A Deep Learning-Based Magnetic Resonance Fingerprint Sequence Parameter Optimization Design Method

Abstract

A deep learning-based method for optimizing magnetic resonance fingerprint sequence parameters, belonging to the field of magnetic resonance fingerprint imaging, is proposed to address the problems of unclear correlation between pulse sequences and quantitative imaging accuracy, and the complexity of pulse sequence parameter calculation. The key technical points of this invention include: constructing a pulse sequence parameter optimization model that correlates the pulse sequence with quantitative imaging accuracy; constructing a pulse sequence parameter generation network based on the physical principles of magnetic resonance pulse excitation; using a Gaussian random sequence with the same scale as the pulse sequence as the network input; setting iterative convergence conditions according to the proposed pulse sequence parameter optimization model; inputting the constructed input sequence into the network, which then generates optimized pulse sequence parameters; evaluating the pulse sequence parameters generated by the network according to the constructed pulse sequence parameter optimization model; and determining whether the set iterative convergence conditions have been met based on the evaluation results of the optimized sequence parameters. If so, the iteration stops and the generated pulse sequence parameters are output. This invention can be used for optimizing the design of magnetic resonance fingerprint pulse sequence parameters, significantly improving the parameter estimation accuracy of magnetic resonance fingerprint imaging without increasing other overhead.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance fingerprint imaging technology, and more specifically to a deep learning-based method for optimizing magnetic resonance fingerprint sequence parameters. Background Technology

[0002] Magnetic resonance imaging (MRI) quantifies tissue parameters such as proton density (PD), spin lattice relaxation time (T1), and spin-spin relaxation time (T2) to provide more accurate information on tissue characteristics, reduce diagnostic subjectivity, and achieve precise disease diagnosis and tracking. MRI fingerprinting technology. [1] As a novel magnetic resonance imaging (MRI) quantitative imaging scheme, this new approach overcomes the shortcomings of traditional MRI schemes, such as long imaging time and difficulties in multi-parameter imaging, through a newly designed data acquisition and post-processing scheme, laying the foundation for further clinical applications of MRI technology. In MRI fingerprinting, randomly or pseudo-randomly varying pulse scanning sequences are often used, resulting in unique response signal evolutions for different human tissues, commonly referred to as tissue MRI fingerprint signals. Simultaneously, a fingerprint dictionary containing theoretical MRI signals of all possible human tissues is constructed based on the Bloch model excited by the MRI signal. Then, based on pattern matching methods, the acquired tissue MRI fingerprint signals are matched with entries in the fingerprint dictionary, achieving simultaneous quantitative imaging of multiple tissue parameters. The pulse sequence directly affects the evolution of the tissue MRI fingerprint signal and plays a decisive role in imaging quality; however, the mechanism by which the pulse sequence affects quantitative imaging results remains unclear. Existing methods mostly rely on experience to generate pulse sequences, lacking a theoretical basis for a correlation model between pulse sequences and quantitative imaging accuracy. In addition, in order to ensure the accuracy of pattern matching, fingerprint signals of a relatively long time frame need to be acquired during magnetic resonance fingerprinting. Therefore, the pulse sequence length is long (>1000) and contains a large number of parameters, resulting in high complexity in solving the pulse sequence parameters of magnetic resonance fingerprinting.

[0003] Therefore, it is necessary to construct a correlation model between pulse sequence and quantitative imaging accuracy, and based on this correlation model, design a computationally efficient magnetic resonance fingerprint sequence parameter optimization method. Summary of the Invention

[0004] This invention proposes a deep learning-based magnetic resonance fingerprint sequence parameter optimization design method to solve the problems of unclear correlation between pulse sequence and quantitative imaging accuracy, and the complexity of pulse sequence parameter solving.

[0005] The technical solution adopted by the present invention to solve the above problems is as follows:

[0006] A deep learning-based method for optimizing magnetic resonance fingerprint sequence parameters includes the following steps:

[0007] Step 1: Construct a pulse sequence parameter optimization model based on the magnetic resonance fingerprinting model. This model includes two parts: the dictionary separability optimal term and the parameter estimation optimal term.

[0008] Step 2: Construct a neural network based on the physical characteristics of magnetic resonance fingerprinting pulse sequences;

[0009] Step 3: Construct the neural network input: Use a Gaussian random sequence with the same scale as the pulse sequence as the network input;

[0010] Step 4: Set sequence parameters to optimize convergence conditions;

[0011] Step 5: Input the input sequence into the network and optimize the magnetic resonance fingerprint pulse sequence parameters;

[0012] Step 6: Evaluate the pulse sequence parameters after network optimization using the optimization model defined in Step 1;

[0013] Step 7: Determine whether the iterative convergence condition has been met. If yes, output the final optimized magnetic resonance fingerprint pulse sequence parameters; otherwise, correct the network parameters according to the backpropagation of the optimization model, and return to Step 5 to continue iterating.

