CN114596378A - Sparse angle CT artifact removing method - Google Patents
Sparse angle CT artifact removing method Download PDFInfo
- Publication number
- CN114596378A CN114596378A CN202210242057.7A CN202210242057A CN114596378A CN 114596378 A CN114596378 A CN 114596378A CN 202210242057 A CN202210242057 A CN 202210242057A CN 114596378 A CN114596378 A CN 114596378A
- Authority
- CN
- China
- Prior art keywords
- image
- sparse
- rsdb
- network
- loss
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 230000004927 fusion Effects 0.000 claims abstract description 35
- 230000006870 function Effects 0.000 claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000000605 extraction Methods 0.000 claims abstract description 11
- 230000008447 perception Effects 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 17
- 238000013527 convolutional neural network Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 230000009191 jumping Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 abstract description 5
- 230000000694 effects Effects 0.000 abstract 1
- 238000002591 computed tomography Methods 0.000 description 51
- 238000005259 measurement Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 230000005855 radiation Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000009795 derivation Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 206010028980 Neoplasm Diseases 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 201000011510 cancer Diseases 0.000 description 1
- 238000013170 computed tomography imaging Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/003—Reconstruction from projections, e.g. tomography
- G06T11/008—Specific post-processing after tomographic reconstruction, e.g. voxelisation, metal artifact correction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2211/00—Image generation
- G06T2211/40—Computed tomography
- G06T2211/424—Iterative
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method for removing a sparse angle CT artifact, which is based on a perception loss and a feature fusion attention residual error network and comprises the steps of preprocessing an input sparse CT image with the artifact and an original CT image; performing feature extraction on the preprocessed sparse CT image and the original CT image to obtain a feature map; performing improved feature fusion attention residual error network training on the obtained feature map by taking the perception loss as a loss function, and taking an original CT image as a label, so that the sparse CT image can remove artifacts but keep the structure of the sparse CT image unchanged; and finally, taking the sparse CT image as input, removing artifacts by using a trained network model, and calculating the peak signal-to-noise ratio and the structural similarity. The method takes the perception loss as a loss function, and can enhance the detail information. And an attention mechanism is added into the characteristic fusion residual error network, so that the performance of the network is improved, and finally, the effect of quickly and effectively removing the artifacts is realized.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to a sparse angle CT artifact removing method.
Background
Computed Tomography (CT) is a key medical imaging technique. However, the potential problem of X-ray radiation in CT imaging is of concern and the radiation dose increases the risk of cancer. Therefore, much research has focused on reducing the radiation dose, which has led to the development of many effective methods. The initial approach focused on reducing X-ray exposure by reducing or adjusting the tube. Methods that reduce the number of projections for a given scan trajectory can cause severe streak artifacts in the CT image. Reconstructing an image from sparsely sampled data may solve the problem of insufficient data due to limited sampling angles. However, reconstruction of sparse view CT projection data (e.g., using FBP) will produce streak artifacts in the CT image.
In recent years, Deep Learning (DL) has become a powerful feature learning method. The natural image problem has made great progress by virtue of the powerful function of deep learning and the end-to-end deep supervised learning method and by means of the convolutional neural network. Such as object detection, image segmentation and image super resolution. Many researchers have used deep learning methods for sparse CT image reconstruction and artifact removal.
Chinese patent application 'CT sparse reconstruction artifact correction method and system based on residual learning' (patent application No. CN201711097204.1, publication No. CN107871332A), learning artifact characteristics by establishing a residual neural network structure to obtain an artifact map; and finally, performing residual error by using the sparsely reconstructed image and the artifact image to restore a clear CT image. However, the method has the problems of long training time, more model parameters and the like, and increases the calculation amount of the network.
Disclosure of Invention
In view of the problems of the above-mentioned techniques and materials in the prior art, the present invention provides a fast, simple and robust method for removing the artifacts of sparse angular CT.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
the invention discloses a method for removing sparse angle CT artifacts, which comprises the following specific steps:
step 2, extracting image characteristics of the preprocessed sparse angle CT data through a VGG19 network so as to facilitate subsequent training;
step 3, constructing a feature fusion attention residual error network, wherein a residual error jumping dense block is used as a foundation of a network shallow layer in the network, a local feature fusion layer is positioned behind two convolution layers, the residual error jumping dense block jumps over the connection of the two local feature fusion layers, a feature fusion result of a former module is used as the input of a latter/next module, the features of local feature fusion are stacked, and residual error information is used for integrating feature information to form a basic network architecture;
step 4, training by taking the sparse CT image as input, the original CT image as a label and the perception loss as a loss function to obtain a trained feature fusion attention residual error network model;
and 5, testing by taking the sparse CT images of the same part as input to obtain an image with removed artifacts, and calculating the peak signal-to-noise ratio and the structural similarity between the output CT image and the labeled CT image.
The invention is further improved in that: the step 1 specifically comprises the following steps:
step 1.1, modifying the sizes of the head sparse CT image and the original CT image in the data set to enable the resolution of each image to be m multiplied by n, wherein m and n are the length and the height of the image respectively, and the sparse CT image and the original CT image are in a format of dcm.
And step 1.2, converting the image with the new size in the step 1.1 into a sensor data type suitable for the Pythrch framework.
The invention is further improved in that: in step 2, three identical images are put together to form three channels to adapt to a VGG19 network, and the preprocessed sparse CT image and the original CT image are subjected to feature extraction through a trained VGG19 model, wherein the sparse CT image is used for artifact removal training, and the original CT image is used for prediction training with streak artifacts.
The invention is further improved in that: the specific operation of constructing the feature fusion attention residual error network in the step 3 is as follows: connecting among a plurality of residual error dense blocks, fusing local features in each RSDB, and transferring the local features to a subsequent residual error block, wherein the f-th RSDB directly introduces local features in the (f-1) -th RSDB or RB and performs feature fusion, if f is 1, for the model adaptation residual error block RB, the local feature fusion LFF is expressed as:
wherein Ff,LFShowing the f-th RSDB local feature fusion result,represents the Concat feature graph Stacking function, Ff-1,LFRepresents the (F-1) th RSDB local feature fusion result, Ff,2The result of shallow feature extraction performed on the f-th RSDB input through two Conv layers is shown as:
where δ denotes the ReLU activation function, Wf,2And Wf,1Respectively represent the weight of the 1 st convolutional layer of the 2 nd convolutional layer, b2And b1Respectively, the 2 nd and 1 st convolution layers, Ff-1Represents the results of the f-1 st RSDB output;
after local feature fusion is carried out in each RSDB, the feature graph obtained by the attention module is transmitted to the next RSDB for the same operation, and the features of the appointed layers in different basic blocks are stacked together to realize jump connection with intervals and effective feature extraction, wherein the feature graph stack is expressed as:
wherein, F0For the 1 st RSDB pre-model to adapt to the characteristic map of the Conv layer in the residual block RB, it can be mentioned in the following FFRN network structure that in order to adapt to the RSDB structure, an RB block is introduced first in the RSDB. Finally, local residual learning is introduced to further improve the information flow, Ff,3The presentation translation layer function uses a convolutional layer with a kernel size of axa.
The attention mechanism module has two ways to obtain characteristics, namely a main road can replace any leading edge convolution module; and the mask channel is mainly used for processing the features to generate soft masks, and becomes a control gate to control new features, so that the features are more concentrated.
Suppose the input of the main road is x, the output is T (x), the mask output is M (x), and the output of the attention module at this time is M (x)
Hi,c(x)=Mi,c(x)*Ti,c(x)
Where i represents the spatial position of the feature and c represents the feature channel.
In the attention module, the mask can be used not only as a feature selector in the forward inference process, but also as a gradient update filter in the backward propagation process.
When the gradient is updated in the reverse propagation way, the parameter theta of the mask part is not updated, but the parameter of the main road is only updatedWhere θ is the parameter of the mask branch and φ is the parameter of the trunk branch. Parameters to trunk branchesAnd (5) derivation and gradient calculation. Therefore, the robustness of the attention module to noise is strong, and the influence of the noise on gradient updating can be effectively reduced. This property makes the network very robust to noise signatures. Each main road is matched with one mask road, and the characteristic mask of the step is learned, so that the mask is more targeted.
