is one of the main treatments for lung cancer, and statistics show that about 70% of lung cancer patients require different levels of radiotherapy.

[0003] Currently, the clinical efficacy of radiotherapy for lung cancer is unsatisfactory, with a typical complication, radiation-induced lung injury (RILI), occurring in up to 20% of cases, and no effective clinical treatment exists. Studies have shown that increasing the radiotherapy dose for patients with locally advanced lung cancer can improve survival benefits, but the occurrence of RILI restricts the clinical application of high-dose radiotherapy for lung cancer. At present, methods to avoid RILI during lung cancer radiotherapy typically involve the physicist following... Figure 1 The illustrated procedure develops a radiotherapy plan and minimizes lung DVH indices during plan optimization to reduce the incidence of RILI as much as possible. However, Figure 1 The process of setting up the radiation field, optimizing the plan, and evaluating the plan is based on two assumptions: ① the function of normal lung tissue is uniformly distributed; ② normal lung tissue has the same radiation sensitivity. Figure 1 The process for developing inverse intensity-modulated radiotherapy (IMRT) plans for lung cancer. The black box indicates that the plan design assumes uniform lung function distribution and equal radiation sensitivity.

[0004] Adaptive radiotherapy, by tracking morphological changes in tumors and organs and establishing a dynamic feedback mechanism to adjust the radiotherapy plan accordingly, can significantly reduce the dose to normal lung tissue. In addition to changes in anatomical information, clinical studies suggest that patients' lung perfusion function undergoes temporal changes during radiotherapy. Lung function in radiotherapy planning mainly refers to the respiratory function of the lungs, which can be assessed based on ventilation and / or perfusion function, with perfusion assessment being the clinical gold standard. The internal environment of lung tissue is complex, and tumors and comorbidities can alter lung perfusion distribution and radiation sensitivity, leading to significant spatial heterogeneity in lung perfusion and injury risk (e.g., ...). Figure 2 (As shown). Existing clinical studies have confirmed that imaging-guided radiotherapy planning can direct radiation dose to low-function areas, effectively reducing the incidence of radiation-induced lung injury (RILI). Timely capture of lung perfusion changes during radiotherapy and use to guide adaptive adjustments to the functional protection plan can further reduce the incidence of RILI. However, the following technical problems urgently need to be solved in the clinical implementation of lung perfusion-guided adaptive radiotherapy plan generation and updating:

[0005] 1) Clinical practice cannot capture and track changes in lung perfusion at a high frequency: Nuclear medicine imaging methods identify lung function through contrast agents, including PET / CT using fluorine-18 labeling and magnetic resonance imaging (MRI) using hyperpolarized gas labeling. 99m SPECT imaging using Tc-labeled serum large particle polymers is an example. PET / CT ventilation imaging has a long scan time and is expensive (approximately 7000-8000 RMB per scan, not covered by medical insurance); MRI ventilation imaging requires specialized scanning equipment and a special supply of hyperpolarizing gas; SPECT imaging has low spatial and temporal resolution and involves the use of ionizing radiation. Due to the high cost and high radiation of nuclear medicine imaging, it is difficult to perform multiple functional imaging sessions on patients, making it impossible to detect changes in lung function between radiotherapy fractions in a timely manner.

[0006] 2) Lack of optimization algorithms for lung function-protective radiotherapy plans that minimize damage: Lung function-protective radiotherapy plans are highly complex and difficult to design, and there is a lack of relevant experience and reference standards. Manually developing these plans is time-consuming and labor-intensive, and may result in suboptimal plans. In addition, existing plan optimization models are based on the assumption that damage and dose are linearly related, which does not conform to the "S"-shaped correlation between actual radiation and damage. The "minimize dose" model cannot obtain a dose distribution that "minimizes damage". Figure 2 Heterogeneous distribution of lung perfusion function in a single patient, with gray intensity representing lung function intensity.

[0007] In recent years, deep learning technology has made significant progress in image segmentation, planning optimization, and cross-modal image conversion, and its application in radiotherapy is becoming increasingly widespread. The National Cancer Center / National Tumor Quality Control Center released the "Physical Practice Guidelines for the Application of Artificial Intelligence Synthetic Images in Radiotherapy (2025 Edition)" (NCC / T-RT 019-2025) in 2025, demonstrating the application prospects of AI-generated images in radiotherapy. Utilizing medical image processing technology to mine image features and generate lung function images has advantages such as low cost and less additional damage to patients. Compared to functional images such as PET and SPECT, image generation technology can mine lung function features from low-cost CT and online radiotherapy images (such as cone-beam CT, online CT, and online MRI), reducing additional radiation and medical costs for patients. It is suitable for multiple imaging sessions throughout the treatment course to track changes in lung function and guide the development of individualized functional adaptive radiotherapy plans. Related research has confirmed the feasibility of using deep learning to mine CT image features to generate lung perfusion images. However, existing methods rely on single-modal images to acquire limited lung tissue features, and the lack of multimodal image datasets and modal gaps leads to low accuracy and unstable performance of lung perfusion image generation networks. Furthermore, research on automated radiotherapy planning algorithms for lung function protection is very limited. Existing algorithms include field angle optimization based on average functional dose and prior knowledge-driven field flux optimization algorithms, as well as field angle optimization algorithms integrating dose scoring and meta-optimization field flux optimization algorithms. However, all of these optimization algorithms assume a linear correlation between dose and lung injury, which does not conform to the true "S"-shaped correlation, and the algorithms cannot directly minimize the probability of lung injury. Moreover, existing planning optimization algorithms do not consider the impact of individualized patient clinical characteristics on the incidence of radiation-induced lung injury, and therefore cannot provide optimal planning sequences for clinical practice.

[0008] This invention addresses the difficulties in acquiring perfusion information and the imperfections in planning optimization algorithms in existing adaptive radiotherapy (RTG) workflows for lung function. It proposes a "method for generating and updating RTG plans based on generative deep models and voxel-level damage optimization." This method constructs a lung perfusion image generation network suitable for scenarios lacking clinical modalities, enabling the capture and dynamic tracking of lung perfusion function. Combined with a biophysically guided voxel-level damage optimization automated radiotherapy planning method, it simultaneously considers the differences in biological effects between high- and low-dose zones and the individualized clinical characteristics of patients at the voxel level. This achieves synergistic optimization of tumor control and reduction of radiation-induced lung injury risk, thereby completing lung perfusion-guided adaptive radiotherapy for lung cancer. Summary of the Invention

[0009] This invention addresses the problem of adaptive radiotherapy planning for lung function in clinical lung cancer radiotherapy, proposing a method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization. The entire process of this method is as follows: Figure 3As shown, by combining the generative deep learning network MAFA-Net with the voxel-level damage optimization algorithm BVIO-Plan, the generation and sequence update of adaptive radiotherapy plans for lung perfusion are realized.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows: The present invention provides a method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization, the specific steps of which are as follows:

[0011] Step 1: Design an image generation network model (MAFA-Net) with missing modality awareness and adaptive attention fusion mechanism for robust lung perfusion generation under conditions of missing multimodal images;

[0012] Step 2: The MAFA-Net model described in Step 1 is trained using a two-stage training strategy to obtain a lung perfusion generation network that is suitable for real clinical data and robust to modality loss.

