CN114004753A - GAN-based low-dose CT (computed tomography) dual-domain combined noise reduction method fusing NDCT noise - Google Patents

GAN-based low-dose CT (computed tomography) dual-domain combined noise reduction method fusing NDCT noise Download PDF

Info

Publication number
CN114004753A
CN114004753A CN202111043544.2A CN202111043544A CN114004753A CN 114004753 A CN114004753 A CN 114004753A CN 202111043544 A CN202111043544 A CN 202111043544A CN 114004753 A CN114004753 A CN 114004753A
Authority
CN
China
Prior art keywords
image
noise
noise reduction
network
ndct
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
Application number
CN202111043544.2A
Other languages
Chinese (zh)
Other versions
CN114004753B (en
Inventor
蹇木伟
靳悦
王芮
王星
陈吉
傅德谦
王振海
张问银
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Linyi University
Shandong University of Finance and Economics
Original Assignee
Linyi University
Shandong University of Finance and Economics
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Linyi University, Shandong University of Finance and Economics filed Critical Linyi University
Priority to CN202111043544.2A priority Critical patent/CN114004753B/en
Publication of CN114004753A publication Critical patent/CN114004753A/en
Application granted granted Critical
Publication of CN114004753B publication Critical patent/CN114004753B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • 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]
    • 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning

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 Processing (AREA)

Abstract

The invention provides a low-dose CT (computed tomography) cyclic multi-time noise reduction method fusing NDCT noise based on a GAN (Gate network), which comprises the following steps of S1: constructing low dose CT images
Figure 100004_DEST_PATH_IMAGE002
And normal dose CT image
Figure 100004_DEST_PATH_IMAGE004
Sample set of
Figure 100004_DEST_PATH_IMAGE006
(ii) a S2: training a network for predicting CT image noise; s3, S4: carrying out primary noise reduction on the LDCT image; s5: removing noise in the NDCT image; s6: training CT images in a time domain and a frequency domain respectively through two generation countermeasure networks (WGANs); s7: optimizing objective functions of two WGANsAnd a better noise reduction effect is obtained. According to the technical scheme, the method mainly comprises two parts, namely, the LDCT is subjected to coarse granularity processing to reduce the burden of the next step; secondly, by utilizing the characteristic that the frequency domain change can separate the detail information and the structure information of the image, the LDCT is further processed with fine granularity by using the dual-domain combined noise reduction, so that the noise reduction effect is improved, and the defects of edge blurring and partial detail loss caused by the noise reduction process are overcome.

Description

GAN-based low-dose CT (computed tomography) dual-domain combined noise reduction method fusing NDCT noise
Technical Field
The invention relates to the technical field of medical image processing, in particular to a low-dose CT (computed tomography) dual-domain combined noise reduction method based on GAN (generic inverse transform) fusion NDCT noise.
Background
The Computed Tomography (CT) technique is widely used in the medical field, the imaging definition of the method is crucial to the accuracy of clinical diagnosis results, the higher the radiation dose is, the higher the definition of the obtained imaging results is, however, the higher the radiation dose can induce canceration, which has certain damage to human body, in order to avoid or reduce the secondary damage to the patient during examination, the low dose CT (ldct) is considered, however, the CT dose is reduced, and simultaneously, a new problem is generated, namely, the low dose CT has more noise than the ordinary dose CT (ndct), which further affects the result judgment. There is therefore a great deal of interest in using low doses while reducing noise in low dose CT imaging results.
Excellent methods for low-dose CT noise reduction have been proposed in succession, and noise reduction methods are mainly classified into three categories: 1) sinusoidal filtering before reconstruction: 2) performing iterative reconstruction; 3) after the reconstructed image is processed, the methods improve the quality of the image to a certain extent, but have some defects, such as edge blurring, artifact introduction, excessive smoothing and the like. With the development of deep learning, people begin to apply a network model to the noise reduction of CT images. Jelmer M. Woltrink et al propose a method based on generation of a countermeasure network, wherein a generator performs noise reduction on LDCT, and an image close to NDCT can be generated by the generator after training, so that the noise reduction effect is obtained. Peltier et al propose a multi-generator countering network noise reduction model for low-dose CT images, fitting different noise distributions with three different generators, respectively, and finally optimizing by a discriminator. Both of these methods have two disadvantages: 1) only the noise in LDCT is considered, and the noise also existing in NDCT is not considered; 2) partial loss of edge information and detail blurring can be caused after image noise reduction, so that the quality of a CT image is reduced, and later clinical use is not facilitated. These disadvantages limit the effectiveness of the process. The frequency domain change can separate the detail information and the structure information of the image, so that the frequency domain image can better represent the edge information and the detail information of the original image.
Disclosure of Invention
In summary, the noise of NDCT is extracted and removed, and then the NDCT participates in the subsequent GAN training, and the GAN training in the time domain and the frequency domain can show better performance. The Coarse-to-fine denoising method is used in the whole system, Coarse-grained processing is performed firstly to obtain a preliminary denoising effect, burden is relieved for further denoising, fine-grained processing is performed afterwards, and the denoising effect is improved. In order to make up the defects of the prior art, the invention provides a low-dose CT double-domain combined noise reduction method based on GAN fusion NDCT noise.
The invention is realized by the following technical scheme: a low-dose CT double-domain combined noise reduction method based on GAN fusion NDCT noise comprises coarse grain processing and fine grain processing, and is characterized by comprising the following specific steps:
first, coarse size treatment
S1: constructed forSample set for training CT image noise reduction network
Figure 244259DEST_PATH_IMAGE001
Wherein
Figure DEST_PATH_IMAGE002
Is a low-dose CT image and is,
Figure 426979DEST_PATH_IMAGE003
normal dose CT images;
s2: training a network for predicting CT image noise: a disclosed deep residual convolutional neural network model (DR-CNN) is cited as a network for predicting noise, and the network input is
Figure 878820DEST_PATH_IMAGE002
The output is a predicted residual image (i.e. a noise image) composed of
Figure DEST_PATH_IMAGE004
Obtaining; by data sets
Figure 463909DEST_PATH_IMAGE005
Performing supervised learning on the DR-CNN to obtain a trained DR-CNN;
s3: using the trained DR-CNN to carry out primary noise reduction on the LDCT image
Figure 12702DEST_PATH_IMAGE002
Inputting the predicted noise image in the trained DR-CNN:
Figure DEST_PATH_IMAGE006
（2）
wherein,
Figure 38427DEST_PATH_IMAGE007
is composed of
Figure DEST_PATH_IMAGE008
Noisy images predicted by network，
Figure 571040DEST_PATH_IMAGE009
Is DR-CNN;
s4: the LDCT is subjected to preliminary noise reduction, and the burden of fine-grained noise reduction treatment is reduced:
Figure DEST_PATH_IMAGE010
（3）
wherein,
Figure 222470DEST_PATH_IMAGE011
the result of LDCT after preliminary noise reduction is obtained;
second, fine particle size processing
S5: when the NDCT is used as a ground route, the noise in the NDCT still influences the noise reduction result, so the method uses the NDCT after noise reduction as the ground route;
Figure DEST_PATH_IMAGE012
（4）
Figure 829032DEST_PATH_IMAGE013
（5）
wherein,
Figure DEST_PATH_IMAGE014
is composed of
Figure 9346DEST_PATH_IMAGE015
The noise image predicted by the network,
Figure DEST_PATH_IMAGE016
and the result after the NDCT initial noise reduction is used as a fine-grained processed ground route.
S6: because the frequency domain image can better reflect the edge information and the detail information of the original image, the method trains the CT images on the time domain and the frequency domain respectively through two generation countermeasure networks, namely, the dual-domain combined noise reduction is carried out; the method uses the open network model WGANs, each WGAN comprises a generator and a discriminator;
s7 Overall loss function of
Figure 701359DEST_PATH_IMAGE017
（16）
Preferably, step S1 includes the steps of:
s1-1, using the Mayo dataset. The data set has 2378 pairs of CT images, i.e., each normal dose CT image has a corresponding low dose CT image.
S1-2: the data set was divided into two parts, with 1902 side CT images as the training set and 476 side CT images as the test set.
Preferably, step S2 includes the steps of:
s2-1: a disclosed deep residual convolutional neural network model (DR-CNN) is cited as a network for predicting noise, and the network input is
Figure 172791DEST_PATH_IMAGE008
The output is a predicted residual image (i.e. a noise image) composed of
Figure DEST_PATH_IMAGE018
Obtaining;
the network all using a small convolution kernel, i.e.
Figure 148706DEST_PATH_IMAGE019
And the convolution kernel enhances the nonlinearity of the network. The first layer uses Conv + ReLU, the second layer uses Conv + BN + ReLU, and then four bypass connection modules are continuously used, so that in order to avoid serious information loss phenomenon along with the increase of network depth, bypass connection is added in each module, 1 group of convolution kernels are arranged in the last convolution layer, channels are converted into 1 channel, and predicted residual images are output;
s2-2: by data sets
Figure DEST_PATH_IMAGE020
The DR-CNN is trained in the same supervised learning mode of the original network by the constructed sample set
Figure 719496DEST_PATH_IMAGE021
Middle reservation
Figure 757859DEST_PATH_IMAGE008
And obtaining corresponding residual image by adopting the mode of obtaining residual image by the original method
Figure 691705DEST_PATH_IMAGE018
Thereby obtaining
Figure 69596DEST_PATH_IMAGE020
A sample set in one-to-one correspondence.
S2-3：
Figure 935921DEST_PATH_IMAGE008
As an input image, a picture is taken,
Figure 336947DEST_PATH_IMAGE018
and (5) as a group, performing supervised learning on the network to obtain the trained DR-CNN.
The loss function of the network is
Figure DEST_PATH_IMAGE022
，
Figure 602712DEST_PATH_IMAGE023
（1）
Where N is the batch size of the BN layer,
Figure DEST_PATH_IMAGE024
is the first in a batch
Figure DEST_PATH_IMAGE025
An input image is input to the image processing device,
Figure DEST_PATH_IMAGE026
is prepared by reacting with
Figure 303952DEST_PATH_IMAGE024
Corresponding to the second in a batch
Figure 731391DEST_PATH_IMAGE025
An NDCT image, Y being a predicted residual image,
Figure 213188DEST_PATH_IMAGE027
is groudtruth.
Preferably, step S5 includes the steps of:
s5-1: the noise in NDCT is first predicted using the preliminary noise-predicted network DR-CNN:
Figure 502218DEST_PATH_IMAGE012
（4）
s5-2: the noise is then removed from the NDCT image:
Figure 385860DEST_PATH_IMAGE013
（5）
wherein,
Figure 249780DEST_PATH_IMAGE014
is composed of
Figure 422135DEST_PATH_IMAGE015
The noise image predicted by the network,
Figure 373911DEST_PATH_IMAGE016
and the result after the NDCT initial noise reduction is used as a fine-grained processed ground route.
Preferably, step S6 includes the steps of:
s6-1: firstly, processing the CT image after the preliminary noise reduction on the time domain
Figure 518584DEST_PATH_IMAGE011
Input to the generator
Figure DEST_PATH_IMAGE028
The method comprises the following steps:
Figure 897613DEST_PATH_IMAGE029
（6）
wherein
Figure DEST_PATH_IMAGE030
For the generator
Figure 1005DEST_PATH_IMAGE028
，
Figure 897417DEST_PATH_IMAGE031
For the generator
Figure 490072DEST_PATH_IMAGE028
Generating an image of (1);
s6-2: will be provided with
Figure 508844DEST_PATH_IMAGE016
As a ground channel
Figure 577163DEST_PATH_IMAGE031
Input together discriminator
Figure DEST_PATH_IMAGE032
Is judged by a discriminator
Figure 542845DEST_PATH_IMAGE031
Whether it is true or not
Figure 255586DEST_PATH_IMAGE016
：
Figure 366630DEST_PATH_IMAGE033
（7）
Wherein
Figure DEST_PATH_IMAGE034
For the generator
Figure 797612DEST_PATH_IMAGE032
Result of discrimination
Figure 504668DEST_PATH_IMAGE035
1 (true) and 0 (false);
defining the loss function of the generation countermeasure network as
Figure DEST_PATH_IMAGE036
：
Figure 258866DEST_PATH_IMAGE037
Figure 100002_DEST_PATH_IMAGE038
（8）
Figure 885019DEST_PATH_IMAGE039
（9）
Wherein
Figure 100002_DEST_PATH_IMAGE040
And
Figure 209822DEST_PATH_IMAGE041
are respectively generators
Figure 238345DEST_PATH_IMAGE028
Discriminator
Figure 660099DEST_PATH_IMAGE032
The output result of (a) is obtained,
Figure DEST_PATH_IMAGE042
is a weighted hyperparameter, set to 10 in the method,
Figure 863678DEST_PATH_IMAGE043
for the slave generator
Figure 3673DEST_PATH_IMAGE028
And a result of random sampling in an output set of the group truth image, wherein Y is a group truth;
s6-3, using the fft.fft2 () function in numpy packet to convert the data in time domain
Figure 832957DEST_PATH_IMAGE011
And
Figure 640376DEST_PATH_IMAGE016
turning to the frequency domain:
Figure 100002_DEST_PATH_IMAGE044
（10）
Figure 749278DEST_PATH_IMAGE045
（11）
wherein
Figure 100002_DEST_PATH_IMAGE046
For the purpose of a Fast Fourier Transform (FFT),
Figure 376568DEST_PATH_IMAGE047
is composed of
Figure 9544DEST_PATH_IMAGE008
After the preliminary noise reduction, the image is transferred to the frequency domain,
Figure 100002_DEST_PATH_IMAGE048
is composed of
Figure 405890DEST_PATH_IMAGE015
After the preliminary noise reduction, transferring to an image on a frequency domain;
s6-4: processing the CT image in the frequency domain with a generation countermeasure network (GAN)
Figure 279168DEST_PATH_IMAGE047
Input to the generator
Figure 534700DEST_PATH_IMAGE049
The method comprises the following steps:
Figure 100002_DEST_PATH_IMAGE050
（12）
wherein
Figure 49995DEST_PATH_IMAGE051
For the generator
Figure 956640DEST_PATH_IMAGE049
，
Figure 100002_DEST_PATH_IMAGE052
For the generator
Figure 797557DEST_PATH_IMAGE049
Generating an image of (1);
s6-5: will be provided with
Figure 602702DEST_PATH_IMAGE048
As a ground channel
Figure 62633DEST_PATH_IMAGE052
Input together discriminator
Figure 699151DEST_PATH_IMAGE053
Is judged by a discriminator
Figure 914232DEST_PATH_IMAGE052
Whether it is true or not
Figure 130974DEST_PATH_IMAGE048
：
Figure 100002_DEST_PATH_IMAGE054
（13）
Wherein
Figure 519230DEST_PATH_IMAGE055
For the generator
Figure 947937DEST_PATH_IMAGE053
Result of discrimination
Figure 100002_DEST_PATH_IMAGE056
1 (true) and 0 (false);
defining the loss function of the generation countermeasure network as
Figure 802761DEST_PATH_IMAGE057
：
Figure 100002_DEST_PATH_IMAGE058
Figure 503869DEST_PATH_IMAGE059
（14）
Figure 100002_DEST_PATH_IMAGE060
（15）
Wherein
Figure 430237DEST_PATH_IMAGE061
And
Figure 100002_DEST_PATH_IMAGE062
are respectively generators
Figure 385555DEST_PATH_IMAGE049
Discriminator
Figure 208017DEST_PATH_IMAGE053
The output result of (a) is obtained,
Figure 6209DEST_PATH_IMAGE043
for the slave generator
Figure 126481DEST_PATH_IMAGE049
And output of grountruth imageAnd (6) giving out the result of random sampling in the set.
Due to the adoption of the technical scheme, compared with the prior art, the invention has the following beneficial effects: the method comprises the steps of firstly predicting the noise of the NDCT at a fine-grained noise reduction part, removing the noise from the NDCT to obtain a noise-reduced NDCT image, and meanwhile inputting the noise-reduced NDCT into a discriminator at a discrimination stage, so that the CT image generated by a generator is better fitted with the noise-reduced NDCT, and a better noise reduction result is achieved; for the low-dose CT noise reduction task, images in a time domain are generally used for training, and the images in the time domain and a frequency domain are used for simultaneously participating in noise reduction, so that the defects of edge blurring and partial detail loss can be overcome in the training process of the frequency domain images while noise reduction is carried out.
Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
fig. 1 is a flow chart of fine-grained noise reduction training.
Detailed Description
In order that the above objects, features and advantages of the present invention can be more clearly understood, a more particular description of the invention will be rendered by reference to the appended drawings. It should be noted that the embodiments and features of the embodiments of the present application may be combined with each other without conflict.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced otherwise than as specifically described herein, and thus the scope of the present invention is not limited by the specific embodiments disclosed below.
The following describes a GAN-based low-dose CT dual-domain combined denoising method with NDCT noise fusion according to an embodiment of the present invention with reference to fig. 1.
The invention provides a GAN-based NDCT noise-fused low-dose CT double-domain combined noise reduction method, which comprises coarse grain processing and fine grain processing and is characterized by comprising the following specific steps of:
first, coarse size treatment
S1: sample set for training CT image noise reduction network
Figure 264201DEST_PATH_IMAGE001
Wherein
Figure 319882DEST_PATH_IMAGE002
Is a low-dose CT image and is,
Figure 11894DEST_PATH_IMAGE003
normal dose CT images;
s1-1, using the Mayo dataset. The data set has 2378 pairs of CT images, i.e., each normal dose CT image has a corresponding low dose CT image.
S1-2: the data set was divided into two parts, with 1902 side CT images as the training set and 476 side CT images as the test set.
S2: training a network for predicting CT image noise: a disclosed deep residual convolutional neural network model (DR-CNN) is cited as a network for predicting noise, and the network input is
Figure 483327DEST_PATH_IMAGE002
The output is a predicted residual image (i.e. a noise image) composed of
Figure 928083DEST_PATH_IMAGE004
Obtaining; by data sets
Figure 92348DEST_PATH_IMAGE005
Performing supervised learning on the DR-CNN to obtain a trained DR-CNN;
s2-1: a disclosed deep residual convolutional neural network model (DR-CNN) is cited as a network for predicting noise, and the network input is
Figure 865132DEST_PATH_IMAGE008
The output is a predicted residual image (i.e. a noise image) composed of
Figure 812360DEST_PATH_IMAGE018
Obtaining;
the network all using a small convolution kernel, i.e.
Figure 190252DEST_PATH_IMAGE019
And the convolution kernel enhances the nonlinearity of the network. The first layer uses Conv + ReLU, the second layer uses Conv + BN + ReLU, and then four bypass connection modules are continuously used, so that in order to avoid serious information loss phenomenon along with the increase of network depth, bypass connection is added in each module, 1 group of convolution kernels are arranged in the last convolution layer, channels are converted into 1 channel, and predicted residual images are output;
s2-2: by data sets
Figure 322156DEST_PATH_IMAGE020
The DR-CNN is trained in the same supervised learning mode of the original network by the constructed sample set
Figure 785498DEST_PATH_IMAGE021
Middle reservation
Figure 788614DEST_PATH_IMAGE008
And obtaining corresponding residual image by adopting the mode of obtaining residual image by the original method
Figure 21012DEST_PATH_IMAGE018
Thereby obtaining
Figure 323817DEST_PATH_IMAGE020
A sample set in one-to-one correspondence.
S2-3：
Figure 274456DEST_PATH_IMAGE008
As an input image, a picture is taken,
Figure 563486DEST_PATH_IMAGE018
and (5) as a group, performing supervised learning on the network to obtain the trained DR-CNN.
The loss function of the network is
Figure 712707DEST_PATH_IMAGE022
，
Figure 389676DEST_PATH_IMAGE023
（1）
Where N is the batch size of the BN layer,
Figure 748982DEST_PATH_IMAGE024
is the first in a batch
Figure 700758DEST_PATH_IMAGE025
An input image is input to the image processing device,
Figure 907748DEST_PATH_IMAGE026
is prepared by reacting with
Figure 427722DEST_PATH_IMAGE024
Corresponding to the second in a batch
Figure 87374DEST_PATH_IMAGE025
An NDCT image, Y being a predicted residual image,
Figure 108419DEST_PATH_IMAGE027
is groudtruth.
S3: using the trained DR-CNN to carry out primary noise reduction on the LDCT image
Figure 91288DEST_PATH_IMAGE002
Inputting the predicted noise image in the trained DR-CNN:
Figure 110059DEST_PATH_IMAGE006
（2）
wherein,
Figure 319324DEST_PATH_IMAGE007
is composed of
Figure 816164DEST_PATH_IMAGE008
The noise image predicted by the network,
Figure 669851DEST_PATH_IMAGE009
is DR-CNN;
s4: the LDCT is subjected to preliminary noise reduction, and the burden of fine-grained noise reduction treatment is reduced:
Figure 859524DEST_PATH_IMAGE010
（3）
wherein,
Figure 556084DEST_PATH_IMAGE011
the result of LDCT after preliminary noise reduction is obtained;
second, fine particle size processing
The LDCT cannot obtain satisfactory noise reduction effect after preliminary noise reduction, and needs fine-grained noise reduction processing.
S5: when the NDCT is used as a ground route, the noise in the NDCT still influences the noise reduction result, so the method uses the NDCT after noise reduction as the ground route;
s5-1: the noise in NDCT is first predicted using the preliminary noise-predicted network DR-CNN:
Figure 777987DEST_PATH_IMAGE012
（4）
s5-2: the noise is then removed from the NDCT image:
Figure 548497DEST_PATH_IMAGE013
（5）
wherein,
Figure 971388DEST_PATH_IMAGE014
is composed of
Figure 92928DEST_PATH_IMAGE015
The noise image predicted by the network,
Figure 603675DEST_PATH_IMAGE016
and the result after the NDCT initial noise reduction is used as a fine-grained processed ground route.
S6: because the frequency domain image can better reflect the edge information and the detail information of the original image, the method trains the CT images on the time domain and the frequency domain respectively through two generation countermeasure networks, namely, the dual-domain combined noise reduction is carried out; the method uses the open network model WGANs, each WGAN comprises a generator and a discriminator; as shown in fig. 1.
S6-1: firstly, processing the CT image after the preliminary noise reduction on the time domain
Figure 556587DEST_PATH_IMAGE011
Input to the generator
Figure 88063DEST_PATH_IMAGE028
The method comprises the following steps:
Figure 445418DEST_PATH_IMAGE029
（6）
wherein
Figure 150069DEST_PATH_IMAGE030
For the generator
Figure 895171DEST_PATH_IMAGE028
，
Figure 535231DEST_PATH_IMAGE031
For the generator
Figure 365784DEST_PATH_IMAGE028
Generating an image of (1);
s6-2: will be provided with
Figure 608546DEST_PATH_IMAGE016
As a ground channel
Figure 473734DEST_PATH_IMAGE031
Input together discriminator
Figure 533963DEST_PATH_IMAGE032
Is judged by a discriminator
Figure 851812DEST_PATH_IMAGE031
Whether it is true or not
Figure 898265DEST_PATH_IMAGE016
：
Figure 290064DEST_PATH_IMAGE033
（7）
Wherein
Figure 334243DEST_PATH_IMAGE034
For the generator
Figure 201705DEST_PATH_IMAGE032
Result of discrimination
Figure 989532DEST_PATH_IMAGE035
1 (true) and 0 (false);
defining the loss function of the generation countermeasure network as
Figure 750684DEST_PATH_IMAGE036
：
Figure 965764DEST_PATH_IMAGE037
Figure 54943DEST_PATH_IMAGE038
（8）
Figure 646461DEST_PATH_IMAGE039
（9）
Wherein
Figure 747273DEST_PATH_IMAGE040
And
Figure 195572DEST_PATH_IMAGE041
are respectively generators
Figure 975309DEST_PATH_IMAGE028
Discriminator
Figure 291889DEST_PATH_IMAGE032
The output result of (a) is obtained,
Figure 575103DEST_PATH_IMAGE042
is a weighted hyperparameter, set to 10 in the method,
Figure 194303DEST_PATH_IMAGE043
for the slave generator
Figure 133441DEST_PATH_IMAGE028
And a result of random sampling in an output set of the group truth image, wherein Y is a group truth;
s6-3, using the fft.fft2 () function in numpy packet to convert the data in time domain
Figure 801182DEST_PATH_IMAGE011
And
Figure 266799DEST_PATH_IMAGE016
turning to the frequency domain:
Figure 260162DEST_PATH_IMAGE044
（10）
Figure 673214DEST_PATH_IMAGE045
（11）
wherein
Figure 206963DEST_PATH_IMAGE046
For the purpose of a Fast Fourier Transform (FFT),
Figure 464769DEST_PATH_IMAGE047
is composed of
Figure 301138DEST_PATH_IMAGE008
After the preliminary noise reduction, the image is transferred to the frequency domain,
Figure 277185DEST_PATH_IMAGE048
is composed of
Figure 614625DEST_PATH_IMAGE015
After the preliminary noise reduction, transferring to an image on a frequency domain;
s6-4: processing the CT image in the frequency domain with a generation countermeasure network (GAN)
Figure 648309DEST_PATH_IMAGE047
Input to the generator
Figure 983475DEST_PATH_IMAGE049
The method comprises the following steps:
Figure 243555DEST_PATH_IMAGE050
（12）
wherein
Figure 322370DEST_PATH_IMAGE051
For the generator
Figure 492451DEST_PATH_IMAGE049
，
Figure 732940DEST_PATH_IMAGE052
For the generator
Figure 745895DEST_PATH_IMAGE049
Generating an image of (1);
s6-5: will be provided with
Figure 362821DEST_PATH_IMAGE048
As a ground channel
Figure 371097DEST_PATH_IMAGE052
Input together discriminator
Figure 844804DEST_PATH_IMAGE053
Is judged by a discriminator
Figure 17159DEST_PATH_IMAGE052
Whether it is true or not
Figure 375459DEST_PATH_IMAGE048
：
Figure 316871DEST_PATH_IMAGE054
（13）
Wherein
Figure 227058DEST_PATH_IMAGE055
For the generator
Figure 886709DEST_PATH_IMAGE053
Result of discrimination
Figure 766809DEST_PATH_IMAGE056
1 (true) and 0 (false);
defining the loss function of the generation countermeasure network as
Figure 890623DEST_PATH_IMAGE057
：
Figure 909395DEST_PATH_IMAGE058
Figure 728446DEST_PATH_IMAGE059
（14）
Figure 490866DEST_PATH_IMAGE060
（15）
Wherein
Figure 469186DEST_PATH_IMAGE061
And
Figure 658859DEST_PATH_IMAGE062
are respectively generators
Figure 482983DEST_PATH_IMAGE049
Discriminator
Figure 580252DEST_PATH_IMAGE053
The output result of (a) is obtained,
Figure 350762DEST_PATH_IMAGE043
for the slave generator
Figure 649019DEST_PATH_IMAGE049
And the result of random sampling in the output set of the groudtuth image.
S7 Overall loss function of
Figure 504980DEST_PATH_IMAGE017
（16）
And continuous optimization is carried out through training, so that a high-quality CT image is obtained, and a good noise reduction effect is achieved.
In the description of the present invention, the terms "plurality" or "a plurality" refer to two or more, and unless otherwise specifically limited, the terms "upper", "lower", and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, and are merely for convenience in describing the present invention and simplifying the description, but do not indicate or imply that the referred device or element must have a specific orientation, be constructed in a specific orientation, and be operated, and thus should not be construed as limiting the present invention; the terms "connected," "mounted," "secured," and the like are to be construed broadly and include, for example, fixed connections, removable connections, or integral connections; may be directly connected or indirectly connected through an intermediate. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
In the description herein, the description of the terms "one embodiment," "some embodiments," "specific embodiments," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention, and various modifications and changes will occur to those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (5)

1. A low-dose CT double-domain combined noise reduction method based on GAN fusion NDCT noise comprises coarse grain processing and fine grain processing, and is characterized by comprising the following specific steps:
first, coarse size treatment
S1: sample set for training CT image noise reduction network
Figure 318177DEST_PATH_IMAGE001
Wherein
Figure 533258DEST_PATH_IMAGE002
Is a low-dose CT image and is,
Figure 560120DEST_PATH_IMAGE003
normal dose CT images;
s2: training predicted CT mapsNetwork like noise: a disclosed deep residual convolutional neural network model (DR-CNN) is cited as a network for predicting noise, and the network input is
Figure 886059DEST_PATH_IMAGE002
The output is a predicted residual image (i.e. a noise image) composed of
Figure 580345DEST_PATH_IMAGE004
Obtaining; by data sets
Figure 966327DEST_PATH_IMAGE005
Performing supervised learning on the DR-CNN to obtain a trained DR-CNN;
s3: using the trained DR-CNN to carry out primary noise reduction on the LDCT image
Figure 965638DEST_PATH_IMAGE002
Inputting the predicted noise image in the trained DR-CNN:
Figure 95268DEST_PATH_IMAGE006
（2）
wherein,
Figure 644062DEST_PATH_IMAGE007
is composed of
Figure 200945DEST_PATH_IMAGE002
The noise image predicted by the network,
Figure 202399DEST_PATH_IMAGE008
is DR-CNN;
s4: the LDCT is subjected to preliminary noise reduction, and the burden of fine-grained noise reduction treatment is reduced:
Figure 135720DEST_PATH_IMAGE009
（3）
wherein,
Figure 522708DEST_PATH_IMAGE010
the result of LDCT after preliminary noise reduction is obtained;
second, fine particle size processing
S5: when the NDCT is used as a ground route, the noise in the NDCT still influences the noise reduction result, so the method uses the NDCT after noise reduction as the ground route;
Figure 250492DEST_PATH_IMAGE011
（4）
Figure 739242DEST_PATH_IMAGE012
（5）
wherein,
Figure 476254DEST_PATH_IMAGE013
is composed of
Figure 734060DEST_PATH_IMAGE003
The noise image predicted by the network,
Figure 632746DEST_PATH_IMAGE014
taking the result of NDCT after preliminary noise reduction as a fine-grained processed ground route;
s6: because the frequency domain image can better reflect the edge information and the detail information of the original image, the method trains the CT images on the time domain and the frequency domain respectively through two generation countermeasure networks, namely, the dual-domain combined noise reduction is carried out; the method uses the open network model WGANs, each WGAN comprises a generator and a discriminator;
s7 Overall loss function of
Figure 608792DEST_PATH_IMAGE015
（16）。
2. The method according to claim 1, wherein the step S1 comprises the steps of:
s1-1, using a Mayo data set; the data set has 2378 pairs of CT images, i.e., each normal dose CT image has a corresponding low dose CT image;
s1-2: the data set was divided into two parts, with 1902 side CT images as the training set and 476 side CT images as the test set.
3. The method according to claim 1, wherein the step S2 comprises the steps of:
s2-1: a disclosed deep residual convolutional neural network model (DR-CNN) is cited as a network for predicting noise, and the network input is
Figure 366139DEST_PATH_IMAGE016
The output is a predicted residual image (i.e. a noise image) composed of
Figure 478452DEST_PATH_IMAGE017
Obtaining;
the network all using a small convolution kernel, i.e.
Figure 813618DEST_PATH_IMAGE018
A convolution kernel enhances the nonlinearity of the network; the first layer uses Conv + ReLU, the second layer uses Conv + BN + ReLU, and then four bypass connection modules are continuously used, so that in order to avoid serious information loss phenomenon along with the increase of network depth, bypass connection is added in each module, 1 group of convolution kernels are arranged in the last convolution layer, channels are converted into 1 channel, and predicted residual images are output;
s2-2: by data sets
Figure 11381DEST_PATH_IMAGE019
The DR-CNN is trained in the same supervised learning mode of the original network by the constructed sample set
Figure 90196DEST_PATH_IMAGE020
Middle reservation
Figure 791436DEST_PATH_IMAGE016
And obtaining corresponding residual image by adopting the mode of obtaining residual image by the original method
Figure 297503DEST_PATH_IMAGE017
Thereby obtaining
Figure 231830DEST_PATH_IMAGE019
A sample set corresponding to each other;
S2-3：
Figure 848756DEST_PATH_IMAGE016
as an input image, a picture is taken,
Figure 935661DEST_PATH_IMAGE017
as a group, performing supervised learning on the network to obtain a trained DR-CNN;
the loss function of the network is
Figure 347051DEST_PATH_IMAGE021
，
Figure 519406DEST_PATH_IMAGE022
（1）
Where N is the batch size of the BN layer,
Figure 940023DEST_PATH_IMAGE023
is the first in a batch
Figure 881434DEST_PATH_IMAGE024
An input image is input to the image processing device,
Figure 480037DEST_PATH_IMAGE025
is prepared by reacting with
Figure 139689DEST_PATH_IMAGE023
Corresponding to the second in a batch
Figure 98417DEST_PATH_IMAGE024
An NDCT image, Y being a predicted residual image,
Figure 894335DEST_PATH_IMAGE026
is groudtruth.
4. The method according to claim 1, wherein the step S5 comprises the steps of:
s5-1: the noise in NDCT is first predicted using the preliminary noise-predicted network DR-CNN:
Figure 647527DEST_PATH_IMAGE027
（4）
s5-2: the noise is then removed from the NDCT image:
Figure 794475DEST_PATH_IMAGE028
（5）
wherein,
Figure 556895DEST_PATH_IMAGE029
is composed of
Figure 456586DEST_PATH_IMAGE030
The noise image predicted by the network,
Figure 646259DEST_PATH_IMAGE031
and the result after the NDCT initial noise reduction is used as a fine-grained processed ground route.
5. The method according to claim 1, wherein the step S6 comprises the steps of:
s6-1: firstly, processing the CT image after the preliminary noise reduction on the time domain
Figure 14924DEST_PATH_IMAGE032
Input to the generator
Figure 315455DEST_PATH_IMAGE033
The method comprises the following steps:
Figure 351544DEST_PATH_IMAGE034
（6）
wherein
Figure 446539DEST_PATH_IMAGE035
For the generator
Figure 56162DEST_PATH_IMAGE033
，
Figure 160384DEST_PATH_IMAGE036
For the generator
Figure 785400DEST_PATH_IMAGE033
Generating an image of (1);
s6-2: will be provided with
Figure 316876DEST_PATH_IMAGE031
As a ground channel
Figure 660132DEST_PATH_IMAGE036
Input together discriminator
Figure 302466DEST_PATH_IMAGE037
Is judged by a discriminator
Figure 47569DEST_PATH_IMAGE036
Whether it is true or not
Figure 733634DEST_PATH_IMAGE031
：
Figure DEST_PATH_IMAGE038
（7）
Wherein
Figure 298607DEST_PATH_IMAGE039
For the generator
Figure 744632DEST_PATH_IMAGE037
Result of discrimination
Figure DEST_PATH_IMAGE040
1 (true) and 0 (false);
defining the loss function of the generation countermeasure network as
Figure 813082DEST_PATH_IMAGE041
：
Figure 437093DEST_PATH_IMAGE043
Figure DEST_PATH_IMAGE044
（8）
Figure 489362DEST_PATH_IMAGE045
（9）
Wherein
Figure DEST_PATH_IMAGE046
And
Figure 473499DEST_PATH_IMAGE047
are respectively generators
Figure 662035DEST_PATH_IMAGE033
Discriminator
Figure 706214DEST_PATH_IMAGE037
The output result of (a) is obtained,
Figure DEST_PATH_IMAGE048
is a weighted hyperparameter, set to 10 in the method,
Figure 495047DEST_PATH_IMAGE049
for the slave generator
Figure 17296DEST_PATH_IMAGE033
And a result of random sampling in an output set of the group truth image, wherein Y is a group truth;
s6-3, using the fft.fft2 () function in numpy packet to convert the data in time domain
Figure 591496DEST_PATH_IMAGE032
And
Figure 806577DEST_PATH_IMAGE031
turning to the frequency domain:
Figure DEST_PATH_IMAGE050
（10）
Figure 833439DEST_PATH_IMAGE051
（11）
wherein
Figure DEST_PATH_IMAGE052
For the purpose of a Fast Fourier Transform (FFT),
Figure 376022DEST_PATH_IMAGE053
is composed of
Figure 804730DEST_PATH_IMAGE016
After the preliminary noise reduction, the image is transferred to the frequency domain,
Figure DEST_PATH_IMAGE054
is composed of
Figure 190712DEST_PATH_IMAGE030
After the preliminary noise reduction, transferring to an image on a frequency domain;
s6-4: processing the CT image in the frequency domain with a generation countermeasure network (GAN)
Figure 704870DEST_PATH_IMAGE053
Input to the generator
Figure 834500DEST_PATH_IMAGE055
The method comprises the following steps:
Figure DEST_PATH_IMAGE056
（12）
wherein
Figure 101402DEST_PATH_IMAGE057
For the generator
Figure 658285DEST_PATH_IMAGE055
，
Figure DEST_PATH_IMAGE058
For the generator
Figure 394160DEST_PATH_IMAGE055
Generating an image of (1);
s6-5: will be provided with
Figure 61902DEST_PATH_IMAGE054
As a ground channel
Figure 215934DEST_PATH_IMAGE058
Input together discriminator
Figure 943718DEST_PATH_IMAGE059
Is judged by a discriminator
Figure 432468DEST_PATH_IMAGE058
Whether it is true or not
Figure 638322DEST_PATH_IMAGE054
：
Figure DEST_PATH_IMAGE060
（13）
Wherein
Figure 630548DEST_PATH_IMAGE061
For the generator
Figure 794814DEST_PATH_IMAGE059
Result of discrimination
Figure DEST_PATH_IMAGE062
1 (true) and 0 (false);
defining the loss function of the generation countermeasure network as
Figure 488969DEST_PATH_IMAGE063
：
Figure 498513DEST_PATH_IMAGE065
Figure DEST_PATH_IMAGE066
（14）
Figure 610826DEST_PATH_IMAGE067
（15）
Wherein
Figure DEST_PATH_IMAGE068
And
Figure 414834DEST_PATH_IMAGE069
are respectively generators
Figure 354540DEST_PATH_IMAGE055
Discriminator
Figure 167775DEST_PATH_IMAGE059
The output result of (a) is obtained,
Figure 400174DEST_PATH_IMAGE049
for the slave generator
Figure 906241DEST_PATH_IMAGE055
And the result of random sampling in the output set of the groudtuth image.
CN202111043544.2A 2021-09-07 2021-09-07 GAN-based NDCT noise-fused low-dose CT double-domain combined noise reduction method Active CN114004753B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111043544.2A CN114004753B (en) 2021-09-07 2021-09-07 GAN-based NDCT noise-fused low-dose CT double-domain combined noise reduction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111043544.2A CN114004753B (en) 2021-09-07 2021-09-07 GAN-based NDCT noise-fused low-dose CT double-domain combined noise reduction method

Publications (2)

Publication Number Publication Date
CN114004753A true CN114004753A (en) 2022-02-01
CN114004753B CN114004753B (en) 2022-07-01

Family

ID=79921220

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111043544.2A Active CN114004753B (en) 2021-09-07 2021-09-07 GAN-based NDCT noise-fused low-dose CT double-domain combined noise reduction method

Country Status (1)

Country Link
CN (1) CN114004753B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115187470A (en) * 2022-06-10 2022-10-14 成都飞机工业（集团）有限责任公司 Double-domain iterative noise reduction method based on 3D printing inner cavity
CN116862790A (en) * 2023-06-30 2023-10-10 中北大学 Neural network-based low-dose CT image noise reduction method and device

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5331550A (en) * 1991-03-05 1994-07-19 E. I. Du Pont De Nemours And Company Application of neural networks as an aid in medical diagnosis and general anomaly detection
CN111429379A (en) * 2020-04-01 2020-07-17 山东大学 Low-dose CT image denoising method and system based on self-supervision learning
CN113205461A (en) * 2021-04-02 2021-08-03 上海慧虎信息科技有限公司 Low-dose CT image denoising model training method, denoising method and denoising device

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5331550A (en) * 1991-03-05 1994-07-19 E. I. Du Pont De Nemours And Company Application of neural networks as an aid in medical diagnosis and general anomaly detection
CN111429379A (en) * 2020-04-01 2020-07-17 山东大学 Low-dose CT image denoising method and system based on self-supervision learning
CN113205461A (en) * 2021-04-02 2021-08-03 上海慧虎信息科技有限公司 Low-dose CT image denoising model training method, denoising method and denoising device

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
徐曾春等: "基于改进型WGAN的低剂量CT图像去噪方法", 《光学与光电技术》 *
朱永成等: "基于字典学习的低剂量X-ray CT图像去噪", 《东南大学学报(自然科学版)》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115187470A (en) * 2022-06-10 2022-10-14 成都飞机工业（集团）有限责任公司 Double-domain iterative noise reduction method based on 3D printing inner cavity
CN116862790A (en) * 2023-06-30 2023-10-10 中北大学 Neural network-based low-dose CT image noise reduction method and device

Also Published As

Publication number Publication date
CN114004753B (en) 2022-07-01

Similar Documents

Publication Publication Date Title
CN114004753B (en) GAN-based NDCT noise-fused low-dose CT double-domain combined noise reduction method
CN110889895B (en) Face video super-resolution reconstruction method fusing single-frame reconstruction network
CN115409733A (en) Low-dose CT image noise reduction method based on image enhancement and diffusion model
CN115953494B (en) Multi-task high-quality CT image reconstruction method based on low dose and super resolution
WO2022045210A1 (en) Image processing device, image processing method, learning device, learning method, and program
CN114187181B (en) Dual-path lung CT image super-resolution method based on residual information refining
Jiang et al. Wasserstein generative adversarial networks for motion artifact removal in dental CT imaging
More et al. Denoising Medical Images Using Deep Learning in IoT Environment.
Liu et al. Unsupervised CT metal artifact reduction by plugging diffusion priors in dual domains
Chan et al. An attention-based deep convolutional neural network for ultra-sparse-view CT reconstruction
CN112767273B (en) Low-dose CT image restoration method and system applying feature decoupling
CN111462004B (en) Image enhancement method and device, computer equipment and storage medium
WO2022047627A1 (en) Deep learning prediction method and application thereof
CN117974468A (en) Multi-mode medical image fusion method for global and local feature interaction parallelism
Sun et al. A lightweight dual-domain attention framework for sparse-view CT reconstruction
Patil et al. Classification of 3D magnetic resonance images of brain using discrete wavelet transform
CN114202464B (en) X-ray CT local high-resolution imaging method and device based on deep learning
Sharma et al. Brain region segmentation using low MSE based active contour model and convolutional neural network
Lu et al. Low-dose CT denoising via neural architecture search
Wu et al. Urine red blood cells generation using StyleGAN2 network
Mahmoodabadi et al. Wavelets and fuzzy relational classifiers: A novel diffusion-weighted image analysis system for pediatric metabolic brain diseases
JP7566696B2 (en) IMAGE PROCESSING APPARATUS, IMAGE PROCESSING METHOD, LEARNING APPARATUS, LEARNING METHOD, AND PROGRAM
Liu et al. Recent Advances in Sparse and Ultra-Sparse Reconstruction for Medical Imaging
Reddy et al. Enhancing Medical Imaging: Noise Reduction and Super Resolution with Transfer Learning
CN117726705B (en) Deep learning method for simultaneous low-dose CT reconstruction and metal artifact correction

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