Slice Adaptive Determination Method for Radial Sampling Trajectories in Magnetic Resonance Imaging

Abstract

The present invention relates to a method for adaptively determining slices for a radial sampling trajectory in magnetic resonance imaging, comprising the following steps: Step 1, obtain magnetic resonance data and construct training and test data for a deep reinforcement learning model; Step 2, perform pre-training of a reconstruction network based on the training data obtained in Step 1; Step 3, construct the environment and decision network of a reinforcement learning system according to the pre-trained reconstruction network obtained in Step 2, so as to achieve high-performance active undersampling of k-space for radial slice adaption. The present invention can improve the utilization ability of k-space data, improve the problem that the current single-objective reward function is not sufficient to represent the change of image quality, and make the goal of reinforcement learning clear by re-establishing the training mode of the reconstruction network, thereby improving the quality of magnetic resonance imaging.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic and communication engineering, and relates to an active undersampling method for magnetic resonance imaging, in particular to a slice adaptive determination method for a radial sampling trajectory in magnetic resonance imaging. Background Art

[0002] Magnetic resonance imaging is a medical imaging technology applied to radiology and is one of the most widely used imaging technologies in current clinical medicine and basic life science research. Compared with other medical imaging technologies, the advantages of magnetic resonance imaging are its non-ionizing radiation, high soft tissue contrast, high spatial resolution, no human trauma, rich imaging parameters, etc.

[0003] Although magnetic resonance imaging has the above-mentioned multiple advantages, the data of magnetic resonance imaging is sequentially acquired in the k-space, and the time for this data acquisition is very slow due to hardware conditions. Therefore, realizing fast MRI imaging and ensuring the quality of magnetic resonance imaging are the key points and difficulties in the field of magnetic resonance imaging research and have great clinical application value.

[0004] However, in the existing technologies for realizing fast MRI imaging, the method for obtaining excellent undersampling trajectories based on the deep reinforcement learning method has the following problems: 1. It only models for the Cartesian sampling mode and makes trajectory decisions for the Cartesian sampling mode, lacking the ability to fully utilize the k-space data; 2. The input data of the training mode of the reconstruction network has poor correlation, and the input data is a combination of all possible k-space data, and the learning goal of the reinforcement learning is not clear, affecting the final optimization learning of the undersampling trajectory with a specific acceleration ratio; 3. The reward function used by the system is only a single-objective reward function, and the evaluation of the change in image quality is not sufficient, affecting the learning performance of the reinforcement learning system.

[0005] After retrieval, no published patent documents identical or similar to the present invention have been found. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and propose a slice adaptive determination method for a radial undersampling trajectory in magnetic resonance imaging, which can improve the ability to utilize k-space data, improve the problem that the current single-objective reward function is not sufficient to represent the change in image quality, and make the goal of the reinforcement learning clear by re-establishing the training mode of the reconstruction network, thereby improving the quality of magnetic resonance imaging.

[0007] The present invention solves its practical problems by adopting the following technical solutions:

[0008] A slice adaptive determination method for a radial sampling trajectory in magnetic resonance imaging includes the following steps:

[0009] Step 1: Obtain magnetic resonance data and construct the training and test data for the deep reinforcement learning model;

[0010] Step 2: Perform pre-training of the reconstruction network based on the training data obtained in Step 1;

[0011] Step 3: Construct the environment and decision network of the reinforcement learning system based on the pre-trained reconstruction network obtained in Step 2, and then achieve high-performance k-space active undersampling adaptive to radial slices.

[0012] Moreover, the specific method of Step 1 is as follows:

[0013] Collect simulated magnetic resonance data. The fully sampled data k1 is saved in the k-space, where the k-space is the frequency domain space for saving magnetic resonance data. Subsequently, the corresponding fully sampled image data x1 is generated through inverse Fourier transform. Store the above two types of data to construct the relevant training and test data for the deep reinforcement learning model;

[0014] Moreover, the specific method of Step 2 includes:

[0015] (1) Construct a k-space undersampling mask M according to k1. The shape of M is the same as that of k1, and all data points in M are set to 0;

[0016] (2) Let the target acceleration ratio be N1 times, and N2 be the total number of radial spokes contained in a single fully sampled data k1. Randomly select |N2 / N1| spokes for each fully sampled data k1, and according to the data positions of the selected spokes, assign 1 to the corresponding positions in M, and the remaining positions are 0;

[0017] (3) Multiply M by the fully sampled data k1 to obtain the undersampled data k2, and perform inverse Fourier transform on k2 to obtain the undersampled image data x2. Use x2 as the input of the reconstruction network during the training process, and use x1 as the image ground truth for pre-training the reconstruction network to train the pre-training reconstruction network.

[0018] Moreover, the environment of the reinforcement learning system constructed in Step 3 is as follows:

[0019] Under the framework of reinforcement learning, construct an interactive environment for radial sampling trajectory learning. The initial state of the environment is the state where only 45° spokes are collected;

[0020] The environment of the reinforcement learning system can evaluate the current action decision based on the improvement degree of the reconstructed image quality, and can transfer the current undersampled data state combined with the sampling action to the next undersampled data state. The termination state is the state with |N2 / N1| spokes;

[0021] Among them, the degree of improvement in image reconstruction quality is called the reward. The reward R includes three evaluation indicators, structural similarity, peak signal-to-noise ratio, and normalized mean square error, and the importance of the three indicators is integrated through parameters. The calculation method is R = p1*S + p2*P + p3*N, where S is the change in structural similarity, P is the change in peak signal-to-noise ratio, N is the change in normalized mean square error, and p, p1, p2, and p3 are importance integration parameters respectively. p is a preset value, p1 = the order of magnitude where P is located / the order of magnitude where S is located, p2 = p, and p3 = -1 * the order of magnitude where P is located / the order of magnitude where N is located.

[0022] Moreover, the decision network constructed in step 3 is as follows:

[0023] The decision network includes a feature extraction module and an action value prediction module; the feature extraction module is connected to the action value prediction module. The feature extraction module is based on a convolutional neural network, and the action value prediction module is based on a fully connected network; the input of the feature extraction module is the reconstructed image calculated by the pre-trained reconstruction network obtained in step 2, and the output is the feature map of the reconstructed image; the input of the action value prediction module is the one-dimensional expansion of the feature map, and the output is the predicted value corresponding to different actions;

[0024] Among them, the predicted value is called the Q value. The Q value represents the cumulative discounted expectation of the reward in deep reinforcement learning; during the training process, the Q value combines the temporal difference calculation value and the current reward value. The current reward value is mapped to Qtarget in a temporal difference manner, and the Smooth L1 loss is calculated between Qtarget and the Q value predicted by the decision network and fed back to each layer of the decision network to optimize the parameters of the decision network. After training, the network has the ability to make sampling decisions based on the current undersampled reconstruction data.

[0025] Moreover, after step 3, the following steps are further included:

[0026] Step 4: Train the decision network constructed in step 3 to obtain a trained decision network;

[0027] Step 5: Test the trained decision network obtained in step 4 to verify the network effect.

[0028] Step 6: Integrate the trained sampling trajectory decision model into the hardware device and determine the radial slice adaptive sampling trajectory during the actual acquisition process in combination with the magnetic resonance device.

[0029] Moreover, the specific method of step 4 is:

[0030] Use the learning rate l, the number of training times ln. The training of the decision network uses the replay unit technology and the target network parameter freezing technology, and the capacity of the replay unit is lw.

[0031] Advantages and beneficial effects of the present invention:

[0032] 1. The present invention proposes a method for determining a slice-adaptive radial sampling trajectory of magnetic resonance imaging based on a convolutional neural network. Adaptive radial undersampling is performed on each slice to obtain an adaptive radial undersampling trajectory, improving the utilization ability of k-space data and the quality of magnetic resonance imaging.

[0033] 2. The input of the convolutional neural network based on the present invention is a reconstructed image obtained by a reconstruction network using the aforementioned pre-training method. It can extract features from the reconstructed image and output the value of different actions. This method can extract features of different slices, generate unique sampling trajectories for different slices, be adaptable to different slices, and improve the quality of magnetic resonance imaging.

[0034] 3. The training method of the pre-training network of the present invention is a training method set according to the acceleration ratio. Therefore, the image quality calculated by the reward function is more conducive to improving the image quality under the preset acceleration ratio. The pre-training reconstruction network obtained by the training method of the present invention has high computational efficiency, and the decision network is lightweight and efficient. Description of the Drawings

[0035] Figure 1 is a flowchart of a slice-adaptive determination method for a radial sampling trajectory of magnetic resonance imaging according to the present invention;

[0036] Figure 2 is a schematic diagram of a decision network based on a convolutional neural network according to the present invention. Detailed Embodiment

[0037] The following further details the embodiments of the present invention with reference to the drawings:

[0038] A slice-adaptive determination method for a radial sampling trajectory of magnetic resonance imaging, as Figure 1 shown, includes the following steps:

[0039] Step 1, obtain magnetic resonance data and construct training and test data of a deep reinforcement learning model;

[0040] The specific method of Step 1 is:

[0041] Collect simulated magnetic resonance data. The fully sampled data k1 collected is stored in the k-space. The k-space is the frequency domain space for storing magnetic resonance data. Subsequently, the corresponding fully sampled image data x1 is generated through inverse Fourier transform. Store the above two types of data to construct relevant training and test data of the deep reinforcement learning model;

[0042] In this embodiment, the acquisition method of this data set is full sampling, and the data is stored in the frequency domain form.

[0043] Step 2: Perform pre-training of the reconstruction network based on the training data obtained in Step 1;

[0044] The specific method of Step 2 includes:

[0045] (1) Construct a k-space undersampling mask M according to k1. The shape of M is the same as that of k1, and all data points in M are set to 0;

[0046] (2) Let the target acceleration ratio be N1 times, and N2 be the total number of radial spokes contained in a single fully sampled data k1 (one spoke represents a piece of data in the radial data). Randomly select |N2 / N1| (rounded) spokes for each fully sampled data k1, and according to the data positions of the selected spokes, assign the corresponding positions in M as 1, and the remaining positions as 0;

[0047] (3) Multiply M by the fully sampled data k1 to obtain the undersampled data k2, and perform an inverse Fourier transform on k2 to obtain the undersampled image data x2. Use x2 as the input of the reconstruction network during training, and x1 as the image ground truth for pre-training the reconstruction network, and perform the training of the pre-training reconstruction network.

[0048] Step 3: Construct the environment and decision network of the reinforcement learning system based on the pre-trained reconstruction network obtained in Step 2;

[0049] The environment of the reinforcement learning system constructed in Step 3 is as follows:

[0050] Under the framework of reinforcement learning, construct an interactive environment for learning radial sampling trajectories. The initial state of the environment is the state of only collecting spokes at 45°. The environment can evaluate the improvement degree of the current action decision (the position index of a certain spoke) based on the quality of the reconstructed image (the reconstructed image is obtained through the pre-trained reconstruction network in Step 2), and can transfer the current undersampled data state combined with the sampling action to the next undersampled data state. The terminal state is the state with |N2 / N1| spokes. The improvement degree of the image reconstruction quality is called the reward. The reward R includes three evaluation indicators: structural similarity, peak signal-to-noise ratio, and normalized mean square error, and the importance of the three indicators is integrated through parameters. The calculation method is R = p1*S + p2*P + p3*N, where S is the change in structural similarity, P is the change in peak signal-to-noise ratio, N is the change in normalized mean square error, and p, p1, p2, p3 are the importance integration parameters respectively. p is a preset value, p1 = the order of magnitude where P is located / the order of magnitude where S is located, p2 = p, p3 = -1 * the order of magnitude where P is located / the order of magnitude where N is located.

[0051] The decision network constructed in Step 3 is as follows:

[0052] The decision-making network includes a feature extraction module and an action value prediction module. The feature extraction module is based on a convolutional neural network, and the action value prediction module is based on a fully connected network. The input of the feature extraction module is the reconstructed image calculated by the pre-trained reconstruction network obtained in step 2, and the output is the feature map of the reconstructed image. The input of the action value prediction module is the one-dimensional expansion of the feature map, and the output is the predicted values corresponding to different actions. The predicted value is called the Q value, and in deep reinforcement learning, the Q value represents the cumulative discounted expectation of the reward. During the training process, the Q value is combined with the temporal difference calculation value and the current reward value (taking time t as an example, Q t =R t +Q t+1 ). The current reward value is mapped to Qtarget in a temporal difference manner. The Smooth L1 loss is calculated between Qtarget and the Q value predicted by the decision-making network and fed back to each layer of the decision-making network to optimize the parameters of the decision-making network. After the training is completed, the network has the ability to make sampling decisions based on the current undersampled reconstruction data.

[0053] Figure 2 FIG. is a schematic diagram of the decision-making network based on a convolutional neural network according to the present invention, which is divided into a feature extraction module and an action decision module. The left part is the feature extraction module (consisting of m1 convolutional layers, and each convolutional layer includes a convolutional kernel, a non-linear activation function, and a pooling layer), and the right part is the action decision module (consisting of an input layer, m2 hidden layers, and an output layer). After the data is input into the feature extraction module, features are extracted and the features are expanded. After the expansion, the action decision module is used to estimate the value of each action.

[0054] x, y, z... represent the convolutional depth, and h, j... represent the hidden layers of the fully connected network.

[0055] Step 4: Train the decision-making network constructed in step 3 to obtain a trained decision-making network;

[0056] The specific method of step 4 is as follows:

[0057] Use a learning rate of l, a training number of ln. The training of the decision-making network uses the replay unit technology and the target network parameter freezing technology, and the capacity of the replay unit is lw;

[0058] In this embodiment, step 4 uses the established data set, the pre-trained reconstruction network, and the interactive environment for learning the radial sampling trajectory to train the constructed decision-making network. The learning rate used during training is 10 -4 , the training number is 10 6 , and the capacity of the replay unit is 4000. According to this, the decision-making network is configured to be trained to obtain a convergent model.

[0059] Step 5: Test the trained decision network obtained in Step 4 to verify the network effect.

[0060] Step 6: Integrate the trained sampling trajectory decision model into the hardware device, and determine the radial slice adaptive sampling trajectory during the actual acquisition process in combination with the magnetic resonance device.

[0061] Specifically, when the magnetic resonance device performs radial k-space data acquisition, initial data acquisition is performed (acquiring 45° spokes), and based on the initial undersampled k-space data, the reconstructed image state is obtained using the pre-trained reconstruction network; then the reconstructed image state is input into the decision network for forward propagation to complete the current sampling decision; finally, the decision is looped until the specified acceleration ratio is met (reaching the termination state), and high-performance k-space active undersampling adaptive to the radial slice can be achieved.

[0062] The following gives a comparison table of the results of the method of the present invention after training and the existing methods in the field. Among them, Pineda represents the method of Pineda et al. (Pineda L, Basu S, Romero A, et al. Active MR k-space sampling with reinforcement learning[C]. Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, 2020: 23-33.). The table is divided into comparisons of the structural similarity SSIM, peak signal-to-noise ratio PSNR, and normalized mean square error NMSE. The higher the SSIM and PSNR, the better, and the lower the NMSE, the better.

[0063] The table is divided into comparison results of 4-fold acceleration ratio and 8-fold acceleration ratio. At 4-fold acceleration ratio, the SSIM of the present invention is 3.18% higher than that of the Pineda method, the PSNR is 1.23 dB higher than that of the Pineda method, and the NMSE is 0.77% lower than that of the Pineda method; at 8-fold acceleration ratio, the SSIM of the present invention is 1.82% higher than that of the Pineda method, the PSNR is 0.65 dB higher than that of the Pineda method, and the NMSE is 0.69% lower than that of the Pineda method; the present invention is superior to the Pineda method in both acceleration ratios and three evaluation indicators, and obtains higher image reconstruction quality.

[0064]

[0065] The innovation of the present invention lies in:

[0066] 1. The present invention proposes a slice adaptive determination method for a radial sampling trajectory in magnetic resonance imaging. By obtaining an optimized radial undersampling trajectory, the quality of magnetic resonance imaging can be improved. The partially observable Markov decision process for the radial adaptive sampling trajectory determination problem is modeled as follows: State space: The state space consists of historical spoke information and original k-space data; Action space: The action space consists of the position information of all spokes; Observation space: The undersampled image obtained by the pre-trained reconstruction network based on the data in the state space, and the undersampled reconstructed image is used as the observation value; State transition: The state transition process is to update the data at the specified spoke position according to the action.

[0067] 2. The training method of the pre-trained reconstruction network for the specified acceleration ratio radial sampling trajectory adopted by the present invention is adjusted according to the preset acceleration ratio, and the training method is set according to the target of the acceleration ratio. For example, when the acceleration ratio target is N1 times, randomly select N2 / N1 pieces of data from each radial k-space data as the undersampled training data, perform inverse Fourier transform on the undersampled training data to obtain the undersampled image data, and input it into the reconstruction network.

[0068] 3. Multi-objective reward function: The reward function includes three evaluation indicators, structural similarity, peak signal-to-noise ratio, and normalized mean square error, and the importance of the three indicators is integrated through parameters. The reward function R includes three evaluation indicators, structural similarity, peak signal-to-noise ratio, and normalized mean square error, and the importance of the three indicators is integrated through parameters. The calculation method is R = p1*S + p2*P + p3*N, where S is the change in structural similarity, P is the change in peak signal-to-noise ratio, N is the change in normalized mean square error, and p, p1, p2, p3 are the importance integration parameters respectively. p is a preset value, p1 = the order of magnitude where P is located / the order of magnitude where S is located, p2 = p, p3 = -1 * the order of magnitude where P is located / the order of magnitude where N is located.

[0069] 4. The lightweight decision network based on convolutional neural network proposed by the present invention has a convolutional depth of less than 4 layers and a fully connected hidden layer of less than 3 layers.

