An extremely sparse view CT reconstruction method based on end-to-end neural network

Abstract

The present invention belongs to the field of medical image processing, and discloses a sparse angle computed tomography (CT) reconstruction method based on an end-to-end neural network, which is used to solve the problem that traditional reconstruction methods cannot reconstruct clear CT images under extremely sparse sampling viewing angle conditions. The end-to-end CT reconstruction method of the present invention first interpolates the sparse viewing angle sinogram A acquired by the CT device using bilinear interpolation to obtain a preliminarily restored full viewing angle sinogram B; then inputs it into a trained sinogram recovery network to obtain an optimized sinogram C; then uses a filtered back-projection algorithm (Filtered Back-Projection, FBP) to preliminarily reconstruct the sinogram C into a CT image D; finally, a trained dual-stream feature fusion recovery network is used to optimize the reconstructed CT image to obtain a high-quality CT image. The method of the present invention can directly reconstruct the sparse viewing angle sinogram, and the reconstructed CT image has higher structural similarity and clarity.

Description

Technical Field

[0001] The present invention relates to a CT image reconstruction method, and in particular to an extreme sparse viewing angle CT reconstruction method based on an end-to-end neural network. Background Art

[0002] CT is an imaging method that uses X-rays, gamma rays, etc. to perform cross-sectional scanning of objects. Thanks to the ability of X-rays, gamma rays, etc. to penetrate objects, CT imaging can directly obtain image information of internal tissues of the human body without actual trauma. It has the characteristics of fast scanning time, clear imaging and visualization of the internal objects, and is widely used in clinical medical diagnosis and industrial non-destructive testing. However, in actual situations, when you want to obtain high-definition CT images, you usually need to use repeated or high-dose ray light sources to scan objects, which will cause patients to be exposed to X-rays for a long time, which may cause metabolic abnormalities and even induce cancer. One of the effective ways to solve the above problem is to reduce the number of viewing angles of the CT device to scan around and perform sparse viewing angle CT reconstruction. However, when the rotating viewing angle of the light source when scanning the object does not reach the full viewing angle required for its reconstruction, the reconstructed CT image will produce stripe artifacts, and the stripe artifacts generated will become more serious as the viewing angle decreases. At present, the sparse viewing angle CT reconstruction algorithms that have been widely studied can be divided into traditional iterative reconstruction algorithms based on compressed sensing and reconstruction algorithms based on deep learning. The iterative reconstruction algorithm needs to reconstruct multiple projection and back-projection processes, which greatly increases the algorithm's computational cost and time. In addition, the regularization terms and balance parameters of its constraints need to be selected through a large number of manual experiments, which greatly limits its practical application. Reconstruction algorithms based on deep learning usually only use traditional convolution modules to extract local features of images, which is difficult to better remove global artifact information, or use convolutional neural networks to replace the iterative solution update process of some sub-problems in iterative reconstruction to speed up the solution. However, in essence, it still needs to establish multiple projection and back-projection processes, and it is still inevitable to have huge computational time complexity. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of existing technologies and provide an extremely sparse perspective CT reconstruction method based on an end-to-end neural network. The CT reconstruction algorithm can directly reconstruct the sinusoidal graph of extremely sparse perspective, and can reconstruct more global detail information and edge information while having better real-time performance.

[0004] The technical solution of the present invention to solve the above technical problems is:

[0005] An extremely sparse view CT reconstruction method based on an end-to-end neural network comprises the following steps:

[0006] S1, using the detector in the CT imaging system to rotate at the same angle, uniformly undersample and collect the corresponding extremely sparse viewing angle sinusoidal graph into the image set A, and collect the corresponding full viewing angle sinusoidal graph into the image set After;

[0007] S2, using a bilinear interpolation method to interpolate and fill in the data of the missing viewing angles, to obtain an initial full-view sinusoidal image set B;

[0008] S3. Put the images in the image set B into the trained sinusoidal image restoration network. First, perform preliminary feature extraction through the shallow feature extraction layer. The shallow feature extraction layer is composed of a CBR convolution block with a convolution kernel size of 3×3. The CBR convolution block contains a convolution operation, a batch normalization layer, and a ReLU activation function. Then enter the deep feature extraction stage, use the fence-type Transformer structure for advanced feature modeling, use three depth-separable convolutions to calculate the Q, K, and V values ​​of the row and column staggered axis area in the input feature map, and combine multi-head attention and softmax functions to calculate the adaptive weights in the corresponding area of ​​the feature map. Finally, use the sub-pixel convolution upsampling layer for resolution restoration. The sub-pixel convolution upsampling layer is composed of a 3×3 CBR convolution block and a pixel random shuffle layer. The CBR convolution block expands the number of channels of the feature map, and the pixel random shuffle layer randomly assigns the feature map of the expanded channel to the spatial dimension to expand the spatial resolution of the feature map. Finally, the optimized sinusoidal image set C is output;

[0009] S4, based on the differentiable FBP algorithm built in Pytorch, the optimized sinusoidal images in the image set C are initially reconstructed into the CT image set D;

[0010] S5. Put the CT image set D into the trained two-stream feature fusion restoration network. The two-stream feature fusion restoration network is composed of an image restoration subnetwork, an edge enhancement subnetwork and a feature fusion layer. The feature fusion layer fuses the feature maps F1 and F2 output by the two subnetworks, and performs channel dimension reduction on the fused feature map F through convolution operation to obtain the optimized CT image.

[0011] Preferably, in steps S3 and S5, the training of the sinusoidal graph restoration network and the dual-stream feature fusion restoration network includes the following steps:

[0012] S5-1, using the FBP algorithm to reconstruct the full-view sinusoidal image in the image set After into a CT image, and put it into the image set Label;

[0013] S5-2, build a dual-stream feature fusion restoration network for improving the quality of CT images, and then combine the bilinear interpolation method in S2, the sinusoidal image restoration network in S3 and the FBP algorithm in S4 to form a dual-domain end-to-end neural network, where the network in the projection domain is the sinusoidal image restoration network, and the network in the CT image domain is the dual-stream feature fusion restoration network;

[0014] S5-3. Use the extremely sparse perspective sinusoidal image in image set A as the input of the dual-domain end-to-end neural network, calculate the loss function value between the output of the sinusoidal image restoration network in the projection domain and the image set After, and simultaneously calculate the loss function value between the CT image finally output by the dual-stream feature fusion restoration network and the image set Label. Backpropagate the loss functions of both to update the network weights and parameters. The trained optimizer is Adam. When the set training rounds are reached, stop updating the weights and bias parameters of the dual-domain network to obtain a fully trained sinusoidal image restoration network and a dual-stream feature fusion restoration network.

[0015] Preferably, in step S5-3, the number of extremely sparse viewing angles may be 1 / 10, 1 / 30, or 1 / 60 of the 180 full viewing angles required in parallel X-ray CT reconstruction.

[0016] Preferably, in step S5, the image restoration subnetwork has the same structure as the sinusoidal image restoration network in S3; the edge enhancement subnetwork includes a Laplacian edge weighted convolution block, a multi-scale convolution block, a channel attention layer, a downsampling 2×2 maximum pooling layer, and a sub-pixel convolution upsampling layer; wherein the Laplacian edge weighted convolution block is composed of a convolution layer with a convolution kernel size of 3×3 and an initial weight of a Laplacian edge detection operator and a pixel weighted point multiplication layer, and the weighted edge information is obtained by pixel-by-pixel multiplication of the output edge features of the convolution with the original input initial reconstructed CT image; the multi-scale convolution block is The feature map is convolved with convolution layers of different sizes to obtain feature information of different scales, and the convolution kernel sizes include 1×1, 3×3 and 5×5; the channel attention layer is composed of an adaptive maximum pooling layer, a convolution layer with a convolution kernel size of 3×3, a Sigmoid activation function and a pixel weighted multiplication layer; the sub-pixel convolution upsampling layer is composed of a convolution layer with a convolution kernel size of 3×3, a batch normalization layer, a ReLU activation function and a pixel random shuffle layer; the number of downsampling channels of the edge enhancement subnetwork is 32, 64, 128, 256, and 512, respectively, and the number of upsampling channels is opposite to that of downsampling.