[0013] Step 3: Based on localized CT images and fused with pulmonary function radiomics biomarker feature maps (RFM) PCA As input, and optional input 18 F-FDG PET images, lung perfusion images obtained by training MAFA-Net in step 2 and registered with localization CT, and functional lung regions are determined based on the lung perfusion images (functional lung masks can be obtained by taking the first 40% to 60% of the perfusion value).

[0014] Step 4: Based on the lung perfusion image and functional lung region obtained in Step 3, construct a biophysical information-guided voxel-level damage optimization automatic planning algorithm (BVIO-Plan). Under the constraints of target area prescription dose and organ at risk limit, the algorithm establishes an objective function at the voxel level with nonlinear "dose-damage" as the core term, and incorporates the lung perfusion image and functional lung region weights, as well as the patient's individualized clinical characteristic index, as objective function terms.

[0015] Step 5: Using the field angle optimization index and weighted search strategy, perform the field angle optimization and field flux optimization of BVIO-Plan described in Step 4, and output the radiotherapy plan parameters.

[0016] Step 6: CT images obtained through offline repositioning or online postural correction. 18 F-FDG PET, as an optional auxiliary modality, generates perfusion images and functional lung regions through step 3, and steps 4 and 5 are repeated to obtain an adaptively updated radiotherapy planning sequence between fractions.

[0017] Furthermore, in step 1, an image generation network model MAFA-Net with missing modality awareness and adaptive attention fusion mechanism is designed for robust lung perfusion generation under conditions of missing multimodal images. The network structure is as follows: Figure 4 As shown. This model designs a Learnable Patchify module for missing modality images, introducing Token Drop and Token Filling operations to explicitly discard missing modality tokens and fill them with learnable tokens, enabling the network to maintain stable inference and training even in scenarios with missing modalities. An adaptive cross-modal attention fusion mechanism is introduced to achieve bidirectional modal interaction. When a modality is missing, the fusion weights automatically change to rely solely on the interaction of available modalities, ensuring improved efficiency in utilizing complementary information when modalities are complete and more stable network output when modalities are missing. This network uses localized CT images as fixed input and employs RFM. PCA In addition, optional 18 The F-FDG PET image is used as input; and when a PET modality is missing, the MAFA-Net generates a patch using the Learnable Patchify module. 18 The network uses alternative token features corresponding to the F-FDG PET modality to ensure stable extraction of key features even when the modality is missing. The fused token sequence is then restored to the image space using the Unpatchify reconstruction module, which robustly outputs lung perfusion images registered with the localization CT.

[0018] In step 1, a feature augmentation strategy is adopted for missing multimodal images: First, a Learnable Patchify module is designed to generate and supplement missing features using learnable parameters. 18 The alternative token sequences corresponding to the F-FDG PET modality are used instead of simply replacing them with zero values; secondly, a lung function radiomics feature map (RFM) is constructed. PCA As a supplement to the pseudo-functional modality image. The RFM PCA Construction method as follows Figure 5 As shown, the specific steps are as follows:

[0019] Specifically, the CT images and corresponding delineation files are read, and the images and binary masks are imported into memory and downsampled. The edges of the delineated structures are filled using mirror voxels. Radiomic features of lung tissue are extracted using a sliding window approach. Abnormal feature values ​​are detected and processed using the iForest algorithm. The extracted feature vectors are reconstructed into a three-dimensional radiomics feature matrix (RFM) and upsampled to the original CT image resolution. The types of pulmonary function radiomics biomarkers used in this step include: ① Histogram features: variance, skewness, kurtosis, median, entropy; ② Gray-level co-occurrence matrix features: joint energy, summation mean, total variance, entropy sum, correlation, autocorrelation, clustering trend, cluster shadow, cluster significance, information correlation; ③ High gray-level run dominance, gray-level non-uniformity, short run high gray-level dominance. Correlation analysis is performed between the resampled RFM and SPECT lung perfusion images, selecting the 3-5 RFM types with the highest Spearman correlation coefficients. Finally, principal component analysis was used to reduce the dimensionality of the RFM to a single RFM that characterizes the heterogeneity of lung perfusion function. PCA By employing sliding-window voxel-level radiomics feature extraction, iForest abnormal feature value detection and correction, and three-dimensional RFM reconstruction and upsampling, the precision and robustness of lung perfusion heterogeneity characterization are improved, and functional feature pseudomodalities that can be used as input for deep learning are generated to effectively supplement clinical multimodal imaging scenarios.

[0020] Step 2: A two-stage training strategy is adopted to address the characteristics of missing multimodal imaging data in clinical practice. The first stage employs a training strategy combining supervised and unsupervised learning, using paired CT scans, 18 F-FDG PET and SPECT lung perfusion image datasets and unlabeled images 18 Optimization is performed on F-FDG PET and CT images. The network learns key features and label data information from the labeled dataset through supervised training, and learns potential feature information from the unlabeled dataset through unsupervised training to enhance the model's adaptability to incomplete data. The loss function used in this stage is as follows:

[0021] (1)

[0022] The dataset That is, CT and RFM with complete modalities PCA , 18 F-FDG PET and SPECT images; dataset That is, CT, RFM PCA and 18 F-FDG PET image; MSE represents the mean squared error between the real image and the generated image. During training, the parameters of the teacher model... Set as student model parameters The exponential moving average (EMA) value; This provides stable pseudo-labels for training student models. .

[0023] The second stage uses all image data for supervised and unsupervised training of missing modalities to improve the model's performance in handling missing modalities. Its goal is to fully utilize information from different datasets, guiding the model to learn the features of missing modalities through supervised learning, while combining unsupervised learning methods to enhance the model's robustness and generalization ability. The loss function used is as follows:

[0024] (2)

[0025] The dataset That is, CT, RFM PCA and SPECT images; Here, represents the weight parameters, and MSE represents the mean squared error between the real and generated images. To improve the student model's performance... 18 The robustness of F-FDG PET modality loss is demonstrated by incorporating a Learnable Patchify (LP) operation into the loss function, enabling modality patching. Approximate feature substitution under missing conditions.

[0026] After the online training is completed, it can be used CT and RFM PCA ,and 18 F-FDG PET images generate lung perfusion images; or 18 When F-FDG PET images are missing, use Excavation CT and RFM PCA Image information is used to generate lung perfusion images. This step employs a "supervised-unsupervised" joint training strategy, which can fully utilize full-modal labeled images, missing multimodal labeled images, and unlabeled multimodal images to achieve collaborative learning of complete and missing modal scenes. At the same time, the network combines lung function CT image features and learnable block operations to overcome the problems of low model accuracy and unstable performance caused by single-modal lung tissue features, insufficient multimodal image data, and modality missing features in traditional lung function generation methods.

