A slice-adaptive active undersampling method for magnetic resonance imaging

Abstract

The present invention relates to a slice-adaptive magnetic resonance imaging active undersampling method, comprising the following specific steps: step 1, constructing a data set required for the active undersampling method; step 2, preprocessing the data; step 3, constructing a parameter-free interactive environment under the framework of reinforcement learning; step 4, constructing a Transformer-based sampling decision network to map the undersampled image in the data state and its corresponding mask in different ways; step 5, training the constructed sampling decision network using the established data set, preprocessing operation and parameter-free interactive environment; step 6, performing real-time active undersampling based on the trained sampling decision network model. The present invention can reconstruct a stable and efficient interactive environment and a sampling decision network that can efficiently model content information and timing information in the data state under the framework of deep reinforcement learning, thereby achieving high-performance active undersampling.

Description

Technical Field

[0001] The invention belongs to the field of information and communication engineering, and relates to an active undersampling method for magnetic resonance imaging, in particular to an active undersampling method for magnetic resonance imaging oriented to slice adaptation. Background Art

[0002] Compared with computed tomography (CT), magnetic resonance imaging (MRI) has the advantages of zero radiation exposure and good soft tissue imaging contrast. However, the scanning of k-space data is extremely time-consuming. The long scanning process will increase the discomfort of the examinee, cause motion artifacts and waste precious medical diagnosis time. Therefore, shortening the scanning time is crucial to the development of MRI equipment.

[0003] Generally speaking, a two-dimensional Cartesian scan is a process of frequency encoding in one dimension (row / column) and phase encoding in another dimension (column / row) to obtain the k-space complex matrix data K∈￡ F×P The scanning time T is expressed as T = TR × P, where TR is the repetition time. Since the scanning time mainly depends on the number of phase encodings, the number of phase encoding samples can be reduced to C < P to speed up the scanning process. The k-space undersampled data containing C phase encodings is defined as Then the acceleration factor AF of the scanning process is P / C, and the number of k-space phase selection combinations is Assume P = 368, C = 70, then

[0004] The actual k-space acquisition process has a significant impact on the phase encoding decision time t p Has strict requirements, namely t p The phase encoding time interval cannot be exceeded Especially in the fast spin echo sequence (FSE), there are multiple phase encodings in one TR time period. The typical phase encoding time interval is In this case, it is required that p <15ms.

[0005] According to the level of data adaptation, phase encoding decision methods can be divided into data-independent, data set-adaptive, and slice-adaptive methods. Data-independent or data set-adaptive methods determine the undersampling mask in a priori or offline manner, so there is no decision time for each phase. Slice-adaptive methods can customize the subset of optimal phase encoding for each slice to be scanned, thereby overcoming the distribution deviation of slices relative to the dataset. Therefore, this method is expected to obtain the optimal sampling trajectory for each slice, thereby obtaining the most accurate reconstructed image. Since the phase selection of slice-adaptive methods is achieved through feature analysis of each phase and slice, the decision efficiency of slice-adaptive methods is difficult to guarantee.

[0006] Data-independent undersampling methods can determine phase subsets without modeling MRI images. Representative methods include random undersampling combined with a uniform probability density function and variable density undersampling.

[0007] The dataset adaptation method uses the characteristics of the entire training data to determine the phase subset. The fixed undersampling mask composed of the determined phase subset can be used for undersampling of all test data. Representative methods include reconstruction error-driven variable density method, iterative low error transfer method, phase-related reconstruction network and learning-based undersampling pattern optimization method (LOUPE)

[0008] The slice adaptive method can generate a phase subset for each k-space slice to be reconstructed, and the slice to be reconstructed is acquired and reconstructed based on this phase subset. Representative methods include eigenvalue-based methods, uncertainty-based methods, and reinforcement learning-based methods proposed by Pineda et al. The latter two methods use deep learning, and the last method is better than the first two methods. Among them, the reinforcement learning-based method proposed by Pineda et al. has achieved great success in the field of slice adaptive phase decision (active undersampling). However, the method proposed by Pineda et al. is inefficient, and its phase decision time is about 1,333ms, which is far longer than the typical phase encoding time interval (maximum allowable phase decision time) of 15ms. Therefore, this method cannot be used in the actual acquisition process at present. In addition, this method does not fully refine the timing information, which limits the ability of phase decision.

[0009] After searching, no published patent documents identical or similar to the present invention were found. Summary of the invention

[0010] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a slice-adaptive active undersampling method for magnetic resonance imaging, which can reconstruct a stable and efficient interactive environment and a sampling decision network that can efficiently model content information and timing information in the data state under the framework of deep reinforcement learning, thereby achieving high-performance active undersampling.

[0011] The present invention solves the practical problem by adopting the following technical solutions:

[0012] A slice-adaptive magnetic resonance imaging active undersampling method includes the following specific steps:

[0013] Step 1, using a magnetic resonance device to collect a large amount of magnetic resonance data to construct a data set required by the active undersampling method;

[0014] Step 2: preprocess each data to be input into the network in the data set obtained in step 1;

[0015] Step 3: Build a parameter-free interactive environment under the framework of reinforcement learning;

[0016] Step 4: Build a sampling decision network based on Transformer to map the undersampled images and their corresponding masks in the data state in different ways;

[0017] Step 5: Train the constructed sampling decision network using the established data set, preprocessing operations, and parameter-free interactive environment;

[0018] Step 6: Based on the trained sampling decision network model, perform real-time active under-sampling.

[0019] Moreover, the data set of step 1 is collected in a full sampling manner, the collection coordinate system is a Cartesian coordinate system, and the data is stored in the frequency domain.

[0020] Moreover, the specific steps of step 2 include:

[0021] (1) First, construct a low-frequency k-space undersampling mask. The mask is symmetrical about the center row / column of the k-space and contains L rows / columns with all 1s and the rest are 0s.

[0022] (2) Then the mask is multiplied by the k-space data to obtain the initial low-frequency under-sampled data. The combination of an under-sampled data and its corresponding under-sampled mask is called a data state.

[0023] Moreover, the parameter-free interactive environment of step 3 includes two parts, data state transfer and image reconstruction; the constructed parameter-free interactive environment can add the phase encoding corresponding to the action to the current undersampled data according to the action selected by the decision network, so as to realize the transfer of data state; the constructed parameter-free interactive environment can reconstruct the k-space image to the spatial domain by inverse Fourier transform, and evaluate the value of the action taken according to the degree of improvement of the image quality before and after the state transfer; wherein, the degree of improvement of the image quality is quantified by the difference of the image quality evaluation index, and this difference is called the reward R.

[0024] Moreover, the specific steps of step 4 include:

[0025] (1) For an undersampled image, it is first decomposed into image blocks, and then each decomposed image block is mapped to a one-dimensional vector, called an image block vector, and there are c image block vectors;

[0026] (2) For the mask corresponding to the undersampled image, it is directly mapped to a vector of the same dimension as the above one-dimensional vector, which is called the mask vector;

[0027] (3) In addition, a category vector is set, so there are a total of c+2 input vectors of dimension d. In order to maintain the spatial information of the data state, a position vector is constructed and then embedded into the input vector by addition, finally forming a d×(c+2) matrix; this d×(c+2) matrix is ​​input into N Transformer encoding layers for different types of feature interactions to predict phase encoding. A Transformer encoding layer includes a normalization layer, a multi-head attention layer, a normalization layer and a multi-layer perceptron in sequence; the output matrix dimension of the last Transformer encoding layer is still d×(c+2), and the features corresponding to the category vector in the output matrix are input into a mapping head with a normalization layer and a fully connected layer to realize the mapping of action to Q value, where Q value represents the cumulative discounted expectation of reward value in deep reinforcement learning;

[0028] (4) The Q value training process combines the time difference and the current reward value. First, the current reward value is mapped to Q in the form of time difference. target ; then Q target The Q value predicted by the decision network is used to smooth the L1 loss and feed it back to each layer of the decision network, so that the parameters of the decision network can be optimized, enabling it to make high-performance sampling decisions based on the current k-space data state.

[0029] Moreover, the specific method of step 5 is: the learning rate used during training is lr, the number of training times is ln, the training of the decision network uses the playback unit technology, the capacity of the playback unit is sr, and the training decision network is configured accordingly to obtain a converged model.

[0030] Advantages and beneficial effects of the present invention:

[0031] 1. The present invention proposes a Transformer-based slice adaptive magnetic resonance imaging active undersampling method, which performs adaptive undersampling on each slice, makes phase encoding decisions that can ensure reconstruction accuracy, and accelerates MRI scanning.

[0032] 2. The input of the Transformer-based deep neural network of the present invention includes not only the image block of the current reconstructed image (undersampled image given by the environment), but also the historical phase decision results. The input of the entire data state is combined with the high-performance Transformer design, which can simultaneously utilize the image content information and the sampling timing information contained in the phase decision. The interaction and fusion of the two types of information improves the performance of the decision network.

[0033] 3. The present invention generates input images for the decision network by performing IFFT (Inverse Fast Fourier Transform) on undersampled k-space data, which has efficient and stable reward estimation. In contrast, in the method proposed by Pineda, the input image of the decision network is obtained through a time-consuming reconstruction network (large-scale cascaded fully convolutional ResNet). Since each phase decision corresponds to a state transition, the reconstruction network has to run multiple times, resulting in low efficiency of the method proposed by Pineda.

[0034] 4. The sampling decision network of the present invention is lightweight and efficient. The lightweight decision network and the efficient environment for generating input images for the decision network enable the proposed active undersampling method to perform real-time slice adaptive phase encoding decisions within a time interval less than the maximum allowed phase encoding interval. In contrast, the current optimal method proposed by Pineda et al. cannot perform phase encoding selection within the maximum allowed phase encoding time interval.

[0035] 5. The present invention shortens the decision time to 9ms by setting up a stable and efficient reinforcement learning environment and a fast and high-performance phase encoding decision network. Compared with the current most advanced method of Pineda et al. with a decision time of up to 1,333ms, the decision efficiency is not only about 150 times faster, but also more accurate images can be reconstructed on the selected phase subset. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the slice-oriented adaptive magnetic resonance imaging active undersampling method of the present invention;

[0037] FIG2( a ) is a schematic diagram of the active undersampling decision network based on Transformer of the present invention - the overall structure of the decision network;

[0038] FIG2( b ) is a schematic diagram of the Transformer-based active undersampling decision network of the present invention - a multi-head attention module, a multi-layer perceptron and a mapping head;

[0039] Figure 3 This is a diagram of the inverse Fourier transform reconstruction result of the phase subset selected by the decision network when AF=3.3 of the present invention, wherein the first row is the reconstructed image and the second row is the true value image. DETAILED DESCRIPTION

[0040] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0041] A slice-adaptive active undersampling method for magnetic resonance imaging uses a magnetic resonance device to collect fully sampled k-space data to construct a data set required for the active undersampling method; in the training phase, this data is constructed into a data state including undersampled data and its corresponding mask, and input into a decision network to enable it to interact with the environment; the decision network model is optimized using back propagation of loss so that the network has the ability to select key phase encoding, that is, has the ability of high-performance active undersampling.

[0042] like Figure 1 As shown, the specific steps include:

[0043] Step 1, using a magnetic resonance device to collect a large amount of magnetic resonance data to construct a data set required by the active undersampling method;

[0044] The data set of step 1 is collected in a full sampling manner, the collection coordinate system is a Cartesian coordinate system, and the data is stored in the frequency domain.

[0045] The spatial domain form data obtained by inverse Fourier transform of frequency domain form data can well express the content information of the image.

[0046] Step 2: preprocess each data to be input into the network in the data set obtained in step 1;

[0047] The specific steps of step 2 include:

[0048] (1) First, construct a low-frequency k-space undersampling mask. The mask is symmetrical about the center row / column of the k-space and contains L rows / columns with all 1s and the rest are 0s.

[0049] (2) Then the mask is multiplied by the k-space data to obtain the initial low-frequency under-sampled data. The combination of an under-sampled data and its corresponding under-sampled mask is called a data state.

[0050] In this embodiment, the step 2 performs the following preprocessing on each data to be input into the network in the above data set: first, a low-frequency k-space undersampling mask is constructed, the mask is symmetrical about the center column of the k-space, contains 30 columns of all 1s, and the rest are 0s; then, the mask is multiplied by the k-space data to obtain the initial low-frequency undersampling data. The combination of an undersampling data and its corresponding undersampling mask is called a data state.

[0051] Step 3: Build a parameter-free interactive environment under the framework of reinforcement learning;

[0052] The parameter-free interactive environment of step 3 includes two parts, data state transfer and image reconstruction; the constructed parameter-free interactive environment can add the phase encoding corresponding to the action to the current undersampled data according to the action selected by the decision network, so as to realize the transfer of data state; the constructed parameter-free interactive environment can reconstruct the k-space image into the spatial domain by inverse Fourier transform, and evaluate the value of the taken action according to the degree of improvement of the image quality before and after the state transfer; wherein, the degree of improvement of the image quality is quantified by the difference of the image quality evaluation index, and this difference is called the reward R.

[0053] In this embodiment, available evaluation indicators include SSIM, PSNR, and NMSE.

[0054] Step 4: Build a sampling decision network based on Transformer to map the undersampled images and their corresponding masks in the data state in different ways;

[0055] The specific steps of step 4 include:

[0056] (1) For an undersampled image, it is first decomposed into image blocks, and then each decomposed image block is mapped to a one-dimensional vector, called an image block vector, and there are c image block vectors;

[0057] (2) For the mask corresponding to the undersampled image, it is directly mapped to a vector of the same dimension as the above one-dimensional vector, which is called the mask vector;

[0058] (3) In addition, a category vector is set, so there are a total of c+2 input vectors of dimension d. In order to maintain the spatial information of the data state, a position vector is constructed and then embedded into the input vector by addition, finally forming a d×(c+2) matrix; this d×(c+2) matrix is ​​input into N Transformer encoding layers for different types of feature interactions to predict phase encoding. A Transformer encoding layer includes a normalization layer, a multi-head attention layer, a normalization layer and a multi-layer perceptron in sequence; the output matrix dimension of the last Transformer encoding layer is still d×(c+2), and the features corresponding to the category vector in the output matrix are input into a mapping head with a normalization layer and a fully connected layer to realize the mapping of action to Q value, where Q value represents the cumulative discounted expectation of reward value in deep reinforcement learning;

[0059] (4) The Q value training process combines the time difference and the current reward value. First, the current reward value is mapped to Q in the form of time difference. target ; then Q targetThe Q value predicted by the decision network is used to smooth the L1 loss and feed it back to each layer of the decision network, so that the parameters of the decision network can be optimized, enabling it to make high-performance sampling decisions based on the current k-space data state.

[0060] In this embodiment, step 4 constructs a sampling decision network based on Transformer, and maps the undersampled image and its corresponding mask in the data state in different ways. For the undersampled image, it is first decomposed into image blocks, and then each decomposed image block is mapped to a one-dimensional vector, called an image block vector, and there are 920 image block vectors; for the mask corresponding to the undersampled image, it is directly mapped to a vector of the same dimension as the above one-dimensional vector, called a mask vector; in addition, a category vector is set. Therefore, there are 922 input vectors with a dimension of 512. In order to maintain spatial information, a position vector is constructed and then embedded into the input vector by addition, and finally a 512×922 matrix is ​​formed. The matrix is ​​input into 4 Transformer encoding layers for different types of feature interactions to predict phase encoding. A Transformer encoding layer includes a normalization layer, a multi-head attention layer, another normalization layer, and a multi-layer perceptron in sequence. Among them, the multi-layer perceptron has only one hidden layer. The output matrix dimension of the last Transformer encoding layer is still 512×922. The features corresponding to the category vector are input into a mapping head with a normalization layer and a fully connected layer to realize the mapping from action to Q value, where Q value represents the cumulative discounted expectation of reward value in deep reinforcement learning. The training process of Q value combines time difference and current reward value (taking time t as an example, Q t =R t +Q t+1 ), first map the current reward value to Q in a time-difference manner target ; then Q target The Q value predicted by the decision network is used to smooth the L1 loss and feed it back to each layer of the decision network, so that the parameters of the decision network can be optimized, enabling it to make high-performance sampling decisions based on the current k-space data state.

[0061] As shown in Figure 2(a) and Figure 2(b), Figure 2(a) focuses on the overall structure of the decision network: three different types of vectors are set to encode data states, among which the image block vector extracts image content information; the mask vector extracts undersampled temporal information; the category vector summarizes the features in the Transformer encoding layer to achieve Q value mapping. Figure 2(b) Multi-head attention first maps the input vector to obtain Q, K, and V, and then divides it into multiple attention heads. Multiple attention heads extract features of multiple dimensions, and finally summarize the features; the multi-layer perceptron further interacts and fuses the features of the multi-head attention through multiple fully connected layers; the mapping head realizes the mapping of category vector features to Q values ​​through normalization layers and fully connected layers.

[0062] Step 5: Use the established data set, preprocessing operations, and parameter-free interactive environment to train the constructed sampling decision network.

[0063] The specific method of step 5 is: the learning rate used during training is lr, the number of training times is ln, the training of the decision network uses the playback unit technology, the capacity of the playback unit is sr, and the training decision network is configured accordingly to obtain a converged model.

[0064] In this embodiment, step 5 uses the established data set, preprocessing operations, and parameter-free interactive environment to train the constructed sampling decision network. The learning rate used in training is 10 -4 , the number of training times is 10 7 ,The training of the decision network also uses the replay unit technology, and the capacity of the replay unit is 20,000. ,According to this configuration, the decision network is trained to obtain a converged model.

[0065] Step 6: Integrate the trained sampling decision network model into the hardware device and combine it with the magnetic resonance scanning device to make sampling decisions during the actual acquisition process.

[0066] Specifically, when the magnetic resonance device collects k-space data, it first collects data according to the low-frequency mask defined above and pre-processes the collected data; then the data state is input into the sampling decision network for forward propagation to complete the current sampling decision; finally, the decision is looped until the specified acceleration factor is met, thereby achieving high-performance k-space active undersampling that is adaptive to the slice.

[0067] The experimental results of the present invention are shown in the attached Figure 3 As shown, according to the active undersampling method proposed in the present invention, the sampling trajectory of MRI data is obtained, and the data is collected and reconstructed using inverse Fourier transform based on this trajectory to obtain the reconstructed image in the first row, the second row is the corresponding MRI image (true value image) when fully sampled, and the SSIM in the first row of images is obtained by calculating the first row of images and the corresponding images in the second row.

[0068] Figure 3 The undersampling ratio AF of the image in the figure is 3.3. It can be seen that when only the inverse Fourier transform is used to reconstruct the phase subset selected by the proposed slice-oriented adaptive active undersampling method, from the perspective of objective indicators, the SSIM can reach above 0.8 or even 0.9. Using this inverse Fourier transform reconstruction result as the input of the reconstruction neural network, a higher performance neural network reconstructed image can be output. From the reconstructed image, the image contrast is almost consistent with the true value, retaining most of the detail information, without introducing additional noise or artifacts, and has a high performance restoration degree. The parameter-free environment proposed in this method can stably and efficiently realize state transfer and action evaluation, and the proposed Transformer-based decision network can fully model image content information and sampling timing information. The combination of the two realizes the generation of high-performance undersampling trajectories while greatly improving the efficiency of active undersampling decisions.

[0069] The working principle of the present invention is:

[0070] The present invention proposes a slice-adaptive magnetic resonance imaging active undersampling method, comprising the following steps:

[0071] (1) using a magnetic resonance device to collect and store fully sampled Cartesian coordinate k-space magnetic resonance data, thereby constructing a data set required for the active undersampling method;

[0072] (2) Each data to be input into the network in the above data set is preprocessed as follows: first, a low-frequency k-space undersampling mask is constructed; then, low-frequency pre-acquisition is performed based on this mask to obtain initial undersampling data. The low-frequency pre-acquisition data is a small portion of k-space data. The combination of an undersampling data and its corresponding mask is called a data state;

[0073] When preprocessing the data, a predefined low-frequency mask is applied to each slice, that is, the mask is applied to the k-space of the slice to obtain the corresponding undersampled k-space. The mask and its corresponding k-space data are jointly constructed into a data state, which contains the content information of the undersampled data and the timing information of the sampling, and can provide sufficient information for the decision network.

[0074] (3) Under the framework of reinforcement learning, a parameter-free interactive environment is constructed, which can evaluate the current sampling action decision (a row / column k-space phase encoding index) according to the degree of improvement in reconstruction quality, and can transfer the current undersampled data state combined with the sampling action (k-space combination) to the next undersampled data state;

[0075] (4) Construct a sampling decision network based on Transformer, where the input of the network is a data state and the output is the expected value of the reward corresponding to different phase encodings. The optimal action is selected by the maximum value strategy for the output of the sampling decision network, and the action is evaluated according to the above parameter-free environment. The evaluation result is fed back to the sampling decision network in the form of the degree of improvement in reconstruction quality. The feedback is passed to each layer of the decision network through back propagation, and the parameter optimization of the decision network can be completed, so that it has the ability to make high-performance sampling decisions based on the current k-space data state;

[0076] In the framework of reinforcement learning, the present invention constructs a parameter-free interactive environment and a Transformer-based sampling decision network. The obvious feature of the interactive environment is that the evaluation and transfer of data states are realized in a parameter-free form. The obvious feature of the decision network is the coordinated use of undersampled k-space data and its corresponding mask.

[0077] The present invention combines the undersampled k-space with the sampling action in a superposition manner to achieve the transfer of the data state. The inverse Fourier transform is used for reconstruction, and the sampling action is evaluated according to the image quality difference before and after the data state transfer.

[0078] There are different ways to map the undersampled images and their corresponding masks in the data state. For the undersampled image, it is first decomposed into image blocks, and then each decomposed image block is mapped to a one-dimensional vector, called an image block vector; for the mask corresponding to the undersampled image, it is directly mapped to a vector of the same dimension as the above one-dimensional vector, called a mask vector. In addition, a category vector is set to summarize the features from each vector at the feature layer of the network and finally form an action decision.

[0079] The image block vector, mask vector, and category vector interact and fuse features in the subsequent Transformer encoding layer, utilizing each other's feature expressions in the encoding layer.

[0080] (5) Using the established data set, preprocessing operations, and parameter-free interactive environment to train the constructed sampling decision network to obtain a trained network model;

[0081] (6) The trained sampling decision network model is integrated into the hardware device. When acquiring k-space data, the data is first acquired according to the low-frequency mask defined above, and the acquired data is preprocessed. Then, the data state is input into the sampling decision network for forward propagation to complete the current sampling decision. Finally, the decision is looped until the specified acceleration factor is met, thereby realizing high-performance k-space active undersampling that is adaptive to the slice.

[0082] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0083] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0084] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0085] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

Claims

1. A slice-adaptive active undersampling method for magnetic resonance imaging, characterized in that: The specific steps include: Step 1, using a magnetic resonance device to collect a large amount of magnetic resonance data to construct a data set required by the active undersampling method; Step 2: preprocess each data to be input into the network in the data set obtained in step 1; Step 3: Build a parameter-free interactive environment under the framework of reinforcement learning; Step 4: Build a sampling decision network based on Transformer to map the undersampled images and their corresponding masks in the data state in different ways; Step 5: Train the constructed sampling decision network using the established data set, preprocessing operations, and parameter-free interactive environment; Step 6: Based on the trained sampling decision network model, perform real-time active undersampling; The specific steps of step 2 include: (1) First, construct a low-frequency k-space undersampling mask. The mask is symmetrical about the center row / column of the k-space and contains L rows / columns with all 1s and the rest are 0s. (2) Then the mask is multiplied by the k-space data to obtain the initial low-frequency under-sampled data. The combination of an under-sampled data and its corresponding under-sampled mask is called a data state. The parameter-free interactive environment of step 3 includes two parts, data state transfer and image reconstruction; the constructed parameter-free interactive environment can add the phase encoding corresponding to the action to the current undersampled data according to the action selected by the decision network to realize the transfer of data state; the constructed parameter-free interactive environment can reconstruct the k-space image into the spatial domain by inverse Fourier transform, and evaluate the value of the taken action according to the degree of improvement of image quality before and after the state transfer; wherein the degree of improvement of image quality is quantified by the difference of image quality evaluation index, and this difference is called reward R; The specific steps of step 4 include: (1) For an undersampled image, it is first decomposed into image blocks, and then each decomposed image block is mapped to a one-dimensional vector, called an image block vector, and there are c image block vectors; (2) For the mask corresponding to the undersampled image, it is directly mapped to a vector of the same dimension as the above one-dimensional vector, which is called the mask vector; (3) Set a category vector, that is, there are c+2 input vectors with dimension d; construct a position vector and then embed it into the input vector by adding it, finally forming a d×(c+2) matrix; the d×(c+2) matrix is ​​input into N Transformer encoding layers for different types of feature interactions to predict phase encoding. A Transformer encoding layer includes a normalization layer, a multi-head attention layer, a normalization layer and a multi-layer perceptron in sequence; the output matrix dimension of the last Transformer encoding layer is still d×(c+2), and the features corresponding to the category vector in the output matrix are input into a mapping head with a normalization layer and a fully connected layer to realize the mapping of action to Q value, where Q value represents the cumulative discounted expectation of reward value in deep reinforcement learning; (4) The Q value training process combines the time difference and the current reward value. First, the current reward value is mapped to Q in the form of time difference. target ; then Q target The Q value predicted by the decision network is used to smooth the L1 loss and feed it back to each layer of the decision network, so that the parameters of the decision network can be optimized, enabling it to make high-performance sampling decisions based on the current k-space data state.

2. The slice-adaptive magnetic resonance imaging active undersampling method according to claim 1, characterized in that: The data set of step 1 is collected in a full sampling manner, the collection coordinate system is a Cartesian coordinate system, and the data is stored in the frequency domain.

3. The slice-adaptive magnetic resonance imaging active undersampling method according to claim 1, characterized in that: The specific method of step 5 is: the learning rate used during training is lr, the number of training times is ln, the training of the decision network uses the playback unit technology, the capacity of the playback unit is sr, and the training decision network is configured accordingly to obtain a converged model.

Images