A high-precision CT metal artifact correction method based on self-supervised learning

Abstract

The application discloses a high-precision CT metal artifact correction method based on self-supervised learning, and belongs to the technical field of computer medical imaging, specifically: a data set for training a network is created, and the data set is divided into a training set and a test set; a model-driven image domain subnetwork based on a tight frame, a data-driven sine domain subnetwork based on a Transformer, and a novel coupling mechanism for connecting the two subnetworks are constructed; a loss function is designed for optimizing the coupled model-data double-driven network; the data set and the loss function are used to train and optimize the coupled double-driven network until the training is completed, and an optimal network model is obtained; and the test set image is input into the network model trained in step 3, so that a metal artifact correction image is obtained. The application has high-precision metal artifact correction performance while having a small amount of network parameters, and relieves the demand of the training network for paired CT images.

Description

Technical Field

[0001] This invention belongs to the field of computer medical imaging technology, and specifically relates to a CT metal artifact correction method based on self-supervised learning, which can be used for high-precision CT metal artifact correction in clinical practice. Background Technology

[0002] Computed tomography (CT) is a non-destructive testing technique widely used in medicine, industry, and many other fields. In medicine, this technique uses X-rays to scan specific parts of the human body. Based on the different absorption capacities of various human tissues to X-rays, cross-sectional images are reconstructed from the scan data. When metal or other high-density materials are present in the human body, they contaminate the corresponding projection data, resulting in metal artifacts in the reconstructed image. The presence of metal artifacts severely degrades image quality, thus affecting doctors' diagnoses. Therefore, researching techniques for correcting CT metal artifacts and improving the quality of reconstructed images is of great significance in the field of radiological diagnosis.

[0003] Existing CT metal artifact correction techniques are mainly divided into traditional metal artifact correction methods and deep learning-based metal artifact correction methods. Traditional metal artifact correction methods are further divided into three categories: physical effect correction, sinusoidal domain (X-ray projection domain) interpolation, and iterative reconstruction. Physical effect correction models the underlying beam hardening physical effects and then directly corrects the metal-affected areas in the sinusoidal domain. When the metal implant is a high-atomic-number metal, the correction will produce unsatisfactory results. Sinusoidal domain interpolation uses various interpolation techniques to replace the metal-contaminated parts in the original projection data. However, this method easily introduces secondary artifacts, causing severe structural deformation in areas with severe metal artifacts in the reconstructed image. Iterative reconstruction techniques reconstruct clean CT images from the observed sinusoidal plot; this method is usually very time-consuming and involves cumbersome parameter tuning. Based on the domain knowledge utilized, deep learning-based methods are mainly divided into three categories: image domain methods, sinusoidal domain methods, and dual-domain methods. Image domain networks treat metal artifact correction as an image restoration problem, directly correcting metal artifacts without using sinusoidal domain data. In image domain networks, consistency constraints are often ignored. The sinusoidal domain method replaces damaged data with data generated by a deep convolutional neural network to produce metallic trace regions, but this method typically introduces severe secondary artifacts. To address the problems of the aforementioned single-domain methods, the dual-domain method emerged and has become the mainstream approach. However, current dual-domain methods generally lack model interpretability, have a large number of network parameters, and require paired CT images.

[0004] In summary, although there are numerous existing methods for correcting CT metal artifacts, each has its own limitations. Therefore, based on existing methods, a novel CT metal artifact correction method is proposed to further improve the quality of reconstructed images. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing methods by proposing a high-precision CT metal artifact correction method based on self-supervised learning. This method mainly solves the problems of insufficient interpretability, large GPU space required for training the network, large number of network parameters, need for paired CT images, and low accuracy of metal artifact correction in previous algorithms. The goal is to reduce network parameters, improve framework interpretability, alleviate the need for paired CT images, and improve the accuracy of metal artifact correction.

[0006] The present invention provides a high-precision CT metal artifact correction method based on self-supervised learning, comprising the following steps:

[0007] (1) Construct a self-supervised CT metal artifact correction network model, namely a coupled model-data dual-drive network model. This network model includes a Transformer-based data-driven sinusoidal domain sub-network (SS-Net) and a compact frame-based model-driven image domain sub-network (IM-Net). The IM-Net contains a prior artifact attention threshold generation module (PATG) for coupling the two networks.

[0008] (2) Training the network:

[0009] (2.1) Prepare the dataset and synthesize the metal contamination s using the collected artifact-free sine maps s. ma By performing filtering and backprojection operations on them, artifact-free CT image x and metal-contaminated CT image x are obtained. ma The CT images and sine waves are cropped to a set size, and then divided into training and testing sets. The maximum number of training iterations is set, and the training parameters are initialized.

[0010] (2.2) The sine curve of the metal artifact s ma and metallic traces m Perform linear interpolation to obtain s LI Then obtain s LI and metallic traces m The data are fed into a Transformer-based data-driven sinusoidal domain subnetwork, and the output is an enhanced sine curve s. out , for s out The reconstructed CT image x is obtained by performing a filtering back projection operation. out ;

[0011] (2.3) Calculate the loss of the sinusoidal domain subnetwork. If the maximum number of training iterations for the sinusoidal domain network has not been reached, repeat step 2.2. Update the network parameters, and the optimal network model of the sinusoidal domain subnetwork is obtained after training.

[0012] (2.4) Input s into the trained sinusoidal domain network model ma and tr m Network output x out and x out Image x with metal contamination ma Subtracting element by element yields the sinusoidal domain metal artifact information e. s Simultaneously, using the image obtained through linear interpolation and beam hardening correction, metal artifacts e are extracted. LI and e BHC Metal artifacts e s e BHC e LI x ma The non-metallic mask m is input into a model-driven image domain subnetwork based on a tight frame, and the output is a CT image with reduced artifacts.

[0013] (2.5) Calculate the image domain loss. If the maximum number of training iterations has not been reached, repeat step 2.4 to update the network parameters.

[0014] (3) Input e into the constructed optimal network model in the image domain. s e BHC e LI x ma And m, the network output is the CT image after metal artifact correction.

[0015] Due to the adoption of the above technical solution, the technical effects achieved by the present invention are as follows:

[0016] Unlike previous joint training of dual-domain metal artifact correction networks, this invention adopts a separate training method based on the proposed coupling mechanism to train the two sub-networks, thereby reducing the GPU memory required during training. Specifically, the data-driven sinusoidal domain sub-network is trained first, and then the optimal model of its sinusoidal domain network is connected to the model-driven image domain sub-network through the proposed coupling module, and then the image domain network is trained. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the present invention;

[0018] Figure 2 This is a structural diagram of the data-driven sinusoidal domain subnetwork based on Transformer according to the present invention;

[0019] Figure 3 This is a structural diagram of the model-driven image domain subnetwork based on a tight frame according to the present invention;

[0020] Figure 4 This is a schematic diagram showing the comparison results of different artifact correction methods on the synthetic dataset in the embodiment;

[0021] Figure 5 This is a schematic diagram showing the comparison results of different artifact correction methods on the clinical dataset in the example. Detailed Implementation

[0022] The following will describe clearly and completely, with reference to the accompanying drawings, the implementation steps of the high-precision CT metal artifact correction algorithm based on self-supervised learning of the present invention, and demonstrate the effectiveness of the method of the present invention through experimental results.

[0023] This invention provides a high-precision CT metal artifact correction method based on self-supervised learning, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0024] The metal contamination s is obtained by simulating the collected artifact-free sine curve s. ma Then, the artifact-free sine wave and the metal contamination sine wave are subjected to filtered back projection (FBP) to obtain the corresponding artifact-free CT image x and the metal contamination CT image x. ma All images were resized to 416×416, and the sine curve was adjusted to 641×640.

[0025] Construct a data-driven sinusoidal domain subnetwork based on Transformer. See also Figure 2 The network consists of a Transformer-based trace-aware network (Tr-AwareNet) and two sinusoidal refinement networks (SinoNet) with the same U-Net architecture. The two identical SinoNets form a dual-branch structure.

[0026] The input to the sinusoidal domain subnetwork is the synthesized metal contamination sinusoid s. ma and metallic traces m Given a sine curve of metal contamination s ma Linear interpolation is performed to obtain s LI Then, the input is fed into Tr-AwareNet, and a residual connection is set between the input and output of Tr-AwareNet to obtain the output result s. prior Next, we will... prior The images are concatenated with two different metal trace channels and input into two different branches constructed from SinoNet, resulting in two thinned sine waves. FBP is then applied to each of these waves to obtain the reconstructed CT image x. rec1 and x rec2The following section provides a detailed introduction to the proposed sinusoidal domain subnetwork:

[0027] See Figure 2 Tr-AwareNet consists of three parts: an encoder, a dual-branch module, and a decoder. The input sine wave is s. LI The intermediate feature f is obtained after encoding. m f m The inputs are fed into the remote branch and the short branch respectively to obtain the output f. l and f s f l and f s f is obtained by fusion fuse That is, take the metallic trace portion of the long-range branch and the non-metallic trace portion of the short-range branch. Then, f fuse The signal is fed into the decoder to obtain the final output f. last The residual connection can be represented as s. prior =f last +s LI The final output result s is obtained. prior .

[0028] The long-range branch is the core of Tr-AwareNet, mainly composed of two modules: a trace-aware multi-head self-attention (Tr-Aware MSA) module and a fully connected feedforward network (FFN). Feature vector f m Corresponding matrix form Divide F into d feature blocks. These d feature blocks can cover the entire feature map F. These d feature blocks can be unfolded into a one-dimensional feature map F. in =[F1,...,F d The process involves calculating Q, K, and V using the standard multi-head self-attention method. This process can be represented as:

[0029]

[0030] Where LN(·) represents layer normalization, and This represents the projection matrix of the h-th head in the Transformer. Inputting Q, K, and V into the Tr-Aware MSA module, the output can be represented as:

[0031]

[0032] Where MSA(·) represents the Tr-Aware MSA network step. Next, we will... When input into an FFN network, the output can be represented as:

[0033]

[0034] Where the output F out =[F′1,...,F′ d [Can be fused into a feature matrix F] l Its corresponding vector form is f l .

[0035] The specific steps of Tr-Aware MSA are as follows: First, analyze the Tr traces of metal. m Matrix form Tr m The operation yields a non-metallic trace, i.e., Tr. n =1-Tr m To match the size of feature F, Tr n Resize and normalize to obtain Tr′ n Its corresponding vector form is tr′ n Then, similarly for Tr′ n Divide the data into d blocks, calculate the average of the elements in each block, and stack these blocks into a vector. Then, thresholding is applied to u as follows:

[0036]

[0037] Where ρ is the threshold. Next, stacking d u values ​​yields matrix U′. Assuming Tr-Aware MSA has H heads, the self-attention calculation for the h-th head can be expressed as:

[0038]

[0039] in The output is used for normalization, and σ is a very small negative value. All header outputs are concatenated by channels and linearly projected to produce the final output. This process can be represented as:

[0040] MultiHead(Q′,K′,V′)=Concat(head1,...,head h )Z A (6)

[0041] head h =Attention(Q) h ,K h V h Z A It is the projection matrix. Q′, K′, and V′ are the channel concatenations of all their corresponding heads.

[0042] See Figure 2The proposed dual-branch architecture consists of two SinoNets, and based on this structure, an FBP consistency loss can be constructed. The outputs of the aforementioned Tr-AwareNet are then... prior The two channels are stitched together with two different metallic traces, and then the two dual-channel sinusoids are input into the corresponding two branches to obtain two thinned sinusoids. FBP reconstruction is then performed on the two thinned sinusoids to obtain two reconstructed CT images x. rec1 and x rec2 In this invention, the reconstructed CT image obtained from the first branch is used as the output of the sinusoidal domain subnetwork.

[0043] To optimize the Transformer-based data-driven sinusoidal domain subnetwork, several loss functions are introduced, including sinusoidal domain loss, reconstruction loss, and FBP consistency loss. Their specific forms are as follows:

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] in It is a loss function in the sinusoidal domain. and It is a reconstruction loss. This is the FBP consistency loss. prior Represents the output of the Tr-AwareNet network. Represents the FBP operation, x rec1 and x rec2 These are the FBP reconstruction results for the two branches, and γ1 and γ2 are the weight parameters for balancing different loss terms.

[0050] Figure 3 The image domain network is driven by a model based on a compact frame. A CT image with metal contamination can be divided into two parts: a non-metallic part and a metallic part. The goal is to preserve the non-metallic part while removing the metallic part. Therefore, the metal artifact decomposition model can be constructed as follows:

[0051] m⊙y=m⊙x+m⊙e (12)

[0052] Where m is the non-metallic mask, x represents the CT image without metal artifacts, e represents the metal artifact, and ⊙ represents element-wise multiplication. To estimate x and e from y, an optimization model is constructed using prior knowledge of x and e through a regularizer. For the metal artifact component, a tight reference frame is used. To characterize the structural information of metallic artifacts. The optimized model can be represented as:

[0053]

[0054] The first item is a data fidelity item used to improve the accuracy of decomposition; the second and third items are regularization items.

[0055] To solve this optimization model, the proximal gradient operator is used to alternately update x and e. In the (t+1)th iteration, the updates of x and e can be expressed as follows:

[0056] 1) In the (t+1)th iteration, update the CT image: given a CT image x with reduced metal artifacts. (t) and estimated metal artifacts e (t) By calculating the quadratic approximation of the optimization model defined in equation (13), x can be iteratively updated as follows:

[0057]

[0058] in η1 is the step size. The update rule for x can be expressed as:

[0059]

[0060] in It is a proximal operator related to the prior term R1(·).

[0061] 2) In the (t+1)th iteration, update the metal artifact: given x (t+1) and e (t) By calculating the quadratic approximation of the optimization model defined in equation (13), e can be iteratively updated:

[0062]

[0063] in η2 is the step size. Similar to the update rules for x and e:

[0064]

[0065] Because W satisfies the compact property The above expression can be rewritten as:

[0066] e (t+1) =W Tsoft(We (t+0.5) ,ε) (18)

[0067] Where soft(μ,ε)=sign(μ)max(|μ|-ε,0) is a soft thresholding function. Similar to image updates, e (t+0.5) =(1-η2m)⊙e (t) +η2m⊙(yx (t+1) In this invention, a proposed prior artifact attention threshold generation module is used to generate the threshold ε.

[0068] Network architecture design:

[0069] 1) X-Net: In the (t+1)th stage, using x (t) and e (t) Get x (t+0.5) Then send it in Implemented in the network Operator, to obtain output x (t+1) See also Figure 3 ,network It is a residual convolution structure.

[0070] 2) E-Net: Similar to x, e (t+0.5) Input In, used to achieve The operator yields the output e. (t+1) See also Figure 3 , It is a process of tightening the standard frame.

[0071] 3) To train the image domain network, the following loss function was designed:

[0072]

[0073] Where x is the output of the FBP operation of s, and w1, w2, w3 and w4 are weight parameters that balance different loss terms.

[0074] A novel coupling mechanism, the Prior Artifact Attention Threshold Generation (PATG) module, is constructed to connect the image domain subnetwork and the sinusoidal domain subnetwork. PATG, located within E-Net, consists of an attention mechanism and convolutional layers. The input to this module is a four-channel metallic artifact containing e-values ​​generated iteratively by the network. (t+0.5) Artifacts extracted by linear interpolation method e LI Artifacts extracted by beam hardening correction method BHC Artifacts e extracted from sinusoidal domain networks s .

[0075] The training parameters for the above embodiment are set as follows: the Adam optimizer is used to train the network, the epoch for the sinusoidal domain network is set to 150, and the initial learning rate is set to 1×10⁻⁶. -4 The learning rate is halved at epochs 50 and 100, with the sinusoidal domain network having an epoch of 100 and an initial learning rate of 3 × 10⁻⁶. -4 The loss function weights are halved at the 50th epoch. The weight parameters of the network's loss function are set as follows: γ1 = γ2 = 1, w1 = 0.1, w2 = w3 = 5 × 10⁻⁶. -4 w4 = 0.01 and ρ = 0.2, H = 8 and c k =32. The network is implemented using the deep learning framework PyTorch, based on an NVIDIA RTX 3080Ti GPU and an i9-10850K CPU (3.60Hz).

[0076] To accurately analyze the effectiveness of the method proposed in this invention, the metal contamination CT images were corrected based on the above method, and the advanced self-supervised SSCNet and unsupervised algorithm ADN were used as comparison algorithms.

