A method, apparatus and system for reconstructing full-angle cone-beam CT projection images

Abstract

The application discloses a kind of full-angle cone beam CT projection image's recovery method, device and system, belong to medical image processing technical field, the method includes: the pixel matrix corresponding to sparse angle cone beam CT projection image is normalized;Normalized image is input 3D-2D channel attention interpolation network, utilizes the 3D feature extractor therein to extract the spatial feature of normalized image and outputs 3D feature map;The spatial feature and channel dimension of 3D feature map are merged using the inter-projection fusion module therein, and 3D feature map is reshaped into 2D feature map;2D projection generator therein is used to recover the missing projection corresponding to 2D feature map to obtain full-angle projection image;Full-angle projection image is reconstructed using FDK algorithm, and high-quality cone beam CT image is recovered;The application can remove reconstruction image artifact while saving fine structure content in image, to ensure the consistency of projection domain and reconstruction domain.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image processing, and more particularly relates to a full-angle cone-beam CT projection image restoration method, device and system. BACKGROUND

[0002] Cone-beam CT (CBCT) is an important medical imaging technology, which is widely used in cancer diagnosis, image-guided surgery, etc. CBCT is an imaging technology that directly obtains a three-dimensional medical image. The reconstruction principle of CBCT is to first collect hundreds of 2D X-ray projections at a fixed angular interval within a certain angle, such as 360 degrees, and then use the FDK algorithm to reconstruct a 3D CT image from the collected hundreds of 2D projections. In order to reduce the radiation dose of CBCT, the CT scanning interval can be increased, that is, the scanning period is kept unchanged, the scanning angle interval is increased, and the number of acquired projections is reduced. Using this way to scan and reconstruct is sparse-angle cone-beam CT reconstruction. However, sparse-angle cone-beam CT will make the acquired sparse projections incomplete, and the image reconstructed by directly using the traditional reconstruction algorithm such as FDK algorithm will have serious strip-like artifacts and the image details will be seriously blurred. These artifacts and blurring will affect the doctor's diagnosis and may cause serious medical accidents. Therefore, how to improve the sparse-angle cone-beam CT reconstruction image quality is a hot spot in current medical image research.

[0003] At present, the methods for improving the sparse-angle cone-beam CT image quality mainly include: based on the reconstruction domain and based on the projection domain. The method based on the reconstruction domain directly processes the reconstructed CBCT image, which has fast processing speed, is easy to integrate into the CT system, and is widely applied; however, it is not deeply integrated with the CT imaging system, and cannot utilize the rich original information of the projection domain image to optimize the reconstruction image quality. The method based on the projection domain restores the full-angle projection before reconstruction, which has great potential to remove artifacts while preserving image structure, and the quality of the reconstructed image depends on whether the projection domain interpolation algorithm can restore relatively accurate projection images.

[0004] At present, researchers have proposed using deep learning networks to process sparse angle projections to restore full angle projections, including deep learning networks such as HDNet (Hybrid-Domain Network), DualCNN (Dual Convolutional Neural Network), HDNet first restores full angle projections using a traditional linear interpolation algorithm and then corrects the restored projections using a network; DualCNN directly interpolates full angle projections using a network. However, these interpolation methods still have some problems. First, HDNet and DualCNN do not consider the consistency of the projection domain and the reconstruction domain in restoring full angle projections, that is, only the projection domain information is focused on when restoring the projections, and the reconstruction information is not used to guide the restoration of the projections. Second, the current interpolation network uses pixel-level loss, which can easily lead to over-smoothing of the projection image, resulting in loss of important detail structures in the reconstructed image. Third, the current network does not fully utilize important spatial information between and within projection images, and cannot effectively restore relatively accurate projections. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a full-angle cone-beam CT projection image restoration method, device and system, which aims to propose a 3D-2D channel attention interpolation network guided by FDK reconstruction, which combines the FDK algorithm to guide the network to focus on the projection information required in the reconstruction process, removes the artifacts in the reconstructed image while preserving the fine structure content in the image, thereby ensuring the consistency of the projection domain and the reconstruction domain, thereby solving the technical problem of low efficiency of the existing cone-beam CT projection image restoration.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present application, a full-angle cone-beam CT projection image restoration method is provided, comprising:

[0007] S1: normalizing the pixel matrix corresponding to the sparse angle cone-beam CT projection image to obtain a normalized image;

[0008] S2: inputting the normalized image into a 3D-2D channel attention interpolation network, which includes a 3D feature extractor, an inter-projection fusion module and a 2D projection generator connected in turn; using the 3D feature extractor to extract the spatial features of the normalized image and output a 3D feature map;

[0009] S3: using the inter-projection fusion module to merge the spatial features and channel dimensions of the 3D feature map, thereby reshaping the 3D feature map into a 2D feature map;

[0010] S4: using the 2D projection generator to restore the missing projections corresponding to the 2D feature map to obtain a full angle projection image;

[0011] S5: reconstructing the full-angle projection image by using the FDK algorithm to obtain a high-quality cone beam CT image.

[0012] In one of the embodiments, the S1 comprises:

[0013] using the formula normalizing the pixel matrix corresponding to the sparse-angle cone beam CT projection image;

[0014] wherein P' n is the nth pixel value in the pixel matrix corresponding to the sparse-angle cone beam CT projection image, P' n is the nth pixel value in the pixel matrix corresponding to the normalized image, P' max is the maximum pixel value in the pixel matrix corresponding to the sparse-angle cone beam CT projection image; and P' min is the minimum pixel value in the pixel matrix corresponding to the sparse-angle cone beam CT projection image.

[0015] In one of the embodiments, the 3D feature extractor is an encoder-decoder architecture, the encoder of which comprises multiple three-dimensional convolution units including three-dimensional convolution layers and Leaky ReLU activation functions; and the decoder of which comprises multiple three-dimensional deconvolution units including three-dimensional deconvolution layers and Leaky ReLU activation functions.

[0016] The 2D projection generator is an encoder-decoder architecture, the encoder of which comprises multiple two-dimensional convolution units including two-dimensional convolution layers and Leaky ReLU activation functions; and the decoder of which comprises multiple two-dimensional deconvolution units including two-dimensional deconvolution layers and Leaky ReLU activation functions.

[0017] Each layer of the 3D feature extractor and the 2D projection generator is embedded with a channel attention module to capture the dependency between the input feature map channels.

[0018] In one of the embodiments, the encoder of the 3D feature extractor comprises five three-dimensional convolution units; and the decoder of the 3D feature extractor comprises four three-dimensional deconvolution units.

[0019] The encoder of the 2D projection generator comprises five two-dimensional convolution units; and the decoder of the 3D feature extractor comprises four two-dimensional deconvolution units.

[0020] In one of the embodiments, the high-quality cone beam CT image is represented as:

[0021]

[0022] Wherein, P represents the input sparse angle cone-beam CT projection image, f θ represents the 3D-2D channel attention interpolation network; AC 1-5 represents the three-dimensional convolution unit; BC 1-4 represents the three-dimensional deconvolution unit; DC 1-5 represents the two-dimensional convolution unit; EC 1-4 represents the two-dimensional deconvolution unit; IPF represents the inter-projection fusion module.

[0023] In one of the embodiments, the channel attention module is used to calculate the global mean of the input feature map along the channel to obtain the average feature map; the average feature map is extracted using the convolution layer and the Leaky ReLU activation function and integrated into a one-dimensional vector; the one-dimensional vector is activated using the Sigmoid function; the input feature map is multiplied with the attention map along the channel, and the output feature map is finally obtained; the output feature map can reflect the channel dependency of the input feature map.

[0024] In one of the embodiments, further comprising:

[0025] using a joint loss function training the deep learning network to obtain the 3D-2D channel attention interpolation network;

[0026]

[0027]

[0028]

[0029] Wherein, M is the number of interpolation projections, andrespectively represent the interpolation projection generated by the network and the real reference projection; RF(*) represents the frequency domain representation of the projection image after filtering; is used to reduce the pixel-level difference between the interpolation projection and the real reference projection, and is obtained by calculating the average absolute error between the projection images; is used to reduce the difference after filtering the frequency domain of the projection image.

