Magnetic resonance CEST image reconstruction method, device and equipment based on deep learning

Abstract

The present invention discloses a magnetic resonance CEST image reconstruction method, device, medium and equipment based on deep learning, the method comprising: obtaining magnetic resonance CEST downsampled k-space data of an object to be imaged; obtaining a trained deep neural network, the network consisting of a data sharing module and a plurality of iteration modules; and reconstructing a CEST source image using a neural network. The method combines coil sensitivity encoding and neural network priors, and makes full use of redundant information in the CEST frequency dimension, thereby improving the CEST imaging speed while ensuring image quality. In addition, the present invention also proposes a multi-channel CEST data simulation method, which greatly reduces the method's reliance on the acquisition of a large amount of training data.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic resonance imaging, and in particular to a rapid chemical exchange saturation transfer imaging technique combining deep neural networks and parallel imaging reconstruction. Background Art

[0002] Chemical Exchange Saturation Transfer (CEST) imaging accumulates and amplifies the magnetic resonance (MR) signals generated by endogenous protons by selectively saturating specific endogenous protons and utilizing the phenomenon of mutual exchange between water protons and endogenous protons. Therefore, it can sensitively detect the spatial distribution of low-concentration metabolites in vivo at the molecular level and realize imaging of important physiological parameters such as metabolites and pH. At present, CEST imaging technology has shown great potential in the diagnosis and treatment of various diseases. For example, amide proton transfer (APT) technology has been used for tumor diagnosis, glutamate chemical exchange saturation transfer technology has been used for epilepsy diagnosis, and glycosaminoglycan chemical exchange saturation transfer technology has been used for osteoarthritis diagnosis.

[0003] However, the high richness of CEST imaging information comes at the cost of a long imaging time. To ensure the quality and reliability of CEST imaging, it is often necessary to collect multiple frames of images within a larger saturation offset frequency range. Therefore, although CEST imaging has a good application prospect, the disadvantage of long imaging time seriously restricts its development and clinical promotion.

[0004] To solve the above problems, parallel imaging methods represented by variable acceleration sensitivity encoding (vSENSE) are used to accelerate CEST imaging. Such methods effectively reduce the CEST imaging time by utilizing the spatial sensitivity differences of multi-channel coils, but their image reconstruction quality is easily affected by factors such as noise amplification, and the achievable acceleration multiples are limited; compressed sensing (CS) reconstruction methods based on image sparsity reduce the noise in CEST images, but such methods need to undergo a time-consuming nonlinear iterative process in the image reconstruction link, and the reconstructed images often lose detail information. These drawbacks limit the application of CS algorithms in the field of CEST imaging; in recent years, deep learning algorithms based on CEST-PROPELLER undersampling have been used to accelerate CEST imaging, but this method can only obtain CEST contrast maps, which are difficult to meet the actual quantitative analysis needs of CEST.

[0005] In addition, the structural information contained in CEST source images of different frequencies is highly similar, and the above-mentioned existing image reconstruction algorithms have failed to fully explore and utilize the redundant information in the frequency dimension. Therefore, how to fully utilize the redundant information of the data itself and achieve faster CEST imaging while ensuring image quality and reconstruction efficiency is still an urgent problem to be solved. Summary of the invention

[0006] The object of the present invention is to provide a magnetic resonance CEST image reconstruction method, device, medium and equipment with short imaging time, good reconstruction effect and high clinical and commercial value, so as to solve the problem of long CEST imaging time.

[0007] The specific technical solutions adopted by the present invention are as follows:

[0008] In a first aspect, the present invention provides a magnetic resonance CEST image reconstruction method based on deep learning, which comprises:

[0009] S1. For a target object to be subjected to magnetic resonance CEST imaging, multi-channel under-sampled k-space data and a corresponding coil sensitivity map of the target object are acquired; the multi-channel under-sampled k-space data is composed of under-sampled k-space data frames at all CEST saturation offset frequencies acquired;

[0010] S2, obtaining a trained deep neural network; the input of the deep neural network is multi-channel undersampled k-space data and corresponding coil sensitivity maps, and the output is a CEST source image reconstructed by the network;

[0011] S3. Input the multi-channel under-sampled k-space data and the corresponding coil sensitivity map obtained in S1 into the trained deep neural network to obtain a reconstructed CEST source image.

[0012] As a preferred embodiment of the first aspect, the deep neural network in S2 is composed of a data sharing module and a plurality of iteration modules cascaded after the data sharing module;

[0013] The data sharing module fills the missing parts of each frame of undersampled k-space data with the valued data at the corresponding position of the undersampled k-space data of the adjacent frames to obtain the filled k-space data, and then performs inverse Fourier transform and multi-channel merging on the filled k-space data, and uses the obtained aliased image as the output S of the module. 1 ;

[0014] Each iterative module has the same structure. The input of any k-th iterative module includes the output S of the previous cascade module. k , the multi-channel under-sampled k-space data and the coil sensitivity map, which outputs Sk+1 for:

[0015]

[0016] Where: k = 1, ... K; γ k is the network learnable weight coefficient; the encoding operator E = MFC, where M represents the k-space undersampling mask matrix, F represents the Fourier encoding matrix, and C represents the coil sensitivity map matrix in the network input; E * represents the adjoint matrix of E; Y represents the multi-channel undersampled k-space data in the network input; and They represent the i-th group of learnable 3D convolution kernels and 3D deconvolution kernels in the k-th iterative module respectively; For distribution in and A learnable activation function between .

