CN112419203B - Diffusion weighted image compressed sensing recovery method and device based on countermeasure network - Google Patents
Diffusion weighted image compressed sensing recovery method and device based on countermeasure network Download PDFInfo
- Publication number
- CN112419203B CN112419203B CN202011451621.3A CN202011451621A CN112419203B CN 112419203 B CN112419203 B CN 112419203B CN 202011451621 A CN202011451621 A CN 202011451621A CN 112419203 B CN112419203 B CN 112419203B
- Authority
- CN
- China
- Prior art keywords
- image
- dwi
- undersampled
- discriminator
- generator
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 91
- 238000009792 diffusion process Methods 0.000 title claims abstract description 74
- 238000011084 recovery Methods 0.000 title claims description 21
- 238000012549 training Methods 0.000 claims abstract description 24
- 238000013528 artificial neural network Methods 0.000 claims abstract description 12
- 238000003062 neural network model Methods 0.000 claims abstract description 12
- 238000002597 diffusion-weighted imaging Methods 0.000 claims description 125
- 238000002598 diffusion tensor imaging Methods 0.000 claims description 41
- 238000012545 processing Methods 0.000 claims description 24
- 238000004458 analytical method Methods 0.000 claims description 14
- 239000000835 fiber Substances 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 abstract description 7
- 238000005457 optimization Methods 0.000 description 10
- 230000008569 process Effects 0.000 description 10
- 230000006870 function Effects 0.000 description 8
- 238000005070 sampling Methods 0.000 description 8
- 238000013135 deep learning Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000003384 imaging method Methods 0.000 description 6
- 238000010801 machine learning Methods 0.000 description 6
- 238000012805 post-processing Methods 0.000 description 5
- 238000011158 quantitative evaluation Methods 0.000 description 5
- 238000007781 pre-processing Methods 0.000 description 4
- 230000000007 visual effect Effects 0.000 description 4
- 238000005259 measurement Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 3
- 230000000052 comparative effect Effects 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000006073 displacement reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000008485 antagonism Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001351 cycling effect Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000001727 in vivo Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000002595 magnetic resonance imaging Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000013421 nuclear magnetic resonance imaging Methods 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 230000005855 radiation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- 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 application discloses a diffusion weighted image compressed sensing restoration method and device based on an countermeasure network, wherein the method comprises the following steps: acquiring a diffusion weighted image DWI airspace image; undersampling the DWI airspace image to obtain an undersampled airspace image; inputting the undersampled spatial image into a generator, and inputting the DWI spatial image into a discriminator; alternately training the generator and the discriminator to perform countermeasure, and obtaining network parameters corresponding to the neural network of the generator and/or the discriminator; and recovering the undersampled DWI which needs to be recovered by using the obtained network parameters. The method solves the problem that the existing artifact and noise removing method does not have learning ability, calculates by using the neural network model, so that the neural network model has learning type, improves the calculation efficiency, and saves hardware calculation resources to a certain extent.
Description
Technical Field
The present invention relates to the field of image processing, and in particular, to a diffusion weighted image compressed sensing restoration method and apparatus based on an countermeasure network.
Background
The magnetic resonance diffusion imaging technology (diffusion magnetic resonance imaging, dMRI) is the only imaging means at present capable of detecting the microstructure of human tissue without damage and radiation, and detecting the situation of water molecule diffusion displacement in biological tissue, so as to infer microscopic information such as tissue fiber direction, tissue structure integrity, tissue size, volume ratio and the like. And acquiring a diffusion weighted image DWI according to a preset diffusion encoding gradient (a measurement parameter b value and a measurement direction bvec). However, DWI imaging processes are extremely motion sensitive, and are subject to motion disturbances during long scans, and many studies have been devoted to developing a method for accelerating DWI acquisition rate, typically by using Compressed Sensing (CS) theory, reducing signal acquisition and restoring imaging quality by computer post-processing methods.
Reducing the amount of acquired signals inevitably causes aliasing artifacts to the DWI, which results in an inability to accurately measure the tissue-in-vivo microscopic tissue structure characteristics. In order to remove artifacts and recover images, one method selects a structured signal obtained from a small number of random measurements to recover a high-dimensional signal, from which the reconstructed spatial image, i.e., the DWI will have a high probability of being free of artifacts; another approach focuses on post-processing of the spatial image, i.e., converting the reconstruction-based algorithm to estimate the original signal from the interfered observations, in removing aliasing artifacts and noise from the spatial image and recovering a high quality DWI.
