A Deep Learning-Based Method and System for Artifact Suppression in Cardiac Magnetic Resonance Cine Imaging

Abstract

This invention provides a method and system for suppressing artifacts in cardiac magnetic resonance imaging (MRI) cine imaging based on deep learning. The method includes: acquiring cardiac MRI cine images with black band artifacts or blood flow artifacts and preprocessing them; performing joint image computation to obtain artifact-free images; constructing a deep learning model, where the input to the artifact suppression network is the cardiac cine image with black band artifacts or blood flow artifacts, and the output is the cardiac cine image with artifacts suppressed; training the deep learning model using the preprocessed images with and without artifacts until the parameters of the artifact suppression network reach the target conditions, obtaining and saving the trained deep learning model; and performing artifact removal on the cardiac cine image with the black band artifacts or blood flow artifacts to be removed, obtaining the artifact-removed cardiac cine image. This invention overcomes the limitation of traditional phase cycling methods that easily introduce blood flow artifacts, while simultaneously suppressing both black band artifacts and blood flow artifacts.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, to a method and system for suppressing artifacts in cardiac magnetic resonance cinema imaging based on deep learning. Background Technology

[0002] Cardiac magnetic resonance imaging (MRI) uses the principles of magnetic resonance to image the anatomical structure and cardiac function of the heart. It is a safe and radiation-free technique. It can provide images with various contrast levels and has become the gold standard for the clinical diagnosis of many heart diseases.

[0003] Cardiac cine imaging records the contraction and relaxation of the heart, providing dynamic imaging throughout the entire cardiac cycle, and has been widely used for cardiac structure and function analysis. Currently, the Balanced Steady State Free Precession (bSSFP) sequence is one of the most commonly used sequences in clinical cardiac cine imaging. It offers fast imaging speed, high signal-to-noise ratio, and good contrast, thus its widespread application in cardiac cine imaging.

[0004] However, bSSFP sequences are quite sensitive to resonance frequencies. When a specific location is subjected to magnetic field inhomogeneity, causing a shift in the resonance frequency and resulting in low signal intensity, this further creates black band artifacts in the image. Since magnetic field inhomogeneity is difficult to completely avoid, black band artifacts appear in most cardiac cine imaging, posing certain difficulties for the analysis of cardiac structure and function.

[0005] Patent document CN111833410A (application number: CN202010646786.X) discloses a deep learning-based X-ray scattering suppression method, including the following steps: S1: Downsampling chest CT data; S2: Removing metal artifacts from the downsampled chest CT data using a region growing algorithm; S3: Dividing the data after removing metal artifacts into pixels according to a threshold, and relabeling the pixels of the divided chest CT data; S4: Repeating the above processing method to preprocess a large amount of chest CT data; S5: Using the preprocessed data as a training set and a test set for deep learning training to obtain the optimal network model. This invention is used for the correction of black band artifacts and blood flow artifacts in cardiac magnetic resonance imaging (MRI), and its application scenario differs from the aforementioned patent.

[0006] Currently, the traditional method for suppressing black band artifacts is through phase-cycle imaging. Phase-cycle imaging involves cyclically changing the phase increment in the bSSFP sequence, causing the position of the black band artifacts on the image to shift, thereby suppressing the artifacts by combining multiple images. Although the phase-cycle imaging method can suppress black band artifacts, it significantly increases scanning time and reduces scanning efficiency because it requires the simultaneous acquisition of multiple images.

[0007] In most cases, acquiring multiple cardiac cine images requires holding one's breath to avoid mismatches between images caused by respiratory motion, which can lead to problems in image co-processing. However, the need for multiple images significantly increases the time patients need to hold their breath, reducing scanning comfort and increasing the risk of introducing motion artifacts. Furthermore, phase-cycled methods require changing the phase increment; therefore, at least at one phase increment, the acquired image will have black band artifacts covering the blood flow area, resulting in severe blood flow artifacts. Fusing multiple images further introduces blood flow artifacts into the final image. Therefore, the effectiveness of phase-cycled methods in removing black band artifacts has limitations.

[0008] In summary, although the phase-cycled method can suppress black band artifacts, its drawbacks include increased scan time, reduced scan comfort, and the introduction of blood flow artifacts. Therefore, we need to further explore and develop new methods to address these issues. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for suppressing artifacts in cardiac magnetic resonance cinema imaging based on deep learning.

[0010] The method for suppressing artifacts in cardiac magnetic resonance cine imaging based on deep learning provided by the present invention includes:

[0011] Step S1: Acquire cardiac magnetic resonance cine images with black band artifacts or blood flow artifacts, perform preprocessing operations on the images, and perform joint image calculations to obtain artifact-free images;

[0012] Step S2: Construct a deep learning model. The input of the artifact suppression network in the model is a heart movie image with black band artifacts or blood flow artifacts, and the output is a heart movie image after black band artifacts or blood flow artifacts are suppressed.

[0013] Step S3: Use the preprocessed images with and without artifacts to train the deep learning model until the parameters of the artifact suppression network reach the target conditions, obtain the trained deep learning model and save it.

[0014] Step S4: Based on the trained deep learning model, perform artifact removal on the cardiac cinema image with black band artifacts or blood flow artifacts to be removed, and obtain the cardiac cinema image with artifact removal.

[0015] Preferably, the preprocessing of the acquired cardiac magnetic resonance cine images includes:

[0016] First, adjust the k-space size to fix the physical resolution of the image based on the image resolution at the time of image acquisition;

[0017] Secondly, crop the image size to a uniform size;

[0018] Then, for images with different bias frequencies, the remaining images are registered using 0Hz as the reference image;

[0019] Finally, the image intensity is normalized to the range [0,1].

[0020] Preferably, the image joint computation process includes: obtaining training labels for the artifact suppression network: selecting multiple sets of cardiac film images with a bias frequency near 0Hz, and obtaining artifact-free images by averaging.