[0070] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0072] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

Claims

1. A slice adaptive determination method for radial sampling trajectories in magnetic resonance imaging, characterized in that: It includes the following steps: Step 1: Obtain magnetic resonance data and construct training and test data for the deep reinforcement learning model; Step 2: Perform pre-training of the reconstruction network based on the training data obtained in Step 1; Step 3: Construct the environment and decision network of the reinforcement learning system based on the pre-trained reconstruction network obtained in Step 2, and then achieve high-performance k-space active undersampling adaptive to radial slices; Step 4: Train the decision network constructed in Step 3 to obtain a trained decision network; Step 5: Test the trained decision network obtained in Step 4 to verify the network effect; Step 6: Integrate the trained sampling trajectory decision model into the hardware device, and combine with the magnetic resonance device to determine the radial slice adaptive sampling trajectory during the actual acquisition process; The specific method of Step 2 includes: (1) Construct a k-space undersampling mask M according to k1. The shape of M is the same as that of k1, and all data points in M are set to 0; (2) Let the target acceleration ratio be N1 times, and N2 be the total number of radial spokes contained in a single fully sampled data k1. Randomly select |N2 / N1| spokes for each fully sampled data k1, and according to the data positions of the selected spokes, assign 1 to the corresponding positions in M, and the remaining positions are 0; (3) Multiply M by the fully sampled data k1 to obtain the undersampled data k2, perform inverse Fourier transform on k2 to obtain the undersampled image data x2, use x2 as the input of the reconstruction network during the training process, and use x1 as the image ground truth for the pre-trained reconstruction network to train the pre-trained reconstruction network; The environment of the reinforcement learning system constructed in Step 3 is as follows: Under the framework of reinforcement learning, construct an interactive environment for radial sampling trajectory learning. The initial state of the environment is the state of only collecting spokes at 45°; The environment of the reinforcement learning system can evaluate the current action decision based on the improvement degree of the reconstructed image quality, and can transfer the current undersampled data state combined with the sampling action to the next undersampled data state. The termination state is the state with |N2 / N1| spokes; The decision network constructed in Step 3 is as follows: The decision network includes a feature extraction module and an action value prediction module; the feature extraction module is connected to the action value prediction module. The feature extraction module is based on a convolutional neural network, and the action value prediction module is based on a fully connected network; the input of the feature extraction module is the reconstructed image calculated by the pre-trained reconstruction network obtained in Step 2, and the output is the feature map of the reconstructed image; the input of the action value prediction module is the one-dimensional expansion of the feature map, and the output is the predicted value corresponding to different actions.

2. The slice adaptive determination method for a radial sampling trajectory for magnetic resonance imaging according to claim 1, wherein: The specific method of Step 1 is: Collect simulated magnetic resonance data. The fully sampled data k1 collected is stored in the k-space. The k-space is the frequency domain space for storing magnetic resonance data. Subsequently, perform inverse Fourier transform to generate the corresponding fully sampled image data x1, store the above two kinds of data, and construct relevant training and test data for the deep reinforcement learning model.

3. A method for adaptively determining slices for a radial sampling trajectory in magnetic resonance imaging according to claim 1, characterized in that: The degree of improvement in the image reconstruction quality is called the reward. The reward R includes three evaluation indicators: structural similarity, peak signal-to-noise ratio, and normalized mean square error. The importance of the three indicators is integrated through parameters, and the calculation method of this integration is R＝p1*S+p2*P+p3*N , where S is the change in structural similarity, P is the change in peak signal-to-noise ratio, N is the change in normalized mean square error, and p, p1, p2, and p3 are importance integration parameters respectively. p is a preset value, p1 = the order of magnitude where P is located / the order of magnitude where S is located, p2 = p, and p3 = -1 * the order of magnitude where P is located / the order of magnitude where N is located.

4. A slice adaptive determination method for a radial sampling trajectory for magnetic resonance imaging according to claim 1, wherein: The predicted value is called the Q value, and the Q value represents the cumulative discounted expectation of the reward in deep reinforcement learning; During the training process, the Q value combines the temporal difference calculation value and the current reward value. The current reward value is mapped to Qtarget in a temporal difference manner. The Smooth L1 loss is calculated between Qtarget and the Q value predicted by the decision network and fed back to each layer of the decision network to optimize the parameters of the decision network. After training, the network has the ability to make sampling decisions based on the current undersampled reconstruction data.

5. A slice adaptive determination method for a radial sampling trajectory for magnetic resonance imaging according to claim 1, characterized in that: The specific method of step 4 is as follows: Using a learning rate of l, a training count of ln, the decision network is trained using the replay unit technique and the target network parameter freezing technique, and the capacity of the replay unit is lw.

Images