[0027] Step 3, based on the image generation network obtained in Step 2 and The system automatically adjusts the input modality based on the type of the patient's multimodal images to acquire localized CT images and corresponding RFM. PCA Feature maps are a required input. 18 When F-FDG PET is available, it is used as an optional input; 18When F-FDG PET is missing, it uses a missing modality completion mechanism to generate alternative input features instead of using zero values.

[0028] Step 4: The core optimization of the BVIO-Plan incorporates a nonlinear "dose-lung injury" relationship at the voxel level. This nonlinear relationship characterizes the nonlinear growth of lung injury risk with dose, thereby shifting the radiotherapy planning optimization from a "minimize dose" to a "minimize injury" model. The BVIO-Plan algorithm includes field angle optimization and field flux optimization. Details are as follows:

[0029] Optimization of firing field angle: A two-step strategy of firing field angle optimization index and weight search is adopted. First, based on the patient's physical (anatomical, spatial location) and biological information (lung function distribution, probability of complications in normal tissues, individualized clinical characteristics), and considering the impact of firing field angle on planning quality, an objective function for firing field angle optimization is constructed as follows:

[0030]

[0031]

[0032] in For shooting field a p The biophysical effect index of neutron field i; Indicates the set of organs at risk; k represents the voxel location index; V j V T V N V F V L These represent the sets of voxels representing organ j, target area, normal tissue, functionally weighted lung, and anatomical lung, respectively. Indicates field a p The dose of subfield i deposited at voxel k; P is the functional weighting coefficient of pulmonary somatosensor k. k The perfusion value is represented by r(⋅), which is a voxel-level injury risk function that can be calculated based on the existing NTCP model. ε is a clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model. j w N w F The weights are Π; Π represents the set of selected field angles {a1, a2}. 2, …, a N}, a p Ψ represents the p-th field of attack, and N represents the total number of selected fields of attack; Π The total biophysical effect index represents the set of angles Π; Indicates field a pThe total number of subfields; m represents the weight of the angle interval term in field selection; δ represents the regularization parameter to avoid singularities. By minimizing The optimal set of shooting angles is obtained.

[0033] Fire field flux optimization: Based on the above optimal fire field angle set, the patient's biological information (functional distribution, probability of complications in normal tissues, individualized clinical characteristics) and the physical information of CT images (anatomical and spatial location) are integrated to guide flux optimization. The objective function for optimization is as follows:

[0034]

[0035]

[0036] Where x represents the intensity distribution map of the radiation field; T represents the set of voxels in the target area; 𝑆 is the set of organs at risk, s∈𝑆 represents the set of voxels of a certain organ at risk; 𝐹 represents the functionally weighted set of lung voxels; 𝐿 represents the set of lung tissue voxels; D represents the total number of voxels; D represents the dose-response matrix; d represents the dose distribution map; , The point dose and functional weight at voxel i are represented by w; the weighting factor is represented by D. T D s and D F These represent the target reference dose, dose constraint limits for organs at risk, and function-weighted lung, respectively; H represents the Heaviside function; r(⋅) is the voxel-level injury risk function, which can be calculated based on the existing NTCP model; ε is the clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model; α and β represent the weights of the injury constraint terms; TCP represents the tumor control rate. For the set of planned quality indicators; Let represent the value of the t-th plan quality indicator. , These are the upper and lower thresholds for this indicator; P t and F t These represent the priority and score of the quantitative plan quality indicators, respectively; the objective function employs a two-layer optimization strategy: the outer layer minimizes p with w as the optimization variable. bps The optimal w is obtained and used to optimize the intensity distribution x of the inner field, ultimately yielding a lung function protection plan sequence with minimal damage.

[0037] Compared to existing technologies, this step has the following advantages: 1) Based on the nonlinear relationship between dose and organ damage, it takes into account the biological differences between high and low doses at the voxel level, achieving synergistic optimization of tumor coverage and minimizing radiation-induced lung injury, ensuring that the plan meets current clinical requirements while further reducing radiation damage to normal lung tissue; 2) It solves the problem that traditional lung function protection radiotherapy planning algorithms linearize the "dose-damage" relationship and are difficult to reflect the actual damage pattern, avoiding additional damage caused by excessive minimization of high and low dose areas; 3) It changes the shortcomings of traditional methods that only reduce the risk of lung injury through physical dose indicators, achieving synergistic optimization of dose distribution and tissue damage, which can further improve the quality of radiotherapy planning and the level of lung function protection.

[0038] Step 6: Make full use of the repositioning CT images acquired offline during the patient's radiotherapy and / or online position-guided CT scans, and incorporate them when available. 18 F-FDG PET images; based on the above images, update the patient's lung perfusion images, and repeat steps 4 and 5 to obtain a functionally adaptive radiotherapy planning sequence between fractions.

[0039] 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 method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization.

[0040] A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. The present invention proposes an adaptive radiotherapy method for lung perfusion function based on generative depth models and voxel-level damage optimization. The lung perfusion images are updated based on fractional CT images, and the plan is optimized at the voxel level in conjunction with the patient's biophysical information and individualized clinical characteristics to achieve lung function-guided adaptive radiotherapy.

[0043] 2. The lung perfusion function adaptive radiotherapy method proposed in this invention can automatically generate lung perfusion images and optimize the lung function protection radiotherapy plan based on the patient's missing multimodal images, reducing the dependence on repeated nuclear medicine functional imaging, and has the advantages of low cost and wide clinical applicability.

[0044] 3. The clinical application scenarios of MAFA-Net compatible with multimodal image deficiencies proposed in this invention, in... 18Even when modal images such as F-FDG PET are missing, high-quality lung perfusion images can still be generated through feature completion. While ensuring clinical applicability, it overcomes the problems of single-modal lung tissue features, insufficient multimodal image data, low model accuracy, and unstable performance caused by modal missing images.

[0045] 4. The MAFA-Net proposed in this invention fully utilizes full-modality labeled data, missing modality labeled data, and unlabeled data through supervised-unsupervised joint training, missing modality completion mechanism, and pulmonary function radiomics feature map supplementation, thereby reducing the network's dependence on large-scale complete functional image labeled data.

[0046] 5. The BVIO-Plan algorithm implemented in this invention introduces a nonlinear relationship between dose and organ damage at the voxel level, taking into account the biological differences between high- and low-dose regions; at the same time, it integrates individualized clinical characteristics of patients to achieve synergistic optimization of tumor coverage and minimization of radiation-induced lung injury, providing clinical patients with a radiotherapy plan that minimizes lung function protection.

