WO2015167307A1

WO2015167307A1 - Magnetic resonance imaging device and method for generating magnetic resonance image

Info

Publication number: WO2015167307A1
Application number: PCT/KR2015/004446
Authority: WO
Inventors: 세르기비치 미구킨아르헴; 알렉산드로비치 코로브첸코드미트리; 유리에비치 시로텐코미하일; 알투로비치 가브릴유크키릴; 굴라카프라빈; 최상천; 니콜라예비치 리차고브미하일; 최양림
Original assignee: 삼성전자 주식회사
Priority date: 2014-04-30
Filing date: 2015-04-30
Publication date: 2015-11-05

Abstract

A magnetic resonance imaging device comprises: an acquisition unit for acquiring an undersampled spectrum on a k-space; and a recovery unit for generating a target image through the undersampled spectrum, wherein the recovery unit can comprise: a first sub recovery unit for performing an initial recovery for data corresponding to a non-sampled position on the k-space through a slit Bregman technique or an approximation sparse encoding technique; a second sub recovery unit which decomposes, in multiple bands, the initially recovered spectrum on the k-space into a plurality of individual spectra, and which carries out a dictionary learning recovery by alternating a sparse approximation and recovery of frequencies measured for images corresponding to the decomposed bands; and an image generation unit for merging the recovered images of the respective frequency bands so as to generate the target image.

Description

Magnetic resonance imaging apparatus and method of generating magnetic resonance imaging

The present invention relates to medical imaging. More specifically, the present invention relates to a method and system for recovering a subject image from undersampled data.

The present invention can be applied to an MRI scanner that accelerates image acquisition, and can be implemented as software or hardware component of a medical device.

Magnetic resonance imaging (MRI) is a non-invasive and non-ionized imaging technique that can visualize both anatomical structures and physiological functions and is widely used for diagnosis. However, the MRI imaging technique is a relatively slow imaging modality that sequentially acquires the data used, the so-called k-space samples of the spatial Fourier transform of the object in time.

That is, the target patient in the magnetic resonance imaging apparatus may feel psychological anxiety in an enclosed space, and acquire an image uneconomically. Therefore, there is a need for a technique for acquiring images at high speed in a magnetic resonance imaging apparatus.

A magnetic resonance imaging apparatus according to an embodiment of the present invention aims to acquire a magnetic resonance image in a shortened time.

Magnetic resonance imaging apparatus according to an embodiment of the present invention includes an acquirer for obtaining an undersampled spectrum on the k-space; And a recovery unit generating a target image through the undersampled spectrum, wherein the recovery unit is configured to perform segmented Braggman or approximate sparse coding on data corresponding to the non-sampled position on the k-space. A first sub recovery unit performing initial recovery; Multiband decomposition the initial recovered spectrum on the k-space into a plurality of individual spectra, and sparse approximating and restoring the measured frequency for the images corresponding to the resolved band in advance A second sub recovery unit performing dictionary learing recovery; And an image generator for merging the restored images of the respective frequency bands to generate a target image.

For example, the first sub-recovery unit collects a data set of full sampled k-spaces, applies a sampling mask, obtains undersampled data according to an index of the sampling mask, and Train the model to predict the sparse code for the undersampled data using the full sampled data for error calculation, and approximate from the undersampled data using the trained model The approximated sparse coding may be performed by predicting the sparse code.

For example, the sparse approximation extracts overlapping patches of non-uniform size from an input image corresponding to an input magnetic resonance signal, performs a sparse decomposition of the patches into sparse codes using a dictionary, and Patches may be restored by sparse code and dictionary, and the image may be restored by the recovered patch.

For example, the dictionary may be learned in advance based on patches extracted from a set of full sampled training magnetic resonance images.

For example, the dictionary can be learned individually for each frequency band.

For example, the sparse decomposition can be performed by minimizing non-zero components in a patch representation as a linear combination of dictionary elements.

For example, the aspect ratio of the patches of non-uniform size may be determined according to the amount of anisotropy of the sampling scheme.

For example, the patches may have a rectangular shape for orthogonal one-dimensional sampling and other anisotropic schemes, and have the longest dimension in the direction of most information loss in the undersampled k-space.

For example, the patches may have a square shape of an isotropic sampling method.

For example, the pre-based recovery may be performed for each k-space corresponding to a different frequency band.

For example, the decomposed k-spaces corresponding to the frequency bands consist of frequencies taken from the initially recovered k-space corresponding to the current band, and the other frequencies are filled with zero and marked as measured. Can be.

A method of generating a magnetic resonance image from undersampled data by a magnetic resonance apparatus according to another embodiment of the present invention, the method comprising: obtaining, by a receiver coil, an undersampled spectrum in k-space; Initial retrieval by initial guessing the unsampled position in the k-space; Multiband decomposition the initial recovered spectrum into a plurality of individual spectra; Performing dictionary-based recovery on the images corresponding to each of the decomposed bands; And merging the recovered images through the dictionary-based recovery to generate a target image: a. The initial restoring comprises optimizing a cost function including a simplified LO norm; b. Said multiband decomposition step applies to the spectrum initially recovered in k-space to obtain the number of individual spectra for various frequency bands; c. The pre-based recovery step includes alternating sparse approximation and recovery of the measured frequency; d. The generating of the target image may generate the target image as a sum of the restoration results of the respective frequency bands.

For example, the normalization of the L0 norm can be simplified through optimization of the L1 norm.

For example, the optimization of the l1 norm may be implemented through a split Bregman technique or an approximate sparse coding method.

According to another aspect of the present invention, there is provided a magnetic resonance imaging system comprising: an acquirer configured to acquire an undersampled spectrum in k-space by a receiver coil; And receive and store the undersampled k-space transmitted from the acquisition unit, perform initial recovery by initially charging a non-sampled location in the k-space, and store the plurality of initial recovered spectra. A recovery server for multiband decomposition into individual spectra, performing pre-based recovery of an image corresponding to the resolved band, merging the recovered images, and generating a destination image, a. The initial recovery is performed by optimizing a cost function including a simplified LO norm; b. The multiband decomposition is applied to the spectrum initially recovered in k-space to obtain the number of individual spectra for various frequency bands, c. The pre-based recovery is performed by alternating sparse approximation and recovery of the measured frequency, d. The target image may be generated by adding the recovery results of the respective frequency bands.

For example, the magnetic resonance system may further include an operator console configured to receive a merge result from the recovery server and display the merge result through a display.

For example, the operator console receives an input from an operator to perform an image recovery, generates an image recovery command, sends the image recovery command to the recovery server, and the recovery server receives a command from the operator console. To restore the image.

The present invention can quickly recover MRI from undersampled Fourier spectral data. In addition, the present invention performs sparse normalization by a precomputed dictionary with patches of non-uniform size, thereby accelerating the calculation algorithm and improving recovery accuracy.

1 is a diagram illustrating a magnetic resonance apparatus 100 according to an embodiment of the present invention.

2 is a flowchart illustrating a method for fast recovery of an MR image according to an embodiment of the present invention.

3 is a flowchart specifically illustrating split Bregman based initialization according to an embodiment of the present invention.

4 is a flowchart illustrating an approximate sparse coding method according to an embodiment of the present invention.

5 is a flowchart illustrating a pre-based recovery method according to an embodiment of the present invention.

6 illustrates a system in which an exemplary embodiment of the present invention may be implemented.

7 is a diagram illustrating a magnetic resonance imaging system according to an embodiment of the present invention.

8 is a diagram illustrating a magnetic resonance imaging system which communicates through a communication unit according to an embodiment of the present invention.

A first aspect of the invention, the acquisition unit for obtaining the undersampled spectrum on the k-space; And a recovery unit generating a target image through the undersampled spectrum, wherein the recovery unit is configured to perform segmented Braggman or approximate sparse coding on data corresponding to the non-sampled position on the k-space. A first sub recovery unit performing initial recovery; Alternately multi-band decomposition the initial recovered spectrum on the k-space into a plurality of individual spectra, sparse approximating and restoring the measured frequency for the images corresponding to the resolved band A second sub recovery unit performing dictionary learing recovery; And an image generator for merging the restored images of the respective frequency bands and generating a target image.

In addition, the first sub-recovery unit collects a data set of full-sampled k-spaces, applies a sampling mask, prepares undersampled data according to an index of the sampling mask, and an error. Train the model to predict the sparse code for the undersampled data using the full sampled data for calculation, and approximate the sparse from the undersampled data using the trained model. The sparse coding may be performed by predicting a sparse code.

In addition, the sparse approximation extracts overlapping patches of non-uniform size from an input image corresponding to an input magnetic resonance signal, performs sparse decomposition of the patches into sparse codes using a dictionary, and the sparse Repairing patches by code and dictionary, and restoring an image by the restored patch.

In addition, the dictionary may be learned in advance based on patches extracted from a set of full sampled training magnetic resonance images.

In addition, the dictionary can be learned individually for each frequency band.

The sparse decomposition can also be performed by minimizing non-zero components in a patch representation as a linear combination of dictionary elements.

In addition, the aspect ratio of the patches of non-uniform size may be determined according to the amount of anisotropy of the sampling scheme.

In addition, the patches have a rectangular shape for orthogonal one-dimensional sampling and other anisotropic schemes and may have the longest dimension in most information loss directions in the undersampled k-space.

In addition, the patches may have a square shape of an isotropic sampling method.

In addition, the pre-based recovery may be performed for each k-space corresponding to a different frequency band.

Further, the decomposed k-spaces corresponding to the major band may consist of frequencies taken from the initially restored k-space corresponding to the current band, and the other frequencies may be marked as zero filled and measured. .

A second aspect of the invention includes the steps of: obtaining a spectrum in which a receiver coil is undersampled in k-space; Initial retrieval by initial guessing the unsampled position in the k-space; Multiband decomposition the initial recovered spectrum into a plurality of individual spectra; Performing dictionary-based recovery on the images corresponding to each of the decomposed bands; And merging the recovered images through the dictionary-based recovery to generate a target image: a. The initial restoring comprises optimizing a cost function including a simplified LO norm; b. Said multiband decomposition step applies to the spectrum initially recovered in k-space to obtain the number of individual spectra for various frequency bands; c. The pre-based recovery step includes alternating sparse approximation and recovery of the measured frequency; d. The generating of the target image may provide a method for generating a magnetic resonance image, wherein the target image is generated as a sum of the restoration results of the respective frequency bands.

In addition, the normalization of the L0 norm can be simplified through optimization of the L1 norm.

In addition, the optimization of the l1 norm may be implemented through a split Bregman technique or an approximate sparse coding method.

In addition, the split Bregman technique can recover measured k-space samples of the spectrum recovered during the inner Bregman iteration.

In addition, when the procedure of the split Bregman technique is completed, the method may further include recovering k-space samples measured in the recovered spectrum.

In addition, the approximate sparse coding method includes: a. Collecting a data set of full sampled k-spaces; b. Applying a sampling mask and preparing undersampled data according to an index of the sampling mask; c. Training the model to predict the sparse code for the undersampled data using the full sampled data for error calculation; d. Predicting the approximated sparse code from the undersampled data using the trained model.

The sparse decomposition can also be performed by minimizing a number of non-zero components in a patch representation as a linear combination of dictionary elements.

Further, the decomposed k-spaces corresponding to the frequency band may consist of frequencies taken from the initially restored k-space corresponding to the current band, and the other frequencies may be filled with zero and marked as measured. .

The method may further include displaying a target image on a display.

According to a third aspect of the present invention, there is provided an apparatus, comprising: an acquirer for obtaining an undersampled spectrum in k-space by a receiver coil; And receive and store the undersampled k-space transmitted from the acquisition unit, perform initial recovery by initially charging a non-sampled location in the k-space, and store the plurality of initial recovered spectra. A recovery server for multiband decomposition into individual spectra, performing pre-based recovery of an image corresponding to the resolved band, merging the recovered images, and generating a destination image, a. The initial recovery is performed by optimizing a cost function including a simplified LO norm; b. The multiband decomposition is applied to the spectrum initially recovered in k-space to obtain the number of individual spectra for various frequency bands, c. The pre-based recovery is performed by alternating sparse approximation and recovery of the measured frequency, d. A magnetic resonance imaging system may be provided by adding the recovery results of the respective frequency bands to generate a target image.

The magnetic resonance system may further include an operator console configured to receive a merge result from the recovery server and to display the merge result through a display.

In addition, the operator console receives an input from an operator to execute an image recovery, generates an image recovery command, transmits the image recovery command to the recovery server, and the recovery server receives the command from the operator console You can perform video recovery.

Advantages and features of the present invention, and methods of achieving them will be apparent with reference to the embodiments described below in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but can be implemented in various different forms, and only the embodiments make the disclosure of the present invention complete, and the general knowledge in the art to which the present invention belongs. It is provided to fully inform the person having the scope of the invention, which is defined only by the scope of the claims.

Terms used herein will be briefly described and the present invention will be described in detail.

The terms used in the present invention have been selected as widely used general terms as possible in consideration of the functions in the present invention, but this may vary according to the intention or precedent of the person skilled in the art, the emergence of new technologies and the like. In addition, in certain cases, there is also a term arbitrarily selected by the applicant, in which case the meaning will be described in detail in the description of the invention. Therefore, the terms used in the present invention should be defined based on the meanings of the terms and the contents throughout the present invention, rather than the names of the simple terms.

When any part of the specification is to "include" any component, this means that it may further include other components, except to exclude other components unless otherwise stated. In addition, the term "part" as used herein refers to a hardware component, such as software, FPGA or ASIC, and "part" plays certain roles. However, "part" is not meant to be limited to software or hardware. The “unit” may be configured to be in an addressable storage medium and may be configured to play one or more processors. Thus, as an example, a "part" refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, procedures, Subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays and variables. The functionality provided within the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts".

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention.

As used herein, “image” may mean multi-dimensional data composed of discrete image elements (eg, pixels in a 2D image and voxels in a 3D image). For example, the image may include a medical image of an object obtained by an X-ray apparatus, a CT apparatus, an MRI apparatus, an ultrasound diagnostic apparatus, and another medical imaging apparatus.

Also, as used herein, an "object" may include a person or an animal, or a part of a person or an animal. For example, the subject may include organs such as the liver, heart, uterus, brain, breast, abdomen, or blood vessels. Also, the "object" may include a phantom. Phantom means a material having a volume very close to the density and effective atomic number of an organism, and may include a sphere phantom having properties similar to the body.

In addition, in the present specification, the "user" may be a doctor, a nurse, a clinical pathologist, a medical imaging expert, or the like, and may be a technician who repairs a medical device, but is not limited thereto.

In addition, in the present specification, "MR image: Magnetic Resonance image" means an image of an object acquired using the principle of nuclear magnetic resonance.

In addition, in the present specification, the "pulse sequence" means a continuation of a signal repeatedly applied in a magnetic resonance imaging system. The pulse sequence may vary according to a time parameter of the RF pulse, for example, a repetition time (TR) and an echo time (Time to Echo, TE).

In addition, in this specification, the "pulse sequence schematic diagram" describes the order of signal generation occurring in the magnetic resonance imaging system. For example, the pulse sequence schematic diagram may be a schematic diagram showing an order in which RF pulses, gradient magnetic fields, magnetic resonance signals, etc. are applied over time.

A magnetic resonance imaging system is an apparatus for obtaining an image of a tomography region of an object by expressing intensity of a magnetic resonance signal (MR) with respect to a radio frequency (RF) signal generated in a magnetic field of a specific intensity in contrast. For example, if a subject is placed in a strong magnetic field and then irradiates the subject with an RF signal that resonates only a particular nucleus (e.g., hydrogen nucleus, etc.), the magnetic resonance signal is emitted from the particular nucleus. The imaging system may receive this magnetic resonance signal to obtain an MR image. The magnetic resonance signal refers to an RF signal emitted from the object. The magnitude of the magnetic resonance signal may be determined by the concentration of predetermined atoms (eg, hydrogen, etc.) included in the object, a relaxation time T1, a relaxation time T2, and a flow of blood flow.

In this specification, k-space may mean a spatial frequency matrix having a mathematical relationship with an image. In the case of digitized MRI signal data, each point in k-space represents a spatial frequency, and signal strength may be added and used. The under-sampled k-spatial data may refer to k-spatial data that does not satisfy the Nyquist criterion. Full-sampled k-spatial data may refer to k-spatial data that satisfies the Nyquist criterion.

In general, a magnetic resonance imaging apparatus may apply a signal having a certain ratio of frequency and energy to an atomic nucleus inside a patient in a space where a magnetic field is applied, and examine an internal structure of the human body using energy induced by a resonance reaction. .

Landmark compression sensing (CS) technology, which was initiated in US patent application (US 2006/0029279, Donoho), suggests that sparse imaging techniques can recover magnetic resonance (MR) images from small sampling data. Proved. That is, for the recovery of magnetic resonance (MR) images, the k-space does not need to be sampled completely according to the phase encoding phase of the pulse sequence. Thus, compressed sensing (CS) is the keynote for hardware-based acceleration of MRI.

Various algorithms can be used for compressed sensing (CS) recovery. For example, the compressed sensing (CS) recovery algorithm includes Total Variaton [B. Liu et al., Compression Sensing MRI with Normalized Sensitivity Encoding (SENSE) Recovery, Magnetic Resonance in Medicine 61, 145-152 (2009)] or L1-normalized Optimization, Recomputing Using Bregman Iteration [S. Ravishankar and Y. Bresler, Image Recovery from Highly Undersampled k-Spatial Data by Prior Learning, IEEE Trans. On Med. Imag. 30, 1028-1041 (2011)] or a precomputed dictionary [A. Migukin et al., 4f optical phase search: background correction and sparse normalization of objects with binary amplitude, Appl. Opt. 52, A269-A280 (2013).

As mentioned above, US patent application US 2006/0029279 is a breakthrough in compression sensing. U.S. Patent Application (US 2006/0029279) describes a technique that can recover signals / images with acceptable quality using fewer measurements than the unknown. The main idea is to find a sparse representation of the object of interest on an overcomplete basis. However, the US patent application (US 2006/0029279) needs to specify a sparsifying transform, due to the too general variational formulation of the optimization problem.

US 2007/0110290 describes a technique for recovering medical images from undersampled data measured in k-space by an iterative minimization algorithm. In this patent application, recovery is forced optimization of formulated and / or preformulated functions, including sparse representation norms of test images derived from datasets and / or data fidelity terms. (constrained optimization). It is important to note that the minimization algorithm can include conjugate gradient sub-algorithm and / or Bregman Iteration.

US patent application (US 2012/0177128) describes an adaptation procedure for designing a dictionary for sparse representation of a signal / image by the K-SVD algorithm. Based on one or more training signals, the described signal processing system provides one dictionary comprising reference signals-atoms-so that each training signal is represented as a sparse linear combination of signal-atoms of the dictionary. To be possible. The use of overcomplete and sparse representations has been demonstrated through several efficient methods, for example, efficient methods include Maxium A Posteriori (MAP) estimation, Maximum Likelihood (ML) method, independent. Matching Pursuit (MP), Orthogonal Matching Pursuit (OMP), Fundamental Pursuit with Independent Component Analysis (ICA) Algorithm, and Method of Optimal Direction (MOD) Basis Pursuit (BP), and Focal Under-determined System Solver (FOCUSS) algorithms.

US patent application US 2012/0259590 describes a CS technique for recovering one signal / image from a plurality of measurement vectors when a signal jointly corresponds to a plurality of sparse vectors. This patent application jointly extracts information from a plurality of measurement vectors of sparse signals and calculates a subset having at least one element of a joint support based on the plurality of measurement vectors. to provide. Sparse recovering is formed by computing a succinct representation or sparse approximation of a given vector by a linear combination of prime numbers from a collection of known vectors.

Disadvantages of the aforementioned method can be mentioned in various aspects. Some techniques are biased towards very narrow specific tasks: prior information, target attributes and / or types of data used are known to the target object. An important point of each CS algorithm is initialization: almost all authors publish results for the default zero-filling case. In general, the dictionary is calculated repeatedly with patches of uniform size.

[Magnetic Resonance Device for MRI Recovery from Undersampled Data]

The magnetic resonance apparatus 100 may include an acquisition unit 110 and a recovery unit 120. The recovery unit 120 may include a first sub recovery unit 130, a second sub recovery unit 140, and an image generator 150.

The acquirer 110 may acquire an undersampled spectrum on k-space. The acquirer 110 may acquire the undersampled data on the k-space through the magnetic resonance signal received by the receiving coil. In detail, the acquirer 110 obtains a magnetic resonance signal from an object by applying an RF signal of a specific frequency to the object, but does not obtain a magnetic resonance signal for some spatial frequencies, and obtains a magnetic resonance signal for some spatial frequencies. Can be obtained.

Magnetic resonance imaging is a time consuming procedure and is therefore important for speeding up the acquisition of magnetic resonance imaging. The magnetic resonance signal is sampled in spectral space representing the two-dimensional Fourier transform of the magnetic resonance image. In the case of acquiring samples sequentially, the total acquisition time depends on the number of samples, so one of the ways of facilitating the acquisition of magnetic resonance images is to narrow the sampled frequency and recover the magnetic resonance images only by the observed frequency. will be.

The present invention describes a method for rapid recovery of magnetic resonance images in an undersampled k-space. The method of generating a magnetic resonance image (or magnetic resonance image recovery method) requires one sampling mask, which is one undersampled k-space and a binary matrix, where "1" of the sampling mask is at that frequency (at that location). Means that the point in k-space is sampled, and "0" means that the frequency (point in k-space at that location) is not sampled. Through this method, the recovered image is generated and output.

The acquirer 110 according to an exemplary embodiment may shorten the data acquisition time by acquiring an undersampled spectrum on k-space corresponding to some spatial frequencies.

The recovery unit 120 may generate the target image through the undersampled spectrum.

In detail, the recovery unit 120 according to an embodiment of the present invention may include a first sub recovery unit 130, a second sub recovery unit 140, and an image generation unit 150.

The first sub recovery unit 130 may perform the initial recovery described in step 102 of FIG. 2. The first sub-recovery unit 130 may perform initial recovery through the split-bragman technique described with reference to FIG. 3 or the approximate sparse encoding technique described with reference to FIG. 4.

The first sub-recovery unit 130 may perform initial recovery on the data corresponding to the non-sampled position on the k-space through a split Bregman technique or an approximate sparse encoding technique. A detailed description of the split Braggman technique or the approximate sparse encoding technique will be given later.

The second sub recovery unit 140 may perform multi-band decomposition described in step 103 of FIG. 2. In addition, the second sub recovery unit 140 may perform dictionary learning reconstruction described with reference to step 104 of FIG. 2 and FIG. 5.

The second sub recovery unit 140 may multi-band decompose the initial recovered spectrum on k-space into a plurality of individual spectra.

In addition, the second sub recovery unit 140 may perform dictionary learing recovery on the images corresponding to the decomposed band. For example, the second sub recovery unit 140 alternately performs sparse approximation and recovery of the measured frequency to perform dictionary learing recovery.

The image generator 150 may generate the target image by merging the restored images of the respective frequency bands. The image generator 150 may receive a restored image of each frequency band decomposed from the second sub-recovery unit 140 and merge the recovered images to generate a target image. The image generator 150 may transmit the target image to the display unit.

The magnetic resonance apparatus according to an embodiment of the present invention combines the characteristics of the high performance and fast convergence of the split-bragman iteration algorithm with the advantages of and dictionary learning. Hereinafter, specific operations will be described.

In step 101, the magnetic resonance apparatus may obtain the undersampled K-spatial data. For example, the magnetic resonance apparatus may obtain the spectral data in k-space by the receiver coil in accordance with a sampling scheme defined as a sampling mask to form the undersampling spectrum to be processed.

In step 102, the magnetic resonance scanner may perform initial recovery. For example, the magnetic resonance scanner may perform initial reconstruction by filling an empty guess (ie, an unsampled position) of k-space with an initial guess.

In step 103, the magnetic resonance apparatus may perform decomposition into a plurality of frequency bands. The magnetic resonance apparatus may decompose the k-space initial spectrum (the first recovered spectrum) into frequency bands to generate a plurality of individual spectra corresponding to each frequency band.

In step 104, the magnetic resonance scanner may perform pre-based recovery for each frequency band. For example, the magnetic resonance scanner may recover each of the obtained spectra using a pre-based compression sensing algorithm.

In step 105, the magnetic resonance scanner may merge recovered images corresponding to different frequency bands. For example, a magnetic resonance apparatus merges all recovered images corresponding to different frequency bands using a sum operator, and the recovered final result, that is, one blurring and denoised ) A full spectrum of k-space may be generated in response to the MR signal.

In the described procedure, a learning-based approach is used to generate the dictionary used for image reconstruction. The dictionary described above is pre-learned based on a set of fully sampled training MR images.

Hereinafter, each step will be described in more detail.

[Initial reconstruction]

Naive object recovery refers to the inverse Fourier transform of the undersampled spectral data without preprocessing. However, this recovery is strongly impaired by aliasing effects, especially due to zeros in the measurement in the case of general under sampling. A simple way to avoid aliasing and help the recovery algorithm for fast convergence is to fill those empty samples in the Fourier domain by some non-zero value. This procedure is called initialization or initial recovery.

There are many ways to take initial data: constant values, changing gradients, or any complex value. The blank samples are taken from the Fourier spectrum of the aliased object preprocessed in the image region anyway, for example as a normalized deconvolution, blind deconvolution using split Braggman iteration algorithm or approximate sparse coding. It can be achieved by deblurring, noise reduction, smoothing, and sharpening, which are recovered by some complex methods. It significantly suppresses the intrinsic aliasing effect and increases both convergence speed and recovery quality. In addition, some conventional information about the object, or object model, or data acquisition model may be used.

In one of the embodiments of the present invention, a split Braggman algorithm [T. Goldstein and S. Osher, l1-Segmented Bregman Method for Normalized Problems, Imag. Sciences 2, SIAM J. 323-343 (2009), can be used in the initialization step 102. This algorithm provides an efficient and fast convergence speed for the method according to the invention.

[D. L. Donoho, Compression Sensing, IEEE Trans. Inf. Theory 52, 1289-1306 (2006)], the CS recovery problem is first simplified and formulated into the following basic seeking noise cancellation problem:

Equation 1 is related to a general CS and is a result derived from a convex cost function that is minimized. Where x is a vector of the object image to be recovered, F _u is a partially sampled Fourier transform, y is a vector of measured undersampled spectra, Ψ is a sparsifying transform, and λ is a normalization parameter variable and, ℓ1 gambling (norm) ∥..∥ ₁ is defined as the sum of the absolute values of the items (items) of the vector, ℓ2 gambling (norm) ∥..∥ ₂ is defined as the square root of the sum of squares of the vector entries do.

[Y. The l1 and l2 norms according to Wang et al., Fast Algorithm for Total Variation Normalization and Image Deblurring, CAAM Technical Report (2007)] can be separated as follows.

Where d is a sparse object approximation. D is an approximation of x. Thus it follows the split Braggman iteration algorithm:

Equation 3.1

Equation 3.2

Equation 3.3

Where μ is a normalization parameter and b ^k is an update of the Bregman parameter.

A flowchart of the split Braggman initialization algorithm is shown in FIG. 3. [T. Goldstein and S. Osher, l1-Segmented Bregman Method for Normalized Problems, Imag. As described in Sciences 2, SIAM J. 323-343 (2009)], this algorithm includes two nested loops.

Referring to FIG. 3, in step 201, the magnetic resonance scanner may initialize auxiliary variables. For example, the magnetic resonance apparatus initializes auxiliary variables of the sparse object approximation d ⁰ and the Bregman parameter b ⁰ with a zero vector.

In step 202, the magnetic resonance scanner may determine whether the outer loop is executed. That is, the magnetic resonance scanner may determine whether the outer loop is completed. If the outer loop has not been performed, the magnetic resonance apparatus may proceed to step 203 and if the outer loop has been performed, the magnetic resonance apparatus may proceed to step 209.

In step 203, the magnetic resonance apparatus may determine whether the inner loop is executed. In other words, the magnetic resonance scanner may determine whether the inner loop is complete. If the inner loop has not been performed, the magnetic resonance apparatus proceeds to the inner loop algorithm, and if so, the magnetic resonance apparatus can proceed to step 208.

In step 204, the magnetic resonance scanner can update the solution image by minimizing the l2 norm. In step 205, the magnetic resonance scanner may recover the measured k-space sample in the recovered spectrum. In step 206, the magnetic resonance scanner may reduce the l1 norm (TV) of the solution. In step 207, the magnetic resonance scanner may update the Bragman parameter.

For example, during the inner loop process, the above-described split Braggman routines (Equation 3.1 to Equation 3.3) may be performed, and the inner loop may include:

recalculating the target object estimate x ^{k + 1} by optimizing the norm norm; (Step 204)

recalculating the sparse object approximation d ^{k + 1} by solving the l1-normalization problem; (Step 205) and

Updating the Begman parameter b ^{k + 1} . (Step 207).

The l1 norm optimization can be realized via a soft threshold with a 1 / 2μ equal threshold (step 206).

In order to increase the recovery accuracy and reduce the number of iterations, the recovered object can be further updated by recovering the k-space sample measured in the recovered Fourier spectrum of x ^{k + 1} (step 205).

In step 208, the magnetic resonance scanner may update the auxiliary inner loop independent variable used to recomput the object estimate (step 204).

After performing the step 208, the algorithm may return to step 202.

As discussed above, once the outer loop (step 202) is complete, the algorithm proceeds to step 209 where the recovered object can be updated again by recovering the k-space samples measured in the recovered Fourier spectrum. Thereafter, a split Braggman based initialization algorithm is completed.

The derivative operator can be used as sparsifying transform Ψ. In such cases, the overall variation is usually normalized.

[Approximate Sparse Coding for Initial Guess Calculation]

According to one embodiment of the invention, the initial guess may be calculated by an approximate sparse encoding method in the initialization step 102. The main idea of this method is to use a trainable model for prediction of sparse codes with fewer calculations than real algorithms. As one of the variations of the learnable model for sparse code prediction, a Learned Coordinate Descent algorithm (LCoD) can be used. For example, the Learned Coordinate Descent algorithm (LCoD) is the same as that disclosed in Karol Gregor and Yann LeCun, Fast Approximation of Sparse Coding, ICML 2010.

4 is a flow diagram illustrating an algorithm of model training for sparse code prediction.

To train the model sparse coding dictionary, we need to pre-define the training set and sampling mask. The training set should contain fully sampled data.

In step 301, the magnetic resonance scanner can secure a full sampled k-space from the training dataset. For example, at step 301, the full sampled k-space may be patched from the data set.

In operation 302, the magnetic resonance apparatus may apply a sampling mask to obtain undersampled data that matches an index of the sampling mask, and fill missing values with zero. For example, in step 302, a predefined sampling mask is applied to the k-space to obtain an undersampled k-space, so that missing values are filled with zeros.

In step 303, the magnetic resonance apparatus may recover the two images by applying an inverse Fourier transform to the undersampled k-space and the k-space obtained in step 301. For example, in step 303, two images are recovered by applying an inverse Fourier transform: one in the undersampled k-space (with artifacts with zero filling), the other in full sampling Image in k-space.

In step 304, the magnetic resonance scanner may extract a patch with an inverse Fourier transform from the two recovered images. The magnetic resonance apparatus can recover two images by applying an inverse Fourier transform to the undersampled k-space and the full sampled k-space. For example, at step 304, a small image patch can be extracted from both of these images. The size of the extracted patch may be the same as the size of the dictionary element.

Steps 305 to 308 the loop executes for all patches.

In step 305, the magnetic resonance apparatus may perform sparse code prediction on each patch extracted from the undersampled image using the releasable model. That is, at step 305, the sparse code can be predicted by a trainable model.

In step 306, the magnetic resonance scanner may perform recovery of the image patch from the predicted code using a predefined dictionary. For example, in step 306, the image patch can be recovered using this predicted code and a predefined dictionary.

In operation 307, the magnetic resonance apparatus may calculate an error between the recovered patch and the patch extracted from the sampling data image acquired in operation 301. For example, in step 307 this recovered image patch is used to calculate the error compared to the image patch extracted from the fully sampled image.

In step 308, the magnetic resonance scanner may update the trainable model parameters using the calculated error. For example, at step 308, using the calculated error, the model parameters are updated with some optimization method, such as, for example, a gradient descent method. This process is repeated until the training process converges.

[Pre-Based Recovery]

In step 104 of FIG. 2, based on the sparse approximation of one input signal using the learned dictionary. In step 104 of FIG. 2, the optimization equation of Equation 4 is used:

Where l0 norm is ₀ , where ₀ is defined as a number of nonzero elements of a vector.

In a dictionary-based approach, sparsifying transforms are represented by sparse decomposition of inputs from dictionaries to elementary elements. The solution to the optimization equation of Equation 4 is performed by the iterative algorithm of steps 401 to 408 shown in FIG.

The magnetic resonance apparatus obtains an image 403 by applying an inverse Fourier transform (step 402) to the input undersampled k-space (step 401). Then, until the solution converges (step 404), the next step is performed. The duplicate patch set 407 is extracted from the image (step 406). Sparse decomposition (step 408) is then performed for each of the patches. In step 408, the optimization equation of Equation 5 is used:

Where the columns of X are vectorized patches, the columns of z are sparse code vectors, D is a pre-transformation matrix, and δ is a normalization parameter.

In this task, the elements of dictionary 409 are represented as columns of matrix D, and each column of X aims to be represented as a linear combination of only a small number of non-zero components and dictionary elements (thus, Heat is sparse). This problem can be solved by a greedy orthogonal matching tracking algorithm.

Sparse code z 410 is obtained for each patch, and the recovered patches 412 are computed as a linear combination of sparse code and dictionary elements computed with coefficients 411. The intermediate recovered image 414 is then obtained by merging the recovered patch with the overlap (step 413). In the next step, the intermediate recovered image 414 is transformed back into frequency space by applying a Fourier transform (step 415), and the intermediate k-spatial data 416 is obtained.

The recovered frequencies are then recovered (step 417) in the intermediate k-space (step 416) by the measured data. The k-space with the recovered observation frequency (step 418) is transformed into image space by an inverse Fourier transform (step 402). Once the solution has converged (step 404), the image (step 403) is output as a recovery result (step 405), otherwise a new iteration begins.

[Non-uniform size patch]

The method according to the invention operates with various sampling techniques. The procedure of extracting a patch from an input image creates a set of square pieces of the image. The optimal aspect ratio (relative ratio of width to height) of the patch depends on the k-space sampling scheme. A common method of sampling in an MRI apparatus is to acquire a phase encoding line. These lines represent rows in k-space.

Thus, information about the frequency is lost along the y direction (vertical coordinates). This loss affects the image (obtained in the undersampled k-space) within the vertical aliasing artifact. Since the fixed number of dictionary elements is difficult to encode such an amount of information with only a small number of elements, the size of the patch should not be very large in both dimensions.

The use of a fixed amount of information (fixed few pixels) in a patch may be more optimal to cover more data in the vertical direction using one patch. This is applicable because aliasing artifacts are distributed in a vertical direction with a Cartesian sampling scheme while skipping a phase coded line. A covering strategy, such as a fixed amount of information per patch, can be performed by applying rectangular patches of non-uniform size.

For example, a rectangular patch of non-uniform size can be applied by increasing the patch height while reducing the patch width. Non-square rectangular patches can achieve better recovery quality than square patches, in which case the number of pixels of the rectangular patch is the same or less than the square patches.

For non-orthogonal isotropic random sampling, square patches are recommended. The recovery quality will be high. In the general case, the aspect ratio of the patch depends on the proportion of missing information along the x direction and the y direction (amount of anistropy).

[Multiband Decomposition]

In the method according to the invention, one of the methods for utilizing multiple structures of data is to recover, at step 103, a plurality of image components corresponding to multiple frequency bands. In this method, the initially recovered spectrum in step 102 is divided, corresponding to a group of several subbands.

For example, the initially recovered spectrum in step 102 is divided corresponding to a group of low, middle and high frequency bands.

In such division, several new (dissolved) k-spaces emerge, each containing only frequencies from the corresponding band, and the other frequencies are filled with zeros and displayed as measurements. Decomposing the MR signal into frequency bands yields various contractions of the in-band signal (within each band, the signal has a simple form), which in turn increases the signal's sparsity and quality of final recovery.

This truncated k-space is treated as input to the main loop of pre-based recovery in step 104 and processed in parallel. Once each band image is recovered, the final recovered image is obtained at step 105 by summing the recovered images of all bands. In the dictionary learning phase, each training image is decomposed into frequency bands in the same way, and a unique dictionary is learned for each of these bands separately.

In the recovery phase, each band is recovered using the corresponding dictionary. In the case of measuring all frequencies of the current band, recovery of the band is performed simply by applying an inverse Fourier transform. The choice of band shape depends on the sampling method.

For example, for orthogonal one-dimensional sampling, the band is rectangular stripes in one k-space aligned with the sampling lines. For radial-symmetric or isotropic random sampling, the band is represented by a concentric ring band around zero frequency.

[System for MRI Recovery from Undersampled Data]

6 is a view for explaining a magnetic resonance imaging system according to an embodiment of the present invention.

Referring to FIG. 6, the MRI system 500 may include an acquisition unit 501, a recovery server 502, and an operator console 503.

The acquirer 501 may acquire the undersampled k-space. For example, the acquirer 501 may acquire spectral data of k-space according to a sampling scheme defined as a sampling mask. The acquirer 501 may be obtained through a receiver coil. The acquirer 501 may be implemented by separate medical equipment that is distinct from the recovery server 502 or the operator console 503.

The recovery server 502 can include a processor 504 and a memory 505. In addition, the processor 504 may include a recovery unit 508.

The recovery unit 508 according to an embodiment of the present invention may recover the undersampled MRI data. The recovery unit 508 may perform initial reconstruction by filling an empty position of the k-space with an initial guess. The recovery unit 508 may decompose the k-space initialization spectrum into frequency bands to generate a plurality of individual spectrums corresponding to each frequency band. The recovery unit 508 may recover each of the plurality of individual spectra using a pre-based compression sensing algorithm.

The recovery unit 508 merges all recovered images corresponding to different frequency bands using a sum operator, and k in response to the restored final result, that is, one smear-free and noise-reduced MR signal. Generate full spectrum of space

The processor 504 may perform various operations for implementing the present invention in addition to the recovery unit 508.

The memory 505 may store a file including k-space data obtained by obtaining the under-sampled k-space through the acquirer 501.

The operator console 503 can include a display 506 and a controller 507.

The controller 507 may execute a program for performing image recovery. The result of the program can be output to the display 506.

Therefore, the magnetic resonance imaging system 500 according to the exemplary embodiment of the present invention may acquire k-spatial data in a shortened time.

7 is a diagram illustrating a magnetic resonance imaging system according to an embodiment of the present invention. Referring to FIG. 7, the magnetic resonance imaging system may include a gantry 20, a signal transceiver 30, a monitor 40, a system controller 50, and an operating unit 60.

The gantry 20 blocks electromagnetic radiation generated by the main magnet 22, the gradient coil 24, the RF coil 26, and the like from radiating to the outside. A static magnetic field and a gradient magnetic field are formed in the bore in the gantry 20, and an RF signal is irradiated toward the object 10.

The main magnet 22, the gradient coil 24 and the RF coil 26 may be disposed along a predetermined direction of the gantry 20. The predetermined direction may include a coaxial cylindrical direction or the like. The object 10 may be positioned on a table 28 that can be inserted into the cylinder along the horizontal axis of the cylinder.

The main magnet 22 generates a static magnetic field or a static magnetic field for aligning the directions of the magnetic dipole moments of the nuclei contained in the object 10 in a constant direction. As the magnetic field generated by the main magnet is stronger and more uniform, a relatively precise and accurate MR image of the object 10 may be obtained.

The gradient coil 24 includes X, Y, and Z coils that generate gradient magnetic fields in the X-, Y-, and Z-axis directions that are perpendicular to each other. The gradient coil 24 may provide the location information of each part of the object 10 by inducing resonance frequencies differently for each part of the object 10.

The RF coil 26 may radiate an RF signal to the patient and receive an MR signal emitted from the patient. Specifically, the RF coil 26, after transmitting the RF signal of the same frequency as the frequency of the precession toward the atomic nucleus existing in the precession patient, stops transmitting the RF signal, and in the nucleus existing in the patient It can receive the MR signal emitted from.

For example, the RF coil 26 generates an electromagnetic signal, for example, an RF signal having a radio frequency corresponding to the type of the nuclear nucleus, such as an RF signal, in order to transition a nuclear nucleus from a low energy state to a high energy state. 10) can be applied. When an electromagnetic signal generated by the RF coil 26 is applied to a certain nucleus, the nucleus can transition from a low energy state to a high energy state. Subsequently, when the electromagnetic wave generated by the RF coil 26 disappears, the atomic nucleus to which the electromagnetic wave is applied may radiate an electromagnetic wave having a Lamor frequency while transitioning from a high energy state to a low energy state. In other words, when the application of the electromagnetic wave signal to the atomic nucleus is stopped, in the atomic nucleus to which the electromagnetic wave is applied, a change in the energy level from a high energy to a low energy may occur and an electromagnetic wave having a lamore frequency may be emitted. The RF coil 26 may receive an electromagnetic wave signal radiated from atomic nuclei inside the object 10.

The RF coil 26 may be implemented as one RF transmission / reception coil having a function of generating an electromagnetic wave having a radio frequency corresponding to a type of atomic nucleus and a function of receiving an electromagnetic wave radiated from the atomic nucleus. Further, it may be implemented as a transmitting RF coil having a function of generating an electromagnetic wave having a radio frequency corresponding to the type of atomic nucleus and a receiving RF coil having a function of receiving electromagnetic waves radiated from the atomic nucleus.

In addition, the RF coil 26 may be fixed to the gantry 20, it may be a removable form. The detachable RF coil 26 may include RF coils for portions of the object, including head RF coils, chest RF coils, leg RF coils, neck RF coils, shoulder RF coils, wrist RF coils and ankle RF coils, and the like. have.

In addition, the RF coil 26 may communicate with an external device in a wired and / or wireless manner, and may also perform dual tune communication according to a communication frequency band.

Also, the RF coil 26 may include a birdcage coil, a surface coil, and a transverse electromagnetic coil (TEM coil) according to the structure of the coil.

In addition, the RF coil 26 may include a transmission-only coil, a reception-only coil, and a transmission / reception combined coil according to an RF signal transmission / reception method.

In addition, the RF coil 26 may include various types of RF coils, such as 16 channels, 32 channels, 72 channels, and 144 channels.

The gantry 20 may further include a display 29 located outside the gantry 20 and a display (not shown) located inside the gantry 20. Predetermined information may be provided to the user or the object through the displays located inside and outside the gantry 20.

The signal transceiver 30 may control the gradient magnetic field formed in the gantry 20, that is, the bore according to a predetermined MR sequence, and control transmission and reception of the RF signal and the MR signal.

The signal transceiver 30 may include a gradient magnetic field amplifier 32, a transmission / reception switch 34, an RF transmitter 36, and an RF receiver 38.

The gradient amplifier 32 drives the gradient coil 24 included in the gantry 20, and outputs a pulse signal for generating the gradient magnetic field under the control of the gradient magnetic field controller 54. Can be supplied to By controlling the pulse signal supplied from the gradient amplifier 32 to the gradient coil 24, gradient magnetic fields in the X-axis, Y-axis, and Z-axis directions can be synthesized.

The RF transmitter 36 and the RF receiver 38 may drive the RF coil 26. The RF transmitter 36 may supply an RF pulse of a Lamore frequency to the RF coil 26, and the RF receiver 38 may receive an MR signal received by the RF coil 26.

The RF receiver 38 according to an embodiment of the present invention may include the acquirer 110 of FIG. 1, and acquire the undersampled spectrum on the k-space through the acquirer 110. The acquirer 110 according to an exemplary embodiment may shorten the data acquisition time by acquiring an undersampled spectrum on k-space corresponding to some spatial frequencies.

The transmit / receive switch 34 may adjust the transmit / receive direction of the RF signal and the MR signal. For example, the RF signal may be irradiated to the object 10 through the RF coil 26 during the transmission mode, and the MR signal from the object 10 may be received through the RF coil 26 during the reception mode. . The transmission / reception switch 34 may be controlled by a control signal from the RF controller 56.

The monitoring unit 40 may monitor or control the gantry 20 or devices mounted on the gantry 20. The monitoring unit 40 may include a system monitoring unit 42, an object monitoring unit 44, a table control unit 46, and a display control unit 48.

The system monitoring unit 42 includes a state of a static magnetic field, a state of a gradient magnetic field, a state of an RF signal, a state of an RF coil, a state of a table, a state of a device measuring body information of an object, a state of a power supply, a state of a heat exchanger, It can monitor and control the condition of the compressor.

The object monitoring unit 44 monitors the state of the object 10. In detail, the object monitoring unit 44 may include a camera for observing the movement or position of the object 10, a respiratory meter for measuring the respiration of the object 10, an ECG meter for measuring the electrocardiogram of the object 10, Or it may include a body temperature measuring instrument for measuring the body temperature of the object (10).

The table controller 46 controls the movement of the table 28 in which the object 10 is located. The table controller 46 may control the movement of the table 28 according to the sequence control of the sequence controller 50. For example, in moving imaging of an object, the table controller 46 may continuously or intermittently move the table 28 according to the sequence control by the sequence controller 50, thereby. The object may be photographed with an FOV larger than the field of view (FOV) of the gantry.

The display controller 48 controls the displays located on the outside and the inside of the gantry 20. In detail, the display controller 48 may control on / off of a display located on the outside and the inside of the gantry 20 or a screen to be output to the display. In addition, when the speaker is located inside or outside the gantry 20, the display controller 48 may control on / off of the speaker or sound to be output through the speaker.

The system controller 50 includes a sequence controller 52 for controlling a sequence of signals formed in the gantry 20, and a gantry controller 58 for controlling the gantry 20 and devices mounted on the gantry 20. can do.

The sequence controller 52 includes a gradient magnetic field controller 54 that controls the gradient magnetic field amplifier 32, and an RF controller 56 that controls the RF transmitter 36, the RF receiver 38, and the transmit / receive switch 34. can do. The sequence controller 52 may control the gradient amplifier 32, the RF transmitter 36, the RF receiver 38, and the transmit / receive switch 34 according to a pulse sequence received from the operating unit 60. Here, the pulse sequence includes all information necessary for controlling the gradient magnetic field amplifier 32, the RF transmitter 36, the RF receiver 38, and the transmit / receive switch 34, for example, the gradient. Information on the intensity of the pulse signal applied to the coil 24, the application time, the application timing (timing) and the like may be included.

The operating unit 60 may command pulse sequence information to the system controller 50 and control the operation of the entire magnetic resonance imaging system.

The operating unit 60 may include an image processor 62, an output unit 64, and an input unit 66 that receive and process an MR signal received by the RF receiver 38.

The image processor 62 may process MR signals received from the RF receiver 38 to generate MR image data of the object 10.

The image processor 62 receives the MR signal received by the RF receiver 38 and applies various signal processing such as amplification, frequency conversion, phase detection, low frequency amplification, filtering, and the like to the received MR signal.

The image processor 62, for example, arranges digital data in a k space of a memory (also referred to as a Fourier space or a frequency space), and performs the two-dimensional or three-dimensional Fourier transform on the data to perform an image. Can be reconstructed with data.

In addition, the image processor 62 may perform a composition process, a difference calculation process, or the like on the reconstructed image data. The composition process may be an addition process for pixels, a maximum projection (MIP) process, or the like. In addition, the image processor 62 may store not only the image data to be reconstructed, but also image data subjected to the synthesis process or the difference calculation process to a memory (not shown) or an external server.

In addition, various signal processings applied to the MR signal by the image processor 62 may be performed in parallel. For example, signal processing may be applied in parallel to a plurality of MR signals received by the multi-channel RF coil to reconstruct the plurality of MR signals into image data.

The image processor 62 according to an embodiment of the present invention may include the recovery unit 120 of FIG. 1 to generate the target image through the undersampled spectrum. In addition, the recovery unit 120 may include the first sub recovery unit 130, the second sub recovery unit 140, and the image generator 150 of FIG. 1.

The first sub-recovery unit 130 may perform initial recovery on the data corresponding to the non-sampled position on the k-space through a split Bregman technique or an approximate sparse encoding technique.

In addition, the second sub recovery unit 140 may perform dictionary learing recovery on the images corresponding to the decomposed band.

Therefore, the magnetic resonance image may be obtained at high speed through the image processor 62.

The output unit 64 may output the image data or the reconstructed image data generated by the image processor 62 to the user. In addition, the output unit 64 may output information necessary for the user to operate the MRI system, such as a user interface (UI), user information, or object information. The output unit 64 is a speaker, a printer, a CRT display, an LCD display, a PDP display, an OLED display, a FED display, an LED display, a VFD display, a digital light processing (DLP) display, a flat panel display (PFD), 3D Display, transparent display, and the like.

The user may input object information, parameter information, scan conditions, pulse sequences, information on image composition or difference calculation, etc. using the input unit 66. Examples of the input unit 66 may include a keyboard, a mouse, a trackball, a voice recognition unit, a gesture recognition unit, a touch screen, and the like.

7 illustrates the signal transceiver 30, the monitor 40, the system controller 50, and the operating unit 60 as separate objects, the signal transceiver 30, the monitor 40, and the system. Functions performed by each of the controller 50 and the operating unit 60 may be performed in another object. For example, although the image processing unit 62 described above converts the MR signal received by the RF receiver 38 into a digital signal, the conversion to the digital signal is performed by the RF receiver 38 or the RF coil 26 directly. It can also be done.

The gantry 20, the RF coil 26, the signal transmitting and receiving unit 30, the monitoring unit 40, the system control unit 50, and the operating unit 60 may be connected to each other wirelessly or by wire. The apparatus may further include an apparatus (not shown) for synchronizing clocks with each other. Communication between the gantry 20, the RF coil 26, the signal transmitting and receiving unit 30, the monitoring unit 40, the system control unit 50, and the operating unit 60 is performed at a high speed such as low voltage differential signaling (LVDS). Asynchronous serial communication such as a digital interface, a universal asynchronous receiver transmitter (UART), mistaken serial communication, or a low-delay network protocol such as a controller area network (CAN), optical communication, or the like may be used.

The communication unit 70 may be connected to at least one of the gantry 20, the signal transmission and reception unit 30, the monitoring unit 40, the system control unit 50, and the operating unit 60 shown in FIG. 7.

The communication unit 70 may exchange data with a hospital server or another medical device in the hospital connected through a picture archiving and communication system (PACS), and use digital imaging and communications (DICOM). Medicine) can communicate data according to the standard.

As shown in FIG. 8, the communication unit 70 may be connected to the network 80 by wire or wirelessly to communicate with the server 92, the medical device 94, or the portable device 96.

In detail, the communication unit 70 may transmit and receive data related to diagnosis of the object through the network 80, and may also transmit and receive medical images photographed by the medical device 94, such as CT, MRI, and X-ray. In addition, the communication unit 70 may receive a diagnosis history or a treatment schedule of the patient from the server 92 and use the same to diagnose the object. In addition, the communication unit 70 may perform data communication with not only the server 92 and the medical device 94 in the hospital, but also a portable device 96 such as a mobile phone, a PDA, a notebook or the like of a doctor or a customer.

In addition, the communication unit 70 may transmit an abnormality or medical image quality information of the magnetic resonance imaging system to the user through the network 80 and receive feedback from the user.

The communication unit 70 may include one or more components that enable communication with an external device, and may include, for example, a short range communication module 72, a wired communication module 74, and a wireless communication module 76. have.

The short range communication module 72 refers to a module for performing short range communication with a device located within a predetermined distance. Local area communication technology according to an embodiment of the present invention includes wireless LAN, Wi-Fi, Bluetooth, Zigbee, WFD (Wi-Fi Direct), UWB (ultra wideband), infrared communication (IrDA) , Infrared Data Association (BLE), Bluetooth Low Energy (BLE), Near Field Communication (NFC), and the like, but are not limited thereto.

The wired communication module 74 refers to a module for performing communication using an electrical signal or an optical signal, and a wired communication technology according to an embodiment of the present invention uses a pair cable, a coaxial cable, an optical fiber cable, or the like. Wired communication technology may be included.

The wireless communication module 76 transmits and receives a wireless signal with at least one of a base station, an external device, and a server on a mobile communication network. Here, the wireless signal may include various types of data according to transmission and reception of a voice call signal, a video call call signal, or a text / multimedia message.

Meanwhile, the above-described embodiments of the present invention can be written as a program that can be executed in a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium.

The computer-readable recording medium may be a magnetic storage medium (for example, a ROM, a floppy disk, a hard disk, etc.), an optical reading medium (for example, a CD-ROM, a DVD, etc.) and a carrier wave (for example, the Internet). Storage medium).

Although embodiments of the present invention have been described above with reference to the accompanying drawings, those skilled in the art to which the present invention pertains may implement the present invention in other specific forms without changing the technical spirit or essential features thereof. You will understand that. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

Claims

In the magnetic resonance imaging apparatus,

an acquisition unit for obtaining an undersampled spectrum on k-space; And

It includes a recovery unit for generating a target image through the undersampled spectrum,

The recovery unit may include: a first sub recovery unit configured to perform initial recovery on data corresponding to the non-sampled position on the k-space through a split Bregman technique or an approximate sparse encoding technique;

Alternately multi-band decomposition the initial recovered spectrum on the k-space into a plurality of individual spectra, sparse approximating and restoring the measured frequency for the images corresponding to the resolved band A second sub recovery unit performing dictionary learing recovery; And

And an image generator for merging the restored images of the respective frequency bands to generate a target image.
The method of claim 1, wherein the first sub recovery unit,

Collect a set of full sampled k-space data, apply a sampling mask, prepare undersampled data according to the index of the sampling mask, and full sample for error calculation The model is trained to predict sparse code for undersampled data, and the approximated sparse code is estimated from undersampled data using the trained model to approximate sparse code. Magnetic resonance imaging apparatus, characterized in that for performing the encoding.
The method of claim 1, wherein the sparse approximation extracts unevenly sized overlapping patches from an input image corresponding to an input magnetic resonance signal, and performs a sparse decomposition of the patches into sparse codes using a dictionary. And restoring patches by the sparse code and the dictionary, and restoring an image by the recovered patches.
2. The magnetic resonance imaging apparatus of claim 1, wherein the dictionary is previously trained based on patches extracted from a set of full sampled training magnetic resonance images.
The magnetic resonance imaging apparatus of claim 1, wherein the dictionary is individually learned for each frequency band.
2. The magnetic resonance imaging apparatus of claim 1, wherein the sparse decomposition is performed by minimizing non-zero components in a patch representation as a linear combination of dictionary elements.
The method of claim 1, wherein the decomposed k-spaces corresponding to the frequency bands are composed of frequencies taken from the initially recovered k-space corresponding to the current band, wherein the other frequencies are zero filled and measured. Magnetic resonance imaging apparatus, characterized in that displayed.
A method of generating a magnetic resonance image from data undersampled by a magnetic resonance apparatus,

The receiver coil obtaining an undersampled spectrum in k-space;

Initial retrieval by initial guessing the unsampled position in the k-space;

Multiband decomposition the initial recovered spectrum into a plurality of individual spectra;

Performing dictionary-based recovery on the images corresponding to each of the decomposed bands; And

Generating a target image by merging the restored image through the dictionary learning recovery;

a. The initial restoring comprises optimizing a cost function including a simplified LO norm;

b. Said multiband decomposition step applies to the spectrum initially recovered in k-space to obtain the number of individual spectra for various frequency bands;

c. The pre-based recovery step includes alternating sparse approximation and recovery of the measured frequency;

d. The generating of the target image may include generating the target image as a sum of the restoration results of the respective frequency bands.
10. The method of claim 8, wherein the normalization of the Norm norm is simplified through optimization of the Norm norm.
10. The method of claim 9, wherein the optimization of the l1 norm is implemented through a segmented Bregman technique or an approximate sparse coding method.
The method of claim 10, wherein the approximate sparse encoding method is:

a. Collecting a data set of full sampled k-spaces;

b. Applying a sampling mask and preparing undersampled data according to an index of the sampling mask;

c. Training the model to predict the sparse code for the undersampled data using the full sampled data for error calculation;

d. Predicting an approximated sparse code from the undersampled data using the trained model.
The method of claim 8, wherein the sparse approximation extracts unevenly sized overlapping patches from an input image corresponding to an input magnetic resonance signal, and performs a sparse decomposition of the patches into sparse codes using a dictionary. Restoring patches by the sparse code and the dictionary, and restoring the images by the recovered patches.
9. The method of claim 8, wherein the decomposed k-spaces corresponding to the frequency bands consist of frequencies taken from the initially restored k-space corresponding to the current band, wherein the other frequencies are zero filled and measured. Magnetic resonance image generation method characterized in that the display.
In a magnetic resonance imaging system,

An acquisition unit for obtaining an undersampled spectrum in k-space by the receiver coil; And

Receive and store the undersampled k-space transmitted from the acquiring unit, perform initial recovery by initial guessing a non-sampled location in the k-space, and store the initial recovered spectrum in a plurality of individual A recovery server for multiband decomposition into spectra, performing pre-based recovery of an image corresponding to the resolved band, merging the recovered images, and generating a target image;

a. The initial recovery is performed by optimizing the cost function, including a simplified 10 norm,

b. The multiband decomposition is applied to the spectrum initially recovered in k-space to obtain the number of individual spectra for the various frequency bands,

c. The pre-based recovery is performed by alternating sparse approximation and recovery of the measured frequency,

d. And a target image is generated by adding the recovery results of the respective frequency bands.
The method of claim 14,

The magnetic resonance system further comprises an operator console configured to receive a merge result from the recovery server and display the merge result through a display.