The current artifact and noise removal method includes: the use of linear or nonlinear filters to remove complex random artifacts and noise in various combinations in corrupted images, and further the use of local signal-to-noise ratio information and edge information to estimate the artifacts and noise, these methods tend to be low performing and suffer from the following drawbacks: limiting to a particular task; specific optimization methods are required; without learning ability, optimization tasks and the like need to be repeatedly performed.
Aiming at the problem that the existing artifact and noise removal method does not have learning capability, no effective solution is proposed in the related art at present.
Disclosure of Invention
The application provides a diffusion weighted image compressed sensing restoration method based on an countermeasure network, which aims to solve the problem that the existing artifact and noise removal method does not have learning ability.
According to one aspect of the present application, there is provided a diffusion weighted image compressed sensing restoration method based on an countermeasure network, including: acquiring a diffusion weighted image DWI airspace image; undersampling the DWI airspace image to obtain an undersampled airspace image; inputting the undersampled spatial image into a generator, inputting the DWI spatial image into a discriminator, wherein the generator is coupled with the discriminator, and is used for generating an image according to the undersampled spatial image and inputting the generated image into the discriminator; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the discriminator are divided into neural network models; alternately training the generator and the discriminator to perform countermeasure, and obtaining network parameters corresponding to the neural network of the generator and/or the discriminator; and recovering the undersampled DWI which needs to be recovered by using the obtained network parameters.
Further, before undersampling the DWI spatial domain image, the method further includes: and normalizing at least the DWI spatial domain image, and normalizing the DWI spatial domain image data to [ -1,1].
Further, performing undersampling processing on the DWI spatial domain image to obtain the undersampled spatial domain image includes: converting the DWI airspace image into a spatial signal; selecting undersampled templates, wherein different undersampled templates produce different forms of artifacts; processing the spatial signal using the selected undersampled templates; and converting the processed space signal into DWI to obtain the undersampled airspace image.
Further, after recovering the undersampled DWI that needs to be recovered using the obtained network parameters, the method further includes: the recovered diffusion tensor is reconstructed from the DWI to image DTI and analyzed for diffusion characteristics.
Further, the diffusion tensor imaging DTI after reconstruction recovery from DWI includes: reconstructing the recovered diffusion tensor imaging DTI from the DWI by least squares fitting; and/or analyzing the diffusion characteristics includes: calculating at least one of the following parameters from the DTI: fractional anisotropy FA, average diffusion coefficient MD, included angle error between fibers, and analysis is performed according to the calculated parameters.
According to another aspect of the present application, there is also provided a diffusion weighted image compressed sensing retrieval device based on an countermeasure network, including: the acquisition module is used for acquiring a diffusion weighted image DWI airspace image; the undersampling module is used for undersampling the DWI airspace image to obtain an undersampled airspace image; the input module is used for inputting the undersampled airspace image into a generator and inputting the DWI airspace image into a discriminator, wherein the generator is coupled with the discriminator, and is used for generating an image according to the undersampled airspace image and inputting the generated image into the discriminator; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the discriminator are divided into neural network models; the training module is used for alternately training the generator and the discriminator to fight against the neural network, so as to obtain network parameters corresponding to the neural network of the generator and/or the discriminator; and the recovery module is used for recovering the undersampled DWI which needs to be recovered by using the obtained network parameters.
Further, the method further comprises the following steps: and the processing module is used for carrying out normalization processing on at least the DWI spatial domain image and normalizing the DWI spatial domain image data to [ -1,1].
Further, the processing module is configured to: converting the DWI airspace image into a spatial signal; selecting undersampled templates, wherein different undersampled templates produce different forms of artifacts; processing the spatial signal using the selected undersampled templates; and converting the processed space signal into DWI to obtain the undersampled airspace image.
Further, the method further comprises the following steps: and the analysis module is used for reconstructing the restored diffusion tensor imaging DTI from the DWI and analyzing the diffusion characteristic.
Further, the analysis module is used for reconstructing the restored diffusion tensor imaging DTI from the DWI through least square fitting; and/or for calculating from the DTI at least one of the following parameters: fractional anisotropy FA, average diffusion coefficient MD, included angle error between fibers, and analysis is performed according to the calculated parameters.
The application adopts the following steps: acquiring a diffusion weighted image DWI airspace image; undersampling the DWI airspace image to obtain an undersampled airspace image; inputting the undersampled spatial image into a generator, inputting the DWI spatial image into a discriminator, wherein the generator is coupled with the discriminator, and is used for generating an image according to the undersampled spatial image and inputting the generated image into the discriminator; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the discriminator are divided into neural network models; alternately training the generator and the discriminator to perform countermeasure, and obtaining network parameters corresponding to the neural network of the generator and/or the discriminator; and recovering the undersampled DWI which needs to be recovered by using the obtained network parameters. The method solves the problem that the existing artifact and noise removing method does not have learning ability, calculates by using the neural network model, so that the neural network model has learning type, improves the calculation efficiency, and saves hardware calculation resources to a certain extent.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application, illustrate and explain the application and are not to be construed as limiting the application. In the drawings:
FIG. 1 is a flow chart of a diffusion weighted image compressed sensing retrieval method based on an countermeasure network in accordance with an embodiment of the present invention;
FIG. 2 is a schematic illustration of an implementation flow in accordance with an alternative embodiment of the invention;
FIG. 3 is a diagram of a network model architecture according to an alternative embodiment of the present invention;
fig. 4 is a visual result of a DWI after removal of aliasing artifacts and a diffusion analysis of the same in accordance with an alternative embodiment of the invention.
Detailed Description
It should be noted that, in the case of no conflict, the embodiments and features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings in conjunction with embodiments.
In order to make the present application solution better understood by those skilled in the art, the following description will be made in detail and with reference to the accompanying drawings in the embodiments of the present application, it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, shall fall within the scope of the present application.
It should be noted that the terms "first," "second," and the like in the description and claims of the present application and the above figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate in order to describe the embodiments of the present application described herein.
It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article or apparatus that comprises an element.
The following embodiments can be applied to the technical field of medical image processing, in particular to a compressed sensing recovery method based on generation of diffusion weighted images of an countermeasure network, and can be applied to a post-processing method for denoising and artifact removal and image recovery of diffusion weighted images acquired by compressed sensing by a computer.
In this embodiment, there is provided a diffusion weighted image compressed sensing recovery method based on an countermeasure network, and fig. 1 is a flowchart of a diffusion weighted image compressed sensing recovery method based on an countermeasure network according to an embodiment of the present invention, as shown in fig. 1, the method includes the steps of:
step S102, obtaining a diffusion weighted image DWI airspace image;
step S104, undersampling the DWI airspace image to obtain an undersampled airspace image;
step S106, inputting the undersampled airspace image into a generator, and inputting the DWI airspace image into a discriminator;
in this step, the generator is coupled to the arbiter, and the generator is configured to perform image generation according to the undersampled spatial image, and input the generated image into the arbiter; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the arbiter are divided into neural network models.
Step S108, alternately training the generator and the discriminator to fight against the neural network, and obtaining network parameters corresponding to the neural network of the generator and/or the discriminator;
and step S110, recovering the undersampled DWI which needs to be recovered by using the obtained network parameters.
The above steps deal with noise and artifacts caused by compressed sensing and use deep learning technology. The image recovery method based on deep learning is characterized in that a large amount of data is utilized for learning optimization, the method can be applied to solving the problem of non-convex optimization, and the network parameters of nonlinear transformation obtained after training can rapidly complete an image recovery task. The image restoration method based on deep learning is applied to the compressed sensing restoration of the DWI.
The method solves the problem that the existing artifact and noise removing method does not have learning ability, calculates by using the neural network model, so that the method has learning type, improves the calculation efficiency and saves hardware calculation resources to a certain extent.
As an alternative embodiment, to facilitate model training and convergence, at least the DWI spatial image may be normalized to normalize the DWI spatial image data to [ -1,1] before undersampling the DWI spatial image. With this preferred embodiment, the model is more rapid in training and convergence.
As an alternative embodiment, the undersampling of the DWI spatial domain image to obtain the undersampled spatial domain image may be performed in the following steps: converting the DWI airspace image into a spatial signal; selecting undersampled templates, wherein different undersampled templates produce different forms of artifacts; processing the spatial signal using the selected undersampled templates; and converting the processed space signal into DWI to obtain the undersampled airspace image.
After recovery is performed, as an alternative embodiment, the recovered diffusion tensor imaging DTI can also be reconstructed from the DWI and analyzed for diffusion characteristics. With this embodiment, the effects of the above method can be determined in a wide variety of ways, for example, the diffusion tensor imaging DTI recovered from DWI reconstruction can include: reconstructing the recovered diffusion tensor imaging DTI from the DWI by least squares fitting; and/or analyzing the diffusion characteristics may include: calculating at least one of the following parameters from the DTI: fractional anisotropy FA, average diffusion coefficient MD, included angle error between fibers, and analysis is performed according to the calculated parameters.
In this embodiment, there is further provided a diffusion weighted image compression perception restoration device based on an antagonism network, where the device corresponds to the method steps in the above method, and each method step may be implemented by one module or may also be implemented by several modules, and in this embodiment, the device includes: the acquisition module is used for acquiring a diffusion weighted image DWI airspace image; the undersampling module is used for undersampling the DWI airspace image to obtain an undersampled airspace image; the input module is used for inputting the undersampled airspace image into a generator and inputting the DWI airspace image into a discriminator, wherein the generator is coupled with the discriminator, and is used for generating an image according to the undersampled airspace image and inputting the generated image into the discriminator; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the discriminator are divided into neural network models; the training module is used for alternately training the generator and the discriminator to fight against the neural network, so as to obtain network parameters corresponding to the neural network of the generator and/or the discriminator; and the recovery module is used for recovering the undersampled DWI which needs to be recovered by using the obtained network parameters.
Optionally, the apparatus may further include: and the processing module is used for carrying out normalization processing on at least the DWI spatial domain image and normalizing the DWI spatial domain image data to [ -1,1]. Optionally, the processing module is configured to: converting the DWI airspace image into a spatial signal; selecting undersampled templates, wherein different undersampled templates produce different forms of artifacts; processing the spatial signal using the selected undersampled templates; and converting the processed space signal into DWI to obtain the undersampled airspace image. Optionally, the apparatus may further include: and the analysis module is used for reconstructing the restored diffusion tensor imaging DTI from the DWI and analyzing the diffusion characteristic. Optionally, the analysis module is configured to reconstruct the recovered diffusion tensor imaging DTI from the DWI by least squares fitting; and/or for calculating from the DTI at least one of the following parameters: fractional anisotropy FA, average diffusion coefficient MD, included angle error between fibers, and analysis is performed according to the calculated parameters.
The modules in these devices correspond to the functions performed by the steps in the methods, and are not described in detail herein. The feature integration in the above embodiment will be described below with reference to a preferred embodiment.
Aiming at the defects of the traditional post-processing image recovery method for performing compressed sensing denoising and artifact removal based on a filter and a machine learning method, the invention aims to provide a method for performing compressed sensing recovery on a diffusion weighted image based on a deep learning method. The method can quickly recover and obtain high-quality DWI under the condition of serious aliasing artifact caused by highly undersampled DWI signals, and can better meet the requirement on the DWI in the process of analyzing diffusion characteristics.
Fig. 2 is a schematic flow chart of an implementation according to an alternative embodiment of the present invention, as shown in fig. 2, after normalization from the tag, K-space undersampling is performed, and then processing of the generator G and the arbiter D is performed, where a loss function is involved, and finally the tag/prediction is performed.
FIG. 3 is a diagram of a network model architecture according to an alternative embodiment of the present invention, and as shown in FIG. 3, cascade connection of 4x4 convolutions and multi-layer sampling are performed, and then feature extraction, feature fusion and classification judgment are performed. The following describes what is referred to in fig. 2 and 3.
The present embodiment is a compressed sensing recovery method based on generating a diffusion weighted image of an countermeasure network, which is characterized in that in the process of accelerating the acquisition of the diffusion weighted image by adopting a compressed sensing method, more serious aliasing artifacts are caused, and in order to meet the image quality requirement of analyzing diffusion characteristics through the diffusion weighted image and reconstruct high-quality diffusion tensor imaging, the technical scheme of the present invention includes the following steps: step one, initializing parameters for generating an countermeasure network for better training and faster convergence, and preprocessing input data; step two, undersampling a diffusion weighted image DWI; step three, constructing a generating countermeasure network, respectively generating a generator G and a discriminator D, training the G and the D, and estimating a nonlinear transformation parameter theta; and fourthly, rapidly recovering undersampled DWI by using the network parameters theta obtained by training, reconstructing the recovered diffusion tensor imaging DTI from the DWI by a least square fitting method, and further analyzing diffusion characteristics for evaluating the reconstruction performance of the compressed sensing recovery method based on generating the diffusion weighted image of the countermeasure network.
The discriminator D is a network with a two-class network as a framework, the last layer is a full-connection layer, a random gradient optimization method based on first-order gradient is adopted during training, and the purpose is to discriminate u and uThe DWI obtained by G prediction.
During training, a random gradient optimization method based on first-order gradient is adopted to optimize an objective function, and an optimization target is minimizedAnd u f And finally, the balance between G and D is achieved.
In the first step, the preprocessing step comprises normalizing the data and enhancing the random data, and normalizing the data to [ -1,1] is beneficial to model training and convergence.
In the second step, an undersampling process of the DWI is simulated to obtain an initial undersampled airspace image u, and the undersampling process is performed by the following steps:
(a) DWI spatial domain image u f Converting into a k-space signal x;
(b) Selecting undersampling templates, such as Gaussian sampling, poisson sampling, radial sampling and the like, wherein different sampling templates lead to DWI to generate different forms of artifacts; (c) processing the k-space signal with an undersampled template; (d) Converting the undersampled k-space signal into DWI, wherein the undersampled DWI is used as input data u; the above processing steps can be expressed as:
u＝F T (M(F(u f )))
wherein F represents Fourier transform, M represents undersampled templates, F T Representing the inverse fourier transform.
In the third step, the steps are as follows:
(a) The purpose of the generator (G) is to generate a sufficiently realistic image "spoof" arbiter (D) whose loss function can be written as:
which comprises countering the lossTo improve the reconstruction quality, content MSE losses are addedWherein i represents the ith data; u represents input data of G; />Representing the reconstructed airspace DWI; u (u) f Representing a fully sampled image, also known as a label; v denotes a frequency domain signal of the input data of G, f vfrequency domain signal representing a fully sampled image, +.>A frequency domain signal representing the reconstructed DWI; the goal of G is to minimize this loss, i.e. minimize +.>And u is equal to f Is the gap between (1);
(b) The purpose of the arbiter is to discriminate as much as possible the data generated by the generatorIf false, judge the full sampling data u f For false, its loss function can be written as:
(c) By alternate training of the generator and the arbiter, the network parameter θ of G can be determined G And network parameter θ of D D The game process can be expressed as:
wherein p is data Representing the distribution of real data, p G Representing the distribution θ of the generated data G And theta D Representing parameters related to the generator and the discriminator respectively, and finally obtaining the recovered DWI through G and D gamesAnd all network parameters θ.
In the fourth step, the step of analyzing the diffusion characteristics of the DWI includes:
(a) Calculating a diffusion vector imaging DTI by using the DWI, the bval and the bvec through least square fitting;
(b) And calculating from the DTI to obtain fractional anisotropy FA, average diffusion coefficient MD and included angle error among fibers, and evaluating the effect of the diffusion weighted image compression perception artifact removal method based on the generation of the countermeasure network through the indexes.
In the fourth step, the diffusion tensor d of DTI needs to be calculated, and the calculation formula is as follows:
d＝(H T H) -1 H T U
where H represents a diffusion gradient matrix, associated with imaging acquisition parameters. U represents DW images in different directions. From the diffusion tensor d, a number of parameters describing the tissue properties can be calculated. Comprising the following steps: fractional Anisotropy (FA), which is used to quantitatively reflect the inconsistency of water molecules in the diffusion direction, the range of which is 0-1, the average diffusion coefficient (MD), which is used to reflect the average diffusion displacement of water molecules, the included angle Angdev between fiber directions, the smaller the included angle error, the better the recovery effect, the above evaluation index formula is defined as follows:
MD＝(λ 1 +λ 2 +λ 3 )/3
wherein lambda is 1 ，λ 2 ，λ 3 The 3 eigenvalues representing the diffusion tensor,and->The principal directions of the fibers calculated from the label DWI and the predicted DWI are shown, respectively.
The following two embodiments are described in connection with practical effects.
Example 1: as shown in fig. 1, 2 and third, a method for de-artifacting a compressed perceived diffusion weighted image based on generating an countermeasure network. The method of the invention comprises the following steps:
1) Preprocessing experimental data, including: normalizing the data, normalizing the data range to [ -1,1]; random data enhancement (random overturn, random brightness change, random zoom-in and zoom-out, random noise and the like) is used in the training process, so that the robustness of the model is improved.
2) DWI spatial domain image u f Converting the k-space signal into a k-space signal x, selecting an undersampled template M, and processing the k-space signal by using the undersampled template; converting the undersampled k-space signal into DWI, wherein the undersampled DWI is used as input data u; the above processing steps can be expressed as:
u＝F T (M(F(u f )))
wherein F represents Fourier transform, M represents undersampled templates, F T Representing the inverse fourier transform.
3) Training is started, data u is put into G, u f Putting the model into a discriminator D, optimizing the model by adopting a random gradient optimization method based on a first-order gradient, wherein the loss function is as follows:
4) U is put into a generator G, G is optimized once by adopting a random gradient optimization method based on first-order gradient, and the loss function is as follows:
5) Cycling through steps 3) and 4) until the loss does not drop further within 10 epochs, then the approximation is considered to converge.
6) Calculating recovered DWIWith full sampling DWiu f The adopted evaluation function has a mean square error MSE, a peak signal to noise ratio PSNR and a structural similarity degree SSIM, and the formula is defined as follows:
where i, j represents the (i, j) th value in the image, μ x ，μ y The average of x, y,variance of x, y, sigma, respectively xy Is the covariance of x and y, c 1 ＝(k 1 L) 2 ,c 2 ＝(k 2 L) 2 Is two constants, avoid dividing 0, k 1 ＝0.01,k 2 =0.03 is a default value.
7) And calculating the DTI by least square fitting, and calculating the included angle between the FA, MD and fiber direction.
Example 2:
the scheme in the embodiment 1 is experimentally verified by combining specific DWI data, and the compressed sensing diffusion weighted image artifact removal method based on the generation of the countermeasure network is compared with the machine learning method and the deep learning method, and the specific steps are as follows:
comparison of experimental content and results:
the invention is respectively compared with a classical machine learning method ADMM-net and a newer deep learning method DAGAN. The DWI data used were from isolated infant hearts, and experiments were performed on DWI at 10%,20%,30%,40%,50% acquisition rates, respectively, with lower sampling rates causing more severe aliasing artifacts to the DWI. The three methods all adopt the same experimental data (comprising the same training set, verification set and test set) and data preprocessing methods, and the quantitative evaluation of DWI recovery effects of different methods and quantitative analysis results of errors of included angles of FAs, MDs and fibers of different methods are realized in the same hardware server and software configuration by using PSNR and SSIM. Each method is trained independently, quantitative evaluation indexes are calculated by using verification set data, and the visual result shows that the test set data are used.
The quantitative evaluation index in the verification set results were as follows:
table 1 quantitative evaluation results of comparative experiments
As can be seen from the quantitative evaluation results of the comparative tests in Table 1, the method is more prominent in PSNR and SSIM. As can be seen from the visual results of fig. 4, the method better removes aliasing artifacts and noise on DWI, and simultaneously recovers high-quality DWI, and in the subsequent analysis of diffusion characteristics, the recovery effect of the method is more prominent.
The above-described embodiments have the following advantages over the prior art:
1. the CS is utilized to accelerate DWI acquisition;
2. while quickly recovering aliasing artifacts and noise of DWI caused by the highly undersampled operation.
The embodiment adopts a compressed sensing diffusion weighted image artifact removing method based on a generated countermeasure network, adopts the generated countermeasure network to replace the traditional machine learning method to complete the task of removing the artifact and noise caused by compressed sensing on the DWI and recovering the high-quality DWI, verifies the effectiveness of the invention through the visual FA, MD and fiber included angle errors, and ensures the quality of the diffusion weighted image while accelerating the acquisition speed of the diffusion weighted image.
The above embodiment provides a compressed sensing recovery method based on generating a diffusion weighted image of an countermeasure network, which solves the problem of aliasing artifacts and noise caused by compressed sensing by adopting the method for generating the countermeasure network, replaces the existing machine learning, overcomes the defect that the machine learning method needs repeated iterative optimization and longer processing time, and compared with the newer deep learning method, the method can learn richer texture features and deeper semantic features, thereby recovering finer image details, and has better performance on DWI and following FA, MD and current direction included angle errors. The embodiment can be used as a post-processing step for accelerating the acquisition speed of diffusion nuclear magnetic resonance imaging.
The foregoing is merely exemplary of the present application and is not intended to limit the present application. Various modifications and changes may be made to the present application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. which are within the spirit and principles of the present application are intended to be included within the scope of the claims of the present application.
Claims (10)
1. A diffusion weighted image compressed sensing restoration method based on a countermeasure network, comprising:
acquiring a diffusion weighted image DWI airspace image;
undersampling the DWI airspace image to obtain an undersampled airspace image;
inputting the undersampled spatial image into a generator, inputting the DWI spatial image into a discriminator, wherein the generator is coupled with the discriminator, and is used for generating an image according to the undersampled spatial image and inputting the generated image into the discriminator; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the discriminator are divided into neural network models;
alternately training the generator and the discriminator to perform countermeasure, and obtaining network parameters corresponding to the neural network of the generator and/or the discriminator;
and recovering the undersampled DWI which needs to be recovered by using the obtained network parameters.
2. The method of claim 1, wherein prior to undersampling the DWI spatial domain image, the method further comprises:
and normalizing at least the DWI spatial domain image, and normalizing the DWI spatial domain image data to [ -1,1].
3. The method according to claim 1 or 2, wherein undersampling the DWI spatial domain image to obtain the undersampled spatial domain image comprises:
converting the DWI airspace image into a spatial signal;
selecting undersampled templates, wherein different undersampled templates produce different forms of artifacts;
processing the spatial signal using the selected undersampled templates;
and converting the processed space signal into DWI to obtain the undersampled airspace image.
4. A method according to any one of claims 1 to 3, characterized in that after recovering the undersampled DWI that needs to be recovered using the network parameters obtained, the method further comprises:
the recovered diffusion tensor is reconstructed from the DWI to image DTI and analyzed for diffusion characteristics.
5. The method of claim 4, wherein the step of determining the position of the first electrode is performed,
the diffusion tensor imaging DTI recovered from DWI reconstruction includes: reconstructing the recovered diffusion tensor imaging DTI from the DWI by least squares fitting; and/or the number of the groups of groups,
analyzing the diffusion characteristics includes: calculating at least one of the following parameters from the DTI: fractional anisotropy FA, average diffusion coefficient MD, included angle error between fibers, and analysis is performed according to the calculated parameters.
6. A diffusion weighted image compressed sensing retrieval device based on a countermeasure network, comprising:
the acquisition module is used for acquiring a diffusion weighted image DWI airspace image;
the undersampling module is used for undersampling the DWI airspace image to obtain an undersampled airspace image;
the input module is used for inputting the undersampled airspace image into a generator and inputting the DWI airspace image into a discriminator, wherein the generator is coupled with the discriminator, and is used for generating an image according to the undersampled airspace image and inputting the generated image into the discriminator; the discriminator is used for discriminating that the image generated by the generator is false and discriminating that the DWI spatial domain image is false; the generator and the discriminator are divided into neural network models;
the training module is used for alternately training the generator and the discriminator to fight against the neural network, so as to obtain network parameters corresponding to the neural network of the generator and/or the discriminator;
and the recovery module is used for recovering the undersampled DWI which needs to be recovered by using the obtained network parameters.
7. The apparatus as recited in claim 6, further comprising:
and the processing module is used for carrying out normalization processing on at least the DWI spatial domain image and normalizing the DWI spatial domain image data to [ -1,1].
8. The apparatus of claim 6 or 7, wherein the processing module is configured to:
converting the DWI airspace image into a spatial signal;
selecting undersampled templates, wherein different undersampled templates produce different forms of artifacts;
processing the spatial signal using the selected undersampled templates;
and converting the processed space signal into DWI to obtain the undersampled airspace image.
9. The apparatus according to any one of claims 6 to 8, further comprising:
and the analysis module is used for reconstructing the restored diffusion tensor imaging DTI from the DWI and analyzing the diffusion characteristic.
10. The apparatus of claim 9, wherein the device comprises a plurality of sensors,
the analysis module is used for reconstructing the restored diffusion tensor imaging DTI from the DWI through least square fitting; and/or for calculating from the DTI at least one of the following parameters: fractional anisotropy FA, average diffusion coefficient MD, included angle error between fibers, and analysis is performed according to the calculated parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011451621.3A CN112419203B (en) | 2020-12-07 | 2020-12-07 | Diffusion weighted image compressed sensing recovery method and device based on countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011451621.3A CN112419203B (en) | 2020-12-07 | 2020-12-07 | Diffusion weighted image compressed sensing recovery method and device based on countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112419203A CN112419203A (en) | 2021-02-26 |
| CN112419203B true CN112419203B (en) | 2023-07-25 |
Family
ID=74776527
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011451621.3A Active CN112419203B (en) | 2020-12-07 | 2020-12-07 | Diffusion weighted image compressed sensing recovery method and device based on countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112419203B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113052925A (en) * | 2021-04-02 | 2021-06-29 | 广东工业大学 | Compressed sensing reconstruction method and system based on deep learning |
| CN113654952B (en) * | 2021-06-28 | 2022-07-19 | 中山大学 | Diffusion coefficient measuring device and method based on right-angle triangular liquid tank |
| US20230270347A1 (en) * | 2022-02-25 | 2023-08-31 | International Business Machines Corporation | Automatic Determination of B-Values from Diffusion-Weighted Magnetic Resonance Images |
| CN114972145B (en) * | 2022-05-26 | 2024-06-28 | 厦门大学 | Diffusion tensor reconstruction method and system based on deep learning and electronic equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111077523A (en) * | 2019-12-13 | 2020-04-28 | 南京航空航天大学 | Inverse synthetic aperture radar imaging method based on generation countermeasure network |
| WO2020132463A1 (en) * | 2018-12-21 | 2020-06-25 | Cornell University | Machine learning for simultaneously optimizing an under-sampling pattern and a corresponding reconstruction model in compressive sensing |
| CN111445546A (en) * | 2020-03-03 | 2020-07-24 | 东软医疗***股份有限公司 | Image reconstruction method and device, electronic equipment and storage medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110942489B (en) * | 2018-09-25 | 2023-04-25 | 西门子医疗***有限公司 | Magnetic resonance diffusion tensor imaging method and device and fiber bundle tracking method and device |
-
2020
- 2020-12-07 CN CN202011451621.3A patent/CN112419203B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020132463A1 (en) * | 2018-12-21 | 2020-06-25 | Cornell University | Machine learning for simultaneously optimizing an under-sampling pattern and a corresponding reconstruction model in compressive sensing |
| CN111077523A (en) * | 2019-12-13 | 2020-04-28 | 南京航空航天大学 | Inverse synthetic aperture radar imaging method based on generation countermeasure network |
| CN111445546A (en) * | 2020-03-03 | 2020-07-24 | 东软医疗***股份有限公司 | Image reconstruction method and device, electronic equipment and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 卷积神经网络重建欠采的磁共振图像;王一达;宋阳;谢海滨;童睿;李建奇;杨光;;磁共振成像(06);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112419203A (en) | 2021-02-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112419203B (en) | Diffusion weighted image compressed sensing recovery method and device based on countermeasure network | |
| Yuan et al. | Fast hyperspectral anomaly detection via high-order 2-D crossing filter | |
| Zhang et al. | Joint image denoising using adaptive principal component analysis and self-similarity | |
| CN104574298B (en) | A kind of noise-reduction method of more b values diffusion weightings images based on mutual information | |
| CN107274462B (en) | Classified multi-dictionary learning magnetic resonance image reconstruction method based on entropy and geometric direction | |
| Li et al. | No-reference quality assessment of deblurred images based on natural scene statistics | |
| CN114359526B (en) | Cross-domain image style migration method based on semantic GAN | |
| CN110599519A (en) | Anti-occlusion related filtering tracking method based on domain search strategy | |
| CN103699874B (en) | Crowd abnormal behavior identification method based on SURF (Speed-Up Robust Feature) stream and LLE (Locally Linear Embedding) sparse representation | |
| CN103279744B (en) | Imitation fingerprint detection methods based on multiple dimensioned three mode texture feature and system | |
| CN108932705A (en) | A kind of image processing method based on matrix variables variation self-encoding encoder | |
| CN106096505B (en) | The SAR target identification method indicated based on Analysis On Multi-scale Features collaboration | |
| CN113592769B (en) | Abnormal image detection and model training method, device, equipment and medium | |
| CN112836602B (en) | Behavior recognition method, device, equipment and medium based on space-time feature fusion | |
| WO2014162300A1 (en) | Acceleration of low-rank mri data acquisition | |
| EP2531956B1 (en) | Image recognition system based on cascaded over-complete dictionaries | |
| CN113865859A (en) | Multi-scale multi-source heterogeneous information fusion gearbox state fault diagnosis method | |
| CN107944497A (en) | Image block method for measuring similarity based on principal component analysis | |
| CN112270255A (en) | Electroencephalogram signal identification method and device, electronic equipment and storage medium | |
| CN115578262A (en) | Polarization image super-resolution reconstruction method based on AFAN model | |
| Tripathi | Facial image noise classification and denoising using neural network | |
| CN106940904B (en) | Attendance checking system based on recognition of face and speech recognition | |
| CN116595465A (en) | High-dimensional sparse data outlier detection method and system based on self-encoder and data enhancement | |
| CN113592972B (en) | Magnetic resonance image reconstruction method and device based on multi-mode aggregation | |
| Babacan et al. | Reference-guided sparsifying transform design for compressive sensing MRI |
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 |