Since the mask value is [0,1], refer to the residual learning structure of Resnet. The output of the module is improved into a residual error structure:
Hi,c(x)=(1+Mi,c(x))*Ti,c(x)
the invention is further improved in that: the perceived loss in step 4 is expressed as:
whereinAndfeatures of sparse CT and original CT images, respectively, extracted through a VGG19 network of j layers, where WjIs the weight coefficient of the j-th layer, GjAnd HjRespectively representing the length and width of the j-th network.
The loss of absolute value mainly measures the difference between the pixel values of the neural network output image and the reference image, and its formula is as follows:
m and N are the number of rows and columns of the image, which calculates the expectation of the absolute value of the difference in pixels between the convolutional neural network output image I and the reference image R, where I (k, l) and R (k, l) represent the input image pixel and the output image pixel of the ith row and ith column, respectively;
the loss function in the network combines absolute loss and perceptual loss, expressed as follows:
Lloss=Lmae+λLfeat
where λ is the weighting factor occupied by the perceptual loss.
The invention has the beneficial effects that:
1. the method is based on the perception loss, and takes the perception loss and the image characteristic loss as loss functions in a trained network model, so that the loss of the network is reduced, and the quality of the output CT image is improved;
2. the invention combines a characteristic fusion attention residual error network, adopts a residual error learning method, and realizes the prediction of the streak artifact by using the image with the streak artifact as a label after subtracting the output CT image from the input sparse CT image with the streak artifact;
3. compared with the traditional residual error network, the network of the invention makes full use of local characteristics, introduces an attention module and learns more detailed information;
4. the invention does not change the size of the image in the training process, and perfectly keeps the details and edge information of the image;
5. compared with the traditional residual error network, the network of the invention has simple structure, and consumes less memory resource and less training time;
6. the images in the training process can adopt other sizes, such as 256 multiplied by 256, 512 multiplied by 512 and the like, so that the applicability of the network is greatly expanded.
Drawings
FIG. 1 is a general flow diagram of a method of an embodiment of the invention.
Fig. 2 is an RB module of the present invention.
Fig. 3 is an RSDB block of the present invention.
FIG. 4 is an attention mechanism module of the present invention.
Fig. 5 is an overall structural view of the present invention.
Detailed Description
In order to make the objects, technical solutions and novel points of the present invention more clear, the present invention will be described in detail with reference to the accompanying drawings and specific embodiments.
The invention discloses a method for removing sparse angle CT artifacts
step 1.1, intercepting original CT data (dcm format) obtained from a hospital into the same size, wherein the pixel size is m × n, m and n are the length and height of an image respectively, and m is 512;
and 1.2, converting the adjusted dcm data into a sensor data type suitable for the Pythrch frame.
Step 2, extracting image characteristics of the preprocessed sparse angle CT data through a VGG19 network so as to facilitate subsequent training; the specific operation is as follows:
the preprocessed data are all single-channel medical images, and the input of the selected VGG19 network is a three-channel color image, so the dimension needs to be added to the input image to adapt to the network. The method comprises the following steps: three identical images are selected to form three-channel (512, 512, 3) input, and feature extraction is carried out through an open-source trained VGG19 model imagenet-VGG-verydep-19 version.
Step 3, constructing a feature fusion attention residual error network, wherein a residual error jumping dense block is used as a foundation of a network shallow layer in the network, a local feature fusion layer is positioned behind two convolution layers, the residual error jumping dense block jumps over the connection of the two local feature fusion layers, a feature fusion result of a former module is used as the input of a latter/next module, the features of local feature fusion are stacked, and residual error information is used for integrating feature information to form a basic network architecture;
the Residual Dense Block (RDB) can extract multi-level features in the original image, further improving network performance. RDB is a residual error dense block proposed in RDN, all layers of RDB are fully utilized by local dense connection, accumulated features are adaptively saved by Local Feature Fusion (LFF), and shallow features and deep features are combined together by global residual error learning to perform Global Feature Fusion (GFF) to obtain global dense features.
Connecting between blocks, fusing local features in each RSDB, and finally transmitting to a subsequent residual block; the f-th RSDB directly introduces local features in the (f-1) -th RSDB or RB and performs feature fusion, wherein if f is 1, the model adaptive residual block RB is represented as:
wherein Ff,LFShowing the f-th RSDB local feature fusion result,represents the Concat feature graph Stacking function, Ff-1,LFRepresents the (F-1) th RSDB local feature fusion result, Ff,2The result of shallow feature extraction performed on the f-th RSDB input through two Conv layers can be expressed as:
where δ denotes the ReLU activation function, Wf,2And Wf,1Respectively represent the weight of the 1 st convolutional layer of the 2 nd convolutional layer, b2And b1Respectively, the 2 nd and 1 st convolution layers, Ff-1Representing the results of the f-1 RSDB output.
After local feature fusion is carried out in each RSDB, the feature graph obtained by the attention module is transmitted to the next RSDB for the same operation, and the features of the appointed layers in different basic blocks are stacked together to realize jump connection with intervals and effective feature extraction, wherein the feature graph stack is expressed as:
wherein, F0For the 1 st RSDB pre-model to adapt to the characteristic map of the Conv layer in the residual block RB, it can be mentioned in the following FFRN network structure that in order to adapt to the RSDB structure, an RB block is introduced first in the RSDB. Finally, local residual learning is introduced to further improve the information flow, Ff,3The presentation translation layer function may use convolutional layers with cores of 1x1 or 3x3 size, with different core sizes serving different functions.
The attention mechanism module has two ways to obtain characteristics, namely a main road which can replace any leading edge convolution module; and the mask channel is mainly used for processing the features to generate soft masks, and becomes a control gate to control new features, so that the features are more concentrated.
Suppose the input of the main trunk is x, the output is the main trunk branch feature map T (x), the output of the mask branch (mask) is M (x), and the output of the attention module at this time is M (x)
Hi,c(x)=Mi,c(x)*Ti,c(x)
Where i represents the spatial position of the feature and c represents the feature channel.
In the attention module, the mask can be used not only as a feature selector in the forward inference process, but also as a gradient update filter in the backward propagation process.
When the gradient is updated in the reverse propagation way, the parameter theta of the mask part is not updated, but the parameter of the main road is only updatedWhere θ is the parameter of the mask branch and φ is the parameter of the trunk branch. Parameters to trunk branchesAnd (5) derivation and gradient calculation. Therefore, the robustness of the attention module to noise is strong, and the influence of the noise on gradient updating can be effectively reduced. This property makes the network very robust to noise signatures. Each main road is matched with one mask road, and the characteristic mask of the step is learned, so that the mask is more targeted.
Since the mask value is [0,1], refer to the residual learning structure of Resnet. The output of the module is improved into a residual error structure:
Hi,c(x)=(1+Mi,c(x))*Ti,c(x)
the convolutional neural network in this example is shown in fig. 2, where the Residual Block (RB) structure is: the first layer is a 3 × 3 convolutional layer, followed by a ReLU layer, then a 3 × 3 convolutional layer, a 3 × 1 convolutional layer, a 1 × 3 convolutional layer. As shown in fig. 3, the dense Residual Skip (RSDB) block structure is: the first layer is a 3 × 3 convolutional layer, followed by a ReLU layer, then a 3 × 3 convolutional layer, a ReLU layer, a tie layer, then a 3 × 1 convolutional layer, 1 × 3 convolutional layer. The receptive field is increased, and meanwhile, the calculation speed is accelerated. As shown in fig. 4, the attention module mechanism.
And 4, taking the sparse CT as input, the original CT as a label, the perception loss as a loss function, training a feature fusion residual error network, calculating the upper and lower problem losses according to the perception difference between the sparse CT and the original CT image and the features of the corresponding convolutional layer, and updating the image value of the reconstructed CT image along with the change of the background loss. The invention adopts an iterative optimization method to correct the sparse CT image. According to the artifact removing result, adjusting the context loss function, the network parameter and the ratio of the feature sample size as necessary, and repeating the network training until the artifact is effectively removed.
The image problem can be regarded as an image transformation task that accepts an input image and then outputs the transformed image. Such as noise extraction, super-resolution reconstruction, colorization of images, etc. in image processing, an input image is a degraded low-quality image (noise, low resolution, graying) and the output is a color, high-resolution, high-quality image.
The traditional convolutional neural network-based loss function is defined by using pixel basic loss, two images are basically compared at a pixel, and if the two images are perceived to be the same but the pixels are different, the pixel loss causes great deviation. The perception loss realizes perception comparison based on the characteristics of the high-level convolutional neural network without the basic loss of pixels.
The perceptual loss function is defined as follows:
whereinAndfeatures of sparse CT and original CT images, respectively, extracted through a VGG19 network of j layers, where WjIs the weight coefficient of the j-th layer, CjAnd HjRespectively representing the length and width of the j-th network.
The loss of absolute value mainly measures the difference between the pixel values of the neural network output image and the reference image, and its formula is as follows:
m and N are the number of rows and columns of the image, which calculates the expectation of the absolute value of the difference in pixels between the convolutional neural network output image I and the reference image R. Where I (k, l) and R (k, l) represent the input image pixel and the output image pixel, respectively, of the kth row and the l column. The resulting image details of the neural network with MAE loss are better.
Loss function in the network:
Lloss=Lmae+λLfeat
the loss function in the network combines absolute loss and perceptual loss, where λ is the weighting factor occupied by the perceptual loss.
This embodiment uses a Pythrch framework in a Python environment on a GeForce GTX 1080Ti processor. For network optimization, the present invention uses an Adam optimizer and a nonlinear activation function ReLU. During network training, the number of the past iterations of the invention can be set to 100, and the learning rate is set to be adaptive. To achieve accurate convergence, the step size is set to 2.
And (5) testing the sparse CT images of the same part as input to obtain an image with the artifact removed, and calculating the peak signal-to-noise ratio and the structural similarity between the output CT image and the labeled CT image. The Peak Signal-to-Noise Ratio (PSNR) is defined as follows:
m and N represent the length and width of the image, L is the maximum value of the gray level in the image, im is the image to be measured, and ref is the reference image. The PSNR calculates the difference between the measurement image im and the reference image ref, and the larger the PSNR value is, the smaller the difference between the measurement image and the reference image is, i.e., the smaller the noise in the measurement image im is.
Structural Similarity Index Measurement (SSIM) is defined as follows:
SSIM(im,ref)=L(im,ref)×C(im,ref)×S(im,ref)
uimand urefRepresenting the mean, σ, of the measurement and reference images, respectivelyimAnd σrefRespectively representing the standard deviation, σ, of the measured image and the reference imageim,refRepresenting the covariance of the measurement image and the reference image, C1、C2And C3Is a constant. The structural similarity is used for comparing the similarity between the measured image im and the reference image ref, the similarity between the images is calculated from three aspects of image brightness, contrast and structure, the value range of the structural similarity is between 0 and 1, and the higher the structural similarity is, the more similar the measured image and the reference image are.
Although the present invention has been described with reference to the preferred embodiments, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (6)
1. A method for sparse angular CT artifact removal, characterized by: the method comprises the following specific steps:
step 1, preprocessing sparse angle CT data with fringe artifacts and original CT data, and processing the sparse angle CT data and the original CT data into a dcm-format image with a specified size;
step 2, extracting image characteristics of the preprocessed sparse angle CT data through a VGG19 network so as to facilitate subsequent training;
step 3, constructing a feature fusion attention residual error network;
step 4, training by taking the sparse CT image as input, the original CT image as a label and the perception loss as a loss function to obtain a trained feature fusion attention residual error network model;
and 5, testing by taking the sparse CT images of the same part as input to obtain an image with removed artifacts, and calculating the peak signal-to-noise ratio and the structural similarity between the output CT image and the labeled CT image.
2. The method of sparse angular CT artifact removal as claimed in claim 1, wherein: the step 1 specifically comprises the following steps:
step 1.1, modifying the sizes of the head sparse CT image and the original CT image in the data set to enable the resolution of each image to be m multiplied by n, wherein m and n are the length and the height of the image respectively, and the sparse CT image and the original CT image are in a format of dcm.
And step 1.2, converting the image with the new size in the step 1.1 into a sensor data type suitable for the Pythrch framework.
3. The method of sparse angular CT artifact removal as claimed in claim 1, wherein: in step 2, three identical images are put together to form three channels to adapt to a VGG19 network, and the preprocessed sparse CT image and the original CT image are subjected to feature extraction through a trained VGG19 model, wherein the sparse CT image is used for artifact removal training, and the original CT image is used for prediction training with streak artifacts.
4. The method of sparse angular CT artifact removal as claimed in claim 1, wherein: the specific operation of constructing the feature fusion attention residual error network in the step 3 is as follows: connecting among a plurality of dense residual error jumping modules, fusing local features in each RSDB and transmitting the fused local features to a subsequent residual error block, wherein the F-th RSDB directly introduces the local features and F in the (F-1) th RSDB or RBf,2And performing feature fusion, wherein if f is 1, the model adaptive residual block RB is represented as:
wherein Ff,LFShowing the f-th RSDB local feature fusion result,represents the Concat feature graph Stacking function, Ff-1,LFRepresents the (F-1) th RSDB local feature fusion result, Ff,2The result of shallow feature extraction performed on the f-th RSDB input through two Conv layers is shown as:
where δ denotes the ReLU activation function, Wf,2And Wf,1Respectively represent the 2 ndWeight of convolution layer 1 st convolution layer, b2And b1Respectively, the 2 nd and 1 st convolution layers, Ff-1Represents the results of the f-1 st RSDB output;
after local feature fusion is carried out in each RSDB, the feature graph obtained by the attention module is transmitted to the next RSDB for the same operation, and the features of the appointed layers in different basic blocks are stacked together to realize jump connection with intervals and effective feature extraction, wherein the feature graph stack is expressed as:
wherein, F0For the 1 st RSDB pre-model to adapt to the characteristic map of the Conv layer in the residual block RB, it can be mentioned in the following FFRN network structure that in order to adapt to the RSDB structure, an RB block is introduced first in the RSDB. Finally, local residual learning is introduced to further improve the information flow, Ff,3The presentation translation layer function uses a convolutional layer with a kernel size of axa.
5. The method of sparse angular CT artifact removal as claimed in claim 4, wherein: suppose the input of the main road is x, the output is T (x), the mask output is M (x), and the output of the attention module is:
Hi,c(x)=Mi,c(x)*Ti,c(x)
wherein i represents the spatial position of the feature and c represents the feature channel;
when mask is used as the gradient update filter in the back propagation process:
when the gradient is updated in the reverse propagation way, the parameter theta of the mask part is not updated, but the parameter of the main road is only updatedWhere θ is a parameter of the mask branch and φ is a parameter of the trunk branch;
since the mask value is [0,1], referring to the residual learning structure of Resnet, the output of this module is the residual structure:
Hi,c(x)=(1+Mi,c(x))*Ti,c(x)。
6. the method of sparse angular CT artifact removal as claimed in claim 1, wherein: the perceived loss in step 4 is expressed as:
whereinAndfeatures of sparse CT and original CT images, respectively, extracted through a VGG19 network of j layers, where WjIs the weight coefficient of the j-th layer, CjAnd HjRespectively representing the length and width of the j-th network.
The loss of absolute value mainly measures the difference between the pixel values of the neural network output image and the reference image, and its formula is as follows:
m and N are the number of rows and columns of the image, which calculates the expectation of the absolute value of the difference in pixels between the convolutional neural network output image I and the reference image R, where I (k, l) and R (k, l) represent the input image pixel and the output image pixel of the kth row and the l column, respectively,
the loss function in the network combines absolute loss and perceptual loss, expressed as follows:
Lloss=Lmae+λLfeat
where λ is the weighting factor occupied by the perceptual loss.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210242057.7A CN114596378A (en) | 2022-03-11 | 2022-03-11 | Sparse angle CT artifact removing method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210242057.7A CN114596378A (en) | 2022-03-11 | 2022-03-11 | Sparse angle CT artifact removing method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114596378A true CN114596378A (en) | 2022-06-07 |
Family
ID=81809137
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210242057.7A Pending CN114596378A (en) | 2022-03-11 | 2022-03-11 | Sparse angle CT artifact removing method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114596378A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115103118A (en) * | 2022-06-20 | 2022-09-23 | 北京航空航天大学 | High dynamic range image generation method, device, equipment and readable storage medium |
CN118351210A (en) * | 2024-06-17 | 2024-07-16 | 南昌睿度医疗科技有限公司 | CT image artifact removal method, system, storage medium and electronic equipment |
CN118351210B (en) * | 2024-06-17 | 2024-08-27 | 南昌睿度医疗科技有限公司 | CT image artifact removal method, system, storage medium and electronic equipment |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107871332A (en) * | 2017-11-09 | 2018-04-03 | 南京邮电大学 | A kind of CT based on residual error study is sparse to rebuild artifact correction method and system |
WO2021042270A1 (en) * | 2019-09-03 | 2021-03-11 | 中山大学 | Compression artifacts reduction method based on dual-stream multi-path recursive residual network |
CN114004912A (en) * | 2021-11-08 | 2022-02-01 | 南京邮电大学 | CBCT image artifact removing method |
-
2022
- 2022-03-11 CN CN202210242057.7A patent/CN114596378A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107871332A (en) * | 2017-11-09 | 2018-04-03 | 南京邮电大学 | A kind of CT based on residual error study is sparse to rebuild artifact correction method and system |
WO2021042270A1 (en) * | 2019-09-03 | 2021-03-11 | 中山大学 | Compression artifacts reduction method based on dual-stream multi-path recursive residual network |
CN114004912A (en) * | 2021-11-08 | 2022-02-01 | 南京邮电大学 | CBCT image artifact removing method |
Non-Patent Citations (2)
Title |
---|
SHIPENG XIE: "《Artifact Removal in Sparse-Angle CT Based on Feature Fusion Residual Network》", 《IEEE TRANSACTIONS ON RADIATION AND PLASMA MEDICAL SCIENCES》, vol. 5, no. 2, 9 June 2020 (2020-06-09), pages 261 - 271 * |
吴从中;陈曦;季栋;詹曙;: "结合深度残差学习和感知损失的图像去噪", 中国图象图形学报, no. 10, 16 October 2018 (2018-10-16) * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115103118A (en) * | 2022-06-20 | 2022-09-23 | 北京航空航天大学 | High dynamic range image generation method, device, equipment and readable storage medium |
CN115103118B (en) * | 2022-06-20 | 2023-04-07 | 北京航空航天大学 | High dynamic range image generation method, device, equipment and readable storage medium |
CN118351210A (en) * | 2024-06-17 | 2024-07-16 | 南昌睿度医疗科技有限公司 | CT image artifact removal method, system, storage medium and electronic equipment |
CN118351210B (en) * | 2024-06-17 | 2024-08-27 | 南昌睿度医疗科技有限公司 | CT image artifact removal method, system, storage medium and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108734659B (en) | Sub-pixel convolution image super-resolution reconstruction method based on multi-scale label | |
Wang et al. | Esrgan: Enhanced super-resolution generative adversarial networks | |
CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
CN115049936B (en) | High-resolution remote sensing image-oriented boundary enhanced semantic segmentation method | |
CN113012172A (en) | AS-UNet-based medical image segmentation method and system | |
Chen et al. | MICU: Image super-resolution via multi-level information compensation and U-net | |
CN112669248A (en) | Hyperspectral and panchromatic image fusion method based on CNN and Laplacian pyramid | |
Qu et al. | TransFuse: A unified transformer-based image fusion framework using self-supervised learning | |
Ma et al. | SD-GAN: Saliency-discriminated GAN for remote sensing image superresolution | |
CN117408924A (en) | Low-light image enhancement method based on multiple semantic feature fusion network | |
CN117333750A (en) | Spatial registration and local global multi-scale multi-modal medical image fusion method | |
CN111222453B (en) | Remote sensing image change detection method based on dense connection and geometric structure constraint | |
CN115511767A (en) | Self-supervised learning multi-modal image fusion method and application thereof | |
CN115578262A (en) | Polarization image super-resolution reconstruction method based on AFAN model | |
CN116977208A (en) | Low-illumination image enhancement method for double-branch fusion | |
CN115293968A (en) | Super-light-weight high-efficiency single-image super-resolution method | |
CN114596378A (en) | Sparse angle CT artifact removing method | |
Lai et al. | Generative focused feedback residual networks for image steganalysis and hidden information reconstruction | |
CN117893409A (en) | Face super-resolution reconstruction method and system based on illumination condition constraint diffusion model | |
CN117593275A (en) | Medical image segmentation system | |
CN115546030B (en) | Compressed video super-resolution method and system based on twin super-resolution network | |
CN114937154B (en) | Significance detection method based on recursive decoder | |
Qin et al. | Remote sensing image super-resolution using multi-scale convolutional neural network | |
Li et al. | Super-resolution of fisheye rectified image based on deep multi-path cascaded network | |
CN111951177B (en) | Infrared image detail enhancement method based on image super-resolution loss function |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |