CN109559359A - Artifact minimizing technology based on the sparse angular data reconstruction image that deep learning is realized - Google Patents
Artifact minimizing technology based on the sparse angular data reconstruction image that deep learning is realized Download PDFInfo
- Publication number
- CN109559359A CN109559359A CN201811137448.2A CN201811137448A CN109559359A CN 109559359 A CN109559359 A CN 109559359A CN 201811137448 A CN201811137448 A CN 201811137448A CN 109559359 A CN109559359 A CN 109559359A
- Authority
- CN
- China
- Prior art keywords
- artifact
- data
- reconstruction
- layers
- sparse angular
- 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.)
- Granted
Links
- 238000013135 deep learning Methods 0.000 title claims abstract description 28
- 238000005516 engineering process Methods 0.000 title claims abstract description 15
- 238000013528 artificial neural network Methods 0.000 claims abstract description 62
- 238000012549 training Methods 0.000 claims abstract description 33
- 238000000034 method Methods 0.000 claims abstract description 30
- 238000012360 testing method Methods 0.000 claims abstract description 25
- 238000003062 neural network model Methods 0.000 claims abstract description 4
- 238000003384 imaging method Methods 0.000 claims description 55
- 238000004422 calculation algorithm Methods 0.000 claims description 16
- 230000000694 effects Effects 0.000 claims description 11
- 230000014509 gene expression Effects 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000000903 blocking effect Effects 0.000 claims description 4
- 235000013399 edible fruits Nutrition 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 3
- 210000005036 nerve Anatomy 0.000 claims description 3
- 230000001537 neural effect Effects 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 2
- 238000011160 research Methods 0.000 abstract description 5
- 238000012517 data analytics Methods 0.000 abstract description 3
- 239000010410 layer Substances 0.000 description 97
- 230000008030 elimination Effects 0.000 description 9
- 238000003379 elimination reaction Methods 0.000 description 9
- 238000002591 computed tomography Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 7
- 230000008901 benefit Effects 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 238000004800 variational method Methods 0.000 description 2
- 206010028980 Neoplasm Diseases 0.000 description 1
- 235000008331 Pinus X rigitaeda Nutrition 0.000 description 1
- 235000011613 Pinus brutia Nutrition 0.000 description 1
- 241000018646 Pinus brutia Species 0.000 description 1
- 206010052428 Wound Diseases 0.000 description 1
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 201000011510 cancer Diseases 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 208000032839 leukemia Diseases 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000002759 z-score normalization Methods 0.000 description 1
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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention provides a kind of artifact minimizing technology of sparse angular data reconstruction image realized based on deep learning, passes through acquisition sparse angular data for projection and complete Angles Projections data；It is rebuild respectively using sparse angular data for projection and complete data for projection；The neural network of artifact is gone in building；Using the image of reconstruction as training data；Trained neural network model is saved, and wherein by test image investment；The noise that obtained neural network forecast in step E is subtracted using test image, can be obtained clear image.The deep learning method having outstanding performance in computer vision at present is introduced into the artifact removal research of sparse angular projection analytic reconstruction image, the characteristics of using Inception-resnet network, the fine and various neural network of ability to express is constructed, the artifact suitable for sparse angular data analytic reconstruction image removes.
Description
Technical field
The present invention relates to a kind of artifact minimizing technologies of sparse angular data reconstruction image realized based on deep learning.
Background technique
CT scan (Computed tomography, CT) passes through development in more than 30 years, imaging speed
Degree and image quality have significant progress, and as CT scan is using more and more extensive, people are for CT scan to human body
Radiation increasingly pay attention to, medical research shows that x-ray bombardment may induce the diseases such as pathobolism, cancer, leukaemia
Disease.In conventional CT examination, dose of radiation suffered by patient has arrived very important stage.However the reduction of CT scan dosage
It is usually associated with the decline of image quality, research hotspot instantly is exactly to guarantee imaging while reducing CT scan dosage as far as possible
Quality.
Sparse angular projection acquisition mainly realizes the purpose for reducing scanning dose, but its by reducing the CT scan time
Analytic reconstruction can bring serious strip artifact, can be to scarce under the conditions of sparse sampling according to the redundancy of CT data for projection
It loses data and repair and estimate to obtain complete data for projection collection, and rebuild with this, under this thinking, Li et al. utilizes dictionary learning
The data for projection that method estimation is lost, and rebuild, achieve the result better than TV method.Zhang etc. utilizes S2etram's
Direction difference approach carries out the missing data that difference removes acquisition projection data spatial to data for projection, then carries out related reconstruction,
This method compares traditional linear difference and achieves preferable experiment effect.Based on compressed sensing (Compressed
Sensing, CS) method in, Rudin is removed for the Gaussian noise in data for projection using full variational method, in mooring
The noise of pine distribution then uses improved full variational method.Ma etc. is obtained using the local smoothing method of non-local mean as constraint
More preferable than TV result.Hu etc. proposes a kind of 2 layers of progressive iterative model, by the prior-constrained introducing three dimensional CT weight of zero norm
It builds, is obtained under conditions of height is sparse and be much better than the prior-constrained result of a norm.Furthermore iterative reconstruction algorithm is also dilute
A solution of Angles Projections is dredged, but relative to parsing class algorithm for reconstructing, excessive calculation amount limits it and answers extensively
With.
In existing all kinds of methods based on analytic reconstruction, there are still following problems:
One, the existing all kinds of methods based on analytic reconstruction thoroughly can not effectively inhibit structural large scale strip artifact,
Iterative reconstruction approach takes long time.
Two, in the neural network of the non-resnet type of tradition, the gradient explosion and the gradient disappearance problem that occur.Generally
For network it is deeper, the ability to express of network is stronger, but in practice, when the neural network number of plies is more than a fixed number
Mesh, the ability to express that just will appear network are not further added by, and are even not so good as shallower network.This is because the gradient updating of network
Rule be according to chain rule, the gradient of front layer from all layers of gradient below product, in this way after the number of plies increases,
The gradient of front layer will be made to become very small or very big, referring to 0.9^100 ≈ 0.00002,1.1^100 ≈ 13781.
In this way training when just will appear front layer and update that amplitude is too small or amplitude causes greatly very much can not finally restrain.
Three, the overall feeling open country of Inception-resnet-v2 network is still not big enough, can not capture strip artifact completely
Feature.
Four, traditional neural network has full articulamentum, will lead to network parameter huge amount, and time consumption for training is very long,
And it is easy over-fitting.
Therefore the deep learning method having outstanding performance in computer vision at present is introduced into and sparse angular is parsed
It is had a very important significance in the removal research of reconstruction image artifact.
Summary of the invention
The object of the present invention is to provide a kind of artifacts of sparse angular data reconstruction image realized based on deep learning to go
Except method solves the problems, such as that current analytic reconstruction class method existing in the prior art can not effectively inhibit large scale strip artifact.
In the present invention, Inception indicates the Inception network basic module of Google, Inception-resnet table
Show the new basic network module that Google combines residual error network and Inception network design to come out, specific structure such as Fig. 2 institute
Show.
In the present invention, FDK algorithm is 3 D pyramidal CT algorithm for reconstructing very classical in CT image reconstruction field, three wounds
The upper-case first letters of beginning name word are composed its name.
The technical solution of the invention is as follows:
A kind of artifact minimizing technology for the sparse angular data reconstruction image realized based on deep learning, is based on deep learning
Constructing one combines the sparse angular reconstruction from projections imaging artifact of full convolutional network and Inception-resnet network module to disappear
Except neural network, going artifact and noise to be predicted and remove it for sparse angular CT image specifically includes following step
Suddenly,
Several pairs of S1, acquisition data for projection, each pair of data for projection includes sparse angular data for projection and complete Angles Projections
Data；
S2, it is rebuild, will be rebuild respectively using the sparse angular data for projection and complete data for projection of step S1 acquisition
Data set patchi (i=1,2,3,4,5 ...) afterwards is saved, and each pair of data set includes sparse angular backprojection reconstruction data set
With complete backprojection reconstruction data set；
S3, building sparse angular reconstruction from projections imaging artifact eliminate neural network；
S4, using the image rebuild in step S2 as training dataset, complete backprojection reconstruction data set does contrasting data,
Neural network is eliminated to the sparse angular reconstruction from projections imaging artifact constructed in step S3 to be trained, and it is preferable to save training effect
Model, be denoted as model model；
The trained sparse angular reconstruction from projections imaging artifact that S5, obtaining step S4 are saved eliminates neural network model
Model, and wherein by test image test_img investment, it finally predicts and saves test image test_img and corresponding puppet
Shadow figure noise_img；
S6, obtained artifact figure noise_ in step S5 is subtracted using the test image test_img saved in step S5
Clear image clean_img can be obtained in img.
Further, in step S1, the acquisition angles interval of the sparse angular data for projection of acquisition is not less than 4 ° and not
Greater than 8 °, complete Angles Projections data acquisition angles interval is not more than 0.5 °, i.e. the group number of data acquisition is in less than 90
Group is greater than in the range of 45 groups.
Further, in step S2 using FDK algorithm for reconstructing to the step S1 sparse angular data for projection acquired and complete
Data for projection is rebuild respectively, specifically,
S21, institute's acquired projections data in two-dimensional projection's matrix, that is, step S1 are calculated with weighted factor and is multiplied, be modified
So that it is met Central slice theorem, obtains revised data for projection；
S22, the revised data for projection of step S21 is done and is filtered line by line using slope filtering method；
S23, image is gone out as backprojection reconstruction to the filtered data of step S22.
Further, sparse angular reconstruction from projections imaging artifact is constructed in step S3 eliminate neural network specific structure are as follows:
Convolution -> BN layers -> relu- > convolution -> BN layers -> relu- > Resblock- > Res_inception_block1- >
Res_inception_block2- > Res_inception_block3- > BN layers -> relu- > convolution；
Wherein, it is normalized for i.e. batch for BN layers, to sparse angular reconstruction from projections imaging puppet when neural metwork training
The output data that shadow eliminates neural network middle layer carries out Z-score normalized, that is, subtracts mean value except variance；Relu is sparse
Angles Projections reconstruction image artifact eliminates the activation primitive that neural network uses, and expression formula is relu (x)=max (0, x)；
Concat representing matrix splices layer, i.e., the input of multiple branches is stitched together as a whole in a certain specified dimension
It is input to next layer；What Resblock was indicated is the basic module of a residual error network；Res_inception_block1,Res_
Inception_block2, Res_inception_block 3 respectively indicates the basic module of an Inception_resnet,
The different internal structure of digital different expressions；
Further, in step S3,
The normal connection path of the Resblock: input -> convolution -> BN layers -> relu- > convolution -> sum layers, then its
A parallel link is had more, is directly connected to sum layers from input, is added with the output of normal connection path, as next layer
Input；
First connection path of the Res_inception_block1: input -> convolutional layer -> concat layers i.e. channel
Splice layer；Article 2 connection path: input -> convolutional layer -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution ->
Concat layers；Third path: input -> convolutional layer -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution -> concat
Layer；Then the input of the output and parallel link path by above-mentioned three paths after concat layers of splicing passes through sum layers
It is added, result is as next layer of input；
First connection path of the Res_inception_block2: input -> convolution -> concat layers；Article 2
Connection path: input -> convolution -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution -> concat layers；It then will be above-mentioned
Output of two paths after concat layers of splicing is added with the input in parallel link path at sum layers, result conduct
Next layer of input；
First connection path of the Res_inception_block3: input -> convolution -> concat layers；Article 2
Connection path: input -> convolution -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution；Then above-mentioned two paths are passed through
Cross concat layers splicing after output be added under sum layers with the input in parallel link path, result is as next layer
Input.
Further, step S4 specifically,
S41, in step S2 rebuild after data set patchi (i=1,2,3,4,5 ...), carry out blocking processing, then
It puts into sparse angular reconstruction from projections imaging artifact and eliminates neural network；
S42, it is eliminated by data set patchi (i=1,2,3,4,5 ...) investment sparse angular reconstruction from projections imaging artifact
Data are normalized before neural network, i.e., training sample obtained in step S2 are normalized, i.e.,Wherein mean is the mean value of patchi, and std is its standard deviation；It will
Patchi* investment sparse angular reconstruction from projections imaging artifact is eliminated neural network and is trained；And put into institute in step S3
The sparse angular reconstruction from projections imaging artifact of building is eliminated neural network and is trained, and training includes propagated forward and reversely passes
It broadcasts；The final model model for saving training and obtaining, supplying step S5 test use.
Further, in step S42, sparse angular reconstruction from projections imaging artifact is eliminated specific in neural network training process
It is to take mean square error mse as last loss function is trained to be corrected to network parameter, expression formula is as follows:
Wherein, n is the size of every a batch investment data volume Batch-size, and w, h are the length and width of each sample, D respectively
It is the effect picture clean_ subtracted after sparse angular reconstruction from projections imaging artifact elimination neural network prediction artifact and noise
Img, X are correlation datas, i.e. the sparse angular backprojection reconstruction data set rebuild in step S2 and complete Angles Projections rebuild number
According to the error image of collection；In training process, sparse angular reconstruction from projections imaging artifact eliminates neural network according to loss function
Numerical value successively calculates renewal amount by back-propagation algorithm and Adam optimization algorithm to update the parameter of network gradually to optimize mould
Type model.
Further, step S5 specifically,
S51, the model model that training obtains in step S4 is saved；
S52, test image test_img is put into progress artifact figure noise_img prediction among model model, and saved
The artifact figure noise_img of prediction.
Further, step S6 specifically,
The artifact figure noise_img and test image test_img predicted in S61, obtaining step S5.
S62, artifact figure noise_img is subtracted with test image test_img, required clear image can be obtained
clean_img。
The beneficial effects of the present invention are:
One, the artifact minimizing technology for the sparse angular data reconstruction image that this kind is realized based on deep learning, will exist at present
The artifact removal that the deep learning method having outstanding performance in computer vision is introduced into sparse angular projection analytic reconstruction image is ground
In studying carefully, the characteristics of using Inception-resnet network, the fine and various neural network of ability to express can be constructed,
Artifact suitable for sparse angular data analytic reconstruction image removes；
Two, the method for the present invention is replaced originally in the overall structure of Inception-resnet network using resnet
Simple convolutional layer in Inception-resnet, to construct the neural network of deeper, so that the overall feeling of network is wild
It is bigger, more meet the feature of strip artifact overall situation distribution.
Three, the present invention uses the mode of the full convolutional network of FCN, that is, convolutional layer is used only without the use of full articulamentum, due to
The features such as weight of convolution is shared, part connects, the parameter of network greatly reduces, and greatly reduces the training time, also improves
Anti- over-fitting ability, and designed image can be easier to the network of image by FCN, save many troubles.
Detailed description of the invention
Fig. 1 is the artifact minimizing technology for the sparse angular data reconstruction image that the embodiment of the present invention is realized based on deep learning
Flow diagram.
Fig. 2 is the structural schematic diagram of the basic module of Inception-resnet in embodiment.
Fig. 3 is that the basic structure of the elimination neural network of sparse angular reconstruction from projections imaging artifact used in embodiment is shown
It is intended to.
Fig. 4 is the structural schematic diagram of corresponding Resblock module in embodiment.
Fig. 5 is corresponding Res_inception_block1 basic structure schematic diagram in Fig. 3.
Fig. 6 is corresponding Res_inception_block2 basic structure schematic diagram in Fig. 3.
Fig. 7 is corresponding Res_inception_block3 basic structure schematic diagram in Fig. 3.
Fig. 8 is the schematic diagram of convolution mode in embodiment.
Wherein, Input table shows input, and Output indicates output；
The Inception network basic module of Inception expression Google；
Inception-resnet indicates that Google combines residual error network (resnet) and Inception network design to come out
New basic network module；
Conv 1*1 indicates that it is a convolutional layer, and convolution kernel is having a size of 1*1；Conv 1*3 indicates that it is a convolution
Layer, convolution kernel is having a size of 1*3；Conv 3*1 indicates that it is a convolutional layer, and convolution kernel is having a size of 3*1；Conv 3*3 indicates it
It is a convolutional layer, convolution kernel is having a size of 3*3；Conv 1*7 indicates that it is a convolutional layer, and convolution kernel is having a size of 1*7；Conv
7*1 indicates that it is a convolutional layer, and convolution kernel is having a size of 7*1；
Relu is amendment linear unit, indicates that sparse angular reconstruction from projections imaging artifact eliminates one kind that neural network uses
Activation primitive；
BN layers expression be batch normalization layer (Batch normalization)；
Filter concat representing matrix splices layer, i.e., the input of multiple branches exists in a certain specified dimension splicing
Next layer is input to together as an entirety；
What Resblock was indicated is the basic module of a residual error network；
Res_inception_block1, Res_inception_block 2, Res_inception_block 3 difference
Indicate the basic module of an Inception_resnet, number is different to indicate different internal structure (convolution in general
Combination it is different).
Specific embodiment
The preferred embodiment that the invention will now be described in detail with reference to the accompanying drawings.
Embodiment
A kind of artifact minimizing technology for the sparse angular data reconstruction image realized based on deep learning is sparse by acquiring
Angles Projections data and complete Angles Projections data；Weight is carried out respectively using sparse angular data for projection and complete data for projection
It builds；It constructs sparse angular reconstruction from projections imaging artifact and eliminates neural network；Using the image of reconstruction as training data；Save instruction
The model perfected, and wherein by test image investment；Obtained sparse angular in step E, which is subtracted, using test image projects weight
The noise that image artifacts eliminate neural network prediction is built, clear image can be obtained.It will be showed in computer vision at present prominent
Deep learning method out is introduced into the artifact removal research of sparse angular projection analytic reconstruction image, utilizes Inception-
The characteristics of resnet network, constructs the fine and various sparse angular reconstruction from projections imaging artifact of an ability to express and eliminates nerve
Network, the artifact suitable for sparse angular data analytic reconstruction image remove.
A kind of artifact minimizing technology of sparse angular data reconstruction image realized based on deep learning of embodiment, is based on
Deep learning constructs one and the sparse angular reconstruction from projections imaging artifact of full convolutional network and Inception-resnet is combined to disappear
Except neural network, goes artifact and noise to be predicted for sparse angular CT image and remove it；Such as Fig. 1, specifically include
Following steps:
Five pairs of S1, acquisition data for projection Ai (i=1,2,3,4,5), per internally comprising one group of sparse angular data for projection Ai_
Low and one group of complete Angles Projections data Ai_high.
The acquisition angles interval of the sparse angular data for projection acquired in step S1 is not less than 4 ° and no more than 8 °, complete
Angles Projections data acquisition angles interval is not more than 0.5 °, i.e. the group number of data acquisition is to be in be greater than 45 groups less than 90 groups
In range.
S2, it is rebuild using the five couples of data for projection Ai acquired in step S1, the data set after reconstruction is saved as five
To S2i (i=1,2,3,4,5), each pair of data set is divided into sparse angular backprojection reconstruction data set S2i_low and complete Angles Projections
Data set for reconstruction S2i_high, single layer image size is 512*512 in each pair of data set.
What reconstruction utilized in step S2 is FDK algorithm classical in analytic reconstruction algorithm, specifically,
S21, it calculates two-dimensional projection's matrix weighted factor and is multiplied, being modified, which makes it substantially meet center slice, determines
Reason.
S22, it revised data for projection is done filters line by line, wherein using slope filtering mode.
S23, image is gone out as backprojection reconstruction to filtered data.
S3, building sparse angular reconstruction from projections imaging artifact eliminate neural network.
S31, constructed sparse angular reconstruction from projections imaging artifact elimination neural network overall structure specific structure are for example attached
Fig. 3
Convolution -> BN layers (batch-normalization) layer -> relu- > convolution -> BN layers -> relu- > Resblock- >
Res_inception_block1- > Res_inception_block2- > Res_inception_block3- > BN layers -> relu-
> convolution.
Wherein, it is normalized for i.e. batch for BN layers, to sparse angular reconstruction from projections imaging puppet when neural metwork training
The output data that shadow eliminates neural network middle layer carries out Z-score normalized, that is, subtracts mean value except variance；Relu is sparse
Angles Projections reconstruction image artifact eliminates the activation primitive that neural network uses, and expression formula is relu (x)=max (0, x)；
Concat representing matrix splices layer, i.e., the input of multiple branches is stitched together as a whole in a certain specified dimension
It is input to next layer；What Resblock was indicated is the basic module of a residual error network；Res_inception_block1,Res_
Inception_block 2, Res_inceptio n_block 3 respectively indicate the basic mould of an Inception_resnet
Block, number is different to indicate different internal structures；
Resblock described in S311, step S31 is as shown in Fig. 4, normal connection path: input -> convolution -> BN layers ->
Relu- > convolution -> sum layers, then it has more a parallel link compared to common convolutional network, is directly connected to from input
It sum layers, is added with the output of normal connection path, as next layer of input.
Res_inception_block1 specific structure described in S312, step S31 is as shown in figure 5, first link road
Diameter: input -> convolutional layer -> concat layers of Filter (matrix splicing layer).Article 2 connection path: input -> convolutional layer -> BN
Layer -> relu- > convolution -> BN layers -> relu- > convolution -> concat layers.Third path: input -> convolutional layer -> BN layers ->
Relu- > convolution -> BN layers -> relu- > convolution -> concat layers.Then by above-mentioned three paths after concat layers of splicing
Output and the input in parallel link path be added by sum layers, result is as next layer of input.
Res_inception_block2 specific structure described in S313, step S31 is as shown in fig. 6, first link road
Diameter: input -> convolution -> concat layers.Article 2 connection path: input -> convolution -> BN layers -> relu- > convolution -> BN layers ->
Relu- > convolution -> concat layers.Then the output by above-mentioned two paths after the splicing of concat floor and parallel link road
The input of diameter is added at sum layers, and result is as next layer of input.
Res_inception_block3 specific structure described in S314, step S31 is as shown in fig. 7, first link road
Diameter: input -> convolution -> concat layers.Article 2 connection path: input -> convolution -> S2N- > relu- > convolution -> S2N- >
Relu- > convolution.Then the input of the output and parallel link path by above-mentioned two paths after concat layers of splicing exists
Sum layers lower to be added, and result is as next layer of input.
Sparse angular reconstruction from projections imaging artifact eliminates the convolution mode of neural network using boundary extential form convolution,
So that each layer of input is identical as Output Size.
S4, using the data set Bi_low (i=1,2,3,4,5) saved in step S2 as sparse angular backprojection reconstruction figure
As the training dataset of artifact elimination neural network, data set Bi_high (i=1,2,3,4,5) does contrasting data, to step S3
The sparse angular reconstruction from projections imaging artifact of middle building is eliminated neural network and is trained, and the preferable model of training effect is saved,
It is denoted as model.
S41, blocking processing is carried out to the data obtained Bi in step S2 (i=1,2,3,4,5), blocking will every figure
Small block data collection patchi (i=1,2,3,4,5) as being divided into several 128*128 puts into sparse angular reconstruction from projections imaging puppet again
Shadow eliminates neural network, to learn.
S42, it needs to carry out at Z-score normalization data before by data patchi (i=1,2,3,4,5) investment
Reason, i.e.,Wherein mean is the mean value of patchi, and std marks for it
It is quasi- poor；Patchi* investment sparse angular reconstruction from projections imaging artifact is eliminated neural network to be trained, and saves training effect
Preferable model is denoted as model.
In step S42, it is specifically to take in neural network training process that sparse angular reconstruction from projections imaging artifact, which is eliminated,
Square error mse is corrected network parameter as last loss function is trained, and expression formula is as follows:
Wherein, n is the size of every a batch investment data volume Batch-size, and w, h are the length and width of each sample, D respectively
It is the effect picture clean_ subtracted after sparse angular reconstruction from projections imaging artifact elimination neural network prediction artifact and noise
Img, X are correlation datas, i.e. the sparse angular backprojection reconstruction data set rebuild in step S2 and complete Angles Projections rebuild number
According to the error image of collection；In training process, sparse angular reconstruction from projections imaging artifact eliminates neural network according to loss function
Numerical value successively calculates renewal amount by back-propagation algorithm and Adam optimization algorithm to update the parameter of network gradually to optimize mould
Type model.
Trained model in S5, obtaining step S4 will need to carry out the picture of artifact removal processing as test image
Test_img puts into model, finally predicts and save the test_img and corresponding artifact figure noise_img.
The model model saved in first obtaining step S42 is comprised the concrete steps that in step S5, will be carried out artifact and is handled
Picture test_img investment model among carry out the prediction of artifact figure, and save the artifact figure noise_img of prediction.
S6, artifact figure noise_img is subtracted using the test image test_img saved in step E, it is clear to can be obtained
Image clean_img.
S61, the noise_img and its test_img used predicted in the step E is obtained.
S62, its corresponding artifact figure noise_img is subtracted with test chart test_img, required clear effect can be obtained
Fruit schemes clean_img.
It has been used in the elimination neural network of sparse angular reconstruction from projections imaging artifact constructed by embodiment many advanced
The thought of neural network model, such as resnet, Inception-resnet, FCN etc..Specific structure is as matched shown in Fig. 2；
Sparse angular reconstruction from projections imaging artifact is eliminated neural network and is inspired by the thought of resnet in embodiment, uses
The skill of " short connection " alleviates gradient disperse, the gradient explosion etc. that conventional depth network occurs as shown in the right one side of something of Fig. 2
Situation.Gradient disperse refers to, conventional network structure is stacked on after the network number of plies increases to a certain degree, sparse angular projection
The performance that reconstruction image artifact eliminates neural network will not only improve, but also may be worse.Gradient disperse is depth nerve
How a relatively conventional phenomenon in network is deepening network, while improving the ability to express of network, is evading gradient disperse
Phenomenon is a hot spot in deep learning.Two non-conterminous hidden layers are directly connected with each other by parallel link, in this way in net
In the case that network is very deep, even if going wrong in certain layers, but due to the presence of parallel link, so that the hidden layer knot of front
Fruit can be input to below by short connection cross-layer, can't excessively influence whole effect.In Inception-resnet base
In this module, this advantage of resnet is absorbed.
The characteristics of Inception is the parallel of a variety of convolution kernel combinations, that is, compared with " width ".Previous common net
Network can only use a number of convolution kernel of like combinations mode when extracting different characteristic, and Inception structure is
The extraction to different characteristic is carried out with several convolution kernels of several various combination modes, this structure schemes sparse angular CT
Structural artifact, which differs larger feature with other noise characteristics and has, as in preferably agrees with.
Inception-resnet basic module combines the above resnet and Inception, as shown in Fig. 2, can be with
The advantages of playing the two.It is eliminated in neural network structure in whole sparse angular reconstruction from projections imaging artifact, use is multiple
Inception-resnet block combiner can construct deeper network and whole ability to express is promoted very much, and by
It is more efficient for the mode of the extraction of different characteristic in the presence of Inception, and the problem that embodiment is studied, it is sparse
Angle artifact mixes with noise, that is, the feature senior middle school low level to be predicted has, and agrees with very much with Inception-resnet base
The network of this module building.As shown in table 1, it is trained and tests using several different networks, be followed successively by from top to bottom
Simple residual error network, sparse angular reconstruction from projections imaging artifact eliminates mind in the convolutional network stacked merely and embodiment
Through network, 9 layers of number of plies for referring to basic block that it is used, such as Resblock, Inception block.Using psnr and
Ssim does evaluation index.
Sparse angular reconstruction from projections imaging artifact, which is eliminated, to be opened in neural network overall structure by the full convolutional neural networks of FCN
The characteristics of hair, benefit are exactly to replace connecting entirely with convolution, and part connects, parameter sharing, so that the parameter of training subtracts significantly
It is few, reduce the required hardware requirement of experiment and a possibility that certain levels reduce over-fitting.Full convolutional network is because take
Disappeared full connection, thus its for input size also without rigid requirement, more suitable for field of image processing.
Sparse angular reconstruction from projections imaging artifact eliminates the convolution mode of neural network using boundary extential form convolution,
As shown in Figure 8.So that each layer of input is identical as Output Size.Utilize sparse angular data reconstruction image and complete angle number
Training data is done according to reconstruction image, enables sparse angular reconstruction from projections imaging artifact to eliminate neural network accurately pre-
Measure the artifact and other noises in sparse angular image.
Deep learning goes the whole concept of artifact that can be divided into two kinds, one is directly from the image with artifact and noise it is straight
It connects and is mapped to original image, another kind is to be mapped to noise therein and artifact from the image with noise and artifact, then again
The noise predicted is subtracted using the figure with noise and artifact and artifact obtains the raw information of image.And it is ground according to correlation
Study carefully, goes the mode of artifact for second, network topology structure is than the first simpler namely artifact noise than normal tissue knot
Structure is easier to be learned to, so this method uses the latter；
Sparse angular reconstruction from projections imaging artifact elimination neural network takes mse as last loss letter in embodiment
Number, expression formula are as follows:
Wherein, n is the size of every a batch investment data volume Batch-size, and w, h are the length and width of each sample, D respectively
It is the effect picture subtracted after sparse angular reconstruction from projections imaging artifact elimination neural network prediction artifact and noise, X is pair
Than figure, i.e., corresponding complete angular image.In training process, sparse angular reconstruction from projections imaging artifact eliminates neural network root
According to the numerical value of loss function, successively calculated according to back-propagation algorithm (Back Propagation, BP) and Adam optimization algorithm
Renewal amount updates weight and biasing；
Sparse angular reconstruction from projections imaging artifact eliminates each parameter setting of neural network: the number of iterations is 150 times, and batch is big
Small is 48, initial learning rate be 0.001, every 30 iterative attenuations be itself 60 percent, loss function use mse, optimization
Device creates sparse angular reconstruction from projections imaging artifact elimination neural network for Adam with this and is trained and tests.Final 90 angle
The test psnr of data for projection reconstruction image reaches 36.5915, ssim and reaches 0.8818, as shown in table 1.
1 embodiment of table and the training of different networks and test result
Explanation is made that the specific implementation of this method above, certain this method can also there are other a variety of specific embodiment parties
Formula, those skilled in the art can make various changes and deformation, but this on the premise of without prejudice to spirit of the invention
A little change should be contained in the required protection scope limited of the application patent with deformation.
Claims (9)
1. a kind of artifact minimizing technology for the sparse angular data reconstruction image realized based on deep learning, it is characterised in that: base
The sparse angular backprojection reconstruction for combining full convolutional network and Inception-resnet network module is constructed in deep learning
Image artifacts eliminate neural network, go artifact and noise to be predicted for sparse angular CT image and remove it, specifically
Include the following steps,
Several pairs of S1, acquisition data for projection, each pair of data for projection includes sparse angular data for projection and complete Angles Projections data；
S2, it is rebuild respectively using the sparse angular data for projection and complete data for projection of step S1 acquisition, after reconstruction
Data set patchi (i=1,2,3,4,5 ...) is saved, and each pair of data set includes sparse angular backprojection reconstruction data set and complete
Standby backprojection reconstruction data set；
S3, building sparse angular reconstruction from projections imaging artifact eliminate neural network；
S4, using the image rebuild in step S2 as training dataset, complete backprojection reconstruction data set does contrasting data, to step
The sparse angular reconstruction from projections imaging artifact constructed in rapid S3 is eliminated neural network and is trained, and the preferable mould of training effect is saved
Type is denoted as model model；
The trained sparse angular reconstruction from projections imaging artifact that S5, obtaining step S4 are saved eliminates neural network model
Model, and wherein by test image test_img investment, it finally predicts and saves test image test_img and corresponding puppet
Shadow figure noise_img；
S6, obtained artifact figure noise_img in step S5 is subtracted using the test image test_img saved in step S5,
Clear image clean_img can be obtained.
2. the artifact minimizing technology for the sparse angular data reconstruction image realized as described in claim 1 based on deep learning,
It is characterized by: the acquisition angles interval of the sparse angular data for projection of acquisition is not less than 4 ° and no more than 8 ° in step S1,
Complete Angles Projections data acquisition angles interval is not more than 0.5 °, i.e. the group number of data acquisition is to be in be greater than 45 less than 90 groups
In the range of group.
3. the artifact minimizing technology for the sparse angular data reconstruction image realized as described in claim 1 based on deep learning,
It is characterized by: using FDK algorithm for reconstructing to the step S1 sparse angular data for projection acquired and complete projection number in step S2
According to being rebuild respectively, specifically,
S21, institute's acquired projections data in two-dimensional projection's matrix, that is, step S1 are calculated with weighted factor and is multiplied, being modified makes it
Meet Central slice theorem, obtains revised data for projection；
S22, the revised data for projection of step S21 is done and is filtered line by line using slope filtering method；
S23, image is gone out as backprojection reconstruction to the filtered data of step S22.
4. the artifact minimizing technology for the sparse angular data reconstruction image realized as described in claim 1 based on deep learning,
It is characterized by: constructing sparse angular reconstruction from projections imaging artifact in step S3 eliminates neural network specific structure are as follows:
Convolution -> BN layers -> relu- > convolution -> BN layers -> relu- > Resblock- > Res_inception_block1- > Res_
Inception_block2- > Res_inception_block3- > BN layers -> relu- > convolution；
Wherein, it normalizes for i.e. batch for BN layers, disappears when neural metwork training to sparse angular reconstruction from projections imaging artifact
Except the output data of neural network middle layer carries out Z-score normalized, that is, subtract mean value except variance；Relu is sparse angular
Reconstruction from projections imaging's artifact eliminates the activation primitive that neural network uses, and expression formula is relu (x)=max (0, x)；concat
Representing matrix splices layer, i.e., the input of multiple branches is stitched together in a certain specified dimension and is input to as a whole
Next layer；What Resblock was indicated is the basic module of a residual error network；Res_inception_block1,Res_
Inception_block2, Res_inception_block3 respectively indicate the basic module of an Inception_resnet,
The different internal structure of digital different expressions.
5. the artifact minimizing technology for the sparse angular data reconstruction image realized as claimed in claim 4 based on deep learning,
It is characterized by: in step S3,
The normal connection path of the Resblock: input -> convolution -> BN layers -> relu- > convolution -> sum layers, then it has more
One parallel link, is directly connected to sum layers from input, is added with the output of normal connection path, as next layer of input；
First connection path of the Res_inception_block1: input -> convolutional layer -> concat layers of i.e. channel splicing
Layer；Article 2 connection path: input -> convolutional layer -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution -> concat layers；
Third path: input -> convolutional layer -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution -> concat layers；Then will
Output of above-mentioned three paths after concat layers of splicing is added with the input in parallel link path by sum layers, is tied
Fruit is as next layer of input；
First connection path of the Res_inception_block2: input -> convolution -> concat layers；Article 2 connection
Path: input -> convolution -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution -> concat layers；Then by above-mentioned two
Output of the path after concat layers of splicing is added with the input in parallel link path at sum layers, and result is as next
The input of layer；
First connection path of the Res_inception_block3: input -> convolution -> concat layers；Article 2 connection
Path: input -> convolution -> BN layers -> relu- > convolution -> BN layers -> relu- > convolution；Then above-mentioned two paths are passed through
Output after concat layers of splicing is added under sum layers with the input in parallel link path, and result is defeated as next layer
Enter.
6. the artifact of the sparse angular data reconstruction image as described in any one in claim 1-5 realized based on deep learning is gone
Except method, it is characterised in that: step S4 specifically,
S41, to the data set patchi (i=1,2,3,4,5 ...) after being rebuild in step S2, carry out blocking processing, then put into
Sparse angular reconstruction from projections imaging artifact eliminates neural network；
S42, data set patchi (i=1,2,3,4,5 ...) investment sparse angular reconstruction from projections imaging artifact is being eliminated into nerve
Data are normalized before network, i.e., training sample obtained in step S2 are normalized, i.e.,Wherein mean is the mean value of patchi, and std is its standard deviation；It will
Patchi* investment sparse angular reconstruction from projections imaging artifact is eliminated neural network and is trained；And put into institute in step S3
The sparse angular reconstruction from projections imaging artifact of building is eliminated neural network and is trained, and training includes propagated forward and reversely passes
It broadcasts；The final model model for saving training and obtaining, supplying step S5 test use.
7. the artifact minimizing technology for the sparse angular data reconstruction image realized as claimed in claim 6 based on deep learning,
It is characterized by: it is specifically to take in neural network training process that sparse angular reconstruction from projections imaging artifact, which is eliminated, in step S42
Mean square error mse is corrected network parameter as last loss function is trained, and expression formula is as follows:
Wherein, n is the size of every a batch investment data volume Batch-size, and w, h are the length and width of each sample respectively, and D is to subtract
Sparse angular reconstruction from projections imaging artifact is gone to eliminate effect picture clean_img, X after neural network prediction artifact and noise
It is correlation data, i.e. the sparse angular backprojection reconstruction data set rebuild in step S2 and complete Angles Projections data set for reconstruction
Error image；In training process, sparse angular reconstruction from projections imaging artifact eliminates neural network according to the numerical value of loss function, borrows
Back-propagation algorithm and Adam optimization algorithm is helped to calculate renewal amount successively to update the parameter of network gradually with Optimized model
model。
8. the artifact of the sparse angular data reconstruction image as described in any one in claim 1-5 realized based on deep learning is gone
Except method, it is characterised in that: step S5 specifically,
S51, the model model that training obtains in step S4 is saved；
S52, test image test_img is put into progress artifact figure noise_img prediction among model model, and saves prediction
Artifact figure noise_img.
9. the artifact of the sparse angular data reconstruction image as described in any one in claim 1-5 realized based on deep learning is gone
Except method, it is characterised in that: step S6 specifically,
The artifact figure noise_img and test image test_img predicted in S61, obtaining step S5.
S62, artifact figure noise_img is subtracted with test image test_img, required clear image clean_ can be obtained
img。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811137448.2A CN109559359B (en) | 2018-09-27 | 2018-09-27 | Artifact removal method for reconstructing image by sparse angle data based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811137448.2A CN109559359B (en) | 2018-09-27 | 2018-09-27 | Artifact removal method for reconstructing image by sparse angle data based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109559359A true CN109559359A (en) | 2019-04-02 |
| CN109559359B CN109559359B (en) | 2023-06-02 |
Family
ID=65864849
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811137448.2A Active CN109559359B (en) | 2018-09-27 | 2018-09-27 | Artifact removal method for reconstructing image by sparse angle data based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109559359B (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110211194A (en) * | 2019-05-21 | 2019-09-06 | 武汉理工大学 | A method of sparse angular CT imaging artefacts are removed based on deep learning |
| CN110477951A (en) * | 2019-08-30 | 2019-11-22 | 浙江大学 | Ultrafast composite plane wave imaging method based on broadband acoustic metamaterial |
| CN110503699A (en) * | 2019-07-01 | 2019-11-26 | 天津大学 | A kind of CT projection path reduce in the case of CT image rebuilding method |
| CN110751701A (en) * | 2019-10-18 | 2020-02-04 | 北京航空航天大学 | X-ray absorption contrast computed tomography incomplete data reconstruction method based on deep learning |
| CN110849462A (en) * | 2019-12-05 | 2020-02-28 | 武汉科技大学 | Tandem mill vibration signal separation method based on sparse feature similarity |
| CN111009019A (en) * | 2019-09-27 | 2020-04-14 | 北京航空航天大学 | Differential phase contrast CT incomplete data reconstruction method based on deep learning |
| CN111028160A (en) * | 2019-11-21 | 2020-04-17 | 西北工业大学 | Remote sensing image noise suppression method based on convolutional neural network |
| CN111223161A (en) * | 2020-01-02 | 2020-06-02 | 京东数字科技控股有限公司 | Image reconstruction method and device and storage medium |
| CN112215910A (en) * | 2019-07-12 | 2021-01-12 | 佳能医疗系统株式会社 | X-ray system, image processing device, and program |
| CN112348936A (en) * | 2020-11-30 | 2021-02-09 | 华中科技大学 | Low-dose cone-beam CT image reconstruction method based on deep learning |
| CN112381904A (en) * | 2020-11-26 | 2021-02-19 | 南京医科大学 | Limited angle CT image reconstruction method based on DTw-SART-TV iterative process |
| CN113052936A (en) * | 2021-03-30 | 2021-06-29 | 大连理工大学 | Single-view CT reconstruction method integrating FDK and deep learning |
| CN113256497A (en) * | 2021-06-21 | 2021-08-13 | 中南大学 | Image reconstruction method and system |
| CN113505536A (en) * | 2021-07-09 | 2021-10-15 | 兰州理工大学 | Optimized traffic flow prediction model based on space-time diagram convolution network |
| CN114140442A (en) * | 2021-12-01 | 2022-03-04 | 北京邮电大学 | Deep learning sparse angle CT reconstruction method based on frequency domain and image domain degradation perception |
| WO2022110530A1 (en) * | 2020-11-24 | 2022-06-02 | 佛山读图科技有限公司 | Tomographic image reconstruction method based on spect data sampling and noise characteristics |
| CN114723842A (en) * | 2022-05-24 | 2022-07-08 | 之江实验室 | Sparse visual angle CT imaging method and device based on depth fusion neural network |
| CN115049753A (en) * | 2022-05-13 | 2022-09-13 | 沈阳铸造研究所有限公司 | Cone beam CT artifact correction method based on unsupervised deep learning |
| CN118300729A (en) * | 2024-06-05 | 2024-07-05 | 中国海洋大学 | Sound field rapid forecasting method based on depth bilateral sparse sampling |
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 |
| CN108122265A (en) * | 2017-11-13 | 2018-06-05 | 深圳先进技术研究院 | A kind of CT reconstruction images optimization method and system |
| US20180240219A1 (en) * | 2017-02-22 | 2018-08-23 | Siemens Healthcare Gmbh | Denoising medical images by learning sparse image representations with a deep unfolding approach |
-
2018
- 2018-09-27 CN CN201811137448.2A patent/CN109559359B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180240219A1 (en) * | 2017-02-22 | 2018-08-23 | Siemens Healthcare Gmbh | Denoising medical images by learning sparse image representations with a deep unfolding approach |
| 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 |
| CN108122265A (en) * | 2017-11-13 | 2018-06-05 | 深圳先进技术研究院 | A kind of CT reconstruction images optimization method and system |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110211194A (en) * | 2019-05-21 | 2019-09-06 | 武汉理工大学 | A method of sparse angular CT imaging artefacts are removed based on deep learning |
| CN110503699A (en) * | 2019-07-01 | 2019-11-26 | 天津大学 | A kind of CT projection path reduce in the case of CT image rebuilding method |
| CN112215910A (en) * | 2019-07-12 | 2021-01-12 | 佳能医疗系统株式会社 | X-ray system, image processing device, and program |
| CN110477951A (en) * | 2019-08-30 | 2019-11-22 | 浙江大学 | Ultrafast composite plane wave imaging method based on broadband acoustic metamaterial |
| US20210064993A1 (en) * | 2019-08-30 | 2021-03-04 | Zhejiang University | Method For Ultrafast Compound Plane Wave Imaging Based On Broadband Acoustic Metamaterial |
| US11829872B2 (en) * | 2019-08-30 | 2023-11-28 | Zhejiang University | Method for ultrafast compound plane wave imaging based on broadband acoustic metamaterial |
| CN111009019B (en) * | 2019-09-27 | 2021-07-16 | 北京航空航天大学 | Differential phase contrast CT incomplete data reconstruction method based on deep learning |
| CN111009019A (en) * | 2019-09-27 | 2020-04-14 | 北京航空航天大学 | Differential phase contrast CT incomplete data reconstruction method based on deep learning |
