WO2020118687A1 - 一种自适应参数学习的动态磁共振图像重建方法和装置 - Google Patents

一种自适应参数学习的动态磁共振图像重建方法和装置 Download PDF

Info

Publication number
WO2020118687A1
WO2020118687A1 PCT/CN2018/121193 CN2018121193W WO2020118687A1 WO 2020118687 A1 WO2020118687 A1 WO 2020118687A1 CN 2018121193 W CN2018121193 W CN 2018121193W WO 2020118687 A1 WO2020118687 A1 WO 2020118687A1
Authority
WO
WIPO (PCT)
Prior art keywords
magnetic resonance
dct
image
filter operator
resonance image
Prior art date
Application number
PCT/CN2018/121193
Other languages
English (en)
French (fr)
Inventor
王珊珊
陈艳霞
郑海荣
梁栋
刘新
肖韬辉
柯子文
Original Assignee
深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Priority to PCT/CN2018/121193 priority Critical patent/WO2020118687A1/zh
Publication of WO2020118687A1 publication Critical patent/WO2020118687A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/002D [Two Dimensional] image generation

Definitions

  • the present application belongs to the technical field of image processing, and in particular relates to an adaptive parameter learning dynamic magnetic resonance image reconstruction method and device.
  • Magnetic resonance Magnetic ResonanceImaging, referred to as MRI
  • MRI Magnetic resonance
  • cardiac, perfusion, and functional imaging are items that require high real-time performance, and magnetic resonance imaging is often unsatisfactory.
  • patients may feel uncomfortable and introduce motion artifacts.
  • the purpose of this application is to provide an adaptive parameter learning dynamic magnetic resonance image reconstruction method and device.
  • the under-sampling image can be efficiently reconstructed through the established magnetic resonance image reconstruction model to obtain the required image, which can effectively shorten the magnetic field Resonance scan time.
  • This application provides an adaptive parameter learning dynamic magnetic resonance image reconstruction method and device implemented as follows:
  • An adaptive parameter learning dynamic magnetic resonance image reconstruction method includes:
  • the DCT filter operator, the TV filter operator and the regularization parameters in the magnetic resonance reconstruction model are updated using the back propagation algorithm in the network to realize the optimization of the magnetic resonance reconstruction model Training until the magnetic resonance image reconstruction model meets the preset accuracy requirements;
  • magnetic resonance image reconstruction is performed.
  • the establishment of a magnetic resonance image reconstruction model includes:
  • DCT filtering is used in the spatial domain
  • TV filtering is used in the time domain to enhance the sparsity of the image, and the sparsity expression formula is obtained;
  • the multiple sub-problem solving formulas are networked to obtain a magnetic resonance image reconstruction model.
  • a joint sparse model is used, DCT filtering is used in the spatial domain, and TV filtering is used in the time domain to enhance the sparsity of the image, and the sparsity expression formula is obtained, including:
  • arg min f(x) represents the set of all independent variables x that make the function f(x) take its minimum value
  • N t ⁇ x 1 ,x 2 ,...,x T ⁇ , indicating that there are T frames of images in the time direction
  • y represents the undersampled k-space data
  • F represents the Fourier transform
  • represents the filter operator
  • represents the parameter.
  • Image representing the spatial domain Represents the image in the time domain
  • ⁇ 1 represents the DCT filter operator
  • ⁇ 2 represents the TV filter operator
  • DCT represents the discrete cosine transform
  • TV represents the total variation transformation, ⁇ 1 , ⁇ 2 To balance the parameters of the data fidelity and regularization terms.
  • the target formula is:
  • z represents the auxiliary variable
  • represents the Lagrange multiplier
  • g( ⁇ ) represents an approximate function of l 2
  • p norm p norm
  • represents the penalty parameter
  • the multiple sub-problem solving formulas are:
  • T transpose
  • represents the scaling factor of the Lagrange multiplier
  • k ⁇ ⁇ 1,2,...,N t ⁇ represents the number of iterations in gradient descent
  • n represents the n-th layer
  • ⁇ 1 (1-l r ⁇ )
  • ⁇ 2 l r ⁇ , which represent the weight parameters that can be learned
  • l r represents the step size
  • H( ⁇ ) represents the gradient of g( ⁇ )
  • D 1 represents the transformation matrix.
  • the multiple sub-problem solving formulas are networked to obtain:
  • I represents the identity matrix
  • C 1 and C 2 represent the two convolutional layers
  • w 1 corresponds to the combination of DCT and TV filter
  • the size is 3*3*1*L
  • b 1 represents the offset of L dimension vector
  • w 2 corresponds to the combination of DCT and TV filter
  • the size is 3*3*L
  • b 2 represents a 1-dimensional offset vector
  • S PLF ( ⁇ ) represents a Controlled piecewise linear function
  • N C is a parameter used to control the points in the piecewise linear function
  • Recon represents the reconstruction layer
  • Addition represents the overlay layer
  • Conv1 represents the convolution layer
  • Nonlinear represents the nonlinear transformation layer
  • Conv2 represents the volume Multilayer
  • Multi represents the multiplier update layer.
  • An adaptive parameter learning dynamic magnetic resonance image reconstruction device includes:
  • the replacement module is used to use the DCT filter operator in the spatial domain and the TV filter operator in the time domain as the regularization term of CS-MRI;
  • the establishment module is used to establish a magnetic resonance image reconstruction model according to the regularization terms introduced by the DCT filter operator and the TV filter operator;
  • the reconstruction module is used to reconstruct the sample image through the established magnetic resonance reconstruction model to obtain the reconstructed image
  • a calculation module used to calculate the difference between the fully sampled image corresponding to the sample image and the reconstructed image
  • the iterative update module is used to update the DCT filter operator, TV filter operator and regularization parameters in the magnetic resonance reconstruction model based on the difference using the back propagation algorithm in the network to achieve the Optimal training of the magnetic resonance reconstruction model until the magnetic resonance image reconstruction model meets the preset accuracy requirements;
  • the application module is used to reconstruct the magnetic resonance image according to the magnetic resonance image reconstruction model that meets the preset accuracy requirements.
  • the establishment module includes;
  • the generating unit is used to use the joint sparse model, use the DCT filter in the spatial domain, and use the TV filter in the time domain to enhance the sparsity of the image, and obtain the sparsity expression formula;
  • a first conversion unit configured to perform fusion conversion on the sparse expression formula to obtain a target formula
  • a second conversion unit for converting the target formula into a plurality of sub-problem solving formulas using an alternating direction multiplier method
  • the network unit is used to network the multiple sub-problem solving formulas to obtain a magnetic resonance image reconstruction model.
  • the generating unit is specifically configured to set the filter operator of the following formula as a DCT operator and a TV operator:
  • arg min f(x) represents the set of all independent variables x that make the function f(x) take its minimum value
  • y represents undersampled k-space data
  • A PF, where P is an undersampled matrix
  • F represents Fourier transform
  • represents a filter operator
  • represents a parameter.
  • ⁇ 1 represents the DCT filter operator
  • ⁇ 2 represents the TV filter operator
  • DCT represents the discrete cosine transform
  • TV represents the total variation transformation
  • ⁇ 1 and ⁇ 2 are the fidelity terms of balanced data Parameters with regularization terms.
  • the target formula is:
  • z represents the auxiliary variable
  • represents the Lagrange multiplier
  • g( ⁇ ) represents an approximate function of l 2
  • p norm p norm
  • represents the penalty parameter
  • the multiple sub-problem solving formulas are:
  • T transpose
  • represents the scaling factor of the Lagrange multiplier
  • k ⁇ ⁇ 1,2,...,N t ⁇ represents the number of iterations in gradient descent
  • n represents the n-th layer
  • ⁇ 1 (1-l r ⁇ )
  • ⁇ 2 l r ⁇ , which represent the weight parameters that can be learned
  • l r represents the step size
  • H( ⁇ ) represents the gradient of g( ⁇ )
  • D 1 represents the transformation matrix.
  • the multiple sub-problem solving formulas are networked to obtain:
  • I represents the identity matrix
  • C 1 and C 2 represent the two convolutional layers
  • w 1 corresponds to the combination of DCT and TV filter
  • the size is 3*3*1*L
  • b 1 represents the offset of L dimension vector
  • w 2 corresponds to the combination of DCT and TV filter
  • the size is 3*3*L
  • b 2 represents a 1-dimensional offset vector
  • S PLF ( ⁇ ) represents a Controlled piecewise linear function
  • N C is a parameter used to control the points in the piecewise linear function
  • Recon represents the reconstruction layer
  • Addition represents the overlay layer
  • Conv1 represents the convolution layer
  • Nonlinear represents the nonlinear transformation layer
  • Conv2 represents the volume Multilayer
  • Multi represents the multiplier update layer.
  • a terminal device includes a processor and a memory for storing processor-executable instructions. When the processor executes the instructions, the following steps are implemented:
  • magnetic resonance image reconstruction is performed.
  • a computer-readable storage medium having computer instructions stored thereon, the instructions implement the following steps when executed:
  • magnetic resonance image reconstruction is performed.
  • the adaptive parameter learning dynamic magnetic resonance image reconstruction method and device provided in this application, because the magnetic resonance image is sparse and there is noise, by using the DCT filter operator in the spatial domain and the TV filter operator in the time domain as regular Parameters to establish a magnetic resonance image reconstruction model, which can remove the redundancy in the spatial direction and the time direction respectively, therefore, it can effectively improve the reconstruction effect. Furthermore, the adjustment of the DCT filter operator and the TV filter operator in the time domain can be effectively improved through the self-learning method of the network, which can effectively improve the accuracy of the magnetic resonance image reconstruction model.
  • the magnetic resonance image reconstruction model established by the above method can Efficiently reconstruct highly undersampled images to obtain images with high reconstruction accuracy and reconstruction speed, which can effectively shorten the time of magnetic resonance scanning without losing spatial resolution.
  • FIG. 1 is a method flowchart of an embodiment of an adaptive parameter learning dynamic magnetic resonance image reconstruction method provided by this application;
  • FIG. 3 is a schematic diagram of a neural network model obtained after networking provided in this application.
  • FIG. 4 is a schematic structural diagram of a terminal device provided by this application.
  • FIG. 5 is a structural block diagram of an adaptive parameter learning dynamic magnetic resonance image reconstruction device provided by the present application.
  • CS-MRI Compressed Sensing-Magnetic Resonance Imaging
  • CS-MRI Compressed Sensing-Magnetic Resonance Imaging
  • wavelet sparse regularization and total variation sparse regularization for the study of regularization terms
  • adaptive dictionary learning further, studies have shown that low-rank models are effective for dynamic magnetic resonance imaging.
  • the existing sparse and low rank-based methods have better reconstruction effects, they still cannot reconstruct the signal well.
  • the filter selection is The artificial selection is not representative, and the reconstruction time based on the dictionary learning method is relatively long.
  • the reconstruction algorithm is based on the compressed sensing theory to quickly realize magnetic resonance imaging. Since the magnetic resonance image is sparse and there is noise, in this example, a discrete cosine (Discrete Cosine Transform, DCT) filter is proposed in space, and the total variation (Total Variation, abbreviation) is used in the time direction.
  • DCT Discrete Cosine Transform
  • Total Variation abbreviation
  • the magnetic resonance image is sparse
  • the discrete cosine transform is used to transform the image to the sparse domain.
  • the value of the filter can be obtained by the network autonomous learning, and no human selection is required. Using TV for filtering in the time direction can not only remove noise, but also further enhance the sparsity of the MR image and improve the reconstruction quality of the image.
  • the scanning speed of dynamic magnetic resonance imaging is increased by undersampling.
  • direct reconstruction will produce artifacts, in this example, by using CS-MRI technology to explore the regularization term, combined with deep learning adaptive learning method, making undersampling magnetic resonance image It can quickly reconstruct high-resolution, approximately full-sampled images.
  • FIG. 1 it is a method flowchart of an embodiment of an adaptive parameter learning dynamic magnetic resonance image reconstruction method described in this application.
  • this application provides method operation steps or device structures as shown in the following embodiments or drawings, more or fewer operation steps or module units may be included in the method or device based on conventional or without creative labor .
  • the execution order of these steps or the module structure of the device is not limited to the execution order or module structure shown in the embodiment of the present application and shown in the drawings.
  • the method or the module structure shown in the embodiments or the drawings can be connected to perform sequential execution or parallel execution (for example, parallel processor or multi-threaded processing). Environment, even distributed processing environment).
  • an adaptive parameter learning dynamic magnetic resonance image reconstruction method may include (step 101 to step 106):
  • Step 101 Use the DCT filter operator in the spatial domain and the TV filter operator in the time domain as the CS-MRI regularization terms;
  • Step 102 Establish a magnetic resonance image reconstruction model according to the regularization terms introduced by the DCT filter operator and the TV filter operator;
  • establishing a magnetic resonance image reconstruction model may include:
  • the filter operator of the following formula can be set as the DCT filter operator and the TV filter operator:
  • arg min f(x) represents the set of all independent variables x that make the function f(x) take its minimum value
  • y represents undersampled k-space data
  • A PF, where P is an undersampled matrix
  • F represents Fourier transform
  • represents a filter operator
  • represents a parameter.
  • ⁇ 1 represents the DCT filter operator
  • ⁇ 2 represents the TV filter operator
  • DCT represents the discrete cosine transform
  • TV represents the total variation transformation
  • ⁇ 1 and ⁇ 2 are the fidelity terms of balanced data Parameters with regularization terms.
  • the filter operator in the original formula is selected as the DCT filter operator and the TV filter operator, so that deduplication can be performed in time and space.
  • the target formula can be expressed as:
  • z represents the auxiliary variable
  • represents the Lagrange multiplier
  • g( ⁇ ) represents an approximate function of l 2
  • p norm p norm
  • represents the penalty parameter
  • the target formula can be converted as follows:
  • the sparse induced norms used for sparse regularization include: l 0 norm, l 1 norm, l 2 , 1 norm is defined as the sum of the l 2 norm of the row vector, minimize l
  • represents the fusion of ⁇ 1 and ⁇ 2
  • g( ⁇ ) is an approximate function of l 2, p norm.
  • the augmented Lagrangian function is the objective formula.
  • the conversion from the basic formula to the target formula is completed in the above manner.
  • ADMM Alternating Direction Method of Multipliers
  • Decomposition-Coordination Decomposition-Coordination
  • T transpose
  • represents the scaling factor of the Lagrange multiplier
  • k ⁇ ⁇ 1,2,...,N t ⁇ represents the number of iterations in gradient descent
  • n represents the n-th layer
  • ⁇ 1 (1-l r ⁇ )
  • ⁇ 2 l r ⁇ , which represent the weight parameters that can be learned
  • l r represents the step size
  • H( ⁇ ) represents the gradient of g( ⁇ )
  • D 1 represents the transformation matrix (such as DCT, wavelet transform, etc.).
  • I represents the identity matrix.
  • C 1 and C 2 respectively represent two convolutional layers.
  • w 1 corresponds to the combination of DCT and TV filter, the size is 3*3*1*L
  • b 1 represents the L-dimensional offset vector
  • w 2 corresponds to the combination of DCT and TV filter, the size is 3*3* L
  • b 2 represents a one-dimensional offset vector.
  • S PLF ( ⁇ ) represents a Controlled piecewise linear function
  • N C is a parameter used to control the points in the piecewise linear function
  • Recon represents the reconstruction layer
  • Addition represents the overlay layer
  • Conv1 represents the convolution layer
  • Nonlinear represents the nonlinear transformation layer
  • Conv2 represents the volume Multilayer
  • Multi represents the multiplier update layer.
  • Step 103 Reconstruct the sample image through the established magnetic resonance reconstruction model to obtain a reconstructed image
  • a neural network of the magnetic resonance reconstruction model can be obtained, and the sample image can be reconstructed based on the neural network to obtain the reconstructed image.
  • Step 104 Calculate the difference between the fully sampled image corresponding to the sample image and the reconstructed image
  • Step 105 Using the back propagation algorithm in the network according to the difference, update the DCT filter operator, TV filter operator and regularization parameters in the magnetic resonance reconstruction model to realize the reconstruction of the magnetic resonance Optimal training of the model until the magnetic resonance image reconstruction model meets the preset accuracy requirements;
  • the loss value can be calculated by the standard mean square error between the reconstructed image and the original image, the expression is as follows:
  • represents the number of training sets.
  • the network model can be backpropagated to update the parameters, so as to optimize and update each parameter in the network model based on the loss value, so that the finally determined network model parameters, It can ensure higher reconstruction accuracy of the network model.
  • the update process of the back propagation parameters can be the gradient update of the Loss layer, which can be expressed as:
  • Step 106 Perform magnetic resonance image reconstruction based on the magnetic resonance image reconstruction model that meets the preset accuracy requirements.
  • the regularization term in the CS-MRI model was improved, including using the discrete cosine filter operator (DCT) in the spatial domain and the total variation filter operator (TV) in the time domain to perform dynamic magnetic resonance imaging.
  • DCT discrete cosine filter operator
  • TV total variation filter operator
  • the magnetic resonance image reconstruction model established by the above method can efficiently reconstruct highly undersampled images to obtain images
  • the magnetic resonance image is sparse.
  • DCT filtering is introduced in the spatial domain of the regularization term, and the time domain Introducing TV filtering, the combination of the two can greatly reduce the redundancy of the image.
  • iterative reconstruction method is used for solving, because the real projection data will be used in the iterative process, so the reconstruction result will be more accurate in theory.
  • the regularization parameters of the image can be learned independently, without artificial setting.
  • represents a filter operator (for example: wavelet transform DWT, discrete cosine transform DCT, total variation TV), and ⁇ 1 and ⁇ 2 are parameters that balance the fidelity and regularization terms of the data.
  • Image representing the spatial domain Representing images in the time domain
  • ⁇ and ⁇ are two regularization parameters.
  • ⁇ 1 represents the DCT filter operator
  • ⁇ 2 represents the TV filter operator.
  • DCT means discrete cosine transform
  • TV means total variation transform. among them, Represents the sum of finite difference operators of the previous frame The finite difference operator representing the next frame.
  • the sparse induced norms used for sparse regularization include: l 0 norm, l 1 norm, l 2 , 1 norm is defined as the sum of the l 2 norm of the row vector, minimize l 2, The purpose of the 1 norm is to select as few non-zero row vectors as possible. However, this can only be applied to vector norms.
  • matrix norms for sparse regularization for example: using l 2,p matrix norms (0 ⁇ p ⁇ 1) for feature selection can be achieved More sparse solutions.
  • choose l 2, p norm, in order to simplify the description, another ⁇ ⁇ .
  • the above formula 2 can be simplified to:
  • represents the fusion of ⁇ 1 and ⁇ 2
  • g( ⁇ ) is an approximate function of l 2, p norm.
  • is the scaling factor of the Lagrange multiplier
  • k ⁇ ⁇ 1,2,...,N t ⁇ represents the number of iterations in gradient descent
  • n the nth layer
  • ⁇ 1 (1-l r ⁇ )
  • ⁇ 2 l r ⁇ , which represents the weight parameters that can be learned
  • l r represents the step size
  • H( ⁇ ) represents the gradient of g( ⁇ )
  • D 1 represents the transformation matrix.
  • the formula is the reconstruction process of the forward-propagated image
  • I represents the identity matrix
  • C 1 and C 2 represent the two convolutional layers
  • w 1 corresponds to the combination of DCT and TV filter
  • the size is 3*3*1*L
  • B 1 represents the L-dimensional offset vector.
  • w 2 corresponds to the combination of DCT and TV filter, the size is 3*3*L, b 2 represents a 1-dimensional offset vector, S PLF ( ⁇ ) represents a Controlled piecewise linear function, N C is a parameter used to control the point in the piecewise linear function, as shown in Figure 3,
  • DCTV-Net in Figure 3 represents the network flow chart, where Recon represents the reconstruction layer and Addition represents Overlay layer, Conv1 represents a convolutional layer, Nonlinear represents a nonlinear transformation layer, Conv2 represents a convolutional layer, and Multi represents a multiplier update layer.
  • the loss value can be calculated by the standard mean square error between the reconstructed image and the original image.
  • the expression is as follows:
  • represents the number of training sets
  • the forward arrow in Fig. 2 is the result of one iteration of the reconstruction
  • the reverse arrow represents the process of back propagation to update the parameters.
  • the update process of the back propagation parameters can be the gradient update of the Loss layer, which can be expressed as:
  • the sparse constraint on the magnetic resonance image is made through sparse network in the time-frequency domain, so that a higher acceleration factor can be obtained without losing the spatial resolution.
  • a large number of solution parameters in CS-MRI are replaced with CNN, which can avoid the contingency caused by human choice.
  • the method used in this example has higher reconstruction accuracy than the traditional deep learning method. Under the same acceleration factor, the reconstruction accuracy of this example is higher and the reconstruction effect is better.
  • the neural network is used to learn these regularization parameters, and the trained model is obtained through offline training. The online test only takes about 3 seconds to obtain a highly undersampled MRI reconstructed image.
  • FIG. 4 is a hardware block diagram of a computer terminal of an adaptive parameter learning dynamic magnetic resonance image reconstruction method according to an embodiment of the present invention.
  • the terminal device 10 may include one or more (only one is shown in the figure) processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) , A memory 104 for storing data, and a transmission module 106 for communication functions.
  • the structure shown in FIG. 4 is merely an illustration, which does not limit the structure of the above electronic device.
  • the computer terminal 10 may also include more or fewer components than those shown in FIG. 4, or have a different configuration from that shown in FIG.
  • the memory 104 may be used to store software programs and modules of application software, such as program instructions/modules corresponding to the adaptive parameter learning dynamic magnetic resonance image reconstruction method in the embodiments of the present invention.
  • the processor 102 runs the software stored in the memory 104 Programs and modules to perform various functional applications and data processing, that is, to realize the above-mentioned adaptive parameter learning dynamic magnetic resonance image reconstruction method.
  • the memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory.
  • the memory 104 may further include memories remotely provided with respect to the processor 102, and these remote memories may be connected to the computer terminal 10 through a network. Examples of the above network include but are not limited to the Internet, intranet, local area network, mobile communication network, and combinations thereof.
  • the transmission module 106 is used to receive or send data via a network.
  • the above specific example of the network may include a wireless network provided by a communication provider of the computer terminal 10.
  • the transmission module 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through the base station to communicate with the Internet.
  • the transmission module 106 may be a radio frequency (Radio Frequency) module, which is used to communicate with the Internet in a wireless manner.
  • Radio Frequency Radio Frequency
  • the above-mentioned device may be as shown in FIG. 5 and may include:
  • the replacement module 501 is used to use the DCT filter operator in the spatial domain and the TV filter operator in the time domain as the regularization term of CS-MRI;
  • the establishment module 502 is used to establish a magnetic resonance image reconstruction model according to the regularization terms introduced by the DCT filter operator and the TV filter operator;
  • the reconstruction module 503 is used to reconstruct the sample image through the established magnetic resonance reconstruction model to obtain a reconstructed image
  • the calculation module 504 is configured to calculate the difference between the fully sampled image corresponding to the sample image and the reconstructed image
