A cross-domain iterative network-based MRI undersampling image reconstruction method

Abstract

The present invention proposes an MRI under-sampling image reconstruction method based on a cross-domain iterative network, which includes preparing a set of fully sampled and under-sampled MRI k-space data and image data to form a training set and a validation set data; constructing a CNN densely connected network that crosses the k-space domain and the image domain, and predefining a loss function; inputting the training set data into the network for training; using the validation set to test the performance and generalization ability of the network, and determining the network model and parameters; and inputting the under-sampled k-space data into the trained image reconstruction network to reconstruct the image. The loss function considers both the image domain loss and the k-space domain loss, and uses multi-supervised training to effectively avoid the loss of supervisory information during gradient backpropagation. The present invention can restore the detailed information of the image while eliminating the high-frequency oscillation artifacts introduced by under-sampling, thereby achieving high-quality images while accelerating MRI sampling.

Description

Technical Field

[0001] The present invention relates to the technical fields of magnetic resonance imaging (MRI), deep learning, undersampling reconstruction, and the like, and in particular to an MRI undersampling image reconstruction method based on a cross-iterative network in a K-space domain and an image domain. Background Art

[0002] MRI is a non-invasive, non-ionizing radiation, arbitrary section direction, multi-parameter imaging technology with high soft tissue resolution, which can provide high-resolution structural and functional information for clinical diagnosis. Traditional MRI needs to collect 2D or 3D data matrix in Cartesian coordinate system, that is, k-space domain data, and the corresponding image domain information can be obtained after inverse Fourier transform. Each row of k-space data corresponds to a phase encoding step. The acquisition of different phase encoding data requires waiting for the recovery of longitudinal magnetization, that is, waiting time TR. According to different MRI imaging methods, TR ranges from hundreds of milliseconds to several seconds. Therefore, MRI is a relatively time-consuming imaging technology. Conventional MRI, the theoretical premise is that the imaging part is still, otherwise motion artifacts will be introduced. Therefore, during the MRI examination, the examinee needs to maintain a body position and stay still for one to several minutes. Abdominal examination requires multiple breath holding acquisitions; parts with natural physiological movements such as the heart require the development of special imaging methods; at the same time, MRI data acquisition requires patients to be in a semi-enclosed magnetic space, which is a challenge for most patients to cooperate for a long time, so MRI urgently needs to develop fast imaging technology.

[0003] At present, the method of combining multi-coil unit parallel imaging has accelerated the imaging speed, with an acceleration factor of about 2, but these methods rely on special hardware or sequences, and there is limited room for further improving the sampling speed. Compressed sensing MRI (Compressed Sensing-M, CS-MRI) is a technology that can speed up imaging by undersampling k-space data at a much lower rate than under the Nyquist sampling theorem, without the need for specific hardware and sequences. However, CS-MRI technology also has some limitations. For example, the CS-MRI sampling trajectory needs to meet the irrelevance criterion, and the typical acceleration factor is between 2.5 and 3; CS-MRI uses fewer sparse coefficients in the global coefficient transformation, which has the potential to hide image detail structures and even introduce blurring artifacts, and cannot accurately describe complex biological tissue structures.

[0004] Convolutional neural network (CNN) combines the three ideas of local receptive field, weight sharing and time or space subsampling to obtain a certain degree of displacement, scale and deformation invariance, which is more similar to biological neural network, reduces the complexity of network model and the number of weights. In recent years, it has been used in the field of MRI undersampling reconstruction. The nonlinear relationship between undersampled image and reference image (full sampled image) is obtained by training the network, which basically does not make full use of the collected k-space data information, and the image details cannot be reconstructed when the acceleration factor is high. Summary of the invention

[0005] Based on the requirements for accelerated imaging and existing technical problems, the present invention proposes an MRI under-sampling image reconstruction method based on a cross-iterative network in the k-space domain and the image domain.

[0006] The technical solution of the present invention is implemented as follows: a method for reconstructing MRI undersampling images based on a cross-domain iterative network, comprising the following steps:

[0007] Step 1: Retrospectively preprocess the fully sampled k-space data to obtain training set and validation set data;

[0008] Step 2: Construct a cross-domain iterative neural network (KI network). The KI network is composed of N modules in cascade. Each module consists of a k-space domain CNN network Kcnn and an image domain CNN network Icnn. The two networks are connected through FT (Fourier transform) or inverse Fourier transform IFT (inverse Fourier transform); predefine the loss function loss of the KI network;

[0009] Step 3: Train the KI network. Input the training set into the KI network for training. If the loss function is the smallest, it is used as the trained intermediate network. If the loss function is not the smallest, back propagate to update the network parameters and return to the KI network for training.

[0010] Step 4: Detect the KI network, input the verification set into the intermediate network trained in step 3, obtain the reconstructed image, detect the generalization ability and overfitting of the neural network, and output the network and its parameters as the image reconstruction network if the evaluation criteria are met; if the evaluation criteria are not met, adjust the network parameters and the number of layers, expand the training set and return to step 2;

[0011] Step 5: Reconstruct the undersampled k-space image. Input the undersampled k-space data into the image reconstruction network obtained in step 4 to reconstruct a high-quality image.

[0012] Furthermore, in step 1, the fully sampled k-space data is randomly undersampled to obtain undersampled k-space data k u, perform inverse Fourier transform IFT on the fully sampled k-space data to obtain the target image data x. The training set consists of multiple groups of undersampled k-space data k. u , fully sampled k-space data and target image data x; the validation set consists of multiple sets of undersampled k-space data k u and target image data x.

[0013] Furthermore, in step 2, both Kcnn and Icnn networks use residual neural networks, which are composed of an input layer, multiple hidden layers and an output layer. The activation function of the hidden layer is the linear rectified function (Rectified Linear Unit, ReLU), and the activation function of the output layer is the softmax function.

[0014] Furthermore, the hidden layer of the Icnn network contains a channel-wise attention layer CA (channel-wise attention layer), which is used to adjust the weights of each channel; a data consistency layer DC (Data consistency layer) is added before the output of the Icnn network.

[0015] Furthermore, the target data set of the Kcnn network is the fully sampled k-space data, and the target data of the Icnn network is the target image data x; the input layer data of the first module Kcnn network is the undersampled k-space data k in the training set. u , the complex k-space data is divided into real and imaginary parts and input into the network respectively; the input layer data of the Kcnn network in the nth module is the image data x output by the Icnn network in the n-1th module n-1 The data after FT (x n-1 ) and the k-space data k output by the Kcnn network in all previous modules n-1 , k n-2 …k1.

[0016] Furthermore, the input data of the Icnn network in the nth module is the k-space data k output by the Kcnn network in the nth module. n The data after IFT (k n ) and the image data x output by the Icnn network in all previous modules n-1 , x n-2 …x1.

[0017] Furthermore, in step 2, the loss function of the KI network is predefined as:

[0018]

[0019] in,

[0020]

[0021] Among them, W k and W I is the weight factor, K is the fully sampled k-space data in the training set, is the k-space data k output by the nth module Kcnn network n ; I is the image x after IFT of the fully sampled k-space data in the training set, is the image data x output by the nth module Icnn network n , M is the total number of training set data groups.

[0022] Furthermore, in step 3, the undersampled k-space data k in the training set are u Input the KI network and use the fully sampled k-space data and target image data x as target data to train the KI network.

[0023] Furthermore, in step 4, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) between the reconstructed image and the target image are selected as the evaluation basis, and the evaluation criteria are that the PSNR is not less than 35dB and the structural similarity is not less than 0.95.

[0024] Beneficial effects of the present invention:

[0025] The present invention proposes an MRI under-sampling image reconstruction method based on a k-space domain and image domain cross-iterative network, which comprises the following steps: preparing a set of fully sampled and under-sampled MRI k-space data and image data to form a training set and a validation set data; constructing a CNN dense connection network that crosses the k-space domain and the image domain; inputting the training set data into the network for training; using the validation set to test the performance and generalization ability of the network, and determining the network model and parameters; and inputting the under-sampled k-space data into the trained image reconstruction network to reconstruct the image.

[0026] Dense connections are used in the iterative process of the network module, which can effectively utilize the information of the intermediate layer; when constructing the k-space domain network, the step-by-step distribution characteristics of MRI k-space data are taken into consideration, and a residual network is used; when constructing the image domain network, a channel attention mechanism is introduced to effectively adjust the weights of each channel, realize the optimization extraction of features, and effectively restore the detailed information of the image; in the loss function, both the image domain loss and the k-space domain loss are considered, and multi-supervised training is used to effectively avoid the loss of supervised information during gradient return. The k-space domain residual network used in the present invention can alleviate the uneven distribution of data intensity in the k-space, and can suppress high-frequency oscillation artifacts in the image. The present invention can restore the detailed information of the image while eliminating the high-frequency oscillation artifacts introduced by undersampling, thereby achieving undersampling to accelerate the MRI sampling speed while obtaining high-quality images. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 is a flow chart of the present invention;

[0029] Figure 2 The structural diagram of the cross-domain iteration network;

[0030] Figure 3 It is the structure diagram of Kcnn network in k-space domain;

[0031] Figure 4 This is the structure diagram of the image domain network Icnn;

[0032] Figure 5 is the k-space undersampling template and the reconstructed image. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] A cross-domain iterative network-based MRI undersampling image reconstruction method, the process is as follows Figure 1 As shown, the specific steps are as follows:

[0035] Step 1: Preprocess the k-space data to obtain training set and validation set data.

[0036] The fully sampled k-space data is randomly undersampled, and the sampling template is as follows: Figure 5 As shown, the undersampled k-space data k is obtained. u ; Perform inverse Fourier transform IFT on the fully sampled k-space data to obtain the corresponding image x, referred to as target image data x, which is used as label data.

[0037] The training set consists of multiple groups of undersampled k-space data, fully sampled k-space data and target image data x.

[0038] The validation set consists of multiple sets of undersampled k-space data and target image data x.

[0039] Step 2: Construct a cross-domain iterative neural network (KI network), such as Figure 2 shown.

[0040] 1) The KI network is composed of N modules in cascade. Each module consists of a k-space domain CNN network Kcnn and an image domain CNN network Icnn. The Kcnn network is connected to the Icnn network through the inverse Fourier transform IFT, and the Icnn network is connected to the Kcnn network through the Fourier transform FT. The Kcnn network in each module is cascaded with the Kcnn network in the previous module; the Icnn network in each module is cascaded with the Icnn network in the previous module;

[0041] 2) If Figure 3 and Figure 4 As shown in the figure, both Kcnn and Icnn networks are residual neural networks, which are composed of an input layer, multiple hidden layers and an output layer. The activation function of the hidden layer is the linear rectification function (Rectified Linear Unit, ReLU), and the activation function of the output layer is the softmax function.

[0042] 3) If Figure 4 As shown in the figure, the hidden layer of the Icnn network contains a channel-wise attention layer (CA), which is used to adjust the weights of each channel; a data consistency layer (DC) is added before the output of the Icnn network. The DC layer is calculated as follows:

[0043]

[0044]

[0045] in: is the k-space data of the image obtained by the Icnn network before the DC layer after FT, and k is the k-space data obtained by the Icnn network after the DC layer, corresponding to the FT of the image data output by the Icnn network. x ,c y is the subscript of k-space data, and the random undersampling template Mask is a matrix composed of 0 and 1, such as Figure 5 As shown; parameter λ is 10 6 .

[0046] 4) The target data sets of Kcnn and Icnn networks, i.e., the labeled data, are the fully sampled k-space data and its IFT-transformed image data x.

[0047] like Figure 2 and Figure 3As shown in the figure, the input data of the first module Kcnn network is the undersampled k-space data in the training set, and the complex k-space data is divided into real and imaginary parts and input into the network respectively; the input data of the Kcnn network in the nth module is the image data x output by the Icnn network in the n-1th module n-1 The data after FT (x n-1 ) and the k-space data k output by the Kcnn network in all previous modules n-1 , k n-2 …k1.

[0048] like Figure 2 and Figure 3 As shown, the input data of the Icnn network in the nth module is the image data k output by the Kcnn network in the nth module n The data after IFT (x n ) and the image data x output by the Icnn network in all previous modules n-1 , x n-2 …x1.

[0049] 5) The loss function loss defined by the KI network is as follows:

[0050]

[0051] in,

[0052]

[0053] Among them, W k Take 0.1, W I Take 0.99, K is the fully sampled k-space data in the training set, is the k-space data k output by the nth module Kcnn network n ; I is the image x after IFT of the fully sampled k-space data in the training set, is the image data x output by the nth module Icnn network n , M is the total number of training set data groups.

[0054] Step 3: Train the KI network.

[0055] like Figure 1 and Figure 2 As shown, the undersampled k-space data k in the training set u The KI network is input, and the full sampled k-space and the image data x obtained after IFT transformation are used as target image data to train the neural network.

[0056] Step 4: Detection neural network (KI network).

[0057] like Figure 1As shown, the undersampled k-space data of the validation set and the corresponding target image data x data are input into the trained intermediate network to obtain the reconstructed image.

[0058] The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) between the reconstructed image and the target image were selected as the evaluation criteria.

[0059] The evaluation criteria are that the PSNR is not less than 35dB and the structural similarity is not less than 0.95.

[0060] If the evaluation criteria are met, the network and its parameters are output and used as an image reconstruction network;

[0061] If the evaluation criteria are not met, the network parameters and the number of layers are adjusted, the training set is expanded, and the second step is returned to ensure the generalization ability and overfitting of the neural network. Expanding the training set means increasing the number of training set data groups.

[0062] The final number of network modules used is 6.

[0063] Step 5: Reconstruct the undersampled k-space image.

[0064] The undersampled k-space data is input into the above image reconstruction network to reconstruct a high-quality image, such as Figure 5 As shown in FIG. 1 , the k-space is undersampled template Mask and the reconstructed image.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A MRI undersampling image reconstruction method based on a cross-domain iterative network, characterized in that: The following steps are involved: Step 1: Preprocess the fully sampled k-space data to obtain training set and validation set data; Step 2: Construct the KI network. The KI network is composed of N modules in cascade. Each module consists of a k-space domain CNN network Kcnn and an image domain CNN network Icnn. The two networks are connected through FT or IFT, and the loss function loss of the KI network is predefined. Step 3: Train the KI network. Input the training set into the KI network for training. If the loss function is the smallest, it is used as the trained intermediate network. If the loss function is not the smallest, back propagate to update the network parameters and return to the KI network for training. Step 4: Detect the KI network, input the verification set into the intermediate network trained in step 3, obtain the reconstructed image, detect the generalization ability and overfitting of the neural network, and output the network and its parameters as the image reconstruction network if the evaluation criteria are met; if the evaluation criteria are not met, adjust the network parameters and the number of layers, expand the training set and return to step 2; Step 5: Reconstruct the undersampled k-space image, input the undersampled k-space data into the image reconstruction network obtained in step 4, and reconstruct a high-quality image; In step 1, the fully sampled k-space data is randomly undersampled to obtain undersampled k-space data k u , perform IFT on the fully sampled k-space data to obtain the target image data x, and the training set consists of multiple groups of under-sampled k-space data k u , fully sampled k-space data and target image data x; the validation set consists of multiple sets of undersampled k-space data k u and target image data x; In step 2, both Kcnn and Icnn networks use residual neural networks, which are composed of an input layer, multiple hidden layers and an output layer. The activation function of the hidden layer is a linear rectification function, and the activation function of the output layer is a softmax function. The hidden layer of the Icnn network contains a channel attention layer CA, which is used to adjust the weights of each channel; a data fidelity layer DC layer is added before the output of the Icnn network. The DC layer is calculated as follows: in: is the k-space data of the image obtained by the Icnn network before the DC layer after FT, k is the k-space data obtained by the Icnn network after the DC layer, corresponding to the FT of the image data output by the Icnn network; c x ,c y is the subscript of k-space data, the random undersampling template Mask is a matrix composed of 0 and 1, and the parameter λ is 10 6 ; The target data set of Kcnn is the fully sampled k-space data, and the target data set of Icnn network is the target image data x; the input layer data of the first module Kcnn network is the undersampled k-space data k in the training set u , the complex k-space data is divided into real and imaginary parts and input into the network respectively; the input layer data of the Kcnn network in the nth module is the image data x output by the Icnn network in the n-1th module n-1 The data after FT (x n-1 ) and the k-space data k output by the Kcnn network in all previous modules n-1 , k n-2 …k1; The input layer data of the Icnn network in the nth module is the k-space data k output by the Kcnn network in the nth module n The data after IFT (k n ) and the image data x output by the Icnn network in all previous modules n-1 , x n-2 …x1; In step 2, the loss function of the predefined KI network is: in, Among them, W k and W I is the weight factor, K is the fully sampled k-space data in the training set, is the k-space data k output by the nth module Kcnn network n ; I is the target image data x after IFT of the fully sampled k-space data in the training set, is the image data x output by the nth module Icnn network n , M is the total number of training set data groups.

2. The MRI undersampling image reconstruction method based on a cross-domain iterative network according to claim 1, characterized in that: In step 3, the undersampled k-space data k in the training set are u Input the KI network and use the fully sampled k-space data and target image data x as target data to train the KI network.

3. The MRI undersampling image reconstruction method based on a cross-domain iterative network according to claim 1, characterized in that: In step 4, the PSNR and structural similarity between the reconstructed image and the target image are selected as the evaluation basis, and the evaluation criteria are that the PSNR is not less than 35dB and the structural similarity is not less than 0.95.

Images