| CN110751701A (en) * | 2019-10-18 | 2020-02-04 | 北京航空航天大学 | X-ray absorption contrast computed tomography incomplete data reconstruction method based on deep learning |
| CN110751701B (en) * | 2019-10-18 | 2021-03-30 | 北京航空航天大学 | X-ray absorption contrast computed tomography incomplete data reconstruction method based on deep learning |
| CN111028160A (en) * | 2019-11-21 | 2020-04-17 | 西北工业大学 | Remote sensing image noise suppression method based on convolutional neural network |
| CN110849462B (en) * | 2019-12-05 | 2021-07-27 | 武汉科技大学 | Tandem mill vibration signal separation method based on sparse feature similarity |
| CN110849462A (en) * | 2019-12-05 | 2020-02-28 | 武汉科技大学 | Tandem mill vibration signal separation method based on sparse feature similarity |
| CN111223161A (en) * | 2020-01-02 | 2020-06-02 | 京东数字科技控股有限公司 | Image reconstruction method and device and storage medium |
| CN111223161B (en) * | 2020-01-02 | 2024-04-12 | 京东科技控股股份有限公司 | Image reconstruction method, device and storage medium |
| WO2022110530A1 (en) * | 2020-11-24 | 2022-06-02 | 佛山读图科技有限公司 | Tomographic image reconstruction method based on spect data sampling and noise characteristics |
| CN112381904A (en) * | 2020-11-26 | 2021-02-19 | 南京医科大学 | Limited angle CT image reconstruction method based on DTw-SART-TV iterative process |
| CN112381904B (en) * | 2020-11-26 | 2024-05-24 | 南京医科大学 | Limited angle CT image reconstruction method based on DTw-SART-TV iterative process |
| CN112348936B (en) * | 2020-11-30 | 2023-04-18 | 华中科技大学 | Low-dose cone-beam CT image reconstruction method based on deep learning |
| CN112348936A (en) * | 2020-11-30 | 2021-02-09 | 华中科技大学 | Low-dose cone-beam CT image reconstruction method based on deep learning |
| CN113052936A (en) * | 2021-03-30 | 2021-06-29 | 大连理工大学 | Single-view CT reconstruction method integrating FDK and deep learning |
| CN113256497A (en) * | 2021-06-21 | 2021-08-13 | 中南大学 | Image reconstruction method and system |
| CN113256497B (en) * | 2021-06-21 | 2021-09-24 | 中南大学 | Image reconstruction method and system |
| CN113505536A (en) * | 2021-07-09 | 2021-10-15 | 兰州理工大学 | Optimized traffic flow prediction model based on space-time diagram convolution network |
| CN114140442A (en) * | 2021-12-01 | 2022-03-04 | 北京邮电大学 | Deep learning sparse angle CT reconstruction method based on frequency domain and image domain degradation perception |
| CN115049753A (en) * | 2022-05-13 | 2022-09-13 | 沈阳铸造研究所有限公司 | Cone beam CT artifact correction method based on unsupervised deep learning |
| CN115049753B (en) * | 2022-05-13 | 2024-05-10 | 沈阳铸造研究所有限公司 | Cone beam CT artifact correction method based on unsupervised deep learning |
| CN114723842B (en) * | 2022-05-24 | 2022-08-23 | 之江实验室 | Sparse visual angle CT imaging method and device based on depth fusion neural network |
| CN114723842A (en) * | 2022-05-24 | 2022-07-08 | 之江实验室 | Sparse visual angle CT imaging method and device based on depth fusion neural network |
| CN118300729A (en) * | 2024-06-05 | 2024-07-05 | 中国海洋大学 | Sound field rapid forecasting method based on depth bilateral sparse sampling |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109559359B (en) | 2023-06-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109559359A (en) | Artifact minimizing technology based on the sparse angular data reconstruction image that deep learning is realized | |
| WO2021159948A1 (en) | Deep learning-based low dose pet 3d reconstruction method | |
| Kang et al. | Deep convolutional framelet denosing for low-dose CT via wavelet residual network | |
| US11308587B2 (en) | Learning method of generative adversarial network with multiple generators for image denoising | |
| CN109166161B (en) | Low-dose CT image processing system based on noise artifact suppression convolutional neural network | |
| CN108961237B (en) | Low-dose CT image decomposition method based on convolutional neural network | |
| CN111429379B (en) | Low-dose CT image denoising method and system based on self-supervision learning | |
| CN108492269A (en) | Low-dose CT image de-noising method based on gradient canonical convolutional neural networks | |
| CN110223255B (en) | Low-dose CT image denoising and recursion method based on residual error coding and decoding network | |
| CN109949235A (en) | A kind of chest x-ray piece denoising method based on depth convolutional neural networks | |
| CN110930318A (en) | Low-dose CT image repairing and denoising method | |
| CN110517198A (en) | High frequency sensitivity GAN network for LDCT image denoising | |
| Song et al. | Denoising of MR and CT images using cascaded multi-supervision convolutional neural networks with progressive training | |
| CN116645283A (en) | Low-dose CT image denoising method based on self-supervision perceptual loss multi-scale convolutional neural network | |
| Du et al. | X-ray CT image denoising with MINF: A modularized iterative network framework for data from multiple dose levels | |
| Li et al. | An adaptive self-guided wavelet convolutional neural network with compound loss for low-dose CT denoising | |
| CN116630738A (en) | Energy spectrum CT imaging method based on depth convolution sparse representation reconstruction network | |
| CN108596995A (en) | A kind of PET-MRI maximum a posteriori joint method for reconstructing | |
| Khaleghi et al. | Metal artifact reduction in computed tomography images based on developed generative adversarial neural network | |
| Xie et al. | Contextual loss based artifact removal method on CBCT image | |
| Wang et al. | Texture‐aware dual domain mapping model for low‐dose CT reconstruction | |
| CN112258438B (en) | LDCT image recovery method based on unpaired data | |
| Liang et al. | A self-supervised deep learning network for low-dose CT reconstruction | |
| CN116245969A (en) | Low-dose PET image reconstruction method based on deep neural network | |
| CN116894783A (en) | Metal artifact removal method for countermeasure generation network model based on time-varying constraint |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |