Metal artifact correction network model and metal artifact removal method

Abstract

A metal artifact correction network model and a metal artifact removal method. The metal artifact correction network model is provided with: a learnable artifact coder, configured to perform extraction and feature coding on artifact information in an uncorrected metal artifact-containing CT image to obtain a corresponding multi-scale artifact coded feature map; a learnable artifact remover, configured to remove the artifact information in the uncorrected metal artifact-containing CT image; and a learnable artifact generator, configured to generate artifact information on an artifact-free CT image paired with the uncorrected metal artifact-containing CT image, wherein the network structure of the learnable artifact remover is the same as that of the learnable artifact generator. The present invention can obtain a metal artifact-reduced image with a good artifact correction effect.

Description

A metal artifact correction network model and its metal artifact removal method Technical Field

[0001] This invention relates to the field of medical imaging quality control technology, and in particular to a metal artifact correction network model and a method for removing metal artifacts. Background Technology

[0002] Computed tomography (CT) is an imaging technique that obtains tomographic images of an object's cross-sectional structure in a non-invasive, non-open manner. Due to its high imaging speed and high resolution, CT has been widely used in medical imaging diagnosis and is gradually becoming the preferred method for modern clinical medical imaging examinations. However, during CT scans, the implantation of high atomic number metallic materials in patients causes drastic attenuation of X-rays. Simultaneously, due to various physical effects such as X-ray beam hardening, scattering, and photon starvation, the signal received by the detector exhibits abrupt changes, resulting in discontinuous projection data. These abnormal signals cause alternating bright and dark stripe artifacts in the reconstructed image, known as metallic artifacts.

[0003] Existing metal artifact correction methods can be broadly categorized into three types: projection completion methods, iterative reconstruction methods, and deep learning methods. The first type, projection completion methods, involves interpolating the metal projection trajectory to recover the metal-damaged projection data in the CT projection. Filtering and backprojecting the corrected projection then yields an image free of metal artifacts. This method is generally simple and effective, but because it discards hidden structural information from the metal projection trajectory, it often results in overly smoothed details around the metal, especially when the metal shape is large or numerous. The second type, iterative reconstruction, obtains high-quality images by iteratively solving an optimization model. Metal artifact correction methods based on iterative reconstruction employ two approaches: one uses projection data with missing metal trajectories to iteratively reconstruct the CT image, and the other adds constraint regularization terms during the iterative optimization process to remove the influence of metal artifacts during reconstruction. This method reconstructs high-quality images, but requires multiple preprojection and reconstruction operations, resulting in significant computational load and time consumption, making its practical application and promotion very difficult. The third type, deep learning, due to its powerful feature learning capabilities, is increasingly being applied to MAR tasks.

[0004] Currently, deep learning-based MAR algorithms primarily utilize supervised learning, aiming to train transformations between images with artifacts or projections to images without artifacts or projections. Supervised networks are trained using simulated paired data, offering stable results but requiring large labeled datasets, and performance degrades in real-world applications. Unsupervised learning uses unpaired clinical data to train networks adversarially, avoiding domain differences caused by simulation, but suffers from training instability, and test results are somewhat uncertain. Semi-supervised MAR methods typically train the network model on simulated paired data and fine-tune it using unpaired clinical data. Semi-supervised methods attempt to find a balance between simulation and clinical data, combining the advantages of both supervised and unsupervised training methods, but often yield unsatisfactory results in both domains. Furthermore, the introduction of unpaired data introduces uncertainty and potential risks to network learning.

[0005] Therefore, it is essential to provide a metal artifact correction network model and a metal artifact removal method to address the shortcomings of existing technologies. Summary of the Invention

[0006] The primary objective of this invention is to overcome the shortcomings of existing technologies by providing a metal artifact correction network model. This metal artifact correction network model produces images with better artifact correction performance.

[0007] The above-mentioned objectives of the present invention are achieved through the following technical measures:

[0008] Provides a metal artifact correction network model, with the following settings:

[0009] Learnable artifact encoders extract and encode artifact information from uncorrected CT images containing metal artifacts to obtain corresponding multi-scale artifact coding feature maps.

[0010] Learnable artifact remover – removes artifact information from uncorrected CT images containing metal artifacts;

[0011] Learnable artifact generator—generates artifact information on artifact-free CT images paired with the uncorrected CT images containing metal artifacts;

[0012] The network structure of the learnable artifact remover is the same as that of the learnable artifact generator.

[0013] Preferably, both the network structure of the learnable artifact remover and the learnable artifact generator are U-net structures with skip links.

[0014] Preferably, both the learnable artifact remover and the learnable artifact generator include eight second residual modules, four second downsampling blocks, four spatial self-attention modules, four first upsampling blocks for artifact encoding operations, and four second upsampling blocks.

[0015] Preferably, the learnable artifact encoder described above includes five first downsampling blocks, one depth convolutional layer, three first residual modules, and three convolutional modules.

[0016] Preferably, the aforementioned Spatial Self-Attention Module (SAM) uses the multi-scale artifact-encoded feature map obtained by the learnable artifact encoder to scale and modulate the feature map of skip connections.

[0017] Preferably, the downsampling operator in the first downsampling block is a convolution operation with a stride of 2.

[0018] The downsampling operator in the second downsampling block is a convolution operation with a stride of 2.

[0019] Preferably, the upsampling operation in the first upsampling block is implemented by interpolation using the Upsample function.

[0020] Preferably, the upsampling operator in the second upsampling block is a deconvolution operation.

[0021] Preferably, the normalization function of the learnable artifact encoder is SpectralNorm, and the activation function is LeakyRelu.

[0022] Preferably, the normalization function of both the learnable artifact remover and the learnable artifact generator is InstanceNorm, and the activation function is LeakyRelu.

[0023] Preferably, the first residual module includes a convolutional layer and an activation function.

[0024] Preferably, the second residual module includes a convolutional layer, a normalization function, and an activation function.

[0025] In the learnable artifact remover and the learnable artifact generator respectively:

[0026] The four first upsampling blocks are defined as first upsampling block 1, first upsampling block 2, first upsampling block 3 and first upsampling block 4 in the transmission direction from front to back, and the first upsampling block 1, first upsampling block 2, first upsampling block 3 and first upsampling block 4 are connected in sequence. The multi-scale artifact coding feature map of the learnable artifact encoder is input from the first upsampling block 1.

[0027] Four of the second residual modules are respectively concatenated with the four second upsampling blocks using the concatenation function Concat to obtain four first concatenation blocks. The four first concatenation blocks are defined as first concatenation block 1, first concatenation block 2, first concatenation block 3 and first concatenation block 4 in the transmission direction from front to back, and first concatenation block 1, first concatenation block 2, first concatenation block 3 and first concatenation block 4 are connected in sequence.

[0028] The other four second residual modules are respectively spliced ​​with the four second downsampling blocks to obtain four second splicing blocks; the four second splicing blocks are defined as second splicing block 1, second splicing block 2, second splicing block 3 and second splicing block 4 in the transmission direction from front to back; and second splicing block 1, second splicing block 2, second splicing block 3 and second splicing block 4 are connected in sequence, and the output of second splicing block 4 is also transmitted to first splicing block 1;

[0029] The four spatial self-attention modules are defined as spatial self-attention module 1, spatial self-attention module 2, spatial self-attention module 3 and spatial self-attention module 4, respectively.

[0030] The output of the first upsampling block 1 and the output of the second stitching block 4 are respectively input to the spatial self-attention module 1, and the output of the spatial self-attention module 1 is input to the first stitching block 1;

[0031] The output of the first upsampling block 2 and the output of the second stitching block 3 are respectively input to the spatial self-attention module 2, and the output of the spatial self-attention module 2 is input to the first stitching block 2;

[0032] The output of the first upsampling block 3 and the output of the second stitching block 2 are respectively input to the spatial self-attention module 3, and the output of the spatial self-attention module 3 is input to the first stitching block 3;

[0033] The output of the first upsampling block 4 and the output of the second stitching block 1 are respectively input to the spatial self-attention module 4, and the output of the spatial self-attention module 4 is input to the first stitching block 4.

[0034] When the uncorrected CT image containing metal artifacts is input from the residual module side of the second stitching block 1 in the learnable artifact remover, the output of the first stitching block 4 is added to the uncorrected CT image containing metal artifacts to output an artifact-free CT image.

[0035] In the learnable artifact generator, the artifact-free CT image is input from the residual module side of the second stitching block 1, and the output of the first stitching block 4 is added to the artifact-free CT image to output the artifact-added CT image.

[0036] In the learnable artifact encoder, an uncorrected CT image containing metal artifacts enters from a first downsampling block, then is input to the depth convolutional layer, then sequentially to four more first downsampling blocks, then sequentially to three first residual modules, then sequentially to three convolutional modules, and finally outputs from the last convolutional module to obtain a multi-scale artifact coding feature map.

[0037] A second objective of this invention is to overcome the shortcomings of existing technologies and provide a method for removing metal artifacts. This method yields a metal artifact-free image with better artifact correction.

[0038] The above-mentioned objectives of the present invention are achieved through the following technical measures:

[0039] A method for removing metal artifacts is provided, which uses the aforementioned metal artifact correction network model.

[0040] The method for removing metal artifacts according to the present invention comprises the following steps:

[0041] S1. Input the CT image of the metal artifact to be removed into the trained learnable artifact encoder to obtain the artifact coding feature map. ;

[0042] S2. Input the CT image to be removed from the metal artifacts into the trained learnable artifact remover for decoding. During the decoding process, the trained learnable artifact remover progressively embeds the CT image to be removed from the metal artifacts into the artifact encoding feature map obtained in S1. This yields the final image with removed metal artifacts.

[0043] The metal artifact removal method of the present invention performs joint training on the learnable artifact encoder, the learnable artifact remover, and the learnable artifact generator based on a pre-constructed training database to obtain the trained learnable artifact encoder, the trained learnable artifact remover, and the trained learnable artifact generator.

[0044] Preferably, the training database is configured with multiple pairs of paired data, and each pair of paired data consists of images containing metal artifacts. and images containing metal artifacts Paired artifact-free images composition.

[0045] Preferably, the above joint training includes the following steps:

[0046] A1. Images containing metal artifacts Input the learnable artifact encoder In the process, artifact coding feature maps are obtained. ;

[0047] A2. Using the same image containing metal artifacts Input the learnable artifact remover Decode the image, then use the learnable artifact remover. The artifact-coded feature map obtained from S1 is gradually embedded during the decoding process. Obtain the image with removed metal artifacts. ,and ;

[0048] A3. The artifact coding feature map obtained in A1 Set it to 0, and then the image containing metal artifacts Input the learnable artifact remover Decoding is performed, and during the decoding process, artifact-coded feature maps that are initially set to 0 are gradually embedded. To obtain an image that retains metal artifacts ,and ;

[0049] A4. Images without metal artifacts from the same paired data. Input the learnable artifact generator Decoding is performed, and the artifact-coded feature map obtained from S1 is gradually embedded during the decoding process. To obtain an image containing metal artifacts ,and ;

[0050] A5. Obtain the artifact coding feature map from A1. Set to 0, then use the metal-free artifact images from the same paired data. Input the learnable artifact generator Decoding is performed, and during the decoding process, artifact-coded feature maps that are initially set to 0 are gradually embedded. To obtain an image that retains no metal artifacts ,and ;

[0051] A6. Construct a total loss function consisting of the artifact removal process loss function, the artifact generation process loss function, and the overall bidirectional embedding framework consistency loss function;

[0052] A7. Update the learnable artifact encoder based on the total loss function obtained in A6 and using stochastic gradient descent. The parameters, the learnable artifact remover The parameters, the learnable artifact generator Parameters;

[0053] A8. Determine if the number of training iterations is greater than or equal to the set number of training iterations. If it is, proceed to A9; otherwise, return to A1.

[0054] A9. Training ends.

[0055] Preferably, the loss function in the above artifact removal process is constrained using the L1 norm, and the artifact-free image obtained by constraining A2 using the L1 norm is... Images with no metal artifacts Obtain the image with removed metal artifacts. With artifact-free images Error between Furthermore, the loss function for artifact removal is expressed by equation (1);

[0056] ...Formula (1).

[0057] Preferably, the loss function in the above artifact generation process is constrained by a perceptual loss function, an adversarial loss function, and a feature loss function. The metal artifact image A4 obtained by constraining the perceptual loss function, adversarial loss function, and feature loss function is then constrained. Images with metal artifacts At the feature level, images containing metal artifacts are obtained. Images with metal artifacts The error between them, and this error is caused by the perceived loss error. Combating loss error and feature loss error composition.

[0058] Preferably, the above-mentioned perceived loss error Calculated by equation (2);

[0059] ...Equation (2);

[0060] The feature is the output of the next activation layer of the pooling layer in a VGG-19 network pre-trained on ImageNet using ImageNet data.

[0061] Preferably, the above-mentioned counter-loss error Calculated by equation (3);

[0062] ...Equation (3);

[0063] To indicate that the data comes from an image containing metal artifacts , To represent data without artifacts in an image , To indicate that the data comes from a learnable artifact generator Generated image containing metal artifacts , Here, E is the scale discriminator and E is the expected value.

[0064] Preferably, the above-mentioned feature loss error Calculated by equation (4);

[0065] ...Equation (4);

[0066] T is the scale discriminant. Total number of floors, The number of elements in each layer. For scale discriminator Extract the first Features of the sub-sampling block;

[0067] Preferably, the overall bidirectional embedding frame consistency loss function described above uses L1 norm constraints, and the image containing metal artifacts obtained by using L1 norm constraints A4... The image obtained with A3 retains the metal artifacts. To obtain an image containing metal artifacts With the preservation of images containing metal artifacts The error between them; and the metal artifact image obtained by constraining A4. The image obtained with A3 retains the metal artifact-free image. To obtain an image containing metal artifacts With maintaining images free of metal artifacts The error between them, and the overall bidirectional embedding framework consistency loss function is represented by equation (5):

[0068] ...Equation (5);

[0069] in, and All are hyperparameters;

[0070] Preferably, the total loss function is represented by equation (6);

[0071] ...Equation (6);

[0072] in, , , , and This is the default value.

[0073] This invention discloses a metal artifact correction network model and a method for removing metal artifacts. The metal artifact correction network model includes a learnable artifact encoder that extracts and encodes artifact information from uncorrected CT images containing metal artifacts, obtaining corresponding multi-scale artifact encoding feature maps; a learnable artifact remover that removes artifact information from the uncorrected CT images containing metal artifacts; and a learnable artifact generator that generates artifact information on artifact-free CT images paired with the uncorrected CT images containing metal artifacts. The network structures of the learnable artifact remover and the learnable artifact generator are identical. This invention constructs an artifact encoder for separating metal artifacts from anatomical structures in images containing metal artifacts. The artifact encoder can extract multi-scale metal artifact encoding feature maps. Then, the artifact removal and artifact generation processes are simultaneously modeled to enhance the effective training of the artifact encoder. Specifically, a learnable artifact remover and a learnable artifact generator with identical network structures were constructed. The learnable artifact remover removes artifact information from uncorrected CT images and incorporates a multi-channel spatial self-attention module. This module uses multi-scale artifact encoding feature maps to scale and modulate the feature maps of skip connections to adapt to changes in artifact information at different scales and to assist in embedding artifact encoding information during decoding. The learnable artifact generator generates artifact information on paired artifact-free CT images to aid in the training of the artifact encoder. The artifact generator takes an image without metal artifacts as input, while the artifact remover takes an image containing metal artifacts as input. Both encoders encode their respective inputs, ultimately achieving the goals of adding and subtracting artifacts. This invention can obtain metal artifact-free images with good artifact correction effects. Attached Figure Description

[0074] The invention will be further described with reference to the accompanying drawings, but the contents of the drawings do not constitute any limitation on the invention.

[0075] Figure 1 shows a framework diagram of a metal artifact correction network model.

[0076] Figure 2 is a framework diagram of a learnable artifact encoder.

[0077] Figure 3 is a structural diagram of the first downsampling block.

[0078] Figure 4 is a structural diagram of the first residual module.

[0079] Figure 5 shows the structure of the convolution module.

[0080] Figure 6 is a structural diagram of a learnable artifact generator.

[0081] Figure 7 is a structural diagram of a learnable artifact remover.

[0082] Figure 8 is a structural diagram of the first upsampling block.

[0083] Figure 9 is a structural diagram of the second residual module.

[0084] Figure 10 is a structural diagram of the second downsampling block.

[0085] Figure 11 is a structural diagram of the second upsampling block.

[0086] Figure 12 is a structural diagram of the head module.

[0087] Figure 13 is a structural diagram of the tail module.

[0088] Figure 14 is a flowchart of the joint training process.

[0089] Figure 15 shows the CT image of the object to be removed from the metal artifact and the final image after the metal artifact was removed. Detailed Implementation

[0090] The technical solution of the present invention will be further described in conjunction with the following embodiments.

[0091] Example 1

[0092] A metal artifact correction network model, as shown in Figure 1, is configured with:

[0093] Learnable artifact encoders extract and encode artifact information from uncorrected CT images containing metal artifacts to obtain corresponding multi-scale artifact coding feature maps.

[0094] Learnable artifact remover – removes artifact information from uncorrected CT images containing metal artifacts;

[0095] Learnable artifact generator – generates artifact information on artifact-free CT images paired with uncorrected CT images containing metal artifacts.

[0096] The network structure of the learnable artifact remover is the same as that of the learnable artifact generator. Both the learnable artifact remover and the learnable artifact generator use a U-net structure with skip links.

[0097] The learnable artifact encoder consists of five first downsampling blocks, one depthwise convolutional layer, three first residual modules, and three convolutional modules. The upsampling operation in the first upsampling block is implemented using the Upsample function for interpolation. The normalization function of the learnable artifact encoder is SpectralNorm, which is beneficial for the learnable artifact encoder. It exhibits good insensitivity to input perturbations, thus contributing to training stability. The activation function is LeakyReLU. The first residual module includes convolutional layers and activation functions.

[0098] As shown in Figure 2, in the learnable artifact encoder, an uncorrected CT image containing metal artifacts enters from a first downsampling block (Figure 3), then is input to a deep convolutional layer, then sequentially to four other first downsampling blocks, then sequentially to three first residual modules (Figure 4), then sequentially to three convolutional modules (Figure 5), and finally outputs from the last convolutional module to obtain a multi-scale artifact coding feature map.

[0099] Both the learnable artifact remover and the learnable artifact generator include eight second residual modules, four second downsampling blocks, four spatial self-attention modules, four first upsampling blocks for artifact encoding operations, and four second upsampling blocks. The spatial self-attention module (SAM) uses the multi-scale artifact-encoded feature map obtained from the learnable artifact encoder to scale and modulate the feature map of skip connections. The downsampling operator in the first downsampling block is a convolution operation with a stride of 2. The downsampling operator in the second downsampling block is also a convolution operation with a stride of 2. The upsampling operator in the second upsampling block is a deconvolution operation. The normalization function for both the learnable artifact remover and the learnable artifact generator is InstanceNorm, i.e., instance normalization, to facilitate the conversion from an image style with artifacts to an image style without artifacts. The normalization method corresponds to the normalization method of the artifact encoder E, and the spectral normalization function SpectralNorm is selected. Then, the spatial self-attention module (SAM) is used to scale and modulate the feature map of skip connections in the U-net structure. and Modulation is performed, and the modulated features are: Among them, subscript Indicates the first Each scale. Superscript n=1 indicates that the artifact comes from the artifact generation process, and n=2 indicates that the artifact comes from the artifact removal process. and Yes The hyperparameters are predicted after 3x3 convolution. Spatial Self-Attention (SAM) module modulation is used to automatically learn and scale the embedded features to better match feature variations at different feature scales. Modulated features. The features are concatenated according to the conventional U-net architecture and the features decoded after upsampling by G.UpBlock.

[0100] The activation function is LeakyReLU. The second residual module includes convolutional layers, a normalization function, and an activation function.

[0101] In the learnable artifact remover and learnable artifact generator respectively:

[0102] The four first upsampling blocks (as shown in Figure 8) are defined sequentially from front to back according to the transmission direction as first upsampling block 1, first upsampling block 2, first upsampling block 3, and first upsampling block 4. These blocks are then connected sequentially to learn the multi-scale artifact coding feature map of the artifact encoder. Input from the first upsampling block 1.

[0103] Four of the second residual modules (as shown in Figure 9) are concatenated with four second upsampling blocks (as shown in Figure 11) respectively using the concatenation function Concat to obtain four first concatenation blocks. The four first concatenation blocks are defined as first concatenation block 1, first concatenation block 2, first concatenation block 3 and first concatenation block 4 in the transmission direction from front to back, and first concatenation block 1, first concatenation block 2, first concatenation block 3 and first concatenation block 4 are connected in sequence.

[0104] The other four second residual modules are respectively spliced ​​with the four second downsampling blocks (as shown in Figure 10) to obtain four second splicing blocks; the four second splicing blocks are defined as second splicing block 1, second splicing block 2, second splicing block 3 and second splicing block 4 in the transmission direction from front to back; and second splicing block 1, second splicing block 2, second splicing block 3 and second splicing block 4 are connected in sequence, and the output of second splicing block 4 is also transmitted to first splicing block 1.

[0105] The four spatial self-attention modules are defined as spatial self-attention module 1, spatial self-attention module 2, spatial self-attention module 3, and spatial self-attention module 4, respectively.

[0106] The output of the first upsampling block 1 and the output of the second stitching block 4 are respectively input to the spatial self-attention module 1, and the output of the spatial self-attention module 1 is input to the first stitching block 1.

[0107] The output of the first upsampling block 2 and the output of the second splicing block 3 are respectively input to the spatial self-attention module 2, and the output of the spatial self-attention module 2 is input to the first splicing block 2.

[0108] The output of the first upsampling block 3 and the output of the second splicing block 2 are respectively input to the spatial self-attention module 3, and the output of the spatial self-attention module 3 is input to the first splicing block 3.

[0109] The output of the first upsampling block 4 and the output of the second stitching block 1 are respectively input to the spatial self-attention module 4, and the output of the spatial self-attention module 4 is input to the first stitching block 4.

[0110] The present invention also connects a head module (as shown in Figure 12) to one side of the second residual module of the second splicing block 1, and a tail module (as shown in Figure 13) to one side of the second residual module of the first splicing block 4.

[0111] When using the learnable artifact remover, uncorrected CT images containing metal artifacts The output of the first stitching block 4 is added to the uncorrected CT image containing metal artifacts, and the result is an artifact-free CT image. The input is from the residual module side of the second stitching block 1. As shown in Figure 7.

[0112] Artifact-free CT images in a learnable artifact generator The input is from the residual module side of the second stitching block 1. The output of the first stitching block 4 is added to the artifact-free CT image to output the artifact-added CT image. As shown in Figure 6.

[0113] It should be noted that in Figures 2-5 and 8-13, K represents the kernel size, S represents the stride, P represents the padding size, and SF represents the upsampling scaling factor.

[0114] This metal artifact correction network model constructs an artifact encoder to separate metal artifacts from anatomical structures in images containing metal artifacts. The artifact encoder can extract multi-scale metal artifact encoding feature maps. Then, the artifact removal and artifact generation processes are modeled simultaneously to enhance the effective training of the artifact encoder. Specifically, a learnable artifact remover and a learnable artifact generator with identical network structures are constructed. The learnable artifact remover removes artifact information from uncorrected CT images and incorporates a multi-channel spatial self-attention module. This module uses multi-scale artifact encoding feature maps to scale and modulate the feature maps of skip connections to adapt to changes in artifact information at different scales and to assist in embedding artifact encoding information during decoding. The learnable artifact generator generates artifact information on paired artifact-free CT images to aid in the training of the artifact encoder. The artifact generator takes an image without metal artifacts as input, while the artifact remover takes an image containing metal artifacts as input. Both encoders encode their respective inputs, ultimately achieving the goals of adding and subtracting artifacts. This invention can obtain metal artifact-free images with good artifact correction effects.

[0115] Example 2

[0116] A method for removing metal artifacts is provided, which uses the metal artifact correction network model of Example 1.

[0117] In the metal artifact removal method of the present invention, a learnable artifact encoder, a learnable artifact remover, and a learnable artifact generator are jointly trained according to a pre-constructed training database to obtain a trained learnable artifact encoder, a trained learnable artifact remover, and a trained learnable artifact generator.

[0118] The training database is configured with multiple pairs of paired data, and each pair consists of images containing metal artifacts. and images containing metal artifacts Paired artifact-free images Composition. Each pair of paired data refers to the metal artifact-free image corresponding to the same slice of the same object. and images containing metal artifacts .

[0119] The joint training includes the following steps, as shown in Figure 14:

[0120] A1. Images containing metal artifacts Input learnable artifact encoder In the process, artifact coding feature maps are obtained. ;

[0121] A2. Using the same image containing metal artifacts Input-learnable artifact remover Decode the image, and then you can learn the artifact remover. The artifact-coded feature map obtained from S1 is gradually embedded during the decoding process. Obtain the image with removed metal artifacts. ,and ;

[0122] A3. Obtain the artifact coding feature map from A1. Set it to 0, then the image containing metal artifacts will be displayed. Input-learnable artifact remover Decoding is performed, and during the decoding process, artifact-coded feature maps that are initially set to 0 are gradually embedded. To obtain an image that retains metal artifacts ,and ;

[0123] A4. Images without metal artifacts from the same paired data. Input-learnable artifact generator Decoding is performed, and the artifact-coded feature map obtained from S1 is gradually embedded during the decoding process. To obtain an image containing metal artifacts ,and ;

[0124] A5. Obtain the artifact coding feature map from A1. Set to 0, then use the metal-free artifact images from the same paired data. Input-learnable artifact generator Decoding is performed, and during the decoding process, artifact-coded feature maps that are initially set to 0 are gradually embedded. To obtain an image that retains no metal artifacts ,and ;

[0125] A6. Construct a total loss function consisting of the artifact removal process loss function, the artifact generation process loss function, and the overall bidirectional embedding framework consistency loss function;

[0126] A7. Update the learnable artifact encoder based on the total loss function obtained in A6 and using stochastic gradient descent. Parameters, Learnable Artifact Removal Parameters, learnable artifact generator Parameters;

[0127] A8. Determine if the number of training iterations is greater than or equal to the set number of training iterations. If it is, proceed to A9; otherwise, return to A1.

[0128] A9. Training ends.

[0129] The artifact removal process of this invention uses the L1 norm to constrain the loss function, and the artifact-free image obtained by constraining A2 using the L1 norm is... Images with no metal artifacts Obtain the image with removed metal artifacts. With artifact-free images Error between Furthermore, the loss function for artifact removal is expressed by equation (1);

[0130] ...Formula (1).

[0131] The loss function for artifact generation is constrained using a perceptual loss function, an adversarial loss function, and a feature loss function. The metal artifact image A4 is then constrained using these three loss functions. Images with metal artifacts At the feature level, images containing metal artifacts are obtained. Images with metal artifacts The error between them, and this error is caused by the perceived loss error. Combating loss error and feature loss error composition;

[0132] Perceived loss error Calculated by equation (2);

[0133] ...Equation (2);

[0134] The feature is the output of the next activation layer of the pooling layer in a VGG-19 network pre-trained on ImageNet using ImageNet data.

[0135] It should be noted that the perceptual loss function of this invention is used to measure perceptual similarity.

[0136] Combating loss error Calculated by equation (3);

[0137] ...Equation (3);

[0138] To indicate that the data comes from an image containing metal artifacts , To represent data without artifacts in an image , To indicate that the data comes from a learnable artifact generator Generated image containing metal artifacts , Let E be the scale discriminator and E be the expected value.

[0139] It should be noted that the adversarial loss of this invention utilizes a multi-scale discriminator. To regularize The rationality of which Select the standard discriminator, LayerDiscriminator.

[0140] Feature loss error Calculated by equation (4);

[0141] ...Equation (4);

[0142] T is the scale discriminant. Total number of floors, The number of elements in each layer. For scale discriminator Extract the first Characteristics of the sampling block under the layer.

[0143] The overall bidirectional embedding frame consistency loss function uses L1 norm constraints, and the image with metal artifacts is obtained by using L1 norm constraints on A4. The image obtained with A3 retains the metal artifacts. To obtain an image containing metal artifacts With the preservation of images containing metal artifacts The error between them; and the metal artifact image obtained by constraining A4. The image obtained with A3 retains the metal artifact-free image. To obtain an image containing metal artifacts With maintaining images free of metal artifacts The error between them, and the overall bidirectional embedding framework consistency loss function is represented by equation (5):

[0144] ...Equation (5);

[0145] in, and All of these are hyperparameters.

[0146] It should be noted that the artifact consistency loss constraint of this invention preserves the image with artifacts. With the generated artifact image Consistent, rather than with Maintain consistency. Because The statement indicates that an image containing artifacts is generated without embedding artifact coding. Under the proposed fully symmetric bidirectional embedding framework, the corresponding target should be an image containing artifacts generated by embedding artifact coding. This is done to ensure consistency in the generated artifact images. Similarly, This refers to an image generated without artifact encoding, but its target should be an image generated with artifact encoding that is free of artifacts. .

[0147] The total loss function is expressed by equation (6);

[0148] ...Equation (6);

[0149] in, , , , and This is the default value.

[0150] It should be noted that the present invention , , , and It was determined based on experiments and experience. In this invention, it is set to... , , , , .

[0151] The method for removing metal artifacts consists of the following steps:

[0152] S1. Input the CT image of the metal artifact to be removed into the trained learnable artifact encoder to obtain the artifact coding feature map. ;

[0153] S2. Input the CT image to be removed from the metal artifacts into the trained learnable artifact remover for decoding. During the decoding process, the trained learnable artifact remover progressively embeds the CT image to be removed from the metal artifacts into the artifact encoding feature map obtained in S1. This yields the final image with removed metal artifacts.

[0154] This metal artifact removal method is based on the aforementioned metal artifact correction network model. The invention first trains the metal artifact correction network model, and because this model has a completely symmetrical architecture, in practical applications, only the trained learnable artifact encoder and the trained learnable artifact remover are needed to identify, extract, and remove artifacts from the input CT image to be removed from. This method demonstrates good accuracy and practicality, ultimately achieving the goals of adding and subtracting artifacts. Furthermore, this invention yields metal artifact-free images with good artifact correction results.

[0155] Example 3

[0156] A method for removing metal artifacts based on the metal artifact correction network model of Example 1, which has the same features as Example 2, except that: , , , and The specific values ​​are as follows: , , , , .

[0157] This embodiment uses 95 sets of patient data without metal artifacts obtained from dental CBCT scanners in dental clinics and 8802 sets of paired data obtained from clinical dental implant metal mask simulation for training. Then, it is trained in the PyTorch deep learning framework according to the joint training in Embodiment 2, using the Adam optimizer and a learning rate of 2×10. -5 linearly decreases to 1×10 -7 The network batch size is 1, and the training times are 50, resulting in a trained learnable artifact encoder and a trained learnable artifact remover.

[0158] The CT images containing metal artifacts obtained by scanning with a CBCT scanner in a dental clinic at a tube voltage of 85kVp and a tube current of 9mAs are used as CT images to be removed from the CT images. The metal artifacts are removed from the CT images to be removed from the CT images according to the metal artifact removal method steps in Example 2, resulting in Figure 15.

[0159] In Figure 15, the left side is the CT image to be removed from the metal artifact, and the right side is the final image after removing the metal artifact. As can be seen from Figure 15, the uncorrected CT image has many alternating bright and dark stripe artifacts. However, the corrected image obtained by the present invention effectively suppresses the radial stripe artifacts caused by metal, while effectively preserving the tooth structure. Therefore, the present invention can obtain a metal artifact removal image with better artifact correction effect.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A metal artifact correction network model, characterized in that, The settings are as follows: Learnable artifact encoders extract and encode artifact information from uncorrected CT images containing metal artifacts to obtain corresponding multi-scale artifact coding feature maps. Learnable artifact remover – removes artifact information from uncorrected CT images containing metal artifacts; Learnable artifact generator—generates artifact information on artifact-free CT images paired with the uncorrected CT images containing metal artifacts; The network structure of the learnable artifact remover is the same as that of the learnable artifact generator.

2. The metal artifact correction network model according to claim 1, characterized in that: The learnable artifact encoder includes five first downsampling blocks, one deep convolutional layer, three first residual modules, and three convolutional modules.

3. The metal artifact correction network model according to claim 2, characterized in that: Both the network structure of the learnable artifact remover and the learnable artifact generator are U-net structures with skip links. Both the learnable artifact remover and the learnable artifact generator include eight second residual modules, four second downsampling blocks, four spatial self-attention modules, four first upsampling blocks for artifact encoding operations, and four second upsampling blocks.

4. The metal artifact correction network model according to claim 3, characterized in that: The Spatial Self-Attention Module (SAM) uses the multi-scale artifact-encoded feature map obtained by the learnable artifact encoder to scale and modulate the feature map of skip connections. The downsampling operator in the first downsampling block is a convolution operation with a stride of 2; The downsampling operator in the second downsampling block is a convolution operation with a stride of 2; The upsampling operation in the first upsampling block is implemented by interpolation using the Upsample function; The upsampling operator in the second upsampling block is a deconvolution operation; The normalization function of the learnable artifact encoder is SpectralNorm, and the activation function is LeakyRelu; The normalization function for both the learnable artifact remover and the learnable artifact generator is InstanceNorm, and the activation function for both is LeakyRelu; The first residual module includes a convolutional layer and an activation function; The second residual module includes a convolutional layer, a normalization function, and an activation function.

5. The metal artifact correction network model according to claim 3, characterized in that, In the learnable artifact remover and the learnable artifact generator respectively: The four first upsampling blocks are defined as first upsampling block 1, first upsampling block 2, first upsampling block 3 and first upsampling block 4 in the transmission direction from front to back, and the first upsampling block 1, first upsampling block 2, first upsampling block 3 and first upsampling block 4 are connected in sequence. The multi-scale artifact coding feature map of the learnable artifact encoder is input from the first upsampling block 1. Four of the second residual modules are respectively concatenated with the four second upsampling blocks using the concatenation function Concat to obtain four first concatenation blocks. The four first concatenation blocks are defined as first concatenation block 1, first concatenation block 2, first concatenation block 3 and first concatenation block 4 in the transmission direction from front to back, and first concatenation block 1, first concatenation block 2, first concatenation block 3 and first concatenation block 4 are connected in sequence. The other four second residual modules are respectively spliced ​​with the four second downsampling blocks to obtain four second splicing blocks; the four second splicing blocks are defined as second splicing block 1, second splicing block 2, second splicing block 3 and second splicing block 4 in the transmission direction from front to back; and second splicing block 1, second splicing block 2, second splicing block 3 and second splicing block 4 are connected in sequence, and the output of second splicing block 4 is also transmitted to first splicing block 1; The four spatial self-attention modules are defined as spatial self-attention module 1, spatial self-attention module 2, spatial self-attention module 3 and spatial self-attention module 4, respectively. The output of the first upsampling block 1 and the output of the second stitching block 4 are respectively input to the spatial self-attention module 1, and the output of the spatial self-attention module 1 is input to the first stitching block 1; The output of the first upsampling block 2 and the output of the second stitching block 3 are respectively input to the spatial self-attention module 2, and the output of the spatial self-attention module 2 is input to the first stitching block 2; The output of the first upsampling block 3 and the output of the second stitching block 2 are respectively input to the spatial self-attention module 3, and the output of the spatial self-attention module 3 is input to the first stitching block 3; The output of the first upsampling block 4 and the output of the second stitching block 1 are respectively input to the spatial self-attention module 4, and the output of the spatial self-attention module 4 is input to the first stitching block 4; When the uncorrected CT image containing metal artifacts is input from the residual module side of the second stitching block 1 in the learnable artifact remover, the output of the first stitching block 4 is added to the uncorrected CT image containing metal artifacts to output an artifact-free CT image. In the learnable artifact generator, the artifact-free CT image is input from the residual module side of the second stitching block 1, and the output of the first stitching block 4 is added to the artifact-free CT image to output the artifact-added CT image. In the learnable artifact encoder, an uncorrected CT image containing metal artifacts enters from a first downsampling block, then is input to the depth convolutional layer, then sequentially to four more first downsampling blocks, then sequentially to three first residual modules, then sequentially to three convolutional modules, and finally outputs from the last convolutional module to obtain a multi-scale artifact coding feature map.

6. A method for removing metal artifacts, characterized in that: The metal artifact correction network model described in any one of claims 1 to 5 is used.

7. The method for removing metal artifacts according to claim 6, characterized in that, The process is as follows: S1. Input the CT image of the metal artifact to be removed into the trained learnable artifact encoder to obtain the artifact coding feature map. ； S2. Input the CT image to be removed from the metal artifacts into the trained learnable artifact remover for decoding. During the decoding process, the trained learnable artifact remover progressively embeds the CT image to be removed from the metal artifacts into the artifact encoding feature map obtained in S1. This yields the final image with removed metal artifacts.

8. The method for removing metal artifacts according to claim 7, characterized in that: The learnable artifact encoder, the learnable artifact remover, and the learnable artifact generator are jointly trained based on a pre-built training database to obtain the trained learnable artifact encoder, the trained learnable artifact remover, and the trained learnable artifact generator. The training database contains multiple pairs of paired data, and each pair of paired data consists of images containing metal artifacts. and images containing metal artifacts Paired artifact-free images composition.

9. The method for removing metal artifacts according to claim 8, characterized in that: The joint training includes the following steps: A1. Images containing metal artifacts Input the learnable artifact encoder In the process, artifact coding feature maps are obtained. ； A2. Using the same image containing metal artifacts Input the learnable artifact remover Decode the image, then use the learnable artifact remover. The artifact-coded feature map obtained from S1 is gradually embedded during the decoding process. Obtain the image with removed metal artifacts. ,and ； A3. The artifact coding feature map obtained in A1 Set it to 0, and then the image containing metal artifacts Input the learnable artifact remover Decoding is performed, and during the decoding process, artifact-coded feature maps that are initially set to 0 are gradually embedded. To obtain an image that retains metal artifacts ,and ； A4. Images without metal artifacts from the same paired data. Input the learnable artifact generator Decoding is performed, and the artifact-coded feature map obtained from S1 is gradually embedded during the decoding process. To obtain an image containing metal artifacts ,and ； A5. Obtain the artifact coding feature map from A1. Set to 0, then use the metal-free artifact images from the same paired data. Input the learnable artifact generator Decoding is performed, and during the decoding process, artifact-coded feature maps that are initially set to 0 are gradually embedded. To obtain an image that retains no metal artifacts ,and ； A6. Construct a total loss function consisting of the artifact removal process loss function, the artifact generation process loss function, and the overall bidirectional embedding framework consistency loss function; A7. Update the learnable artifact encoder based on the total loss function obtained in A6 and using stochastic gradient descent. The parameters, the learnable artifact remover The parameters, the learnable artifact generator Parameters; A8. Determine if the number of training iterations is greater than or equal to the set number of training iterations. If it is, proceed to A9; otherwise, return to A1. A9. Training ends.

10. The method for removing metal artifacts according to claim 9, characterized in that: The loss function for the artifact removal process is constrained using the L1 norm, and the artifact-free image obtained by constraining A2 using the L1 norm is... Images with no metal artifacts Obtain the image with removed metal artifacts. With artifact-free images Error between Furthermore, the loss function for artifact removal is expressed by equation (1); ...Equation (1); The loss function for the artifact generation process is constrained by a perceptual loss function, an adversarial loss function, and a feature loss function. The metal artifact image A4 is then constrained using these three loss functions. Images with metal artifacts At the feature level, images containing metal artifacts are obtained. Images with metal artifacts The error between them, and this error is caused by the perceived loss error. Combating loss error and feature loss error composition; The perceived loss error Calculated by equation (2); ...Equation (2); The feature is the output of the next activation layer of the pooling layer in a VGG-19 network pre-trained on ImageNet, which is a network pre-trained using ImageNet data; The anti-loss error Calculated by equation (3); ...Equation (3); To indicate that the data comes from an image containing metal artifacts ， To represent data without artifacts in an image ， To indicate that the data comes from a learnable artifact generator Generated image containing metal artifacts ， Here, E is the scale discriminator and E is the expected value. The feature loss error Calculated by equation (4); ...Equation (4); T is the scale discriminant. Total number of floors, The number of elements in each layer. For scale discriminator Extract the first Features of the sub-sampling block; The overall bidirectional embedding framework consistency loss function uses L1 norm constraints, and the image containing metal artifacts is obtained by using L1 norm constraints on A4. The image obtained with A3 retains the metal artifacts. To obtain an image containing metal artifacts With the preservation of images containing metal artifacts The error between them; and the metal artifact image obtained by constraining A4. The image obtained with A3 retains the metal artifact-free image. To obtain an image containing metal artifacts With maintaining images free of metal artifacts The error between them, and the overall bidirectional embedding framework consistency loss function is represented by equation (5): ...Equation (5); in, and All are hyperparameters; The total loss function is represented by equation (6); ...Equation (6); in, 、 、 、 and This is the default value.

Images