A dual-view CT reconstruction method based on dual-domain supervision of projection and tomography
The dual-domain supervision method addresses the challenges of high radiation and incomplete projection in CT reconstruction by using a network architecture to enhance dual-view CT reconstruction accuracy and reliability, achieving high-precision images with reduced radiation dose.
Patent Information
- Application Number
- CN202411404516.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-10-10
AI Technical Summary
Traditional CT reconstruction methods require full-circle scanning, leading to high radiation doses and incomplete projection information, resulting in artifacts and reduced reconstruction quality, while dual-view methods face significant challenges in accurate three-dimensional reconstruction due to severe projection deficiencies.
A dual-domain supervision method combining projection and tomographic domains using a network architecture comprising a projection completion network, a projection residual learning network, and a tomographic refinement network, with a reprojection loss to enhance accuracy and reliability, allowing reconstruction from only two orthogonal projections.
Achieves high-precision three-dimensional reconstruction with reduced radiation dose by leveraging dual-domain supervision, improving image quality and reducing artifacts through specialized network architectures and reprojection loss.
Smart Images

Figure CN118918211B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of medical imaging and artificial intelligence, and relates to a dual-viewing angle CT reconstruction method based on projection and tomography dual-domain supervision. Background Art
[0002] CT technology is one of the core technologies of modern medical diagnosis. Through CT scanning, detailed sections of organs and tissues in the human body can be obtained. It has been widely used in cancer and trauma diagnosis, preoperative surgical planning, cardiovascular disease treatment, etc., providing doctors with a basis for diagnosis, thereby ensuring the accuracy and safety of treatment measures. However, the traditional CT reconstruction method requires full-circle scanning to obtain projection data of the human body at all angles, which will cause the human body to receive higher doses of radiation and damage human health.
[0003] In order to reduce the number of projection acquisitions and reduce the radiation dose received by the human body, CT reconstruction methods with limited view and sparse view came into being. However, due to the incomplete projection information obtained by these two methods, the reconstructed image will have artifacts, affecting the reconstruction quality. With the development of artificial intelligence technology, deep neural networks are applied to CT reconstruction in the absence of projections, and the structural priors of the object to be reconstructed are learned through a large number of data samples, thereby making up for the lack of projection information. Patent CN116843825A reports a progressive CBCT sparse view reconstruction method, which learns CT reconstruction of sparse views in a progressive learning process by constructing a multi-stage convolutional neural network. However, regardless of the reconstruction method of sparse view or limited view, a large number of projection data still need to be obtained. In order to further reduce the scanning time and the radiation dose received by the human body, some scholars have proposed a dual-view CT reconstruction scheme, that is, only two orthogonal projections are used to complete the three-dimensional reconstruction, but due to the serious lack of projections, the reconstruction is extremely difficult. Patent CN117392259A reports a method for cross-modal reconstruction of CT images with dual view, combining the diffusion model and the GAN network, and reconstructing the projections of two orthogonal views into a three-dimensional CT image across the domain. This method uses CT true value images for single-domain supervision in the tomographic domain and directly generates tomographic data from projection data. This cross-dimensional conversion may introduce conversion errors.
[0004] Different from the existing single-domain supervision scheme, the present invention proposes a dual-domain supervision method combining the projection domain and the fault domain to realize dual-view CT reconstruction. Specifically, the present invention uses the projection truth value of 180 angles in the projection domain as supervision to complete the orthogonal dual-view projection by 180°. Then the FDK algorithm is used for preliminary reconstruction, converting from the projection domain to the fault domain. Subsequently, real fault data is used to supervise learning in the fault domain. Such a dual-domain supervision method avoids the conversion error of cross-domain supervision, and the introduction of the FDK reconstruction algorithm maintains the accuracy of cross-domain conversion. The present invention proposes a dual-domain supervised network to realize dual-view CT reconstruction, which includes three parts: a projection completion network, a projection residual learning network, and a fault refinement network. The projection completion network is based on the Swin Transformer structure, and effectively captures the local details and global features of the projection through sliding window technology and self-attention mechanism; the projection residual learning network and the fault refinement network both adopt a two-way parallel encoder-decoder architecture, and the two ways learn structural features and high-frequency details respectively, so as to improve the accuracy of completing projections and reconstructing faults. In order to further ensure the reliability of the reconstructed fault, the present invention designs a reprojection loss, that is, by reprojecting the stereoscopic prediction of the fault refinement network and using the projection true value in the corresponding range as supervision, the potential errors in the reconstruction process are effectively corrected. The present invention avoids the conversion error of cross-domain supervision, further improves the accuracy and reliability of the reconstructed three-dimensional fault by designing a special network architecture and introducing the reprojection loss, and has a higher reconstruction quality. Summary of the invention
[0005] The present invention provides a dual-view CT tomographic reconstruction method based on projection and tomographic dual-domain supervision. The three-dimensional tomography of the corresponding part can be reconstructed by only using the projection views of two orthogonal angles of the human body, which can achieve low-radiation dose medical imaging. The dual-view reconstruction model proposed in the present invention includes three parts: a projection completion network, a projection residual learning network, and a tomographic refinement network. The overall architecture is as follows: Figure 1 As shown in Figure 2, through dual supervision of the projection domain and the tomography domain, high-precision reconstruction of dual-view CT is achieved through collaboration.
[0006] The technical solution of the present invention:
[0007] A dual-view CT reconstruction method based on dual-domain supervision of projection and tomography, the steps are as follows:
[0008] Step 1: Constructing a “projection-tomography” CT dataset
[0009] The projection data of a certain part of the human body is collected at intervals of 1°, and the tomography of the part is reconstructed. The posteroanterior (PA) view is defined as 0°, and the lateral (LAT) view is defined as 90°. 0° and 90° are selected as the input of the dual-view CT reconstruction network, the 0-180° projection data is the projection truth value, the 90-270° projection data is the reprojection truth value, and the reconstructed tomography is used as the tomography truth value. By collecting data of several cases of this part of the human body, multiple sets of data are constructed to complete the data set production, which is used to train and verify the dual-view CT reconstruction network.
[0010] Step 2: Build a dual-domain supervised dual-view CT reconstruction network
[0011] (1) Building a projection completion network: The projection completion network is responsible for completing the input 0° and 90° projections into 180 projections with an interval of 1° for FDK reconstruction to achieve the conversion from the projection domain to the fault domain.
[0012] The projection completion network consists of an encoder-decoder architecture based on Swin Transformer and a shallow feature extraction module. The specific structure is as follows Figure 2 As shown in Figure 2, the network input is 0° and 90° projections, and the number of channels is 2.
[0013] The encoder first uses the Patch Partition module to convert the input 128 The 0° and 90° projections of 128 resolution are split into 4 4 non-overlapping blocks, i.e. patches. The feature channel dimension is then expanded by the Linear Embedding module, so that each patch is converted into a feature vector with a higher dimension. These high-dimensional feature vectors are then fed into two consecutive Swin Transformer Block modules, which enhance the network's ability to capture local details and integrate global features through their self-attention mechanism and sliding window. Downsampling is then performed through the Patch Merging module to merge adjacent patches and increase feature depth. The above process of processing through two Swin Transformer Block modules and then downsampling by the Patch Merging module is repeated 3 times in the encoder. At the bottom of the encoder, it is connected to the decoder through two consecutive Swin Transformer Block modules.
[0014] The decoder first performs upsampling through the Patch Expanding module to expand the feature size and reduce the feature depth, and then passes through two consecutive Swin Transformer Block modules. This process is repeated three times in the decoder. In order to compensate for the loss of spatial information caused by downsampling, skip connections are introduced between layers with the same feature size in the encoder and decoder.
[0015] Since the division of patches will cause the network to lose a large number of shallow features, a shallow feature learning module is designed, which consists of a convolutional layer, a downsampling layer and two upsampling layers. The input 0° and 90° projections first pass through a convolutional layer to increase the channel dimension to 180, and then pass through a downsampling layer composed of residual and convolution to perform 2x downsampling. For the output of the decoder, it is upsampled twice by 2x through two upsampling layers composed of residual and deconvolution. Layers with the same feature size are connected by jump connections. The input 0° and 90° projections are processed by the encoder-decoder based on Swin Transformer and the shallow feature learning module, and finally the completed 180 projections are output.
[0016] (2) Build a projection residual learning network: Experiments have found that the difference between the true value projection and the completed projection contains the structure and detail information of the projection. Therefore, a projection residual learning network is designed to minimize the difference between the true value projection and the completed projection, thereby improving the accuracy of projection completion. The specific structure is as follows: Figure 3 shown.
[0017] The projection residual learning network is a two-way parallel encoder-decoder architecture based on the residual attention module. The network is divided into a main path and a branch path, and the two paths have the same structure. The input of the network is 180 completed projections output by the projection completion network, the number of input channels is 180, and the supervision of the network is the difference between the true projection value and the completed projection. The main path of the network is responsible for learning the structural features of the projection difference, and the branch path of the network is responsible for learning the high-frequency details of the projection difference.
[0018] The main path of the network is downsampled four times and upsampled four times with skip connections. Before each downsampling and after each upsampling, it passes through a residual attention module. The number of convolution kernels of the residual attention block is scaled accordingly at different feature scales. The architecture of the branch network is the same as the main path, focusing on learning the high-frequency quantity of the projection difference. The learned high-frequency features are fused with the features learned by the main path of the network. The residual attention module is designed as follows: through a series of 3 3 convolutional layers for deep feature extraction, and each convolutional layer is equipped with a batch normalization layer and a ReLU activation function. A densely connected structure is used to merge the output of each convolutional layer with the output of all previous convolutional layers to enhance feature transfer and utilization efficiency. Finally, the spatial attention layer is introduced to extract spatial features through the maximum pooling layer, average pooling layer, and convolution layer. After aggregation, the convolution layer and Sigmoid activation are used to generate an attention weight map, which is weighted by multiplying the features of the input spatial attention layer element by element to highlight key features, refine the model's attention to important areas in the image, and optimize detail recovery and image quality.
[0019] In this way, the two paths learn the structural features and detail information of the projection difference respectively. Finally, the projection difference output by the projection residual learning network is added to the completed projection output by the projection completion network to obtain a high-precision completed projection.
[0020] (3) Building a fault refinement network: Experiments have found that the faults obtained by FDK reconstruction after the completed projection are still relatively rough. Therefore, a fault refinement network is designed to refine the rough faults reconstructed by FDK, learn the missing details, and improve the accuracy of the faults.
[0021] The fault refinement network uses the same two-way parallel encoder-decoder architecture as the projection residual learning network. The main path and branch paths of the network learn the structural features and high-frequency details of the fault respectively. The specific structure is as follows: Figure 3 The input of the fault refinement network is the fault obtained by FDK reconstruction of the high-precision completed projection. The number of input channels is the number of reconstructed fault layers, 128, and the network supervision is the true value of the fault.
[0022] Step 3: Train a dual-domain supervised dual-view CT reconstruction network
[0023] The dual-domain supervised dual-view CT reconstruction network consists of three parts: a projection completion network, a projection residual learning network, and a tomographic refinement network. The training sequence is as follows: first, the projection completion network is trained separately to obtain a preliminary projection completion result; then, the parameters of the projection completion network are fixed, and the projection residual learning network is trained. The learned projection residual is added to the preliminary completion projection to obtain an accurate projection completion result; finally, on the basis of fixing the parameters of the first two parts of the network, the accurate projection completion result is reconstructed by FDK as the input of the tomographic refinement network, and the tomographic refinement network is trained to obtain a fine tomographic image.
[0024] (1) Training the projection completion network: The projection completion network completes the input dual-view projection to a 180-degree projection. The quality of its training results will directly affect the subsequent network performance, so the projection completion network is first trained separately. The training process is as follows: From a set of data established in step 1, select 0° and 90° in the true projection value as the input of the projection completion network, and the 0-180° projection predicted by the network is named The network supervision is the 0-180° projection truth value, named The loss function used is composed of MSE loss and SSIM loss, as shown in formula (1).
[0025] (1)
[0026] in , for The loss weight coefficient is . Using the multiple sets of data established in step 1, repeat the above training process. After training to a certain number of generations, Convergence, projection completion network training is complete. At this point, the network parameters are fixed for the next step of training.
[0027] (2) Training the projection residual learning network: The projection residual learning network learns the difference between the predicted projection and the true projection value to improve the accuracy of the completed projection. The training process is as follows: The 0-180° projection predicted by the projection completion network is used as input. The 0-180° projection difference predicted by the projection residual learning network is named , the main supervision of the network is the difference between the completed projection and the true value projection = - , the high-frequency quantity of the 0-180° projection difference predicted by the network branch is named , the network branch supervision is High frequency ,in For extraction The loss function used is MSE loss, as shown in formula (2).
[0028] (2)
[0029] Train to After convergence, the network parameters are fixed for the next step of training.
[0030] (3) Training the fault refinement network: Add the predicted projection and projection difference trained in the first two steps to obtain an accurate 180-degree angle complement projection, and then use the FDK algorithm to reconstruct the preliminary fault result. The fault refinement network learns the missing details of the preliminary fault result and obtains a fine fault image. The training process is as follows: Taking the fault result reconstructed by FDK as input, the fault predicted by the main path of the fault refinement network is named , the main network supervision is the fault truth value The high-frequency fault quantity predicted by the network branch is named , the network branch supervision is the high frequency quantity of the fault truth value .
[0031] In order to ensure the accuracy of the predicted fault, the present invention specially designs a reprojection loss function. For the three-dimensional predicted by the network, it is projected at 1° intervals to obtain its 90-270° projection data. (equivalent to 0-180° projection after 90° rotation of the stereo, and the true projection value of 90-270° obtained in step 1 Supervision. The reprojection loss introduces information from other stereoscopic perspectives, which helps to improve the reliability of predicted faults. Its definition is shown in formula (3).
[0032] (3)
[0033] in , for The loss weight coefficient of the fault refinement network is composed of MSE loss and reprojection loss. The composition is shown in formula (4).
[0034] (4)
[0035] in , , for The loss weight coefficient. When convergence occurs, the network training is complete.
[0036] Effects and benefits of the present invention:
[0037] (1) The present invention proposes a dual-view CT reconstruction method based on projection and tomography dual-domain supervision, which avoids the conversion error of cross-domain supervision and completes high-precision reconstruction of dual-view CT.
[0038] (2) This paper proposes a projection completion network, a projection residual learning network, and a tomographic refinement network to work together to complete dual-domain supervised dual-view CT reconstruction. A special architecture is designed for these three networks, and a reprojection loss is introduced to improve the reconstruction quality.
[0039] (3) The present invention can reconstruct the tomographic image of the corresponding part by only using the projection of two viewing angles of the human body, which greatly reduces the radiation dose received by the human body. This is of great significance for improving the safety of treatment and protecting the health of the human body. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 This is the overall architecture diagram of the dual-view CT reconstruction network based on dual-domain supervision of the present invention.
[0041] Figure 2 It is a detailed architecture diagram of the projection completion network of the present invention.
[0042] Figure 3 It is a detailed architecture diagram of the projection residual learning network and fault refinement network of the present invention. DETAILED DESCRIPTION
[0043] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.
[0044] First, under the premise of ensuring the privacy and rights of the human body, we contacted the cooperative hospital to obtain about 500 projection and tomographic data of a certain part of the human body (such as the chest cavity). According to step 1, we built a data set for parameter training of the network.
[0045] Then, follow step 2 to build a dual-domain supervised dual-view CT reconstruction network: Build a projection completion network, according to step 2 and combined Figure 2 Build a Swin Transformer-based encoder-decoder with a shallow feature learning module; build a projection residual learning network, according to step 2 and combined Figure 3 A two-way parallel encoder-decoder architecture based on the residual attention module is built, and the number of input channels is set to 180; a fault refinement network is built, whose architecture is the same as the projected residual learning network, and the number of input channels is set to 128.
[0046] Secondly, train the dual-domain supervised dual-view CT reconstruction network according to step 3. Train the three networks in the corresponding order: first train the projection completion network to obtain projection data of sufficient viewing angles, then train the projection residual learning network to improve the accuracy of the completed projection, reconstruct the FDK algorithm to achieve the conversion from the projection domain to the fault domain, and finally train the fault refinement network to obtain fine and clear faults. Specific details: 450 human body data were selected as the training set and 50 as the validation set. According to the above process, three stages of training were carried out. After each stage, the network parameters of all previous stages were fixed. After each stage has a good effect on the validation set and the loss function tends to converge, it enters the next stage of training. It takes 40 hours to complete the overall training under the RTX 4090 GPU.