[0030] In one of the embodiments, the frequency domain representation RF(*) of the projection image after filtering is calculated using the formula RF(P t )=|α(u,v)×F t (u,v)|;

[0031]

[0032] where H, W are the height and width of the projection image respectively, (x, y) is the spatial domain coordinate of the projection, P t (x, y) is the pixel value of the projection at (x, y); (u, v) is the frequency domain coordinate of the projection, F t (u, v) is the frequency domain complex value of the projection at (u, v); a(u, v) is the coefficient of the Ramp-Filter weight matrix at (u, v).

[0033] According to another aspect of the present application, there is provided a full-angle cone-beam CT projection image restoration device for performing the full-angle cone-beam CT projection image restoration method, comprising:

[0034] a normalization module configured to normalize the pixel matrix corresponding to the sparse-angle cone-beam CT projection image to obtain a normalized image;

[0035] an extraction module configured to input the normalized image into a 3D-2D channel attention interpolation network, which comprises, in sequence, a 3D feature extractor, an inter-projection fusion module and a 2D projection generator; the 3D feature extractor is configured to extract the spatial features of the normalized image and output a 3D feature map;

[0036] a merging module configured to merge the spatial features and channel dimensions of the 3D feature map using the inter-projection fusion module, so as to reshape the 3D feature map into a 2D feature map;

[0037] a recovery module configured to recover the missing projection corresponding to the 2D feature map using the 2D projection generator to obtain a full-angle projection image;

[0038] a reconstruction module configured to perform reconstruction processing on the full-angle projection image using an FDK algorithm to obtain a high-quality cone-beam CT image.

[0039] According to another aspect of the present application, there is provided a full-angle cone-beam CT projection image restoration system comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method when executing the computer program.

[0040] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects compared with the prior art:

[0041] (1) The application inputs the normalized image corresponding to the sparse angle cone beam CT projection pattern into the 3D-2D channel attention interpolation network, uses the 3D feature extractor therein to extract the spatial features of the normalized image and outputs the 3D feature map, then uses the inter-projection fusion module to combine the spatial features and channel dimensions of the 3D feature map, so as to reshape the 3D feature map into a 2D feature map, then uses the 2D projection generator to restore the missing projection corresponding to the 2D feature map, and obtains a full-angle projection image, finally combines the FDK algorithm guided network to pay attention to the projection information required in the reconstruction process, and finally obtains a high-quality cone beam CT image. The method of the application can restore a high-quality full-angle projection, remove the artifacts of the reconstructed image while preserving the fine structure content in the image, thereby ensuring the consistency of the projection domain and the reconstruction domain; in addition, the CBCT image reconstructed by the restored full-angle projection can effectively suppress the artifacts in the CT image and retain more CT image structures compared with other methods. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the full-angle cone beam CT projection image restoration method provided by the application;

[0043] Figure 2 is a structural schematic diagram of the FDK reconstruction guided 3D-2D channel attention interpolation network;

[0044] Figure 3 and Figure 4 is a comparison chart of the qualitative results of the present method and the prior art at 1 / 4 and 1 / 8 doses;

[0045] Figure 5 and Figure 6 is a comparison chart of the quantitative results and the region of interest at 1 / 4 dose of the present method and the prior art;

[0046] Figure 7 and Figure 8 is a comparison chart of the quantitative results and the region of interest at 1 / 8 dose of the present method and the prior art. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0048] As shown in Figure 1 , the present application provides a full-angle cone beam CT projection image restoration method, comprising:

[0049] S1: Normalize the pixel matrix corresponding to the sparse-angle cone beam CT projection image to obtain a normalized image;

[0050] S2: Input the normalized image into a 3D-2D channel attention interpolation network, which includes a 3D feature extractor, an inter-projection fusion module and a 2D projection generator connected in turn; use the 3D feature extractor to extract the spatial features of the normalized image and output a 3D feature map;

[0051] S3: Use the inter-projection fusion module to merge the spatial features and channel dimensions of the 3D feature map, so as to reshape the 3D feature map into a 2D feature map;

[0052] S4: Use the 2D projection generator to restore the missing projection corresponding to the 2D feature map to obtain a full-angle projection image;

[0053] S5: Use the FDK algorithm to reconstruct the full-angle projection image to obtain a high-quality cone beam CT image.

[0054] Each step will be described in detail as follows:

[0055] Step 1: Projection normalization: normalize the sparse-angle cone beam CT projection image matrix.

[0056] The sparse-angle CBCT projection image matrix is normalized, and the formula is as follows:

[0057]

[0058] where P' n is the nth pixel value of the CBCT projection image matrix, P n is the normalized pixel value, P' min and P' max are the minimum and maximum values of the pixel matrix, respectively. n n min max

[0059] Step 2: Projection feature extraction: 3D network based on channel attention mechanism extracts spatial features of input projection.

[0060] This step is projection feature extraction, that is, a 3D feature extractor is used to extract the spatial features of the input continuous sparse-angle projection. Considering the characteristics of the CT imaging system, the input continuous projection has a rotational motion relationship, and the use of a 3D convolution kernel can capture the motion change characteristics between projections. In order to measure the mutual dependence between feature channels, a channel attention module is embedded at each layer of the 3D feature extractor, allowing the 3D feature extractor to extract more useful features. The input of the 3D feature extractor is continuous sparse-angle projection, and the output is a 3D feature map.

[0061] ​​​​Step 3: Inter-projection feature fusion: merging the angle and channel dimensions of spatial features to reshape the 3D feature map into a 2D feature map.

[0062] This step is inter-projection feature fusion, that is, using an inter-projection fusion module to reshape the 3D feature map into a 2D feature map. The 3D feature map output by the 3D feature extractor contains rich spatial feature information. In order to make full use of these feature information, the fusion module merges the angle and channel dimensions of the feature map through the reshaping operation. The input of the inter-projection fusion module is a 3D feature map, and the output is a 2D feature map.

[0063] Step 4: Missing projection generation: a 2D network based on channel attention mechanism generates projections of missing angles.

[0064] This step is missing projection generation, that is, using a 2D projection generator to generate projection images of missing angles from a 2D feature map. In order to adapt to the need to output different numbers of projections under different dose conditions, the 2D projection generator can dynamically adjust the number of channels of the feature map to meet this need. For example, for 1 / 4 dose sparse projection, the number of missing projections output by the generator is 3, and for 1 / 8 dose, the number is 7. The input of the 2D projection generator is a 2D feature map, and the output is a projection image of a missing angle.

[0065] Figure 2 is a schematic diagram of the structure of the FDK reconstruction guided 3D-2D channel attention interpolation network. In another embodiment, the 3D feature extractor is an encoder-decoder architecture, the encoder of which is a plurality of three-dimensional convolution units including a three-dimensional convolution layer and a Leaky ReLU activation function; the decoder of which is a plurality of three-dimensional deconvolution units including a three-dimensional deconvolution layer and a Leaky ReLU activation function; the 2D projection generator is an encoder-decoder architecture, the encoder of which is a plurality of two-dimensional convolution units including a two-dimensional convolution layer and a Leaky ReLU activation function; the decoder of which is a plurality of two-dimensional deconvolution units including a two-dimensional deconvolution layer and a Leaky ReLU activation function; wherein each layer of the 3D feature extractor and the 2D projection generator is embedded with a channel attention module to capture the dependency between the channels of the input feature map.

[0066] Preferably, after the sparse angle continuous projection image is input into the network, a 3D feature extractor is first used to extract the feature information of the input projection. The 3D feature extractor is an encoder-decoder architecture. The encoder thereof is 5 three-dimensional convolution units (three-dimensional convolution layer + Leaky ReLU activation function), each unit being denoted as AC 1-5 ; the decoder thereof is 4 three-dimensional deconvolution units (three-dimensional deconvolution layer + Leaky ReLU activation function), each unit being denoted as BC 1-4). The 3D feature maps outputted by the 3D feature extractor are inputted into an inter-projection fusion module to transform into 2D feature maps, which is denoted as IPF. Then the transformed 2D feature maps are inputted into a 2D projection generator to recover the missing projections. The 2D projection generator is an encoder-decoder architecture. Its encoder is 5 two-dimensional convolution units (two-dimensional convolution layer + Leaky ReLU activation function), each unit is denoted as DC 1-5 ); its decoder is 4 two-dimensional deconvolution units (two-dimensional deconvolution layer + Leaky ReLu activation function), each unit is denoted as EC 1-4 .

[0067] The input sparse angle projection image is denoted as P, and the output interpolated projection image is The network interpolation process can be described by the formula as follows:

[0068]

[0069] In an embodiment, a channel attention module is embedded in each layer of the interpolation network to capture the dependencies between the channels of the feature maps. Inputting the extracted feature maps into the channel attention module can make full use of the global information of the feature maps. The process mainly includes the following steps. 1) Calculate the global mean of the input feature maps along the channel; 2) further extract the features of the average feature maps by a convolution layer and a LeakyReLU activation function, and then integrate them into a one-dimensional vector; 3) activate the one-dimensional vector by a Sigmoid function; 4) point multiply the original input feature maps and the attention map along the channel. The output feature maps obtained finally can focus on the channel dependencies of the input feature maps.

[0070] In an embodiment, the interpolation network in the present application is trained by using a joint loss function, which includes Loss and frequency domain filtering loss which is used to reduce the pixel-level difference between the interpolated projection and the reference projection, and is obtained by calculating the mean absolute error between the projection images. By reducing the difference between the projection images after frequency domain filtering, the interpolation network can focus on the projection image information that is beneficial to reconstruction. The joint loss and Loss and frequency domain filtering loss The specific calculation process is as follows:

[0071]

[0072]

[0073]

[0074] wherein M is the number of interpolated projections, and respectively represent the interpolated projection generated by the network and the real reference projection. RF(*) represents the filtered frequency domain representation of the projection image. To obtain the filtered frequency domain representation of the projection image, a two-dimensional fast Fourier transform is used to transform the projection into the frequency domain:

[0075]

[0076] where H, W are the height and width of the projection image respectively, (x, y) is the spatial domain coordinate of the projection, P t (x, y) is the pixel value of the projection at (x, y); (u, v) is the frequency domain coordinate of the projection, F t (u, v) is the frequency domain complex value of the projection at (u, v). In order to make the network pay attention to the projection information used for reconstruction, the Ramp-Filter filter operator in the FDK reconstruction is multiplied with the projection frequency domain transformation to enhance the high frequency information of the frequency domain transformation and suppress the low frequency information, the process is as follows:

[0077] RF(P t )=|α(u, v)×F t (u, v)|

[0078] where α(u, v) is the coefficient of the Ramp-Filter weight matrix at (u, v).

[0079] Step 5: Reconstruction: The recovered full-angle projection image is reconstructed into a cone beam CT image by using the FDK algorithm. Specifically, the full-angle projection image recovered by the interpolation network is input into the FDK algorithm to reconstruct a CBCT image.

[0080] The present application has the following three points as the biggest difference from the existing sparse angle cone beam projection recovery method:

[0081] First, considering that the image quality of the cone beam CT reconstruction depends on the filtered projection image information, the present application designs a frequency domain filtering loss function to make the interpolation network generate a projection image that is beneficial to reconstruction;

[0082] Second, the present application designs a 3D-2D interpolation network structure, which can capture the motion change relationship between the projections, and adaptively output the corresponding number of projections according to different dose levels.

[0083] Third, the present application proposes a combination mode of convolution layer and channel attention mechanism module.

[0084] The following is the experimental evaluation of the method of the present application:

[0085] The test is carried out on 1 / 4 and 1 / 8 dose sparse-angle walnut cone-beam CT data, wherein 42 cone-beam CT projection data of walnuts are contained, and the latest sparse-angle cone-beam CT processing method HDNet (Hybrid-Domain Network) and the projection domain interpolation method in DualCNN (Dual Convolutional Neural Network) are compared, so as to illustrate the effectiveness of the interpolation network guided by the FDK reconstruction of the method. The training data set of the comparative method is consistent with the training data set of the method. The following is a specific description of the training data.

[0086] The training data of the interpolation network come from the projection images of the 7th to 42nd walnut data, the number of full-angle projections of each walnut is 600, the number of sparse-angle projections of 1 / 4 and 1 / 8 dose is 150 and 75 respectively, and the projection image size is 972*768. In each iteration process, 4 sparse-acquired projection images are input, and for 1 / 4 dose, the interpolation network outputs 3 missing projections in the middle, and for 1 / 8 dose, the output is 7 missing projections in the middle.

[0087] The Adam optimizer is adopted in the training process, the learning rate is set to 2*10 -5 , and the total training number is 150 epochs.

[0088] The validation set comes from the projection images and reconstructed images of the 6th walnut, and the test set comes from the reconstructed images of the 1st to 5th walnuts, 200 reconstructed images of each walnut, and a total of 1000 images for quantification, Figure 3 and Figure 4 quantitative results of 1 / 4 and 1 / 8 dose are shown respectively. The 1st walnut is used to show the qualitative results, and the 1 / 4 and 1 / 8 dose results are shown in Figure 5 , Figure 6 and Figure 7 , Figure 8 respectively.

[0089] Figure 3 and Figure 4 quantify the root mean squared error (RMSE, Root Mean Squared Error), peak signal-to-noise ratio (PSNR, Peak signal-to-noise ratio) and structural similarity (SSIM, Structural Similarity) of the images generated by the method and the comparative method. Figure 5 and Figure 6 are the HDNet, DualCNN and the method for processing 1 / 4 dose sparse-angle cone-beam CT images and the corresponding regions of interest, Figure 7 and Figure 8Results for 1 / 8 dose. From Figure 3 and Figure 4 It can be seen that the method of the present application achieves the highest quantification values at both 1 / 4 and 1 / 8 dose. The qualitative and quantitative assessments are consistent, as shown by the black and white arrows in Figure 5-8 Compared with other methods, the method of the present application not only effectively suppresses the artifacts in sparse angular cone-beam CT images, but also restores some of the fine structures in the images, while preventing over-smoothing and making the images appear more natural.

[0090] Those skilled in the art will easily understand that the above description is only preferred embodiments of the present application and is not intended to limit the present application, and any modifications, equivalent replacements and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of reconstruction of full-angle cone-beam CT projection images, characterized by, The method comprises: S1: normalizing a pixel matrix corresponding to a sparse-angle cone-beam CT projection image to obtain a normalized image; S2: inputting the normalized image into a 3D-2D channel attention interpolation network, which comprises, connected in sequence: a 3D feature extractor, an inter-projection fusion module, and a 2D projection generator; extracting spatial features of the normalized image by using the 3D feature extractor and outputting a 3D feature map; S3: merging the spatial features and channel dimensions of the 3D feature map by using the inter-projection fusion module, so as to reshape the 3D feature map into a 2D feature map; S4: restoring a missing projection corresponding to the 2D feature map by using the 2D projection generator to obtain a full-angle projection image; S5: performing reconstruction processing on the full-angle projection image by using an FDK algorithm to obtain a high-quality cone-beam CT image; Utilizing a joint loss function training a deep learning network to obtain the 3D-2D channel attention interpolation network; where M is the number of interpolated projections, with respectively represent the network-generated interpolated projection and the ground truth reference projection; RF(*) represents the frequency domain representation of the filtered projection image; is used to reduce the pixel-level difference between the interpolated projection and the ground truth reference projection, which is obtained by calculating the mean absolute error between the projection images; is used to reduce the difference of the filtered projection images in the frequency domain.

2. The full-angle cone-beam CT back-projection image reconstruction method of claim 1, wherein, The S1 comprises: Using the formula Normalizing the pixel matrix corresponding to the sparse angular cone beam CT projection image; wherein P n is the nth pixel value in the pixel matrix corresponding to the sparse angle cone beam CT projection image, P n is the nth pixel value in the pixel matrix corresponding to the normalized image, P' max is the maximum pixel value in the pixel matrix corresponding to the sparse angle cone beam CT projection image; P' min is the minimum pixel value in the pixel matrix corresponding to the sparse angle cone beam CT projection image.

3. The full-angle cone-beam CT projection image restoration method of claim 1, wherein, The 3D feature extractor is an encoder-decoder architecture, the encoder of which comprises a plurality of three-dimensional convolution units comprising a three-dimensional convolution layer and a LeakyReLU activation function; and the decoder of which comprises a plurality of three-dimensional deconvolution units comprising a three-dimensional deconvolution layer and a LeakyReLU activation function; The 2D projection generator is an encoder-decoder architecture, the encoder of which comprises a plurality of two-dimensional convolution units comprising a two-dimensional convolution layer and a LeakyReLU activation function; and the decoder of which comprises a plurality of two-dimensional deconvolution units comprising a two-dimensional deconvolution layer and a Leaky ReLu activation function; Wherein, each layer of the 3D feature extractor and the 2D projection generator is embedded with a channel attention module to capture the dependency between the input feature map channels.

4. The full-angle cone-beam CT projection image reconstruction method of claim 3, wherein, The encoder of the 3D feature extractor comprises 5 three-dimensional convolution units; and the decoder of the 3D feature extractor comprises 4 three-dimensional deconvolution units. The encoder of the 2D projection generator comprises 5 two-dimensional convolution units; and the decoder of the 3D feature extractor comprises 4 two-dimensional deconvolution units.

5. The full-angle cone-beam CT back-projection image reconstruction method of claim 4, wherein, The high-quality cone-beam CT image is represented as: wherein P denotes the input sparse angular cone-beam CT projection image, f θ denotes the 3D-2D channel attention interpolation network; AC 1-5 denotes the three-dimensional convolution unit; BC 1-4 denotes the three-dimensional deconvolution unit; DC 1-5 denotes the two-dimensional convolution unit; EC 1-4 denotes the two-dimensional deconvolution unit; IPF denotes the inter-projection fusion module.

6. The full-angle cone-beam CT back-projection image reconstruction method of claim 3, wherein, The channel attention module is used to calculate the global mean of the input feature map along the channel to obtain an average feature map; extract features of the average feature map by using a convolution layer and a Leaky ReLU activation function and integrate them into a one-dimensional vector; activate the one-dimensional vector by using a Sigmoid function; and point multiply the input feature map and the attention mapping along the channel to finally obtain an output feature map; the output feature map can reflect the channel dependency of the input feature map.

7. The full-angle cone-beam CT back-projection image reconstruction method of claim 1, wherein, A filtered frequency domain representation RF(*) of the projection image is computed using the formula RF(P t ) = |a(u,v) x F t (u,v) |; where H, W are the height and width of the projected image, respectively, (x, y) is the spatial domain coordinate of the projection t (x, y) is the pixel value projected at (x, y); (u, v) is the frequency domain coordinate of the projection, F t (u, v) is the frequency domain complex value projected at (u, v); a(u, v) is the coefficient of the Ramp-Filter weight matrix at (u, v).

8. A device for restoring full-angle cone-beam CT projection images, characterized in that, A device for performing the full-angle cone-beam CT projection image restoration method of any one of claims 1-7, comprising: a normalization module configured to normalize a pixel matrix corresponding to a sparse-angle cone-beam CT projection image to obtain a normalized image; The extraction module is configured to input the normalized image into a 3D-2D channel attention interpolation network, which comprises, in sequence, a 3D feature extractor, an inter-projection fusion module and a 2D projection generator; the 3D feature extractor is configured to extract spatial features of the normalized image and output a 3D feature map; The merging module is configured to merge the spatial features and channel dimensions of the 3D feature map by using the inter-projection fusion module, so as to reshape the 3D feature map into a 2D feature map; The recovery module is configured to recover missing projections corresponding to the 2D feature map by using the 2D projection generator, to obtain a full-angle projection image; The reconstruction module is configured to perform reconstruction processing on the full-angle projection image by using an FDK algorithm, to obtain a high-quality cone beam CT image.

9. A full-angle cone-beam CT projection image reconstruction system comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7.

Images