WO2019148610A1

WO2019148610A1 - Multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency

Info

Publication number: WO2019148610A1
Application number: PCT/CN2018/080011
Authority: WO
Inventors: 朱高杰; 罗海; 吴子岳; 周翔; 刘霞; 王超; 陈梅泞
Original assignee: 奥泰医疗系统有限责任公司
Priority date: 2018-01-30
Filing date: 2018-03-22
Publication date: 2019-08-08
Also published as: CN108459289A; CN108459289B

Abstract

A multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency comprises: acquiring multi-channel pre-scan data; generating, according to the multi-channel pre-scan data, a convolution kernel based on data consistency; acquiring diffusion-weighted magnetic resonance imaging data of multiple excitations; calculating reconstructed data according to the convolution kernel and the diffusion-weighted magnetic resonance imaging data of each excitation; combining the reconstructed data to obtain a combined image; updating the combined image to obtain an updated image; checking whether iteration has reached a predetermined condition; and if the iteration reaches the predetermined condition, ending the iteration; or if the iteration has not reached the predetermined condition, performing phase recovery on the updated image. The method improves the efficiency of image sampling, and obtains a relatively stable reconstructed image.

Description

A multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency

Technical field

The present invention relates to the field of magnetic resonance imaging technologies, and in particular, to a multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency.

Background technique

Magnetic resonance imaging is a technique for imaging using nuclear magnetic resonance phenomena of hydrogen protons. The human body contains a single proton nucleus, such as a widely existing hydrogen nucleus, whose protons have a spin motion. The spin motion of charged nuclei is physically similar to individual small magnets, and the directional distribution of these small magnets is random under the influence of external conditions. When the human body is placed in an external magnetic field, these small magnets will be rearranged according to the magnetic field lines of the external magnetic field, specifically: in two directions parallel or anti-parallel to the magnetic field lines of the external magnetic field, the above parallel to the magnetic field lines of the external magnetic field The direction is called the positive longitudinal axis, and the direction parallel to the magnetic field lines of the external magnetic field is referred to as the negative longitudinal axis, and the nucleus has only the longitudinal magnetization component, which has both the direction and the amplitude.

The radio frequency (RF) pulse of a specific frequency is used to excite the nucleus in the external magnetic field, and the spin axis of these nucleuses is deviated from the positive longitudinal axis or the negative longitudinal axis to generate resonance, which is the magnetic resonance phenomenon. After the spin axis of the excited nuclei is deviated from the positive or negative longitudinal axis, the nucleus has a transverse magnetization component. After stopping the radio frequency pulse, the excited nuclei emit echo signals, and the absorbed energy is gradually released in the form of electromagnetic waves. The phase and energy level are restored to the state before the excitation, and the echo signals emitted by the nuclei are spatially encoded. The image can be reconstructed after further processing.

Magnetic resonance diffusion imaging is a new technology that relies on the motion of water molecules in the body to provide image contrast. The diffusion of water molecules in the tissue conforms to a random thermal motion model, and the magnitude and direction of the diffusion are affected by biomacromolecules in the biofilm and tissue. When a gradient magnetic field is present, the diffusion motion of the water molecules causes the phase loss of the magnetization vector, resulting in a decrease in the magnetic resonance signal. The extent to which magnetic resonance signals are reduced depends on the tissue type, structure, physical and physiological state, and microenvironment.

In the above process, the gradient magnetic field specifically used to influence the thermal motion of water molecules is called a diffusion sensitive gradient. Diffusion-sensitive gradients can significantly improve the sensitivity of various sequences to the random thermal motion of water molecules, which in turn helps to observe the diffusion properties of water molecules, but this gradient magnetic field is also very sensitive to other types of motion such as head movement. A single-shot diffusion imaging technique that acquires all data for imaging after a single signal excitation. This method can effectively shorten the scanning time and avoid introducing more macro motion to affect the image. However, the single-excitation scanning method uses a long echo chain, which is easy to cause magnetic-sensitive artifacts and geometric deformation; the data available for a single acquisition is limited, so the resolution of the image is low, which is not conducive to the diagnosis of fine structures.

In order to reduce image artifacts and geometric distortion, and to effectively improve image spatial resolution, a widely adopted strategy is to use multiple-excitation magnetic resonance diffusion imaging techniques. The main challenge of this technology is to effectively deal with the phase error caused by macroscopic motion between the data acquired after different excitations. According to different data collection methods, macro motion correction can be divided into two categories: the first type needs to collect navigation echo signals before normal data acquisition, and this signal will be used to correct the imaging data acquired by each subsequent excitation. The second type does not require the acquisition of a navigation echo signal, but corrects the phase of each other by exciting the relationship between the data each time. Compared with the method of collecting navigation echoes, the sampling method that does not require navigation echo has higher data collection efficiency, and can also avoid the problem of mismatch between navigation echo and actual imaging data.

In 2013, Nan-kuei Chen et al. proposed the MUSE (Multiplexed Sensitivity-encoding) technology. This technique uses SENSE parallel imaging technology to estimate the phase difference between different excitation data due to macroscopic motion, and on the other hand combines different excitation data to reconstruct the final image. This method results in higher image resolution, higher signal-to-noise ratio, and significantly reduced motion artifacts. Compared with the technique of using navigation echo, this method is more stable in clinical performance. In 2016, Hua Guo et al. used POCSENSE parallel imaging technology to replace the previous SENSE parallel imaging technology, and proposed POCS-ICE (POCS-Enhanced Inherent Correction of Motion-Induced Phase Errors for high resolution Multishot). Diffusion MRI). This technology has similar performance to MUSE.

It is noted that magnetic resonance parallel imaging technology has greatly assisted the multi-excitation diffusion magnetic resonance imaging. In 2016, Wentao Liu et al. proposed a multi-excitation diffusion imaging technique based on parallel imaging GRAPPA technology and using navigation echo. This method proposes the concept of virtual channel. It is considered that the data of multiple excitations received by each real channel can be regarded as undersampled data from multiple virtual channels, and these undersampled data can pass through one kind. The GRRAPA algorithm after K spatial rearrangement is reconstructed. At the same time, phase errors caused by macroscopic motion between different excitation data can be corrected by navigation echoes. This method does not require an explicit estimation of the phase difference compared to the SENSE-based method described above. However, this method relies on navigation echoes for motion correction, which reduces the efficiency of sampling.

Magnetic resonance parallel imaging technology, according to the data processed by the algorithm, can be divided into an image domain and a K-space domain. The image domain-based parallel imaging technique SENSE resolves the image curling artifacts due to undersampling according to the known spatial distribution of coil sensitivity, and restores the state without curling. This method is highly dependent on the sensitivity of the coil. For clinical applications, it is difficult to obtain higher coil sensitivity at lower signal-to-noise ratios, complex tissue structures, and the like.

In order to overcome the above various shortcomings of the existing magnetic resonance imaging methods, it is necessary to propose a new magnetic resonance imaging method based on this.

Summary of the invention

The present invention aims to provide a multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency, which does not rely on navigation echo data for motion correction, and does not depend on the coil sensitivity of the image domain, thereby improving image sampling efficiency. And get a more stable reconstructed image.

In order to achieve the above object, the technical solution adopted by the present invention is as follows:

A multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency, comprising:

Step 101: Collect multi-channel pre-scan data, where the multi-channel pre-scan data is fully sampled K-space data;

Step 102: Generate a convolution kernel based on data consistency according to the multi-channel pre-scan data.

Step 103: separately acquiring diffusion-weighted magnetic resonance imaging data of multiple excitations, and each of the excited diffusion-weighted magnetic resonance imaging data is under-sampled K-space data;

Step 104: Calculate reconstruction data according to the convolution kernel and the diffusion-weighted magnetic resonance imaging data of each excitation;

Step 105: Synthesize the reconstructed data to obtain a composite image.

Step 106: Update the composite image to obtain an updated image.

Step 107: Check if the iteration reaches a predetermined condition;

Step 108: If the iteration reaches the predetermined condition, the iteration is terminated; if the iteration does not reach the predetermined condition, the updated image is phase-recovered, and the updated multi-excited diffusion-weighted magnetic resonance imaging data is acquired, and the process returns to step 103.

Preferably, the method for generating a data consistency based convolution kernel according to the multi-channel pre-scan data is:

among them,

Corresponding to the grid for the multi-channel pre-scan data

a K-space data point at the location; R _r is an extraction operator; K _ij is a set of convolution kernels to be solved;

Rewrite equation (1) as a matrix: x=Gx

Wherein, the matrix x represents data points on all K-space grids, and the matrix G is the data consistency-based convolution kernel.

Preferably, the method for calculating reconstruction data according to the convolution kernel and the diffusion-weighted magnetic resonance imaging data of each excitation is: diffusing weighted magnetic resonance of the convolution kernel and the each excitation The imaging data is convolved to obtain the reconstructed data.

Preferably, the method for calculating the reconstructed data according to the convolution kernel and the diffusion-weighted magnetic resonance imaging data of each excitation is:

Where x represents the sampled data on the single-shot K-space grid, y represents the unsampled data on the single-shot K-space grid, I is the image data before sampling, and λ(ε) is used to control the image data before and after sampling. Consistency.

Where x represents the sampled data on the single-shot K-space grid, y represents the unsampled data on the single-shot K-space grid, I is the image data before sampling, and λ ₁ and λ _{2 are} used to control the image before and after sampling. The consistency of the data, the function R(x) represents the regularization term.

Preferably, the regularization term is an L1 regularization term, or an L2 regularization term.

Preferably, the L1 regularization term is: R(x)=||x|| ₂ , and the L2 regularization term is: R(x)=||ψ{IFFT(x)}|| ₁ , wherein , IFFT (x) is a discrete Fourier inverse transform function.

Preferably, the method for synthesizing the reconstructed data to obtain a composite image is:

Where I _k is the image corresponding to the Kth excitation, Hann(I _k ) represents the Hanning filtering of the image, and I _avg is the composite image.

Preferably, the method for updating the composite image to obtain an updated image is:

Where I _avg is the composite image,

Represents the composite image acquired during the nth iteration, and η is used to control the degree of update of the composite image.

Preferably, the method for checking whether the iteration reaches a predetermined condition is: detecting whether the iteration converges, or detecting whether the number of iterations reaches a predetermined upper limit;

The method for detecting whether the iteration converges is:

Where τ is a predetermined constant,

Represents the composite image acquired during the nth iteration.

The data consistency-based multiple-excitation diffusion-weighted magnetic resonance imaging method provided by the embodiment of the present invention has higher scanning efficiency and can be avoided in clinically because it does not rely on additional navigation echo data for motion correction. The problem of mismatch between the echo data and the imaging data provides a more stable image quality in the clinic. At the same time, the present invention is a parallel imaging method based on K-space data consistency. Unlike the image domain parallel imaging method SENSE which relies on image domain coil sensitivity, it can avoid image reconstruction errors caused by coil sensitivity estimation deviation, and provide more clinically. For a stable result. In addition, the data reconstruction method proposed in the embodiment of the present invention can conveniently and effectively integrate various known information, such as L2 regularization, which is beneficial to speed up the reconstruction convergence speed and improve the image quality.

DRAWINGS

1 is a flowchart of a method according to an embodiment of the present invention;

2 is a schematic diagram of a calculation process based on a pre-scan data convolution kernel in an embodiment of the present invention.

Detailed ways

In order to make the objects, technical solutions and advantages of the present invention more comprehensible, the present invention will be further described in detail with reference to the accompanying drawings.

Step 101: Acquire multi-channel pre-scan data, which is full-sampled K-space data. Data is received using a multi-channel receive coil. The acquired data can be derived from a variety of scan sequences, and it is recommended to use an echo plane sequence scan of the same type as diffusion weighted imaging. The size of the data generated by the scan can be expressed as: N _x * N _y * N _c . Where N _x represents the number of rows of data collected, N _y represents the number of columns of data, and N _c represents the number of receiving channels.

Step 102: Generate a data consistency-based convolution kernel according to the multi-channel pre-scan data. The self-calibration data comes from the center position of the multi-channel data collected in the above steps, and the data size can be expressed as: N _a *N _y *N _c . Wherein, N _a is the width calibration data from the default direction of the phase encoding direction along the row. Convolution kernels based on data consistency can be solved by the following equation:

among them,

Corresponding to the grid for the multi-channel pre-scan data

a K-space data point at the location; R _r is an extraction operator that extracts all data points around the target point;

Representing the K space grid

Extracted data points around, not included

This point; K _ij is a set of convolution kernels to be solved. The process of equation (1) can be represented by Figure 2. In Figure 2, the black data points represent the K-space locations that have been acquired, and the red data points represent the unoccupied K-space locations. The size of each set of convolution kernels is a four-dimensional array: W _x *W _y *N _c *N _c . In this step, since the pre-scan data is all sampled data, in equation (1)

And extraction operator

All of them are known, and the convolution kernel K _ij is an unknown quantity, and the convolution kernel can be calculated by solving the above linear equation.

Rewrite equation (1) as a matrix form, expressed as:

x=Gx formula (2)

Wherein, the matrix x represents data points on all K-space grids, and the matrix G is a convolution operator representing the corresponding position, that is, the data consistency-based convolution kernel. Equation (2) represents the self-correction process for convolution kernel calculations, ie for fully sampled data points, each data point can be reconstructed from the convolution kernel and its surrounding data points.

Step 103: separately acquire the diffusion-weighted magnetic resonance imaging data of the multiple excitations, and the diffusion-weighted magnetic resonance imaging data of each excitation is the under-sampled K-space data; at this time, the receiving coil is the same as the receiving coil used for the pre-scanning.

Step 104: Calculate the reconstruction data according to the convolution kernel and the diffusion-weighted magnetic resonance imaging data of each excitation.

Specifically, the convolution kernel is convolved with the diffusion-weighted magnetic resonance imaging data of each excitation to acquire the reconstruction data. The physical meaning of this convolution is to reconstruct the undersampled K-space data points through the convolution kernel. This convolution process can be expressed as:

y=Gx formula (3)

Where x represents the sampled data on the single-shot K-space grid, y represents the unsampled data on the single-shot K-space grid, and G is the convolution kernel calculated in step 102.

Equation (2) and formula (3) describe the data consistency of the convolution kernel self-correction phase and the data under-sampling phase, respectively. In order to avoid data noise and control correction errors, the above problem can be transformed into an optimization problem:

Equation (4) transforms the reconstruction process of undersampled data into an optimization problem. Therefore, it is convenient to set constraints on the optimization problem based on known knowledge and turn the problem into:

Where x represents the sampled data on the single-shot K-space grid, y represents the unsampled data on the single-shot K-space grid, I is the image data before sampling, and λ ₁ and λ _{2 are} used to control the image before and after sampling. The consistency of the data, the function R(x) represents the regularization term of known information.

The regularization term can be in the image domain or in the K space domain. The regularization term is an L1 regularization term, or an L2 regularization term. Typical regularization items include:

L1 regularization term: R(x)=||x|| ₂ formula (5-1)

L2 regularization term: R(x)=||ψ{IFFT(x)}|| ₁ formula (5-2)

Where IFFT(x) is a discrete Fourier inverse transform function.

Step 105: Synthesize the reconstructed data to obtain a composite image, and the process of synthesizing may be described by the following formula:

Where I _k is the image corresponding to the Kth excitation, Hann(I _k ) represents the Hanning filtering of the image, and I _avg is the composite image. Equation (6-1) indicates that the low-pass phase corresponding to the single-shot data is calculated, and Equation (6-2) synthesizes the data of the multiple-excitation by the low-pass phase.

Step 106: Update the composite image in the current iteration process to obtain an updated image. Specifically, the following formula may be used:

Wherein, I _avg represents the synthesized multiple excitation data calculated by the formula (6), that is, the composite image,

Represents the composite image acquired during the nth iteration, and η is used to control the degree of update of the data before and after, that is, to control the degree of update of the composite image.

Step 107: Check whether the iteration reaches a predetermined condition, specifically, whether the iteration is converged, or whether the number of iterations reaches a predetermined upper limit. The upper limit of the number of iterations comes from a fixed value defined in advance. To detect whether the iteration is convergent, the following formula can be used:

Where τ is a predefined iteration constraint, which is a predetermined constant,

Represents the composite image acquired during the nth iteration.

Step 108: If the iteration reaches a predetermined condition, that is, if the difference between two consecutive iterations is less than τ, the iteration is terminated; if the iteration does not reach the predetermined condition, phase recovery is performed on the updated image, and the diffusion weight of the updated multiple excitation is obtained. The magnetic resonance imaging data is returned to step 103. The process of phase recovery can be described by the following formula:

That is, the low frequency phase information corresponding to each excitation data is reconfigured to the updated image data to obtain updated data for each excitation. Pass this data to the next iteration cycle.

The data consistency-based multiple-excitation diffusion-weighted magnetic resonance imaging method provided by the embodiment of the invention directly collects multiple excitation data, and then synthesizes the data repeatedly excited by data consistency, and corrects the motion between different excitation data. The resulting phase error results in a high resolution composite image. The important clinical significance of this method is: a) no additional navigation echo data, so higher acquisition efficiency, shorter scan time; b) no additional navigation echo data, can avoid navigation echo data and The imaging data is mismatched and the imaging results are more robust. The present invention obtains a convolution kernel by pre-scanning data, and then uses the convolution kernel to generate undersampled data in the data synthesis stage. The benefits of this implementation include: a) avoiding the use of coil sensitivity calculations in the image domain to prevent reconstruction errors due to incorrect estimation of coil sensitivity; b) the method will utilize motion correction between data generation steps of convolution kernels and different excitation data The steps are integrated into one, and the two calculations can be processed simultaneously in the same process, which reduces the amount of calculation; c) the method can conveniently add the limitation of existing knowledge to the reconstruction process, which is beneficial to speed up the reconstruction convergence and provide more stability. Reconstructed image.

The above is only a specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily think of changes or substitutions within the technical scope of the present invention. It should be covered by the scope of the present invention.

Claims

A multi-excitation diffusion-weighted magnetic resonance imaging method based on data consistency, characterized in that it comprises:

Step 101: Collect multi-channel pre-scan data, where the multi-channel pre-scan data is fully sampled K-space data;

Step 102: Generate a convolution kernel based on data consistency according to the multi-channel pre-scan data.

Step 103: separately acquiring diffusion-weighted magnetic resonance imaging data of multiple excitations, and each of the excited diffusion-weighted magnetic resonance imaging data is under-sampled K-space data;

Step 104: Calculate reconstruction data according to the convolution kernel and the diffusion-weighted magnetic resonance imaging data of each excitation;

Step 105: Synthesize the reconstructed data to obtain a composite image.

Step 106: Update the composite image to obtain an updated image.

Step 107: Check if the iteration reaches a predetermined condition;

Step 108: If the iteration reaches the predetermined condition, the iteration is terminated; if the iteration does not reach the predetermined condition, the updated image is phase-recovered, and the updated multi-excited diffusion-weighted magnetic resonance imaging data is acquired, and the process returns to step 103.
The data consistency-based multiple-excitation diffusion-weighted magnetic resonance imaging method according to claim 1, wherein the method for generating a data consistency-based convolution kernel according to the multi-channel pre-scan data is:

among them,
Corresponding to the grid for the multi-channel pre-scan data
a K-space data point at the location; R r is an extraction operator; K ij is a set of convolution kernels to be solved;

Rewrite equation (1) as a matrix: x=Gx

Wherein, the matrix x represents data points on all K-space grids, and the matrix G is the data consistency-based convolution kernel.
The data consistency-based multiple excitation diffusion weighted magnetic resonance imaging method according to claim 2, wherein said calculating is based on said convolution kernel and said diffusion-weighted magnetic resonance imaging data for each excitation The method of reconstructing data is:

The convolution kernel is convolved with the diffusion-weighted magnetic resonance imaging data of each excitation to obtain the reconstruction data.
The data consistency-based multiple excitation diffusion weighted magnetic resonance imaging method according to claim 2, wherein said calculating is based on said convolution kernel and said diffusion-weighted magnetic resonance imaging data for each excitation The method of reconstructing data is:

Where x represents the sampled data on the single-shot K-space grid, y represents the unsampled data on the single-shot K-space grid, I is the image data before sampling, and λ(ε) is used to control the image data before and after sampling. Consistency.
The data consistency-based multiple excitation diffusion weighted magnetic resonance imaging method according to claim 2, wherein said calculating is based on said convolution kernel and said diffusion-weighted magnetic resonance imaging data for each excitation The method of reconstructing data is:

Where x represents the sampled data on the single-shot K-space grid, y represents the unsampled data on the single-shot K-space grid, I is the image data before sampling, and λ 1 and λ 2 are used to control the image before and after sampling. The consistency of the data, the function R(x) represents the regularization term.
The data consistency-based multiple excitation diffusion weighted magnetic resonance imaging method according to claim 5, wherein the regularization term is an L1 regularization term, or an L2 regularization term.
The data consistency based multiple excitation diffusion weighted magnetic resonance imaging method according to claim 6, wherein

The L1 regularization term is: R(x)=||x|| 2

The L2 regularization term is: R(x)=||ψ{IFFT(x)}|| 1

Where IFFT(x) is a discrete Fourier inverse transform function.
The data consistency-based multiple-excitation diffusion-weighted magnetic resonance imaging method according to claim 2, wherein the method of synthesizing the reconstructed data to obtain a composite image is:

Where I k is the image corresponding to the Kth excitation, Hann(I k ) represents the Hanning filtering of the image, and I avg is the composite image.
The data consistency-based multiple-excitation diffusion-weighted magnetic resonance imaging method according to claim 8, wherein the method of updating the composite image to obtain an updated image is:

Where I avg is the composite image,
Represents the composite image acquired during the nth iteration, and η is used to control the degree of update of the composite image.
The data consistency-based multiple-shot diffusion-weighted magnetic resonance imaging method according to claim 9, wherein the method of checking whether the iteration reaches a predetermined condition is: detecting whether the iteration converges, or detecting whether the number of iterations reaches The upper limit of the reservation;

The method for detecting whether the iteration converges is:

Where τ is a predetermined constant,
Represents the composite image acquired during the nth iteration.