An adaptive noise reduction method for chemical exchange saturation transfer magnetic resonance imaging sequences

Through the combination of singular value decomposition and data-driven truncation criteria, the problem of noise interference in CEST-MRI technology is solved, high-quality CEST image noise reduction is achieved, and the signal-to-noise ratio and imaging quality are significantly improved.

CN118134795BActive Publication Date: 2024-11-22厦门医学院附属第二医院
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410089563.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-23
Publication Date
2024-11-22
Estimated Expiration
2044-01-23

AI Technical Summary

Technical Problem

CEST-MRI technology has failed to effectively utilize the correlation between imaging sequences and data redundancy for denoising due to noise interference, affecting quantitative analysis and imaging quality.

Method used

Singular value decomposition combined with appropriate data-driven truncation criterion is used to separate meaningful z-spectral features, remove noise through matrix low-rank approximate decomposition, and build high-quality CEST images.

Benefits of technology

It significantly improves the signal-to-noise ratio of CEST images, improves imaging quality, simplifies the denoising process, and can achieve noise reduction by only inputting the Z spectrum of each image voxel, which performs better than the classic single image noise reduction technology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118134795B_ABST
    Figure CN118134795B_ABST
Patent Text Reader

Abstract

The present invention relates to a chemical exchange saturation transfer imaging (CEST-MRI) imaging technology, provides a chemical exchange saturation transfer magnetic resonance imaging sequence adaptive denoising method, constructs a matrix low-rank approximation CEST image denoising technology, decomposes a matrix into the product of two low-rank matrices, that is, C=UV, where U is an m×r matrix, V is an r×n matrix, and r is the rank of the matrix. This decomposition method can be implemented by singular value decomposition (SVD), that is, C=UΣV T , where Σ is an r×r diagonal matrix, and the elements on the diagonal are singular values. The present invention reconstructs high-quality CEST images based on the singular value decomposition method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a nuclear magnetic resonance image processing method, in particular to noise reduction imaging. Background Art

[0002] Chemical exchange saturation transfer imaging (CEST-MRI) is a sensitive imaging technique used to detect compounds containing exchangeable protons. Such exchangeable protons can be selectively saturated by radiofrequency pulses, and the saturation is then transferred to the bulk water signal via proton chemical exchange, resulting in a significant sensitivity improvement. CEST-MRI has been shown to quantify low concentration metabolites such as peptides and proteins, creatine and glutamate, and injected solutes such as iopamide, glucose, and glucose derivatives.

[0003] Since the CEST-MRI effect is essentially small and is also affected by noise, the quantitative analysis and imaging of the CEST-MRI effect are affected. In order to improve the CEST-MRI signal-to-noise ratio, the imaging parameters and the saturation pulse parameters that affect the size of the CEST-MRI effect are generally optimized. Nevertheless, these CEST-MRI effects are affected by noise, whether in clinical field strength or ultra-high field.

[0004] CEST imaging technology usually uses radio frequency pulses to saturate water protons at different chemical shifts to obtain a CEST-MRI sequence. Extracting the value of each voxel on this sequence image can obtain a CEST z-spectrum. In this way, hundreds or thousands of z-spectra can be generated in one experiment. These z-spectra can be arranged in order to construct a z-spectrum matrix, which contains a large amount of redundant data. A common method for denoising using data redundancy is to use a low-rank approximation of structured signals. In signal processing, matrix low-rank approximation can be used for signal denoising, decomposing a high-dimensional signal matrix into two low-dimensional matrices, thereby removing noise and interference in the signal. The low-rank approximation of the Z-spectrum has been used to improve the signal-to-noise ratio of CEST images and iterative motion correction of skeletal muscle creatine CEST spectra.

[0005] Usually, CEST imaging noise reduction adopts a single image noise reduction method without considering the correlation between CEST imaging sequences and image data redundancy. Summary of the invention

[0006] The denoising algorithm of the present invention combines the identification of z-spectral redundancy by singular value decomposition with an appropriate data-driven truncation criterion to separate meaningful z-spectral features from noise features. The latter is crucial for applications with abundant in vivo data, where the number of necessary components cannot be determined a priori. The given truncation criterion facilitates a direct assessment of data quality, providing guidance for CEST-MRI signal-to-noise ratios. The applicability of the algorithm is not limited to the studied low-power z-spectra, but includes any CEST-related method, providing an additional reference dimension for exploiting data redundancy.

[0007] The present invention constructs a matrix low-rank approximation CEST image denoising technology. The basic idea of ​​matrix low-rank approximation is to decompose a matrix into the product of two low-rank matrices, that is, C=UV, where U is an m×r matrix and V is an r×n matrix, and r is the rank of the matrix. This decomposition method can be implemented by singular value decomposition (SVD), that is, C=UΣV T , where Σ is an r×r diagonal matrix with singular values ​​on the diagonal.

[0008] (1) Constructing z-spectrum matrix

[0009] In the CEST experiment, each saturation frequency offset is used as a variable, and the z value corresponding to each voxel is a different observed value. A single CEST image z(Δω) with spatial dimensions u×v×n i ) can be rearranged into a vector z(Δω) with a length of m≤u×v×n i ). Here m is the number of voxels actually considered, which may be less than the total number of voxels. A complete dataset {Δω1, Δω2, …, Δω n}, and then reshaped into a z-spectral matrix C of size m×n.

[0010]

[0011] (2) Singular value decomposition of z-spectral matrix C

[0012] Singular value decomposition is a nonlinear filtering noise reduction method. Since the singular value of noise is smaller than the singular value of useful information, the purpose of noise reduction can be achieved by performing singular value decomposition on the z-spectrum matrix C, removing smaller singular values, and reorganizing the image matrix. Suppose the z-spectrum matrix is ​​an m×n-order matrix C with a rank of r. According to the principle of singular value decomposition, the decomposition form of the matrix C can be expressed as:

[0013]

[0014] Among them, U and V represent m×n and n×n orthogonal matrices, Σ r×rIt is a diagonal matrix with non-zero singular values ​​in descending order, that is, λ1≥λ2…≥λ r , 0 represents the zero matrix. Singular values ​​of zero do not provide any information when reconstructing the matrix, so some very small or even close to zero singular values ​​are also ignored. In this way, an approximate representation of the z-spectral matrix C can be obtained.

[0015]

[0016] Among them, Σ s×s is Σ r×r Subdiagonal matrix.

[0017] (3) Truncation criteria for singular values

[0018] The adaptive noise reduction method of the invention is to determine the diagonal matrix Σ by the variation characteristics of non-zero singular values. s×s The number of non-zero singular values ​​in the image. The image denoising result mainly depends on the number s of non-zero singular values ​​retained. If too few components are retained, the actual information will be lost, resulting in missing or deformed resonance, while too many components will reduce the denoising effect. Therefore, an s close to the optimal value is crucial for the performance of denoising. The truncation criterion for retaining the number of non-zero singular values ​​in the invention is to calculate the second-order difference of the logarithm of the non-zero singular value and find the number s of non-zero singular values ​​corresponding to the maximum second-order difference, and the specific calculation is as follows.

[0019] Δ(Δlog(λ k ))=log(λ k+2 )-2log(λ k+1 )+log(λ k )

[0020]

[0021] The present invention reconstructs high-quality CEST images based on the singular value decomposition method. Without any further processing of the CEST image, the singular value decomposition method performs significantly better than the classic single image denoising technique. The singular value decomposition method proposed in the present invention can achieve CEST image denoising by only inputting the Z spectrum of each image voxel. Compared with the classic single image denoising technique, the proposed technique solves the task of CEST image denoising with only a simpler procedure. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The figure shows the process of constructing the z-spectrum matrix C. The region of interest of the image has 360 pixels in total, each pixel corresponds to a z-spectrum of length 101, and the pixel vectors sequentially construct a z-spectrum matrix C of size 360×101.

[0023] Figure 2, shows the logarithmic image of non-zero singular values, and number represents the ordinal number of non-zero singular values ​​from large to small.

[0024] Figure 3 It shows that the singular value decomposition method significantly reduces the noise of the z-spectrum. The continuous curve represents the z-spectrum after denoising, and the volatility is significantly reduced.

[0025] Figure 4 It shows that the peak signal-to-noise ratio of the CEST image sequence after noise reduction is higher than that of the CEST image sequence without noise reduction. The horizontal axis represents the RF pulse bias frequency, and the vertical axis represents the peak signal-to-noise ratio of the CEST image obtained corresponding to each RF pulse bias frequency. The curve at the top is the peak signal-to-noise ratio after denoising, and all the values ​​are greater than the signal-to-noise ratio curve without denoising. DETAILED DESCRIPTION

[0026] (1) In vivo CEST image sequence acquisition

[0027] In vivo CEST image sequences were acquired on a 7.0T small animal MRI scanner (Agilent Technologies, Santa Clara, CA, USA). Eight-week-old male Sprague-Dawley rats weighing 250 g were used to make the animal model. In the experimental study, rats were anesthetized with isoflurane mixed with O2 at a rate of 1 L / min; 4.0% isoflurane was used for induction of anesthesia, and 2.0-3.0% isoflurane was used for maintenance. The respiratory rate and body temperature during the 7T scan were 60–70 times / min and 38.5–39.5°C, respectively. The B0 map was shimmed before the experiment to eliminate the signal interference of B0 field inhomogeneity. The parameters were as follows: repetition time (TR) = 6000 ms, echo time (TE) = 40 ms, array = frequency offset, slice thickness = 2 mm, field of view (FOV) = 64 × 64 mm, matrix size = 64 × 64, spatial resolution = 1 × 1 mm, average = 1. An echo planar imaging (EPI) readout sequence was used to obtain CEST images, where continuous wave (CW) RF irradiation was implemented on the scanner. Some parameters of the EPI sequence were: saturation power = 1.2 μt, RF pulse duration = 5000 ms, pulse shape = rectangular RF pulse, number of shots = single shot. The uniformly densely sampled CEST images had 101 frequency offsets, which were uniformly distributed between -6 and 6 ppm. A total of 101 CEST images with saturation pulses and 101 images without saturation pulses were collected.

[0028] (2) Constructing the z-spectral matrix of the CEST image sequence

[0029] First, the acquired saturation pulse CEST image is standardized. Secondly, the region of interest is delineated, and for each voxel in the region of interest, a one-dimensional z(Δω) with a length of 101 is obtained. i ) vector. Here, the region of interest has 360 voxels in total, so we obtain m = 360 one-dimensional z(Δω) with a length of 101 i ) vectors. These vectors are sequentially constructed into a z-spectrum matrix C of size 360×101. Figure 1 shown.

[0030] (3) CEST image sequence denoising

[0031] Perform singular value decomposition on the z-spectrum matrix C obtained in step (2) to obtain matrices U, V and eigenvalues. According to the truncation criterion of non-zero singular values, such as Figure 2 As shown, the first 7 largest eigenvalues ​​are selected. In this way, the z-spectral matrix of the denoised CEST image sequence can be obtained:

[0032]

[0033] By calculating the matrix U, V and eigenvalue, the matrix is ​​calculated. Arbitrarily choose the matrix C and The same row vector, such as Figure 3 As shown, the singular value decomposition method obviously reduces the noise of the z-spectrum.

[0034] According to step (2) to obtain the z-spectral matrix C, for the matrix The noise-reduced CEST image sequence is obtained in reverse. The peak signal-to-noise ratio is calculated as follows: Figure 4 As shown in Figure 3, the peak signal-to-noise ratio of the CEST image sequence after denoising is significantly higher than that of the CEST image sequence without denoising.

[0035] The present invention also provides a chemical exchange saturation transfer magnetic resonance imaging sequence adaptive noise reduction device, and the device is provided with a module of the above-mentioned operating program.

[0036] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited to this. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A method for adaptive noise reduction of chemical exchange saturation transfer magnetic resonance imaging sequence, characterized in that: Decompose a matrix into the product of two low-rank matrices, that is, C = UV, where U is an m×r matrix and V is an r×n matrix, r is the rank of the matrix; this is achieved through singular value decomposition (SVD), that is, C = UΣV T , where Σ is an r×r diagonal matrix with singular values ​​on the diagonal; The specific steps include: (1) Constructing z-spectrum matrix In the CEST experiment, each saturation frequency offset is used as a variable, and the z value corresponding to each voxel is a different observed value; a single CEST image z(Δω i ) is rearranged into a vector z(Δω) with a length of m≤u×v×n i ), m is the number of voxels actually considered; a complete dataset {Δω1, Δω2, …, Δω n }, reshaped into a z-spectral matrix C of size m×n; (2) Singular value decomposition of z-spectral matrix C Perform singular value decomposition on the z-spectrum matrix C, remove smaller singular values, and reconstruct the image matrix; suppose the z-spectrum matrix is ​​an m×n-order matrix C with a rank of r. According to the singular value decomposition principle, the decomposition form of the matrix C can be expressed as: Among them, U and V represent m×n and n×n orthogonal matrices, Σ r×r It is a diagonal matrix with non-zero singular values ​​in descending order, i.e., λ1≥λ2...≥λ r , 0 represents the zero matrix; where Σ s×s is Σ r×r Subdiagonal matrix; (3) Truncation criteria for singular values The truncation criterion for retaining the number of non-zero singular values ​​is to calculate the second-order difference of the logarithm of the non-zero singular values ​​and find the number s of non-zero singular values ​​corresponding to the maximum second-order difference. The specific calculation is as follows: Δ(Δlog(λ k ))=log(λ k+2 )-2log(λ k+1 )+log(λ k ) 2. The method for adaptive noise reduction of chemical exchange saturation transfer magnetic resonance imaging sequence according to claim 1, characterized in that An animal model was prepared with the following parameters: repetition time (TR) = 6000 ms, echo time (TE) = 40 ms, array = frequency offset, slice thickness = 2 mm, field of view (FOV) = 64×64 mm, matrix size = 64×64, spatial resolution = 1×1 mm, average = 1; an echo planar imaging (EPI) readout sequence was used to obtain CEST images, wherein continuous wave (CW) RF irradiation was implemented on the scanner; some parameters of the echo planar imaging sequence were: saturation power = 1.2 μt, RF pulse duration = 5000 ms, pulse shape = rectangular RF pulse, number of transmissions = single transmission; the uniformly densely sampled CEST images had 101 frequency offsets, and the frequency offsets were uniformly distributed between -6 and 6 ppm, and a total of 101 CEST images with saturation pulses and 101 images without saturation pulses were collected.

3. The method for adaptive noise reduction of chemical exchange saturation transfer magnetic resonance imaging sequence according to claim 2, characterized in that The acquired saturation pulse CEST image is normalized. Secondly, the region of interest is delineated, and for each voxel in the region of interest, a one-dimensional z(Δω) with a length of 101 is obtained. i ) vector; obtain m = 360 one-dimensional z(Δω) with a length of 101 i ) vectors, which are sequentially constructed into a z-spectrum matrix C of size 360×101.

4. The method for adaptive noise reduction of chemical exchange saturation transfer magnetic resonance imaging sequence according to claim 3, characterized in that In the singular value decomposition of the z-spectrum matrix C in step (2), a singular value of zero does not have any information when reconstructing the matrix. The approximate representation of the z-spectrum matrix C is: By calculating the matrix U, V and eigenvalue, the matrix is ​​calculated. Arbitrarily choose the matrix C and Same row vector.