KR20210095526A - Apparatus and method for reconstructing mr parameter map - Google Patents

Abstract

Provided are an apparatus for reconstructing a magnetic resonance (MR) parameter map, which reconstructs a high-quality MR parameter map from undersampled k-space data, and a method thereof. According to the present invention, the apparatus comprises: an initial parameter mapping unit receiving and converting undersampled k-space data obtained according to a predetermined sequence into a sparse MR contrast image in a predetermined manner and estimating a pattern of the sparse MR contrast image according to a pre-learned pattern estimation method to obtain an initial MR parameter map; an MR signal modeling unit storing an MR signal model in which a relationship between a predetermined reference MR parameter map and the required MR parameter map to be reconstructed; and a parameter map reconstruction unit receiving the initial MR parameter map or a previously reconstructed MR parameter map to remove artifacts according to the pre-learned pattern estimation method, reconstructing the artifact-removed MR parameter map according to the MR signal model to obtain a reconstructed MR parameter map, and repeating reconstruction until the obtained reconstructed MR parameter map satisfies a predetermined condition.

Description

Magnetic resonance parameter map reconstruction apparatus and method {APPARATUS AND METHOD FOR RECONSTRUCTING MR PARAMETER MAP}

The present invention relates to an apparatus and method for reconstructing a magnetic resonance parameter map, and to a magnetic resonance parameter map reconstruction apparatus and method capable of generating a magnetic resonance parameter map from k-space data obtained by being undersampled.

Magnetic Resonance Imaging (MRI) is a representative medical image that can obtain non-invasive and high-quality images of the inside of the human body. Accordingly, it is used for diagnosis of various diseases or disorders, and in particular, it is useful for diagnosis of musculoskeletal systems such as bones, muscles, ligaments, tendons and cartilage due to excellent contrast with respect to soft tissues.

And magnetic resonance parameter maps (hereinafter referred to as MR parameter maps) that provide diagnostic information by indicating quantitative tissue characteristics of the human body such as _{T 1} , T ₂ or T ^* _{2 relaxation characteristics using MRI (or magnetic resonance characteristics)} Also called constant) can be obtained. MR parameter maps can be used in precision medicine because they provide useful information about a patient's disease diagnosis.

A conventional method for obtaining the MR parameter map is reverse recovery (inversion recovery), look for a reorganization T ₁ map-been proposed such as the rocker (look-locker) and VFA (variable flip angle) method, reconstruction T ₂ map A multi-echo spin-echo method has been proposed for , and several methods such as a multi-echo gradient-echo method for ^{T *} _{2 map reconstruction have been proposed.}

However, the existing methods have to repeatedly perform an MRI scan by applying different scan parameters in order to obtain a quantitative MR parameter map, so that the scan time increases. In addition, if the scan time becomes longer, it not only causes patient discomfort, but also increases the possibility of motion-related artifacts in the reconstructed image, so there is a limitation in that quantitative MR parameter maps cannot be easily applied to actual clinical protocols.

Korean Patent Registration No. 10-1886575 (Registered on August 1, 2018)

An object of the present invention is to provide a magnetic resonance parameter map reconstruction apparatus and method for reconstructing a high-quality MR parameter map from undersampled k-space data.

Another object of the present invention is to provide an apparatus and method for reconstructing a magnetic resonance parameter map capable of reconstructing an MR parameter map at high speed using an artificial neural network and a model-based data maintenance technique.

The magnetic resonance parameter map reconstruction apparatus according to an embodiment of the present invention for achieving the above object receives k-space data obtained by under-sampled according to a predetermined sequence and converts it into a sparse MR contrast image in a predetermined manner, , an initial parameter mapping unit configured to obtain an initial MR parameter map by estimating a pattern of a sparse MR contrast image according to a pre-learned pattern estimation method; an MR signal modeling unit storing in advance an MR signal model in which a relationship between a predetermined reference MR parameter map and a required MR parameter map requiring reconstruction is modeled; and the initial MR parameter map or the previously reconstructed reconstructed MR parameter map is applied, artifacts are removed according to a pre-learned pattern estimation method, and the artifact-removed MR parameter map is reconstructed according to the MR signal model, reconstructed MR parameters and a parameter map reconstruction unit that acquires a map and repeats reconstruction until the obtained reconstruction MR parameter map satisfies a predetermined condition.

The parameter map reconstruction unit is implemented as a pre-trained artificial neural network to receive an initial MR parameter map or a previously reconstructed reconstructed MR parameter map, and a pattern estimation method learned from the applied initial MR parameter map or the previously reconstructed reconstructed MR parameter map. a learning-based reconstruction unit that reconstructs the MR parameter map by removing aliasing artifacts accordingly; a model-based reconstruction unit that receives the artifact-removed MR parameter map and reconstructs the MR parameter map so that a data relationship of the artifact-free MR parameter map is maintained according to the MR signal model; and a repetition determining unit that transmits the reconstructed reconstructed MR parameter map to the model-based reconstruction unit when it is determined that the reconstructed reconstructed MR parameter map satisfies a predetermined condition to be repeated.

The learning-based reconstruction unit divides the applied initial MR parameter map or the previously reconstructed reconstructed MR parameter map into the reference MR parameter map and at least one required MR parameter map, the reference MR parameter map or the at least one request a plurality of neural networks for removing artifacts of corresponding MR parameter maps in the MR parameter maps; and a combiner for combining and outputting the outputs of the plurality of neural networks.

The model-based reconstruction unit applies a coil sensitivity map and a radio frequency transmission field map, which are parameters specified for performing a sequence used to obtain the k-space data together with the MR signal model and the k-space data, to obtain MR The parameter map can be reconstructed.

The iterative determination unit may output the current reconstructed MR parameter map if it is determined that the iteratively reconstructed reconstructed reconstructed MR parameter map converges, or the difference between the currently acquired reconstructed MR parameter map and the previous reconstructed MR parameter map is less than a predetermined reference value. there is.

The initial parameter mapping unit may include: a parameter map conversion unit that receives the k-space data and converts it into the sparse MR contrast image by filling a region corresponding to the missing data in the k-space data with 0 according to a predetermined method; and an initial parameter map acquisition unit implemented as an artificial neural network to acquire the initial MR parameter map according to a pattern estimation method previously learned from a sparse MR contrast image.

The initial parameter map obtaining unit may obtain an initial MR parameter map including the reference MR parameter map and the requested MR parameter map.

In a magnetic resonance parameter map reconstruction method according to another embodiment of the present invention for achieving the above other object, k-space data obtained by under-sampling according to a predetermined sequence is applied and converted into a sparse MR contrast image in a predetermined manner. and obtaining an initial MR parameter map by estimating a pattern of a sparse MR contrast image according to a pre-learned pattern estimation method; The initial MR parameter map or the previously reconstructed reconstructed MR parameter map is received, artifacts are removed according to a pre-learned pattern estimation method, and the artifact-removed MR parameter map is combined with a predetermined reference MR parameter map and the required MR that requires reconstruction. reconstructing the relationship between the parameter maps according to the modeled MR signal model to obtain a reconstructed MR parameter map, and repeating the reconstruction until the obtained reconstructed MR parameter map satisfies a predetermined condition.

Therefore, an apparatus and method for reconstructing a magnetic resonance parameter map according to an embodiment of the present invention is to use an artificial neural network and a model-based data retention technique to obtain a high-quality MR parameter map from k-space data obtained by under-sampling for a short time. can be reconstructed, shortening the imaging time, greatly improving patient convenience, and reducing imaging costs as well as various errors caused by a difference in imaging time.

1 shows a schematic structure of an apparatus for reconstructing a magnetic resonance parameter map according to an embodiment of the present invention.
FIG. 2 is a view for explaining the operation of the magnetic resonance parameter map reconstruction apparatus of FIG. 1 .
FIG. 3 shows an example of an artificial neural network constituting the initial parameter map acquisition unit of FIG. 1 .
4 shows an example of an artificial neural network constituting the learning-based reconstruction unit of FIG. 1 .
FIG. 5 shows a detailed configuration of a parameter map reconstruction unit including the learning-based reconstruction unit of FIG. 4 .
6 is a diagram for explaining a method for reconstructing a magnetic resonance parameter map according to an embodiment of the present invention.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings illustrating preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, the present invention will be described in detail by describing preferred embodiments of the present invention with reference to the accompanying drawings. However, the present invention may be embodied in various different forms, and is not limited to the described embodiments. In addition, in order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components unless otherwise stated. In addition, terms such as "... unit", "... group", "module", and "block" described in the specification mean a unit that processes at least one function or operation, which is hardware, software, or hardware. and a combination of software.

1 shows a schematic structure of an apparatus for reconstructing a magnetic resonance parameter map according to an embodiment of the present invention, and FIG. 2 is a view for explaining an operation of the apparatus for reconstructing a magnetic resonance parameter map of FIG. 1 .

Referring to FIG. 1 , the magnetic resonance parameter map reconstruction apparatus according to the present embodiment includes a data acquisition unit 100 , an initial parameter mapping unit 200 , a parameter map reconstruction unit 300 , and an MR signal modeling unit 400 . can do.

The data acquisition unit 100 acquires undersampled multi-channel k-space data (b) by imaging an imaging target according to a predetermined sequence corresponding to the MR parameter map using the MRI apparatus. ^{The initial parameter mapping unit 200 obtains a sparse MR contrast image (H *} b) by transforming the undersampled multi-channel k-space data (b) in a predetermined method, and obtains it according to a pre-learned pattern estimation method An initial MR parameter map (x ₀ ) is obtained from the sparse MR contrast image (H ^{* b).} And the parameter map reconstruction unit 300 according to the pre-learned pattern estimation method and a predetermined magnetic resonance signal model (hereinafter referred to as MR signal model) the initial MR parameter map (x ₀ ) or the previously obtained MR parameter map (x _k ) of aliasing artifacts (aliasing artifacts) and the MR parameter map reconstructed by repeating the process of removing artifacts, such as noise or the like, and reconstruct the MR parameter map the artifact is removed (D _R (x _k)) based on the MR signal model ( x _k+1 ).

Each configuration of the magnetic resonance parameter map reconstruction apparatus of FIG. 1 will be described in detail with reference to FIG. 2 . First, the data acquisition unit 100 is implemented as a magnetic resonance imaging apparatus (hereinafter referred to as an MRI apparatus) and uses various predetermined sequences, Undersampling k-space data (b) is obtained. In order to reduce the scan time of the MR parameter map, the data acquisition unit 100 samples the undersampled k-space data (b) at a frequency lower than the Nyquist rate instead of acquiring all data for the k-space. ) can be obtained. Also, the data acquisition unit 100 may acquire undersampled multi-channel k-space data based on a signal sensed through a multi-coil of the MRI apparatus. .

_{In FIG. 2 , a spoiled gradient echo (SPGR) T 1} (T ₁ SPGR) k-space obtained by the data acquisition unit 100 according to a variable flip angle (VFA) MR signal model among various MR parameter maps. Data is shown as an example. Here, the SPGR T ₁ k-spatial data (b) is multi-channel obtained through multiple coils for _{different flip angles (α 1} to α _n ) at a plurality of scan times (n), respectively, as shown in FIG. 2 . It appears as k-space data, and it can be seen that missing data (or missing data) is generated in each of the k-space data as the undersampling is performed, thereby generating a missing line or region.

The initial parameter mapping unit 200 converts the undersampled multi-channel k-space data (b) into a sparse MR contrast image (H ^* b), and learns in advance from the transformed sparse MR contrast image (H ^{* b)} An initial MR parameter map (x ₀ ) is obtained according to the obtained pattern estimation method.

The initial parameter mapping unit 200 may include a contrast image converting unit 210 and an initial parameter map obtaining unit 220 .

The contrast image conversion unit 210 receives the undersampled multi-channel k-space data (b) obtained by the data acquisition unit 100, and performs an inverse Fourier transform on the applied undersampled k-space data (b). ^{A sparse MR contrast image (H *} b) is obtained by transforming according to a predetermined transformation technique such as et al. In this case, the sparse MR contrast image (H ^* b) is an image in which unsampled data in k-space is zero-filled as the undersampled multi-channel k-space data (b) is transformed.

In addition, the contrast image conversion unit 210 applies coil sensitivity maps (C) indicating the sensitivity of multi-coils among the parameters set by the data acquisition unit 100 to undersampled multi-channels. ^{A precise sparse MR contrast image (H *} b) can be obtained from the k-space data (b).

In FIG. 2 , since the obtained sparse MR contrast image (H ^* b) is _{obtained by transforming the SPGR T 1} k-space data (b), it is expressed as an _{SPGR T 1 parameter map.}

The initial parameter map obtaining unit 220 estimates the pattern of the sparse MR contrast image (H ^* b) obtained by the contrast image conversion unit 210 according to a pre-learned pattern estimation method, and the initial MR parameter map (x ₀ or x _k=0 ). As shown in FIG. 2 , the initial parameter map acquisition unit 220 may be implemented as an artificial neural network in which a pattern estimation method such as CNN has been previously learned.

_{In addition, the initial MR parameter map x 0} obtained by the initial parameter map acquisition unit 220 is an MR parameter map corresponding to a sequence for acquiring k-space data b and at least one MR parameter map of a different type. It may be obtained in the form of a mapping map indicating a shape or a relationship between different MR parameter maps. As an example, here, it is assumed that the _{initial MR parameter map (x 0 ) is combined with a T 1} MR parameter map required in response to a sequence for acquiring k-space data (b) and a different M _{0 MR parameter map.} . Here, the T ₁ MR parameter map is a required MR parameter map and may be referred to as a required MR parameter map, and the M ₀ MR parameter map is the most basic MR parameter map, and an MR parameter according to an MR signal model (S) to be described later It may be referred to as a reference MR parameter map for maintaining a relationship between maps. In FIG. 2 , as an example, two MR parameter maps of T ₁ and M ₀ are obtained from the _{initial MR parameter map (x 0 ), but other MR parameter maps such as T 2} and T ₂ ^* are obtained, or a plurality of An MR parameter map may be configured to be obtained.

Meanwhile, the parameter map reconstruction unit 300 removes artifacts based on the pre-learned pattern estimation method, and repeats the process of reconstructing so that the relationship between the required MR parameter map and the reference MR parameter map is maintained according to a predetermined MR signal model. to output the reconstructed MR parameter map (x _k+1). The parameter map reconstruction unit 300 may include a learning-based reconstruction unit 310 , a model-based reconstruction unit 320 , and an iteration determining unit 330 .

The learning-based reconstruction unit 310 receives the initial MR parameter map (x ₀ ) or the previously reconstructed MR parameter map (x _k ) obtained by the initial parameter map obtainer 220 , and receives the applied initial MR parameter map (x k ). ₀ ) or the previously reconstructed MR parameter map (x _k ) to remove artifacts according to a pre-learned pattern estimation scheme. At this time, the learning-based reconstruction unit 310 distinguishes the required MR parameter map and the reference MR parameter map constituting the _{initial MR parameter map (x 0} ) or the previously reconstructed MR parameter map (x _{k ), and divides the required MR parameter} After removing artifacts from each of the map and the reference MR parameter map, they can be combined and output.

Then, the model-based reconstruction unit 320 receives the artifact-removed MR parameter map D _R (x _k ), and acquires the undersampled k-space data (b) and the data obtained by the data acquisition unit 100 . part of the parameters set on the _{(100) (C, B 1} +) and the MR signal modeling unit 400, the MR parameter map (D _R (x _k)), the artifacts are removed based on the MR signal model (s) model in by reconstructing to obtain a reconstructed MR parameter map (x _{k+1 ).}

Here, the learning-based reconstruction unit 310 _{removes artifacts included in the initial MR parameter map (x 0} ) by estimating the blank space of the ^{sparse MR contrast image (H *} b) by the initial parameter map acquisition unit 220 . Since the initial MR parameter map (x ₀ ) is obtained, the obtained initial MR parameter map (x ₀ ) includes a large number of artifacts, and removing artifacts from the _{previously reconstructed MR parameter map (x k ) is iteratively} this is because by removing the artifact as well as to improve the quality of the MR parameter map, artifacts may additionally occur in the process of reconfiguring the MR parameter map (D _R (x _k)) by the model-based reconstruction unit 320 . Not only is it intended to be able to reconstruct a learning-based reconstruction unit 310 by removing the artifact, the model-based reconstruction unit 320 is more easily MR parameter map (D _R (x _k)).

And a model-based reconstruction unit 320 is reconstructing the MR parameter map (D _R (x _k)) based on the MR signal model (S) being applied to the MR signal modeling unit (400) for obtaining the k- space data This is to obtain a _{reconstructed MR parameter map (x k+1} ) while the required MR parameter map according to the MR sequence and the reference MR parameter map maintain a data relationship according to the MR signal model S.

Meanwhile, the iteration determining unit 330 determines whether to apply the reconstruction MR parameter map (x _k+1 ) obtained from the model-based reconstruction unit 320 again to the learning-based reconstruction unit 310 . Although the artifacts are removed in the learning-based reconstruction unit 310 and reconstructed in the model-based reconstruction unit 320 to obtain the reconstruction MR parameter map (x _k+1 ), only one time of removal and reconstruction of the artifacts is required. It is difficult to obtain high-quality MR parameter maps. The repeat the determining unit 330 determines whether or not by analyzing the reconstructed MR parameter map (x _{k + 1),} reconstructed MR parameter map (x _{k + 1)} to what is applied repeatedly to a learning-based reconstruction unit (310). In this case, the repetition determining unit 330 may determine whether to repeat according to a predetermined condition. For example, the iteration determining unit 330 repeatedly applies the reconstructed MR parameter map (x _k+1 ) to the learning-based reconstructor 310 until the reconstructed MR parameter map applied from the model-based reconstructor 320 converges. can do. However, since it may take a very long time for the reconstructed MR parameter map to completely converge, the iterative determining unit 330 may determine between the currently obtained reconstructed MR parameter map (x _k+1 ) and the previously obtained reconstructed MR parameter map (x _k ). If the difference of is greater than or equal to a predetermined reference value, _{the process of applying the reconstruction MR parameter map (x k+1} ) to the learning-based reconstruction unit 310 is repeated, and if it is less than the reference value, the obtained reconstruction MR parameter map (x _k+1 ) It may be configured to output.

Here, the parameter map reconstruction unit 300 may iteratively obtain a _{reconstructed MR parameter map (x k+1) based on iterative gradient descent.}

The MR signal modeling unit 400 stores the MR signal model S corresponding to the MR parameter map requiring reconstruction. Here, required of the MR parameter map hayeoteumeuro the case of using the VFA model for obtaining SPGR image to obtain a T ₁ map home, MR signal modeling unit 400 is to T ₁ and the M ₀ parameter corresponding to T ₁ mapping It is possible to obtain and store the MR signal model (S) in advance.

Hereinafter, the operational meaning of each configuration according to the purpose of the MR parameter map reconstruction apparatus of FIG. 1 will be described.

The MR parameter map reconstruction apparatus as shown in FIG. 1 aims to reconstruct a _{required MR parameter map (here, a T 1 map as an example) (x) from the acquired undersampled k-space data (b).} First, if the relationship between the undersampled k-space data (b) and the MR parameter map (x) is expressed using the operator (A), it can be expressed as in Equation (1).

Here, the operator (A) is the magnetic resonance signal model (S), the coil sensitivity map (C), the Fourier used by the MR parameter map reconstruction apparatus of FIG. 1 to obtain the MR parameter map (x) from the k-space data (b). As a function configured to include both a Fourier transform function (F) and a k-space sampling matrix (M), it can be expressed by Equation (2).

However, as described above, since the k-space data obtained by the data acquisition unit 100 is undersampled multi-channel k-space data (b), the MR parameter map is obtained from the undersampled multi-channel k-space data (b). Since reconstructing (x) is an ill-posed problem, there is a limit in that it is difficult to define an explicit operator (A) for reconstructing the MR parameter map (x) from the k-space data (b). Accordingly, by approaching the operator (A) of Equation 1 as a least-squares problem, it can be expressed by Equation 3 in a way of minimizing together with a regularization term (R).

where ≫ ₂ is an L2-norm function, and λ is a regularization variable for adjusting the balance between data consistency and the regularization term (R).

In general, the normalization term (R) is, for example, a wavelet transform, a total variation and a total generalized variation in a transformed domain to reconstruct using sparsity. is devised However, in the case of reconstructing MRI from undersampled k-space data using normalization functions using these predefined features, due to over-smoothing problems or residual aliasing artifacts due to excessive normalization, There are limits.

In order to overcome the above limitation, the apparatus for reconstructing an MR parameter map according to the present embodiment shown in FIG. 1 applies regularization using a learning-based artificial neural network to reconstruct an MR parameter map from undersampled k-space data.

The regularization term (R(x)) using the learning-based artificial neural network can be expressed as Equation (4).

Here, D _R (x) is a function representing the operation of the artificial neural network to obtain the artifact-free MR parameter map.

_{In Equation 4, the subtraction (R} (x)) between the MR parameter map (x) and the de-artifacted MR parameter map (DR (x)) means artifacts such as residual aliasing and noise. Therefore, by substituting the normalization term (R (x)) is calculated as the difference between mathematical MR parameter maps of the formula 4 (x) and the MR parameter map the artifact is removed (D _R (x)) for equation (3) to equation (5) can be expressed again.

And the least-squares problem of Equation 5 can be solved by iterative gradient descent as in Equation 6.

In Equation 6, r _{k denotes the difference between the k-space data obtained by applying the MR parameter map (x k} ) according to the number of iterations (k) to the operator (A) and the undersampled k-space data (b). , and J _A is a Jacobian matrix for the operator (A), and may be expressed as in Equation 7.

where x _k is the reconstructed image at the k (k = 0, …, N _k )th iteration, μ _k is the step size at k, and α _n is the nth flip angle (flip). angle), n is the number of MRI scans for parameter mapping, v _k ^m is _{the m-th variable of x k} , and m is the number of parameter maps to be reconstructed. Since the T ₁ map is assumed as an example here, the variables (v ¹ , v ² _{) of x k} ^{are (v 1} , v ² ) = (T ₁ ,M ₀ ) with the T1 map and MRI.

B₁ ⁺(x,y)로 보정될 수 있다. 여기서 x와 y는 영상의 공간적 좌표를 의미한다. 또한 상기한 바와 같이, SPGR을 획득하는 VFA 자기 공명 신호 모델의 경우, MR 신호 모델(S)은 수학식 8과 같이 나타낸다. _{On the other hand, if the B 1} ⁺ map, which is a radio frequency transmission field map, is additionally considered, the flip angle (α _n ) is α _n (x,y)=α _n (x,y)

It can be corrected as B ₁ ⁺ (x,y). Here, x and y mean the spatial coordinates of the image. Also, as described above, in the case of the VFA magnetic resonance signal model for obtaining the SPGR, the MR signal model (S) is expressed as in Equation (8).

Here, TR means a repetition time.

And if the MR signal model S of Equation 8 is differentiated by T ₁ and M _{0 , which are variables of x k} , respectively, it can be expressed as Equation 9.

However, here, the MR signal model S and _{the variables of x k} may be changed according to the MR sequence for acquiring k-space data and the MR parameter map to be acquired. For example, if you want to acquire a _{T 2} map or a T ₂ ^* map, the variables (v ¹ , v ² _{) of x k} are (T ₂ , M ₀ ) or (T ₂ ^* , M ₀ ), and the MR signal model ( S) may be a multi spin-echo or gradient-echo sequence.

The MR parameter map reconstruction apparatus of FIG. 1 is a device configured based on Equation 6, and the parameter map reconstruction unit 300 performs artificial neural network-based normalization from k-space data at every iteration, a coil sensitivity map (C), and additional radio The frequency transmission field map (B ₁ ⁺ ) is configured to reconstruct the magnetic resonance parameter map using the MR signal model (S).

However, in order to find a solution according to Equation 6, an appropriate initial parameter map (x ₀ or x _k=0 ) must be obtained, and thus, in FIG. 1 , the initial parameter mapping unit 200 obtains the _{initial parameter map (x 0 )} .

In the case of obtaining MRI from the undersampled k-space data (b), the k-space data (b) is substituted for the ^{operator (A * ), which is the inverse of the operator (A) in Equation 1, and is sampled in the k-space} A sparse MRI with zero-filled data can be easily obtained. However, in the MR parameter map, sparse MR contrast images cannot be obtained by ^{directly substituting k-space data (b) into the inverse operator (A *} ) due to the complexity and nonlinearity of the MR signal model (S). Accordingly, in FIG. 1 , the contrast image conversion unit 210 of the initial parameter mapping unit 200 applies a coil sensitivity map (C) to the undersampled k-space data (b) to obtain a sparse MR contrast image (H ^* b). ) is obtained.

^{Here, the operator (H *} ) representing the operation of the contrast image conversion unit 210 is an inverse function of the operator (H) for generating multi-channel k-space data from the MR contrast image to be acquired by the magnetic resonance parameter map reconstruction apparatus. , operator (H) may be expressed by Equation (10).

Here, C denotes a coil sensitivity map, F denotes a Fourier transform, and M denotes a k-space sampling matrix.

And when the operation of the initial parameter map obtainer 220 obtained by estimating the initial MR parameter map (x ₀ ) from the sparse MR contrast image (H ^* _{b) is expressed as a function (D M} ), the sparse MR contrast The relationship between the image (H ^* b) and the initial MR parameter map (x _{k = 0} ) may be expressed as in Equation (11).

FIG. 3 shows an example of an artificial neural network constituting the initial parameter map acquisition unit of FIG. 1 .

As described above, the initial parameter map acquisition unit 220 may be implemented as an artificial neural network, and FIG. 3 illustrates a case in which it is implemented as a convolutional neural network (CNN) as an example. As shown in FIG. 3 , the initial parameter map acquisition unit 220 includes a plurality of convolution blocks connected in series, and each of the plurality of convolution blocks is a convolution layer CL. ), a normalization layer (RL), and a nonlinear activation function layer (AL).

The initial parameter map obtaining unit 220 _{estimates the initial MR parameter map (x k=0} ) according to the learned pattern estimation method by receiving the ^{sparse MR contrast image (H *} b), and in this case, the initial MR parameter map ( x _k=0 ) may be set together with an initial variable (x _k=0 = (T ₁ , M ₀ ) _k=0 of the required MR parameter map, that is, the required MR parameter map and the reference MR parameter map are to be designated. can

4 shows an example of an artificial neural network constituting the learning-based reconstruction unit of FIG. 1 , and FIG. 5 shows a detailed configuration of the parameter map reconstruction unit including the learning-based reconstruction unit of FIG. 4 .

로 표현될 수 있다. 여기서는 2개의 변수(T₁, M₀)만을 고려하여, 2개의 신경망(311, 312)만을 도시하였으나, 요구 파라미터 맵의 종류가 증가되면, 즉 변수의 개수가 증가되면 신경망의 개수 또한 증가될 수 있다.Referring to FIG. 4 , the learning-based reconstruction unit 310 for removing artifacts from the previously obtained MR parameter map (x _{k ) divides the MR parameter map (x k} ) according to the variables (T ₁ , M ₀ ) _k by dividing the , may include a number of neural networks 311 and 312 configured to remove artifacts for each corresponding variable. In this case, the learning-based reconstruction unit 310 that removes artifacts including a plurality of neural networks 311 and 312 is

can be expressed as Here, only the two neural networks 311 and 312 are shown in consideration of only the two variables (T ₁ , M ₀ ). However, if the type of the required parameter map is increased, that is, if the number of variables is increased, the number of neural networks can also be increased. there is.

_{The learning-based reconstruction unit 310 classifies the MR parameter map (x k} ) according to the variable ((T ₁ , M ₀ ) _k ) as shown in FIG. 4 , and the divided variable ((T ₁ , M ₀ ) _k ) The reason why artifacts are removed by using different neural networks 311 and 312 for ? is because the MR parameter maps (T ₁ map, M ₀ map) according to each variable have different tissue properties. Each of the two neural networks 311 and 312 includes a plurality of convolution blocks connected in series similarly to the initial parameter map acquisition unit 220 of FIG. 3 , and each convolution block is a convolution block. It may include a solution layer (CL), a normalization layer (RL), and a nonlinear activation function layer (AL). In addition, an adder (AD) that adds the _{variables ((T 1} , M ₀ ) _k ) input to the output after aliases are removed through a plurality of convolution blocks to have a residual learning structure ) may be included.

And learning-based reconstruction unit 310 combines (Concatenate) the output of the two neural networks 311, 312 may further include a coupling portion 313 for outputting the MR parameter map (D _R (x _k)) .

When performing learning based reconstruction unit 310, the artifacts are removed the MR parameter map (D _R (x _k)) to repeat gradient descent by applying the MR signal model (S) for in the method, equation (6) is Equation (12) can be rewritten as

As shown in equation (12), the parameter map reconstruction unit 300 includes a previous reconstructed MR parameter map (x _k) and a learning-based reconstruction unit 310, the MR parameter map (D _R (x _k)), remove artifacts, and ( Reconstructed using A(x _k )-b). where λ _k and μ _k are learnable regularization and increase/decrease magnitude variables according to the number of iterations (k). And (A(x _k )-b) may be calculated by applying the MR signal model (S) according to Equations 7 and 8.

According to Equation 12, the model-based reconstruction unit 320 in the parameter map reconstruction unit 300 may be configured as shown in FIG. 5 .

When the sparse MR contrast image is obtained from the k-space data (b) in the MR parameter map reconstruction apparatus, the iterative gradient descent unit 330 repeats the obtained sparse MR contrast image through the model-based reconstruction unit 320 . by making repeated in accordance with the law MR, but also be configured to obtain the parameter map, in this _{case, (D R (x k +} 1)) MR parameter map because of the complexity and nonlinearity of the MR signal model (S), as described above the not only can occur if not reconstructed, a very long time it takes to obtain the MR parameter maps of the required level _{(D R (x k + 1} )).

On the other hand, the MR parameter map reconstruction apparatus allows the model-based reconstruction unit 320 to remove the aliasing artifact at every iteration, so that the calculation is possible even in the complexity and nonlinearity of the MR signal model S, and the calculation time can be greatly reduced. there is. The initial parameter map obtaining unit 220 also obtains an initial MR parameter map (x ₀ ^{) from the sparse MR contrast image (H *} b) and transmits it to the parameter map reconstructor 300, so that the parameter map reconstruction unit 300 By making it possible to greatly reduce the number of repetitions k, it is possible to easily obtain the _{MR parameter map D R} (x _{k+1 ).}

However, in the MR parameter map reconstruction apparatus according to the present embodiment, the initial parameter map acquisition unit 220 of the initial parameter mapping unit 200 and the learning-based reconstruction unit 310 of the parameter map reconstruction unit 300 are implemented as artificial neural networks. , prior learning is required.

Accordingly, the MR parameter map reconstruction apparatus may further include a learning unit 500 as shown in FIG. 1 . The learner 500 is used only when the MR parameter map reconstruction apparatus is trained, and may be removed when the MR parameter map reconstruction apparatus is actually used.

Here, the learning unit 500 calculates the initial loss (L _{M ) of the initial parameter map acquisition unit 220 and the reconstruction loss (L R} ) of the parameter map reconstruction unit 300 including the learning-based reconstruction unit 310 , respectively. to _{obtain a total loss L T} , and backpropagating the obtained total loss L _T , the initial parameter map obtainer 220 and the learning-based reconstruction unit 310 may be trained.

First, the initial loss (L _M ) may be obtained according to Equation (13).

)에 대입하여 획득된 희소 MR 대조도 영상 사이의 평균 절대 오차(mean absolute error)로 계산될 수 있다.Here, N _S is the number of training data. According to Equation 13, the initial loss (L _M ) is the initial MR parameter map (x ₀ ) estimated by the initial parameter map obtainer 220 from the sparse MR contrast image (H ^* b) using the operator (

) can be calculated as the mean absolute error between the obtained sparse MR contrast images.

Meanwhile, the reconstruction loss (L _R ) may be calculated according to Equation (14).

Here, l _L and l _S mean the weight of a structural similarity (SSIM) loss function with l1, respectively, and m _t is a reference MR parameter map that is training data. The reconstruction loss (L _R ) is added by calculating the mean absolute error and the structural similarity loss function between the reference MR parameter map and the reconstructed reference MR parameter map. By combining the mean absolute error and the structural similarity loss function, it is possible to reduce the error while preserving the detail of the MR parameter map.

From Equations 13 and 14, the total loss L _T can be calculated as Equation 15.

When the total loss L _T is obtained according to Equation 15, the initial parameter map obtainer 220 and the learning-based reconstruction unit 310 are trained by backpropagating the obtained total loss L _{T .}

6 is a diagram for explaining a method for reconstructing a magnetic resonance parameter map according to an embodiment of the present invention.

1 to 5, the MR parameter map reconstruction method of FIG. 6 will be described. First, the MRI apparatus acquires undersampled multi-channel k-spatial data (b) by imaging a subject according to a predetermined sequence. do (S10). Then, the obtained undersampled multi-channel k-space data (b) is converted into a sparse MR contrast image (H ^* b) according to a predetermined transformation technique such as an inverse Fourier transform (S20). When the sparse MR contrast image (H ^* b) is obtained, an initial MR parameter map (x ₀ ) is obtained by estimating the pattern of the obtained sparse MR contrast image (H ^{* b) according to a pre-learned pattern estimation method.} (S30).

When the initial MR parameter map (x _{0 ) is obtained, artifacts included in the MR parameter map (x 0} ) are removed according to a pre-learned pattern estimation method ( S40 ). And normalizing the MR signal model (S) for a given period MR parameter map (D _R (x ₀₎₎ to remove the artifacts are based reconstructs the MR parameter map (x ₁₎ (S50). If the MR parameter map (x ₁ ) is reconstructed, it is determined whether the reconstructed MR parameter map (x ₁ ) satisfies a condition according to the least squares method, and whether or not repetition is estimated ( S60 ). That is, whether the reconstructed MR parameter map (x ₁ ) converges or whether the difference is less than a predetermined reference value compared with the previously reconstructed MR parameter map (x _{0 ) may be determined to determine whether to repeat.}

If it is determined that it needs to be repeated, _{artifacts are removed according to the pattern estimation method learned in advance for the reconstructed MR parameter map (x k} ) again (S40), and the next MR parameter is normalized based on the MR signal model (S). The map (x _k+1 ) is reconstructed (S50).

However, if it is determined not to repeat, the reconstructed MR parameter map is output (S70).

The method according to the present invention may be implemented as a computer program stored in a medium for execution by a computer. Here, the computer-readable medium may be any available medium that can be accessed by a computer, and may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, and read dedicated memory), RAM (Random Access Memory), CD (Compact Disk)-ROM, DVD (Digital Video Disk)-ROM, magnetic tape, floppy disk, optical data storage, and the like.

Although the present invention has been described with reference to the embodiment shown in the drawings, which is merely exemplary, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom.

Accordingly, the true technical protection scope of the present invention should be defined by the technical spirit of the appended claims.

100: data acquisition unit 200: initial parameter mapping unit
210: contrast image conversion unit 220: initial parameter map acquisition unit
300: parameter map reconstruction unit 310: learning-based reconstruction unit
320: model-based reconstruction unit 330: iteration discrimination unit
400: MR signal modeling unit 500: learning unit

Claims (20)

The k-space data obtained by being undersampled according to a predetermined sequence is received and converted into a sparse MR contrast image in a predetermined manner, and the pattern of the sparse MR contrast image is estimated according to a pre-learned pattern estimation method to estimate the initial MR an initial parameter mapping unit for obtaining a parameter map;
an MR signal modeling unit storing in advance an MR signal model in which a relationship between a predetermined reference MR parameter map and a required MR parameter map requiring reconstruction is modeled; and
Reconstructed MR parameter map by receiving the initial MR parameter map or the previously reconstructed reconstructed MR parameter map, removing artifacts according to a pre-learned pattern estimation method, and reconstructing the artifact-removed MR parameter map according to the MR signal model A magnetic resonance parameter map reconstruction apparatus comprising: a parameter map reconstruction unit that obtains , and repeats reconstruction until the obtained reconstruction MR parameter map satisfies a predetermined condition. The method of claim 1, wherein the parameter map reconstruction unit
Implemented as a pre-trained artificial neural network to receive an initial MR parameter map or a previously reconstructed reconstructed MR parameter map, and remove aliasing artifacts according to a pattern estimation method learned from the applied initial MR parameter map or previously reconstructed reconstructed MR parameter map a learning-based reconstruction unit to reconstruct the MR parameter map;
a model-based reconstruction unit that receives the artifact-removed MR parameter map and reconstructs the MR parameter map so that a data relationship of the artifact-free MR parameter map is maintained according to the MR signal model; and
A magnetic resonance parameter map reconstruction apparatus comprising a repetition determining unit that passes the reconstructed reconstruction MR parameter map to the model-based reconstruction unit when it is determined that the reconstructed reconstruction MR parameter map satisfies a predetermined condition to be repeated. The method of claim 2, wherein the learning-based reconstruction unit
By dividing the applied initial MR parameter map or the previously reconstructed reconstructed MR parameter map into the reference MR parameter map and at least one required MR parameter map, a corresponding one of the reference MR parameter map or the at least one required MR parameter map multiple neural networks that remove artifacts of the MR parameter map; and
Magnetic resonance parameter map reconstruction apparatus comprising a combiner for outputting by combining the outputs of the plurality of neural networks. The method of claim 3, wherein the model-based reconstruction unit
Reconstructing the MR parameter map by applying the coil sensitivity map, the radio frequency transmission field map, which are parameters specified to perform the sequence used to acquire the k-space data together with the MR signal model and the k-space data Magnetic resonance parameter map reconstruction device. The method of claim 3, wherein the repetition determining unit
If the reconstructed reconstructed MR parameter map is determined to converge or the difference between the currently acquired reconstructed MR parameter map and the previous reconstructed MR parameter map is less than a predetermined reference value, the magnetic resonance parameter map reconstruction outputting the current reconstructed MR parameter map Device. The method of claim 5, wherein the parameter map reconstruction unit
Reconstructed MR parameter map (x _{k + 1)} the previous reconstructed MR parameter map (x _k) and the MR parameter map remove artifacts (D _R (x _k)) and the MR parameter map according to the number of iterations _(k) (x k _{) to the operator (A) and the difference (A(x k} )-b) between the undersampled k-space data and the k-space data obtained by applying the equation according to the iterative gradient descent method

(Where λ _k and μ _k are learnable regularization and increase/decrease magnitude variables according to the number of iterations (k).)
is calculated according to
A is an operator for transforming the MR parameter map into k-space data,

(where S is the MR signal model, C is the coil sensitivity map, F is the Fourier transform function, and M is the k-space sampling matrix)
is expressed as
J _A is a Jacobian matrix for the operator (A),

(where α _n is the nth flip angle, n is the number of MRI scans for parameter mapping, v _k ^m is _{the mth variable of x k} , which is the reference MR parameter map and the required parameter map, and m is the It is the number of parameter maps to be used.)
A magnetic resonance parameter map reconstruction device calculated according to 7. The method of claim 6, wherein the MR signal model is
If the requested MR parameter map is a T ₁ parameter map and the reference MR parameter map is an M ₀ parameter map, the equation

(Where TR is the repetition time.)
A magnetic resonance parameter map reconstruction device obtained according to The method of claim 3, wherein the initial parameter mapping unit
a parameter map conversion unit that receives the k-space data and converts it into the sparse MR contrast image by filling a region corresponding to the missing data in the k-space data with zeros according to a predetermined method; and
A magnetic resonance parameter map reconstruction apparatus comprising: an initial parameter map acquisition unit implemented as an artificial neural network and acquiring the initial MR parameter map according to a pattern estimation method previously learned from a sparse MR contrast image. The method of claim 8, wherein the initial parameter map acquisition unit
A magnetic resonance parameter map reconstruction apparatus for obtaining an initial MR parameter map including the reference MR parameter map and the desired MR parameter map. According to claim 6, The magnetic resonance parameter map reconstruction apparatus
Further comprising a learning unit for learning the initial parameter mapping unit and the parameter map reconstruction unit, wherein the learning unit

(N _S is the number of training data, l _L and l _S mean the weight of the structural similarity (SSIM) loss function with l1, respectively, and m _t is the reference MR parameter map that is the training data.)
A magnetic resonance parameter map reconstruction device that backpropagates by acquiring a loss according to The k-space data obtained by being undersampled according to a predetermined sequence is received and converted into a sparse MR contrast image in a predetermined manner, and the pattern of the sparse MR contrast image is estimated according to a pre-learned pattern estimation method to estimate the initial MR obtaining a parameter map;
The initial MR parameter map or the previously reconstructed reconstructed MR parameter map is received, artifacts are removed according to a pre-learned pattern estimation method, and the artifact-removed MR parameter map is combined with a predetermined reference MR parameter map and the required MR that requires reconstruction. Reconstructing the relationship between the parameter maps according to the modeled MR signal model to obtain a reconstructed MR parameter map, and repeating the reconstruction until the obtained reconstructed MR parameter map satisfies a predetermined condition. Reconstruction method. The method of claim 11, wherein repeating the reconstruction comprises:
Implemented as a pre-trained artificial neural network to receive an initial MR parameter map or a previously reconstructed reconstructed MR parameter map, and remove aliasing artifacts according to a pattern estimation method learned from the applied initial MR parameter map or previously reconstructed reconstructed MR parameter map to do;
receiving the de-artifacted MR parameter map, and reconstructing the MR parameter map so that a data relationship of the de-artifacted MR parameter map is maintained according to the MR signal model; and
When it is determined that the reconstructed reconstructed MR parameter map satisfies a predetermined condition and it is determined to be repeated, the magnetic resonance parameter map reconstruction method comprising the step of transferring the reconstructed reconstructed MR parameter map to the step of removing the artifact. 13. The method of claim 12, wherein removing the artifact comprises:
dividing the applied initial MR parameter map or the previously reconstructed reconstructed MR parameter map into the reference MR parameter map and at least one required required MR parameter map;
removing each artifact of the reference MR parameter map or the at least one required MR parameter map by using a plurality of neural networks in which a pattern estimation method has been previously learned; and
A method for reconstructing a magnetic resonance parameter map, comprising the step of combining and outputting an artifact-removed reference parameter map and at least one required MR parameter map. The method of claim 13, wherein reconstructing the MR parameter map comprises:
Reconstructing the MR parameter map by applying the coil sensitivity map, the radio frequency transmission field map, which are parameters specified to perform the sequence used to acquire the k-space data together with the MR signal model and the k-space data Magnetic resonance parameter map reconstruction method. 14. The method of claim 13, wherein the delivering comprises:
If the reconstructed reconstructed MR parameter map is determined to converge or the difference between the currently acquired reconstructed MR parameter map and the previous reconstructed MR parameter map is less than a predetermined reference value, the magnetic resonance parameter map reconstruction outputting the current reconstructed MR parameter map method. The method of claim 15, wherein reconstructing the MR parameter map comprises:
Reconstructed MR parameter map (x _{k + 1)} the previous reconstructed MR parameter map (x _k) and the MR parameter map remove artifacts (D _R (x _k)) and the MR parameter map according to the number of iterations _(k) (x k _{) to the operator (A) and the difference (A(x k} )-b) between the undersampled k-space data and the k-space data obtained by applying the equation according to the iterative gradient descent method

(Where λ _k and μ _k are learnable regularization and increase/decrease magnitude variables according to the number of iterations (k).)
is calculated according to
A is an operator for transforming the MR parameter map into k-space data,

(where S is the MR signal model, C is the coil sensitivity map, F is the Fourier transform function, and M is the k-space sampling matrix)
is expressed as
J _A is a Jacobian matrix for the operator (A),

(where α _n is the nth flip angle, n is the number of MRI scans for parameter mapping, v _k ^m is _{the mth variable of x k} , which is the reference MR parameter map and the required parameter map, and m is the It is the number of parameter maps to be used.)
The magnetic resonance parameter map reconstruction method calculated according to 17. The method of claim 16, wherein the MR signal model is
If the requested MR parameter map is a T ₁ parameter map and the reference MR parameter map is an M ₀ parameter map, the equation

(Where TR is the repetition time.)
A magnetic resonance parameter map reconstruction method obtained according to 14. The method of claim 13, wherein obtaining the initial MR parameter map comprises:
receiving the k-space data and converting it into the sparse MR contrast image by filling a region corresponding to the missing data in the k-space data with zeros according to a predetermined method; and
and acquiring the initial MR parameter map from a sparse MR contrast image using an artificial neural network in which a pattern estimation method has been previously learned. The method of claim 18, wherein the acquiring from the sparse MR contrast image comprises:
A magnetic resonance parameter map reconstruction method for obtaining an initial MR parameter map including the reference MR parameter map and the desired MR parameter map. The method of claim 16 , wherein the magnetic resonance parameter map reconstruction method comprises:
Further comprising a learning step, wherein the learning step is

(N _S is the number of training data, l _L and l _S mean the weight of the structural similarity (SSIM) loss function with l1, respectively, and m _t is the reference MR parameter map that is the training data.)
A magnetic resonance parameter map reconstruction method that backpropagates by acquiring losses according to

Images