[0047] 6. The BVIO-Plan algorithm implemented in this invention overcomes the limitation of traditional methods that only indirectly reduce the risk of lung injury through physical dose indicators, and realizes direct synergistic optimization of dose distribution and tissue damage probability, thereby improving the biological rationality and planning quality of lung function protection radiotherapy plans.

[0048] 7. This invention forms a standardized closed-loop process from image preprocessing and perfusion functional map generation to voxel-level damage optimization and adaptive radiotherapy planning updates. It is compatible with existing radiotherapy planning system processes, reduces repeated manual trials and subjective differences, reduces planning time, and improves the consistency and scalability of the plan. Attached Figure Description

[0049] Figure 1 The process of developing a reverse intensity-modulated radiotherapy (IMRT) plan for lung cancer.

[0050] Figure 2 Heterogeneous distribution of lung perfusion function in individual patients.

[0051] Figure 3 This invention proposes a generative deep model and voxel-level damage optimization method for generating and updating adaptive radiotherapy plans for lung function.

[0052] Figure 4 The proposed invention is an image generation network (MAFA-Net) with missing modality awareness and adaptive attention fusion mechanism.

[0053] Figure 5 Characterizing the lung perfusion process based on CT radiomics.

[0054] Figure 6 A voxel-level damage optimization automatic planning algorithm based on biophysical information (BVIO-Plan).

[0055] Figure 7 The process of optimizing the field flux of the BVIO-Plan algorithm. Figure 8 CT images, lung perfusion images, and radiomics features of different patients. Detailed Implementation

[0056] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0057] Example 1: This invention provides a method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization. The specific steps are as follows:

[0058] Step 1: Design an image generation network model (MAFA-Net) with missing modality awareness and adaptive attention fusion mechanism for robust lung perfusion generation under conditions of missing multimodal images;

[0059] Step 2: The MAFA-Net model described in Step 1 is trained using a two-stage training strategy to obtain a lung perfusion generation network that is suitable for real clinical data and robust to modality loss.

[0060] Step 3: Based on localized CT images and fused with pulmonary function radiomics biomarker feature maps (RFM) PCA As input, and optional input 18 F-FDG PET images, lung perfusion images obtained by training MAFA-Net in step 2 and registered with localization CT, and functional lung regions are determined based on the lung perfusion images (functional lung masks can be obtained by taking the first 40% to 60% of the perfusion value).

[0061] Step 4: Based on the lung perfusion image and functional lung region obtained in Step 3, construct a biophysical information-guided voxel-level damage optimization automatic planning algorithm (BVIO-Plan). Under the constraints of target area prescription dose and organ at risk limit, the algorithm establishes an objective function at the voxel level with nonlinear "dose-damage" as the core term, and incorporates the lung perfusion image and functional lung region weights, as well as the patient's individualized clinical characteristic index, as objective function terms.

[0062] Step 5: Using the field angle optimization index and weighted search strategy, perform the field angle optimization and field flux optimization of BVIO-Plan described in Step 4, and output the radiotherapy plan parameters.

[0063] Step 6: CT images obtained through offline repositioning or online postural correction. 18 F-FDG PET, as an optional auxiliary modality, generates perfusion images and functional lung regions through step 3, and steps 4 and 5 are repeated to obtain an adaptively updated radiotherapy planning sequence between fractions.

[0064] In step 1, an image generation network model, MAFA-Net, with missing modality awareness and adaptive attention fusion mechanism is designed for robust lung perfusion generation under conditions of missing multimodal images. The network structure is as follows: Figure 4 As shown. This model designs a Learnable Patchify module for missing modal images, introducing Token Drop and Token Filling operations to explicitly discard missing modal tokens and fill them in with learnable tokens, enabling the network to maintain stable inference and training even in scenarios with missing modalities. An adaptive cross-modal attention fusion mechanism is introduced to achieve bidirectional modal interaction. When a modality is missing, the fusion weights automatically change to rely solely on the interaction of available modalities, ensuring improved efficiency in utilizing complementary information when modalities are complete and more stable network output when modalities are missing. This network uses localized CT images as fixed input and employs RFM. PCA In addition, optional 18 The F-FDG PET image is used as input; and when a PET modality is missing, the MAFA-Net generates a patch using the Learnable Patchify module. 18 The network uses alternative token features corresponding to the F-FDG PET modality to ensure stable extraction of key features even when the modality is missing. The fused token sequence is then restored to the image space using the Unpatchify reconstruction module, which robustly outputs lung perfusion images registered with the localization CT.

[0065] In step 1, a feature augmentation strategy is adopted for missing multimodal images: First, a Learnable Patchify module is designed to generate and supplement missing features using learnable parameters. 18 The alternative token sequences corresponding to the F-FDG PET modality are used instead of simply replacing them with zero values; secondly, a lung function radiomics feature map (RFM) is constructed. PCA As a supplement to the pseudo-functional modality image. The RFM PCA Construction method as follows Figure 5 As shown, the specific steps are as follows: Figure 5Characterizing the lung perfusion process based on CT radiomics.

[0066] Specifically, the CT images and corresponding delineation files are read, and the images and binary masks are imported into memory and downsampled. The edges of the delineated structures are filled using mirror voxels. Radiomic features of lung tissue are extracted using a sliding window approach. Abnormal feature values ​​are detected and processed using the iForest algorithm. The extracted feature vectors are reconstructed into a 3D RFM and upsampled to the original CT image resolution. The types of pulmonary function radiomics biomarkers used in this step include: ① Histogram features: variance, skewness, kurtosis, median, entropy; ② Gray-level co-occurrence matrix features: joint energy, summation mean, total variance, entropy sum, correlation, autocorrelation, clustering trend, cluster shadow, cluster significance, information correlation; ③ High gray-level run dominance, gray-level non-uniformity, and short run high gray-level dominance. Correlation analysis is performed between the resampled RFM and SPECT lung perfusion images, selecting the 3-5 RFM types with the highest Spearman correlation coefficients. Finally, principal component analysis is used to reduce the dimensionality of the RFM into a single RFM characterizing the heterogeneity of lung perfusion function. PCA By employing sliding-window voxel-level radiomics feature extraction, iForest abnormal feature value detection and correction, and three-dimensional RFM reconstruction and upsampling, the precision and robustness of lung perfusion heterogeneity characterization are improved, and functional feature pseudomodalities that can be used as input for deep learning are generated to effectively supplement clinical multimodal imaging scenarios.

[0067] Step 2: A two-stage training strategy is adopted to address the characteristics of missing multimodal imaging data in clinical practice. The first stage employs a training strategy combining supervised and unsupervised learning, using paired CT scans, 18 F-FDG PET and SPECT lung perfusion image datasets and unlabeled images 18 Optimization is performed on F-FDG PET and CT images. The network learns key features and label data information from the labeled dataset through supervised training, and learns potential feature information from the unlabeled dataset through unsupervised training to enhance the model's adaptability to incomplete data. The loss function used in this stage is as follows:

[0068] (1)

[0069] The dataset That is, CT and RFM with complete modalities PCA , 18 F-FDG PET and SPECT images; dataset That is, CT, RFM PCA and 18F-FDG PET image; MSE represents the mean squared error between the real image and the generated image. During training, the parameters of the teacher model... Set as student model parameters The exponential moving average (EMA) value; This provides stable pseudo-labels for training student models. .

[0070] The second stage uses all image data for supervised and unsupervised training of missing modalities to improve the model's performance in handling missing modalities. Its goal is to fully utilize information from different datasets, guiding the model to learn the features of missing modalities through supervised learning, while combining unsupervised learning methods to enhance the model's robustness and generalization ability. The loss function used is as follows:

[0071] (2)

[0072] The dataset That is, CT, RFM PCA and SPECT images; Here, represents the weight parameters, and MSE represents the mean squared error between the real and generated images. To improve the student model's performance... 18 The robustness of F-FDG PET modality loss is demonstrated by incorporating a Learnable Patchify (LP) operation into the loss function, enabling modality patching. Approximate feature substitution under missing conditions.

[0073] After the online training is completed, it can be used CT and RFM PCA ,and 18 F-FDG PET images generate lung perfusion images; or 18 When F-FDG PET images are missing, use Excavation CT and RFM PCA Image information is used to generate lung perfusion images.

[0074] Step 3, based on the image generation network obtained in Step 2 and The system automatically adjusts the input modality based on the type of the patient's multimodal images to acquire localized CT images and corresponding RFM. PCA Feature maps are a required input. 18 When F-FDG PET is available, it is used as an optional input; 18 When F-FDG PET is missing, it uses a missing modality completion mechanism to generate alternative input features instead of using zero values.

[0075] Step 4: The core optimization of the BVIO-Plan incorporates a nonlinear "dose-lung injury" relationship at the voxel level. This nonlinear relationship characterizes the nonlinear growth of lung injury risk with dose, thereby shifting the radiotherapy planning optimization from a "minimize dose" to a "minimize injury" model. The BVIO-Plan algorithm flow is as follows: Figure 6 As shown, the content includes optimization of the field angle and optimization of the field throughput. Figure 6 A voxel-level damage optimization automatic planning algorithm based on biophysical information (BVIO-Plan).

[0076] Specifically as follows:

[0077] Optimization of firing field angle: A two-step strategy of firing field angle optimization index and weight search is adopted. First, based on the patient's physical (anatomical, spatial location) and biological information (lung function distribution, probability of complications in normal tissues, individualized clinical characteristics), and considering the impact of firing field angle on planning quality, an objective function for firing field angle optimization is constructed as follows:

[0078]

[0079]

[0080] in For shooting field a p The biophysical effect index of neutron field i; Indicates the set of organs at risk; k represents the voxel location index; V j V T V N V F V L These represent the sets of voxels representing organ j, target area, normal tissue, functionally weighted lung, and anatomical lung, respectively. Indicates field a p The dose of subfield i deposited at voxel k; P is the functional weighting coefficient of pulmonary somatosensor k. k The perfusion value is represented by r(⋅), which is a voxel-level injury risk function that can be calculated based on the existing NTCP model. ε is a clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model. j w N w F The weights are Π; Π represents the set of selected field angles {a1, a2}. 2, …, a N}, a p Ψ represents the p-th field of attack, and N represents the total number of selected fields of attack; Π The total biophysical effect index represents the set of angles Π; Indicates field a p The total number of subfields; m represents the weight of the angle interval term in field selection; δ represents the regularization parameter to avoid singularities. By minimizing The optimal set of shooting angles is obtained.

[0081] Fire field flux optimization: Based on the above optimal fire field angle set, the patient's biological information (functional distribution, probability of complications in normal tissues, individualized clinical characteristics) and the physical information of CT images (anatomical structure, spatial location) are integrated to guide flux optimization. Specific optimization steps are as follows: Figure 7 As shown, its objective function for optimization is as follows:

[0082]

[0083]

[0084] Where x represents the intensity distribution map of the radiation field; T represents the set of voxels in the target area; 𝑆 is the set of organs at risk, s∈𝑆 represents the set of voxels of a certain organ at risk; 𝐹 represents the functionally weighted set of lung voxels; 𝐿 represents the set of lung tissue voxels; D represents the total number of voxels; D represents the dose-response matrix; d represents the dose distribution map; , The point dose and functional weight at voxel i are represented by w; the weighting factor is represented by D. T D s and D F These represent the target reference dose, dose constraint limits for organs at risk, and function-weighted lung, respectively; H represents the Heaviside function; r(⋅) is the voxel-level injury risk function, which can be calculated based on the existing NTCP model; ε is the clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model; α and β represent the weights of the injury constraint terms; TCP represents the tumor control rate. For the set of planned quality indicators; Let represent the value of the t-th plan quality indicator. , These are the upper and lower thresholds for this indicator; P t and F t These represent the priority and score of the quantitative plan quality indicators, respectively; the objective function employs a two-layer optimization strategy: the outer layer minimizes p with w as the optimization variable. bps The optimal w is obtained and used to optimize the intensity distribution x of the inner field, ultimately yielding a lung function protection plan sequence with minimal damage. Figure 7 The process of optimizing the field flux of the BVIO-Plan algorithm.

[0085] Step 6: Make full use of the repositioning CT images acquired offline during the patient's radiotherapy and / or online position-guided CT scans, and incorporate them when available. 18 F-FDG PET images; based on the above images, update the patient's lung perfusion images, and repeat steps 4 and 5 to obtain a functionally adaptive radiotherapy planning sequence between fractions.

[0086] Example 2: Based on the actual situation of lung cancer patients during radiotherapy, the "adaptive radiotherapy planning generation and updating method based on generative depth model and voxel-level damage optimization" proposed in this invention was applied. Figure 3 This invention proposes a generative deep model and voxel-level damage optimization method for generating and updating adaptive radiotherapy plans for lung function.

[0087] The process for generating and updating adaptive radiotherapy plans for lung function is as follows:

[0088] Step 1, Patient Simulation Positioning. Based on the patient's treatment position, a radiotherapy-specific thermoplastic mold is used for fixation, followed by a large-aperture CT scan to obtain the patient's positioning CT image.

[0089] Step 2: Modeling the lung perfusion image generation network MAFA-Net based on generative deep learning.

[0090] Step 2.1, missing multimodal image collection. Collected localization and diagnostic images include, but are not limited to, localization CT scans. 18 F-FDG PET, SPECT perfusion images, etc., allow for the presence of 18 F-FDG PET or SPECT modality missing.