[0017] Preferably, in step S5-3, the loss function of the training network includes four functions, which are as follows: sino It is composed of mean square error loss function; dual domain feature fusion network loss function Loss imgIt consists of the mean square error loss function, the multi-scale structural similarity loss function and the edge loss function based on the Gaussian Laplace operator; the multi-scale structural similarity loss function is:

[0018]

[0019] where ε p ,ε g Respectively represent the mean of the reconstructed CT image and the labeled CT image, σ p ,σ g Respectively represent the standard deviation of the reconstructed CT image and the labeled CT image, σ pg It represents the covariance between the two, β m ,γ m The parameter is used to balance the proportion of the two parameters, c 1 ,c 2 is a constant to prevent the equation from being divided by 0; the edge loss function based on the Gaussian Laplace operator is:

[0020]

[0021] Where n is the total number of pixels, the subscript i is the pixel position of the convolution output image; * is a two-dimensional discrete convolution operation; f(x,y), f gt (x, y) are the dual-stream feature fusion recovery network output and the CT label image respectively;

[0022]

[0023] ΔG σ (x, y) is the Gaussian kernel of the second-order derivative, that is, the mean of the L1 norm between the feature maps after the convolution of the two is calculated as the loss function value; the mean square error loss function is:

[0024]

[0025] The total loss function of the final network is:

[0026]

[0027] That is, the network loss function Loss is restored by the sinusoidal graph sino And the dual domain feature fusion network loss function Loss img Together constitute.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The extremely sparse view CT reconstruction method based on an end-to-end neural network of the present invention introduces the self-attention mechanism in the Transformer into the dual-stream feature fusion restoration network for the sinusoidal image restoration network and the CT image restoration. Compared with the previous CT reconstruction algorithm that only uses traditional convolution to extract local feature information in the image, the method of the present invention combines the convolution block and the Transformer enhanced network to simultaneously extract global and local features, so as to capture richer relationships between upper and lower feature blocks, and improve the network's ability to suppress global artifacts, secondary artifacts and restore detailed structures.

[0030] 2. The extremely sparse view CT reconstruction method based on an end-to-end neural network of the present invention constructs an edge weighted convolution module based on the Laplacian edge detection operator in a dual-stream feature fusion recovery network, and further optimizes the weighted edge feature information obtained by combining multi-scale convolution blocks and channel attention mechanisms. Finally, the feature information output by the image recovery subnetwork combined with the Transformer is fused in dimension. The network enhances the recovery of its edge structure information while restoring the main structure, thereby further improving the visual perception of the image.

[0031] 3. In addition to using the mean square error loss function when constructing the loss function, the extreme sparse view CT reconstruction method based on the end-to-end neural network of the present invention also utilizes a multi-scale loss function to constrain the reconstructed image and constructs an edge loss function based on the Gaussian Laplace operator to further improve the reconstructed edge effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of the extremely sparse perspective CT reconstruction method based on an end-to-end neural network of the present invention.

[0033] Figure 2 The invention discloses a dual-stream feature fusion restoration network in the extremely sparse perspective CT reconstruction method based on an end-to-end neural network, comprising an edge enhancement subnetwork and an image restoration subnetwork.

[0034] Figure 3 The figure is a flowchart of the training and specific test execution case of the extremely sparse perspective CT reconstruction method based on the end-to-end neural network of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below in conjunction with examples, but the embodiments of the present invention are not limited thereto.

[0036] See also Figure 1 The extremely sparse viewing angle CT reconstruction method based on an end-to-end neural network of the present invention comprises the following steps:

[0037] S1, using the detector in the CT imaging system to rotate at the same angle, uniformly undersample and collect the corresponding extremely sparse viewing angle sinusoidal graph into the image set A, and collect the corresponding full viewing angle sinusoidal graph into the image set After;

[0038] S2, using a bilinear interpolation method to interpolate and fill in the data of the missing viewing angles, to obtain an initial full-view sinusoidal image set B;

[0039] S3. Put the images in the image set B into the trained sinusoidal image restoration network. First, perform preliminary feature extraction through the shallow feature extraction layer. The shallow feature extraction layer is composed of a CBR convolution block with a convolution kernel size of 3×3. The CBR convolution block contains a convolution operation, a batch normalization layer, and a ReLU activation function. Then enter the deep feature extraction stage, use the fence-type Transformer structure for advanced feature modeling, use three depth-separable convolutions to calculate the Q, K, and V values ​​of the row and column staggered axis area in the input feature map, and combine multi-head attention and softmax functions to calculate the adaptive weights in the corresponding area of ​​the feature map. Finally, use the sub-pixel convolution upsampling layer for resolution restoration. The sub-pixel convolution upsampling layer is composed of a 3×3 CBR convolution block and a pixel random shuffle layer. The CBR convolution block expands the number of channels of the feature map, and the pixel random shuffle layer randomly assigns the feature map of the expanded channel to the spatial dimension to expand the spatial resolution of the feature map. Finally, the optimized sinusoidal image set C is output;

[0040] S4, based on the differentiable FBP algorithm built in Pytorch, the optimized sinusoidal images in the image set C are initially reconstructed into the CT image set D;

[0041] S5. Put the CT image set D into the trained two-stream feature fusion restoration network. The two-stream feature fusion restoration network is composed of an image restoration subnetwork, an edge enhancement subnetwork and a feature fusion layer. The feature fusion layer fuses the feature maps F1 and F2 output by the two subnetworks, and performs channel dimension reduction on the fused feature map F through convolution operation to obtain the optimized CT image.

[0042] See also Figure 1 In steps S3 and S5, the training of the dual-stream feature fusion recovery network and the sinusoidal graph recovery network includes the following steps:

[0043] S5-1, using the FBP algorithm to reconstruct the full-view sinusoidal image in the image set After into a CT image, and put it into the image set Label;

[0044] S5-2, build a dual-stream feature fusion restoration network for improving the quality of CT images, and then combine the bilinear interpolation method in S2, the sinusoidal image restoration network in S3 and the FBP algorithm in S4 to form a dual-domain end-to-end neural network, where the network in the projection domain is the sinusoidal image restoration network, and the network in the CT image domain is the dual-stream feature fusion restoration network;

[0045] S5-3. Use the extremely sparse perspective sinusoidal image in image set A as the input of the dual-domain end-to-end neural network, calculate the loss function value between the output of the sinusoidal image restoration network in the projection domain and the image set After, and simultaneously calculate the loss function value between the CT image finally output by the dual-stream feature fusion restoration network and the image set Label. Backpropagate the loss functions of both to update the network weights and parameters. The trained optimizer is Adam. When the set training rounds are reached, stop updating the weights and bias parameters of the dual-domain network to obtain a fully trained sinusoidal image restoration network and a dual-stream feature fusion restoration network.

[0046] When reconstructing CT images under extremely sparse viewing conditions, the resulting images are almost left with only interlaced global stripe artifact information. When faced with feature information with global characteristics and some smaller weak features, the traditional fully convolutional network used for image restoration has difficulty representing these features. The global artifact features exceed the receptive field of its own network convolutional block, and thus cannot effectively suppress global artifacts and secondary artifacts in the image, resulting in poor structural similarity of the restored image, low clarity, and the image still carries weak stripe artifacts. At the same time, the traditional fully convolutional neural network lacks attention to image edge information when used for CT reconstruction, and it is difficult for the traditional convolution module to extract only the edge feature information of the image, resulting in the reconstructed CT image edge effect is not strong.

[0047] For this purpose, see Figure 2The image restoration subnetwork has the same structure as the sinusoidal image restoration network in S3; the edge enhancement subnetwork includes a Laplacian edge weighted convolution block Laplacian_Conv, a multi-scale convolution block MSC, a channel attention layer SE, a downsampling 2×2 maximum pooling layer, and a sub-pixel convolution upsampling layer; wherein the Laplacian edge weighted convolution block consists of a convolution layer with a convolution kernel size of 3×3 and an initial weight of the Laplacian edge detection operator and a pixel weighted point multiplication layer, and the weighted edge information is obtained by pixel-by-pixel multiplication of the output edge features of the convolution with the original input initial reconstructed CT image; the multi-scale convolution block The convolution block is realized by convolving the feature map with convolution layers of different sizes to obtain feature information of different scales, and the convolution kernel sizes include 1×1, 3×3 and 5×5; the channel attention layer is composed of an adaptive maximum pooling layer, a convolution layer with a convolution kernel size of 3×3, a Sigmoid activation function and a pixel weighted dot product layer; the sub-pixel convolution upsampling layer is composed of a convolution layer with a convolution kernel size of 3×3, a batch normalization layer, a ReLU activation function and a pixel random shuffle layer; the number of downsampling channels of the edge enhancement subnetwork is 32, 64, 128, 256, and 512, respectively, and the number of upsampling channels is opposite to that of downsampling.

[0048] Through the above settings, the global characteristics of the main image of the Transformer associated feature map are utilized, and the Laplacian convolution block is combined with the multi-scale channel attention module to further enhance and fuse the semantic edge information of different scales, which can further enhance the recovery of the edge structure based on the restored image.

[0049] In addition, the loss functions of the training network include 4, which are as follows: Sine graph recovery network loss function Loss sino It is composed of mean square error loss function; dual domain feature fusion network loss function Loss img It consists of the mean square error loss function, the multi-scale structural similarity loss function and the edge loss function based on the Gaussian Laplace operator; the multi-scale structural similarity loss function is:

[0050]

[0051] where ε p ,ε g Respectively represent the mean of the reconstructed CT image and the labeled CT image, σ p ,σ g Respectively represent the standard deviation of the reconstructed CT image and the labeled CT image, σ pg It represents the covariance between the two, β m ,γ m The parameter is used to balance the proportion of the two parameters, c 1 ,c 2is a constant to prevent the equation from being divided by 0; the edge loss function based on the Gaussian Laplace operator is:

[0052]

[0053] Where n is the total number of pixels, the subscript i is the pixel position of the convolution output image; * is a two-dimensional discrete convolution operation; f(x,y), f gt (x, y) are the dual-stream feature fusion recovery network output and the CT label image respectively;

[0054]

[0055] ΔG σ (x, y) is the Gaussian kernel of the second-order derivative, that is, the mean of the L1 norm between the feature maps after the convolution of the two is calculated as the loss function value; the mean square error loss function is:

[0056]

[0057] The total loss function of the final network is:

[0058]

[0059] That is, the network loss function Loss is restored by the sinusoidal graph sino And the dual domain feature fusion network loss function Loss img Together constitute.

[0060] See also Figure 3 The CT reconstruction method of the present invention is described below with a specific case:

[0061] First, the sparse view sinusoidal data of the corresponding angle of the object is collected by the CT device, and the sparse view sinusoidal data is placed in the image set A; the bilinear interpolation method is used to interpolate the data of the missing view of the image in the image set A, and the interpolated sinusoidal data is placed in the image set B; then the images in the image set B are placed in the trained sinusoidal recovery network to obtain the optimized sinusoidal data image set C; then the FBP algorithm is used to reconstruct the sinusoidal in the image set C to obtain the CT image set D; finally, the initial reconstructed CT image in the image set D is placed in the trained two-stream feature fusion recovery network to obtain the optimized CT image; and the training process is based on the sinusoidal loss Loss sino and CT image loss img Jointly train dual-domain networks.

[0062] The above is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An extremely sparse view CT reconstruction method based on an end-to-end neural network, It is characterized in that The following steps are involved: S1, using the detector in the CT imaging system to rotate at the same angle, uniformly undersample and collect the corresponding extremely sparse viewing angle sinusoidal graph into the image set A, and collect the corresponding full viewing angle sinusoidal graph into the image set After; S2, using a bilinear interpolation method to interpolate and fill in the data of the missing viewing angles, to obtain an initial full-view sinusoidal image set B; S3. Put the images in the image set B into the trained sinusoidal image restoration network. First, perform preliminary feature extraction through the shallow feature extraction layer. The shallow feature extraction layer is composed of a CBR convolution block with a convolution kernel size of 3×3. The CBR convolution block contains a convolution operation, a batch normalization layer, and a ReLU activation function. Then enter the deep feature extraction stage, use the fence-type Transformer structure for advanced feature modeling, use three depth-separable convolutions to calculate the Q, K, and V values ​​of the row and column staggered axis area in the input feature map, and combine multi-head attention and softmax functions to calculate the adaptive weights in the corresponding area of ​​the feature map. Finally, use the sub-pixel convolution upsampling layer for resolution restoration. The sub-pixel convolution upsampling layer is composed of a 3×3 CBR convolution block and a pixel random shuffle layer. The CBR convolution block expands the number of channels of the feature map, and the pixel random shuffle layer randomly assigns the feature map of the expanded channel to the spatial dimension to expand the spatial resolution of the feature map. Finally, the optimized sinusoidal image set C is output; S4, based on the differentiable FBP algorithm built in Pytorch, the optimized sinusoidal images in the image set C are initially reconstructed into the CT image set D; S5. Put the CT image set D into the trained two-stream feature fusion restoration network. The two-stream feature fusion restoration network is composed of an image restoration subnetwork, an edge enhancement subnetwork and a feature fusion layer. The feature fusion layer fuses the feature maps F1 and F2 output by the two subnetworks, and performs channel dimension reduction on the fused feature map F through convolution operation to obtain the optimized CT image.

2. The extremely sparse view CT reconstruction method based on end-to-end neural network according to claim 1, It is characterized in that In steps S3 and S5, the training of the sinusoidal graph restoration network and the dual-stream feature fusion restoration network includes the following steps: S5-1, using the FBP algorithm to reconstruct the full-view sinusoidal image in the image set After into a CT image, and put it into the image set Label; S5-2, build a dual-stream feature fusion restoration network for improving the quality of CT images, and then combine the bilinear interpolation method in S2, the sinusoidal image restoration network in S3 and the FBP algorithm in S4 to form a dual-domain end-to-end neural network, where the network in the projection domain is the sinusoidal image restoration network, and the network in the CT image domain is the dual-stream feature fusion restoration network; S5-3. Use the extremely sparse perspective sinusoidal image in image set A as the input of the dual-domain end-to-end neural network, calculate the loss function value between the output of the sinusoidal image restoration network in the projection domain and the image set After, and simultaneously calculate the loss function value between the CT image finally output by the dual-stream feature fusion restoration network and the image set Label. Backpropagate the loss functions of both to update the network weights and parameters. The trained optimizer is Adam. When the set training rounds are reached, stop updating the weights and bias parameters of the dual-domain network to obtain a fully trained sinusoidal image restoration network and a dual-stream feature fusion restoration network.

3. The extremely sparse view CT reconstruction method based on end-to-end neural network according to claim 2, It is characterized in that In step S5-3, the number of extremely sparse viewing angles may be 1 / 10, 1 / 30, or 1 / 60 of the 180 full viewing angles required in parallel X-ray CT reconstruction.

4. The extreme sparse view CT reconstruction method based on end-to-end neural network according to claim 1, It is characterized in that In step S5, the image restoration subnetwork has the same structure as the sinusoidal image restoration network in S3; the edge enhancement subnetwork includes a Laplacian edge weighted convolution block, a multi-scale convolution block, a channel attention layer, a downsampling 2×2 maximum pooling layer, and a sub-pixel convolution upsampling layer; wherein the Laplacian edge weighted convolution block is composed of a convolution layer with a convolution kernel size of 3×3 and an initial weight of the Laplacian edge detection operator and a pixel weighted point multiplication layer, and the weighted edge information is obtained by pixel-by-pixel multiplication of the output edge features of the convolution with the original input initial reconstructed CT image; the multi-scale convolution block is obtained by different The invention realizes the feature map convolution by using convolution layers of different sizes to obtain feature information of different scales, and the convolution kernel sizes include 1×1, 3×3 and 5×5; the channel attention layer is composed of an adaptive maximum pooling layer, a convolution layer with a convolution kernel size of 3×3, a Sigmoid activation function and a pixel weighted multiplication layer; the sub-pixel convolution upsampling layer is composed of a convolution layer with a convolution kernel size of 3×3, a batch normalization layer, a ReLU activation function and a pixel random shuffle layer; the number of downsampling channels of the edge enhancement subnetwork is 32, 64, 128, 256 and 512 respectively, and the number of upsampling channels is opposite to that of downsampling.

5. The extremely sparse view CT reconstruction method based on end-to-end neural network according to claim 1, It is characterized in that In step S5, the feature fusion layer includes a CBR convolution block and a 1×1 convolution block.

6. The extreme sparse view CT reconstruction method based on end-to-end neural network according to claim 2, It is characterized in that The loss functions of the training network include 4, which are as follows: Sine graph recovery network loss function Loss sino It is composed of mean square error loss function; dual domain feature fusion network loss function Loss img It consists of the mean square error loss function, the multi-scale structural similarity loss function and the edge loss function based on the Gaussian Laplace operator; the multi-scale structural similarity loss function is: where ε p ,ε g Respectively represent the mean of the reconstructed CT image and the labeled CT image, σ p ,σ g Respectively represent the standard deviation of the reconstructed CT image and the labeled CT image, σ pg It represents the covariance between the two, β m ,γ m The parameter is used to balance the proportion of the two parameters, c 1 ,c 2 is a constant to prevent the equation from being divided by 0; the edge loss function based on the Gaussian Laplace operator is: Where n is the total number of pixels, the subscript i is the pixel position of the convolution output image; * is a two-dimensional discrete convolution operation; f(x,y), f gt (x, y) are the dual-stream feature fusion recovery network output and the CT label image respectively; ΔG σ (x, y) is the Gaussian kernel of the second-order derivative, that is, the mean of the L1 norm between the feature maps after the convolution of the two is calculated as the loss function value; the mean square error loss function is: The total loss function of the final network is: That is, the network loss function Loss is restored by the sinusoidal graph sino And the dual domain feature fusion network loss function Loss img Together constitute.

Images