CN112669400B - Dynamic MR reconstruction method based on deep learning prediction and residual error framework - Google Patents
Dynamic MR reconstruction method based on deep learning prediction and residual error framework Download PDFInfo
- Publication number
- CN112669400B CN112669400B CN202011445536.6A CN202011445536A CN112669400B CN 112669400 B CN112669400 B CN 112669400B CN 202011445536 A CN202011445536 A CN 202011445536A CN 112669400 B CN112669400 B CN 112669400B
- Authority
- CN
- China
- Prior art keywords
- signal
- dynamic
- residual
- layer
- space
- 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
Abstract
The application belongs to the technical field of imaging, and provides a dynamic MR reconstruction method based on deep learning prediction and residual error architecture, which comprises the following steps: and inputting the dynamic MR undersampled data of the K space into a neural network model for processing to obtain a reconstructed dynamic MR image. According to the method, dynamic MR undersampled data are predicted and residual error separated through a residual error learning network module, and sparse coding is performed on residual error signals through a networked contraction soft threshold algorithm module, so that a high-quality dynamic MR image is reconstructed. Compared with the traditional reconstruction method, the method greatly shortens the reconstruction time of the dynamic MR image, can learn some hyper-parameters in a self-adaptive manner without manual selection, and can solve the technical problem that the hyper-parameters are difficult to select in the traditional reconstruction method.
Description
Technical Field
The application belongs to the technical field of imaging, and particularly relates to a dynamic MR reconstruction method based on deep learning prediction and a residual error framework.
Background
Magnetic resonance imaging is a non-invasive imaging technique with good spatial resolution and soft tissue contrast, and is widely used in clinical diagnosis and research, such as monitoring the motion of the heart. Dynamic magnetic resonance imaging attempts to reveal the spatiotemporal distribution of the underlying anatomy and has wide application in the fields of cardiovascular imaging and the like. At present, the cardiac cine imaging technology applied clinically needs a patient to hold a breath for matching during scanning, the overlong scanning time hardly ensures that the patient keeps still during the scanning process, and the whole scanning process is done to cause fatigue of the patient. The long acquisition time is not only a burden, but also makes the mri sensitive to motion artifacts. Therefore, the clinical urgent need exists for a cardiac magnetic resonance imaging method with short scanning time and easy patient coordination.
For many years researchers have proposed a number of methods to accelerate the reconstruction of dynamic magnetic resonance images. Wherein a k-t space based on FOCal underlying system (k-t FOCUSS) is a typical example of exploiting the space-time redundancy of data, which recovers the signal by enhancing the sparsity of the signal in the x-f domain under the framework of compressed sensing. A low rank sparse reconstruction scheme exploits the correlation between voxel temporal distributions by introducing a non-convex spectral norm and a spatio-temporal fully-variant fractional norm. Despite the breakthrough advances made by the above reconstruction methods, these optimization-based methods also face several challenges: first, the regularization function and its parameters, which are problem-specific and difficult to select, must be carefully chosen. Secondly, due to the need for iterative algorithmic solutions, the reconstruction speed of these methods is often slow. Therefore, it is still a popular research field to propose a robust algorithm.
Disclosure of Invention
The embodiment of the invention provides a dynamic MR reconstruction method based on deep learning prediction and a residual error framework, which shortens the reconstruction time of a dynamic MR image to a certain extent, and a neural network model can learn some hyper-parameters in a self-adaptive manner without manual selection, so that the technical problem that the hyper-parameters of a traditional reconstruction module are difficult to select can be solved.
An embodiment of the present application provides a dynamic MR reconstruction method based on deep learning prediction and residual error framework, including:
inputting the dynamic MR undersampled data of the K space into a neural network model for processing to obtain a reconstructed dynamic MR image; the method comprises the following steps:
predicting and residual error separating dynamic MR undersampled data of a K space to obtain a predicted signal and a residual error signal;
transforming the prediction signal and the residual signal of the K space into a prediction signal and a residual signal in an image domain, and transforming the prediction signal and the residual signal in the image domain into a prediction signal and a residual signal in an x-f domain;
carrying out sparse coding processing on the residual signal in the x-f domain to obtain a residual signal after sparse coding processing;
and fusing the prediction signal and the residual signal after sparse coding processing to obtain a reconstructed dynamic MR image.
The neural network model in one embodiment of the application comprises a residual error learning network module, a transformation layer and a contraction soft threshold algorithm module;
predicting and residual separating dynamic MR undersampled data of K space to obtain a predicted signal and a residual signal, wherein the predicted signal and the residual signal comprise:
and carrying out mean value processing on the dynamic MR undersampled data of the K space along the time direction, and taking the obtained time mean value as a prediction signal.
In an embodiment of the present application, predicting and residual separating dynamic MR undersampled data in K space to obtain a prediction signal and a residual signal includes:
and inputting the dynamic MR undersampled data of the K space into the residual error learning network module for learning, and learning a residual error signal in a self-adaptive manner through the residual error learning network module.
Illustratively, the residual learning network module includes a plurality of network layers, an output of a previous layer of two adjacent layers of the plurality of network layers is an input of a next layer, an input of a first layer of the plurality of network layers is dynamic MR undersampled data of K space, and an output of a last layer is a residual signal.
The network layers are sequentially a first three-dimensional volume layer, a first active layer, a first ConvLSTM layer, a second three-dimensional volume layer, a second active layer, a second ConvLSTM layer, a third three-dimensional volume layer, a third active layer, a third ConvLSTM layer, a fourth three-dimensional volume layer and a fourth active layer.
In an embodiment of the present application, dynamic MR undersampled data of a K space is input to a residual learning network module for learning, and a learning residual signal adaptively learned by the residual learning network module includes:
inputting the dynamic MR undersampled data of the K space into a neural network formed by a plurality of network layers for learning, forming a residual error network module by the neural network and the time average value, and adaptively learning a residual error signal through the residual error network module.
In an embodiment of the present application, transforming the prediction signal and the residual signal in the K space into the prediction signal and the residual signal in the image domain, and transforming the prediction signal and the residual signal in the image domain into the prediction signal and the residual signal in the x-f domain includes:
inputting the prediction signal and the residual signal of the K space into a transformation layer to perform two-dimensional Fourier inverse transformation to obtain a prediction signal and a residual signal in an image domain; and performing one-dimensional Fourier transform on the prediction signal and the residual signal in the image domain along the time direction to obtain the prediction signal and the residual signal in the x-f domain.
In an embodiment of the present application, performing sparse coding processing on a residual signal in an x-f domain to obtain a residual signal after the sparse coding processing includes:
and inputting the residual signals in the x-f domain into a networked soft threshold shrinking algorithm module, and performing sparse coding processing on the residual signals by the networked soft threshold shrinking algorithm module to obtain residual signals after sparse coding processing.
Illustratively, the soft threshold algorithm module has a plurality of two-dimensional convolutional layers, an output of a previous layer of two adjacent layers of the plurality of two-dimensional convolutional layers is an input of a next layer, an input of a first layer of the plurality of two-dimensional convolutional layers is a residual signal in an x-f domain, and an output of a last layer is a residual signal after sparse coding processing.
In an embodiment of the present application, inputting dynamic MR undersampled data of a K space into a neural network model for processing, and obtaining a reconstructed dynamic MR image includes:
acquiring a training set, wherein the training set comprises a plurality of groups of training data, each group of training data consists of dynamic MR undersampled data of K space and corresponding full-sampling dynamic MR images, training a neural network model according to the training set, and using the trained neural network model for reconstruction.
Compared with the prior art, the embodiment of the application has the advantages that:
the dynamic MR undersampling data of the K space are input into the neural network model to be processed, and a reconstructed dynamic MR image is obtained. Compared with the traditional reconstruction method, the method greatly shortens the reconstruction time of the dynamic MR image, can learn some hyper-parameters in a self-adaptive manner without manual selection, and can solve the technical problem that the hyper-parameters are difficult to select in the traditional reconstruction method.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or the prior art descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a flowchart of a dynamic MR reconstruction method based on deep learning prediction and residual error architecture according to an embodiment of the present application;
FIG. 2 is a schematic diagram of dynamic MR undersampled data provided by an embodiment of the present application;
FIG. 3 is a schematic diagram of a neural network model according to an embodiment of the present application;
fig. 4 is a schematic structural diagram of a residual learning network module according to an embodiment of the present application;
FIG. 5 is a block diagram of a systolic soft threshold algorithm provided in an embodiment of the present application;
Detailed Description
In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular system structures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.
It will be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
It should also be understood that the term "and/or" as used in this specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items and includes such combinations.
Currently, to accelerate data acquisition for dynamic magnetic resonance, most methods consider undersampling the data in k-space. Undersampling introduces aliasing artifacts in the image domain due to violation of the nyquist sampling theorem. Reconstructing an image through undersampled k-space is an underdetermined inverse problem, and generally some prior information needs to be added for constraint solution. More mature and well-accepted are reconstruction methods based on the theory of Compressed Sensing, such methods are called Compressed Sensing magnetic resonance imaging (CS-MRI). Wherein a k-t space FOCal understated system module (k-t FOCUSS) is a typical example of exploiting the space-time redundancy of data, which recovers a signal by enhancing the sparsity of the signal in the x-f domain under the framework of compressed sensing. However, the k-t FOCUSS module still has the following disadvantages when recovering signals: first, the regularization function and its hyper-parameters, which are problem-specific and difficult to select, must be carefully selected. Secondly, due to the need for iterative algorithmic solutions, the reconstruction speed of these methods is often slow.
Based on the defects, the method is based on a framework of a k-t FOCUSS module, and simulates a Magnetic Resonance (MR) undersampling process by undersampling fully-sampled data; the dynamic MR undersampled data of the K space are input into the neural network model to be trained to obtain a reconstructed dynamic MR image, compared with a traditional K-t FOCUSS model, the reconstruction time of the dynamic MR image is shortened, the neural network model can learn some hyper-parameters in a self-adaptive mode instead of manual selection, and the technical problem that the hyper-parameters of a traditional reconstruction module are difficult to select can be solved.
Referring to fig. 1, fig. 1 is a flowchart of a dynamic MR reconstruction method based on deep learning prediction and residual error architecture according to an embodiment of the present application: the execution main body of the method can be an independent terminal, for example, a mobile phone, a computer, a multimedia device, a streaming media device, a monitoring device and other terminal devices; the method can also be an integrated module in the terminal equipment, and the integrated module can be realized as a certain function in the terminal equipment. The following describes an example in which the method is executed by a terminal device, but the embodiment of the present application is not limited thereto. Referring to fig. 1, fig. 1 is a method for dynamic MR reconstruction based on deep learning prediction and residual error architecture according to an embodiment of the present application, which specifically includes the following steps:
step 110: predicting and residual error separating dynamic MR undersampled data of a K space to obtain a predicted signal and a residual error signal;
in the embodiment, the dynamic MR undersampled data of the K space is subjected to mean processing along the time direction in a neural network model, and the obtained time average value is used as a prediction signal.
For example, please refer to fig. 2, fig. 2 is a schematic diagram of dynamic MR undersampling data according to an embodiment of the present application. Wherein k is x Indicating the phase encoding direction, k y Indicating the readout direction. Simulated in this embodiment are dynamic MR undersampled data at successive times T1 to T3. And carrying out mean value processing on the dynamic MR undersampled data at the continuous time from T1 to T3 along the time direction to obtain a prediction signal.
Referring to fig. 3, fig. 3 is a schematic diagram of a neural network model structure according to an embodiment of the present application; in the embodiment, dynamic MR undersampled data of a K space is input into a residual error learning network module for learning, and a residual error signal is adaptively learned through the residual error learning network module.
Illustratively, the neural network model includes a residual learning network module (Res-CNN), a transformation layer, and a systolic soft threshold algorithm module (soft-CNN);
in one embodiment of the application, K-space dynamic MR undersampled data are input into a neural network model, mean value processing is firstly carried out on the dynamic MR undersampled data along the time direction, and the obtained time mean value is used as a prediction signal; secondly, inputting K-space dynamic MR undersampled data into a Res-CNN self-adaptive learning residual signal; k space dynamic MR full sampling data is a supervision signal in Res-CNN; then inputting the prediction signal and the residual signal into a transformation layer, transforming the prediction signal and the residual signal into an image domain by utilizing two-dimensional Fourier inverse transformation in the transformation layer, performing one-dimensional Fourier transformation on the prediction signal and the residual signal in the image domain along the time direction, and transforming the prediction signal and the residual signal into an x-f domain; and then inputting the residual signal into soft-CNN for sparse coding, and finally fusing the residual signal subjected to sparse coding with the prediction signal to obtain a reconstructed dynamic MR image.
In this embodiment, a Magnetic Resonance scanner is first used to perform Magnetic Resonance scanning on a sample to be detected, so as to obtain dynamic Magnetic Resonance (MR) full sampling data in K space, and then the dynamic MR full sampling data in K space is multiplied by a Mask (Mask) to simulate an MR undersampling process, so as to obtain dynamic MR undersampling data in K space.
Referring to fig. 4, fig. 4 is a schematic structural diagram of a residual learning network module according to an embodiment of the present application; illustratively, the residual learning network module includes a plurality of network layers, an output of a previous layer of two adjacent layers of the plurality of network layers is an input of a next layer, an input of a first layer of the plurality of network layers is dynamic MR undersampled data of K space, and an output of a last layer is a residual signal.
The network layers are sequentially a first three-dimensional volume layer, a first active layer, a first ConvLSTM layer, a second three-dimensional volume layer, a second active layer, a second ConvLSTM layer, a third three-dimensional volume layer, a third active layer, a third ConvLSTM layer, a fourth three-dimensional volume layer and a fourth active layer; the activation function in each activation layer is LeakyReLU.
In the embodiment, dynamic MR undersampled data of a K space is input into a neural network formed by a plurality of network layers for learning, the neural network and a time average value form a residual error network module, and a residual error signal is adaptively learned through the residual error network module.
In this embodiment, a three-dimensional convolution layer and a ConvLSTM layer in a residual learning network module are used to output residual signals, on one hand, the three-dimensional convolution layer is used to maintain the time sequence form of dynamic MR undersampled data of a K space, the time sequence is taken into consideration, and the ConvLSTM layer is used to directly process dynamic MR undersampled data with time sequence; on the other hand, the residual learning network module is used to adaptively perform prediction and separation of residual signals, and in the conventional method for calculating residual signals, the residual signals are obtained by directly subtracting the signal to be calculated and the prediction signal in a relatively simple manner. The embodiment can adaptively learn the residual signal by using the residual learning network module without manually designing the separation of the prediction and the residual signal.
Step 120: the prediction signal and the residual signal in the K space are converted into a prediction signal and a residual signal in an image domain, and the prediction signal and the residual signal in the image domain are converted into a prediction signal and a residual signal in an x-f domain.
Referring to fig. 3, fig. 3 is a schematic diagram of a neural network model structure according to an embodiment of the present application; in the embodiment, the reconstructed dynamic MR image uses dynamic MR undersampled data in K space, which does not satisfy the nyquist sampling theorem, and in order to maintain the quality of the magnetic resonance image reconstructed by using the undersampled data and reduce the number of undersampled signal acquisition points, the sparsity in x-f domain space needs to be used. In the embodiment, a prediction signal and a residual signal obtained in a K space are input into a transformation layer for two-dimensional Fourier inversion to obtain a prediction signal and a residual signal in an image domain; and performing one-dimensional Fourier transform on the prediction signal and the residual signal in the image domain along the time direction to obtain the prediction signal and the residual signal in the x-f domain. In a purely mathematical sense, the fourier transform is processed by converting a function into a series of periodic functions, in this embodiment, a prediction signal and a residual signal, which are separated in K space, are both functions of the frequency distribution of the image; prediction signal and residual signal in the image domain, which are both gray scale distribution functions of the image. The inverse fourier transform means to transform a frequency distribution function of an image into a gray distribution function, and the fourier transform means to transform a gray distribution function of an image into a frequency distribution function of an image.
Step 130, carrying out sparse coding processing on the residual signal in the x-f domain to obtain a residual signal subjected to sparse coding processing;
referring to fig. 3, fig. 3 is a schematic diagram of a neural network model structure according to an embodiment of the present application;
in the embodiment, the residual signal in the x-f domain is input into a networked Soft shrinkage threshold algorithm module (Soft-CNN), and the networked Soft shrinkage threshold algorithm module performs sparse coding processing on the residual signal to obtain the residual signal after the sparse coding processing.
For example, please refer to fig. 5, fig. 5 is a schematic structural diagram of a systolic soft threshold algorithm module according to an embodiment of the present application; the shrinkage soft threshold algorithm module is provided with a plurality of two-dimensional convolution layers, the output of the previous layer in two adjacent layers in the two-dimensional convolution layers is the input of the next layer, the input of the first layer in the two-dimensional convolution layers is a residual signal in an x-f domain, and the output of the last layer is a residual signal after sparse coding processing.
The plurality of two-dimensional convolution layers are sequentially a first two-dimensional convolution layer, an active layer and a second two-dimensional convolution layer, and the active layer comprises an activation function (ReLU).
And step 124, fusing the prediction signal and the residual signal after sparse coding processing to obtain a reconstructed dynamic MR image.
According to the embodiment of the application, in a neural network model, an x-f domain prediction signal obtained by enabling a K-space prediction signal to pass through a transformation layer and an x-f domain sparse coding residual signal obtained by enabling a K-space residual signal to pass through the transformation layer and a shrinkage soft threshold algorithm module are fused to obtain a reconstructed dynamic MR image.
Before the above 110 is performed, the dynamic MR undersampled data needs to be input into the neural network model for training, including:
acquiring a training set, wherein the training set comprises a plurality of groups of training data, each group of training data consists of dynamic MR undersampled data of K space and corresponding full-sampling dynamic MR images, training a neural network model according to the training set, and using the trained neural network model for reconstruction.
When the neural network model is trained, a plurality of groups of pre-obtained K-space dynamic MR undersampled data and corresponding full-sampling dynamic MR images can be input into the pre-constructed neural network model, and the neural network model is trained according to the plurality of groups of pre-obtained K-space dynamic MR undersampled data and corresponding full-sampling dynamic MR images; and re-inputting a group of dynamic MR undersampled data into the trained neural network model, and comparing an output result with a corresponding full-sampling dynamic MR image to verify the reconstruction performance of the neural network model.
In the above embodiments, the description of each embodiment has its own emphasis, and reference may be made to the related description of other embodiments for parts that are not described or recited in any embodiment.
Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
The above-mentioned embodiments are only used for illustrating the technical solutions of the present application, and not for limiting the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present application and are intended to be included within the scope of the present application.
Claims (8)
1. A dynamic MR reconstruction method based on deep learning prediction and residual error architecture is characterized by comprising the following steps:
inputting the dynamic MR undersampled data of the K space into a neural network model for processing to obtain a reconstructed dynamic MR image; the method comprises the following steps:
predicting and residual error separating dynamic MR undersampled data of a K space to obtain a predicted signal and a residual error signal;
transforming the prediction signal and the residual signal of the K space into the prediction signal and the residual signal in an image domain, transforming the prediction signal and the residual signal in the image domain into the prediction signal and the residual signal in an x-f domain;
performing sparse coding processing on the residual signal in the x-f domain to obtain the residual signal subjected to sparse coding processing;
fusing the prediction signal and the residual signal after sparse coding processing to obtain a reconstructed dynamic MR image; the neural network model comprises a residual error learning network module, a transformation layer and a contraction soft threshold algorithm module;
the predicting and residual separating the dynamic MR undersampled data of the K space to obtain a predicted signal and a residual signal comprises:
carrying out mean value processing on the dynamic MR undersampled data of the K space along the time direction, and taking the obtained time mean value as the prediction signal;
the performing sparse coding processing on the residual signal in the x-f domain to obtain the residual signal after sparse coding processing includes:
inputting the residual signal in the x-f domain into a networked soft threshold shrinkage algorithm module, and performing sparse coding processing on the residual signal by the networked soft threshold shrinkage algorithm module to obtain the residual signal after the sparse coding processing.
2. The method according to claim 1, wherein the predicting and residual separating the dynamic MR undersampled data in K space to obtain a predicted signal and a residual signal comprises:
inputting the dynamic MR undersampled data of the K space into the residual error learning network module for learning, and learning the residual error signal in a self-adaptive manner through the residual error learning network module.
3. The method according to claim 2, wherein the residual learning network module comprises a plurality of network layers, an output of a previous layer of two adjacent layers of the plurality of network layers is an input of a next layer, an input of a first layer of the plurality of network layers is the dynamic MR undersampled data of the K space, and an output of a last layer is the residual signal.
4. The method of claim 3, wherein the plurality of network layers are sequentially a first three-dimensional convolutional layer, a first active layer, a first ConvLSTM layer, a second three-dimensional convolutional layer, a second active layer, a second ConvLSTM layer, a third three-dimensional convolutional layer, a third active layer, a third ConvLSTM layer, a fourth three-dimensional convolutional layer, and a fourth active layer.
5. The method according to claim 4, wherein the inputting the dynamic MR undersampled data of the K space into the residual learning network module for learning comprises:
inputting the dynamic MR undersampled data of the K space into a neural network formed by the plurality of network layers for learning, wherein the neural network and the time average value form a residual error network module, and the residual error signal is adaptively learned through the residual error network module.
6. The method of claim 1, wherein transforming the prediction signal and the residual signal of the K-space into the prediction signal and the residual signal in an image domain, and transforming the prediction signal and the residual signal in the image domain into the prediction signal and the residual signal in an x-f domain comprises:
inputting the prediction signal and the residual signal of the K space into a transform layer to perform two-dimensional Fourier inverse transform to obtain the prediction signal and the residual signal in the image domain; and performing one-dimensional Fourier transform on the prediction signal and the residual signal in the image domain along the time direction to obtain the prediction signal and the residual signal in an x-f domain.
7. The method according to claim 1, wherein the contracted soft threshold algorithm module has a plurality of two-dimensional convolutional layers, an output of a previous layer of two adjacent layers of the two-dimensional convolutional layers is an input of a next layer, an input of a first layer of the two-dimensional convolutional layers is the residual signal in the x-f domain, and an output of a last layer is the residual signal after the sparse coding process.
8. The method of claim 1, wherein the method comprises the steps of: the method for inputting the dynamic MR undersampled data of the K space into the neural network model for processing to obtain the reconstructed dynamic MR image comprises the following steps:
acquiring a training set, wherein the training set comprises a plurality of groups of training data, each group of training data consists of dynamic MR undersampled data of K space and corresponding full-sampling dynamic MR images, and training a neural network model according to the training set, and using the trained neural network model for reconstruction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011445536.6A CN112669400B (en) | 2020-12-11 | 2020-12-11 | Dynamic MR reconstruction method based on deep learning prediction and residual error framework |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011445536.6A CN112669400B (en) | 2020-12-11 | 2020-12-11 | Dynamic MR reconstruction method based on deep learning prediction and residual error framework |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112669400A CN112669400A (en) | 2021-04-16 |
| CN112669400B true CN112669400B (en) | 2022-11-29 |
Family
ID=75402092
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011445536.6A Active CN112669400B (en) | 2020-12-11 | 2020-12-11 | Dynamic MR reconstruction method based on deep learning prediction and residual error framework |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112669400B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113421247B (en) * | 2021-06-30 | 2022-06-07 | 首都医科大学附属北京天坛医院 | Method and device for determining brain tissue microstructure estimation parameters and electronic equipment |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111867465A (en) * | 2018-03-12 | 2020-10-30 | 勒维斯公司 | System and method for generating thin image slices from thick image slices |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10133964B2 (en) * | 2017-03-28 | 2018-11-20 | Siemens Healthcare Gmbh | Magnetic resonance image reconstruction system and method |
| EP3451670A1 (en) * | 2017-08-28 | 2019-03-06 | Thomson Licensing | Method and apparatus for filtering with mode-aware deep learning |
| US11029381B2 (en) * | 2018-01-12 | 2021-06-08 | Korea Advanced Institute Of Science And Technology | Method for varying undersampling dimension for accelerating multiple-acquisition magnetic resonance imaging and device for the same |
-
2020
- 2020-12-11 CN CN202011445536.6A patent/CN112669400B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111867465A (en) * | 2018-03-12 | 2020-10-30 | 勒维斯公司 | System and method for generating thin image slices from thick image slices |
Non-Patent Citations (2)
| Title |
|---|
| k-t NEXT: Dynamic MR Image Reconstruction Exploiting Spatio-temporal Correlations;C Qin et al;《arXiv:1907.09425》;20191231;第1-9页 * |
| 基于压缩感知技术的快速动态磁共振成像技术研究;阮钦杰;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20140915;第C030-32页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112669400A (en) | 2021-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | RARE: Image reconstruction using deep priors learned without groundtruth | |
| CN109325985B (en) | Magnetic resonance image reconstruction method, apparatus and computer readable storage medium | |
| Liu et al. | Highly undersampled magnetic resonance image reconstruction using two-level Bregman method with dictionary updating | |
| CN111028306A (en) | AR2U-Net neural network-based rapid magnetic resonance imaging method | |
| Pal et al. | A review and experimental evaluation of deep learning methods for MRI reconstruction | |
| Cheng et al. | Learning data consistency and its application to dynamic MR imaging | |
| Knoll et al. | Deep learning methods for parallel magnetic resonance image reconstruction | |
| CN114998458A (en) | Undersampled magnetic resonance image reconstruction method based on reference image and data correction | |
| CN114299185A (en) | Magnetic resonance image generation method, magnetic resonance image generation device, computer equipment and storage medium | |
| Manimala et al. | Convolutional neural network for sparse reconstruction of MR images interposed with gaussian noise | |
| Li et al. | A modified generative adversarial network using spatial and channel-wise attention for CS-MRI reconstruction | |
| CN111784792A (en) | Rapid magnetic resonance reconstruction system based on double-domain convolution neural network and training method and application thereof | |
| CN113534031A (en) | Image domain data generating method, computer device and readable storage medium | |
| Wang et al. | High fidelity direct-contrast synthesis from magnetic resonance fingerprinting in diagnostic imaging | |
| CN112669400B (en) | Dynamic MR reconstruction method based on deep learning prediction and residual error framework | |
| Xiao et al. | SR-Net: a sequence offset fusion net and refine net for undersampled multislice MR image reconstruction | |
| CN110942496A (en) | Propeller sampling and neural network-based magnetic resonance image reconstruction method and system | |
| Zhou et al. | RNLFNet: Residual non-local Fourier network for undersampled MRI reconstruction | |
| Gan et al. | Image reconstruction for mri using deep cnn priors trained without groundtruth | |
| Jaubert et al. | Deep artifact suppression for spiral real-time phase contrast cardiac magnetic resonance imaging in congenital heart disease | |
| CN112700508A (en) | Multi-contrast MRI image reconstruction method based on deep learning | |
| Bilal et al. | Respiratory motion correction for compressively sampled free breathing cardiac MRI using smooth-norm approximation | |
| CN116626570A (en) | Multi-contrast MRI sampling and image reconstruction | |
| CN112924913B (en) | Space-time coding magnetic resonance imaging super-resolution reconstruction method and system | |
| Ke et al. | Cascaded residual dense networks for dynamic MR imaging with edge-enhanced loss constraint |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| CB03 | Change of inventor or designer information |
Inventor after: Wang Shanshan Inventor after: Peng Zhenkun Inventor after: Zheng Hairong Inventor after: Liu Xin Inventor before: Wang Shanshan Inventor before: Zheng Hairong Inventor before: Peng Zhenkun Inventor before: Liu Xin |
|
| CB03 | Change of inventor or designer information | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |