Generation method and system based on lung DR image to CT breathing motion image

Abstract

The invention relates to the technical field of image generation, in particular to a method and a system for generating a respiratory motion image from a lung DR image to a CT (Computed Tomography), the method comprises the following steps: firstly, acquiring DR data and CT data of a plurality of patients, preprocessing to form a generalization data set, training a generative adversarial network, and obtaining a generalization generative adversarial network; then, acquiring a specific data set and training the generalization generative adversarial network by utilizing multi-stage CT data and DR data of the target patient in a historical respiratory cycle before surgical intervention, thereby obtaining the specific generative adversarial network suitable for the target patient, greatly reducing generation deviation caused by individual difference, and improving the accuracy of the target patient. The matching degree of the generated image and the real physiological state of the patient is improved; and finally, in the interventional operation and radiotherapy process, the real-time 4D CT data of the patient can be obtained only by acquiring the DR data of the lung of the target patient in real time, so that focus positioning and targeted therapy are more timely and accurate.

Description

Technical Field

[0001] This invention relates to the field of image generation technology, specifically to a method and system for generating CT respiratory motion images from lung DR images. Background Technology

[0002] Computed tomography (CT) is a medical imaging technique that uses X-rays and computer processing to generate high-resolution images of human tissue structures. Compared to digital radiography (DR, also known as X-ray film), CT offers higher contrast and clarity, clearly displaying tissue structures and lesions in the form of three-dimensional voxel maps. CT is particularly effective at visualizing bone and calcified tissue. Therefore, it is commonly used to diagnose fractures, bone tumors, and vascular lesions. Furthermore, low-dose CT remains the gold standard for early lung cancer screening.

[0003] 4D CT, on the other hand, involves taking multiple CT scans over a short period, revealing changes in human tissue and structures caused by respiratory movements. This feature of 4D CT supports clinicians in surgical planning, radiotherapy design, and precise lesion localization. Despite these advantages, CT also has significant limitations. The ionizing radiation dose generated during a CT scan is far higher than that of digital X-ray imaging (tens to hundreds of times), increasing the risk of cancer. The scanning process itself is also time-consuming, limiting its real-time application in surgical guidance and emergency situations. Therefore, developing methods to acquire CT images faster with lower radiation doses has become a major focus of current research. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for generating respiratory motion images from lung DR images to CT images.

[0005] The technical solution of this invention is as follows: A method for generating respiratory motion images from lung DR images to CT images includes the following operations: S1. Obtain DR and CT data of lungs from multiple patients, preprocess them to form a generalized dataset, train the generative adversarial network, and obtain the generalized generative adversarial network. The processing operations in training the generative adversarial network are as follows: based on the X-ray image and the corresponding edge feature map, obtain the shallow X-ray information features, the middle X-ray information features, and the deep X-ray information features; concatenate the deep X-ray information features with the deep edge features corresponding to the edge feature map, and obtain the channel enhancement features through multi-channel feature fusion processing; the X-ray image is obtained based on DR data and CT data; After extending the channel enhancement features to three dimensions, the attention enhancement features are obtained through several nonlocal attention processes. Attention-enhanced features, along with shallow X-ray, mid-layer X-ray, and deep X-ray information features, are processed by cross-attention based on residual connections and then reconstructed in three dimensions to obtain pseudo-CT data. Training ends when the training loss value of each pseudo-CT data and the corresponding real CT data is less than the loss threshold, and a generalized generative adversarial network is obtained. S2. Obtain multi-phase CT data of the target patient's lungs within a historical respiratory cycle. Register the first phase CT data in the multi-phase CT data to the other phase CT data to obtain a multi-phase registration displacement field. After sampling and augmentation, use the augmented displacement field to transform the other phase CT data to obtain multiple CT training data. Combine these with the target patient's DR data and, after preprocessing, form a specific dataset. A generalized generative adversarial network is trained using a specific dataset to obtain a specific generative adversarial network. S3. Real-time acquisition of DR data of the target patient's lungs, which is then processed by a specific generative adversarial network to obtain CT-generated data in real time.

[0006] In S1, the methods for obtaining X-ray shallow, mid-level, and deep information features are as follows: The X-ray image and its corresponding edge feature map are processed by two-dimensional convolution to obtain an X-ray convolutional image and an edge convolutional image, respectively. The X-ray convolutional image and the edge convolutional image are then stitched together and downsampled, and subjected to multiple multi-channel feature fusion processes to obtain the X-ray shallow information features. The edge convolutional image is then subjected to multiple multi-channel feature fusion processes to obtain the edge shallow information features. The X-ray shallow information features and the edge shallow information features are then stitched together and downsampled, and subjected to multiple multi-channel feature fusion processes to obtain the X-ray mid-level information features. The edge shallow information features are then subjected to multiple multi-channel feature fusion processes to obtain the edge mid-level information features. The X-ray mid-level information features and the edge mid-level information features are then stitched together and downsampled, and subjected to multiple multi-channel feature fusion processes to obtain the X-ray deep information features.

[0007] The multi-channel feature fusion processing operation is as follows: the stitched downsampled image is convolved in different channels, and then processed by stitching, layer normalization, convolution, and multilayer perceptron to obtain nonlinear features; the shallow nonlinear features and the stitched downsampled image are added element by element to obtain X-ray information features, which are used to perform the next multi-channel feature fusion processing.

[0008] The nonlocal attention operation in S1 is as follows: the channel-enhanced 3D features are processed by inner normalization to obtain inner normalized features; the inner normalized features are processed by 3D convolution at different scales, and then multiplied element-wise and nonlinearly to obtain convolutionally fused nonlinear features; the inner normalized features are processed by 3D convolution, and then multiplied element-wise with the convolutionally fused nonlinear features to obtain convolutionally fused features; the convolutionally fused features are added element-wise with the inner normalized features to obtain the initial attention-enhanced features, which are used to perform the next nonlocal attention process.

[0009] The training loss in S1 includes generator loss and discriminator loss; generator loss includes reconstruction loss of pseudo-CT data, projection loss of pseudo-CT data and corresponding real CT data, perception loss, and adaptive motion region loss.

[0010] The projection loss is calculated using the following formula: , For projection loss, , , These are the projection functions for the cross-section, sagittal plane, and coronal plane, respectively. , These are real CT data and pseudo CT data, respectively. For paradigm 1 processing, Let be the expected function.

[0011] The preprocessing operations in S1 are as follows: resample the CT data and crop out the central region to obtain the key CT data; spatially align the key CT data with the DR data to obtain the X-ray image; perform trilinear interpolation and bilinear interpolation on the CT data and X-ray image respectively, and adjust them to the same size to obtain data pairs; all data pairs form a generalized dataset.

[0012] A system for generating respiratory motion images from lung DR images to CT images, used to implement the aforementioned method for generating respiratory motion images from lung DR images to CT images, includes: The generalized generative adversarial network (GAN) module acquires DR and CT data from multiple patients' lungs. After preprocessing, a generalized dataset is formed, which is then used to train the GAN. The processing operations during GAN training are as follows: Based on the X-ray images and corresponding edge feature maps, shallow, mid, and deep X-ray information features are obtained. The deep X-ray information features are concatenated with the deep edge features corresponding to the edge feature maps, and then subjected to multi-channel feature fusion processing to obtain channel enhancement features. The X-ray images are obtained based on DR and CT data. After expanding the channel enhancement features into three dimensions, several non-local attention processes are performed to obtain attention enhancement features. The attention enhancement features, along with the shallow, mid, and deep X-ray information features, are subjected to cross-attention processing based on residual connections, and then 3D reconstruction is performed to obtain pseudo-CT data. Training ends when the training loss value of each pseudo-CT data point compared to the corresponding real CT data is less than a loss threshold, resulting in the generalized GAN. The specific generative adversarial network (GAN) generation module is used to acquire multi-phase CT data of the target patient's lungs within a historical respiratory cycle. The first phase CT data is registered to the other phase CT data to obtain a multi-phase registration displacement field. After sampling and augmentation, the augmented displacement field is used to transform the other phase CT data to obtain multiple CT training data. These training data are then combined with the target patient's DR data and preprocessed to form a specific dataset. The generalized generative adversarial network is trained using the specific dataset to obtain the specific generative adversarial network. The CT data generation module is used to acquire DR data of the lungs of the target patient in real time, and then process it with a specific generative adversarial network to obtain CT generated data in real time.

[0013] A device for generating respiratory motion images from lung DR images to CT images includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-described method for generating respiratory motion images from lung DR images to CT images.

[0014] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for generating lung DR images to CT respiratory motion images.

[0015] The beneficial effects of this invention are as follows: This invention provides a method for generating respiratory motion images from lung DR images to CT images. First, DR and CT data from multiple patients are acquired and preprocessed to form a generalized dataset. A generative adversarial network (GAN) is then trained to obtain a generalized GAN. Next, using multi-phase CT and DR data from a historical respiratory cycle of the target patient before surgical intervention, a specific dataset is obtained. This dataset is then used to train the generalized GAN to obtain a specific GAN suitable for the target patient, significantly reducing generation bias caused by individual differences and improving the matching degree between the generated images and the patient's actual physiological state. Finally, during interventional surgery and radiotherapy, only the real-time acquisition of the target patient's lung DR data is needed to obtain the patient's real-time 4D CT data, thereby enabling more timely and accurate lesion localization and targeted therapy. This invention provides a method for generating respiratory motion images from lung DR images to CT images. It designs a generative adversarial network (GAN) to enhance the feature fusion capabilities of X-ray images and edge feature maps. This allows patients to avoid excessive radiation exposure during the acquisition of guided medical images while simultaneously obtaining high-quality real-time CT images from X-ray films. The generation effect is controlled based on training loss, ensuring high structural similarity between the generated CT data and real CT data. Furthermore, it can present specific details of lung tissue, such as pulmonary blood vessels, tuberculosis lesions, and bronchi. The respiratory motion process it displays closely approximates the real respiratory process, assisting radiation oncologists in precise lesion localization and targeted therapy. Detailed Implementation

[0016] This embodiment provides a method for generating CT respiratory motion images from lung DR images, including the following operations: S1. Obtain DR and CT data of lungs from multiple patients, preprocess them to form a generalized dataset, train the generative adversarial network, and obtain the generalized generative adversarial network. S2. Acquire multi-phase CT data of the target patient's lungs within a historical respiratory cycle. Register the first phase CT data from the multi-phase CT data to the other phase CT data to obtain a multi-phase registration displacement field. After sampling and augmentation, use the augmented displacement field to transform the other phase CT data to obtain multiple CT training data. Combine this with the target patient's DR data and, after preprocessing, form a specific dataset. Use the specific dataset to train a generalized generative adversarial network to obtain a specific generative adversarial network. S3. Real-time acquisition of DR data of the target patient's lungs, which is then processed by a specific generative adversarial network to obtain CT-generated data in real time.

[0017] The specific steps are detailed below.

[0018] S1. Obtain DR and CT data of lungs from multiple patients, preprocess them to form a generalized dataset, train a generative adversarial network, and obtain a generalized generative adversarial network.

[0019] DR and CT data of lungs from multiple patients were acquired and preprocessed to form a generalized dataset. A generative adversarial network was then trained to obtain a generalized generative adversarial network. This allows for the use of a population generalized model to reduce the training time of patient-specific models and improve the robustness of patient-specific models.

[0020] First, DR and CT data of the lungs of multiple patients were acquired. After preprocessing, the DR and CT data of each patient were used to obtain multiple data pairs, forming a generalized dataset.

[0021] The preprocessing of DR and CT data specifically involves: using the SimpleITK library to convert the original DICOM format DR and CT data into NIFTI format; and resampling the CT data to a resolution of 1×1×1mm. 3 Cut out the central area, with a volume of 320×320×320 mm. 3 Key CT data were obtained; based on DRR technology, the key CT data and DR data were spatially aligned to obtain X-ray images; the CT data were then resized to 256×256×256 mm using trilinear interpolation. 3 The X-ray image was resized to 256×256×256mm using bilinear interpolation. 3 Finally, the CT data and X-ray images, adjusted to the same size, are stored in an h5py file in matrix form to obtain data pairs. All data pairs form a generalized dataset.

[0022] Next, a generative adversarial network is trained using a generalized dataset to achieve the function of generalizing DR data to generate CT data, thus obtaining a generalized generative adversarial network.

[0023] The details of the processing steps in training a generative adversarial network are as follows.

[0024] Step 1: Based on the X-ray image and the corresponding edge feature map, obtain the shallow X-ray information features, the middle X-ray information features, and the deep X-ray information features.

[0025] The specific methods for obtaining X-ray shallow, mid, and deep information features are as follows: The X-ray image and its corresponding edge feature map are processed by two-dimensional convolution to obtain an X-ray convolutional image and an edge convolutional image, respectively. The X-ray convolutional image and the edge convolutional image are then stitched together and downsampled, and subjected to multiple (preferably 3) multi-channel feature fusion processes to obtain the X-ray shallow information features. The edge convolutional image is then subjected to multiple (preferably 3) multi-channel feature fusion processes to obtain the edge shallow information features. The X-ray shallow information features and the edge shallow information features are then stitched together and downsampled, and subjected to multiple (preferably 3) multi-channel feature fusion processes to obtain the X-ray mid-layer information features. The edge shallow information features are then subjected to multiple (preferably 3) multi-channel feature fusion processes to obtain the edge mid-layer information features. The X-ray mid-layer information features and the edge mid-layer information features are then stitched together and downsampled, and subjected to multiple (preferably 27) multi-channel feature fusion processes to obtain the X-ray deep information features.

[0026] The above multi-channel feature fusion operation is as follows: the stitched downsampled image is subjected to convolution processing of different channels, and then processed by stitching, layer normalization, convolution and multilayer perceptron to obtain nonlinear features; the shallow nonlinear features and the stitched downsampled image are added element by element to obtain X-ray information features, which are used to perform the next multi-channel feature fusion processing.

[0027] The stitched downsampled image is obtained by stitching together X-ray convolutional images and edge convolutional images and then downsampling them, or by stitching together shallow X-ray information features and shallow edge information features and then downsampling them, or by stitching together mid-layer X-ray information features and mid-layer edge information features and then downsampling them.

[0028] In the convolution processing of different channels, the result of the convolution processing of the first channel remains unchanged from the stitched downsampled image, that is, no processing is performed on the stitched downsampled image; the convolution processing of the second channel is achieved through 3×3 convolution, and the convolution processing of the third and fourth channels is achieved through orthogonal strip convolution.

[0029] Convolution processing for different channels is achieved using the following formula: , , , , , To stitch downsampled images x Perform partitioning operations with different allocation ratios. , , , These are the first segmentation image, the second segmentation image, the third segmentation image, and the fourth segmentation image, respectively. , , , These are the feature maps for the first, second, third, and fourth channels, respectively. Orthogonal strip convolution is used to extract multi-channel depth features. , , They are 3×3 convolution, 11×1 convolution, and 1×11 convolution, respectively. This is for splicing operations.

[0030] Step 2: The X-ray deep information features are stitched together with the deep edge features corresponding to the edge feature map. After multi-channel feature fusion processing, channel enhancement features are obtained. The X-ray image is obtained based on DR data and CT data.

[0031] The method for obtaining the deep edge features corresponding to the edge feature map is as follows: the edge mid-layer information features are processed by multiple (preferably 3 times) multi-channel feature fusion to obtain the edge deep-layer information features.

[0032] Step 3: Expand the channel enhancement features into three dimensions to obtain channel enhancement three-dimensional features; the channel enhancement three-dimensional features are processed by non-local attention several times (preferably 2 times) to obtain attention enhancement features.

[0033] The nonlocal attention processing operation is as follows: the channel-enhanced 3D features are processed by inner normalization to obtain inner normalized features; the inner normalized features are processed by 3D convolution at different scales, and then multiplied element-wise and nonlinearly processed (preferably using the softmax function) to obtain convolutional fusion nonlinear features; the inner normalized features are processed by 3D convolution, and then multiplied element-wise with the convolutional fusion nonlinear features to obtain convolutional fusion features; the convolutional fusion features are added element-wise with the inner normalized features to obtain the initial attention-enhanced features, which are used to perform the next nonlocal attention processing.

[0034] Step 4: The attention-enhanced features, shallow X-ray information features, mid-layer X-ray information features, and deep X-ray information features are processed by cross-attention based on residual connections, and then reconstructed in three dimensions to obtain pseudo-CT data.

[0035] The specific operation of cross-attention processing based on residual connections is as follows: attention enhancement features and shallow X-ray information features are processed through cross-attention to obtain a first cross-attention feature; the first cross-attention feature and mid-layer X-ray information features are processed through cross-attention to obtain a second cross-attention feature; the second cross-attention feature and deep X-ray information features are processed through cross-attention to obtain a third cross-attention feature, which is used to perform the three-dimensional reconstruction operation.

[0036] Taking the second cross-attention feature as an example, the cross-attention processing operation is as follows: after the X-ray mid-layer information feature is expanded into a three-dimensional feature, the key feature and value feature are obtained. The query feature of the first cross-attention feature is processed by attention, three-dimensional convolution based on residual connection and upsampling to obtain the second cross-attention feature.

[0037] Step 5: When the training loss of each pseudo-CT data and the corresponding real CT data is less than the loss threshold, the training ends and the generalized generative adversarial network is obtained.

[0038] Training loss includes generator loss and discriminator loss.

[0039] The generator loss includes reconstruction loss of pseudo-CT data, projection loss of pseudo-CT data and corresponding real CT data, perception loss, and adaptive motion region loss.

[0040] The projection loss is calculated using the following formula: , For projection loss, , , These are the projection functions for the cross-section, sagittal plane, and coronal plane, respectively. , These are real CT data and pseudo CT data, respectively. For paradigm 1 processing, Let be the expected function.

[0041] Perceptual loss is calculated using the following formula: , In order to perceive loss, for Network processing is used to extract texture and structural information from images. This is handled in paradigm 2.

[0042] The adaptive motion region loss is calculated using the following formula: , For adaptive motion region loss, This refers to the patient's 4D CT data, which includes CT data from multiple time points within a single respiratory cycle. Pseudo CT data To real CT data displacement field, This is for paradigm processing.

[0043] Discriminator loss includes adversarial loss and decoding loss.

[0044] S2. Acquire multi-phase CT data (10 phases) of the target patient within a historical respiratory cycle. Register the first phase CT data from the multi-phase CT data to the other phases to obtain multi-phase registration displacement fields (1080 displacement fields). After sampling and augmentation (which can be achieved using a principal component analysis model PCA-MM), use the augmented displacement fields to transform the other phases of CT data to obtain multiple CT training datasets (1080 CT data to simulate the human morphology present during the patient's complete respiratory process). Combine these with the target patient's DR data, and after preprocessing, form a specific dataset. Use the specific dataset to train a generalized generative adversarial network to obtain a specific generative adversarial network suitable for the target patient. This ensures that the generative adversarial network conforms to the individual respiratory patterns and anatomical characteristics of the target patient, significantly reducing generation bias caused by individual differences, improving the matching degree between the generated images and the patient's real physiological state, and thus improving the accuracy and reliability of subsequent clinical applications based on generated images.

[0045] S3. In clinical practice, DR data (X-ray images) of target patients can be acquired in real time and processed by a specific generative adversarial network to obtain CT data in real time, thereby making lesion localization and targeted treatment more timely and accurate.

[0046] To verify the generation effect of this embodiment, the following implementation was performed.

[0047] Experimental Setup. The experimental development environment used was PyTorch 2.0.1, on a single NVIDIA RTX 4090 graphics processor. The Adam optimizer was used with a learning rate of 0.0002 and a batch size of 1. In the pre-training phase, the dataset was trained for 100 epochs using diverse patient data, taking approximately 15 hours. In the transfer learning phase, the dataset was trained for 20 epochs, taking approximately 3 hours. Experimental parameters are shown in Table 1, and the results are shown in Table 2.

[0048] Table 1 Summary of Experimental Parameters

[0049] Table 2 Summary of Experimental Results

[0050] This embodiment also provides a system for generating respiratory motion images from lung DR images to CT images, used to implement the above-mentioned method for generating respiratory motion images from lung DR images to CT images, including: The generalized generative adversarial network (GAN) module acquires DR and CT data from multiple patients' lungs. After preprocessing, a generalized dataset is formed, which is then used to train the GAN. The processing operations during GAN training are as follows: Based on the X-ray images and corresponding edge feature maps, shallow, mid, and deep X-ray information features are obtained. The deep X-ray information features are concatenated with the deep edge features corresponding to the edge feature maps, and then subjected to multi-channel feature fusion processing to obtain channel enhancement features. The X-ray images are obtained based on DR and CT data. After expanding the channel enhancement features into three dimensions, several non-local attention processes are performed to obtain attention enhancement features. The attention enhancement features, along with the shallow, mid, and deep X-ray information features, are subjected to cross-attention processing based on residual connections, and then 3D reconstruction is performed to obtain pseudo-CT data. Training ends when the training loss value of each pseudo-CT data point compared to the corresponding real CT data is less than a loss threshold, resulting in the generalized GAN. The specific generative adversarial network (GAN) generation module is used to acquire multi-phase CT data of the target patient's lungs within a historical respiratory cycle. The first phase CT data is registered to the other phase CT data to obtain a multi-phase registration displacement field. After sampling and augmentation, the augmented displacement field is used to transform the other phase CT data to obtain multiple CT training data. These training data are then combined with the target patient's DR data and preprocessed to form a specific dataset. The generalized generative adversarial network is trained using the specific dataset to obtain the specific generative adversarial network. The CT data generation module is used to acquire DR data of the lungs of the target patient in real time, and then process it with a specific generative adversarial network to obtain CT generated data in real time.

[0051] This embodiment also provides a device for generating respiratory motion images from lung DR images to CT images, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-described method for generating respiratory motion images from lung DR images to CT images.

[0052] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for generating lung DR images to CT respiratory motion images.

[0053] This embodiment provides a method for generating respiratory motion images from lung DR images to CT images. First, DR and CT data from multiple patients are acquired and preprocessed to form a generalized dataset. A generative adversarial network (GAN) is then trained to obtain a generalized GAN. Next, a specific dataset is obtained using multiple phases of CT and DR data from a historical respiratory cycle of the target patient before surgical intervention. This dataset is then used to train the generalized GAN to obtain a specific GAN suitable for the target patient, significantly reducing generation bias caused by individual differences and improving the matching degree between the generated images and the patient's actual physiological state. Finally, during interventional surgery and radiotherapy, only the real-time acquisition of the target patient's lung DR data is needed to obtain the patient's real-time 4D CT data, thereby enabling more timely and accurate lesion localization and targeted treatment.

[0054] This embodiment provides a method for generating respiratory motion images from lung DR images to CT images. It designs a generative adversarial network to enhance the feature fusion capability of X-ray images and edge feature maps. This allows patients to avoid excessive radiation exposure during the acquisition of guided medical images while obtaining high-quality real-time CT images through X-ray films. The generation effect is controlled according to the training loss, so that the generated CT data has a high structural similarity to real CT data and can present specific details of lung tissue, such as pulmonary blood vessels, pulmonary tuberculosis lesion areas, and bronchi. Moreover, the respiratory motion process it displays is very close to the real respiratory process, which can assist radiation oncologists in accurately locating lesions and targeted treatment.

Claims

1. A method for generating respiratory motion images from lung DR images to CT images, characterized in that, This includes the following operations: S1. Obtain DR and CT data of lungs from multiple patients, preprocess them to form a generalized dataset, train the generative adversarial network, and obtain the generalized generative adversarial network. The processing operations in training the generative adversarial network are as follows: based on the X-ray image and the corresponding edge feature map, obtain the shallow X-ray information features, the middle X-ray information features, and the deep X-ray information features; concatenate the deep X-ray information features with the deep edge features corresponding to the edge feature map, and obtain the channel enhancement features through multi-channel feature fusion processing; the X-ray image is obtained based on DR data and CT data; After extending the channel enhancement features to three dimensions, the attention enhancement features are obtained through several nonlocal attention processes. Attention-enhanced features, along with shallow X-ray, mid-layer X-ray, and deep X-ray information features, are processed by cross-attention based on residual connections and then reconstructed in three dimensions to obtain pseudo-CT data. Training ends when the training loss value of each pseudo-CT data and the corresponding real CT data is less than the loss threshold, and a generalized generative adversarial network is obtained. S2. Obtain multi-phase CT data of the target patient's lungs within a historical respiratory cycle. Register the first phase CT data in the multi-phase CT data to the other phase CT data to obtain a multi-phase registration displacement field. After sampling and augmentation, use the augmented displacement field to transform the other phase CT data to obtain multiple CT training data. Combine these with the target patient's DR data and, after preprocessing, form a specific dataset. A generalized generative adversarial network is trained using a specific dataset to obtain a specific generative adversarial network. S3. Real-time acquisition of DR data of the target patient's lungs, which is then processed by a specific generative adversarial network to obtain CT-generated data in real time.

2. The method for generating CT respiratory motion images from lung DR images according to claim 1, characterized in that, In S1, the methods for obtaining shallow X-ray information features, mid-layer X-ray information features, and deep X-ray information features are as follows: The X-ray image and the corresponding edge feature map are respectively processed by two-dimensional convolution to obtain the X-ray convolution map and the edge convolution map; After the X-ray convolutional image and the edge convolutional image are stitched together, they are downsampled and then subjected to multiple multi-channel feature fusion processes to obtain the shallow X-ray information features. The edge convolutional map is processed through multiple multi-channel feature fusion processes to obtain shallow edge information features; the shallow X-ray information features and the shallow edge information features are concatenated and downsampled, and then processed through multiple multi-channel feature fusion processes to obtain the mid-layer X-ray information features; The shallow edge information features are processed through multiple multi-channel feature fusion processes to obtain the mid-edge edge information features; the mid-edge X-ray information features and the mid-edge edge information features are spliced ​​together and then downsampled, and processed through multiple multi-channel feature fusion processes to obtain the deep X-ray information features.

3. The method for generating CT respiratory motion images from lung DR images according to claim 1 or 2, characterized in that, The operation of multi-channel feature fusion processing is as follows: The stitched downsampled image is convolved in different channels, and then processed by stitching, layer normalization, convolution, and multilayer perceptron to obtain nonlinear features. The shallow nonlinear features are added to the stitched downsampled image element by element to obtain X-ray information features, which are used to perform the next multi-channel feature fusion process.

4. The method for generating CT respiratory motion images from lung DR images according to claim 1, characterized in that, The non-local attention operation in S1 is as follows: The enhanced 3D features of the channel are then processed by internal normalization to obtain internally normalized features; After the internally normalized features are processed by 3D convolution at different scales, they are multiplied element-wise and nonlinearly processed to obtain convolutionally fused nonlinear features. After the inner normalized features are processed by 3D convolution, they are multiplied element-wise with the convolutional fusion nonlinear features to obtain the convolutional fusion features. The convolutional fusion features are then added element-wise with the inner normalized features to obtain the initial attention enhancement features, which are used to perform the next nonlocal attention processing.

5. The method for generating CT respiratory motion images from lung DR images according to claim 1, characterized in that, In S1, the training loss includes generator loss and discriminator loss; generator loss includes reconstruction loss of pseudo-CT data, projection loss of pseudo-CT data and corresponding real CT data, perception loss, and adaptive motion region loss.

6. The method for generating CT respiratory motion images from lung DR images according to claim 5, characterized in that, The projection loss is calculated using the following formula: ， For projection loss, , , These are the projection functions for the cross-section, sagittal plane, and coronal plane, respectively. , These are real CT data and pseudo CT data, respectively. For paradigm 1 processing, Let be the expected function.

7. The method for generating CT respiratory motion images from lung DR images according to claim 1, characterized in that, In S1, the preprocessing operation is as follows: The CT data is resampled and the central region is cropped to obtain the key CT data. The key CT data and DR data are spatially aligned to obtain the X-ray image. The CT data and X-ray image are then subjected to trilinear interpolation and bilinear interpolation respectively, and adjusted to the same size to obtain data pairs. All data pairs form a generalized dataset.

8. A system for generating respiratory motion images from lung DR images to CT images, used to implement the method for generating respiratory motion images from lung DR images to CT images as described in claim 1, characterized in that, include: The generalized generative adversarial network generation module is used to acquire DR and CT data of the lungs of multiple patients. After preprocessing, a generalized dataset is formed, which is then used to train the generative adversarial network to obtain the generalized generative adversarial network. The processing operations in training the generative adversarial network are as follows: Based on the X-ray image and the corresponding edge feature map, obtain the shallow X-ray information features, the mid-layer X-ray information features, and the deep X-ray information features; concatenate the deep X-ray information features with the deep edge features corresponding to the edge feature map, and obtain the channel enhancement features through multi-channel feature fusion processing; the X-ray image is obtained based on DR data and CT data; after expanding the channel enhancement features into three dimensions, it is processed through several non-local attention processes to obtain the attention enhancement features; Attention-enhanced features, along with shallow, mid, and deep X-ray information features, are processed through cross-attention based on residual connections and then reconstructed in 3D to obtain pseudo-CT data. Training ends when the training loss value of each pseudo-CT data and its corresponding real CT data is less than the loss threshold, resulting in a generalized generative adversarial network. The specific generative adversarial network generation module is used to acquire multi-phase CT data of the lungs of the target patient within a historical respiratory cycle. The first phase CT data in the multi-phase CT data is registered to the other phase CT data to obtain a multi-phase registration displacement field. After sampling and augmentation, the augmented displacement field is used to transform the other phase CT data to obtain multiple CT training data. These data, along with the target patient's DR data, are preprocessed to form a specific dataset. A generalized generative adversarial network is trained using a specific dataset to obtain a specific generative adversarial network. The CT data generation module is used to acquire DR data of the lungs of the target patient in real time, and then process it with a specific generative adversarial network to obtain CT generated data in real time.

9. A device for generating respiratory motion images from lung DR images to CT images, characterized in that, The system includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the method for generating lung DR images to CT respiratory motion images as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method for generating lung DR images to CT respiratory motion images as described in any one of claims 1-7.

Images