[0017] As a preferred embodiment of the first aspect, in S2, the deep neural network is pre-trained using a multi-channel CEST data set generated by simulation, and the method for obtaining the multi-channel CEST data set generated by simulation is as follows:

[0018] S11. Acquire a multi-channel MR structural image data set, and perform data preprocessing on each MR structural image therein to obtain a multi-channel MR structural image of uniform size and a corresponding merged MR structural image after channel merging; at the same time, establish a multi-pool Bloch-McConnell mathematical model based on a water molecule pool and an amide proton pool, and traverse the values ​​of the model parameters within the corresponding preset range to generate a z spectrum set containing a large number of z spectra through numerical simulation, wherein each z spectrum covers N ω Frequency points;

[0019] S12, using each merged MR structural image obtained in S11, generating binary masks of the tumor area and the non-tumor area respectively through image segmentation, and then adding different random weak textures to the two types of masks to fuse them into a textured mask;

[0020] S13, traversing each pixel in each of the textured masks obtained in S12, and according to a pre-constructed mapping relationship table between the pixel grayscale value and the solute concentration of the amide proton pool, retrieving the z spectrum corresponding to the pixel grayscale value from the z spectrum set generated in S11, and pairing the z spectrum with the corresponding pixel, so that the paired z spectra of all pixels in each textured mask correspond to form a three-dimensional z spectrum matching map;

[0021] S14, for each multi-channel MR structural image obtained in S11, each channel image is divided into two sub-images according to the dimension number N of the z spectrum. ω Copy N ωThe layers are stacked, and then each stacked channel image is subjected to a dot multiplication operation with the corresponding z spectrum matching map obtained in S13 to form a CEST source image corresponding to each channel. Finally, each multi-channel MR structural image corresponds to a group of multi-channel CEST source images, and the multi-channel CEST source images are combined through channels to form a fully sampled CEST source image.

[0022] S15. For each group of the multi-channel CEST source images, obtain corresponding multi-channel fully sampled k-space data by Fourier transform, and use a pre-generated k-space undersampling mask to undersample the multi-channel fully sampled k-space data to obtain multi-channel undersampled k-space data; and calculate the corresponding coil sensitivity map according to the undersampled k-space data;

[0023] S16. Each group of the multi-channel under-sampled k-space data, the corresponding coil sensitivity map, and the corresponding fully sampled CEST source image together constitute a sample in the multi-channel CEST data set generated by simulation.

[0024] As a preferred embodiment of the first aspect, the learnable activation function is composed of N w Gaussian radial basis functions are weighted combined to obtain the activation function of the i-th activation function in the k-th iteration module. The form is:

[0025]

[0026] Among them, z represents the input of the activation function, and the fixed parameter δ j and σ are used to control the shape of each basis function; the combined weight of each basis function Set as learnable parameters, i=1,…N v ,j=1,…N w .

[0027] As a preferred embodiment of the first aspect, the loss function used during the deep neural network training is:

[0028]

[0029] in represents the set of all learnable parameters of the network, d k represents the set of convolution kernel weight coefficients in the kth iteration module, n is used to count the training set samples, N n is the total number of training samples in the training set; represents the nth group of CEST source images output by the deep neural network, S n Represents the nth group of fully sampled CEST source image labels; Indicates based on The calculated asymmetric effect intensity spectrum of the magnetization transfer rate corresponding to the mth offset frequency is: Indicates that based on S n The calculated asymmetric effect intensity spectrum of the magnetization transfer rate corresponding to the mth offset frequency, N M is the total number of positive and negative frequency pairs; μ(e) represents the weight coefficient related to the number of network training rounds e.

[0030] As a preferred embodiment of the above-mentioned first aspect, in the data sharing module, for any under-sampled k-space data frame, all missing data points in the frame are traversed, and it is determined whether each missing data point has valuable data at the same position in the adjacent frame. If so, the valuable data is filled into the missing data point; if not, no filling is performed.

[0031] As a preferred embodiment of the first aspect, the coil sensitivity map in S1 is directly acquired or calculated using the multi-channel under-sampled k-space data.

[0032] In a second aspect, the present invention provides a magnetic resonance CEST image reconstruction data processing device based on deep learning, which includes a memory and a processor;

[0033] The memory is used to store computer programs;

[0034] The processor is used to implement the deep learning-based magnetic resonance CEST image reconstruction method as described in any one of the schemes of the first aspect when executing the computer program.

[0035] In a third aspect, the present invention provides a computer-readable storage device having a computer program stored thereon. When the computer program is executed by a processor, the deep learning-based magnetic resonance CEST image reconstruction method as described in any one of the schemes of the first aspect is implemented.

[0036] In a fourth aspect, the present invention provides a magnetic resonance imaging device, which includes a magnetic resonance scanner and a control unit;

[0037] The magnetic resonance scanner is used to obtain CEST multi-channel under-sampled k-space data of the target object through a parallel imaging method;

[0038] A computer program is stored in the control unit, and when the computer program is executed, it is used to implement the magnetic resonance CEST image reconstruction method based on deep learning as described in any one of the solutions of the first aspect.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention combines a deep learning algorithm with a parallel imaging algorithm. On the one hand, by using a neural network prior and multi-channel spatial sensitivity difference information, the CEST imaging speed can be significantly improved while ensuring image quality, compared with existing methods, achieving a speed-up of at least 4 times for CEST imaging. On the other hand, the present invention can be directly used to reconstruct multi-channel CEST source images collected clinically, and has good clinical and commercial promotion potential.

[0041] (2) The present invention fully guarantees the reliability of the reconstruction results. Since the present method extracts and optimizes the CEST frequency dimension information with the help of the time-frequency convolution kernel, and uses the reconstruction error of the CEST effect as part of the loss function to constrain the model training, the present invention not only reconstructs the spatial dimension information of the image with high quality, but also reconstructs the CEST features represented by the z spectrum in the frequency dimension. Compared with the existing methods, the reconstruction results based on the present invention can achieve more reliable CEST analysis.

[0042] (3) The multi-channel CEST data simulation method proposed in the present invention can generate a network training data set simply and effectively. This method improves the generalization ability of the deep neural network algorithm and greatly reduces the dependence of the deep neural network algorithm on the collection of a large amount of training data, thereby ensuring the practicability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flowchart of the steps of the magnetic resonance CEST image reconstruction method based on deep learning;

[0044] Figure 2 It is a structural diagram of a deep neural network;

[0045] Figure 3 It is a flow chart of multi-channel CEST data simulation;

[0046] Figure 4 is an overall implementation flow chart of an embodiment of the present invention;

[0047] Figure 5 The APTw image reconstruction result of a brain tumor patient in the embodiment. Figure (a) shows the fully sampled APTw reference image, the GRAPPA algorithm reconstruction result, the BCS algorithm reconstruction result, the VN algorithm reconstruction result and the reconstruction result of the method of the present invention from left to right, Figure (b) shows the k-space undersampling mask used in the experiment, with an acceleration factor of 4, and Figure (c) shows the error diagrams of the GRAPPA algorithm, the BCS algorithm, the VN algorithm and the reconstruction results of the present invention relative to the fully sampled reference image from left to right, as well as the corresponding normalized root mean square error value (nRMSE).

[0048] Figure 6The reconstruction results and CEST analysis results of a 3.5ppm source image of a brain tumor patient in the embodiment are shown. From left to right are the reconstruction results of the GRAPPA algorithm, the BCS algorithm, the VN algorithm and the reconstruction results of the method of the present invention when the acceleration factor R=4; the first line is the 3.5ppm source image of a brain tumor patient reconstructed by each algorithm, the second line is the reconstruction error diagram of each algorithm for the 3.5ppm source image, the third and fourth rows are the z spectrum and MTRasym spectrum of the region of interest calculated based on the reconstruction results of each algorithm, the solid line in the spectrum represents the reference value, and the dotted line represents the reconstruction result of the corresponding algorithm. The quantitative indicators in the figure include peak signal-to-noise ratio (PSNR), normalized root mean square error (nRMSE) and mean absolute error (MAE). DETAILED DESCRIPTION

[0049] The present invention is further described and illustrated below in conjunction with the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0050] like Figure 1 As shown, as a preferred implementation of the present invention, a fast magnetic resonance CEST image reconstruction method based on deep learning is provided, which includes the following steps:

[0051] S1. For a target object to be subjected to magnetic resonance CEST imaging, multi-channel under-sampled k-space data and a corresponding coil sensitivity map are acquired for the target object; the multi-channel under-sampled k-space data is composed of all acquired under-sampled k-space data frames at CEST saturation offset frequencies.

[0052] It should be noted that in step S1, the acquisition method for the multi-channel under-sampled k-space data of the target object can be offline or online. For the offline method, it is only necessary to read the data from the storage device storing the data; while for the online method, it is necessary to use a magnetic resonance imaging device to perform CEST imaging scanning on the target object to obtain the data.

[0053] It should be noted that, in step S1, the coil sensitivity map can be directly acquired or calculated using the corresponding multi-channel under-sampled k-space data, and there is no limitation to this.

[0054] It should be noted that in step S1, the target object may be any object that can be imaged by magnetic resonance CEST, such as the brain of a patient.

[0055] S2. Obtain a trained deep neural network to reconstruct the image of the data obtained in S1, thereby realizing fast multi-channel CEST imaging.

[0056] Before describing the structure of the deep neural network in detail, a brief introduction to its basic theory is given first to facilitate understanding of the principles of the deep neural network involved in the present invention, as follows:

[0057] The CEST source image reconstruction problem involved in the present invention can be described by the following optimization model:

[0058]

[0059] in is the CEST source image to be solved, is the multi-channel under-sampled k-space data actually acquired, N x and N y Indicates the spatial size of the CEST source image, N ω Indicates the total number of CEST saturation offset frequencies, N c represents the number of channels of the coil; the encoding operator E=MFC, M represents the k-space undersampling mask matrix, F represents the Fourier encoding matrix, and C represents the coil sensitivity map; R is the regularization term, and λ is the regularization weight coefficient. The regularization used in the present invention is a generalized form of sparse regularization, and its specific structure is:

[0060]

[0061] Where D represents the sparse convolution kernel, φ represents the nonlinear potential function, the subscript i is used to count different sparse convolution kernels and potential functions, and N v is the total number of sparse convolution kernels or potential functions.

[0062] The above formula (1) can be solved by the gradient descent algorithm:

[0063]

[0064] Where E * The adjoint matrix of represents the transpose of the sparse convolution kernel D, φ i ′ is the activation function (i.e. the first-order derivative of the potential function), α k is the step size of the kth iteration.

[0065] The regularization term weight coefficient λ in the above formula (3) can be implicitly learned by the activation function, so the above formula can be rearranged into the following form:

[0066]

[0067] Therefore, based on the above theory, the structure of the deep neural network can be designed.

[0068] Specifically, expand (4) into a deep neural network, as Figure 2 As shown. The entire deep neural network consists of a data sharing module and K iteration modules in cascade. The data sharing module is used to utilize the relevant information between adjacent frames; each subsequent iteration module performs the gradient descent process shown in formula (4) once, and jointly realizes the iterative optimization of the image, and finally realizes the reconstruction of the CEST source image.

[0069] For the entire deep neural network, the network input is the multi-channel under-sampled k-space data Y of the magnetic resonance CEST imaging target and the corresponding coil sensitivity map C, and the output is the fully sampled CEST source image reconstructed by the network; among them, the input of the data sharing module is the multi-channel under-sampled k-space data Y, and the output is the channel merged aliased image S after k-space data sharing 1 ; Output S of the data sharing module 1 As one of the inputs of the first iterative module, the output S of the first iterative module 2 As one of the inputs of the second iteration module, and so on, the output of the previous iteration module is one of the inputs of the next iteration module, and the output S of the last iteration module is K+1 As the reconstruction result of the entire network, that is, the fully sampled CEST source image. It should be noted that the input of each iterative module includes the output of the previous iterative module, as well as the multi-channel under-sampled k-space data Y and the corresponding coil sensitivity map C.

[0070] The following is an introduction to the functions and implementation details of the data sharing module and iteration module:

[0071] (1) The function of the data sharing module is to fill the missing parts of each frame of undersampled k-space data with the valued data at the corresponding position of the undersampled k-space data of the adjacent frame, so as to obtain the filled k-space data. In the specific implementation process, the data sharing module first traverses all the missing data points in any undersampled k-space data frame, and determines whether there is valued data at the corresponding position of each missing data point in the two adjacent frames: 1) If there is valued data at the corresponding position of the two adjacent frames, the average of the valued data in the two frames is taken to fill the missing data point; 2) If there is valued data at the corresponding position of one of the two adjacent frames, the valued data of the corresponding frame is taken to fill the missing data point; 3) If there is no valued data, no filling is performed.

[0072] It should be noted that for the first frame and the last frame, there is only one adjacent frame, so the data sharing process is only performed with the adjacent frame.

[0073] After completing the above work, the obtained filled k-space data is subjected to inverse Fourier transform and multi-channel merging, and the obtained channel-merged aliased image is used as the output S of the data sharing module. 1 .

[0074] (2) For iterative modules, in order to maintain generality, k (k = 1, ... K) is used to refer to the number of any iterative module. For the kth iterative module, its input is the output S of the previous cascade module. k , the multi-channel under-sampled k-space data Y and the coil sensitivity map C, the output is recorded as S k+1 The specific calculation performed is shown in formula (5):

[0075]

[0076] Where: γ k is the network's learnable weight coefficient; encoding operator E and its accompanying matrix E * The definition of E is as mentioned above; * represents the adjoint matrix of E; Y represents the multi-channel undersampled k-space data in the network input; and They represent the i-th group of learnable 3D convolution kernels and 3D deconvolution kernels in the k-th iterative module, respectively. is the time-frequency convolution kernel, is the time-frequency deconvolution kernel; For distribution in and A learnable activation function between .

[0077] It should be noted that the D i , φ′ i Although they have the same structure, the learnable parameters in each iterative module are usually not shared. Therefore, formula (5) is used to calculate the learnable structure D i , φ′ i The superscript k is added to facilitate differentiation, i.e. They represent the D corresponding to the kth iteration module respectively. i , φ′ i .

[0078] It should be noted that the previous cascade modules of different iteration modules are different. For the k=1th iteration module, the previous cascade module is the data sharing module, but for other k=2,…Kth iteration modules, the previous cascade module is the k-1th iteration module.

[0079] In addition, D in each iteration module i , φ′ i , γ k (i=1,…N v) are all learnable parts of the network, and their specific forms are as follows:

[0080] 1) Convolution kernel D i and the deconvolution kernel

[0081] In order to simultaneously extract and reconstruct the spatial dimension and frequency dimension information of the source image, the convolution kernel D i Use size n x ×n y ×n ω The three-dimensional time-frequency convolution kernel is Where n x 、n y and n ω are the convolution kernel size parameters respectively, and the specific values ​​can be optimized according to the actual situation. To reconstruct the complex source image, D i Contains two channels, one for processing the real and one for processing the imaginary data of the source image matrix.

[0082] In addition, the deconvolution kernel D can be used directly i Each iteration module has N v Group convolution kernel D i and the deconvolution kernel The weight coefficients d of all convolutional kernels contained therein are set to be learned.

[0083] 2) Activation function φ′ i (i=1,…N v ):

[0084] Nonlinear activation function φ′ i Set the convolution kernel D i and the deconvolution kernel , which is used to promote the efficient expression of image information in the sparse domain. In practice, φ′ i N w Gaussian radial basis functions are weighted combined to obtain the i-th (i=1,…N) in the k-th iteration module. v ) activation functions For example, its form is:

[0085]

[0086] Where z represents the input of the activation function and the fixed parameter δ j and σ are used to control the shape of each basis function; the combined weight of each basis function Set as a learnable parameter.

[0087] 3) Data item weight coefficient γ k :

[0088] γ k The initial value of training is set to 1, and no other constraints are imposed on it during network training.

[0089] At this point, the calculation process of formula (4) executed in each iterative module can be represented by the following parameterized nonlinear mapping f:

[0090] S k+1 =f k (S k |θ k ), (7)

[0091] Where S k and S k+1 are the kth iteration module f k Input and output, represents the trainable parameters (d k represents the set of convolution kernel weight coefficients in the kth iteration module). On this basis, the reconstruction process of the entire deep neural network F can be expressed as:

[0092]

[0093] in Represents the set of all trainable parameters in the network.

[0094] Before using the above deep neural network for image reconstruction, it is necessary to optimize the network parameters in advance. The specific method of network parameter optimization can refer to the conventional training method of deep neural network, that is, setting the loss function and using the optimizer to optimize the network trainable parameter set θ until the optimal parameter set is obtained.

[0095] The loss function used in network training can be adjusted and optimized according to actual conditions. A composite loss function provided in the present invention is:

[0096]

[0097] in Represents the set of all learnable parameters of the network, n is used to count the samples in the training set, N n is the total number of training samples in the training set; represents the nth group of CEST source images output by the deep neural network, S n Represents the nth group of fully sampled CEST source image labels; Indicates based on The calculated asymmetric effect intensity spectrum of the magnetization transfer rate corresponding to the mth offset frequency is: Indicates that based on S nThe calculated asymmetric effect intensity spectrum of the magnetization transfer rate corresponding to the mth offset frequency, N M is the total number of positive and negative frequency pairs; v(e) represents the weight coefficient associated with the number of network training rounds e.

[0098] When the optimal network parameter set is obtained After that, the actual undersampled CEST source image can be reconstructed through the network containing the optimal parameters.

[0099] S3, input the multi-channel under-sampled data Y of the target object obtained in S1 into the deep neural network trained in S2 to obtain the reconstructed full-sampled CEST source image The process can be expressed as follows:

[0100]

[0101] The fully sampled CEST source image reconstructed by the above network can be further processed and analyzed by CEST according to actual needs.

[0102] In addition, the training data set of the above-mentioned deep neural network requires large-scale sample construction, which can be obtained through magnetic resonance experiment collection. However, the currently available CEST multi-channel data is relatively scarce, and it is usually difficult to meet the training needs of the network. In response to this problem, the present invention also proposes a method for generating a multi-channel CEST training data set, which uses a readily available multi-channel MR structural image data set to obtain multi-channel CEST source image data through numerical simulation of the Bloch-McConnell model, thereby greatly reducing the dependence of the deep neural network algorithm on the collection of a large amount of training data, ensuring the practicability of the algorithm.

[0103] As another preferred embodiment of the present invention, the above-mentioned deep neural network is pre-trained using a multi-channel CEST data set generated by simulation. The simulation process is as follows Figure 3 As shown, the specific steps are as follows:

[0104] S11. Acquire a multi-channel MR structural image data set, and perform necessary data preprocessing such as cropping and scaling on each MR structural image therein to obtain a multi-channel MR structural image of uniform size (including multiple channel images of different channels) and a corresponding merged MR structural image after channel merging; at the same time, establish a multi-pool Bloch-McConnell mathematical model based on a water molecule pool and an amide proton pool, and traverse the values ​​of the model parameters within a corresponding preset range (the specific range of each model parameter can be set according to actual conditions). Each set of model parameters can be used as a set of basic data for numerical simulation to generate a corresponding z spectrum, thereby generating a z spectrum set containing a large number of z spectra through numerical simulation, wherein each z spectrum covers N ωFrequency points;

[0105] S12, using each merged MR structural image obtained in S11, generating binary masks of the tumor area and the non-tumor area respectively through image segmentation, and then adding different random weak textures to the two types of masks, and then fusing them to form a textured mask.

[0106] It should be noted that the generation of the above-mentioned binary mask can be achieved through manual segmentation, or can be automatically achieved using clustering algorithms such as k-means.

[0107] In addition, the random weak texture can come from a large number of natural scene images containing rich information, or can be randomly generated. However, it is preferred to use natural scene images for generation. The specific generation method of random weak texture is as follows: first, a natural scene image containing rich texture features is obtained, and the image is uniformly grayed, and then the grayed natural scene image is smoothed and filtered, and then superimposed on the binary mask to form a texture mask.

[0108] S13, traverse each pixel in each of the textured masks obtained in S12, and according to a pre-constructed mapping relationship table between the pixel grayscale value and the solute concentration of the amide proton pool, retrieve the z spectrum corresponding to the pixel grayscale value from the z spectrum set generated in S11, and pair the z spectrum with the corresponding pixel, so that each pixel in the textured mask is matched with a z spectrum, thereby correspondingly forming a three-dimensional z spectrum matching map of the paired z spectra of all pixels in each textured mask.

[0109] S14, for each multi-channel MR structural image obtained in S11, each channel image is divided into two sub-images according to the dimension number N of the z spectrum. ω Copy N ω The layers are stacked, and then each stacked channel image is multiplied with the z spectrum matching map obtained in S13 to form a CEST source image corresponding to each channel. Finally, each multi-channel MR structural image corresponds to a group of multi-channel CEST source images, and the multi-channel CEST source images are merged through channels to form a fully sampled CEST source image.

[0110] It should be noted that since the multi-channel MR structural images and the merged MR structural images are matched in groups, and each z spectrum matching map also corresponds to a merged MR structural image, when performing the dot multiplication operation in step S14, it is necessary to take the z spectrum matching map corresponding to this group of multi-channel MR structural images and merged MR structural images for dot multiplication with the stacked channel images.

[0111] S15. For each group of the multi-channel CEST source images, obtain corresponding multi-channel fully sampled k-space data by Fourier transform, and use a pre-generated k-space undersampling mask to undersample the multi-channel fully sampled k-space data to obtain multi-channel undersampled k-space data; and calculate the corresponding coil sensitivity map according to the undersampled k-space data;

[0112] It should be noted that in order to fully utilize the redundant information of the CEST frequency dimension during image reconstruction, the k-space undersampling masks of different frames should have different undersampling patterns. The unsampled k-space parts are filled with 0s to simulate the actual k-space data undersampling process.

[0113] S16. Each group of the multi-channel under-sampled k-space data, the corresponding coil sensitivity map, and the corresponding fully sampled CEST source image together constitute a sample in the multi-channel CEST data set generated by simulation.

[0114] Based on the methods described in S1 to S3 and S11 to S16 above, the following will combine them with specific embodiments to demonstrate their specific technical effects so that those skilled in the art can better understand the essence of the present invention.

[0115] Example

[0116] The deep learning-based rapid magnetic resonance CEST image reconstruction method described in S1 to S3 above is applied to a specific embodiment to demonstrate the technical effects that can be achieved. In addition, in order to illustrate the effectiveness of the proposed multi-channel CEST data set simulation method, this embodiment uses a CEST data set generated by simulation according to the S11 to S16 process to train the network. The specific framework process of the methods described in S1 to S3 and S11 to S16 above is as described above, and will not be fully repeated in this embodiment. The following focuses on showing the specific implementation details and technical effects of each step. The complete process of this embodiment is as follows Figure 4 shown.

[0117] 1. Data preparation

[0118] 1.1 Training Data Simulation

[0119] 1) Preparatory work: On the one hand, the fastMRI open source structural image dataset was obtained, and each structural image was cropped, scaled and other necessary preprocessing was performed to obtain multi-channel MR structural images of uniform size and the corresponding merged MR structural images after channel merging. On the other hand, a three-pool Bloch-McConnell model including a water proton pool, an amide proton pool and a magnetization transfer (MT) pool was established. By traversing the simulation parameter values ​​within the preset range, a z spectrum set containing a large number of z spectra was generated through numerical simulation, where each z spectrum covered 53 frequency points within the range of -6 to 6 ppm.

[0120] 2) Tumor identification, segmentation and texture mask generation: The improved k-means clustering algorithm is used to identify and segment the tumor regions in the combined MR structural images obtained in 1), and binary masks of the tumor region and non-tumor region are generated respectively. Then, different random weak textures are added to the two types of masks to obtain textured masks. The textures come from a large number of natural scene images containing rich information.

[0121] 3) Generate a z-spectrum matching map: traverse each pixel in each of the textured masks, retrieve the z-spectrum corresponding to the pixel grayscale value from the z-spectrum set generated in 1) according to a pre-constructed mapping relationship table between the pixel grayscale value and the solute concentration of the amide proton pool, and pair the z-spectrum with the corresponding pixel; the z-spectrums of all pixels in each textured mask correspond to a three-dimensional z-spectrum matching map.

[0122] 4) Generate a multi-channel CEST source image: Take each group of multi-channel MR structural images in 1) as an unsaturated image (S0), and use the z spectrum matching map generated in 3) to perform saturation modulation on the unsaturated image to obtain a multi-channel CEST source image saturated with a frequency of -6 to 6 ppm. The specific saturation modulation method is: first, the images of each channel of the multi-channel MR structural image are saturated according to the dimension number N of the z spectrum. ω =53 copies and stacks into 53 layers, and then each stacked channel image is multiplied point by point (Hadamard product) with the z spectrum matching map corresponding to the structural image. The multi-channel CEST source image can be Fourier transformed to obtain the multi-channel fully sampled k-space data.

[0123] According to the above steps, this embodiment simulates and generates 9200 sets of multi-channel full-sampling k-space data.

[0124] 1.2 MRI data acquisition

[0125] In order to test the generalization performance of the model on real data, evaluate the effectiveness of the CEST simulation data set, and monitor the model training status in real time, this embodiment obtains the measured CEST data through MR experiments as samples in the test set.

[0126] Specifically, a 3 Tesla Siemens scanner with a 16-channel head coil (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany) was used to scan the brains of two glioma patients. The sequence used in the scanning experiment was a 2D fast spin echo (TSE) CEST imaging sequence, and its specific acquisition parameters were as follows: saturation pulse duration of 1.0 s, intensity of 2 μT, flip angle FA = 90°; echo time (TE) = 6.7 ms; repetition time (TR) = 3 s; field of view (FOV) = 212 × 186 mm 2 ; Resolution = 2.2 × 2.2 mm 2 ; Slice thickness = 5 mm; Acquisition turbine coefficient = 96. A total of 54 frequency offset frames were acquired, including the unsaturated frame S0, and saturated frames with frequencies of 0, ±0.25, ±0.5, ±0.75, ±1, ±1.5, ±2(2), ±2.5(2), ±3(2), ±3.25(2), ±3.5(6), ±3.75(2), ±4(2), ±4.5, ±5, and ±6 ppm (the numbers in brackets represent the number of repeated acquisitions of the corresponding frequency point frames).

[0127] In order to correct the B0 field inhomogeneity, this embodiment calculates the B0 spectrum by direct water saturation shift referencing (WASSR). The TR of the WASSR sequence used is 2s, the saturation pulse intensity is 0.5μT, and a total of 26 frequency points equidistantly distributed within -1.5 to 1.5ppm are collected. Other parameters are consistent with the above-mentioned CEST imaging sequence.

[0128] 1.3 Data preprocessing and dataset construction

[0129] In order to construct the training set, validation set and test set of the network, it is necessary to preprocess the fully sampled k-space data collected from the simulation or MRI experiment by downsampling.

[0130] In this embodiment, the Figure 5 (b) shows a 1D variable density Cartesian undersampling mode, where each k-space data is undersampled by 4 times to obtain multi-channel undersampled k-space data. Then, the corresponding coil sensitivity map is calculated based on the undersampled k-space data using the ESPIRiT algorithm. In addition, the sensitivity encoding (SENSE) algorithm is used to reconstruct the fully sampled k-space data to obtain the corresponding complex-valued fully sampled CEST source image.

[0131] So far, each independent data sample includes 1) multi-channel undersampled k-space data, 2) corresponding coil sensitivity map, and 3) complex valued fully sampled CEST source image, of which the first two will be used as inputs to the neural network, and the complex valued fully sampled CEST source image will be used as the label of the sample (Ground Truth). For the 9200 sets of samples generated by simulation, 7000 sets are randomly selected as training sets, and the rest are used as validation sets. The test set of the model is constructed from the collected samples of the MRI experiment.

[0132] 2. Model building and training

[0133] Use the deep learning framework PyTorch to build Figure 2 The deep neural network shown in the figure has the number of network iteration modules K set to 10, the size of the time-frequency convolution kernel in each iteration module is set to 7×7×7, and the number of basis functions and convolution groups N v is 48, Gaussian radial basis function N w The number is 31.

[0134] The network is trained using the multi-channel CEST dataset generated by simulation. During training, the Adam optimizer is used to optimize the network parameter set θ by minimizing the loss function value shown in formula (9) to obtain the relatively optimal network parameters The weight coefficient μ(e) in the loss function is set to a constant of 15, and all neural networks are trained for 40 iterations on four NVIDIA RTX 2080 Ti GPUs. During the training process, the validation set is used to test the generalization ability of the current model and determine whether the network has reached convergence conditions.

[0135] 3. Image reconstruction and post-processing

[0136] After the above network is trained, it can be used for image reconstruction of actual acquired data. Specifically, this embodiment uses the undersampled k-space data of brain tumor patients and the corresponding coil sensitivity map as the input of the network, and the network output is the reconstructed fully sampled CEST source image. At the same time, in order to compare and evaluate the performance of this algorithm, the GRAPPA (Generalized Autocalibrating Partially Parallel Acquisitions) algorithm, the BCS (Blind Compressed Sensing) algorithm combined with parallel imaging, and the unimproved VN (Variational Network) algorithm are used to reconstruct the same undersampled data.

[0137] In addition, in order to eliminate the influence of B0 field inhomogeneity, the WASSR method is used to perform B0 correction on the CEST source images reconstructed by various algorithms, and then the APTw image is calculated by equation (11):

[0138]

[0139] And taking APTw image as an example, CEST analysis is performed.

[0140] 4. Results Analysis

[0141] Figure 5 The reconstruction results of the algorithm for a CEST image of a brain tumor patient are shown. The experimental results show that, under the same 4-fold undersampling mode, compared with advanced image reconstruction algorithms such as GRAPPA, BCS and VN, the use of this reconstruction algorithm can achieve significantly higher CEST image quality, and the difference between the reconstructed image and the full sampling image is minimal.

[0142] Figure 6 The CEST source image at +3.5ppm reconstructed by this algorithm is shown, as well as the z spectrum and MTRasym spectrum of the region of interest calculated based on the reconstruction results. The experimental results show that compared with other advanced algorithms, this algorithm can better reconstruct the detailed information of the image and has a stronger ability to remove image artifacts. At the same time, the z spectrum and MTRasym spectrum obtained by this algorithm are closer to the reference situation, so the source image reconstructed based on this algorithm can be used for more reliable CEST analysis.

[0143] In addition, the network trained using simulation data in the embodiment can achieve high-quality reconstruction of the actual CEST images. This shows that the network has excellent generalization performance. It also indirectly illustrates the effectiveness of the multi-channel CEST data set simulation method proposed in the present invention, which can well meet the training requirements of deep neural networks.

[0144] Similarly, based on the same inventive concept, another preferred embodiment of the present invention further provides a magnetic resonance CEST imaging data processing device based on deep learning corresponding to the magnetic resonance CEST image reconstruction method based on deep learning provided in the above embodiment, which includes a memory and a processor;

[0145] The memory is used to store computer programs;

[0146] The processor is used to implement the magnetic resonance CEST image reconstruction method based on deep learning as described above when executing the computer program.

[0147] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the deep learning-based magnetic resonance CEST image reconstruction method provided in the above embodiment, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the deep learning-based magnetic resonance CEST image reconstruction method as described above is implemented.

[0148] It is understandable that the above storage medium may include random access memory (RAM) and non-volatile memory (NVM), such as at least one disk storage. At the same time, the storage medium may also be various media that can store program codes, such as a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. Of course, with the widespread use of cloud servers, the above software programs can also be installed on the cloud platform to provide corresponding services, so the computer-readable storage medium is not limited to the form of local hardware.

[0149] It can be understood that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0150] It should also be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In each embodiment provided in this application, the division of steps or modules in the device and method is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0151] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0152] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a magnetic resonance imaging device corresponding to the magnetic resonance CEST image reconstruction method based on deep learning provided in the above embodiment, which includes a magnetic resonance scanner and a control unit. Among them:

[0153] The magnetic resonance scanner is used to obtain k-space data under-sampling frames, k-space under-sampling masks and corresponding coil sensitivity maps of the target object at all frequency offsets by a parallel imaging method;

[0154] A computer program is stored in the control unit, and when the computer program is executed, it is used to implement the magnetic resonance CEST image reconstruction method based on deep learning as described above.

[0155] It should be noted that the magnetic resonance imaging device can be any magnetic resonance scanner that can implement the parallel imaging method, and its structure belongs to the existing technology, and a mature commercial product can be used, and the specific model is not limited. In addition, in addition to storing the above-mentioned computer program, the control unit of the magnetic resonance imaging device should also have the imaging sequence and other software programs necessary to implement CEST imaging.

[0156] Of course, the above-mentioned control unit can be an independent control unit or a control unit provided in the magnetic resonance scanner, that is, the above-mentioned deep learning-based magnetic resonance CEST image reconstruction method can be integrated into the control unit of the magnetic resonance imaging device in the form of a data processing program, so that the magnetic resonance scanner can directly output the reconstruction result without the need for an additional control unit.

[0157] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A magnetic resonance CEST image reconstruction method based on deep learning, characterized in that: include: S1. For a target object to be subjected to magnetic resonance CEST imaging, multi-channel under-sampled k-space data and a corresponding coil sensitivity map of the target object are acquired; the multi-channel under-sampled k-space data is composed of under-sampled k-space data frames at all CEST saturation offset frequencies acquired; S2, obtaining a trained deep neural network; the input of the deep neural network is multi-channel undersampled k-space data and corresponding coil sensitivity maps, and the output is a CEST source image reconstructed by the network; S3, inputting the multi-channel under-sampled k-space data and the corresponding coil sensitivity map obtained in S1 into the trained deep neural network to obtain a reconstructed CEST source image; The deep neural network in S2 is composed of a data sharing module and a plurality of iterative modules cascaded after the data sharing module; The data sharing module fills the missing parts of each frame of undersampled k-space data with the valued data at the corresponding position of the undersampled k-space data of the adjacent frames to obtain the filled k-space data, and then performs inverse Fourier transform and multi-channel merging on the filled k-space data, and uses the obtained aliased image as the output S of the module. 1 ; Each iterative module has the same structure. The input of any k-th iterative module includes the output S of the previous cascade module. k , the multi-channel under-sampled k-space data and the coil sensitivity map, which outputs S k+1 for: Where: k = 1, ... K; γ k is the network learnable weight coefficient; the encoding operator E = MFC, where M represents the k-space undersampling mask matrix, F represents the Fourier encoding matrix, and C represents the coil sensitivity map matrix in the network input; E * represents the adjoint matrix of E; Y represents the multi-channel undersampled k-space data in the network input; and They represent the i-th group of learnable 3D convolution kernels and 3D deconvolution kernels in the k-th iterative module respectively; For distribution in and A learnable activation function between ; In S2, the deep neural network is pre-trained using a multi-channel CEST data set generated by simulation, and the method for obtaining the multi-channel CEST data set generated by simulation is as follows: S11. Acquire a multi-channel MR structural image data set, and perform data preprocessing on each MR structural image therein to obtain a multi-channel MR structural image of uniform size and a corresponding merged MR structural image after channel merging; at the same time, establish a multi-pool Bloch-McConnell mathematical model based on a water molecule pool and an amide proton pool, and traverse the values ​​of the model parameters within the corresponding preset range to generate a z spectrum set containing a large number of z spectra through numerical simulation, wherein each z spectrum covers N ω Frequency points; S12, using each merged MR structural image obtained in S11, generating binary masks of the tumor area and the non-tumor area respectively through image segmentation, and then adding different random weak textures to the two types of masks to fuse them into a textured mask; S13, traversing each pixel in each of the textured masks obtained in S12, and according to a pre-constructed mapping relationship table between the pixel grayscale value and the solute concentration of the amide proton pool, retrieving the z spectrum corresponding to the pixel grayscale value from the z spectrum set generated in S11, and pairing the z spectrum with the corresponding pixel, so that the paired z spectra of all pixels in each textured mask correspond to form a three-dimensional z spectrum matching map; S14, for each multi-channel MR structural image obtained in S11, each channel image is divided into two sub-images according to the dimension number N of the z spectrum. ω Copy N ω The layers are stacked, and then each stacked channel image is subjected to a dot multiplication operation with the corresponding z spectrum matching map obtained in S13 to form a CEST source image corresponding to each channel. Finally, each multi-channel MR structural image corresponds to a group of multi-channel CEST source images, and the multi-channel CEST source images are combined through channels to form a fully sampled CEST source image. S15. For each group of the multi-channel CEST source images, obtain corresponding multi-channel fully sampled k-space data by Fourier transform, and use a pre-generated k-space undersampling mask to undersample the multi-channel fully sampled k-space data to obtain multi-channel undersampled k-space data; and calculate the corresponding coil sensitivity map according to the undersampled k-space data; S16, each group of multi-channel under-sampled k-space data, the corresponding coil sensitivity map and the corresponding fully sampled CEST source image together constitute a sample in the multi-channel CEST data set generated by simulation; The learnable activation function is given by N w Gaussian radial basis functions are weighted combined to obtain the activation function of the i-th activation function in the k-th iteration module. The form is: Among them, z represents the input of the activation function, and the fixed parameter δ j and σ are used to control the shape of each basis function; the combined weight of each basis function Set as learnable parameters, i=1,…N v ,j=1,…N w ; The loss function used during the training of the deep neural network is: in represents the set of all learnable parameters of the network, d k represents the set of convolution kernel weight coefficients in the kth iteration module, n is used to count the training set samples, N n is the total number of training samples in the training set; represents the nth group of CEST source images output by the deep neural network, S n Represents the nth group of fully sampled CEST source image labels; Indicates based on The calculated asymmetric effect intensity spectrum of the magnetization transfer rate corresponding to the mth offset frequency is: Indicates that based on S n The calculated asymmetric effect intensity spectrum of the magnetization transfer rate corresponding to the mth offset frequency, N M is the total number of positive and negative frequency pairs; μ(e) represents the weight coefficient related to the number of network training rounds e; In the data sharing module, for any undersampled k-space data frame, all missing data points in the frame are traversed to determine whether each missing data point has valuable data at the same position in the adjacent frame. If so, the valuable data is filled into the missing data point; if not, no filling is performed.

2. The magnetic resonance CEST image reconstruction method based on deep learning according to claim 1, characterized in that: The coil sensitivity map in S1 is directly acquired or calculated using the multi-channel under-sampled k-space data.

3. A magnetic resonance CEST imaging data processing system based on deep learning, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the deep learning-based magnetic resonance CEST image reconstruction method as described in claim 1 or 2 when executing the computer program.

4. A computer-readable storage device, characterized in that: The storage device stores a computer program, and when the computer program is executed by the processor, the magnetic resonance CEST image reconstruction method based on deep learning as described in claim 1 or 2 is implemented.

5. A magnetic resonance imaging device, characterized in that: including a magnetic resonance scanner and a control unit; The magnetic resonance scanner is used to obtain CEST multi-channel under-sampled k-space data of the target object through a parallel imaging method; The control unit stores a computer program, which, when executed, is used to implement the deep learning-based magnetic resonance CEST image reconstruction method as described in claim 1 or 2.

Images