CN110728727B - Low-dose energy spectrum CT projection data recovery method - Google Patents
Low-dose energy spectrum CT projection data recovery method Download PDFInfo
- Publication number
- CN110728727B CN110728727B CN201910827015.8A CN201910827015A CN110728727B CN 110728727 B CN110728727 B CN 110728727B CN 201910827015 A CN201910827015 A CN 201910827015A CN 110728727 B CN110728727 B CN 110728727B
- Authority
- CN
- China
- Prior art keywords
- domain
- projection
- network
- image
- data
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/003—Reconstruction from projections, e.g. tomography
-
- 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
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)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention relates to a method for recovering low-dose energy spectrum CT projection data, which is mainly technically characterized by comprising the following steps of: constructing a projection domain and image domain dual-domain generation countermeasure network framework, obtaining projection domain and image domain dual-domain combined distribution through the coupling of two generation network models, inputting complete virtual single-energy projection data and virtual single-energy images into two discrimination network models, and obtaining a low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution through the discrimination network models; and training the dual-domain generation countermeasure network framework to obtain a low-dose energy spectrum CT projection domain data recovery model based on dual-domain combined distribution, and finally obtaining a high-quality low-dose energy spectrum CT reconstruction single-energy image with beam hardening artifact eliminated. The invention has reasonable design, effectively eliminates the serious beam hardening artifact caused by the reduction of X-ray dose and beam hardening effect in the low-dose energy spectrum CT and greatly improves the final image reconstruction quality.
Description
Technical Field
The invention belongs to the technical field of computed tomography, and particularly relates to a method for recovering low-dose energy spectrum CT projection data.
Background
The X-ray Computed Tomography (CT) technology has been widely used in the fields of industrial detection and medical diagnosis, but the CT examination is always troubled by artifacts, which bring great difficulty to clinical examination and diagnosis. Meanwhile, it is not negligible that the higher radiation dose of the X-ray in the CT examination may have adverse effect on the human health, so one of the important directions for the development of CT is to reduce the radiation dose to the human body during CT scanning, and reducing the voltage of the X-ray tube is one of the main approaches of low dose, but this method may cause the reduction of the ray penetration capability, resulting in the generation of serious beam hardening artifacts in the reconstructed image.
Beam hardening artifacts caused by the pleochromism of the actual X-ray energy spectrum have a significant effect on the quality of the reconstructed slice. The energy of the radiation generated by the actual X-ray source is not single and has the characteristic of continuous distribution of energy spectrum. When X-rays penetrate through an object to interact with the object, photons with lower energy are easily absorbed by the substance, and the attenuation coefficient is reflected to be higher under low energy and lower under high energy, so that the energy spectrum composition of the rays penetrating through the object is changed. The mean energy of the X-rays after passing through the object is higher than the mean energy before not passing through the object, a phenomenon which we call "beam hardening". The commonly used CT reconstruction filtered back projection algorithm (FBP) is based on an assumption: x-rays are monoenergetic, which causes certain artifacts to appear in the reconstructed image. The artifacts present in the tomographic images directly reconstructed by means of multi-energy projection are called beam hardening artifacts, and the beam hardening artifacts present in the CT images obtained under the low-dose CT condition are more serious.
In recent years, spectral CT with a higher ability to distinguish substances has become a hot spot of current research. The energy spectrum CT can distinguish multiple X-ray photons with different energies, can provide more image information than the conventional CT by utilizing different absorption of substances generated under different X-ray energies, and carries out image reconstruction by utilizing the absorption difference of the substances on the X-ray photons with different energies to obtain a CT reconstructed image under a multi-energy section.
One of the clinically important applications of spectral CT is the virtual single-energy imaging function of the spectral CT technology, the spectral CT can fully acquire and utilize the energy information of the scanned material, and the combination of high-energy and low-energy images of the spectral CT can theoretically acquire a CT image obtained by scanning with any single-energy X-ray source. If the X-ray source is mono-energetic, then there will be no beam hardening artifacts in the scanned CT images, which provides a solution for eliminating the severe beam hardening artifacts present in low dose spectral CT images: the advantage that material information under a plurality of energy sections of a scanned object can be obtained by utilizing the energy spectrum CT is utilized to obtain a virtual single energy map so as to eliminate the problem of serious beam hardening artifacts existing in the low-dose energy spectrum CT.
The beam hardening effect caused by high-density substances existing in a scanned region can actually cause the defect of energy spectrum CT projection domain data, due to the fact that energy spectrum imaging is a projection data space reconstruction method, the defect of projection data can cause considerable influence on subsequent image reconstruction and diagnosis work, and the defect of projection data is completely supplemented, so that serious beam hardening artifacts existing in an image can be effectively eliminated in the subsequent image reconstruction process; it is very difficult to accurately and effectively eliminate artifacts due to beam hardening effects and to reconstruct high quality spectral CT images purely in the image domain of spectral CT.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a method for recovering low-dose energy spectrum CT projection data, which utilizes a depth generation deconvolution neural network to realize the recovery of the low-dose energy spectrum CT projection data and completely eliminates beam hardening artifacts and metal artifacts existing in CT images by supplementing the energy spectrum CT projection data.
The technical problem to be solved by the invention is realized by adopting the following technical scheme:
a method for recovering low-dose energy spectrum CT projection data comprises the following steps:
step 1, acquiring a high-energy segment projection data set P N And a set of low energy segment projection data P L ；
Step 2, carrying out filtering back projection transformation reconstruction on the projection data to obtain a high-energy CT image data set I N And a low energy CT image dataset I L ；
Step 3, constructing a projection domain and image domain double-domain generation countermeasure network framework: the dual-domain generation countermeasure network framework comprises a pair of generation network models and a pair of discrimination network models, wherein the pair of generation network models is; a projection domain-based generation network model P-Gen and an image domain-based generation network model I-Gen, and a pair of discrimination network models is as follows: first discrimination network model D 1 And a second discrimination network model D 2 ；
Step 4, collecting the high-energy segment projection data P N And a set of low energy segment projection data P L Respectively as the input of two generated network models;
and 5: learning the joint distribution of the energy spectrum CT double-domain data through the coupling of two generation network models, and forcibly generating the same weight shared by a plurality of layers before the network model to obtain the double-domain joint distribution P (P, G) of a projection domain and an image domain;
and 6: the last layer of the two generated network models maps the dual-domain joint distribution P (P, G) to respective domain to obtain complete virtual monoenergetic projection data
And a virtual monoenergetic image @, which is used as an auxiliary contrast>
And 7: complete virtual monoenergetic projection data
And virtual monoenergetic projection data labels x 1 Inputting a first discrimination network model D 1 Combining the virtual mono-energetic image>
And virtual monoenergetic image tag x 2 Inputting a second discrimination network model D 2 Is provided with f 1 And f 2 The mapping relations in the two discrimination network models are respectively:
wherein f is 1 (i) And
is D 1 And D 2 Layer i of (2), N 1 And N 2 Is D 1 And D 2 Each discriminative network model maps the input to a probability value: />
Estimating the likelihood that the input is from a true data distribution;
step 8, as long as any one discrimination network model gives a discrimination result that the output of the generated model is false, returning to the step 1, continuing training and iteration on the whole generated network until the discrimination network model can not be generated in a distinguishing way any more, and obtaining a low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution;
step 9, training the dual-domain generated confrontation network framework to obtain a low-dose energy spectrum CT projection domain data recovery model based on dual-domain joint distribution;
step 10, inputting real low-dose energy spectrum CT projection data and corresponding FBP reconstruction images into a model for testing based on a low-dose energy spectrum CT projection domain data recovery model with double-domain combined distribution to obtain fully supplemented virtual single-energy projection domain data, performing FBP reconstruction on the fully supplemented virtual single-energy projection domain data to obtain an energy spectrum CT virtual single-energy image with beam hardening artifact eliminated, performing peak signal-to-noise ratio and structural similarity quantitative comparison analysis on the energy spectrum CT virtual single-energy image and the original CT image, changing the iteration times of network training in step 8 according to a test result, and generating a network layer number M 1 And M 2 The number k of layers shared by the execution weights of the P-Gen and the I-Gen, the size of a convolution kernel, the number of judgment network layers and the data recovery capability of the model are continuously enhanced, finally, a good low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution is obtained, and then real projection domain test data P are obtained N 、P L After inputting, a high-quality low-dose energy spectrum CT reconstruction single-energy image with the beam hardening artifact eliminated is finally obtained.