[0021] Preferably, the artifact suppression network construction process includes: the artifact suppression network architecture comprises a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, ..., the (N-1)th convolutional layer, an upsampling layer, the Nth convolutional layer, and an output layer; wherein the i-th stage convolutional layer corresponds to the Ni-th convolutional layer, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer after the Ni-th convolutional layer as the input of the Nth convolutional layer, where N is a positive integer greater than 2; the input of the artifact suppression network contains the original cardiac film image containing black band artifacts or blood flow artifacts; the output is the image after the black band artifacts and blood flow artifacts are suppressed;

[0022] The loss function for constructing the artifact suppression network is as follows: Given a 3D cardiac magnetic resonance cinematic image x, the artifact suppression network processes the input x to obtain the artifact-suppressed result F. AS (x), y AS It is an AS tag;

[0023] L AS (x)=L recon (F AS (x),y AS )+λ AS L perceptual (F AS (x),y AS )

[0024] Among them, L recon L represents image restoration loss; perceptual λ represents the perceptual loss; λ is the weight that balances the image restoration loss and the perceptual loss.

[0025] Preferably, before the artifact suppression network, an artifact recognition network is added to output an artifact identity map. The input of the artifact recognition network is a cardiac cinema image with black band artifacts or blood flow artifacts, and the output is the corresponding artifact identity map. The input of the artifact suppression network is an image with artifacts and its corresponding artifact identity map, and the output is a cardiac cinema image after the black band artifacts or blood flow artifacts have been suppressed. The specific execution steps include:

[0026] Obtain training labels for the artifact recognition network: Perform joint calculations on images without artifacts and images with artifacts, requiring the joint calculation results to show the location and type of artifacts, which will serve as the artifact identity map.

[0027] The artifact recognition network architecture consists of a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, ..., the (M-1)th convolutional layer, an upsampling layer, the Mth convolutional layer, and an output layer. The i-th stage convolutional layer corresponds to the Mi-th convolutional layer, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer following the Mi-th convolutional layer as the input to the M-th convolutional layer, where M is a positive integer greater than 2. The output layer of the artifact recognition network consists of a 1×1×1 three-dimensional convolutional layer and a sigmoid layer, restricting the artifact recognition output range to [0,1].

[0028] The artifact suppression network construction process includes: the artifact suppression network architecture consists of a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, ..., the (N-1)th convolutional layer, an upsampling layer, the Nth convolutional layer, and an output layer; where the i-th stage convolutional layer corresponds to the Ni-th convolutional layer, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer after the Ni-th convolutional layer as the input of the Nth convolutional layer, where N is a positive integer greater than 2; the artifact suppression network is a dual-input network, with the input containing the original heart film image and the artifact identity map output by the artifact recognition network; based on the artifact identity map, the identity information corresponding to each voxel is obtained, the positions of black band artifacts, blood flow artifacts, and non-artifacts are distinguished, and the output is the image after the black band artifacts and blood flow artifacts are suppressed.

[0029] Define the loss function L of the artifact recognition network. VI (x) and the loss function L of the artifact suppression network AS (x), specifically in the form of:

[0030] Input a 3D cardiac magnetic resonance cinematic image x, and use an artifact recognition network to obtain an artifact recognition map F. VI (x), y VI It is an artifact recognition label; the artifact suppression network pairs (x,F) VI (x)) The two inputs are processed to obtain the artifact-suppressed result F. AS (x,F VI (x)), y AS It is an artifact suppression label;

[0031] L VI (x)=L recon (F VI (x),y VI )+λ VI Lperceptual (F VI (x),y VI )

[0032] L AS (x)=L recon (F AS (x,F VI (x)),y AS )+λ AS L perceptual (F AS (x,F VI (x)),y AS )

[0033] Among them, L recon L represents image restoration loss; perceptual Represents perceived loss; λ VI , λ AS We weigh the image restoration loss and perceptual loss for the artifact recognition network and the artifact suppression network;

[0034] Model training process: The artifact recognition network is trained using preprocessed images with artifacts and artifact identity maps; the artifact suppression network is trained using preprocessed images with artifacts and images without artifacts; until the parameters of the artifact recognition network and the artifact suppression network reach the target conditions, the trained deep learning model is obtained.

[0035] Model processing procedure: The trained deep learning model is saved. The cardiac magnetic resonance imaging data with black band artifacts or blood flow artifacts to be removed are subjected to artifact removal. The image data after artifact removal is obtained by relying on the trained deep learning model and saved to the database. The artifact recognition network outputs artifact identity map and the artifact suppression network outputs artifact-free image, which is used to prepare for clinical cardiac function analysis.

[0036] The deep learning-based cardiac magnetic resonance cinema imaging artifact suppression system provided by the present invention includes:

[0037] Data acquisition module: Acquires cardiac magnetic resonance cine images with black band artifacts or blood flow artifacts;

[0038] Data processing module: performs joint image calculations on cardiac MRI cine images with black band artifacts or blood flow artifacts to obtain artifact-free images; performs preprocessing operations on the images with artifacts and the images without artifacts.

[0039] Model building module: Constructs a deep learning model, which is an artifact suppression network used to suppress black band artifacts or blood flow artifacts in cardiac film images;

[0040] Model training module: The artifact suppression network is trained using preprocessed images with and without artifacts until the parameters of the artifact suppression network reach the target conditions, thus obtaining the trained deep learning model.

[0041] Model processing module: Saves the trained deep learning model, performs artifact removal on cardiac magnetic resonance image data with black band artifacts or blood flow artifacts to be removed, obtains artifact-removed image data based on the trained deep learning model, and saves it to the database; the artifact suppression network outputs artifact-free images to prepare for clinical cardiac function analysis.

[0042] Preferably, the preprocessing of the acquired cardiac magnetic resonance cine images includes:

[0043] First, adjust the k-space size to fix the physical resolution of the image based on the image resolution at the time of image acquisition;

[0044] Secondly, crop the image size to a uniform size;

[0045] Then, for images with different bias frequencies, the remaining images are registered using 0Hz as the reference image;

[0046] Finally, the image intensity is normalized to the range [0,1].

[0047] Preferably, the image joint computation process includes: obtaining training labels for the artifact suppression network: selecting multiple sets of cardiac film images with a bias frequency near 0Hz, and obtaining artifact-free images by averaging.