[0047] After the network training and verification are completed, the system integration and implementation can be carried out. In the system integration phase, the hardware is configured and the software environment is built to ensure that the server has sufficient computing resources. An interface is established to connect with the CT equipment. A data preprocessing module is developed to process the received data into a format acceptable to the network. The trained model is integrated into the system for dual-view CT reconstruction. The three-dimensional tomographic image output by the model is transmitted to the observation device or image viewing system through the interface. During clinical implementation, the CT device is first used to collect the projection data of the human body from the two perspectives of PA and LAT. After the acquisition is completed, the data is transmitted to the CT reconstruction system through a secure interface. After the system receives the data, it uses the preprocessing module to process it into a format suitable for network input. Then, the preprocessed data is input into the dual-view CT reconstruction model to reconstruct the corresponding three-dimensional tomographic image. The reconstructed three-dimensional image is transmitted to the hospital's observation equipment through the interface.
Claims
1. A dual-view CT reconstruction method based on dual-domain supervision of projection and tomography, characterized in that: Here are the steps: Step 1: Constructing a "projection-tomography" CT dataset The projection of a certain part of the human body is collected at intervals of 1°, and the fault of the part is reconstructed; the posteroanterior view is defined as 0°, and the lateral view is defined as 90°; the 0° and 90° projections are selected as the input of the dual-view CT reconstruction network, the 0-180° projection is the projection true value, the 90-270° projection is the re-projection true value, and the reconstructed fault is used as the fault true value; By collecting data of this part of several patients, multiple sets of data are constructed to complete the data set production, which is used to train and verify the dual-view CT reconstruction network; Step 2: Build a dual-domain supervised dual-view CT reconstruction network (1) Building a projection completion network: The projection completion network is responsible for completing the input 0° and 90° projections into 180 projections with an interval of 1° for FDK reconstruction to achieve the conversion from the projection domain to the fault domain; The projection completion network consists of an encoder-decoder architecture based on Swin Transformer and a shallow feature extraction module. The input of the projection completion network is 0° and 90° projections, and the number of channels is 2; The encoder first divides the input 0° and 90° projections into 4 4 non-overlapping blocks, i.e., patches; the feature channel dimension is then expanded by the Linear Embedding module, so that each patch is converted into a feature vector with a higher dimension; then, these feature vectors are fed into two consecutive Swin Transformer Block modules, which enhance the ability of the projection completion network to capture local details and integrate global features through their self-attention mechanism and sliding window; then down-sampled by the Patch Merging module to merge adjacent patches and increase feature depth; the above process of processing by two Swin Transformer Block modules and then down-sampling by the Patch Merging module is repeated 3 times in the encoder; at the bottom of the encoder, it is connected to the decoder through two consecutive Swin Transformer Block modules; The decoder first performs upsampling through the Patch Expanding module to expand the feature size and reduce the feature depth, and then passes through two consecutive Swin Transformer Block modules. This process is repeated three times in the decoder. In order to compensate for the loss of spatial information caused by downsampling, skip connections are introduced between the layers with the same feature size in the encoder and decoder. Since the division of patches will cause the projection completion network to lose a lot of shallow features, a shallow feature learning module is designed; The shallow feature learning module includes a convolution layer, a downsampling layer, and two upsampling layers. The input 0° and 90° projections first pass through a convolution layer to increase the channel dimension to 180, and then pass through a downsampling layer composed of residual and convolution to perform 2x downsampling. For the output of the decoder, it passes through two upsampling layers composed of residual and deconvolution to perform two times of upsampling. Layers with the same feature size are connected by skip connections; the input 0° and 90° projections are processed by the Swin Transformer-based encoder-decoder and shallow feature learning module, and finally 180 complementary projections are output; (2) Build a projection residual learning network: The projection residual learning network is a two-way parallel encoder-decoder architecture based on the residual attention module. The projection residual learning network is divided into a main network path and a branch network, and the two paths have the same structure. The input of the projection residual learning network is the 180 completed projections output by the projection completion network. The number of input channels is 180. The supervision of the completion network is the difference between the true projection value and the completed projection. The main path is responsible for learning the structural features of the projection difference, and the branch network is responsible for learning the high-frequency details of the projection difference. The main path of the network performs four downsampling and four upsampling with skip connections. Before each downsampling and after each