Further, the generated network model P-Gen and the generated network model I-Gen are both realized in the form of a multilayer convolutional neural network and are respectively represented as:
wherein, the first and the second end of the pipe are connected with each other,
is the complete virtual monoenergetic projection data, generated by the P-Gen>
Is a virtual monoenergetic image generated by I-Gen; />
And &>
Is layer I of P-Gen and I-Gen, M 1 And M 2 Is the number of layers of P-Gen and I-Gen, g 1 And g 2 Is a mapping relationship in the generative models of P-Gen and I-Gen.
Further, the specific implementation method of step 9 is as follows: in the process of training a framework of a double-domain generation countermeasure network, firstly, a small amount of data sets are utilized to initially train the network to obtain an initial low-dose energy spectrum CT projection domain data recovery model based on double-domain joint distribution, the effect of the model is tested and observed, then, the complete data sets are utilized to carry out large-scale iterative training on the network, whether the loss function V of the framework basically reaches the minimum value after a plurality of times of alternate iterative training is judged, if yes, the network is converged, and an incomplete projection CT spatial information data recovery model is obtained; changing iteration times and generating network layer number M 1 And M 2 The number k of layers shared by the P-Gen and the I-Gen execution weight, the size of a convolution kernel and the number of layers of a judgment network are calculated to obtain the low-dose energy spectrum CT (computed tomography) projection based on the double-domain joint distributionShadow domain data recovery model.
The invention has the advantages and positive effects that:
the invention adopts a depth generation antagonistic neural network to learn the complex data distribution of an energy spectrum CT projection domain and an image domain and obtain the joint distribution of the complex data distribution, utilizes a GAN weight sharing strategy to combine the structural information of a scanned object displayed by the image domain into the projection domain as rich prior knowledge, realizes the enhancement and the completeness of projection domain data, obtains virtual single-energy projection domain data without beam hardening artifacts, eliminates the serious beam hardening artifacts caused by the reduction of X-ray dose and beam hardening effect in low-dose energy spectrum CT, further realizes the reconstruction function of high-quality energy spectrum CT images under the condition of low dose, and greatly improves the final image reconstruction quality.
Drawings
FIG. 1 is a diagram of a dual-domain generative countermeasure network framework of the present invention;
fig. 2 is a schematic diagram of a training set data preparation process.
Detailed Description
The embodiments of the present invention will be described in detail with reference to the accompanying drawings.
The invention realizes the projection data recovery function of low-dose energy spectrum CT by using a depth-generated antagonistic convolutional neural network (GAN), and completely eliminates beam hardening artifacts and metal artifacts existing in CT images by supplementing energy spectrum CT projection data. The design idea is as follows:
an important advantage of the deep convolutional neural network in deep learning is to extract information layer by layer from raw data to abstract semantic concepts, which makes it a prominent advantage in extracting global features of data. There are two basic models for traditional deep learning: and generating a model and a discrimination model. The deep generation deconvolution-resistant neural network unifies the two models in a framework, the two models play games with each other, alternate iterative training is carried out, and a complex statistical distribution rule of input data is learned. The projection domain and the image domain of the energy spectrum CT present complex data distribution, and the data distribution of the energy spectrum CT projection domain and the corresponding image domain is different, so that a low-dose energy spectrum CT projection data recovery- 'projection domain & image domain' double-domain generation confrontation network framework is established, a generation network weight sharing strategy is adopted, the joint distribution of the two data domains is learned, the GAN also has strong texture detail learning capacity, and in the generation network iterative learning process, the network output effectively retains details in the data distribution.
Because of the serious beam hardening artifact existing in the low-dose energy spectrum CT image, the incomplete (caused by the beam hardening effect) projection data of the scanned object is completely supplemented and enhanced in the projection domain of the energy spectrum CT, and the serious beam hardening artifact existing in the image can be effectively eliminated in the subsequent image reconstruction process. And generating an antagonistic neural network based on depth, analyzing the complex data distribution of a projection domain and an image domain by using a weight sharing strategy, combining the structural information of the scanned object displayed by the image domain into the projection domain as prior knowledge, realizing the recovery and enhancement of the projection domain data, and obtaining complete virtual monoenergetic projection domain data so as to greatly improve the final image reconstruction quality.
Based on the design concept, the method for recovering the low-dose energy spectrum CT projection data comprises the following steps:
step 1, data set preparation: high energy band projection data set P N And a set of low energy segment projection data P L 。
Step 2, carrying out filtering back projection transformation reconstruction on the projection data to obtain a high-energy CT image data set I N Low energy CT image dataset I L Severe beam hardening artifacts may be present in these images.
And 3, constructing a projection domain and an image domain to generate a confrontation network framework.
As shown in FIG. 1, the dual-domain generative countermeasure network framework comprises a pair of generative network models: projection domain based generative model (P-Gen)&Image-domain-based generative model (I-Gen) which can be used for learning and analyzing, respectively, data distribution p of spectral CT projection domain (P) And data distribution p of image domain (I) (ii) a And a pair of discriminant models D 1 And D 2 . The above generation network and judgmentThe other networks are deep convolutional neural networks.
And 4, step 4: p N 、P L As input to the P-Gen network, I N 、I L As input to the I-Gen network. Let g 1 And g 2 Is a mapping relation in the generation models of P-Gen and I-Gen. Both P-Gen and I-Gen are implemented in the form of a multi-layer convolutional neural network, which can be expressed as:
wherein
Is the complete virtual monoenergetic projection data, generated by the P-Gen>
Is a virtual monoenergetic image generated by the I-Gen; />
And &>
Is layer I of P-Gen and I-Gen, M 1 And M 2 Is the number of layers of P-Gen and I-Gen, note that M 1 And M 2 Not necessarily equal.
And 5: through the coupling of the P-Gen and the I-Gen double generation network, the joint distribution of the energy spectrum CT double-domain data is learned: force the first several layers of P-Gen and I-Gen (layers that decode high level semantics) to share the same weight, namely:
wherein
And/or>
Is->
And &>
K is the number of layers implementing the weight sharing constraint that forces the generation network to decode high level semantics in the same way, resulting in a two-domain joint distribution P (P, G) of projection domain and image domain.
Step 6: the weight sharing constraint is not executed on the last layer of the two generation networks, and the last layer of the generation networks maps the dual-domain joint distribution P (P, G) to the respective domain, that is, the final generation result is: complete virtual monoenergetic projection data:
and a virtual monoenergetic image used as an auxiliary contrast>
/>
And 7: P-Gen generated complete virtual monoenergetic projection data:
and virtual monoenergetic projection data labels x 1 Input discrimination network D 1 Virtual monoenergetic image generated by I-Gen>
And virtual monoenergetic image tag x 2 Input discrimination network D 2 . Let f 1 And f 2 The mapping relations in the two discriminant models are respectively:
wherein f is 1 (i) And
is D 1 And D 2 Layer i of (2), N 1 And N 2 Is D 1 And D 2 The number of layers of (a). Each discriminant model maps the input to a probability value: />
The likelihood that the input is from a true data distribution is estimated.
And 8, returning to the generation network as long as any discrimination model gives a discrimination result that the output of the generation model is false, namely returning to the step 1, and continuing training and iteration on the whole generation network until the discrimination model can not be generated distinguishably any more. The generation of the network:
with respective labels x 1 、x 2 Whether a difference exists between the two, and then a low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution is obtained. The corresponding dual domain of the process generates a loss function against the network framework:
V(g 1 ,g 2 ,f 1 ,f 2 )＝V 1 (g 1 ,f 1 )+V 2 (g 2 ,f 2 )
the net training end result is to minimize V:
Wherein, V 1 (g 1 ,f 1 ) Is a generating network P-Gen and a discriminating network D 1 Of (2) a antagonism loss function, V 2 (g 2 ,f 2 ) Is a generating network I-Gen and a discriminating network D 2 In the calculation of the loss function, the label x 1 、x 2 Distribution is obeyed separately: p (x) 1 )、p(x 2 ). The generation network and the discrimination network are updated in an alternating gradient mode and are propagated reversely to train the elimination of the beam hardening artifact in the low-dose energy spectrum CT projection domain, namely a dual-domain generation countermeasure network framework.
And step 9: in the process of training a framework of a double-domain generation countermeasure network, firstly, a small amount of data sets are utilized to initially train the network, an initial low-dose energy spectrum CT projection domain data recovery model based on double-domain joint distribution is obtained, the effect of the model is tested and observed, then, the complete data sets are utilized to carry out large-scale iterative training on the network, whether the loss function V of the framework basically reaches the minimum value after a plurality of times of alternate iterative training is judged, if yes, the network is converged, and an incomplete projection CT spatial information data recovery model is obtained. Changing iteration times and generating network layer number M 1 And M 2 And the number k of layers shared by the P-Gen and the I-Gen execution weight, the size of a convolution kernel, the number of layers of a judgment network and other super parameters further influence the result of the training model.
Step 10, based on the low-dose energy spectrum CT projection domain data recovery model based on the double-domain combined distribution obtained by training in step 9, inputting the real low-dose energy spectrum CT projection data and the corresponding FBP reconstruction image into the model for testing to obtain the completely supplemented virtual mono-energy projection domain data, carrying out FBP reconstruction on the virtual mono-energy projection domain data to obtain an energy spectrum CT virtual mono-energy map with beam hardening artifacts eliminated, carrying out peak signal-to-noise ratio (PSNR) and Structure Similarity (SSIM) quantitative comparison analysis on the virtual mono-energy map and the original CT image with serious beam hardening artifacts, changing the iteration times of network training in step 8 according to the test result, and generating a network layer number M 1 And M 2 K number of layers for performing weight sharing between P-Gen and I-Gen, size of convolution kernel, number of layers of discriminant network, and the likeAnd (4) continuously enhancing the data recovery capability of the model. Finally, a good low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution is obtained, and real projection domain test data P is obtained N 、P L After inputting, a high-quality low-dose energy spectrum CT reconstruction single-energy image with good detail retention and without beam hardening artifacts can be finally obtained.
Examples
The description is made according to an example of a generation-based countermeasure network (GAN) framework given in fig. 1. The generation networks P-Gen and I-Gen and the discrimination network D of the dual-domain generation countermeasure network framework 1 、D 2 The network depth of (2) is 5, that is, the network comprises 5 convolutional layers. In the network structure adopted by the people, the size of each layer of convolution kernel of the generated network is fixed to be 3 multiplied by 3, the number of the convolution kernels is fixed to be 32, the movement step length of the convolution kernels is 1, and the same padding mode is selected to ensure that the size of an image output after convolution and an input image is kept one system, so that each convolution layer can output 32 256 multiplied by 256 characteristic images to enter a ReLU activation function; the convolution kernel size of the last convolution layer is 3 x 3, the number of convolution kernels is 1, the step length is 1, the convolution kernel size is used for outputting a reconstructed image, a 256 x 1 image is still output in the same padding mode, and at the moment, virtual mono-energy projection data and a virtual mono-energy image are generated.
Preparing data set as shown in FIG. 2, using human body model (high density matter exists inside), performing X-ray parallel light scanning projection, realizing energy spectrum CT low dose through low tube voltage, using layered detector model to perform simulated photoelectric conversion, wherein the high energy interval is (60keV, 80keV), the low energy interval is (20keV, 40keV), obtaining low dose energy spectrum CT high and low energy projection data, P N 、P L Carrying out FBP reconstruction on the projection data to obtain a CT image I of the chest part region of the human body N 、I L A total of 100 sheets were used to construct P-Gen and I-Gen entries. The labels input into the discrimination network together with the generation result of the generated network are 100 pieces of virtual mono-energy projection data and 100 pieces of virtual mono-energy CT images.