[0077] Figure 4 The images in the diagram are, in order, a reference image, an input metal contamination image, an AND-corrected image, an SSCNet-corrected image, and an image corrected by the method of this invention.

[0078] For the calibration of clinical data, the model trained on the aforementioned synthetic dataset was directly tested. Figure 5 The images in the table are, in order, the input metal contamination image, the LI-corrected image, the AND-corrected image, the SSCNet-corrected image, and the image corrected by the method of this invention.

[0079] Qualitative and quantitative results show that the correction results of this invention are more accurate, effectively removing complex metal artifacts, especially eliminating residual artifacts at metal boundaries and preserving more tissue structure. Meanwhile, the correction results from clinical datasets demonstrate that this invention has a certain degree of generalization ability.

[0080] In summary, the present invention has the following advantages compared with the prior art:

[0081] 1) Existing MAR (Metal Artifact Correction) techniques typically train metal artifact correction networks using supervised learning. This method requires paired CT images with consistent anatomical structures, which are difficult to obtain in practice. This invention employs a self-supervised learning-based approach to train the network, requiring only sinusoidal domain images.

[0082] 2) This invention constructs a trace-aware network based on Transformer, namely Tr-AwareNet. In this network, traces are used to guide the self-attention calculation in Transformer, thereby more effectively completing the sine curve. At the same time, in the image domain, a tight frame is used to characterize the structural information of metal artifacts.

[0083] 3) Unlike previous dual-domain methods, this invention proposes a novel coupling mechanism to connect the model-driven image domain subnetwork based on a tight-frame architecture and the data-driven sinusoidal domain subnetwork based on Transformer. This coupling mechanism is implemented by a module called PATG, which generates a reliable threshold by utilizing multiple types of information, including sinusoidal domain information. Furthermore, the artifact attention module in PATG adaptively reassigns the weights of each metal artifact by utilizing the interrelationships between these input metal artifacts. This allows PATG to fully utilize and fuse multiple types of information, thereby significantly improving MAR performance.

[0084] 4) This invention employs a separate training approach: first, a Transformer-based data-driven sinusoidal domain subnetwork is trained, then coupled to a compact frame-based model-driven image domain subnetwork, and finally, the image domain network is trained. This sequential training method saves more GPU memory. This strategy allows the network to be trained even with limited GPU resources.

[0085] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention; however, these modifications and changes based on the spirit of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A high-precision CT metal artifact correction method based on self-supervised learning, comprising the following steps: S1 constructs a self-supervised metal artifact correction network model, namely a coupled model-data dual-driven network model. The network model includes a Transformer-based data-driven sinusoidal domain sub-network SS-Net and a compact frame-based model-driven image domain sub-network IM-Net. IM-Net contains a prior artifact attention threshold generation module PATG for coupling the two sub-networks. S2 training network: S2.1 Prepare the dataset and synthesize the metal-contaminated sine map s from the collected artifact-free sine maps s. ma By performing a filtering and backprojection operation, artifact-free CT image x and metal-contaminated CT image x are obtained. ma The CT images and sine waves are cropped to a set size and divided into training and testing sets. The maximum number of training iterations is set, and the training parameters are initialized. S2.2 will display the sine curve of metal pollution. ma and metallic traces m Perform linear interpolation to obtain s LI , will s LI and metallic traces m The input is fed into a Transformer-based data-driven sinusoidal domain subnetwork, and the output is an enhanced sine curve s. out , for s out The reconstructed CT image x is obtained by performing a filtering back projection operation. out ; S2.3 Calculate the loss of the sinusoidal domain subnetwork. If the maximum number of training iterations for the sinusoidal domain network has not been reached, repeat step S2.2 to update the network parameters. The optimal network model of the sinusoidal domain network is obtained after training ends. S2.4 Input s into the trained sinusoidal subnetwork model ma and tr m Network output x out and x out CT images contaminated with metal artifacts x ma Element-wise subtraction yields sinusoidal domain metal artifact information e s Simultaneously, linear interpolation and beam hardening methods are used to obtain the corrected image, and metal artifacts are extracted from both corrected images. LI and e BHC Metal artifacts e s e BHC e LI x ma The non-metallic mask m is input into the model-driven image domain subnetwork IM-Net based on the tight frame, and the output is a CT image with reduced artifacts; S2.5 Calculate the image domain loss. If the maximum number of training iterations has not been reached, repeat step 2.4 to update the network parameters. S3 inputs e into the constructed optimal network model in the image domain. s e BHC e LI x ma And m, the network output is the CT image after metal artifact correction.

2. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 1, characterized in that, Step (1) constructing a self-supervised learning metal correction network: specifically includes the following processes: constructing a model-driven network, a data-driven network, and a coupling mechanism; the model-driven network is a model-driven image domain sub-network based on a tight frame: a metal artifact decomposition model is established, a tight frame learning model is constructed based on the tight frame representation model, the problem is transformed into an easily solvable sub-problem using the alternating direction multiplier method, and the problem is solved using the proximal gradient technique. Then, each operator of the iterative algorithm is converted into a corresponding network module and expanded into a deep neural network architecture to construct a model-driven metal artifact correction network; the data-driven network is a data-driven sine domain sub-network based on Transformer: a trace-aware Transformer sine domain network is proposed in the data-driven sine domain network based on Transformer, using non-metallic traces to guide the multi-head self-attention mechanism in Transformer, thereby ensuring that the network learning will not utilize areas with severe metal contamination in the sine image; the coupling mechanism: is implemented by a module called the prior artifact attention threshold generation module PATG, in which the artifact attention module adaptively reassigns the weights of each metal artifact by utilizing the interrelationships between the input metal artifacts.

3. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 2, characterized in that: The Transformer-based data-driven sinusoidal domain subnetwork SS-Net consists of a Transformer-based trace-aware network Tr-AwareNet and two sinusoidal refinement networks SinoNet with the same U-Net architecture. The two identical SinoNets form a dual-branch structure. The input to the sinusoidal domain subnetwork is a synthesized metal-contaminated sinusoid s. ma and metallic traces m Given the sine curve of metal pollution s ma Linear interpolation is performed to obtain s LI Then, the input is fed into Tr-AwareNet, and a residual connection is set between the input and output of Tr-AwareNet to obtain the output result s. prior , will s prior The images are concatenated with two different metal trace channels and input into two different branches constructed from SinoNet, resulting in two thinned sine waves. FBP is then applied to each of these waves to obtain the reconstructed CT image x. rec1 and x rec2 .

4. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 3, characterized in that: Tr-AwareNet consists of three parts: an encoder, a dual-branch module, and a decoder, with the input sine wave s. LI The intermediate feature f is obtained after encoding. m f m The inputs are fed into the remote branch and the short branch respectively to obtain the output f. l and f s , will f l and f s f is obtained by fusion fuse That is, take the metallic trace portion of the long-range branch and the non-metallic trace portion of the short-range branch, and then, f fuse The output f is fed into the decoder to obtain the final output f. last The residual connection can be represented as s prior =f last +s LI The final output result s is obtained. prior .

5. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 4, characterized in that: The remote branch is the core of Tr-AwareNet, consisting of two modules: a trace-aware multi-head self-attention Tr-AwareMSA module and a fully connected feedforward network FFN, with feature vector f. m Corresponding matrix form Divide F into d feature blocks. These d feature blocks can cover the entire feature map F. These d feature blocks can be unfolded into a one-dimensional feature map F. in =[F1,...,F d The process of obtaining Q, K, and V using the standard multi-head self-attention method can be expressed as follows: Where LN(·) represents layer normalization, and Representing the projection matrix of the h-th head in the Transformer, with Q, K, and V input into the Tr-Aware MSA module, the output can be represented as: Where MSA(·) represents the Tr-Aware MSA network step, next, will When input into an FFN network, the output can be represented as: Where the output F out =[F1′,...,F′ d [Can be fused into a feature matrix F] l Its corresponding vector form is f l .

6. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 5, characterized in that: The specific steps of Tr-Aware MSA are as follows: First, analyze the Tr traces of metal. m Matrix form Tr m The operation yields a non-metallic trace, i.e., Tr. n =1-Tr m In order to match the size of feature F, for Tr n Resize and normalize to obtain Tr′ n Its corresponding vector form is tr′ n Then, similarly for Tr′ n Divide the data into d blocks, calculate the average of the elements in each block, and stack these blocks into a vector. Then, thresholding is applied to u as follows: Where ρ is the threshold, and then, stacking d u values ​​yields matrix U′. Assuming Tr-Aware MSA has H heads, the self-attention calculation for the h-th head can be expressed as: in Used for normalization, σ is a very small negative value. All header outputs are concatenated by channels and linearly projected to produce the final output. This process can be represented as: MultiHead(Q′,K′,V′)=Concat(head1,...,head h )Z A (6) head h =Attention(Q) h ,K h V h Z A It is a projection matrix, where Q′, K′, and V′ are the channel splices of all their corresponding heads.

7. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 1, characterized in that: The Transformer-based data-driven sinusoidal domain subnetwork is optimized by introducing loss functions, including sinusoidal domain loss, reconstruction loss, and FBP consistency loss, as follows: in It is a loss function in the sinusoidal domain. and It is a reconstruction loss. It is the FBP consistency loss, s prior Represents the output of the Tr-AwareNet network. Represents the FBP operation, x rec1 and x rec2 These are the FBP reconstruction results for the two branches, and γ1 and γ2 are the weight parameters for balancing different loss terms.

8. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 2, characterized in that: The construction of the image domain subnetwork IM-Net based on a compact frame model: A CT image with metal contamination can be divided into two parts: a non-metallic part and a metallic part. The goal is to preserve the non-metallic part while removing the metallic part. The metal artifact decomposition model can be constructed as follows: m⊙y=m⊙x+m⊙e (12) Where m is the non-metallic mask, x represents the CT image without metal artifacts, e represents the metal artifact, and ⊙ represents element-wise multiplication; to estimate x and e from y, a regularizer is used to construct an optimization model using prior knowledge of x and e; for the metal artifact part, a tight frame is used. To characterize the structural information of metal artifacts, the optimized model can be represented as: The first item is a data fidelity item used to improve the accuracy of decomposition; the second and third items are regularization items. To solve this optimization model, the proximal gradient operator is used to alternately update x and e. In the (t+1)th iteration, the updates of x and e can be expressed as follows: In the (t+1)th iteration, update the CT image: given a CT image x with reduced metal artifacts. (t) and estimated metal artifacts e (t) By calculating the quadratic approximation of the optimization model defined in equation (13), x can be iteratively updated as follows: in η1 is the step size, and the update rule for x can be expressed as: in It is a proximal operator related to the prior term R1(·). In the (t+1)th iteration, update the metal artifact: given x (t+1) and e (t) By calculating the quadratic approximation of the optimization model defined in equation (13), e can be iteratively updated: in η2 is the step size, similar to the update rules for x and e: Because W satisfies the compact property The above expression can be rewritten as: e (t+1) =W T soft(We (t+0.5) ,ε) (18) Where soft(μ,ε)=sign(μ)max(|μ|-ε,0) is a soft thresholding function, similar to image updating, e (t+0.5) =(1-η2m)⊙e (t) +η2m⊙(yx (t+1) The threshold ε is generated using a proposed prior artifact attention threshold generation module.

9. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 2, characterized in that: The network architecture design of the model-driven image domain subnetwork based on the compact frame includes X-Net and E-Net. X-Net: In the (t+1)th stage, it utilizes x... (t) and e (t) Get x (t+0.5) Then send it in Implemented in the network Operator, to obtain output x (t+1) ;network It is a residual convolutional structure; the E-Net is similar to x, e (t+0.5) Input In, used to achieve The operator yields the output e. (t+1) , It is a compact frame process; the PATG is located in E-Net and consists of an attention mechanism and convolutional layers. The input of this module is a four-channel metal artifact, which contains e generated by network iteration. (t+0.5) Artifacts extracted by linear interpolation method e LI Artifacts extracted by beam hardening correction method BHC Artifacts e extracted from sinusoidal domain networks s .

10. The high-precision CT metal artifact correction method based on self-supervised learning according to claim 4, characterized in that: The loss function for training the model-driven image domain subnetwork IM-Net based on a tight frame is as follows: Where x is the output of the FBP operation of s, and w1, w2, w3 and w4 are weight parameters that balance different loss terms.

Images