[0048] Preferably, the artifact suppression network construction process includes: the artifact suppression network architecture comprises a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, ..., the (N-1)th convolutional layer, an upsampling layer, the Nth convolutional layer, and an output layer; wherein the i-th stage convolutional layer corresponds to the Ni-th convolutional layer, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer after the Ni-th convolutional layer as the input of the Nth convolutional layer, where N is a positive integer greater than 2; the input of the artifact suppression network contains the original cardiac film image containing black band artifacts or blood flow artifacts; the output is the image after the black band artifacts and blood flow artifacts are suppressed;

[0049] The loss function for constructing the artifact suppression network is as follows: Given a 3D cardiac magnetic resonance cinematic image x, the artifact suppression network processes the input x to obtain the artifact-suppressed result F. AS (x), y AS It is an AS tag;

[0050] L AS (x)=L recon (F AS(x),y AS )+λ AS L perceptual (F AS (x),y AS )

[0051] Among them, L recon L represents image restoration loss; perceptual λ represents the perceptual loss; λ is the weight that balances the image restoration loss and the perceptual loss.

[0052] Preferably, before the artifact suppression network, an artifact recognition network is added to output an artifact identity map. The input of the artifact recognition network is a cardiac cinema image with black band artifacts or blood flow artifacts, and the output is the corresponding artifact identity map. The input of the artifact suppression network is an image with artifacts and its corresponding artifact identity map, and the output is a cardiac cinema image after the black band artifacts or blood flow artifacts have been suppressed. The specific execution steps include:

[0053] Obtain training labels for the artifact recognition network: Perform joint calculations on images without artifacts and images with artifacts, requiring the joint calculation results to show the location and type of artifacts, which will serve as the artifact identity map.

[0054] The artifact recognition network architecture consists of a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, ..., the (M-1)th convolutional layer, an upsampling layer, the Mth convolutional layer, and an output layer. The i-th stage convolutional layer corresponds to the Mi-th convolutional layer, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer following the Mi-th convolutional layer as the input to the M-th convolutional layer, where M is a positive integer greater than 2. The output layer of the artifact recognition network consists of a 1×1×1 three-dimensional convolutional layer and a sigmoid layer, restricting the artifact recognition output range to [0,1].

[0055] The artifact suppression network construction process includes: the artifact suppression network architecture consists of a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, ..., the (N-1)th convolutional layer, an upsampling layer, the Nth convolutional layer, and an output layer; where the i-th stage convolutional layer corresponds to the Ni-th convolutional layer, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer after the Ni-th convolutional layer as the input of the Nth convolutional layer, where N is a positive integer greater than 2; the artifact suppression network is a dual-input network, with the input containing the original heart film image and the artifact identity map output by the artifact recognition network; based on the artifact identity map, the identity information corresponding to each voxel is obtained, the positions of black band artifacts, blood flow artifacts, and non-artifacts are distinguished, and the output is the image after the black band artifacts and blood flow artifacts are suppressed.

[0056] Define the loss function L of the artifact recognition network. VI(x) and the loss function L of the artifact suppression network AS (x), specifically in the form of:

[0057] Input a 3D cardiac magnetic resonance cinematic image x, and use an artifact recognition network to obtain an artifact recognition map F. VI (x), y VI It is an artifact recognition label; the artifact suppression network pairs (x,F) VI (x)) The two inputs are processed to obtain the artifact-suppressed result F. AS (x,F VI (x)), y AS It is an artifact suppression label;

[0058] L VI (x)=L recon (F VI (x),y VI )+λ VI L perceptual (F VI (x),y VI )

[0059] L AS (x)=L recon (F AS (x,F VI (x)),y AS )+λ AS L perceptual (F AS (x,F VI (x)),y AS )

[0060] Among them, L recon L represents image restoration loss; perceptual Represents perceived loss; λ VI , λ AS We weigh the image restoration loss and perceptual loss for the artifact recognition network and the artifact suppression network;

[0061] Model training process: The artifact recognition network is trained using preprocessed images with artifacts and artifact identity maps; the artifact suppression network is trained using preprocessed images with artifacts and images without artifacts; until the parameters of the artifact recognition network and the artifact suppression network reach the target conditions, the trained deep learning model is obtained.

[0062] Model processing procedure: The trained deep learning model is saved. The cardiac magnetic resonance imaging data with black band artifacts or blood flow artifacts to be removed are subjected to artifact removal. The image data after artifact removal is obtained by relying on the trained deep learning model and saved to the database. The artifact recognition network outputs artifact identity map and the artifact suppression network outputs artifact-free image, which is used to prepare for clinical cardiac function analysis.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] (1) In cardiac cine imaging using the bSSFP sequence, it is not necessary to acquire multiple phase loop images, but only to act on a single cardiac cine imaging, which shortens the breath-holding time, improves scanning comfort, and reduces the risk of introducing motion artifacts.

[0065] (2) Combined removal of black band artifacts and blood flow artifacts overcomes the limitation of traditional phase circulation schemes that easily introduce blood flow artifacts and improves the quality of magnetic resonance cardiac cine images.

[0066] (3) The deep learning model proposed in this invention can optionally add an artifact recognition network, so that the overall architecture contains two sub-networks. While suppressing artifacts, it provides an artifact identity map that represents the location and type of artifacts. Given a magnetic resonance cardiac film image, the artifact identity map can help understand artifact information and assist in understanding the object and process of artifact suppression. Compared with the black-box output of results by deep learning models, this scheme has better clinical applicability. Attached Figure Description

[0067] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0068] Figure 1 This is a flowchart of a method for joint removal of black band artifacts and blood flow artifacts in cardiac magnetic resonance cinema imaging based on deep learning, according to an embodiment of the present invention.