Inputting a training set into a generated countermeasure network for training, observing whether the loss function can be converged to a minimum value, if not, changing the learning rate in the network, and re-training after the initial value of the number of the convolution kernels is changed until the loss function is converged. And finally, testing an incomplete projection CT spatial information data recovery model by using 20 groups of testing sample units of the human chest (different from the training set), and detecting the artifact elimination performance of the framework by adopting image quality evaluation indexes PSNR and SSIM. By the processing process, beam hardening artifacts and metal artifacts existing in the CT image can be effectively eliminated, and the final image reconstruction quality is improved.
Nothing in this specification is said to apply to the prior art.
It should be emphasized that the embodiments described herein are illustrative rather than restrictive, and thus the present invention is not limited to the embodiments described in the detailed description, but also includes other embodiments that can be derived from the technical solutions of the present invention by those skilled in the art.
Claims (3)
1. A method for recovering low-dose energy spectrum CT projection data is characterized by comprising the following steps:
step 1, acquiring a high-energy segment projection data set P N And a set of low energy segment projection data P L ；
Step 2, carrying out filtering back projection transformation reconstruction on the projection data to obtain a high-energy CT image data set I N And a low energy CT image dataset I L ；
Step 3, constructing a projection domain and image domain double-domain generation countermeasure network framework: the dual-domain generation countermeasure network framework comprises a pair of generation network models and a pair of discrimination network models, wherein the pair of generation network models is; a generating network model P-Gen based on a projection domain and a generating network model I-Gen based on an image domain, and a pair of judging network models are as follows: first discrimination network model D 1 And a second discrimination network model D 2 ；
Step 4, collecting the high-energy segment projection data P N And a set of low energy segment projection data P L Are respectively provided withAs inputs to two generative network models;
and 5: learning the joint distribution of the energy spectrum CT double-domain data through the coupling of two generation network models, and forcibly generating the same weight shared by a plurality of layers before the network model to obtain the projection domain and image domain double-domain joint distribution P (P, G);
step 6: the last layer of the two generated network models maps the dual-domain joint distribution P (P, G) to respective domain to obtain complete virtual single-energy projection data
And a virtual monoenergetic image @, which is used as an auxiliary contrast>
And 7: complete virtual monoenergetic projection data
And virtual monoenergetic projection data labels x 1 Inputting a first discrimination network model D 1 Combining the virtual mono-energetic image>
And virtual monoenergetic image tag x 2 Inputting a second discrimination network model D 2 Is provided with f 1 And f 2 The mapping relations in the two discrimination network models are respectively:
wherein f is 1 (i) And
is D 1 And D 2 Layer i of (2), N 1 And N 2 Is D 1 And D 2 Each discriminative network model maps the input to a probability value: />
Estimating the likelihood that the input is from a true data distribution;
step 8, as long as any one discrimination network model gives a discrimination result that the output of the generated model is false, returning to the step 1, continuing training and iteration of the whole generated network until the discrimination network model can not be generated in a distinguishing way any more, and obtaining a low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution;
step 9, training the dual-domain generation countermeasure network framework to obtain a low-dose energy spectrum CT projection domain data recovery model based on dual-domain combined distribution;
step 10, a low-dose energy spectrum CT projection domain data recovery model based on double-domain combined distribution inputs real low-dose energy spectrum CT projection data and a corresponding FBP reconstruction image into the model for testing to obtain virtual monoenergetic projection domain data which are completely supplemented, carries out FBP reconstruction to obtain an energy spectrum CT virtual monoenergetic image for eliminating beam hardening artifacts, carries out peak signal-to-noise ratio and structural similarity quantitative comparison analysis on the energy spectrum CT virtual monoenergetic image and the original CT image, changes the iteration times of network training in step 8 according to a test result, and generates a network layer number M 1 And M 2 The number k of layers shared by the P-Gen and the I-Gen execution weight, the size of a convolution kernel, the number of layers of a discrimination network and the continuous enhancement of the data recovery capability of the model finally obtain a good low-dose energy spectrum CT projection domain data recovery model based on projection domain-image domain combined distribution, and then test data P of a real projection domain N 、P L After inputting, a high-quality low-dose energy spectrum CT reconstruction single-energy image with the beam hardening artifact eliminated is finally obtained.
2. A method of recovering low dose spectral CT projection data as recited in claim 1, wherein: the generation network model P-Gen and the generation network model I-Gen are both realized in the form of multilayer convolutional neural networks and are respectively expressed as:
wherein the content of the first and second substances,
is the complete virtual monoenergetic projection data, generated by the P-Gen>
Is a virtual monoenergetic image generated by the I-Gen; />
And &>
Is the I-th layer of P-Gen and I-Gen, M 1 And M 2 Is the number of layers of P-Gen and I-Gen, g 1 And g 2 Is a mapping relationship in the generative models of P-Gen and I-Gen.
3. A method for recovering low dose spectral CT projection data as claimed in claim 1, wherein: the specific implementation method of the step 9 is as follows: in the process of training a framework of a dual-domain generation confrontation network, firstly, a small amount of data sets are utilized to initially train the network, an initial low-dose energy spectrum CT projection domain data recovery model based on dual-domain joint distribution is obtained, the effect of the model is tested and observed, then, the complete data sets are utilized to carry out large-scale iterative training on the network, and a loss function V of the framework is judged to be trained through a plurality of times of alternate iterative trainingWhether the data reaches the minimum value after practice or not is judged, if so, network convergence is shown, and an incomplete projection CT spatial information data recovery model is obtained; changing iteration times and generating network layer number M 1 And M 2 And the number k of layers shared by the P-Gen and the I-Gen execution weight, the size of a convolution kernel and the number of layers of a judgment network are obtained, and finally, a low-dose energy spectrum CT projection domain data recovery model based on double-domain joint distribution is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910827015.8A CN110728727B (en) | 2019-09-03 | 2019-09-03 | Low-dose energy spectrum CT projection data recovery method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910827015.8A CN110728727B (en) | 2019-09-03 | 2019-09-03 | Low-dose energy spectrum CT projection data recovery method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110728727A CN110728727A (en) | 2020-01-24 |
| CN110728727B true CN110728727B (en) | 2023-04-18 |
Family
ID=69218933
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910827015.8A Active CN110728727B (en) | 2019-09-03 | 2019-09-03 | Low-dose energy spectrum CT projection data recovery method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110728727B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325809B (en) * | 2020-02-07 | 2021-03-12 | 广东工业大学 | Appearance image generation method based on double-impedance network |
| CN111489404B (en) * | 2020-03-20 | 2023-09-05 | 深圳先进技术研究院 | Image reconstruction method, image processing device and device with storage function |
| WO2021189383A1 (en) * | 2020-03-26 | 2021-09-30 | 深圳先进技术研究院 | Training and generation methods for generating high-energy ct image model, device, and storage medium |
| CN112001978B (en) * | 2020-07-28 | 2023-09-01 | 中国人民解放军战略支援部队信息工程大学 | Method and device for reconstructing image based on dual-energy dual-90-degree CT scanning of generating countermeasure network |
| CN112037146B (en) * | 2020-09-02 | 2023-12-22 | 广州海兆印丰信息科技有限公司 | Automatic correction method and device for medical image artifacts and computer equipment |
| CN112508808B (en) * | 2020-11-26 | 2023-08-01 | 中国人民解放军战略支援部队信息工程大学 | CT double-domain combined metal artifact correction method based on generation countermeasure network |
| CN112435309A (en) * | 2020-12-07 | 2021-03-02 | 苏州深透智能科技有限公司 | Method for enhancing quality and resolution of CT image based on deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108898642A (en) * | 2018-06-01 | 2018-11-27 | 安徽工程大学 | A kind of sparse angular CT imaging method based on convolutional neural networks |
| CN109272472A (en) * | 2018-10-15 | 2019-01-25 | 天津大学 | Noise and artifact eliminating method towards medical power spectrum CT image |
