A fast MRI reconstruction method based on multi-scale pixel deconstruction network

Abstract

A fast MRI reconstruction method based on a multi-scale pixel anti-reconstruction network includes the following steps: Step 1 Input undersampled MRI image x u ; Step 2: The reconstructed MRI image #imgabs1# is obtained by the multi-scale pixel anti-reconstruction network #imgabs0# Figure 1 shows the multi-scale pixel anti-reconstruction network f MS‑PU (·; Θ) structure diagram. The present invention fully utilizes the information redundancy in undersampled MRI images and compresses the undersampled MRI images at different scales based on pixel decomposition. This can effectively reduce the impact of a large number of artifacts in the input undersampled images on the reconstruction effect, and can effectively improve the reconstruction speed and accuracy of MRI images.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, in particular to the field of medical image technology of magnetic resonance imaging, and relates to an MRI reconstruction method, which is mainly used to realize accelerated reconstruction of medical magnetic resonance imaging. Background Art

[0002] Magnetic resonance imaging (MRI) is an imaging technique widely used in clinical diagnosis and treatment. It has the advantages of being non-invasive, high-resolution, safe for living organisms, and having excellent soft tissue contrast. MRI requires sequential acquisition of data samples in the Fourier transform domain (k-space), and the acquisition process is relatively slow. This leads to the following three problems: (1) it results in high examination costs; (2) it greatly affects the comfort and compliance of patients, resulting in local artifacts in MRI images due to involuntary movements of patients; and (3) it is difficult to apply to the screening of diseases that require a long diagnostic time, such as stroke. Since the emergence of MRI in the 1970s, improving imaging speed has been a major ongoing research goal and has also been one of the research hotspots in recent years.

[0003] A common approach to achieving fast MRI is to first undersample the magnetic resonance data at a high rate, then reconstruct it using image processing and other techniques to restore high-quality images. Traditional MRI reconstruction techniques are primarily based on compressed sensing (CS). This reconstruction method has a lengthy iterative process and high computational complexity, resulting in low MRI image reconstruction efficiency.

[0004] With the rapid development of deep learning in natural image processing, the research and application of deep neural networks (DNNs) in medical image processing have also attracted widespread attention. Compared with traditional compressed sensing methods, DNN-based MRI reconstruction methods have higher accuracy and faster speed. However, DNN models typically require training based on large amounts of data, and cutting-edge DNN models are often large in parameter size, often requiring training for several days or longer. Moreover, the reconstruction accuracy of existing DNN models needs to be further improved. Summary of the Invention

[0005] In order to overcome the shortcomings of existing MRI reconstruction methods such as long training cycle and low reconstruction accuracy, the present invention provides a fast MRI reconstruction method based on a multi-scale pixel unshuffle network. This method is based on a deep cascaded convolutional neural network (DC-CNN) model and performs pixel unshuffle (PU) compression at different scales on the undersampled MRI images input to each subnetwork of the DC-CNN, so as to effectively overcome the problem of slow training and prediction speed of the DC-CNN model.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A fast MRI reconstruction method based on a multi-scale pixel anti-reconstruction network comprises the following steps:

[0008] Step 1 Input undersampled MRI image here, represents the complex field, x u Obtained through the following methods:

[0009] x u =F -1 y u =F -1 (M⊙y)=F -1 (M⊙F(x)),

[0010] Among them, y u =M⊙y, represents the fully sampled MRI image to be reconstructed, y = Fx represents the fully sampled data in k-space, F(·) and F -1 (·) represent the two-dimensional Fourier transform matrix and the two-dimensional Fourier transform inverse matrix, respectively. F(·) is used to transform the MRI image from pixel space to k space, and F -1 (·) is used to transform MRI images from k-space to pixel space, M∈{0,1} MxN represents the mask matrix for undersampling y;

[0011] Step 2 is performed by the multi-scale pixel decomposition network f MS-PU (·; Θ): Get the reconstructed MRI image

[0012] Furthermore, in step 2, the multi-scale pixel decomposition network f MS-PU (·; Θ) is formed by cascading Γ PU subnets:

[0013]

[0014] Among them, Θ={Θ1, Θ2,…, Θ γ ,…,Θ Γ} are the parameters of the network, s={s1,s2,…,s γ ,…,s Γ} is the scale of each PU used in the network, 1≤γ≤Γ, The PU scale is s γ PU subnet.

[0015] Furthermore, the PU subnet By PU operator, D γ It is composed of three parts: a reconstruction subnet and a pixel shuffle (PS) operator.

[0016]

[0017] Here, f PU (·,s γ ): Indicates that the PU scale is s γ The PU operator is used to de-reorganize the pixels of the input single-channel image, so that its dimension changes from M×N to and Set M and N to be s γ divisible; Represents the PS operator, which is the inverse operation of the PU operator and is used to convert the input The low-dimensional image of the channels is pixel-wise reorganized to change its dimension from becomes M×N; for The parameter set of the network, d∈{1, 2, ..., D γ}, x (0) =x u ;f γ，d (·;θ γ，d )for The d-th reconstruction subnetwork is defined as follows:

[0018]

[0019] Among them, R γ，d (·;θ d ): is a residual learning module used to learn the lost details in the input MRI image, θ γ，d Its parameter; f DC (·) represents the data consistency operator, let and Respectively represent f DCThe input and output of (·), Calculated as follows:

[0020]

[0021] here, Calculated as follows:

[0022]

[0023] in, Z[i, j] represents the element in the i-th row and j-th column of the matrix Z.

[0024] The technical concept of the present invention is: the existing DNN-based fast MRI reconstruction method requires feature extraction of the entire input MRI image, and the computational efficiency is low. The present invention notes that the input undersampled MRI image is highly redundant, and the presence of this redundant information will not only slow down the network's calculation speed, but also reduce the network's prediction accuracy. In order to reduce the redundancy of the input image, the present invention compresses the input image based on pixel unshuffle (PU), extracts and reconstructs the image compressed by the PU operation, and then uses pixel shuffle (PS) operation to restore the dimension of the input image; and the present invention points out that embedding multi-scale PU and PS operations into a deep cascade convolutional neural network (DC-CNN) model for MRI reconstruction can achieve better reconstruction effects.

[0025] The beneficial effects of the present invention are mainly manifested in: (1) it can accelerate the reconstruction speed of DC-CNN; (2) since the influence of redundant information in the input image is fully removed, a better reconstruction effect can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is the structure diagram of the multi-scale pixel decomposition network proposed in this invention, where x u is the input undersampled MRI image, is the output reconstructed MRI image, The PU scale is s γ PU subnet, x (γ) for The output of f PU (·,s γ ) indicates that the PU scale is s γ Pixel decomposition (PU) operator, f PS (·,s γ ) indicates that the PU scale is s γ Pixel reassembly (PS) operator; f γ，d(·;θ γ，d )for The d-th reconstructed subnetwork, f γ，d (·;θ γ，d ) output, R γ，d (·;θ d ) is the residual learning module, f DC It is the data consistency operator.

[0027] Figure 2 is an example of a residual learning module, consisting of n convolutional layers, where Conv γ，d，j Represents R γ，d (·;θ d )’s j-th convolutional layer.

[0028] Figure 3 are examples of MRI images, where (a) and (b) are an 8-fold undersampled MRI image and a fully sampled MRI image in k-space, respectively, and (c) and (d) are a reconstructed MRI image and a fully sampled MRI image in pixel space, respectively.

[0029] Figure 4 are examples of mask matrices, where (a) is a center undersampling mask matrix and (b) is a Gaussian undersampling mask matrix.

[0030] Figure 5 yes Figure 1 The calculation process examples of the PU operator and PS operator used in the figure are as follows: (a) to (b) to (c) is the calculation process of the PU operator; the PS operator is the inverse operation of the PU operator, and (c) to (b) to (a) is the calculation process of the PS operator.

[0031] Figure 6 The quantization results of the proposed method at different PU scales are compared, where r represents the speedup ratio, r = 8 or r = 16 represents a speedup ratio of 8 or 16; the two numbers separated by “ / ” represent the peak signal-to-noise ratio (PSNR) and structural similarity metric (SSIM) of the reconstructed image respectively; “-” is used to separate the numbers of different s ×n cascaded to represent the specific configuration of the multi-scale PU network proposed in this invention, s ×n represents the configuration of a PU subnet, where the underlined numbers s represents the PU scale of the current PU subnet, and ×n means that the current PU subnet uses n reconstruction subnets.

[0032] Figure 7This is a quantitative comparison of the reconstruction results of the method proposed in this invention and the existing methods UNet, DeepCascade and DuDoRNet, as well as a comparison of model training time.

[0033] Figure 8 This is a visual comparison of the reconstruction results for the acceleration ratio r = 16. (a) is the input undersampled MRI image, (b) is the fully sampled MRI image, (c), (d), (e), and (f) are the reconstructed images of UNet, DeepCascade, DuDoRNet, and the method proposed in this invention, respectively. DETAILED DESCRIPTION

[0034] The present invention will be further described below with reference to the accompanying drawings.

[0035] Reference Figures 1 to 8 A fast MRI method based on a multi-scale pixel anti-reconstruction network (Deep Laplacian network, DeepLap) includes the following steps:

[0036] Step 1 Input undersampled MRI image

[0037] Step 2 is performed by the multi-scale pixel decomposition network f MS-PU (·; Θ): Get the reconstructed MRI image Figure 1 A multi-scale pixel deconstruction network f is given MS-PU (·; Θ) network structure diagram, Figure 2 Given f MS-PU An example of the residual learning module used in (·; Θ).

[0038] In the step 1, represents the complex field, x u Obtained through the following methods:

[0039] x u =F -1 y u =F -1 (M⊙y)=F -1 (M⊙F(x)),

[0040] Among them, y u , =M⊙y, represents the fully sampled MRI image to be reconstructed, y = Fx represents the fully sampled data in k-space, F(·) and F -1 (·) represent the two-dimensional Fourier transform matrix and the two-dimensional Fourier transform inverse matrix, respectively. F(·) is used to transform the MRI image from pixel space to k space, and F -1(·) is used to transform MRI images from k-space to pixel space, M∈{0,1} M×N represents the mask matrix for undersampling y; Figure 3 (a) shows y undersampled by 8 times u For example, Figure 3 (b) and (d) give examples of fully sampled MRI images y and x in k-space and pixel space, respectively; the sampling mask matrix M can use a central undersampling matrix or a Gaussian undersampling matrix, as Figure 4 As shown, Figure 3 (a) Yes Figure 3 (b) K-space MRI image obtained by central undersampling.

[0041] In step 2, the multi-scale pixel anti-reconstruction network f MS-PU (·; Θ) is formed by cascading Γ PU subnets:

[0042]

[0043] Among them, Θ={Θ1, Θ2,..., Θ γ ,…,Θ Γ} are the parameters of the network, s={s1,s2,…,s γ ,…,s Γ} is the PU scale used by the network, 1≤γ≤T, The PU scale is s γ PU subnet.

[0044] PU subnet By PU operator, D γ It is composed of three parts: a reconstruction subnet and a pixel reorganization (PixelShuffle, PS) operator.

[0045]

[0046] Here, f PU (·,s γ ): Indicates that the PU scale is s γ The PU operator is used to de-reorganize the pixels of the input single-channel image, so that its dimension changes from M×N to and Assume that both M and N can be s γ divisible; f PS (·,s γ ): Represents the PS operator, which is the inverse operation of the PU operator and is used to convert the input The low-dimensional image of the channels is pixel-wise reorganized to change its dimension from becomes M×N; for The parameter set of the network, d∈{1, 2, ..., D γ}, x (0) =x u ;f γ，d (·;θ γ，d )for The d-th reconstruction subnetwork is defined as follows:

[0047]

[0048] Among them, R γ，d (·;θ d ): is a residual learning module used to learn the lost details in the input MRI image, θ γ，d Its parameter; f DC (·) represents the data consistency operator, let and Respectively represent f DC The input and output of (·), Calculated as follows:

[0049]

[0050] here, Calculated as follows:

[0051]

[0052] in, Z[i, j] represents the element in the i-th row and j-th column of the matrix Z. Figure 1 The structure diagram of the multi-scale pixel decomposition and reconstruction network proposed in this invention is given, in which the residual learning module R γ，d (·;θ d ) can be flexibly designed according to the existing DC-CNN model. The present invention uses a full convolutional network to construct R γ，d (·;θ d ), see Figure 2 , which is composed of n convolutional layers cascaded, where n is usually set to 5; the key parameters used include: the number of PU subnets Γ, the set of PU scales used by the network s = {s1, s2, ..., s γ ,…,s Γ}、Constitute the reconstructed subnet f of each PU subnet γ，d (·;θ γ，d ) γ, their settings are related to the acceleration ratio r of the input undersampled MRI image. The following takes two commonly used acceleration ratios r = 8 and r = 16 as examples to illustrate their settings:

[0053] (1) Setting the number of PU subnets Γ: When r = 8, Γ can be set to 5; when r = 16, Γ can be set to 4;

[0054] (2) The set of PU scales s = {s1, s2, ..., s γ ,…,s Γ} settings: when r = 8, it can be set to s = {6, 4, 3, 2, 1}, that is, s1 = 6, s2 = 4, s3 = 3, s4 = 2, s5 = 1; when r = 16, it can be set to s = {6, 4, 3, 2}, that is, s1 = 6, s2 = 4, s3 = 3, s4 = 2;

[0055] (3) Reconstructed subnet f of each PU subnet γ，d (·;θ γ，d ) γ Settings: when r=8, D1=2, D2=3, D3=3, D4=3, D5=1; when r=16, D1=4, D2=3, D3=3, D4=2.

[0056] Further, Figure 5 Examples of the calculation processes of the PU operator and PS operator used are given. (a) to (b) to (c) is the calculation process of the PU operator; the PS operator is the inverse operation of the PU operator, and (c) to (b) to (a) is the calculation process of the PS operator.

[0057] Experiments were conducted to verify the beneficial effects of the proposed multi-scale pixel deconstruction network. FastMRI was used as the experimental dataset, and the proton density (PD) sequence of fastMRI was selected for reconstruction. A Gaussian undersampling matrix (with a center sampling ratio of 0.5) was used as the sampling mask. The speedup ratio r was set to 8 and 16, and the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) were used as quantitative evaluation indicators.

[0058] The multi-scale pixel decomposition and reconstruction network proposed in this invention is compared with UNet, DeepCascade, and DuDoRNet.

[0059] Figure 6The PSNR / SSIM comparison of the quantization results of the method proposed in this invention at different PU scales is shown in the figure. s ×n cascaded to represent the specific configuration of the multi-scale PU network proposed in this invention, s ×n represents the configuration of a PU subnet, where the underlined numbers s The experimental results show that the above-mentioned parameter settings of the method proposed in the present invention are reasonable.

[0060] Figure 7 This is a quantitative comparison of the reconstruction results and model training time of the proposed method with the existing methods UNet, DeepCascade, and DuDoRNet. It can be seen that compared with the existing methods, the proposed method has higher computational speed and accuracy.

[0061] Figure 8 This is a visual comparison of the reconstruction results with a speedup ratio of r = 16. (a) is the input undersampled MRI image, (b) is the fully sampled MRI image, and (c), (d), (e), and (f) are the reconstructed images using UNet, DeepCascade, DuDoRNet, and the proposed method, respectively. It can be seen that the proposed method can better reconstruct the details of highly undersampled MRI images.

[0062] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. A fast MRI reconstruction method based on a multi-scale pixel anti-reconstruction network, characterized in that: The method comprises the following steps: Step 1 Input undersampled MRI image here, represents the complex field, x u Obtained through the following methods: x u =F -1 y u =F -1 (M⊙y)=F -1 (M⊙F(x)), Among them, y u =M⊙y, represents the fully sampled MRI image to be reconstructed, y = Fx represents the fully sampled data in k-space, F(·) and F -1 (·) represent the two-dimensional Fourier transform matrix and the two-dimensional Fourier transform inverse matrix, respectively. F(·) is used to transform the MRI image from pixel space to k space, and F -1 (·) is used to transform MRI images from k-space to pixel space, M∈{0,1} M×N represents the mask matrix for undersampling y; Step 2 is performed by multi-scale pixel deconstruction network Get the reconstructed MRI image In step 2, the multi-scale pixel anti-reconstruction network f MS-PU (·; Θ) is formed by cascading Γ PU subnets: Where Θ={Θ1,Θ2,…,Θ γ ,,Θ Γ } are the parameters of the network, s={s1,s2,…,s γ ,,s Γ } is the scale of each PU used in the network, 1≤γ≤Γ, The PU scale is s γ PU subnet; PU subnet By PU operator, D γ It is composed of three parts: a reconstruction subnet and a pixel shuffle (PS) operator. here, Indicates that the PU scale is s γ The PU operator is used to de-reorganize the pixels of the input single-channel image, so that its dimension changes from M×N to and Set M and N to be s γ divisible; Represents the PS operator, which is the inverse operation of the PU operator and is used to convert the input The low-dimensional image of the channels is pixel-wise reorganized to change its dimension from becomes M×N; for The parameter set of the network, d∈{1,2,…,D γ }, for The d-th reconstruction subnetwork is defined as follows: in, is a residual learning module used to learn the lost details in the input MRI image, θ γ,d Its parameter; f DC (·) represents the data consistency operator, let and Respectively represent f DC The input and output of (·), Calculated as follows: here, Calculated as follows: in, Z[i,j] represents the element in the i-th row and j-th column of the matrix Z.

Images