[0069] Figure 2 This is a flowchart illustrating the workflow of an artifact recognition network and an artifact suppression network according to an embodiment of the present invention. Detailed Implementation

[0070] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0071] Example 1:

[0072] This invention discloses a method for joint removal of black band artifacts and blood flow artifacts in cardiac magnetic resonance cinematic imaging based on deep learning. Specifically, the method includes the following steps:

[0073] S1, Image Acquisition: Cardiac cine imaging was acquired from multiple healthy volunteers. For each volunteer, 12 sets of phase-cycled cardiac cine images were acquired, with the resonant frequency adjusted during acquisition by changing the frequency offset of the bSSFP sequence. The offset frequencies corresponding to the 12 sets of phase-cycled images were [-137.5, -110, -82.5, -55, -27.5, 0, 27.5, 55, 82.5, 110, 137.5, 165] Hz, basically covering the frequency response period 1 / TR of the bSSFP sequence, with black band artifacts and corresponding blood flow artifacts in different locations. During scanning, the short-axis view of the heart was the primary focus, supplemented by scanning the long-axis view.

[0074] S2, Image Preprocessing: To ensure that the input images have the same spatial size and intensity range distribution, the same preprocessing operations are performed on each image. First, the k-space size is adjusted according to the resolution at the time of image acquisition to fix the physical resolution of the image at 1.5mm × 1.5mm; second, the image size is cropped to 192 × 192; then, for images with different bias frequencies, 0Hz is used as the reference image, and the remaining 11 groups of images are registered in a rigid registration manner to reduce image mismatch caused by respiratory movements, etc.; finally, the image intensity is normalized to the range of [0,1]. After preprocessing, the spatial dimension of each group of cardiac cine images is 192 × 192, and the temporal dimension is not processed.

[0075] S3, obtain training labels:

[0076] To avoid cardiac blood flow artifacts caused by excessively high bias frequencies, five groups of cardiac cine images at frequencies of [-55, -27.5, 0, 27.5, 55] Hz were selected, and a combined image was obtained by averaging. Here, the traditional phase-cycle combination method was not used to reduce the risk of introducing cardiac blood flow artifacts. Since the artifact locations are different in each group of images, combining them can suppress black band artifacts without introducing blood flow artifacts, and can be used as the AS label for the artifact suppression network.

[0077] S4, Image Segmentation: The preprocessed cardiac cine images are segmented proportionally to obtain training and validation data. Specifically, from the preprocessed MRI data of 28 subjects, the scan data of the first 15 volunteers are selected as the training set, and the scan data of the last 13 volunteers are selected as the test set. Of course, in other embodiments of the present invention, other proportions can be used for segmentation without affecting the purpose of the present invention.

[0078] S5, Building Deep Learning Models:

[0079] The artifact suppression network architecture consists of a series of cascaded layers: an input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, a downsampling layer, a third-stage convolutional layer, a downsampling layer, a fourth-stage convolutional layer, an upsampling layer, a fifth-stage convolutional layer, an upsampling layer, a sixth-stage convolutional layer, an upsampling layer, a seventh-stage convolutional layer, and an output layer. In this system, the i-th stage convolutional layer corresponds to the Mi-th convolutional layer (M=8, i=1, 2, 3). The concatenate operation combines the output of the i-th stage convolutional layer with the output of the upsampled layer preceding the Mi-th convolutional layer, which serves as the input to the Mi-th convolutional layer. Each stage convolutional layer sequentially contains two sets of 3×3×3 three-dimensional convolutions with a kernel size of 3, 32 Group Norm and ReLU. The purpose of using Group Norm is to improve the normalization performance when the batch size is small. The downsampling layer uses max pooling with a stride of 2×2×1 voxels, and the corresponding upsampling layer is a max unpooling operation. The input to the AS network is a cardiac MRI cine image with black band artifacts or blood flow artifacts, and the output is a cardiac cine image with the black band artifacts or blood flow artifacts suppressed.

[0080] S6, Construct the network loss function:

[0081] Given a 3D cardiac magnetic resonance cinematic image x, an artifact suppression network processes the input x to obtain the artifact-suppressed result F. AS (x), y AS It is an AS tag.

[0082] L AS (x)=L recon (F AS (x),y AS )+λ AS L perceptual (F AS (x),y AS )

[0083] Among them, L recon The image restoration loss is represented by the L2 norm difference between the network output and the corresponding label; L perceptual The perceptual loss is represented by the norm difference between the features output from the network output and the corresponding labels at each stage of the VGG19 network; λ is the weight that balances the image restoration loss and the perceptual loss. In this embodiment, λ AS =0.01.

[0084] S7, Training the Model: The original cardiac film image is input into the artifact suppression AS network, with AS labels used as the training target. The AS network is trained using the loss function defined above. Training stops when the validation set loss no longer decreases, and the network parameters of the AS network are saved. In this embodiment, the Adam optimizer is used, the learning rate is kept at 0.0001, and the AS network is iterated 300 times.

[0085] S8, Obtain the image after artifact suppression: Remove the black band artifact and blood flow artifact in cardiac cine imaging based on the trained artifact suppression network.

[0086] Specifically, the above-mentioned method of obtaining artifact-suppressed cardiac cine images based on the final artifact suppression model includes: First, acquiring a single cardiac cine image and performing the same preprocessing as in step S2 to obtain a cardiac cine image with a spatial dimension of 192×192, with no requirement for the temporal dimension; Second, inputting the preprocessed three-dimensional cardiac cine image into the AS network to obtain the artifact-suppressed result.

[0087] Given any cardiac cine image, this invention can obtain not only images with suppressed black band artifacts but also images with suppressed blood flow artifacts. Compared to traditional methods, this approach improves scanning efficiency and provides combined suppression of both black band and blood flow artifacts.

[0088] Example 2:

[0089] like Figure 1 This invention discloses a method for joint removal of black band artifacts and blood flow artifacts in cardiac magnetic resonance cinema imaging based on deep learning. Specifically, the method includes the following steps:

[0090] S1, Image Acquisition: Cardiac cine imaging was acquired from multiple healthy volunteers. For each volunteer, 12 sets of phase-cycled cardiac cine images were acquired, with the resonant frequency adjusted during acquisition by changing the frequency offset of the bSSFP sequence. The offset frequencies corresponding to the 12 sets of phase-cycled images were [-137.5, -110, -82.5, -55, -27.5, 0, 27.5, 55, 82.5, 110, 137.5, 165] Hz, basically covering the frequency response period 1 / TR of the bSSFP sequence, with black band artifacts and corresponding blood flow artifacts in different locations. During scanning, the short-axis view of the heart was the primary focus, supplemented by scanning the long-axis view.

[0091] S2, Image Preprocessing: To ensure that the input images have the same spatial size and intensity range distribution, the same preprocessing operations are performed on each image. First, the k-space size is adjusted according to the resolution at the time of image acquisition to fix the physical resolution of the image at 1.5mm × 1.5mm; second, the image size is cropped to 192 × 192; then, for images with different bias frequencies, 0Hz is used as the reference image, and the remaining 11 groups of images are registered in a rigid registration manner to reduce image mismatch caused by respiratory movements, etc.; finally, the image intensity is normalized to the range of [0,1]. After preprocessing, the spatial dimension of each group of cardiac cine images is 192 × 192, and the temporal dimension is not processed.

[0092] S3, obtaining training labels, can be divided into the following two steps:

[0093] S31: Obtaining AS tags for the artifact suppression phase. To avoid cardiac blood flow artifacts caused by excessively high bias frequencies, five groups of cardiac cine images at [-55, -27.5, 0, 27.5, 55] Hz were selected, and a combined image was obtained by averaging. Here, the traditional phase-cycle combination method was not used to reduce the risk of introducing cardiac blood flow artifacts. Since the artifact locations are different in each group of images, combining them together can suppress black band artifacts without introducing blood flow artifacts, and can be used as AS tags for the artifact suppression phase.

[0094] S32: Obtain the VI label for the artifact recognition stage. For black band artifacts, blood flow artifacts, and non-artifact regions, the VI label should have different intensities to distinguish these three types of regions. Therefore, divide the AS label by 12 sets of images with different bias frequencies, subtract 1, and then use the sigmoid function to limit the numerical range to [0,1] to obtain the artifact identity map. In these images, for voxels in the black band artifact, the corresponding intensity value is high, close to 1, because the intensity value in the AS label is high while the intensity value in the original input image is low; while for voxels in the blood flow artifact, the corresponding intensity value is low, because the intensity value in the AS label is low while the intensity value in the original input image is high; for voxels without artifacts, the corresponding intensity value is around 0.5 because the intensity values ​​of the AS label and the original input image are close. Therefore, the artifact identity map obtained through this calculation method can effectively distinguish the artifact information of image voxels and can be used as the VI label for the artifact recognition stage.

[0095] S4, Image Segmentation: The preprocessed cardiac cine images are segmented proportionally to obtain training and validation data. Specifically, from the preprocessed MRI data of 28 subjects, the scan data of the first 15 volunteers are selected as the training set, and the scan data of the last 13 volunteers are selected as the test set. Of course, in other embodiments of the present invention, other proportions can be used for segmentation without affecting the purpose of the present invention.

[0096] S5, Constructing a Deep Learning Model: The deep learning architecture used in this invention includes two three-dimensional networks: an artifact recognition network and an artifact suppression network, as follows: Figure 2 The network parameters of the two networks are randomly initialized.

[0097] S51: The artifact recognition network architecture includes the following cascaded layers: input layer, first-stage convolutional layer, downsampling layer, second-stage convolutional layer, downsampling layer, third-stage convolutional layer, downsampling layer, fourth-stage convolutional layer, upsampling layer, fifth-stage convolutional layer, upsampling layer, sixth-stage convolutional layer, upsampling layer, seventh-stage convolutional layer, and output layer. In this system, the i-th stage convolutional layer corresponds to the Mi-th convolutional layer (M=8, i=1,2,3). The concatenate operation combines the output of the i-th stage convolutional layer with the output of the upsampled layer preceding the Mi-th convolutional layer as the input of the Mi-th convolutional layer. Each stage convolutional layer sequentially contains two groups of 3×3×3 three-dimensional convolutions, 32 groups of Group Norm and ReLU. The purpose of using Group Norm is to improve the normalization performance when the batch size is small. The downsampling layer uses max pooling with a stride of 2×2×1 voxels, and the corresponding upsampling layer is a max unpooling operation. The output layer of the VI network consists of a 1×1×1 three-dimensional convolutional layer and a sigmoid layer. The purpose of using the sigmoid layer is to limit the VI output range to [0,1].

[0098] S52: Compared to artifact suppression networks, artifact recognition networks have a similar architecture except that the output layer does not contain a sigmoid layer. The AS network is a two-input network, such as... Figure 2 As shown, the input includes the original cardiac film image and the artifact recognition map output by the VI network. The purpose of inputting the artifact recognition map into the AS network is: given any cardiac film image, the artifact recognition map provides the identification information corresponding to each voxel, such as black band artifacts, blood flow artifacts, and non-artifacts; guided by the artifact recognition map, the AS network is assisted in learning artifact suppression more specifically, thereby having better generalization ability.

[0099] S6, Constructing the Network Loss Function: This invention includes two networks: an artifact detection network and an artifact suppression network. The loss functions defined for the VI network and the AS network have the same form, explained in detail below:

[0100] Input a 3D cardiac magnetic resonance cinematic image x, and use an artifact recognition network to obtain an artifact recognition map F. VI (x), y VI It is a VI tag; the artifact suppression network is for (x,F) VI(x)) The two inputs are processed to obtain the artifact-suppressed result F. AS (x,F VI (x)), y AS It is an AS tag.

[0101] L VI (x)=L recon (F VI (x),y VI )+λ VI L perceptual (F VI (x),y VI )

[0102] L AS (x)=L recon (F AS (x,F VI (x)),y AS )+λ AS L perceptual (F AS (x,F VI (x)),y AS )

[0103] Among them, L recon The image restoration loss is represented by the L2 norm difference between the network output and the corresponding label; L perceptual The perceptual loss is represented by the norm difference between the features output from the network output and the corresponding labels at each stage of the VGG19 network; λ is the weight that balances the image restoration loss and the perceptual loss. In this embodiment, λ VI =λ AS =0.01.

[0104] S7, Training the Model: The original cardiac film image is input into the artifact recognition network, with the VI label as the training target. Training is performed using the loss function defined above. When the number of training iterations reaches a certain number and the validation set loss no longer decreases, training is stopped, and the network parameters of the VI network are saved. After the VI network is trained, the VI network is fixed, and the original cardiac film image and the artifact recognition map output by the VI network are input into the artifact suppression AS network, with the AS label as the training target. Training is performed using the loss function defined above. When the validation set loss no longer decreases, training is stopped, and the network parameters of the AS network are saved. In this embodiment, the Adam optimizer is used, the learning rate is kept at 0.0001, the VI network is iterated 700 times, and the AS network is iterated 500 times.

[0105] S8, Obtain the image after artifact suppression: Remove the black band artifact and blood flow artifact in cardiac cine imaging based on the trained artifact recognition network and artifact suppression network.

[0106] Specifically, the process of obtaining artifact-suppressed cardiac cine images based on the final artifact suppression model includes: First, acquiring a single cardiac cine image and performing the same preprocessing as in step S2 to obtain a cardiac cine image with a spatial dimension of 192×192 (no time dimension requirement); Second, inputting the preprocessed 3D cardiac cine image into the VI network to obtain an artifact recognition map, and then combining the artifact recognition map with the cardiac cine image... Figure 1 The same input is fed into the AS network to obtain the result after artifact suppression.

[0107] Based on this invention, given any cardiac film image, not only can an image with jointly suppressed black band artifacts and blood flow artifacts be obtained, but also an artifact recognition map can be generated. The artifact recognition map can characterize the regions and types of artifacts in the original input image, and can be used to understand how the neural network processes each voxel. Compared with the black-box output method of deep learning models, this invention increases the interpretability of the model's artifact processing process and has better clinical applicability.

[0108] Example 3:

[0109] The present invention also provides a deep learning-based cardiac magnetic resonance imaging (MRI) artifact suppression system. The deep learning-based cardiac MRI artifact suppression system can be implemented by executing the process steps of the deep learning-based cardiac MRI artifact suppression method. That is, those skilled in the art can understand the deep learning-based cardiac MRI artifact suppression method as a preferred embodiment of the deep learning-based cardiac MRI artifact suppression system.

[0110] The cardiac magnetic resonance cinematic imaging artifact suppression system based on deep learning provided by the present invention includes: Module M1: acquiring cardiac magnetic resonance cinematic images and preprocessing them, performing joint image computation to obtain training labels, and dividing the acquired images into training and testing sets; Module M2: constructing a deep learning model, the deep learning model including an artifact recognition network and an artifact suppression network, inputting a single three-dimensional cardiac cinematic image into the artifact recognition network to obtain a corresponding artifact recognition map; combining the artifact recognition map and the cardiac cinematic image in the artifact suppression network, simultaneously suppressing black band artifacts and blood flow artifacts to obtain an image after black band artifacts and blood flow artifact suppression; Module M3: constructing a network training loss function, defining a loss function in the form of a combination of reconstruction loss and perception loss; selecting a model optimizer, and training the cardiac cinematic image artifact suppression model; Module M4: saving the trained cardiac magnetic resonance cinematic image artifact suppression model, inputting cinematic image data that has not participated in the deep learning framework training into the model for testing, outputting artifact suppression results, and saving them to a database for preparation for clinical cardiac function analysis.

[0111] The preprocessing of the acquired cardiac magnetic resonance cine images includes: first, adjusting the k-space size to fix the physical resolution of the image at 1.5mm×1.5mm according to the image resolution at the time of image acquisition; second, cropping the image size to 192×192; then, for images with different bias frequencies, using 0Hz as the reference image, registering the remaining images in a rigid registration manner; finally, normalizing the image intensity to the range of [0,1].

[0112] The process of obtaining training labels includes: obtaining training labels for the artifact suppression network: selecting multiple sets of cardiac film images with a bias frequency of 0Hz and obtaining a joint image by averaging; obtaining training labels for the artifact recognition network: dividing the artifact suppression label by the cardiac film images with different bias frequencies, subtracting 1, and then using the sigmoid function to limit the numerical range to [0,1] to obtain the artifact identity map.

[0113] The construction process of the artifact recognition network and artifact suppression network includes: The artifact recognition network architecture sequentially comprises an input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, a downsampling layer, a third-stage convolutional layer, a downsampling layer, a fourth-stage convolutional layer, an upsampling layer, a fifth-stage convolutional layer, an upsampling layer, a sixth-stage convolutional layer, an upsampling layer, a seventh-stage convolutional layer, and an output layer; wherein the i-th stage convolutional layer corresponds to the Mi-th stage convolutional layer, M=8, i=1,2,3, and the output of the i-th stage convolutional layer is combined with the output of the upsampling layer preceding the Mi-th stage convolutional layer as the input of the Mi-th stage convolutional layer; each stage convolutional layer sequentially comprises two sets of 3×3×3 three-dimensional convolutions with a kernel size of 3×3×3 and 32 groups of group normalization. Norm and ReLU functions; the downsampling layer uses max pooling with a stride of 2×2×1 voxels, and the corresponding upsampling layer uses unmax pooling; the output layer of the artifact recognition network consists of a 3D convolutional layer with a kernel size of 1×1×1 and a sigmoid layer, limiting the artifact recognition output range to [0,1]; compared with the artifact recognition network, the artifact suppression network architecture only excludes the sigmoid layer in the output layer; the artifact suppression network is a dual-input network, with the input containing the original cardiac film image and the artifact recognition map output by the artifact recognition network; based on the artifact recognition map, the identity information corresponding to each voxel is obtained, distinguishing the positions of black band artifacts, blood flow artifacts and non-artifacts.

[0114] The loss function L defined by the artifact recognition network VI (x) and the loss function L defined by the artifact suppression network AS (x) has the same form, specifically: input a three-dimensional cardiac magnetic resonance film image x, and use an artifact recognition network to obtain an artifact recognition map F. VI (x), y VIIt is an artifact recognition label; the artifact suppression network pairs (x,F) VI (x)) The two inputs are processed to obtain the artifact-suppressed result F. AS (x,F VI (x)), y AS It is an artifact suppression label;

[0115] L VI (x)=L recon (F VI (x),y VI )+λ VI L perceptual (F VI (x),y VI )

[0116] L AS (x)=L recon (F AS (x,F VI (x)),y AS )+λ AS L perceptual (F AS (x,F VI (x)),y AS )

[0117] Among them, L recon Representing image restoration loss, it quantifies the difference between the network output and the corresponding label. L2, L1, SSIM, or any other metric reflecting vector differences can be used. perceptual Representing perceptual loss, it quantifies the difference between the network output and the features output by the corresponding labels in the high-dimensional feature space of the trained network. It can be represented by the 2-norm, 1-norm, SSIM, or any other metric reflecting vector differences; λ VI , λ AS We weigh the image restoration loss and perceptual loss for the artifact recognition network and the artifact suppression network.

[0118] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0119] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A deep learning-based method for suppressing artifacts in cardiac magnetic resonance cinema imaging, characterized in that, include: Step S1: Acquire cardiac magnetic resonance cine images with black band artifacts or blood flow artifacts, perform preprocessing operations on the images, and perform joint image calculations to obtain artifact-free images; Step S2: Construct a deep learning model. The input of the artifact suppression network in the model is a heart movie image with black band artifacts or blood flow artifacts, and the output is a heart movie image after black band artifacts or blood flow artifacts are suppressed. Step S3: Use the preprocessed images with and without artifacts to train the deep learning model until the parameters of the artifact suppression network reach the target conditions, obtain the trained deep learning model and save it. Step S4: Based on the trained deep learning model, perform artifact removal on the cardiac cinema image with black band artifacts or blood flow artifacts to be removed, and obtain the cardiac cinema image after artifact removal. Before the artifact suppression network, an artifact recognition network is added to output an artifact identity map. The input of the artifact recognition network is a cardiac cinema image with black band artifacts or blood flow artifacts, and the output is the corresponding artifact identity map. The input of the artifact suppression network is an image with artifacts and its corresponding artifact identity map, and the output is a cardiac cinema image with black band artifacts or blood flow artifacts suppressed. The specific execution steps include: Obtain training labels for the artifact recognition network: Perform joint calculations on images without artifacts and images with artifacts, requiring the joint calculation results to show the location and type of artifacts, which will serve as the artifact identity map. The artifact recognition network architecture consists of an input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, a downsampling layer, a third-stage convolutional layer, a downsampling layer, a fourth-stage convolutional layer, an upsampling layer, a fifth-stage convolutional layer, an upsampling layer, a sixth-stage convolutional layer, an upsampling layer, a seventh-stage convolutional layer, and an output layer. The i-th stage convolutional layer corresponds to the Mi-th stage convolutional layer, where M = 8 and i = 1, 2, 3. The output of the i-th stage convolutional layer is combined with the output of the upsampling layer preceding the Mi-th stage convolutional layer, serving as the input to the Mi-th stage convolutional layer. The output layer of the artifact recognition network consists of a three-dimensional convolutional layer with a kernel size of 1×1×1 and a sigmoid layer, which restricts the output range of artifact recognition to [0,1]. The artifact suppression network construction process includes: the artifact suppression network architecture consists of a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, an (N-1)th convolutional layer, an upsampling layer, an Nth convolutional layer, and an output layer; where the i-th stage convolutional layer corresponds to the Ni-th convolutional layer, the output of the i-th stage convolutional layer is combined with the output of the upsampling layer after the Ni-th convolutional layer as the input of the Nth convolutional layer, where N is a positive integer greater than 2; the artifact suppression network is a dual-input network, the input includes the original heart film image and the artifact identity map output by the artifact recognition network; based on the artifact identity map, the identity information corresponding to each voxel is obtained, the positions of black band artifacts, blood flow artifacts and non-artifacts are distinguished, and the output is the image after the black band artifacts and blood flow artifacts are suppressed. Model training process: The artifact recognition network is trained using preprocessed images with artifacts and artifact identity maps; the artifact suppression network is trained using preprocessed images with artifacts and images without artifacts; until the parameters of the artifact recognition network and the artifact suppression network reach the target conditions, the trained deep learning model is obtained. Model processing procedure: The trained deep learning model is saved. The cardiac magnetic resonance imaging data with black band artifacts or blood flow artifacts to be removed are subjected to artifact removal. The image data after artifact removal is obtained by relying on the trained deep learning model and saved to the database. The artifact recognition network outputs artifact identity map and the artifact suppression network outputs artifact-free image, which is used to prepare for clinical cardiac function analysis.

2. The method for suppressing artifacts in cardiac magnetic resonance cinema imaging based on deep learning according to claim 1, characterized in that, Preprocessing of the acquired cardiac magnetic resonance cine images includes: First, adjust the k-space size to fix the physical resolution of the image based on the image resolution at the time of image acquisition; Secondly, crop the image size to a uniform size; Then, for images with different bias frequencies, the remaining images are registered using 0Hz as the reference image; Finally, the image intensity is normalized to the range [0,1].

3. The method for suppressing artifacts in cardiac magnetic resonance cinema imaging based on deep learning according to claim 1, characterized in that, The process of image joint computation includes: obtaining training labels for the artifact suppression network: selecting multiple sets of cardiac film images with a bias frequency of 0Hz, and obtaining artifact-free images by averaging.

4. The method for suppressing artifacts in cardiac magnetic resonance cinema imaging based on deep learning according to claim 1, characterized in that, Define the loss function L of the artifact recognition network. VI (x) and the loss function L of the artifact suppression network AS (x), specifically in the form of: Input a 3D cardiac magnetic resonance cinematic image x, and use an artifact recognition network to obtain an artifact recognition map F. VI (x), y VI It is an artifact recognition label; the artifact suppression network pairs (x,F) VI (x)) The two inputs are processed to obtain the artifact-suppressed result F. AS (x,F VI (x)), y AS It is an artifact suppression label; L VI (x)=L recon (F VI (x),y VI )+λ VI L perceptual (F VI (x),y VI ) L AS (x)=L recon (F AS (x,F VI (x)),y AS )+λ AS L perceptual (F AS (x,F VI (x)),y AS ) Among them, L recon L represents image restoration loss; perceptual Represents perceived loss; λ VI , λ AS We weigh the image restoration loss and perceptual loss for the artifact recognition network and the artifact suppression network.

5. A deep learning-based cardiac magnetic resonance cinematic imaging artifact suppression system, characterized in that, include: Data acquisition module: Acquires cardiac magnetic resonance cine images with black band artifacts or blood flow artifacts; Data processing module: performs joint image calculations on cardiac MRI cine images with black band artifacts or blood flow artifacts to obtain artifact-free images; performs preprocessing operations on the images with artifacts and the images without artifacts. Model building module: Constructs a deep learning model, which is an artifact suppression network used to suppress black band artifacts or blood flow artifacts in cardiac film images; Model training module: The artifact suppression network is trained using preprocessed images with and without artifacts; The training continues until the parameters of the artifact suppression network reach the target conditions, resulting in a trained deep learning model. Model processing module: Saves the trained deep learning model, performs artifact removal on cardiac magnetic resonance image data with black band artifacts or blood flow artifacts to be removed, obtains artifact-removed image data based on the trained deep learning model, and saves it to the database; the artifact suppression network outputs artifact-free images to prepare for clinical cardiac function analysis. Before the artifact suppression network, an artifact recognition network is added to output an artifact identity map. The input of the artifact recognition network is a cardiac cinema image with black band artifacts or blood flow artifacts, and the output is the corresponding artifact identity map. The input of the artifact suppression network is an image with artifacts and its corresponding artifact identity map, and the output is a cardiac cinema image with black band artifacts or blood flow artifacts suppressed. The specific execution steps include: Obtain training labels for the artifact recognition network: Perform joint calculations on images without artifacts and images with artifacts, requiring the joint calculation results to show the location and type of artifacts, which will serve as the artifact identity map. The artifact recognition network architecture consists of an input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, a downsampling layer, a third-stage convolutional layer, a downsampling layer, a fourth-stage convolutional layer, an upsampling layer, a fifth-stage convolutional layer, an upsampling layer, a sixth-stage convolutional layer, an upsampling layer, a seventh-stage convolutional layer, and an output layer. The i-th stage convolutional layer corresponds to the Mi-th stage convolutional layer, where M = 8 and i = 1, 2, 3. The output of the i-th stage convolutional layer is combined with the output of the upsampling layer preceding the Mi-th stage convolutional layer, serving as the input to the Mi-th stage convolutional layer. The output layer of the artifact recognition network consists of a three-dimensional convolutional layer with a kernel size of 1×1×1 and a sigmoid layer, which restricts the output range of artifact recognition to [0,1]. The artifact suppression network construction process includes: the artifact suppression network architecture consists of a cascaded input layer, a first-stage convolutional layer, a downsampling layer, a second-stage convolutional layer, an (N-1)th convolutional layer, an upsampling layer, an Nth convolutional layer, and an output layer; where the i-th stage convolutional layer corresponds to the Ni-th convolutional layer, the output of the i-th stage convolutional layer is combined with the output of the upsampling layer after the Ni-th convolutional layer as the input of the Nth convolutional layer, where N is a positive integer greater than 2; the artifact suppression network is a dual-input network, the input includes the original heart film image and the artifact identity map output by the artifact recognition network; based on the artifact identity map, the identity information corresponding to each voxel is obtained, the positions of black band artifacts, blood flow artifacts and non-artifacts are distinguished, and the output is the image after the black band artifacts and blood flow artifacts are suppressed. Model training process: The artifact recognition network is trained using preprocessed images with artifacts and artifact identity maps; the artifact suppression network is trained using preprocessed images with artifacts and images without artifacts; until the parameters of the artifact recognition network and the artifact suppression network reach the target conditions, the trained deep learning model is obtained. Model processing procedure: The trained deep learning model is saved. The cardiac magnetic resonance imaging data with black band artifacts or blood flow artifacts to be removed are subjected to artifact removal. The image data after artifact removal is obtained by relying on the trained deep learning model and saved to the database. The artifact recognition network outputs artifact identity map and the artifact suppression network outputs artifact-free image, which is used to prepare for clinical cardiac function analysis.

6. The deep learning-based cardiac magnetic resonance cinema imaging artifact suppression system according to claim 5, characterized in that, Preprocessing of the acquired cardiac magnetic resonance cine images includes: First, adjust the k-space size to fix the physical resolution of the image based on the image resolution at the time of image acquisition; Secondly, crop the image size to a uniform size; Then, for images with different bias frequencies, the remaining images are registered using 0Hz as the reference image; Finally, the image intensity is normalized to the range [0,1].

7. The deep learning-based cardiac magnetic resonance cinema imaging artifact suppression system according to claim 5, characterized in that, The process of image joint computation includes: obtaining training labels for the artifact suppression network: selecting multiple sets of cardiac film images with a bias frequency near 0Hz, and obtaining artifact-free images by averaging.

8. The deep learning-based cardiac magnetic resonance cinema imaging artifact suppression system according to claim 5, characterized in that, Define the loss function L of the artifact recognition network. VI (x) and the loss function L of the artifact suppression network AS (x), specifically in the form of: Input a 3D cardiac magnetic resonance cinematic image x, and use an artifact recognition network to obtain an artifact recognition map F. VI (x), y VI It is an artifact recognition label; the artifact suppression network pairs (x,F) VI (x)) The two inputs are processed to obtain the artifact-suppressed result F. AS (x,F VI (x)), y AS It is an artifact suppression label; L VI (x)=L recon (F VI (x),y VI )+λ VI L perceptual (F VI (x),y VI ) L AS (x)=L recon (F AS (x,F VI (x)),y AS )+λ AS L perceptual (F AS (x,F VI (x)),y AS ) Among them, L recon L represents image restoration loss; perceptual Represents perceived loss; λ VI , λ AS We weigh the image restoration loss and perceptual loss for the artifact recognition network and the artifact suppression network.

Images