[0091] Step 2.2, Image Preprocessing

[0092] Specifically, a threshold segmentation method is used to determine the lung tissue region in the CT image, and the trachea and bronchi within the segmented region are manually removed by the radiologist. Then, a registration method based on global-local information is used to achieve accurate registration of multimodal images: ① Rigid registration is performed based on global lung tissue information from CT and SPECT perfusion images, and the registration results are used for calibration. 18 F-FDGPET and SPECT perfusion images; ② Fine-tuning the global rigid registration results based on local information (left or right lung) to further improve registration accuracy. The registered multimodal images are resampled to the same resolution to obtain the spatial correspondence of voxel points between different modal images. Then, all modal images are normalized, and the SPECT images are subjected to three-dimensional Gaussian filtering with a kernel size of [7 77] to obtain spatially aligned, scale-consistent multimodal volume data suitable for network input. Three-dimensional Gaussian filtering of SPECT images can suppress noise and artifacts, thereby improving the stability of subsequent network training and inference.

[0093] Step 2.3, lung perfusion characterization based on pulmonary function radiomics biomarkers.

[0094] Specifically, the CT images and corresponding delineation files are read, and the images and binary masks are imported into memory and downsampled. The edges of the delineated structures are filled using mirror voxels. Radiomic features of lung tissue are extracted using a sliding window approach. Abnormal feature values ​​are detected and processed using the iForest algorithm. The extracted feature vectors are reconstructed into a 3D RFM and upsampled to the original CT image resolution. The types of pulmonary function radiomics biomarkers used in this step include: ① Histogram features: variance, skewness, kurtosis, median, entropy; ② Gray-level co-occurrence matrix features: joint energy, summation mean, total variance, entropy sum, correlation, autocorrelation, clustering trend, cluster shadow, cluster significance, information correlation; ③ High gray-level run dominance, gray-level non-uniformity, and short run high gray-level dominance. Correlation analysis is performed between the resampled RFM and SPECT lung perfusion images, selecting the 3-5 RFM types with the highest Spearman correlation coefficients. Finally, principal component analysis is used to reduce the dimensionality of the RFM into a single RFM characterizing the heterogeneity of lung perfusion function. PCA By using voxel-level radiomics feature extraction via sliding windows, detection and correction of anomalous feature values ​​in iForest, and 3D RFM reconstruction and upsampling, the precision and robustness of lung perfusion heterogeneity characterization are improved, and functional feature pseudomodalities that can be used as inputs for deep learning are generated to effectively supplement clinical multimodal imaging scenarios.

[0095] Step 2.4 Construction of the MAFA-Net Deep Learning Network

[0096] Specifically, the MAFA-Net structure is as follows: Figure 4 As shown, by integrating missing multimodal images and RFM PCA Lung function characteristics were identified and mapped to SPECT lung perfusion images. A two-stage training strategy was employed to address the lack of clinical multimodal imaging data. The first stage combined supervised and unsupervised learning, using paired CT scans... 18 F-FDG PET and SPECT lung perfusion image datasets and unlabeled CT, 18 The network is optimized using F-FDG PET images. Supervised training learns key features and label data information from the labeled dataset, while unsupervised training learns latent feature information from the unlabeled dataset to enhance the model's adaptability to incomplete data. The loss function used in this stage is as follows:

[0097] (1)

[0098] The dataset That is, CT and RFM with complete modalities PCA , 18 F-FDG PET and SPECT images; dataset That is, CT, RFM PCA and 18 F-FDG PET image; MSE represents the mean squared error between the real image and the generated image. During training, the parameters of the teacher model... Set as student model parameters The exponential moving average (EMA) value provides stable pseudo-labels for training the student model. .

[0099] The second stage uses all image data for supervised and unsupervised training of missing modalities to improve the model's performance in handling missing modalities. Its goal is to fully utilize information from different datasets, guiding the model to learn the features of missing modalities through supervised learning, while combining unsupervised learning methods to enhance the model's robustness and generalization ability. The loss function used is as follows:

[0100] (2)

[0101] The dataset That is, CT, RFM PCA and SPECT images; Here, represents the weight parameters, and MSE represents the mean squared error between the real and generated images. To improve the student model's performance... 18 The robustness of F-FDG PET modality loss is demonstrated by incorporating a Learnable Patchify (LP) operation into the loss function, enabling modality patching. Approximate feature substitution under missing conditions.

[0102] After the online training is completed, it can be used CT and RFM PCA ,and 18 F-FDG PET images generate lung perfusion images; or 18 When F-FDG PET images are missing, use Excavation CT and RFM PCA Image information is used to generate lung perfusion images.

[0103] Step 2.5 Generation of lung perfusion images

[0104] Specifically, the same image preprocessing method as in steps 2.2-2.3 was used to extract lung perfusion RFM. PCA Then the CT images were combined with RFM. PCA and optional input 18F-FDG PET images to post-training or The network was used to obtain lung perfusion images.

[0105] Step 3, Multimodal Image Segmentation

[0106] Specifically, the treatment planning system (TPS) is used to delineate the lung tumor and normal organs on the localization CT images, and combined with the generated perfusion images, the threshold method is used to obtain the functional area (the functional area is defined by a threshold of 40% to 60%).

[0107] Step 4: Construction of the Biophysics-Guided Voxel-Level Damage Optimization Automated Planning (BVIO-Plan) Algorithm

[0108] Specifically, radiobiology and clinical studies have confirmed that the physical dose and organ damage in radiotherapy patients are not directly linearly correlated. Therefore, traditional radiotherapy planning optimization algorithms that focus on minimizing dose cannot minimize lung injury and may even cause additional radiation damage to patients. Furthermore, traditional planning optimization algorithms do not consider the impact of individualized clinical characteristics on the incidence of radiation-induced lung injury, making it difficult to achieve optimal protection. The BVIO-Plan algorithm integrates the nonlinear relationship between dose and lung injury at the voxel level, taking into account the differences in damage effects between high and low doses, and incorporates the impact of individualized clinical characteristics on the incidence of lung injury, achieving synergistic optimization of tumor control rate and minimization of lung function impairment.

[0109] Step 4.1, Optimize the shooting angle

[0110] A two-step strategy of optimization index and weighted search for the optimal field angle is adopted. First, based on the patient's physical (anatomical, spatial location) and biological information (lung function distribution, probability of complications in normal tissues, individualized clinical characteristics), and considering the impact of the field angle on the planning quality, an objective function for optimizing the field angle is constructed as follows:

[0111]

[0112]

[0113] in For shooting field a p The biophysical effect index of neutron field i; Indicates the set of organs at risk; k represents the voxel location index; V j V T V N V F V LThese represent the sets of voxels representing organ j, target area, normal tissue, functionally weighted lung, and anatomical lung, respectively. Indicates field a p The dose of subfield i deposited at voxel k; P is the functional weighting coefficient of pulmonary somatosensor k. k The perfusion value is represented by r(⋅), which is a voxel-level injury risk function that can be calculated based on the existing NTCP model. ε is a clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model. j w N w F The weights are Π; Π represents the set of selected field angles {a1, a2}. 2, …, a N}, a p Ψ represents the p-th field of attack, and N represents the total number of selected fields of attack; Π The total biophysical effect index represents the set of angles Π; Indicates field a p The total number of subfields; m represents the weight of the angle interval term in field selection; δ represents the regularization parameter to avoid singularities. By minimizing The optimal set of shooting angles is obtained.

[0114] Step 4.2 Field Flux Optimization

[0115] The steps for optimizing field flux are as follows: Figure 7 As shown, based on the above optimal set of field angles, throughput optimization is guided by integrating the patient's biological information (functional distribution, probability of complications in normal tissues, and individualized clinical characteristics) and the physical information (anatomical and spatial location) of CT images. The objective function for optimization is as follows:

[0116]

[0117]

[0118] Where x represents the intensity distribution map of the radiation field; T represents the set of voxels in the target area; 𝑆 is the set of organs at risk, s∈𝑆 represents the set of voxels of a certain organ at risk; 𝐹 represents the functionally weighted set of lung voxels; 𝐿 represents the set of lung tissue voxels; D represents the total number of voxels; D represents the dose-response matrix; d represents the dose distribution map; , The point dose and functional weight at voxel i are represented by w; the weighting factor is represented by D. T D s and D FThese represent the target reference dose, dose constraint limits for organs at risk, and function-weighted lung, respectively; H represents the Heaviside function; r(⋅) is the voxel-level injury risk function, which can be calculated based on the existing NTCP model; ε is the clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model; α and β represent the weights of the injury constraint terms; TCP represents the tumor control rate. For the set of planned quality indicators; Let represent the value of the t-th plan quality indicator. , These are the upper and lower thresholds for this indicator; P t and F t These represent the priority and score of the quantitative plan quality indicators, respectively; the objective function employs a two-layer optimization strategy: the outer layer minimizes p with w as the optimization variable. bps The optimal w is obtained and used to optimize the intensity distribution x of the inner field, ultimately yielding a lung function protection plan sequence with minimal damage.

[0119] Step 5, Planning, Assessment, and Treatment

[0120] The feasibility of the treatment plan is assessed by the radiation physicist based on the accelerator equipment properties. Then, the clinician evaluates whether the target dose and the dose to organs at risk meet the requirements. At the same time, the degree of protection of functional lungs is assessed by combining the generated perfusion images to obtain a clinically acceptable radiotherapy plan for lung function protection. Finally, the radiation technician administers the irradiation to the patient.

[0121] Step 6, Image acquisition during treatment

[0122] Specifically, image acquisition during treatment includes, but is not limited to, images acquired by accelerator-borne fan-beam CT, repositioning images from large-aperture CT scans, and assessments of tumor status. 18 F-FDG PET images, etc.

[0123] Step 7, Adaptive Plan Modification

[0124] Specifically, based on the image modality obtained in step 6, steps 2-5 are repeated to obtain a functionally adaptive radiotherapy plan. After approval by the clinician, irradiation is performed according to the new plan.

[0125] Step 8: Repeat step 7 until the patient's treatment is completed.

[0126] Compared to existing technologies, this invention utilizes MAFA-Net to construct a robust correlation model between lung function and imaging features under conditions of insufficient or missing multimodal imaging data. It also addresses the problem that traditional lung function-protective radiotherapy plans neglect the differences in biological effects between high and low doses through the BVIO-Plan algorithm. Furthermore, it incorporates the impact of individualized clinical characteristics of patients on the incidence of radiation-induced lung injury, providing a lung function-protective radiotherapy plan sequence with minimal damage for clinical lung cancer. This provides an efficient and easy-to-implement method for adaptive radiotherapy of lung function in lung cancer patients.

[0127] Effectiveness evaluation:

[0128] This invention proposes a method for adaptive radiotherapy planning generation and updating based on generative deep models and voxel-level damage optimization algorithms. Its core comprises two parts: an image generation network model (MAFA-Net) based on missing modality perception and adaptive attention fusion mechanisms, and a voxel-level damage optimization automatic planning algorithm (BVIO-Plan) guided by biophysical information. MAFA-Net uses pulmonary function radiomics biomarkers as pseudo-functional images for lung perfusion generation. The correlation coefficient r between RFM and lung perfusion obtained after feature extraction and outlier detection and processing proposed in this patent is between 0.326 and 0.428. Figure 8 As shown, this provides supplementary information for the missing modal images. Figure 8 CT images, lung perfusion images, and radiomics feature maps of different patients. RFM represents the pulmonary function radiomics feature map, and r represents the Spearman correlation coefficient between RFM and perfusion images.

[0129] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or alternatives made based on the above technical solutions shall fall within the scope of protection of the present invention.

Claims

1. A method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization, characterized in that, The specific steps are as follows: Step 1: Design an image generation network model with missing modality awareness and adaptive attention fusion mechanism (Missing-modality Aware Adaptive Fusion-Attention Image Generation Network, MAFA-Net). Step 2: The MAFA-Net model from Step 1 is trained using a two-stage training strategy to obtain a lung perfusion generation network that is suitable for real clinical data and robust to modality loss. Step 3, based on localized CT images and fused with pulmonary function radiomics biomarker feature maps (RFM) PCA ( ) as input, and input is optional. 18 F-FDG PET images, lung perfusion images registered with localization CT after training with MAFA-Net in step 2, and functional lung regions determined based on the lung perfusion images are high-functioning areas within the lungs, determined according to perfusion thresholds. Step 4: Based on the lung perfusion images and functional lung regions obtained in Step 3, a biophysical information-guided voxel-level damage optimization automatic planning algorithm (BVIO-Plan) is constructed. This algorithm, under the constraints of target area prescription dose and organ-at-risk limits, establishes an objective function at the voxel level with a nonlinear "dose-lung injury" as its core term, incorporating lung perfusion map and functional lung region weights, as well as the patient's individualized clinical characteristic index, as objective function terms. Step 5: Using the field angle optimization index and weighted search strategy, perform field angle and field flux optimization for the BVIO-Plan in Step 4, and output the radiotherapy planning sequence that minimizes damage. Step 6: CT images obtained through offline repositioning or online postural correction. 18 F-FDG PET, as an optional auxiliary modality, generates perfusion images and functional lung regions through step 3, and steps 4 and 5 are repeated to obtain an adaptively updated radiotherapy planning sequence between fractions.

2. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, In step 1, a lung perfusion generation network, MAFA-Net, is designed for clinically missing multimodal images. This model incorporates a Learnable Patchify module for missing modal images, introducing Token Drop and Token Filling operations to explicitly discard missing modal tokens and fill them with learnable tokens, ensuring stable inference and training even in missing modal scenarios. An adaptive cross-modal attention fusion mechanism is introduced to achieve bidirectional modal interaction. When a modality is missing, the fusion weights automatically change to rely solely on the available modalities, ensuring improved utilization of complementary information when modalities are complete and more stable network output when modalities are missing. This network uses localized CT images as fixed input and employs RFM (Reference-Based Motion)... PCA To supplement, with 18 F-FDG PET images are used as input; and when PET modalities are missing, MAFA-Net generates corresponding images using the Learnable Patchify module. 18 The network uses alternative token features corresponding to the F-FDG PET modality to ensure stable extraction of key features even when the modality is missing. The fused token sequence is then restored to the image space using the Unpatchify reconstruction module, which robustly outputs lung perfusion images registered with the localization CT.

3. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, In step 1, a LearnablePatchify module is designed for missing multimodal images, including Token Drop and Token Filling, to generate and display missing data using learnable parameters. 18 The alternative token sequence corresponding to the F-FDGPET modality.

4. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, Step 2 employs a two-stage training strategy tailored to the characteristics of clinically missing multimodal imaging data. The first phase employs a training strategy that combines supervised and unsupervised learning, using paired training. 18 F-FDG PET, CT, and SPECT lung perfusion image datasets and unlabeled images 18 Optimization is performed on F-FDG PET and CT images. The network learns key features and label data information from the labeled dataset through supervised training, and learns potential feature information from the unlabeled dataset through unsupervised training to enhance the model's adaptability to incomplete data. The loss function used in this stage is as follows: (1) The dataset That is, CT and RFM with complete modalities PCA , 18 F-FDG PET and SPECT images; dataset That is, CT, RFM PCA and 18 F-FDG PET image; MSE represents the mean squared error between the real image and the generated image, and the parameters of the teacher model during training. Set as student model parameters The exponential moving average (EMA) value provides stable pseudo-labels for training the student model. , The second stage uses all image data for supervised and unsupervised training of missing modalities to improve the model's performance in handling missing modalities. Its goal is to fully utilize information from different datasets, guiding the model to learn the features of missing modalities through supervised learning, while combining unsupervised learning methods to enhance the model's robustness and generalization ability. The loss function used is as follows: (2) The dataset That is, CT, RFM PCA and SPECT images; The weighting parameter for the unsupervised consistency loss balances the importance of supervised reconstruction and unsupervised constraint terms, and its value is set based on the validation set performance; MSE represents the mean squared error between the real and generated images, used to improve the student model's performance. 18 The robustness of F-FDG PET modality loss is demonstrated by incorporating a Learnable Patchify (LP) operation into the loss function, enabling modality patching. Approximate feature substitution under missing conditions After network training is completed, use CT and RFM PCA ,and 18 F-FDG PET images generate lung perfusion images; or 18 When F-FDG PET images are missing, use Excavation CT and RFM PCA Image information is used to generate lung perfusion images. Indicates missing 18 Learnable alternative tokens for F-FDG PET images.

5. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, Step 3: Automatically adjust the input modality based on the type of the patient's multimodal images to acquire localization CT images and corresponding RFM. PCA Feature maps are a required input. 18 When F-FDG PET is available, it is used as an optional input; 18 When F-FDG PET is missing, a missing modality completion mechanism is used to generate alternative input features.

6. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, In step 4, the core optimization of the BVIO-Plan algorithm incorporates the nonlinear relationship between "dose-lung injury" at the voxel level. This nonlinear relationship is used to characterize the nonlinear growth of lung injury risk with dose, so as to realize the transformation of radiotherapy planning optimization from "minimizing dose" to "minimizing damage".

7. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, In step 4, a two-step strategy of field angle optimization index and weight search is adopted. First, based on the patient's physical and biological information and considering the impact of the field angle on the planning quality, an objective function for angle optimization is constructed. The objective function is as follows: in For shooting field a p The biophysical effect index of neutron field i; Indicates the set of organs at risk; k represents the voxel location index; V j V T V N V F V L These represent the sets of voxels representing organ j, target area, normal tissue, functionally weighted lung, and anatomical lung, respectively. Functionally weighted lung is the set of lung tissues that has been weighted according to perfusion values. Indicates field a p The dose of subfield i deposited at voxel k; P is the functional weighting coefficient of pulmonary somatosensor k. k ε is the perfusion value; r(⋅) is the voxel-level injury risk function; ε is the clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the NTCP model parameters; w j w N w F The weights are Π; Π represents the set of selected field angles {a1, a2}. 2, …, a N }, a p Ψ represents the p-th field of attack, and N represents the total number of selected fields of attack; Π The total biophysical effect index represents the set of angles Π; Indicates field a p The total number of subfields; m represents the weight of the angle interval term in the field selection; δ represents the regularization parameter to avoid singularities, which is minimized by... Obtain the optimal set of shooting angles. Then, based on the optimal set of field angles, the patient's biological information and the physical information of the CT images are integrated to guide throughput optimization. The objective function for optimization is as follows: Where x represents the intensity distribution map of the radiation field; T represents the set of voxels in the target area; 𝑆 is the set of organs at risk, s∈𝑆 represents the set of voxels of a certain organ at risk; 𝐹 represents the functionally weighted set of lung voxels; 𝐿 represents the set of lung tissue voxels; D represents the total number of voxels; D represents the dose-response matrix; d represents the dose distribution map; , represents the point dose and functional weight at voxel i; w represents the weighting factor. D T D s and D F These represent the target reference dose, dose constraint limits for organs at risk, and function-weighted lung, respectively; H represents the Heaviside function; r(⋅) is the voxel-level injury risk function; ε is the clinical characteristic correction factor, obtained by fitting retrospective clinical data parameters, used to correct the parameters of the NTCP model; α and β represent the weights of the injury constraint terms. TCP represents the tumor control rate; For the set of planned quality indicators; Let represent the value of the t-th plan quality indicator. , These are the upper and lower thresholds for this indicator; P t and F t These represent the priority and score of the quantitative plan quality indicators, respectively; the objective function employs a two-layer optimization strategy: the outer layer minimizes p with w as the optimization variable. bps The optimal w is obtained and used to optimize the intensity distribution x of the inner field, ultimately yielding a lung function protection plan sequence with minimal damage.

8. The method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization according to claim 1, characterized in that, In step 6, fully utilize the repositioning CT images acquired offline during the patient's radiotherapy process and / or online position-guided CT, and incorporate them when available. 18 F-FDG PET images; based on the above images, update the patient's lung perfusion images, and repeat steps 4 and 5 to obtain an adaptive radiotherapy planning sequence with inter-fractional functional updates.

9. 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 method for generating and updating adaptive radiotherapy plans for lung function based on generative depth models and voxel-level damage optimization as described in any one of claims 1 to 7. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When executed by a processor, the computer instructions implement the method for generating and updating adaptive radiotherapy plans for lung function based on generative deep models and voxel-level damage optimization as described in any one of claims 1 to 7.

Images