upsampling, a residual attention module is passed. The number of convolution kernels of the residual attention block is scaled accordingly at different feature scales. The architecture of the network branch is the same as that of the main network, focusing on learning the high-frequency quantity of the projection difference. The learned high-frequency features are integrated with the features learned by the main network. The residual attention module is designed as follows: Through a series of 3 3 convolutional layers for deep feature extraction, each convolutional layer is equipped with a batch normalization layer and ReLU activation function; a dense connection structure is used to merge the output of each convolutional layer with the output of all previous convolutional layers to enhance feature transfer and utilization efficiency; Finally, the spatial attention layer is introduced. After the spatial features are extracted and aggregated through the maximum pooling layer, average pooling layer and convolution layer, the attention weight map is generated through the convolution layer and Sigmoid activation. The attention weight map is multiplied element by element with the features of the input spatial attention layer to highlight the key features, refine the model's attention to important areas in the image, and optimize detail recovery and image quality. In this way, the main network and the branch network learn the structural features and detail information of the projection difference respectively; finally, the projection difference output by the projection residual learning network is added to the completed projection output by the projection completion network to obtain a high-precision completed projection; (3) Building a fault refinement network: The fault refinement network adopts the same two-way parallel encoder-decoder architecture as the projection residual learning network. The main network and the branch network learn the structural characteristics and high-frequency details of the fault respectively. The input of the fault refinement network is the fault obtained by FDK reconstruction of the high-precision completed projection. The number of input channels is the number of reconstructed fault layers, 128, and the network supervision is the true value of the fault. Step 3: Train a dual-domain supervised dual-view CT reconstruction network The dual-domain supervised dual-view CT reconstruction network consists of three parts: a projection completion network, a projection residual learning network, and a tomographic refinement network. The training sequence is as follows: first, the projection completion network is trained separately to obtain a preliminary completed projection; then, the parameters of the projection completion network are fixed, and the projection residual learning network is trained. The obtained projection residual is added to the preliminary completed projection to obtain an accurate projection completion result; finally, the parameters of the projection completion network and the projection residual learning network are fixed, and the accurate projection completion result is reconstructed by FDK as the input of the tomographic refinement network, and the tomographic refinement network is trained to obtain a fine tomographic image; (1) Training the projection completion network: From the data set established in step 1, select 0° and 90° in the true projection value as the input of the projection completion network. The 0-180° projection predicted by the projection completion network is named ; The network supervision is the 0-180° projection truth value, named ; The loss function used is composed of MSE loss and SSIM loss, as shown in formula (1): (1) Among them, , for The loss weight coefficient of Using the multiple sets of data established in step 1, repeat the above process of training the projection completion network; after training to a certain number of generations, Convergence, the projection completion network training is completed; at this time, the projection completion network parameters are fixed for the next step of training; (2) Training the projection residual learning network: The 0-180° projection predicted by the projection completion network is used as input, and the 0-180° projection difference predicted by the main path of the projection residual learning network is named , the main supervision of the projection residual learning network is the difference between the completed projection and the true value projection = - , the high-frequency quantity of the 0-180° projection difference predicted by the network branch is named , the branch supervision of the projection residual learning network is High frequency ,in, For extraction The loss function used for high-frequency functions is MSE loss, as shown in formula (2): (2) Training to After convergence, the projection residual learning network parameters are fixed for the next step of training; (3) Training the fault refinement network: Taking the fault results reconstructed by FDK as input, the faults predicted by the main path of the fault refinement network are named , the main network supervision is the fault truth value ; The high-frequency fault quantity predicted by the network branch is named , the network branch supervision is the high frequency quantity of the fault truth value ; In order to ensure the accuracy of the predicted fault, a reprojection loss function is designed; for the stereo predicted by the network, it is projected at 1° intervals to obtain its 90-270° projection data , and the true projection value of 90-270° obtained in step 1 Supervision; the definition of reprojection loss is shown in formula (3): (3) Among them, , for The loss weight coefficient of the fault refinement network is composed of MSE loss and reprojection loss. The composition is shown in formula (4): (4) Among them, , , for The loss weight coefficient is When converged, the dual-domain supervised dual-view CT reconstruction network training is completed.