  • the iterative update module 505 is used to update the DCT filter operator, TV filter operator and regularization parameters in the magnetic resonance reconstruction model based on the difference using the back propagation algorithm in the network to achieve Describe the optimal training of the magnetic resonance reconstruction model until the magnetic resonance image reconstruction model meets the preset accuracy requirements;
  • the application module 506 is used to reconstruct a magnetic resonance image according to a magnetic resonance image reconstruction model that meets preset precision requirements.
  • the establishment module 502 may include: a generating unit that uses a joint sparse model, uses DCT filtering in the spatial domain, and uses TV filtering in the time domain to enhance the sparseness of the image to obtain a sparse expression formula; A conversion unit is used to fuse the sparse expression formula to obtain a target formula; a second conversion unit is used to convert the target formula into a plurality of sub-problem solving formulas using an alternating direction multiplier method; networked The unit is used to network the multiple sub-problem solving formulas to obtain a magnetic resonance image reconstruction model.
  • the above-mentioned generating unit may specifically be used to set the filter operator of the following formula as a DCT operator and a TV operator:
  • arg min f(x) represents the set of all independent variables x that make the function f(x) take its minimum value
  • y represents undersampled k-space data
  • A PF, where P is an undersampled matrix
  • F represents Fourier transform
  • represents a filter operator
  • represents a parameter.
  • Image representing the spatial domain Represents images in the time domain
  • N t ⁇ x 1 ,x 2 ,...,x T ⁇ , representing a total of T frames in the time direction
  • ⁇ 1 represents a DCT filter operator
  • ⁇ 2 represents a TV filter operator
  • DCT stands for Discrete Cosine Transform
  • TV stands for Total Variation Transform
  • ⁇ 1 and ⁇ 2 are the parameters of the fidelity and regularization terms of the balanced data.
  • the above target formula can be expressed as:
  • z represents the auxiliary variable
  • represents the Lagrange multiplier
  • g( ⁇ ) represents an approximate function of l 2
  • p norm p norm
  • represents the penalty parameter
  • T transpose
  • represents the scaling factor of the Lagrange multiplier
  • k ⁇ ⁇ 1,2,...,N t ⁇ represents the number of iterations in gradient descent
  • n represents the n-th layer
  • ⁇ 1 (1-l r ⁇ )
  • ⁇ 2 l r ⁇ , which represent the weight parameters that can be learned
  • l r represents the step size
  • H( ⁇ ) represents the gradient of g( ⁇ )
  • D 1 represents the transformation matrix.
  • networking multiple sub-problem solving formulas can obtain:
  • I represents the identity matrix
  • C 1 and C 2 represent the two convolutional layers
  • w 1 corresponds to the combination of DCT and TV filter
  • the size is 3*3*1*L
  • b 1 represents the offset of L dimension vector
  • w 2 corresponds to the combination of DCT and TV filter
  • the size is 3*3*L
  • b 2 represents a 1-dimensional offset vector
  • S PLF ( ⁇ ) represents a Controlled piecewise linear function
  • N C is a parameter used to control the points in the piecewise linear function
  • Recon represents the reconstruction layer
  • Addition represents the overlay layer
  • Conv1 represents the convolution layer
  • Nonlinear represents the nonlinear transformation layer
  • Conv2 represents the volume Multilayer
  • Multi represents the multiplier update layer.
  • Embodiments of the present application also provide a specific implementation of an electronic device that can implement all the steps in the dynamic parameter imaging method for adaptive parameter learning in the foregoing embodiment.
  • the electronic device specifically includes the following:
  • Processor processor
  • memory memory
  • communication interface Communication and bus
  • the processor, the memory, and the communication interface complete communication with each other through the bus; the processor is used to call a computer program in the memory, and the processor implements the computer program to implement the above embodiment All the steps in the dynamic magnetic resonance image reconstruction method of adaptive parameter learning, for example, when the processor executes the computer program, the following steps are realized:
  • Step 1 Use the DCT filter operator in the spatial domain and the TV filter operator in the time domain as the regularization terms of CS-MRI;
  • Step 2 According to the regularization terms introduced by the DCT filter operator and the TV filter operator, a magnetic resonance image reconstruction model is established;
  • Step 3 Reconstruct the sample image through the established magnetic resonance reconstruction model to obtain the reconstructed image
  • Step 4 Calculate the difference between the fully sampled image corresponding to the sample image and the reconstructed image
  • Step 5 Use the back propagation algorithm in the network according to the difference to update the DCT filter operator, TV filter operator and regularization parameters in the magnetic resonance reconstruction model until the magnetic resonance image reconstruction model meets Preset accuracy requirements;
  • Step 6 Perform magnetic resonance image reconstruction based on the magnetic resonance image reconstruction model that meets the preset accuracy requirements.
  • the DCT filter operator in the spatial domain and the TV filter operator in the time domain are used as regularization parameters to establish a magnetic resonance image reconstruction model.
  • the spatial direction and the time direction remove redundancy, so the reconstruction effect can be effectively improved.
  • the adjustment of the DCT filter operator and the TV filter operator in the time domain can be effectively improved through the self-learning method of the network, which can effectively improve the accuracy of the magnetic resonance image reconstruction model.
  • the magnetic resonance image reconstruction model established by the above method can Efficiently reconstruct highly undersampled images to obtain images with high reconstruction accuracy and reconstruction speed, which can effectively shorten the time of magnetic resonance scanning without losing spatial resolution.
  • An embodiment of the present application also provides a computer-readable storage medium capable of implementing all the steps in the adaptive parameter learning dynamic magnetic resonance image reconstruction method in the foregoing embodiment, the computer-readable storage medium storing a computer program, When the computer program is executed by the processor, all steps of the dynamic parameter resonance method for adaptive parameter learning in the above embodiments are implemented. For example, when the processor executes the computer program, the following steps are realized:
  • Step 1 Use the DCT filter operator in the spatial domain and the TV filter operator in the time domain as the regularization terms of CS-MRI;
  • Step 2 According to the regularization terms introduced by the DCT filter operator and the TV filter operator, a magnetic resonance image reconstruction model is established;
  • Step 3 Reconstruct the sample image through the established magnetic resonance reconstruction model to obtain the reconstructed image
  • Step 4 Calculate the difference between the fully sampled image corresponding to the sample image and the reconstructed image
  • Step 5 Using the back propagation algorithm in the network according to the difference, update the DCT filter operator, TV filter operator and regularization parameters in the magnetic resonance reconstruction model to realize the reconstruction of the magnetic resonance Optimal training of the model until the magnetic resonance image reconstruction model meets the preset accuracy requirements;
  • Step 6 Perform magnetic resonance image reconstruction based on the magnetic resonance image reconstruction model that meets the preset accuracy requirements.
  • the DCT filter operator in the spatial domain and the TV filter operator in the time domain are used as regularization parameters to establish a magnetic resonance image reconstruction model.
  • the spatial direction and the time direction remove redundancy, so the reconstruction effect can be effectively improved.
  • the adjustment of the DCT filter operator and the TV filter operator in the time domain can be effectively improved through the self-learning method of the network, which can effectively improve the accuracy of the magnetic resonance image reconstruction model.
  • the magnetic resonance image reconstruction model established by the above method can Efficiently reconstruct highly undersampled images to obtain images with high reconstruction accuracy and reconstruction speed, which can effectively shorten the time of magnetic resonance scanning without losing spatial resolution.
  • the system, device, module or unit explained in the above embodiments may be specifically implemented by a computer chip or entity, or implemented by a product with a certain function.
  • a typical implementation device is a computer.
  • the computer may be, for example, a personal computer, a laptop computer, an on-board human-machine interaction device, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet A computer, a wearable device, or any combination of these devices.
  • the functions are divided into various modules and described separately.
  • the functions of each module may be implemented in one or more software and/or hardware, or the modules that implement the same function may be implemented by a combination of multiple submodules or subunits.
  • the device embodiments described above are only schematic.
  • the division of the unit is only a division of logical functions.
  • there may be another division manner for example, multiple units or components may be combined or integrated To another system, or some features can be ignored, or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.
  • controller in addition to implementing the controller in the form of pure computer-readable program code, it is entirely possible to logically program method steps to make the controller use logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded To achieve the same function in the form of a microcontroller, etc. Therefore, such a controller can be regarded as a hardware component, and the device for implementing various functions included therein can also be regarded as a structure within the hardware component. Or even, the means for realizing various functions can be regarded as both a software module of an implementation method and a structure within a hardware component.
  • each flow and/or block in the flowchart and/or block diagram and a combination of the flow and/or block in the flowchart and/or block diagram can be implemented by computer program instructions.
  • These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processing machine, or other programmable data processing device to produce a machine that enables the generation of instructions executed by the processor of the computer or other programmable data processing device
  • These computer program instructions may also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, the instructions
  • the device implements the functions specified in one block or multiple blocks of the flowchart one flow or multiple flows and/or block diagrams.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operating steps are performed on the computer or other programmable device to produce computer-implemented processing, which is executed on the computer or other programmable device
  • the instructions provide steps for implementing the functions specified in one block or multiple blocks of the flowchart one flow or multiple flows and/or block diagrams.
  • the computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
  • processors CPUs
  • input/output interfaces network interfaces
  • memory volatile and non-volatile memory
  • the memory may include non-permanent memory, random access memory (RAM) and/or non-volatile memory in computer-readable media, such as read only memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.
  • RAM random access memory
  • ROM read only memory
  • flash RAM flash random access memory
  • Computer-readable media including permanent and non-permanent, removable and non-removable media, can store information by any method or technology.
  • the information may be computer readable instructions, data structures, modules of programs, or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, Magnetic tape cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices.
  • computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.
  • the embodiments of the present specification may be provided as methods, systems, or computer program products. Therefore, the embodiments of the present specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present specification may take the form of computer program products implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.
  • computer usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • Embodiments of this specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules.
  • program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types.
  • the embodiments of the present specification may also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected through a communication network.
  • program modules may be located in local and remote computer storage media including storage devices.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

一种自适应参数学习的动态磁共振图像重建方法和装置,其中,该方法是对CS-MRI模型中的正则化项进行改进,包括在空间域使用DCT和在时间域使用TV对动态磁共振图像进行去冗余,并且利用卷积神经网络来对CS-MRI中大量的参数进行自适应学习,建立磁共振图像重建模型;通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;计算全采样图像与重建图像的差值;根据差值,利用网络中的反向传播算法,对模型中的参数,包括DCT、TV滤波算子和正则化参数等进行更新。该方法可以对高度欠采样图像进行高效的重建,得到重建精度和重建速度很高的图像,从而可以在不损失空间分辨率的情况下有效缩短磁共振扫描的时间。

Description

一种自适应参数学习的动态磁共振图像重建方法和装置 技术领域
本申请属于图像处理技术领域,尤其涉及一种自适应参数学习的动态磁共振图像重建方法和装置。
背景技术
磁共振成像(Magnetic Resonance Imaging,简称为MRI)可以准确获取人体器官和组织的生理功能、解剖结构、病变和功能信息。然而,传统的磁共振信号的扫描和图像的重建需要很长的时间。例如:心脏、灌注和功能成像这些对实时性要求比较高的项目,磁共振成像往往无法满足。进一步的,由于扫描时间长,患者可能会感到不舒服并引入运动伪影。
然而,针对如何缩短磁共振扫描的时间提升重建图像的精度,目前上述提出有效的解决方案。
发明内容
本申请目的在于提供一种自适应参数学习的动态磁共振图像重建方法和装置,通过建立的磁共振图像重建模型可以对欠采样图像进行高效的重建,得到所需的图像,从而可以有效缩短磁共振扫描的时间。本申请提供一种自适应参数学习的动态磁共振图像重建方法和装置是这样实现的:
一种自适应参数学习的动态磁共振图像重建方法,所述方法包括:
将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI的正则化项;
根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
在一个实施方式中,根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型,包括:
利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式;
对所述稀疏性表达公式进行融合转换,得到目标公式;
利用交替方向乘子法,将所述目标公式转换为多个子问题求解公式;
将所述多个子问题求解公式进行网络化,得到磁共振图像重建模型。
在一个实施方式中,利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式,包括:
将如下公式的滤波算子设置为DCT算子和TV算子:
Figure PCTCN2018121193-appb-000001
其中,arg min f(x)表示使得函数f(x)取得其最小值的所有自变量x的集合,
Figure PCTCN2018121193-appb-000002
是重建的磁共振图像,N t={x 1,x 2,...,x T},表示时间方向共有T帧图像,y表示欠采样的k空间数据,A=PF,其中,P为一个欠采样矩阵,F表示傅里叶变换,Φ表示滤波算子,λ表示参数。
得到如下公式:
Figure PCTCN2018121193-appb-000003
其中,
Figure PCTCN2018121193-appb-000004
表示空间域的图像,
Figure PCTCN2018121193-appb-000005
表示时间域的图像,Φ 1表示DCT滤波算子、Φ 2表示TV滤波算子,||·|| DCT表示离散余弦变换,||·|| TV表示总变差变换,λ 1、λ 2为平衡数据保真项与正则化项的参数。
在一个实施方式中,所述目标公式为:
Figure PCTCN2018121193-appb-000006
其中,z表示辅助变量,α表示拉格朗日乘子,g(·)表示l 2,p范数的一个近似函数,ρ表示惩罚参数。
在一个实施方式中,多个子问题求解公式为:
Figure PCTCN2018121193-appb-000007
其中,T表示转置,β表示拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降 中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
Figure PCTCN2018121193-appb-000008
l r表示步长的大小,
Figure PCTCN2018121193-appb-000009
表示更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵。
在一个实施方式中,将所述多个子问题求解公式进行网络化得到:
Figure PCTCN2018121193-appb-000010
其中,I表示单位矩阵,C 1、C 2分别表示两个卷积层,w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量。w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量,S PLF(·)表示一个由点
Figure PCTCN2018121193-appb-000011
控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
一种自适应参数学习的动态磁共振图像重建装置,该装置包括:
替换模块,用于将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI的正则化项;
建立模块,用于根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
重建模块,用于通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
计算模块,用于计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
迭代更新模块,用于根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
应用模块,用于根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
在一个实施方式中,所述建立模块包括;
生成单元,用于利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式;
第一转换单元,用于对所述稀疏性表达公式进行融合转换,得到目标公式;
第二转换单元,用于利用交替方向乘子法,将所述目标公式转换为多个子问题求解公式;
网络化单元,用于将所述多个子问题求解公式进行网络化,得到磁共振图像重建模型。
在一个实施方式中,所述生成单元,具体用于将如下公式的滤波算子设置为DCT算子和TV算子:
Figure PCTCN2018121193-appb-000012
其中,arg min f(x)表示使得函数f(x)取得其最小值的所有自变量x的集合,
Figure PCTCN2018121193-appb-000013
是重建的磁共振图像,y表示欠采样的k空间数据,A=PF,其中,P为一个欠采样矩阵,F表示傅里叶变换,Φ表示滤波算子,λ表示参数。
得到如下公式:
Figure PCTCN2018121193-appb-000014
其中,
Figure PCTCN2018121193-appb-000015
表示空间域的图像,
Figure PCTCN2018121193-appb-000016
表示时间域的图像,N t={x 1,x 2,...,x T},表示时间方向共有T帧图像。Φ 1表示DCT滤波算子、Φ 2表示TV滤波算子,||·|| DCT表示离散余弦变换,||·|| TV表示总变差变换,λ 1、λ 2为平衡数据保真项与正则化项的参数。
在一个实施方式中,所述目标公式为:
Figure PCTCN2018121193-appb-000017
其中,z表示辅助变量,α表示拉格朗日乘子,g(·)表示l 2,p范数的一个近似函数,ρ表示惩罚参数。
在一个实施方式中,多个子问题求解公式为:
Figure PCTCN2018121193-appb-000018
其中,T表示转置,β表示拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
Figure PCTCN2018121193-appb-000019
l r表示步长的大小,
Figure PCTCN2018121193-appb-000020
表示更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵。
在一个实施方式中,将所述多个子问题求解公式进行网络化得到:
Figure PCTCN2018121193-appb-000021
其中,I表示单位矩阵,C 1、C 2分别表示两个卷积层,w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量。w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量,S PLF(·)表示一个由点
Figure PCTCN2018121193-appb-000022
控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
一种终端设备,包括处理器以及用于存储处理器可执行指令的存储器,所述处理器执行所述指令时实现如下步骤:
将空间域的DCT滤波算子和时间域的TV滤波算子,作为正则化参数;
根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
根据所述差值,对所述磁共振重建模型进行反向迭代,以更新DCT滤波算子和TV滤波算子的值,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
一种计算机可读存储介质,其上存储有计算机指令,所述指令被执行时实现如下步骤:
将空间域的DCT滤波算子和时间域的TV滤波算子,作为正则化参数;
根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
根据所述差值,对所述磁共振重建模型进行反向迭代,以更新DCT滤波算子和TV滤波算子的值,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足 预设精度要求;
根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
本申请提供的自适应参数学习的动态磁共振图像重建方法和装置,由于磁共振图像是稀疏的,并且存在噪声,通过将空间域的DCT滤波算子和时间域的TV滤波算子,作为正则化参数,建立磁共振图像重建模型,可以分别在空间方向和时间方向去除冗余,因此,可以有效提升重建效果。进一步的,通过网络自主的学习的方式,对上述DCT滤波算子和时间域的TV滤波算子进行调整,可以有效提升磁共振图像重建模型的精度,通过上述方式建立的磁共振图像重建模型可以对高度欠采样图像进行高效的重建,得到重建精度和重建速度很高的图像,从而可以在不损失空间分辨率的情况下有效缩短磁共振扫描的时间。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案,下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请中记载的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1是本申请提供的自适应参数学习的动态磁共振图像重建方法一种实施例的方法流程图;
图2是本申请提供的自适应参数学习的动态磁共振图像稀疏处理的示意图;
图3是本申请提供的网络化后得到的神经网络的模型示意图;
图4是本申请提供的终端设备的架构示意图;
图5是本申请提供的自适应参数学习的动态磁共振图像重建装置的结构框图。
具体实施方式
为了使本技术领域的人员更好地理解本申请中的技术方案,下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都应当属于本申请保护的范围。
目前加快磁共振成像速度的方式主要有:并行磁共振成像(Parallel MR imaging)与 k空间欠采样。本例主要是针对k空间欠采样来提高磁共振成像速度进行的研究。考虑到,压缩感知磁共振成像(Compressed Sensing-Magnetic Resonance Imaging,简称为CS-MRI)由数据保真项和正则化项组成,对于正则化项的研究有小波稀疏正则化、总变差稀疏正则化和自适应字典学习;进一步的,有研究表明低秩模型对动态磁共振成像是有效的。现有的基于稀疏和低秩的方法虽然有比较好的重建效果,但仍然无法很好地重建信号,基于低秩方法并没有考虑图像中的噪声,而基于稀疏的方法,滤波器的选择是人为选择的,不具有代表性,基于字典学习的方法重建时间比较长。
基于此,在本例中,从实际应用的角度出发,基于压缩感知理论重建算法以快速实现磁共振成像。由于磁共振图像是稀疏的,并且存在噪声,在本例中提出了一种在空间上利用离散余弦(Discrete Cosine Transform,简称为DCT)滤波,在时间方向上利用总变差(Total Variation,简称为TV)滤波,由于其分别在空间方向和时间方向去除冗余,因此,可以有效提升重建效果。
具体的,利用磁共振图像是稀疏的,利用离散余弦变换将图像变换到稀疏域,滤波器的值可以是网络自主学习得到的,不需要人为进行选择。在时间方向上利用TV进行滤波,不仅能去除噪声,还能进一步增强MR图像的稀疏性,提高图像的重建质量。
针对磁共振成像,尤其是动态磁共振成像扫描速度慢的问题,在本例中通过欠采样的方法来提高动态磁共振成像的扫描速度。针对欠采样直接重建会产生伪影的问题,在本例中,通过利用CS-MRI技术,对其中的正则化项进行探索,并结合深度学习自适应学习的方法,使得欠采样的磁共振图像能快速重建出高分辨率的近似于全采样的图像。
如图1所示,是本申请所述一种自适应参数学习的动态磁共振图像重建方法一个实施例的方法流程图。虽然本申请提供了如下述实施例或附图所示的方法操作步骤或装置结构,但基于常规或者无需创造性的劳动在所述方法或装置中可以包括更多或者更少的操作步骤或模块单元。在逻辑性上不存在必要因果关系的步骤或结构中,这些步骤的执行顺序或装置的模块结构不限于本申请实施例描述及附图所示的执行顺序或模块结构。所述的方法或模块结构的在实际中的装置或终端产品应用时,可以按照实施例或者附图所示的方法或模块结构连接进行顺序执行或者并行执行(例如并行处理器或者多线程处理的环境,甚至分布式处理环境)。
具体的,如图1所示,本申请一种实施例提供的一种自适应参数学习的动态磁共振图像重建方法可以包括(步骤101到步骤106):
步骤101:将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI的正 则化项;
步骤102:根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
具体的,根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型可以包括:
S1:利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式;
例如:可以将如下公式的滤波算子设置为DCT滤波算子和TV滤波算子:
Figure PCTCN2018121193-appb-000023
其中,arg min f(x)表示使得函数f(x)取得其最小值的所有自变量x的集合,
Figure PCTCN2018121193-appb-000024
是重建的磁共振图像,y表示欠采样的k空间数据,A=PF,其中,P为一个欠采样矩阵,F表示傅里叶变换,Φ表示滤波算子,λ表示参数。
得到如下公式:
Figure PCTCN2018121193-appb-000025
其中,
Figure PCTCN2018121193-appb-000026
表示空间域的图像,
Figure PCTCN2018121193-appb-000027
表示时间域的图像,N t={x 1,x 2,...,x T},表示时间方向共有T帧图像。Φ 1表示DCT滤波算子、Φ 2表示TV滤波算子,||·|| DCT表示离散余弦变换,||·|| TV表示总变差变换,λ 1、λ 2为平衡数据保真项与正则化项的参数。
即,将原始公式中的滤波算子选定为DCT滤波算子和TV滤波算子,使得可以在时间和空间上进行去冗余。
S2:对所述稀疏性表达公式进行融合转换,得到目标公式;
其中,目标公式可以表示为:
Figure PCTCN2018121193-appb-000028
其中,z表示辅助变量,α表示拉格朗日乘子,g(·)表示l 2,p范数的一个近似函数,ρ表示惩罚参数。
具体的,该目标公式可以是按照如下方式转换得到的:
考虑到在实际应用中,用于稀疏正则化的稀疏诱导范数包括:l 0范数,l 1范数,l 2,1范 数定义为行向量的l 2范数之和,最小化l 2,1范数的目的是选择尽量少的非零行向量。例如:利用l 2,p矩阵范数(0<p<1)做特征选择可以取得更稀疏的解,因此,选择l 2,p范数,并另α=β。相应的,上述的公式:
Figure PCTCN2018121193-appb-000029
可以简化为:
Figure PCTCN2018121193-appb-000030
其中,Φ表示Φ 1与Φ 2的融合,λ l=λ 1=λ 2为正则化参数,g(·)是l 2,p范数的一个近似函数。
利用展开迭代的方法对上述方式进行求解,在图像域引入辅助变量z,可以将公式:
Figure PCTCN2018121193-appb-000031
转换为:
Figure PCTCN2018121193-appb-000032
s.t.z=x
其中,上述的增广拉格朗日函数为:
Figure PCTCN2018121193-appb-000033
该增广拉格朗日函数就是目标公式。通过上述方式就完成了从基础公式到目标公式的转换。
S3:利用交替方向乘子法,将所述目标公式转换为多个子问题求解公式;
其中,交替方向乘子法(Alternating Direction Method of Multipliers,简称为ADMM)是一种求解优化问题的计算框架,适用于求解分布式凸优化问题,特别是统计学习问题。ADMM通过分解协调(Decomposition-Coordination)过程,将大的全局问题分解为多个较小、较容易求解的局部子问题,并通过协调子问题的解而得到大的全局问题的解。
在本例中,通过交替方向乘子法可以将上述目标公式转换为:
多个子问题求解公式为:
Figure PCTCN2018121193-appb-000034
其中,T表示转置,β表示拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
Figure PCTCN2018121193-appb-000035
l r表示步长的大小,
Figure PCTCN2018121193-appb-000036
是更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵(如DCT、小波变换等)。
S4:将所述多个子问题求解公式进行网络化,得到磁共振图像重建模型。
具体的,将上述多个子问题求解公式进行网络化可以得到:
Figure PCTCN2018121193-appb-000037
其中,I表示单位矩阵。C 1、C 2分别表示两个卷积层。w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量,w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量。S PLF(·)表示一个由点
Figure PCTCN2018121193-appb-000038
控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
步骤103:通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
通过上述方式网络化之后,就可以得到磁共振重建模型的神经网络,基于该神经网络可以对样本图像进行重建,以得到重建后的图像。
步骤104:计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
步骤105:根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
进一步的,在重建之后,可以对重建后的图像与原图像通过标准均方误差可以计算loss值,表达式如下:
Figure PCTCN2018121193-appb-000039
其中,Λ表示训练集的个数,基于loss值可以对该网络模型进行反向传播更新参数,以便基于loss值对网络模型中的各个参数进行优化更新,以使得最终确定的网络模型的参数,可以保证网络模型的重建精度更高。
其中,反向传播参数的更新过程,可以是Loss层梯度更新,可以表示为:
Figure PCTCN2018121193-appb-000040
步骤106:根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
在上例中,CS-MRI模型中的正则化项进行改进,包括在空间域使用离散余弦滤波算子(DCT)和在时间域使用总变差滤波算子(TV)对动态磁共振图像进行去冗余,并且利用卷积神经网络来对CS-MRI中大量的参数进行自适应学习,建立磁共振图像重建模型;通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;计算全采样图像与重建图像的差值;根据差值,利用网络中的反向传播算法,对模型中的参数,包括DCT、TV滤波算子和正则化参数等进行更新,实现对磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。通过上述方式建立的磁共振图像重建模型可以对高度欠采样图像进行高效的重建,得到重建精度和重建速度很高的图像,从而可以在不损失空间分辨率的情况下有效缩短磁共振扫描的时间。
下面结合一个具体实施例对上述方法进行说明,然而,值得注意的是,该具体实施例仅是为了更好地说明本申请,并不构成对本申请的不当限定。
为了进行图像重建,引入了正则化项,为了构造好的正则项,往往需要获取图像的特征信息和先验知识。然而,已有的迭代方法需要人为设计正则项的权重参数,而不同的图像所对应的最优参数可能是不同的,这使得这类方法的普适性不是很强,且现有的迭代方法所需的时间比较长。
在本例中,如图2所示,磁共振图像是稀疏的,利用磁共振图像(MRI)本身特有的稀疏低秩等先验信息,在正则化项中的空间域引入DCT滤波,时间域引入TV滤波,这两者的组合可以极大低降图像的冗余。进一步的,利用迭代重建的方法进行求解,因为在迭代过程中会用到真实的投影数据,因此重建结果在理论上将会更为精确。将迭代 过程网络化,可以自主学习图像的正则化参数,不需人为设定。
具体的,利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,可以按照如下公式进行:
Figure PCTCN2018121193-appb-000041
其中,
Figure PCTCN2018121193-appb-000042
是重建的MRI图像,y是欠采样的k空间数据,A=PF,其中,P是一个欠采样矩阵,F是傅里叶变换。Φ表示滤波算子(例如:小波变换DWT、离散余弦变换DCT、总变差TV),λ 1、λ 2是平衡数据保真项与正则化项的参数。
根据图像的稀疏性,可以将上述公式1改为:
Figure PCTCN2018121193-appb-000043
其中,
Figure PCTCN2018121193-appb-000044
表示空间域的图像,
Figure PCTCN2018121193-appb-000045
表示时间域的图像,α和β是两个正则化参数。Φ 1表示DCT的滤波算子、Φ 2表示TV的滤波算子。||·|| DCT表示离散余弦变换,||·|| TV表示总变差变换。
Figure PCTCN2018121193-appb-000046
其中,
Figure PCTCN2018121193-appb-000047
表示前一帧的有限差分算子和
Figure PCTCN2018121193-appb-000048
表示后一帧的有限差分算子。
在实际应用中,用于稀疏正则化的稀疏诱导范数包括:l 0范数,l 1范数,l 2,1范数定义为行向量的l 2范数之和,最小化l 2,1范数的目的是选择尽量少的非零行向量。然而,这都只能应用于向量范数。基于深度学习的出现,特征选择、多视角学习等领域的研究者提出用矩阵范数来做稀疏正则化,例如:利用l 2,p矩阵范数(0<p<1)做特征选择可以取得更稀疏的解。在本例中,选择l 2,p范数,为简化说明,另α=β。相应的,上述公式2可以简化为:
Figure PCTCN2018121193-appb-000049
其中,Φ表示Φ 1与Φ 2的融合,λ l=λ 1=λ 2为正则化参数,g(·)是l 2,p范数的一个近似函数。
利用展开迭代的方法来求解上述公式3,在图像域引入辅助变量z,上述公式3可以转换为:
Figure PCTCN2018121193-appb-000050
s.t.z=x            (公式4)
其中,上述公式4的增广拉格朗日函数为:
Figure PCTCN2018121193-appb-000051
利用交替方向乘子法,将上述公式5转化为若干个子问题来求解,结果如下:
Figure PCTCN2018121193-appb-000052
其中,
Figure PCTCN2018121193-appb-000053
β为拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
Figure PCTCN2018121193-appb-000054
l r表示步长的大小,
Figure PCTCN2018121193-appb-000055
表示更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵。
为了将上述公式6网络化,可以将上述公式6修改为:
Figure PCTCN2018121193-appb-000056
该公式为前向传播图像的重建过程,I表示单位矩阵,C 1、C 2分别表示两个卷积层,w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量。w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量,S PLF(·)表示一个由点
Figure PCTCN2018121193-appb-000057
控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,如图3所示,图3中DCTV-Net表示网络流程图,其中,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
在重建之后,重建后的图像与原图像通过标准均方误差可以计算loss值,表达式如下:
Figure PCTCN2018121193-appb-000058
其中,Λ表示训练集的个数,图2中正向箭头是迭代一次重建的结果,反向箭头表 示反向传播更新参数的过程。
其中,反向传播参数的更新过程,可以是Loss层梯度更新,可以表示为:
Figure PCTCN2018121193-appb-000059
在上例中,通过时频域稀疏网络化对磁共振图像进行稀疏约束,从而使得在不损失空间分辨率的情况下获得更高的加速倍数。进一步的,利用自适应学习的方法,将CS-MRI中大量的求解参数用CNN来替代,这样可以避免人为选择所带来的偶然性。相较于深度学习的重建方法,当数据量有限时,在本例中所使用的方法相较于传统的深度学习方法具有更高的重建精度。在相同加速倍数的条件下,本例的方式重建精度更高,重建效果更好。在本例中,利用神经网络来学习这些正则化参数,并且通过线下训练获得训练好的模型,线上测试仅需要3秒左右的时间就能获得高度欠采样的MRI重建图像。
本申请上述实施例所提供的方法实施例可以在终端设备、计算机终端或者类似的运算装置中执行。以运行在终端设备上为例,图4是本发明实施例的一种自适应参数学习的动态磁共振图像重建方法的计算机终端的硬件结构框图。如图4所示,终端设备10可以包括一个或多个(图中仅示出一个)处理器102(处理器102可以包括但不限于微处理器MCU或可编程逻辑器件FPGA等的处理装置)、用于存储数据的存储器104、以及用于通信功能的传输模块106。本领域普通技术人员可以理解,图4所示的结构仅为示意,其并不对上述电子装置的结构造成限定。例如,计算机终端10还可包括比图4中所示更多或者更少的组件,或者具有与图4所示不同的配置。
存储器104可用于存储应用软件的软件程序以及模块,如本发明实施例中的自适应参数学习的动态磁共振图像重建方法对应的程序指令/模块,处理器102通过运行存储在存储器104内的软件程序以及模块,从而执行各种功能应用以及数据处理,即实现上述的自适应参数学习的动态磁共振图像重建方法。存储器104可包括高速随机存储器,还可包括非易失性存储器,如一个或者多个磁性存储装置、闪存、或者其他非易失性固态存储器。在一些实例中,存储器104可进一步包括相对于处理器102远程设置的存储器,这些远程存储器可以通过网络连接至计算机终端10。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
传输模块106用于经由一个网络接收或者发送数据。上述的网络具体实例可包括计算机终端10的通信供应商提供的无线网络。在一个实例中,传输模块106包括一个网络 适配器(Network Interface Controller,NIC),其可通过基站与其他网络设备相连从而可与互联网进行通讯。在一个实例中,传输模块106可以为射频(Radio Frequency,RF)模块,其用于通过无线方式与互联网进行通讯。
在软件层面,上述装置可以如图5所示,可以包括:
替换模块501,用于将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI的正则化项;
建立模块502,用于根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
重建模块503,用于通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
计算模块504,用于计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
迭代更新模块505,用于根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
应用模块506,用于根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
在一个实施方式中,上述建立模块502可以包括:生成单元,用于利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式;第一转换单元,用于对所述稀疏性表达公式进行融合转换,得到目标公式;第二转换单元,用于利用交替方向乘子法,将所述目标公式转换为多个子问题求解公式;网络化单元,用于将所述多个子问题求解公式进行网络化,得到磁共振图像重建模型。
在一个实施方式中,上述生成单元,具体可以用于将如下公式的滤波算子设置为DCT算子和TV算子:
Figure PCTCN2018121193-appb-000060
其中,arg min f(x)表示使得函数f(x)取得其最小值的所有自变量x的集合,
Figure PCTCN2018121193-appb-000061
是重建的磁共振图像,y表示欠采样的k空间数据,A=PF,其中,P为一个欠采样矩阵,F表示傅里叶变换,Φ表示滤波算子,λ表示参数。
得到如下公式:
Figure PCTCN2018121193-appb-000062
其中,
Figure PCTCN2018121193-appb-000063
表示空间域的图像,
Figure PCTCN2018121193-appb-000064
表示时间域的图像,N t={x 1,x 2,...,x T},表示时间方向共有T帧图像,Φ 1表示DCT滤波算子、Φ 2表示TV滤波算子,||·|| DCT表示离散余弦变换,||·|| TV表示总变差变换,λ 1、λ 2为平衡数据保真项与正则化项的参数。
在一个实施方式中,上述目标公式可以表示为:
Figure PCTCN2018121193-appb-000065
其中,z表示辅助变量,α表示拉格朗日乘子,g(·)表示l 2,p范数的一个近似函数,ρ表示惩罚参数。
在一个实施方式中,上述多个子问题求解公式可以表示为:
Figure PCTCN2018121193-appb-000066
其中,T表示转置,β表示拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
Figure PCTCN2018121193-appb-000067
l r表示步长的大小,
Figure PCTCN2018121193-appb-000068
表示更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵。
在一个实施方式中,将多个子问题求解公式进行网络化可以得到:
Figure PCTCN2018121193-appb-000069
其中,I表示单位矩阵,C 1、C 2分别表示两个卷积层,w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量。w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量,S PLF(·)表示一个由点
Figure PCTCN2018121193-appb-000070
控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
本申请的实施例还提供能够实现上述实施例中的自适应参数学习的动态磁共振图像 重建方法中全部步骤的一种电子设备的具体实施方式,所述电子设备具体包括如下内容:
处理器(processor)、存储器(memory)、通信接口(Communications Interface)和总线;
其中,所述处理器、存储器、通信接口通过所述总线完成相互间的通信;所述处理器用于调用所述存储器中的计算机程序,所述处理器执行所述计算机程序时实现上述实施例中的自适应参数学习的动态磁共振图像重建方法中的全部步骤,例如,所述处理器执行所述计算机程序时实现下述步骤:
步骤1:将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI的正则化项;
步骤2:根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
步骤3:通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
步骤4:计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
步骤5:根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,直至磁共振图像重建模型满足预设精度要求;
步骤6:根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
从上述描述可知,由于磁共振图像是稀疏的,并且存在噪声,通过将空间域的DCT滤波算子和时间域的TV滤波算子,作为正则化参数,建立磁共振图像重建模型,可以分别在空间方向和时间方向去除冗余,因此,可以有效提升重建效果。进一步的,通过网络自主的学习的方式,对上述DCT滤波算子和时间域的TV滤波算子进行调整,可以有效提升磁共振图像重建模型的精度,通过上述方式建立的磁共振图像重建模型可以对高度欠采样图像进行高效的重建,得到重建精度和重建速度很高的图像,从而可以在不损失空间分辨率的情况下有效缩短磁共振扫描的时间。
本申请的实施例还提供能够实现上述实施例中的自适应参数学习的动态磁共振图像重建方法中全部步骤的一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,该计算机程序被处理器执行时实现上述实施例中的自适应参数学习的动态磁共振图像重建方法的全部步骤,例如,所述处理器执行所述计算机程序时实现下述步骤:
步骤1:将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI的正则化项;
步骤2:根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
步骤3:通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
步骤4:计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
步骤5:根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
步骤6:根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
从上述描述可知,由于磁共振图像是稀疏的,并且存在噪声,通过将空间域的DCT滤波算子和时间域的TV滤波算子,作为正则化参数,建立磁共振图像重建模型,可以分别在空间方向和时间方向去除冗余,因此,可以有效提升重建效果。进一步的,通过网络自主的学习的方式,对上述DCT滤波算子和时间域的TV滤波算子进行调整,可以有效提升磁共振图像重建模型的精度,通过上述方式建立的磁共振图像重建模型可以对高度欠采样图像进行高效的重建,得到重建精度和重建速度很高的图像,从而可以在不损失空间分辨率的情况下有效缩短磁共振扫描的时间。
本说明书中的各个实施例均采用递进的方式描述,各个实施例之间相同相似的部分互相参见即可,每个实施例重点说明的都是与其他实施例的不同之处。尤其,对于硬件+程序类实施例而言,由于其基本相似于方法实施例,所以描述的比较简单,相关之处参见方法实施例的部分说明即可。
上述对本说明书特定实施例进行了描述。其它实施例在所附权利要求书的范围内。在一些情况下,在权利要求书中记载的动作或步骤可以按照不同于实施例中的顺序来执行并且仍然可以实现期望的结果。另外,在附图中描绘的过程不一定要求示出的特定顺序或者连续顺序才能实现期望的结果。在某些实施方式中,多任务处理和并行处理也是可以的或者可能是有利的。
虽然本申请提供了如实施例或流程图所述的方法操作步骤,但基于常规或者无创造性的劳动可以包括更多或者更少的操作步骤。实施例中列举的步骤顺序仅仅为众多步骤执行顺序中的一种方式,不代表唯一的执行顺序。在实际中的装置或客户端产品执行时,可以按照实施例或者附图所示的方法顺序执行或者并行执行(例如并行处理器或者多线程处理的环境)。
上述实施例阐明的系统、装置、模块或单元,具体可以由计算机芯片或实体实现,或者由具有某种功能的产品来实现。一种典型的实现设备为计算机。具体的,计算机例如可以为个人计算机、膝上型计算机、车载人机交互设备、蜂窝电话、相机电话、智能电话、个人数字助理、媒体播放器、导航设备、电子邮件设备、游戏控制台、平板计算机、可穿戴设备或者这些设备中的任何设备的组合。
虽然本说明书实施例提供了如实施例或流程图所述的方法操作步骤,但基于常规或者无创造性的手段可以包括更多或者更少的操作步骤。实施例中列举的步骤顺序仅仅为众多步骤执行顺序中的一种方式,不代表唯一的执行顺序。在实际中的装置或终端产品执行时,可以按照实施例或者附图所示的方法顺序执行或者并行执行(例如并行处理器或者多线程处理的环境,甚至为分布式数据处理环境)。术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、产品或者设备不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、产品或者设备所固有的要素。在没有更多限制的情况下,并不排除在包括所述要素的过程、方法、产品或者设备中还存在另外的相同或等同要素。
为了描述的方便,描述以上装置时以功能分为各种模块分别描述。当然,在实施本说明书实施例时可以把各模块的功能在同一个或多个软件和/或硬件中实现,也可以将实现同一功能的模块由多个子模块或子单元的组合实现等。以上所描述的装置实施例仅仅是示意性的,例如,所述单元的划分,仅仅为一种逻辑功能划分,实际实现时可以有另外的划分方式,例如多个单元或组件可以结合或者可以集成到另一个系统,或一些特征可以忽略,或不执行。另一点,所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口,装置或单元的间接耦合或通信连接,可以是电性,机械或其它的形式。
本领域技术人员也知道,除了以纯计算机可读程序代码方式实现控制器以外,完全可以通过将方法步骤进行逻辑编程来使得控制器以逻辑门、开关、专用集成电路、可编程逻辑控制器和嵌入微控制器等的形式来实现相同功能。因此这种控制器可以被认为是一种硬件部件,而对其内部包括的用于实现各种功能的装置也可以视为硬件部件内的结构。或者甚至,可以将用于实现各种功能的装置视为既可以是实现方法的软件模块又可以是硬件部件内的结构。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每 一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器,使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中,使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品,该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上,使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理,从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
在一个典型的配置中,计算设备包括一个或多个处理器(CPU)、输入/输出接口、网络接口和内存。
内存可能包括计算机可读介质中的非永久性存储器,随机存取存储器(RAM)和/或非易失性内存等形式,如只读存储器(ROM)或闪存(flash RAM)。内存是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括,但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带,磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质,可用于存储可以被计算设备访问的信息。按照本文中的界定,计算机可读介质不包括暂存电脑可读媒体(transitory media),如调制的数据信号和载波。
本领域技术人员应明白,本说明书的实施例可提供为方法、系统或计算机程序产品。因此,本说明书实施例可采用完全硬件实施例、完全软件实施例或结合软件和硬件方面的实施例的形式。而且,本说明书实施例可采用在一个或多个其中包含有计算机可用程 序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本说明书实施例可以在由计算机执行的计算机可执行指令的一般上下文中描述,例如程序模块。一般地,程序模块包括执行特定任务或实现特定抽象数据类型的例程、程序、对象、组件、数据结构等等。也可以在分布式计算环境中实践本说明书实施例,在这些分布式计算环境中,由通过通信网络而被连接的远程处理设备来执行任务。在分布式计算环境中,程序模块可以位于包括存储设备在内的本地和远程计算机存储介质中。
本说明书中的各个实施例均采用递进的方式描述,各个实施例之间相同相似的部分互相参见即可,每个实施例重点说明的都是与其他实施例的不同之处。尤其,对于系统实施例而言,由于其基本相似于方法实施例,所以描述的比较简单,相关之处参见方法实施例的部分说明即可。在本说明书的描述中,参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本说明书实施例的至少一个实施例或示例中。在本说明书中,对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且,描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外,在不相互矛盾的情况下,本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
以上所述仅为本说明书实施例的实施例而已,并不用于限制本说明书实施例。对于本领域技术人员来说,本说明书实施例可以有各种更改和变化。凡在本说明书实施例的精神和原理之内所作的任何修改、等同替换、改进等,均应包含在本说明书实施例的权利要求范围之内。

Claims (14)

  1. 一种自适应参数学习的动态磁共振图像重建方法,其特征在于,所述方法包括:
    将空间域的离散余弦DCT滤波算子和时间域的总变差TV滤波算子,作为压缩感知-磁共振成像CS-MRI的正则化项;
    根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
    通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
    计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
    根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
    根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
  2. 根据权利要求1所述的方法,其特征在于,根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型,包括:
    利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式;
    对所述稀疏性表达公式进行融合转换,得到目标公式;
    利用交替方向乘子法,将所述目标公式转换为多个子问题求解公式;
    将所述多个子问题求解公式进行网络化,得到磁共振图像重建模型。
  3. 根据权利要求2所述的方法,其特征在于,利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式,包括:
    将如下公式的滤波算子设置为DCT滤波算子和TV滤波算子:
    Figure PCTCN2018121193-appb-100001
    其中,arg min f(x)表示使得函数f(x)取得其最小值的所有自变量x的集合,
    Figure PCTCN2018121193-appb-100002
    是重建的磁共振图像,N t={x 1,x 2,...,x T},表示时间方向共有T帧图像,y表示欠采样的k空间数据,A=PF,其中,P为一个欠采样矩阵,F表示傅里叶变换,Φ表示滤波算子,λ表示参数;
    得到如下公式:
    Figure PCTCN2018121193-appb-100003
    其中,
    Figure PCTCN2018121193-appb-100004
    表示空间域的图像,
    Figure PCTCN2018121193-appb-100005
    表示时间域的图像,Φ 1表示DCT滤波算子、Φ 2表示TV滤波算子,||·|| DCT表示离散余弦变换,||·|| TV表示总变差变换,λ 1、λ 2为平衡数据保真项与正则化项的参数。
  4. 根据权利要求3所述的方法,其特征在于,所述目标公式为:
    Figure PCTCN2018121193-appb-100006
    其中,z表示辅助变量,α表示拉格朗日乘子,g(·)表示l 2,p范数的一个近似函数,ρ表示惩罚参数。
  5. 根据权利要求4所述的方法,其特征在于,多个子问题求解公式为:
    Figure PCTCN2018121193-appb-100007
    其中,T表示转置,β表示拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
    Figure PCTCN2018121193-appb-100008
    l r表示步长的大小,
    Figure PCTCN2018121193-appb-100009
    表示更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵。
  6. 根据权利要求5所述的方法,其特征在于,将所述多个子问题求解公式进行网络化得到:
    Figure PCTCN2018121193-appb-100010
    其中,I表示单位矩阵,C 1、C 2分别表示两个卷积层,w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量,w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量,S PLF(·)表示一个由点
    Figure PCTCN2018121193-appb-100011
    控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
  7. 一种自适应参数学习的动态磁共振图像重建装置,其特征在于,包括:
    替换模块,用于将空间域的DCT滤波算子和时间域的TV滤波算子,作为CS-MRI 的正则化项;
    建立模块,用于根据引入DCT滤波算子和TV滤波算子的正则化项,建立磁共振图像重建模型;
    重建模块,用于通过建立的磁共振重建模型对样本图像进行重建,得到重建图像;
    计算模块,用于计算所述样本图像对应的全采样图像与所述重建图像之间的差值;
    迭代更新模块,用于根据所述差值,利用网络中的反向传播算法,对所述磁共振重建模型中的DCT滤波算子、TV滤波算子和正则化参数进行更新,实现对所述磁共振重建模型的优化训练,直至磁共振图像重建模型满足预设精度要求;
    应用模块,用于根据满足预设精度要求的磁共振图像重建模型,进行磁共振图像重建。
  8. 根据权利要求7所述的装置,其特征在于,所述建立模块包括:
    生成单元,用于利用联合稀疏模型,在空间域利用DCT滤波,在时间域利用TV滤波来增强图像的稀疏性,得到稀疏性表达公式;
    第一转换单元,用于对所述稀疏性表达公式进行融合转换,得到目标公式;
    第二转换单元,用于利用交替方向乘子法,将所述目标公式转换为多个子问题求解公式;
    网络化单元,用于将所述多个子问题求解公式进行网络化,得到磁共振图像重建模型。
  9. 根据权利要求8所述的装置,其特征在于,所述生成单元,具体用于将如下公式的滤波算子设置为DCT滤波算子和TV滤波算子:
    Figure PCTCN2018121193-appb-100012
    其中,arg min f(x)表示使得函数f(x)取得其最小值的所有自变量x的集合,
    Figure PCTCN2018121193-appb-100013
    是重建的磁共振图像,y表示欠采样的k空间数据,A=PF,其中,P为一个欠采样矩阵,F表示傅里叶变换,Φ表示滤波算子,λ表示参数;
    得到如下公式:
    Figure PCTCN2018121193-appb-100014
    其中,
    Figure PCTCN2018121193-appb-100015
    表示空间域的图像,
    Figure PCTCN2018121193-appb-100016
    表示时间域的图像,N t={x 1,x 2,...,x T},T表示共有T帧图像,Φ 1表示DCT滤波算子、Φ 2表示TV滤波算子,||·|| DCT表示离散 余弦变换,||·|| TV表示总变差变换,λ 1、λ 2为平衡数据保真项与正则化项的参数。
  10. 根据权利要求9所述的装置,其特征在于,所述目标公式为:
    Figure PCTCN2018121193-appb-100017
    其中,z表示辅助变量,α表示拉格朗日乘子,g(·)表示l 2,p范数的一个近似函数,ρ表示惩罚参数。
  11. 根据权利要求10所述的装置,其特征在于,多个子问题求解公式为:
    Figure PCTCN2018121193-appb-100018
    其中,T表示转置,β表示拉格朗日乘子的缩放因子,k∈{1,2,...,N t}表示梯度下降中的迭代次数,n表示第n层,μ 1=(1-l rρ),μ 2=l rρ,分别表示可学习的权重参数,
    Figure PCTCN2018121193-appb-100019
    l r表示步长的大小,
    Figure PCTCN2018121193-appb-100020
    表示更新率,H(·)表示g(·)的梯度,D 1表示变换矩阵。
  12. 根据权利要求11所述的装置,其特征在于,将所述多个子问题求解公式进行网络化得到:
    Figure PCTCN2018121193-appb-100021
    其中,I表示单位矩阵,C 1、C 2分别表示两个卷积层,w 1对应于DCT和TV滤波器的组合,大小为3*3*1*L,b 1表示L维的偏置向量,w 2对应于DCT和TV滤波器的组合,大小为3*3*L,b 2表示1维偏置向量,S PLF(·)表示一个由点
    Figure PCTCN2018121193-appb-100022
    控制的分段线性函数,N C是一个参数,用于控制分段线性函数中的点,Recon表示重建层,Addition表示叠加层,Conv1表示卷积层,Nonlinear表示非线性变换层,Conv2表示卷积层,Multi表示乘子更新层。
  13. 一种终端设备,包括处理器以及用于存储处理器可执行指令的存储器,所述处理器执行所述指令时实现权利要求1至6中任一项所述方法的步骤。
  14. 一种计算机可读存储介质,其上存储有计算机指令,所述指令被执行时实现权利要求1至6中任一项所述方法的步骤。
PCT/CN2018/121193 2018-12-14 2018-12-14 一种自适应参数学习的动态磁共振图像重建方法和装置 WO2020118687A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/121193 WO2020118687A1 (zh) 2018-12-14 2018-12-14 一种自适应参数学习的动态磁共振图像重建方法和装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/121193 WO2020118687A1 (zh) 2018-12-14 2018-12-14 一种自适应参数学习的动态磁共振图像重建方法和装置

Publications (1)

Publication Number Publication Date
WO2020118687A1 true WO2020118687A1 (zh) 2020-06-18

Family

ID=71076764

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/121193 WO2020118687A1 (zh) 2018-12-14 2018-12-14 一种自适应参数学习的动态磁共振图像重建方法和装置

Country Status (1)

Country Link
WO (1) WO2020118687A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102231204A (zh) * 2011-06-20 2011-11-02 南京航空航天大学 一种序列图像自适应正则超分辨率重建方法
CN103871041A (zh) * 2014-03-21 2014-06-18 上海交通大学 基于认知正则化参数构建的图像超分辨率重构方法
CN105022010A (zh) * 2015-07-08 2015-11-04 中国科学院苏州生物医学工程技术研究所 基于正则化迭代的并行磁共振图像重建方法
US10049446B2 (en) * 2015-12-18 2018-08-14 Carestream Health, Inc. Accelerated statistical iterative reconstruction
US20180285695A1 (en) * 2017-03-28 2018-10-04 Siemens Healthcare Gmbh Magnetic Resonance Image Reconstruction System and Method
CN108765293A (zh) * 2018-05-31 2018-11-06 成都信息工程大学 一种基于稀疏表示的磁共振图像重建方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102231204A (zh) * 2011-06-20 2011-11-02 南京航空航天大学 一种序列图像自适应正则超分辨率重建方法
CN103871041A (zh) * 2014-03-21 2014-06-18 上海交通大学 基于认知正则化参数构建的图像超分辨率重构方法
CN105022010A (zh) * 2015-07-08 2015-11-04 中国科学院苏州生物医学工程技术研究所 基于正则化迭代的并行磁共振图像重建方法
US10049446B2 (en) * 2015-12-18 2018-08-14 Carestream Health, Inc. Accelerated statistical iterative reconstruction
US20180285695A1 (en) * 2017-03-28 2018-10-04 Siemens Healthcare Gmbh Magnetic Resonance Image Reconstruction System and Method
CN108765293A (zh) * 2018-05-31 2018-11-06 成都信息工程大学 一种基于稀疏表示的磁共振图像重建方法及系统

Similar Documents

Publication Publication Date Title
CN109671129B (zh) 一种自适应参数学习的动态磁共振图像重建方法和装置
Wu et al. Self-attention convolutional neural network for improved MR image reconstruction
CN114549731B (zh) 视角图像的生成方法、装置、电子设备及存储介质
JP7378642B2 (ja) 生成的非線形人間形状モデル
Terris et al. Image reconstruction algorithms in radio interferometry: From handcrafted to learned regularization denoisers
Beetz et al. Generating subpopulation-specific biventricular anatomy models using conditional point cloud variational autoencoders
CN113906473A (zh) 用于学习可组合的三维(3d)对象的高阶函数网络及其操作方法
Durasov et al. Double refinement network for efficient monocular depth estimation
Li et al. D 3 C 2-Net: Dual-Domain Deep Convolutional Coding Network for Compressive Sensing
WO2020118687A1 (zh) 一种自适应参数学习的动态磁共振图像重建方法和装置
Shamsolmoali et al. Distance-based Weighted Transformer Network for image completion
CN116452715A (zh) 动态人手渲染方法、装置及存储介质
US20230267659A1 (en) Machine-learning techniques for sparse-to-dense spectral reconstruction
Yang et al. Single image super-resolution via a ternary attention network
Iglesias et al. Memetic electromagnetism algorithm for surface reconstruction with rational bivariate Bernstein basis functions
GB2613240A (en) Transformer-based shape models
Faraj et al. VoxMorph: 3-scale freeform deformation of large voxel grids
Repetti et al. A non-convex perspective on calibration and imaging in radio interferometry
Li et al. Parallel magnetic resonance imaging reconstruction algorithm by three-dimension directional Haar tight framelet regularization
Hu et al. Single image super resolution of 3D MRI using local regression and intermodality priors
CN107316343A (zh) 一种基于数据驱动的模型处理方法及设备
Zhang et al. Iterative CT image reconstruction using neural network optimization algorithms
Dong et al. Bayesian Deep Learning for Image Reconstruction: From structured sparsity to uncertainty estimation
Umer et al. Real Image Super-Resolution using GAN through modeling of LR and HR process
Zhou et al. Hflnet: Hierarchical Feature Learning Network for 3d Object Reconstruction

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18942737

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 03.11.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18942737

Country of ref document: EP

Kind code of ref document: A1