Computed tomography image reconstruction method and system based on extremely sparse scanning

Abstract

The invention relates to a computed tomography image reconstruction method and system based on extremely sparse scanning. The method comprises the following steps: acquiring original scanning data of a CT machine and a reconstruction image of a standard algorithm; building a generative adversarial network; designing a loss function, and training the generative adversarial network; a VdCS algorithm and a CT-RDNet network are combined through a feedback path, and an extremely sparse scanning CT image reconstruction framework, namely a CSUF image reconstruction framework, is constructed; acquiring scanning data, namely extremely sparse scanning data, of an extremely sparse scanning mode of the CT machine; and taking the extremely sparse scanning data as input, executing extremely sparse scanning CT image reconstruction by the CSUF image reconstruction framework, and outputting a reconstructed image meeting imaging clinical diagnosis requirements. According to the method, the CT image meeting the imaging clinical diagnosis requirement can be obtained under the extremely sparse scanning condition, and good robustness and engineering practicability are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging, and in particular, to a method and system for computer tomography image reconstruction based on extremely sparse scanning. Background Art

[0002] Computed tomography (CT) has become an indispensable important technology in clinical diagnosis due to its high resolution and fast imaging ability. However, the X-ray ionizing radiation during CT scanning poses a potential hazard to human health. Prolonged or excessive exposure of patients to X-ray radiation increases the risk of cancer. Reducing the number of scans in CT scanning is an effective way to reduce the X-ray radiation dose. However, in the case of extremely sparse scanning (the number of scanning angles in the fan beam / cone beam scanning within the range of [0, 2π) is no more than 20), due to severe undersampling, the reconstructed image contains obvious artifacts and is prone to problems such as loss of high-frequency or detailed information, which weakens the imaging clinical diagnostic value and status of CT images and also forms an obstacle to the wider clinical application and promotion of CT imaging technology.

[0003] Traditional CT image reconstruction methods, including analytical algorithms such as filtered backprojection (FBP) and iterative image reconstruction algorithms such as ART, perform poorly under extremely sparse scanning conditions and are difficult to provide reconstructed images that meet the requirements of imaging clinical diagnosis. In the literature "View-driven Compressed Sensing method of CT image reconstruction" (IOP Conference Series: Materials Science and Engineering, 768, 062003), a CT iterative image reconstruction algorithm applicable to extremely sparse scanning (a CT image reconstruction algorithm based on view-driven compressed sensing, VdCS algorithm) was proposed based on the compressed sensing theory. The VdCS algorithm can relatively effectively achieve image reconstruction under extremely sparse scanning conditions. However, the limitations of the compressed sensing theory are also manifested in the VdCS algorithm. For example, the algorithm has a staircase effect, and the reconstructed image is prone to problems such as loss of high-frequency information or blurred artifacts. In recent years, CT image reconstruction methods based on deep learning have received attention. Compared with traditional algorithms, deep learning-based methods have shown significant advantages in suppressing noise, removing artifacts, and improving image detail retention. However, most existing deep learning-based CT image reconstruction methods are limited to post-processing in the image domain, that is, learning the mapping from low-quality reconstructed images to high-quality images through neural networks. Under extremely sparse scanning conditions, due to the poor quality of the original input images, existing deep learning methods are difficult to accurately learn the above mapping relationship, which may lead to problems such as overly smooth reconstructed images and loss of detailed information or even key lesion information. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention proposes a CT image reconstruction method that combines CT iterative image reconstruction and deep learning. This method can obtain CT images that meet the clinical diagnosis requirements under the condition of extremely sparse scanning (the scale of the scanned data is equivalent to that the number of scanning angles in the fan-beam / cone-beam scanning within the range of [0, 2π) is no more than 20), and has good robustness and engineering practicability.

[0005] The present invention also provides a computer tomography image reconstruction system based on extremely sparse scanning.

[0006] Term Explanation:

[0007] 1. VdCS algorithm, a CT image reconstruction algorithm for view-driven compressed sensing, briefly recorded as the VdCS algorithm.

[0008] 2. The conventional scanning mode of the CT machine refers to that when the CT machine images the tomogram of the scanned human body, it uses typical or standard scanning parameters and sampling strategies within the range recommended by the current mainstream clinical practice or equipment manufacturers to obtain sufficiently dense and evenly distributed scanned data, so as to be able to perform high-quality and complete image reconstruction.

[0009] 3. The extremely sparse scanning mode of the CT machine refers to a scanning strategy in which the CT machine only obtains a very small amount of sparsely distributed scanned data (the scale of the scanned data is equivalent to that the number of scanning angles in the fan-beam / cone-beam scanning within the range of [0, 2π) is no more than 20), including the undersampling scanning mode and the limited-angle scanning mode; among them, the undersampling scanning mode means that the CT machine only obtains sparse scanned data at a very small number of but evenly distributed scanning angles, and the limited-angle scanning mode means that the CT machine only obtains sparse scanned data at a limited number of partial scanning angles.

[0010] The technical solution of the present invention is as follows:

[0011] A computer tomography image reconstruction method based on extremely sparse scanning, including:

[0012] Step 1, obtaining the original scanned data of the CT machine and the reconstructed image of the standard algorithm;

[0013] The original scan data refers to the X-ray attenuation measurement data; several scan angles are evenly selected from the original scan data to obtain sparse scan data, where the sparse scan data refers to the scan data at a small number of scan angles, and the sparse reconstruction images at different numbers of scan angles are obtained using the VdCS algorithm; the sparse reconstruction images are paired with the reconstruction images of the standard algorithm, where pairing means that the sparse reconstruction images at different numbers of scan angles are respectively corresponding one by one to the reconstruction images of the standard algorithm to obtain the training dataset for supervised deep learning.

[0014] Step 2: Construct a CT image restoration and noise reduction network, namely the CT-RDNet network.

[0015] Step 3: Use the CT-RDNet network as the generator to build a generative adversarial network.

[0016] Step 4: Design a loss function and train the generative adversarial network.

[0017] Step 5: Combine the VdCS algorithm and the CT-RDNet network through a feedback path to construct an extremely sparse scan CT image reconstruction framework, namely the CSUF image reconstruction framework.

[0018] Step 6: Obtain the scan data in the extremely sparse scan mode of the CT machine, namely the extremely sparse scan data.

[0019] Step 7: Use the extremely sparse scan data as the input, and the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction to output the reconstruction images that meet the requirements of radiological clinical diagnosis.

[0020] Preferably according to the present invention, the CT-RDNet network is a symmetric U-shaped architecture of an encoder-decoder.

[0021] The encoder of the CT-RDNet network includes 8 downsampling modules, which are used to extract the features of the sparse reconstruction images and downsample the sparse reconstruction images.

[0022] The decoder of the CT-RDNet network includes 7 upsampling modules and an output layer, which are used to reconstruct high-quality images from the features of the extracted sparse reconstruction images.

[0023] The CT-RDNet network uses skip connections to fuse features and a self-attention mechanism to focus on global features; among them, the output feature map of the upsampling module and the output feature map of the downsampling module with the same size are concatenated in the channel direction to obtain fused features, and then the fused features are processed by the self-attention module to obtain the attention feature map, and finally upsampling is performed.

[0024] The input feature map, i.e., the fused feature, obtains three corresponding feature maps through three 1×1 convolutions, denoted as Q, K, and V. The obtained feature maps Q, K, and V are flattened into one-dimensional vectors in the spatial dimension with a length of H×W, and Q is transposed. The transposed Q and K are multiplied and processed using the SoftMax function to obtain attention scores. V and the attention scores are multiplied and the vector is restored to the two-dimensional size of H×W to obtain the attention output. Finally, the attention output is processed using a 1×1 convolution to obtain the attention feature map enhanced by self-attention.

[0025] Preferably according to the present invention, the discriminator of the generative adversarial network includes 5 discriminant modules, a zero-padding layer, and an output layer. The discriminant module is used to extract image features. The zero-padding layer is used to adjust the size of the feature map.

[0026] Preferably according to the present invention, a loss function is designed to train the generative adversarial network, including:

[0027] Loss function is defined as follows:

[0028]

[0029] where E (x～p(x)) represents the expected value of the random variable x sampled from the distribution p(x), x is the input image, is the output image of the CT-RDNet network, y is the reconstructed image of the standard algorithm; D represents the discriminator; c, C1, C2 are constants, N is the number of pixels of the image; μ " and are respectively the means of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network, and are respectively the variances of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network, is the standard deviation of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network; β is the weight parameter.

[0030] Preferably according to the present invention, with extremely sparse scan data as the input, the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction and outputs a reconstructed image that meets the requirements of imaging clinical diagnosis, including:

[0031] (1) Use the VdCS algorithm to perform iterative image reconstruction on the extremely sparse scan data and transfer the reconstructed image to the CT-RDNet network;

[0032] (2) The CT-RDNet network performs image restoration and noise reduction optimization processing on the reconstructed image in step (1);

[0033] (3) The image processed by the CT-RDNet network is provided to the VdCS algorithm as a prior image through the feedback path;

[0034] (4) Using the prior image as the initial estimate, iteratively reconstruct the extremely sparse scan data again, and transfer the image reconstruction result to the CT-RDNet network;

[0035] (5) Further optimize the iterative image reconstruction result of the VdCS algorithm through the CT-RDNet network, and output a reconstructed image that meets the requirements of imaging clinical diagnosis.

[0036] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the computer tomography image reconstruction method based on extremely sparse scanning are implemented.

[0037] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the computer tomography image reconstruction method based on extremely sparse scanning are implemented.

[0038] A computer tomography image reconstruction system based on extremely sparse scanning includes:

[0039] A data acquisition module is configured to: acquire the original scan data of a CT machine and the reconstructed image of a standard algorithm; the original scan data refers to X-ray attenuation measurement data; uniformly select several scan angles from the original scan data to obtain sparse scan data, and the sparse scan data refers to scan data under a small number of scan angle numbers, and use the VdCS algorithm to obtain sparse reconstructed images with different numbers of scan angles respectively; pair the sparse reconstructed images with the reconstructed images of the standard algorithm, and the pairing means that the sparse reconstructed images with different numbers of scan angles and the reconstructed images of the standard algorithm correspond one by one respectively to obtain a training data set for supervised deep learning;

[0040] A generative adversarial network construction module is configured to: construct a CT image restoration and noise reduction network, that is, the CT-RDNet network; use the CT-RDNet network as a generator to build a generative adversarial network;

[0041] A generative adversarial network training module is configured to: design a loss function and train the generative adversarial network;

[0042] An image reconstruction module is configured to: combine the VdCS algorithm and the CT-RDNet network through a feedback path to construct an extremely sparse scan CT image reconstruction framework, that is, the CSUF image reconstruction framework; acquire the scan data in the extremely sparse scan mode of the CT machine, that is, the extremely sparse scan data; use the extremely sparse scan data as the input, and the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction and outputs a reconstructed image that meets the requirements of imaging clinical diagnosis.

[0043] The beneficial effects of the present invention are as follows:

[0044] Compared with the prior art, the method of the present invention can obtain CT images that meet the needs of clinical imaging diagnosis under the condition of extremely sparse scanning (the scale of scanned data is no more than 20 in the number of scanning angles of fan-beam / cone-beam scanning in the range of [0, 2π)), and has good robustness and engineering practicability. Description of the Drawings

[0045] Figure 1 is a schematic flowchart of a computerized tomography image reconstruction method based on extremely sparse scanning according to the present invention;

[0046] Figure 2 is a schematic flowchart of the production of the data set according to the present invention;

[0047] Fig. 3(a) is a schematic structural diagram of the CT-RDNet network of the present invention;

[0048] Fig. 3(b) is a schematic structural diagram of the self-attention module of the CT-RDNet network of the present invention;

[0049] Fig. 3(c) is a schematic structural diagram of the discriminator of the generative adversarial network of the present invention;

[0050] Figure 4 is a schematic structural diagram of an extremely sparse scanning CT image reconstruction framework (CSUF image reconstruction framework) of the present invention; Detailed Embodiments

[0051] The present invention will be further defined below in conjunction with the accompanying drawings of the specification and embodiments, but is not limited thereto.

[0052] Embodiment 1

[0053] A computerized tomography image reconstruction method based on extremely sparse scanning, as Figure 1 shown, includes:

[0054] Step 1, obtaining the original scanned data in the conventional scanning mode of the CT machine and the reconstructed image of the standard algorithm;

[0055] The original scan data refers to the X-ray attenuation measurement data directly collected by the CT machine detector and not yet preprocessed and image-reconstructed; the reconstructed image of the standard algorithm usually refers to the reconstructed image obtained using the image reconstruction algorithm preset or default by the CT device manufacturer for general human tomographic tissue imaging (such as the filtered backprojection algorithm, i.e., the FBP algorithm). The reconstructed image of the standard algorithm of the CT machine often makes a relatively balanced trade-off between spatial resolution and noise suppression and is suitable for the imaging clinical examination needs of most cases. Select a number of scan angles uniformly from the original scan data to obtain sparse scan data. Sparse scan data refers to the scan data at a small number of scan angles. For example, the number of scan angles is 20, 45, 90, 180 respectively. Compared with the original scan data, the data volume of the sparse scan data is greatly reduced; use the VdCS algorithm to obtain sparse reconstructed images at different numbers of scan angles respectively; pair the sparse reconstructed images with the reconstructed images of the standard algorithm. Pairing means corresponding the sparse reconstructed images at different numbers of scan angles with the reconstructed images of the standard algorithm one by one respectively to obtain the training data set for supervised deep learning; the production process of the training data set is as Figure 2 shown.

[0056] Step 2: Construct a CT image restoration and noise reduction network, namely the CT-RDNet network;

[0057] Step 3: Take the CT-RDNet network as the generator and build a generative adversarial network;

[0058] Step 4: Design a loss function and train the generative adversarial network;

[0059] Step 5: Combine the VdCS algorithm and the CT-RDNet network through a feedback path to construct an extremely sparse scan CT image reconstruction framework, namely the CSUF image reconstruction framework; the structure of the CSUF image reconstruction framework is as Figure 4 shown; among them, the CT-RDNet network provides prior image information to the VdCS algorithm through the feedback path.

[0060] Step 6: Obtain the scan data in the extremely sparse scan mode of the CT machine, namely the extremely sparse scan data;

[0061] Step 7: Take the extremely sparse scan data as the input, and the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction (i.e., CSUF image reconstruction) and outputs a reconstructed image that meets the imaging clinical diagnosis requirements.

[0062] Embodiment 2

[0063] A computer tomography image reconstruction method based on extremely sparse scanning according to Embodiment 1, which is characterized in that:

[0064] As shown in Figure 3(a), the CT-RDNet network is a symmetric U-shaped architecture of an encoder-decoder;

[0065] The encoder of the CT-RDNet network includes 8 downsampling modules, which are used to extract the features of the sparse reconstruction image and downsample the sparse reconstruction image;

[0066] The 8 downsampling modules include Down1, Down2, Down3, Down4, Down5, Down6, Down7, and Down8 connected in sequence;

[0067] Down1 includes a convolutional layer and an activation function;

[0068] Both Down2 and Down3 include a convolutional layer, an activation function, and a batch normalization layer;

[0069] Down4, Down5, Down6, and Down7 all include a convolutional layer, an activation function, a batch normalization layer, and a dropout layer;

[0070] Down8 includes a convolutional layer, an activation function, and a dropout layer;

[0071] The convolutional layer has a kernel size of 4×4, a stride of 2, and a padding of 1; the activation function uses the LeakyReLU function with a slope of 0.2; the dropout layer is used to reduce overfitting.

[0072] The decoder of the CT-RDNet network includes 7 upsampling modules and an output layer, which are used to reconstruct a high-quality image from the features of the extracted sparse reconstruction image;

[0073] The 7 upsampling modules include Up1, Up2, Up3, Up4, Up5, Up6, and Up7 connected in sequence;

[0074] Up1, Up2, and Up3 all include a transposed convolutional layer, a batch normalization layer, and a ReLU activation function;

[0075] Up4, Up5, Up6, and Up7 all include a transposed convolutional layer, a batch normalization layer, a ReLU activation function, and a dropout layer;

[0076] The transposed convolutional layer has a kernel size of 4×4, a stride of 2, and a padding of 1; the dropout layer is used to reduce overfitting;

[0077] The output layer contains a transposed convolutional layer with a kernel size of 4×4, a stride of 2, and a padding of 1 and a Tanh activation function.

[0078] The CT-RDNet network uses skip connections to fuse features and a self-attention mechanism to focus on global features. Among them, the output feature map of the first upsampling module is concatenated with the output feature map of the same size in the seventh downsampling module in the channel direction to obtain fused features, which are then processed by the self-attention module to obtain an attention feature map, and finally upsampling is performed.

[0079] Here, the implementation process of the self-attention module (shown in Figure 3(b)) of the present invention is as follows:

[0080] The input feature map, that is, the fused features, respectively obtain three corresponding feature maps through three 1×1 convolutions, denoted as Q, K, and V. The obtained feature maps Q, K, and V are flattened into one-dimensional vectors in the spatial dimension (i.e., height (H) and width (W)), with a length of H×W, and Q is transposed. The transposed Q and K are multiplied and processed using the SoftMax function to obtain attention scores. V and the attention scores are multiplied and the vector is restored to the two-dimensional size of H×W to obtain the attention output. Finally, the attention output is processed using a 1×1 convolution to obtain the attention feature map strengthened by self-attention.

[0081] As shown in Figure 3(c), the discriminator of the generative adversarial network includes 5 discriminant modules, a zero-padding layer, and an output layer. The discriminant modules are used to extract image features. The zero-padding layer is used to adjust the size of the feature map.

[0082] The 5 discriminant modules include Dise1, Dise2, Dise3, Dise4, and Dise5.

[0083] Dise1 includes a convolutional layer, a LeakyReLU activation function, and a dropout layer.

[0084] Dise2, Dise3, Dise4, and Dise5 all include a convolutional layer, a LeakyReLU activation function, a dropout layer, and a batch normalization layer.

[0085] The convolutional layer has a convolutional kernel size of 4×4, a stride of 2, and a padding of 1. The slope of the LeakyReLU activation function is 0.2. The zero-padding layer is used to adjust the size of the feature map. The output layer uses a convolutional layer without bias, with a convolutional kernel size of 4×4, a stride of 1, and a padding of 1. Here, the output size of the discriminator and the corresponding label size are both 16×16.

[0086] Design a loss function to train the generative adversarial network, including:

[0087] The overall loss function of the CT-RDNet network of the present invention fuses the LSGAN loss, the L1 loss, and the SSIM loss. The loss function is defined as follows:

[0088]

[0089] Among them, E (x～p(x)) represents the expected value of the random variable x sampled from the distribution p(x), where x is the input image, is the output image of the CT-RDNet network, y is the reconstructed image of the standard algorithm; D represents the discriminator; c, C1, C2 are constants, which are 100, 0.01, 0.03 respectively; N is the number of pixels of the image; μ y and are the means of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network respectively, and are the variances of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network respectively, is the standard deviation of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network; β is the weight parameter, usually taking the value of 0.4.

[0090] The discriminator loss function of the generative adversarial network of the present invention adopts the LSGAN loss, and the label corresponding to the generated sample is 0, and the label corresponding to the real sample is 1.

[0091] The specific training steps of the generative adversarial network are as follows:

[0092] Data preparation: Using the training data set constructed in Step 1, the sparse reconstructed image is used as the input image, and the reconstructed image of the standard algorithm is used as the target image. Before training, the input and target images are normalized to ensure consistent data distribution, which helps to accelerate network convergence and improve training stability.

[0093] The specific training process includes:

[0094] (1) Initialize the weights of the generator and the discriminator.

[0095] (2) For each batch of input images, execute:

[0096] i. Generate an image: Input the normalized sparse reconstructed image into the generator to generate a preliminary reconstructed image (generated sample);

[0097] ii. Update the discriminator: Input the target image (real sample, label is 1) and the generated image (generated sample, label is 0); into the discriminator, calculate the discriminator loss and perform backpropagation to update the weights of the discriminator to improve its ability to distinguish real images and generated images.

[0098] iii. Updater: Calculate the losses between the generated image and the target image (LSGAN loss, L1 loss, and SSIM loss), perform backpropagation, update the weights of the generator, and optimize the quality of the generated image.

[0099] (3) Stop when the training reaches the termination condition.

[0100] Taking extremely sparse scan data as input, the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction and outputs a reconstructed image that meets the requirements of imaging clinical diagnosis; it includes:

[0101] (1) Use the VdCS algorithm to perform iterative image reconstruction on the extremely sparse scan data. The VdCS algorithm is an iterative image reconstruction algorithm. Reconstruct the extremely sparse scan data to obtain a reconstructed image and pass the reconstructed image to the CT-RDNet network;

[0102] (2) Perform image restoration and noise reduction optimization on the reconstructed image in step (1) through the CT-RDNet network; perform restoration and noise reduction on the CT reconstructed image of the sparse scan through the CT-RDNet;

[0103] (3) The image optimized by the CT-RDNet network is provided to the VdCS algorithm as a prior image through the feedback path;

[0104] (4) Using the prior image as the initial estimate, the VdCS algorithm performs iterative image reconstruction on the extremely sparse scan data again. Similar to step (1), the difference is that the input of the VdCS algorithm used here adds a prior image, and the image reconstruction result is passed to the CT-RDNet network;

[0105] (5) Further optimize the iterative image reconstruction result of the VdCS algorithm through the CT-RDNet network. Similar to step (2), both use the CT-RDNet for restoration and noise reduction processing and output a reconstructed image that meets the requirements of imaging clinical diagnosis.

[0106] Example 3

[0107] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the method for computer tomography image reconstruction based on extremely sparse scanning described in Example 1 or 2.

[0108] Example 4

[0109] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the method for computer tomography image reconstruction based on extremely sparse scanning described in Example 1 or 2.

[0110] Example 5

[0111] A computer tomography image reconstruction system based on extremely sparse scanning, comprising:

[0112] A data acquisition module, configured to: acquire the original scan data of a CT machine and the reconstructed image of a standard algorithm; the original scan data refers to X-ray attenuation measurement data; uniformly select a number of scan angles from the original scan data to obtain sparse scan data, where the sparse scan data refers to scan data under a small number of scan angles, and use the VdCS algorithm to obtain sparse reconstructed images of different numbers of scan angles respectively; pair the sparse reconstructed images with the reconstructed image of the standard algorithm, where pairing means corresponding the sparse reconstructed images of different numbers of scan angles with the reconstructed image of the standard algorithm one by one respectively, to obtain a training data set for supervised deep learning;

[0113] A generative adversarial network construction module, configured to: construct a CT image restoration and noise reduction network, i.e., the CT-RDNet network; use the CT-RDNet network as a generator to build a generative adversarial network;

[0114] A generative adversarial network training module, configured to: design a loss function and train the generative adversarial network;

[0115] An image reconstruction module, configured to: combine the VdCS algorithm and the CT-RDNet network through a feedback path to construct an extremely sparse scanning CT image reconstruction framework, i.e., the CSUF image reconstruction framework; acquire the scan data of the extremely sparse scanning mode of the CT machine, i.e., the extremely sparse scan data; use the extremely sparse scan data as input, and the CSUF image reconstruction framework performs extremely sparse scanning CT image reconstruction and outputs a reconstructed image that meets the requirements of imaging clinical diagnosis.

Claims

1. A computer tomography image reconstruction method based on extremely sparse scanning, characterized in that, Including: Step 1: Obtain the original CT scan data and the reconstructed images of the standard algorithm; The original scan data refers to the X-ray attenuation measurement data; several scan angles are uniformly selected from the original scan data to obtain sparse scan data, where the sparse scan data refers to the scan data under a small number of scan angle numbers. The VdCS algorithm is used to obtain sparse reconstructed images with different numbers of scan angles respectively; the sparse reconstructed images are paired with the reconstructed images of the standard algorithm, where pairing means that the sparse reconstructed images with different numbers of scan angles are respectively corresponding one by one to the reconstructed images of the standard algorithm to obtain a training dataset for supervised deep learning; Step 2: Construct a CT image restoration and noise reduction network, namely the CT-RDNet network; Step 3: Use the CT-RDNet network as the generator to build a generative adversarial network; Step 4: Design a loss function and train the generative adversarial network; Step 5: Combine the VdCS algorithm and the CT-RDNet network through a feedback path to construct an extremely sparse scan CT image reconstruction framework, namely the CSUF image reconstruction framework; Step 6: Obtain the scan data of the extremely sparse scan mode of the CT machine, namely the extremely sparse scan data; Step 7: Using the extremely sparse scan data as the input, the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction and outputs reconstructed images that meet the requirements of imaging clinical diagnosis.

2. The computer tomography image reconstruction method based on extremely sparse scanning according to claim 1, wherein The CT-RDNet network is a symmetric U-shaped architecture of an encoder-decoder; The encoder of the CT-RDNet network includes 8 downsampling modules, which are used to extract the features of the sparse reconstructed images and downsample the sparse reconstructed images; The decoder of the CT-RDNet network includes 7 upsampling modules and an output layer, which are used to reconstruct high-quality images from the features of the extracted sparse reconstructed images; The CT-RDNet network uses skip connections to fuse features and a self-attention mechanism to focus on global features; among them, the output feature map of the upsampling module and the output feature map of the downsampling module with the same size are concatenated in the channel direction to obtain fused features, and then processed by a self-attention module to obtain an attention feature map, and finally upsampled; The input feature map, that is, the fused features, respectively obtain three corresponding feature maps through three 1×1 convolutions, denoted as Q, K, and V; the obtained feature maps Q, K, and V are flattened into one-dimensional vectors in the spatial dimension, with a length of H×W, and Q is transposed; the transposed Q and K are multiplied and processed using the SoftMax function to obtain attention scores; V and the attention scores are multiplied and the vector is restored to a two-dimensional size of H×W to obtain an attention output; finally, the attention output is processed using a 1×1 convolution to obtain an attention feature map strengthened by self-attention.

3. A computer tomography image reconstruction method based on extremely sparse scanning according to claim 1, characterized in that, The discriminator of the generative adversarial network includes 5 discriminant modules, a zero-padding layer, and an output layer; the discriminant modules are used to extract image features; The zero-padding layer is used to adjust the size of the feature map.

4. The computer tomography image reconstruction method based on extremely sparse scanning according to the claim is characterized in that, Design a loss function and train the generative adversarial network; Including: Loss function is defined as follows: Among them, E (x～p(x)) represents the expected value of the random variable x sampled from the distribution p(x), where x is the input image, is the output image of the CT-RDNet network, y is the reconstructed image of the standard algorithm; D represents the discriminator; c, C1, C2 are constants, N is the number of pixels of the image; μ " and are the means of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network respectively, and are the variances of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network respectively, is the standard deviation of the pixel values of the reconstructed image of the standard algorithm and the output image of the CT-RDNet network; β is the weight parameter.

5. A computer tomography image reconstruction method based on extremely sparse scanning according to any one of claims 1-4, characterized in that, Using the extremely sparse scan data as the input, the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction and outputs reconstructed images that meet the requirements of imaging clinical diagnosis; Including: (1) Iteratively reconstruct the extremely sparse scan data using the VdCS algorithm and transfer the reconstructed image to the CT-RDNet network; (2) Restore and denoise the reconstructed image in step (1) through the CT-RDNet network; (3) The image optimized by the CT-RDNet network is provided to the VdCS algorithm as a prior image through the feedback path; (4) Using the prior image as the initial estimate, iteratively reconstruct the extremely sparse scan data again and transfer the image reconstruction result to the CT-RDNet network; (5) Further optimize the iterative image reconstruction result of the VdCS algorithm through the CT-RDNet network and output the reconstructed image that meets the requirements of radiological clinical diagnosis.

6. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the computerized tomography image reconstruction method based on extremely sparse scanning according to any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the computerized tomography image reconstruction method based on extremely sparse scanning according to any one of claims 1-5.

8. A computer tomography image reconstruction system based on extremely sparse scanning, characterized in that, Comprising: A data acquisition module, configured to: acquire the original scan data of the CT machine and the reconstructed image of the standard algorithm; the original scan data refers to the X-ray attenuation measurement data; uniformly select several scan angles from the original scan data to obtain sparse scan data, and the sparse scan data refers to the scan data under a small number of scan angle numbers, and use the VdCS algorithm to obtain sparse reconstructed images with different numbers of scan angles respectively; pair the sparse reconstructed images with the reconstructed images of the standard algorithm, and pairing means corresponding the sparse reconstructed images with different numbers of scan angles to the reconstructed images of the standard algorithm one by one respectively to obtain a training data set for supervised deep learning; A generative adversarial network construction module, configured to: construct a CT image restoration and denoising network, that is, the CT-RDNet network; use the CT-RDNet network as a generator to build a generative adversarial network; A generative adversarial network training module, configured to: design a loss function and train the generative adversarial network; An image reconstruction module, configured to: combine the VdCS algorithm and the CT-RDNet network through the feedback path to construct an extremely sparse scan CT image reconstruction framework, that is, the CSUF image reconstruction framework; acquire the scan data of the extremely sparse scan mode of the CT machine, that is, the extremely sparse scan data; using the extremely sparse scan data as the input, the CSUF image reconstruction framework performs extremely sparse scan CT image reconstruction and outputs the reconstructed image that meets the requirements of radiological clinical diagnosis.

Images