| EP3447721A1 (en) * | 2017-08-24 | 2019-02-27 | Agfa Nv | A method of generating an enhanced tomographic image of an object |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180018757A1 (en) * | 2016-07-13 | 2018-01-18 | Kenji Suzuki | Transforming projection data in tomography by means of machine learning |
-
2019
- 2019-09-03 CN CN201910827015.8A patent/CN110728727B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3447721A1 (en) * | 2017-08-24 | 2019-02-27 | Agfa Nv | A method of generating an enhanced tomographic image of an object |
| CN108898642A (en) * | 2018-06-01 | 2018-11-27 | 安徽工程大学 | A kind of sparse angular CT imaging method based on convolutional neural networks |
| CN109272472A (en) * | 2018-10-15 | 2019-01-25 | 天津大学 | Noise and artifact eliminating method towards medical power spectrum CT image |
Non-Patent Citations (2)
| Title |
|---|
| 史再峰 ; 许泽昊 ; 庞科 ; 曹清洁 ; 姚素英 ; .一种基于开关型加权中值滤波的随机脉冲噪声去除方法.南开大学学报(自然科学版).2018,(第04期),全文. * |
| 满晨龙 ; 史再峰 ; 徐江涛 ; 姚素英 ; .基于区域分割的快速随机喷洒Retinex方法.南开大学学报(自然科学版).2017,(第02期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110728727A (en) | 2020-01-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110728727B (en) | Low-dose energy spectrum CT projection data recovery method | |
| Gong et al. | A computer simulation study comparing lesion detection accuracy with digital mammography, breast tomosynthesis, and cone‐beam CT breast imaging | |
| Dong et al. | Combined iterative reconstruction and image‐domain decomposition for dual energy CT using total‐variation regularization | |
| Wu et al. | Deep learning based spectral CT imaging | |
| Zeng et al. | Spectral CT image restoration via an average image-induced nonlocal means filter | |
| WO2020237873A1 (en) | Neural network-based spiral ct image reconstruction method and device, and storage medium | |
| CN104751429B (en) | A kind of low dosage power spectrum CT image processing methods based on dictionary learning | |
| Shi et al. | A material decomposition method for dual‐energy CT via dual interactive Wasserstein generative adversarial networks | |
| US20180182129A1 (en) | Spectral ct image reconstructing method and spectral ct imaging system | |
| Su et al. | DIRECT‐Net: A unified mutual‐domain material decomposition network for quantitative dual‐energy CT imaging | |
| Shen et al. | Multienergy CT acquisition and reconstruction with a stepped tube potential scan | |
| Xue et al. | A 3D attention residual encoder–decoder least-square GAN for low-count PET denoising | |
| CN109741254A (en) | Dictionary training and Image Super-resolution Reconstruction method, system, equipment and storage medium | |
| Feng et al. | Dual residual convolutional neural network (DRCNN) for low-dose CT imaging | |
| Niu et al. | Iterative reconstruction for photon-counting CT using prior image constrained total generalized variation | |
| WO2023000244A1 (en) | Image processing method and system, and application of image processing method | |
| Clark et al. | Deep learning based spectral extrapolation for dual‐source, dual‐energy x‐ray computed tomography | |
| Cong et al. | Monochromatic image reconstruction via machine learning | |
| Puong et al. | Dual-energy contrast enhanced digital breast tomosynthesis: concept, method, and evaluation on phantoms | |
| Guo et al. | Material decomposition of spectral CT images via attention-based global convolutional generative adversarial network | |
| Xia et al. | Dynamic controllable residual generative adversarial network for low-dose computed tomography imaging | |
| Sheikhi et al. | Deep-learned generation of renal dual-energy CT from a single-energy scan | |
| Gang et al. | Three-material decomposition using a dual-layer flat panel detector in the presence of soft tissue motion | |
| CN116542868A (en) | Method for inhibiting X-chest rib bone image based on attention generation countermeasure network | |
| Chen et al. | CT-Net: Cascaded T-shape network using spectral redundancy for dual-energy CT limited-angle reconstruction |
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 |