A noise reduction method and system suitable for sparse-angle CT reconstruction tomographic images

Abstract

The application provides a noise reduction method and system suitable for sparse angle CT reconstruction tomography image. The method comprises the following steps: S1: sparse angle reconstruction is performed through a CT reconstruction algorithm to obtain a sparse reconstruction tomography image; S2: a feature extraction is performed through a deep learning method to extract a tomographic noise image from the sparse reconstruction tomography image; S3: the sparse reconstruction tomography image is subtracted from the tomographic noise image to obtain a high-quality tomography image; S4: gradient information and edge structure information are calculated from the high-quality tomography image; and S5: an edge-guided total variation method is used to further denoise the high-quality tomography image to obtain a final denoised tomography image. The application can achieve good results in sparse angle CT reconstruction noise reduction, effectively ensure clear structure edges and restore good boundary details.

Description

Technical Field

[0001] This invention relates to the fields of sparse angle CT reconstruction, total variational image denoising algorithm (TVM), convolutional neural network (CNN), and sparse angle CT characterization technology. Specifically, it relates to a denoising method and system suitable for sparse angle CT reconstruction tomographic images. Background Technology

[0002] X-ray computed tomography (CT) is a non-destructive imaging technique widely used for the three-dimensional quantitative evaluation of the internal structure of samples. In in-situ characterization experiments targeting dynamic processes, sparse sampling is necessary due to the discrepancy between the dynamic evolution time and the CT sampling exposure time.

[0003] However, sparse sampling violates the Nyquist-Shannon sampling theorem, and the reconstruction results using the classic Filtered Back Projection (FBP) algorithm are of unsatisfactory quality. Sparsely reconstructed tomographic images exhibit significant stripe artifacts, severely impacting the quantitative characterization of the sample's internal structure. Existing Total Variation Minimization (TVM) methods based on compressed sensing theory, by imposing regularization constraints, enable the reconstructed image to be recovered from the sparse signal. However, TVM, through its uniform penalty of image gradients, tends to result in overly smoothed tomographic images, leading to a certain degree of degradation at edge structures.

[0004] By leveraging the feature extraction and fitting capabilities of convolutional neural networks, effective information can be extracted from noisy sparse reconstructed tomographic images, making denoising of these images possible. However, training neural networks is challenging, and the training results are unstable. Furthermore, using mean squared error loss as the loss function also fails to yield images with sharp boundaries, and high-frequency information in the image is difficult to recover. A single approach is insufficient to achieve ideal results.

[0005] Therefore, we can first extract sparse reconstruction tomographic noise through the network, and then subtract the noise from the tomographic noise to obtain the denoised image of the network as the prior image. Based on the edge structure information of the prior image, we can guide the direction of total variation minimization, thus obtaining sharp edge structures while correctly removing noise. Summary of the Invention

[0006] In view of the deficiencies in the prior art, the purpose of this invention is to provide a noise reduction method and system suitable for sparse angle CT reconstructed tomographic images.

[0007] In a first aspect, according to the present invention, a noise reduction method for sparse angle CT reconstructed tomographic images includes:

[0008] Step S1: Perform sparse angle reconstruction using a CT reconstruction algorithm to obtain a sparse reconstructed tomographic image;

[0009] Step S2: Extract features using deep learning methods to extract fault noise images from the sparse reconstructed tomographic images;

[0010] Step S3: Subtract the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image;

[0011] Step S4: Using the high-quality tomographic image as the prior image, calculate the gradient information and edge structure information of the prior image;

[0012] Step S5: The high-quality tomographic image is further denoised using the edge-guided total variation method to obtain the final denoised tomographic image.

[0013] Preferably, step S2 employs an adversarial training strategy involving alternating iterative generation and discrimination networks to learn the mapping relationship from sparsely reconstructed tomographic images to sparsely noisy images, thereby removing noise from the sparsely reconstructed images.

[0014] Step S2.1: The input is a sparsely reconstructed tomographic image, and the output is a tomographic noise image;

[0015] Step S2.2: Identify the network loss D Defined as: the binary classification cross-entropy loss to determine whether a noisy image obtained through a generator network is real, mathematically expressed as:

[0016] l D =H[1,D(I noise )]+H[0,D(G(I sparse ))]

[0017] Where H represents cross-entropy, D represents the discriminant network, and I noise I represents the label of a real noisy image. sparse represents a sparsely reconstructed tomographic image, and G represents a generative network;

[0018] Step S2.3: Network MSE loss MSE Defined as: the mean square error between the noisy image obtained from the input through the generator network and the true noise label, mathematically expressed as:

[0019]

[0020] Where G represents the generator network;

[0021] Step S2.4: Generate network l G Defined as the joint loss: a weighted sum of the MSE loss and the discrimination network loss, mathematically expressed as:

[0022] l G =l MSE+λl D

[0023] Where λ represents the hyperparameter that balances the weights of the MSE loss and the discriminant network loss;

[0024] Step S2.5: Train the generator network and the discriminator network alternately and iteratively using gradient descent until l G Stablize.

[0025] Preferably, step S3 includes subtracting the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image I denoised by a neural network. net Mathematically, this is represented as:

[0026] I net =I sparse -G(I sparse ).

[0027] Preferably, step S4 includes,

[0028] Step S4.1: Extract the edge structure of the prior image using the Sobel operator;

[0029] Step S4.2: Expand the width of the prior image edge structure, apply a 3×3 Gaussian convolution template to smooth the edges, and assign 1 to pixel values ​​with a gray level greater than 0 to obtain edge information I. * ;

[0030] Step S4.3: Calculate the gradient information of the prior image. Calculate the gradient magnitude of the prior image point by point, and substitute it into the weight function to obtain the weight information ω, which is mathematically expressed as:

[0031]

[0032] in, Let I be the gradient operator, σ be the edge smoothing parameter, and I be the gradient operator. net (x,y) represents a high-quality tomographic image I. net The direction vector;

[0033] The parameter σ is used to control the degree of edge smoothing, and different adjustments are made for different types of images.

[0034] Preferably, step S5 includes,

[0035] S5.1: Using the total variation TV minimization method, mathematically expressed as:

[0036]

[0037] Where TV represents the total variation term;

[0038] Step S5.2: Through edge information I * The edge direction vector e is calculated and used to guide the direction of the total variation correction at the edge details. Mathematically, it is expressed as:

[0039]

[0040]

[0041] Where F represents the noise amplitude in the direction parallel to the edge structure, and EGTV represents the edge-guided total variation term. After correction, the problem of minimizing the total variation at the edge is transformed into minimizing the noise amplitude parallel to the edge structure, ensuring that the edge gradient is not affected while removing noise.

[0042] Step S5.3: Based on the weight ω, combine the total variation with the edge-guided total variation, and the final correction term FTV is expressed as:

[0043] FTV(I net )=TV(ω·I net )+EGTV((1-ω)·I net );

[0044] Step S5.4: The final denoised image is obtained through iterative calculation using the gradient descent method. The process includes:

[0045] 1) Initialize the step size parameter α = 0.01, and the number of iterations N. iter =40;

[0046] 2) Enter the loop, when n <N iter :

[0047] Calculate the variational correction term: FTV(I) net )=TV(ω·I net )+EGTV((1-ω)·I net )

[0048] Calculate the partial derivatives:

[0049] Using gradient descent: β = ||I net ||1 / ||d||1, I n+1 =I n -α·β·d

[0050] 3) Obtain the final denoised image I.

[0051] Secondly, the present invention also provides a noise reduction system suitable for sparse angle CT reconstructed tomographic images, comprising:

[0052] Module M1: Performs sparse angle reconstruction using CT reconstruction algorithms to obtain sparse reconstructed tomographic images;

[0053] Module M2: Extracts fault noise images from the sparse reconstructed tomographic images using deep learning methods;

[0054] Module M3: Subtracts the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image;

[0055] Module M4: Using the high-quality tomographic image as a priori image, calculate the gradient information and edge structure information of the priori image;

[0056] Module M5: Using the edge-guided total variation method, further denoise the high-quality tomographic image to obtain the final denoised tomographic image.

[0057] Preferably, module M2 learns the mapping relationship from sparsely reconstructed tomographic images to sparsely noisy images by employing an adversarial training strategy of alternating iterative generation and discrimination networks, thereby removing noise from the sparsely reconstructed images.

[0058] Module M2.1: The input is a sparsely reconstructed tomographic image, and the output is a tomographic noise image;

[0059] Module M2.2: Loss of the identification network D Defined as: the binary classification cross-entropy loss to determine whether a noisy image obtained through a generator network is real, mathematically expressed as:

[0060] l D =H[1,D(I noise )]+H[0,D(G(I sparse ))]

[0061] Where H represents cross-entropy, D represents the discriminant network, and I noise I represents the label of a real noisy image. sparse represents a sparsely reconstructed tomographic image, and G represents a generative network;

[0062] Module M2.3: Network MSE Loss MSE Defined as: the mean square error between the noisy image obtained from the input through the generator network and the true noise label, mathematically expressed as:

[0063]

[0064] Where G represents the generator network;

[0065] Module M2.4: Generating Networks G Defined as the joint loss: a weighted sum of the MSE loss and the discrimination network loss, mathematically expressed as:

[0066] l G =lMSE +λl D

[0067] Where λ represents the hyperparameter that balances the weights of the MSE loss and the discriminant network loss;

[0068] Module M2.5: Trains the generator and discriminator networks iteratively using gradient descent until l... G Stablize.

[0069] Preferably, module M3 includes subtracting the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image I denoised by a neural network. net Mathematically, this is represented as:

[0070] I net =I sparse -G(I sparse ).

[0071] Preferably, the module M4 includes,

[0072] Step M4.1: Extract the edge structure of the prior image using the Sobel operator;

[0073] Module M4.2: Expands the width of the prior image edge structure, applies a 3×3 Gaussian convolution template to smooth the edges, assigns 1 to pixels with a gray level greater than 0, and obtains edge information I. * ;

[0074] Module M4.3: Calculates the gradient information of the prior image, calculates the gradient magnitude of the prior image point by point, and substitutes it into the weight function to obtain the weight information ω, which is mathematically expressed as:

[0075]

[0076] in, Let I be the gradient operator, σ be the edge smoothing parameter, and I be the gradient operator. net (x,y) represents a high-quality tomographic image I. net The direction vector;

[0077] The parameter σ is used to control the degree of edge smoothing, and different adjustments are made for different types of images.

[0078] Preferably, the module M5 includes,

[0079] M5.1: Uses the total variation TV minimization method, mathematically expressed as:

[0080]

[0081] Where TV represents the total variation term;

[0082] Module M5.2: Through edge information I * The edge direction vector e is calculated and used to guide the direction of the total variation correction at the edge details. Mathematically, it is expressed as:

[0083]

[0084]

[0085] Where F represents the noise amplitude in the direction parallel to the edge structure, and EGTV represents the edge-guided total variation term. After correction, the problem of minimizing the total variation at the edge is transformed into minimizing the noise amplitude parallel to the edge structure, ensuring that the edge gradient is not affected while removing noise.

[0086] Module M5.3: Based on the weight ω, the total variation is combined with the edge-guided total variation, and the final correction term FTV is expressed as:

[0087] FTV(I net )=TV(ω·I net )+EGTV((1-ω)·I net );

[0088] Module M5.4: Iteratively calculates the final denoised image using gradient descent. The process includes:

[0089] 1) Initialize the step size parameter α = 0.01, and the number of iterations N. iter =40;

[0090] 2) Enter the loop, when n <N iter :

[0091] Calculate the variational correction term: FTV(I) net )=TV(ω·I net )+EGTV((1-ω)·I net )

[0092] Calculate the partial derivatives:

[0093] Using gradient descent: β = ||I net ||1 / ||d||1, I n+1 =I n -α·β·d

[0094] 3) Obtain the final denoised image I.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] 1. This invention improves the quality of sparse reconstructed tomographic maps and restores good boundary details by using an edge-guided total variational method combined with a good prior image obtained by deep learning methods.

[0097] 2. This invention overcomes the problem that traditional variational methods are difficult to reconstruct high-quality tomographic images from a sparse perspective, as well as the difficulties in training deep learning and the inability to recover high-frequency information of images, and achieves good results in detail restoration and artifact control.

[0098] 3. The present invention has good noise reduction effect in sparse angle CT reconstruction. In edge details, the FTV method can effectively ensure clear structural edges. Attached Figure Description

[0099] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0100] Figure 1 This is a diagram of a neural network architecture.

[0101] Figure 2 Flowchart of the edge-guided variational method;

[0102] Figure 3 A schematic diagram to guide the variational method;

[0103] Figure 4 This is a diagram showing the reconstruction results. Detailed Implementation

[0104] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0105] This invention provides a method for denoising sparse-angle CT reconstructed tomographic images, referring to... Figure 1 As shown, the specific steps include the following:

[0106] Step S1: Perform sparse angle reconstruction using a CT reconstruction algorithm to obtain a sparse reconstructed tomographic image;

[0107] Specifically, based on the CT projection image obtained from sparse sampling, the filtered back projection algorithm (FBP) is directly used for sparse angle reconstruction to obtain the sparse reconstructed tomographic image I. sparse Mathematically, this is represented as:

[0108] I sparse =FBP(sparseprojections)

[0109] Step S2: The FBP sparse reconstruction suffers from severe artifact noise, which is difficult to eliminate using traditional methods. Therefore, this invention uses deep learning methods for feature extraction to extract the tomographic noise image from the sparsely reconstructed tomographic image;

[0110] Specifically, a GAN convolutional neural network is first used to extract useful information from noisy tomographic images. Furthermore, sparse artifact noise in CT reconstruction has been shown to have a simpler topological distribution, allowing the residual network to learn the stripe noise distribution more accurately.

[0111] Deep learning methods are processes that map input data to output data. In the field of image processing, especially in image generation or conversion tasks, deep learning involves training to learn a mapping process from one class of images to another, and then generating images based on this association.

[0112] Therefore, NG-Net (Noise Generate Net) was designed, which uses an adversarial training strategy of alternating iterative generation and discrimination networks to learn the mapping relationship from sparsely reconstructed tomographic images to sparsely noisy images, and then removes noise from the sparsely reconstructed images:

[0113] Step S2.1: The input is a sparsely reconstructed tomographic image, and the output is a tomographic noise image;

[0114] Step S2.2: Identify the network loss D Defined as: the binary classification cross-entropy loss to determine whether a noisy image obtained through a generator network is real, mathematically expressed as:

[0115] l D =H[1,D(I noise )]+H[0,D(G(I sparse ))]

[0116] Where H represents cross-entropy, D represents the discriminant network, and I noise I represents the label of a real noisy image. sparse represents a sparsely reconstructed tomographic image, and G represents a generative network;

[0117] Step S2.3: Network MSE loss MSE Defined as: the mean square error between the noisy image obtained from the input through the generator network and the true noise label, mathematically expressed as:

[0118]

[0119] Where G represents the generator network;

[0120] Step S2.4: Generate network l GDefined as the joint loss: a weighted sum of the MSE loss and the discrimination network loss, mathematically expressed as:

[0121] l G =l MSE +λl D

[0122] Where λ represents the hyperparameter that balances the weights of the MSE loss and the discriminant network loss;

[0123] Step S2.5: Train the generator network and the discriminator network alternately and iteratively using gradient descent until l G Stablize.

[0124] Using a deep learning-based convolutional neural network, the loss function of the network considers the difference between the generated tomographic noise image and the labeled image, as well as the accuracy in determining whether the two are consistent. However, it is not limited to the above-mentioned loss functions.

[0125] Step S3: Subtract the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image I denoised by a neural network. net Mathematically, this is represented as:

[0126] I net =I sparse -G(I sparse )

[0127] Step S4: Using the high-quality tomographic image denoised by the neural network as the prior image, extract edge and gradient information to guide edge-guided variational denoising, such as... Figure 3 As shown.

[0128] Step S4.1: Extract the edge structure of the prior image using the Sobel operator;

[0129] Step S4.2: Expand the width of the prior image edge structure, apply a 3×3 Gaussian convolution template to smooth the edges, and assign 1 to pixel values ​​with a gray level greater than 0 to obtain edge information I. * ;

[0130] Step S4.3: Calculate the gradient information of the prior image. Calculate the gradient magnitude of the prior image point by point, and substitute it into the weight function to obtain the weight information ω, which is mathematically expressed as:

[0131]

[0132] in, Let I be the gradient operator, σ be the edge smoothing parameter, and I be the gradient operator. net (x,y) represents a high-quality tomographic image I. net The direction vector;

[0133] The parameter σ is used to control the degree of edge smoothing. Different degrees of adjustment are made for different types of images. A smoothing filter is used to expand the edge width of the edge structure of the high-quality tomographic image, and then the edge direction vector e is calculated.

[0134] Step S5: Further denoise the tomographic image using the edge-guided total variation method. Calculate the total variation term TV for the high-quality tomographic image based on the total variation formula; calculate the edge-guided total variation term EGTV based on the edge direction vector e; and obtain the final correction term FTV = (1-ω)·EGTV + ω·TV through iterative iteration to obtain the final denoised tomographic image, as shown below. Figure 4 As shown. Figure 4 To illustrate the reconstruction results, from left to right are: complete angle reconstruction result, FBP method 45-angle reconstruction result, TV method combined 45-angle reconstruction result, and the present invention 45-angle reconstruction result.

[0135] Among them, the edge-guided variational term (EGTV) corrects the total variational term through the edge direction vector e, transforming the overall denoising problem into minimizing the noise in the parallel edge direction; the variational term weights are assigned according to ω to obtain the final variational correction term FTV.

[0136] The following is a detailed description of the invention:

[0137] S5.1: Using the total variation TV minimization method, mathematically expressed as:

[0138]

[0139] Where TV represents the total variation term;

[0140] Step S5.2: Through edge information I * The edge direction vector e is calculated and used to guide the direction of the total variation correction at the edge details. Mathematically, it is expressed as:

[0141]

[0142]

[0143] Where F represents the noise amplitude in the direction parallel to the edge structure, and EGTV represents the edge-guided total variation term. After correction, the problem of minimizing the total variation at the edge is transformed into minimizing the noise amplitude parallel to the edge structure, ensuring that the edge gradient is not affected while removing noise.

[0144] Step S5.3: Based on the weight ω, combine the total variation with the edge-guided total variation, and the final correction term FTV is expressed as:

[0145] FTV(I net)=TV(ω·I net )+EGTV((1-ω)·I net );

[0146] Step S5.4: The final denoised image is obtained through iterative calculation using the gradient descent method. The process includes:

[0147] 1) Initialize the step size parameter α = 0.01, and the number of iterations N. iter =40;

[0148] 2) Enter the loop, when n <N iter :

[0149] Calculate the variational correction term: FTV(I) net )=TV(ω·I net )+EGTV((1-ω)·I net )

[0150] Calculate the partial derivatives:

[0151] Using gradient descent: β = ||I net ||1 / ||d||1, I n+1 =I n -α·β·d

[0152] 3) Obtain the final denoised image I.

[0153] The NG-Net used in this invention employs a GAN network structure, such as... Figure 2 As shown, the network is divided into a generator network and a discriminator network. The generator network employs a modified U-Net network structure, consisting of an encoder and a decoder module. The encoder progressively reduces the feature map using downsampling convolutional layers with a stride of 2. The decoder progressively expands the feature map using deconvolution. A skip connection strategy is used to combine the raw features obtained by the encoder with the high-level features generated by the decoder. The backbone of the convolutional module uses a residual module, consisting of convolutional layers, batch normalization, and an activation function. The convolutional kernel size is 3×3, the stride is 1, and the activation function is LeakyReLU. The first and last layers of the generator network are connected using 1×1 convolutional layers to achieve the change in the number of channels.

[0154] The discriminant network uses four downsampling modules, followed by a global average pooling layer to reduce the input image size to 1×1. Finally, a fully connected layer is used with a Sigmoid activation function to determine whether the image is real or fake.

[0155] This invention can fully leverage the advantages of edge-guided total variation and combine it with the feature extraction capabilities of neural networks to remove sparse stripe noise while maintaining clear details of the reconstructed tomographic edge structure, thereby obtaining excellent sparse reconstructed images.

[0156] This invention also provides a noise reduction system suitable for sparse angle CT reconstructed tomographic images. This noise reduction system can be implemented by executing the steps of the noise reduction method for sparse angle CT reconstructed tomographic images. That is, those skilled in the art can understand the noise reduction method for sparse angle CT reconstructed tomographic images as a preferred embodiment of the noise reduction system. The system specifically includes:

[0157] Module M1: Performs sparse angle reconstruction using CT reconstruction algorithms to obtain sparse reconstructed tomographic images;

[0158] Module M2: Extracts fault noise images from the sparse reconstructed tomographic images using deep learning methods;

[0159] Module M3: Subtracts the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image;

[0160] Module M4: Calculates gradient information and edge structure information from the high-quality tomographic image;

[0161] Module M5: Using the edge-guided total variation method, further denoise the high-quality tomographic image to obtain the final denoised tomographic image.

[0162] The module M2 learns the mapping relationship from sparsely reconstructed tomographic images to sparsely noisy images by using an adversarial training strategy of alternating iterative generation and discrimination networks, thereby removing noise from the sparsely reconstructed images.

[0163] Module M2.1: The input is a sparsely reconstructed tomographic image, and the output is a tomographic noise image;

[0164] Module M2.2: Loss of the identification network D Defined as: the binary classification cross-entropy loss to determine whether a noisy image obtained through a generator network is real, mathematically expressed as:

[0165] l D =H[1,D(I noise )]+H[0,D(G(I sparse ))]

[0166] Where H represents cross-entropy, D represents the discriminant network, and I noise I represents the label of a real noisy image. sparse represents a sparsely reconstructed tomographic image, and G represents a generative network;

[0167] Module M2.3: Network MSE Loss MSEDefined as: the mean square error between the noisy image obtained from the input through the generator network and the true noise label, mathematically expressed as:

[0168]

[0169] Where G represents the generator network;

[0170] Module M2.4: Generative Network G Defined as the joint loss: a weighted sum of the MSE loss and the discrimination network loss, mathematically expressed as:

[0171] l G =l MSE +λl D

[0172] Where λ represents the hyperparameter that balances the weights of the MSE loss and the discriminant network loss;

[0173] Module M2.5: Trains the generator and discriminator networks iteratively using gradient descent until l... G Stablize.

[0174] The module M3 includes subtracting the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image I denoised by a neural network. net Mathematically, this is represented as:

[0175] I net =I sparse -G(I sparse ).

[0176] The module M4 includes,

[0177] Step M4.1: Extract the edge structure of the prior image using the Sobel operator;

[0178] Module M4.2: Expands the width of the prior image edge structure, applies a 3×3 Gaussian convolution template to smooth the edges, assigns 1 to pixels with a gray level greater than 0, and obtains edge information I. * ;

[0179] Module M4.3: Calculates the gradient information of the prior image, calculates the gradient magnitude of the prior image point by point, and substitutes it into the weight function to obtain the weight information ω, which is mathematically expressed as:

[0180]

[0181] in, Let I be the gradient operator, σ be the edge smoothing parameter, and I be the gradient operator. net (x,y) represents a high-quality tomographic image I. net The direction vector;

[0182] The parameter σ is used to control the degree of edge smoothing, and different adjustments are made for different types of images.

[0183] The module M5 includes,

[0184] M5.1: Using the total variation TV minimization method, mathematically expressed as:

[0185]

[0186] Where TV represents the total variation term;

[0187] Module M5.2: Through edge information I * The edge direction vector e is calculated and used to guide the direction of the total variation correction at the edge details. Mathematically, it is expressed as:

[0188]

[0189]

[0190] Where F represents the noise amplitude in the direction parallel to the edge structure, and EGTV represents the edge-guided total variation term. After correction, the problem of minimizing the total variation at the edge is transformed into minimizing the noise amplitude parallel to the edge structure, ensuring that the edge gradient is not affected while removing noise.

[0191] Module M5.3: Based on the weight ω, the total variation is combined with the edge-guided total variation, and the final correction term FTV is expressed as:

[0192] FTV(I net )=TV(ω·I net )+EGTV((1-ω)·I net );

[0193] Module M5.4: Iteratively calculates the final denoised image using gradient descent. The process includes:

[0194] 1) Initialize the step size parameter α = 0.01, and the number of iterations N. iter =40;

[0195] 2) Enter the loop, when n <N iter :

[0196] Calculate the variational correction term: FTV(I) net )=TV(ω·I net )+EGTV((1-ω)·I net )

[0197] Calculate the partial derivatives:

[0198] Using gradient descent: β = ||I net ||1 / ||d||1, I n+1 =I n -α·β·d

[0199] 3) Obtain the final denoised image I.

[0200] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0201] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A noise reduction method suitable for sparse angle CT reconstructed tomographic images, characterized in that, include: Step S1: Perform sparse angle reconstruction using a CT reconstruction algorithm to obtain a sparse reconstructed tomographic image; Step S2: Extract features using deep learning methods to extract fault noise images from the sparse reconstructed tomographic images; Step S3: Subtract the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image; Step S4: Using the high-quality tomographic image as the prior image, calculate the gradient information and edge structure information of the prior image; Step S5: Further denoise the high-quality tomographic image using the edge-guided total variation method to obtain the final denoised tomographic image; Step S2 employs an adversarial training strategy involving alternating iterative generation and discrimination networks to learn the mapping relationship from sparsely reconstructed tomographic images to sparsely noisy images, thereby removing noise from the sparsely reconstructed images. Step S2.1: The input is a sparsely reconstructed tomographic image, and the output is a tomographic noise image; Step S2.2: Identify network loss Defined as: the binary classification cross-entropy loss to determine whether a noisy image obtained through a generator network is real, mathematically expressed as: Where H represents cross-entropy and D represents the discriminant network. Labels representing real noise images. represents a sparsely reconstructed tomographic image, and G represents a generative network; Step S2.3: Network MSE Loss Defined as: the mean square error between the noisy image obtained from the input through the generator network and the true noise label, mathematically expressed as: Where G represents the generator network; Step S2.4: Generate the network Defined as the joint loss: a weighted sum of the MSE loss and the discrimination network loss, mathematically expressed as: in, The hyperparameters representing the weights that balance the MSE loss and the discriminant network loss; Step S2.5: Train the generator network and the discriminator network alternately and iteratively using gradient descent until... Stablize; Step S5 includes, S5.1: Using the total variation TV minimization method, mathematically expressed as: in, Represented as total variation term; Step S5.2: Using edge information The edge direction vector is calculated. This is used to guide the direction of the total variation correction at edge details, and is mathematically represented as: in, Represented as the noise amplitude in the direction parallel to and along the edge structure. Represented as edge-guided total variation, after correction, the problem of minimizing the total variation at the edge is transformed into minimizing the amplitude of noise parallel to the edge structure, ensuring that the edge gradient is not affected while removing noise; Step S5.3: Based on weights By combining total variation with edge-guided total variation, the final correction term Represented as: ； Step S5.4: The final denoised image is obtained through iterative calculation using the gradient descent method. The process includes: (1) Initialize step size parameters Number of iterations ; (2) Enter the loop, when : Calculate the variational correction term: Calculate the partial derivatives: Using gradient descent: , (3) Obtain the final denoised image .

2. The noise reduction method for sparse angle CT reconstructed tomographic images according to claim 1, characterized in that, Step S3 includes subtracting the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image denoised by a neural network. Mathematically, this is represented as: 。 3. The noise reduction method for sparse angle CT reconstructed tomographic images according to claim 1, characterized in that, Step S4 includes, Step S4.1: Extract the edge structure of the prior image using the Sobel operator; Step S4.2: Expand the width of the prior image edge structure, apply a 3×3 Gaussian convolution template to smooth the edges, and assign 1 to pixels with a gray level greater than 0 to obtain edge information. ; Step S4.3: Calculate the gradient information of the prior image, calculate the gradient magnitude of the prior image point by point, and substitute it into the weight function to obtain the weight information. Mathematically, this is represented as: in, Represented as the gradient operator, Represented as edge smoothing parameters, Represented as high-quality tomographic images The direction vector; The parameter σ is used to control the degree of edge smoothing, and different adjustments are made for different types of images.

4. A noise reduction system suitable for sparse angle CT reconstructed tomographic images, characterized in that, include: Module M1: Performs sparse angle reconstruction using CT reconstruction algorithms to obtain sparse reconstructed tomographic images; Module M2: Extracts fault noise images from the sparse reconstructed tomographic images using deep learning methods; Module M3: Subtracts the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image; Module M4: Using the high-quality tomographic image as a priori image, calculate the gradient information and edge structure information of the priori image; Module M5: The high-quality tomographic image is further denoised using the edge-guided total variation method to obtain the final denoised tomographic image; The module M2 learns the mapping relationship from sparsely reconstructed tomographic images to sparsely noisy images by using an adversarial training strategy of alternating iterative generation and discrimination networks, thereby removing noise from the sparsely reconstructed images. Module M2.1: The input is a sparsely reconstructed tomographic image, and the output is a tomographic noise image; Module M2.2: Identifying the Loss of the Network Defined as: the binary classification cross-entropy loss to determine whether a noisy image obtained through a generator network is real, mathematically expressed as: Where H represents cross-entropy and D represents the discriminant network. Labels representing real noise images. represents a sparsely reconstructed tomographic image, and G represents a generative network; Module M2.3: Network MSE Loss Defined as: the mean square error between the noisy image obtained from the input through the generator network and the true noise label, mathematically expressed as: Where G represents the generator network; Module M2.4: Generative Network Defined as the joint loss: a weighted sum of the MSE loss and the discrimination network loss, mathematically expressed as: in, The hyperparameters representing the weights that balance the MSE loss and the discriminant network loss; Module M2.5: Trains the generator network and the discriminator network alternately and iteratively using gradient descent until... Stablize; The module M5 includes, M5.1: Uses the total variation TV minimization method, mathematically expressed as: in, Represented as total variation term; Module M5.2: Through edge information The edge direction vector is calculated. This is used to guide the direction of the total variation correction at edge details, and is mathematically represented as: in, Represented as the noise amplitude in the direction parallel to and along the edge structure. Represented as edge-guided total variation, after correction, the problem of minimizing the total variation at the edge is transformed into minimizing the amplitude of noise parallel to the edge structure, ensuring that the edge gradient is not affected while removing noise; Module M5.3: Based on weight By combining total variation with edge-guided total variation, the final correction term Represented as: ； Module M5.4: Iteratively calculates the final denoised image using gradient descent. The process includes: (1) Initialize step size parameters Number of iterations ; (2) Enter the loop, when : Calculate the variational correction term: Calculate the partial derivatives: Using gradient descent: , (3) Obtain the final denoised image .

5. The noise reduction system for sparse-angle CT reconstructed tomographic images according to claim 4, characterized in that, The module M3 includes subtracting the sparse reconstructed tomographic image from the tomographic noise image to obtain a high-quality tomographic image denoised by a neural network. Mathematically, this is represented as: 。 6. The noise reduction system for sparse angle CT reconstructed tomographic images according to claim 4, characterized in that, The module M4 includes, Step M4.1: Extract the edge structure of the prior image using the Sobel operator; Module M4.2: Expands the width of the prior image edge structure, applies a 3×3 Gaussian convolution template to smooth the edges, and assigns 1 to pixels with a gray level greater than 0 to obtain edge information. ; Module M4.3: Calculates the gradient information of the prior image, calculates the gradient magnitude of the prior image point by point, and substitutes it into the weight function to obtain the weight information. Mathematically, this is represented as: in, Represented as the gradient operator, Represented as edge smoothing parameters, Represented as high-quality tomographic images The direction vector; The parameter σ is used to control the degree of edge smoothing, and different adjustments are made for different types of images.

Images