[0014] Furthermore, the pulse sequence parameter optimization model constructed in step one is as follows:

[0015]

[0016] Where W represents the weight focusing matrix, Let I represent the organization fingerprint dictionary after entry-by-entry normalization, where I represents the unit diagonal matrix, and λ>0 indicates an adjustable hyperparameter. This represents the normalized magnetic resonance fingerprint data. This represents a linear degradation operator that introduces undersampling artifacts and noise.

[0017] Furthermore, the neural network constructed in step two based on the physical characteristics of magnetic resonance fingerprinting pulse sequences is specifically constructed as follows: a sequence parameter generation unit is constructed based on recurrent gated neural units and a smoothing constraint module, and several pulse sequence parameter generation units are connected in series to form a pulse sequence parameter generation network. The constructed network can fully utilize the correlation information between pulse sequence contexts and ensure that the generated pulse sequence parameters meet the physical requirements of magnetic resonance pulse excitation.

[0018] Furthermore, the input to the pulse sequence parameter generation network in step three is... This represents Gaussian white noise with the same scale as the pulse sequence to be generated.

[0019] Furthermore, in step four, based on the proposed pulse sequence parameter optimization model, the relative change in sequence parameter performance is defined as: ∈n =cost n / cost n-1 Let the convergence condition of the iteration be ∈ n <1e -5 .

[0020] Furthermore, in step five, the constructed input sequence is input into the network, and the constructed sequence parameter generation network generates optimized pulse sequence parameters:

[0021]

[0022] in, f represents the input sequence to be constructed. net (·) represents the constructed pulse sequence parameter generation network. This represents the optimized pulse sequence parameters generated by the network. These represent the parameters of the network.

[0023] Further, in step six, the pulse sequence parameters generated by the network are evaluated according to the constructed pulse sequence parameter optimization model.

[0024] Further, in step seven, based on the evaluation results of the generated optimized sequence parameters, it is determined whether the set iterative convergence condition has been met. If yes, the iteration stops and the generated pulse sequence parameters are output; otherwise, the process returns to step five to continue iterating and generate new pulse sequence parameters.

[0025] The present invention has the following beneficial technical effects:

[0026] This invention presents a deep learning-based method for optimizing magnetic resonance fingerprint sequence parameters. It constructs a sequence parameter optimization model that correlates pulse sequences with quantitative imaging accuracy, fully considering the impact of noise and undersampling on quantitative imaging accuracy, and quantitatively evaluating the performance of pulse sequence parameters. Furthermore, based on the excitation principle of magnetic resonance signals, a pulse sequence parameter generation network is proposed, fully utilizing the contextual information of the sequence parameters and employing parallel computing to improve the optimization efficiency. Using the optimized pulse sequence for magnetic resonance fingerprint imaging can significantly improve the parameter estimation accuracy without increasing additional overhead.

[0027] This invention addresses the problems of unclear correlation between pulse sequences and quantitative imaging accuracy, and the complexity of solving for pulse sequence parameters. The key technical aspects of this invention include: constructing a pulse sequence parameter optimization model that correlates pulse sequences with quantitative imaging accuracy; building a pulse sequence parameter generation network based on the physical principles of magnetic resonance pulse excitation; using a Gaussian random sequence with the same scale as the pulse sequence as the network input; setting iterative convergence conditions according to the proposed pulse sequence parameter optimization model; inputting the constructed input sequence into the network, which then generates optimized pulse sequence parameters; evaluating the pulse sequence parameters generated by the network according to the constructed pulse sequence parameter optimization model; and determining whether the set iterative convergence conditions have been met based on the evaluation results of the optimized sequence parameters. If so, the iteration stops and the generated pulse sequence parameters are output. This invention can be used for the optimized design of magnetic resonance fingerprint pulse sequence parameters. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This is a schematic diagram of the pulse sequence parameter generation network structure proposed in this invention.

[0030] Figure 3 This diagram illustrates the comparison between the optimized magnetic resonance fingerprint pulse sequence parameters generated using the method proposed in this invention and the original sequence parameters. "Original" represents the original sequence parameters, and "Ours" represents the optimized sequence parameters obtained using this invention. The sequence length is 400.

[0031] Figure 4 This is a schematic diagram illustrating the correlation between the tissue fingerprint dictionaries generated using the optimized sequence parameters and the original sequence parameters of this invention. The correlation was visualized to intuitively show that the darker the color, the lower the correlation. The optimized sequence parameters of this invention result in the lowest correlation of tissue fingerprint signals, which also means the optimal separability.

[0032] Figure 5 Experimental results demonstrating the quantitative imaging performance verification using the optimized sequence parameters of this invention and the original sequence are presented, showcasing imaging results for three parameter maps under the same conditions: longitudinal relaxation time T1, lateral relaxation time T2, and proton density PD. The experimental results show that the pulse sequence optimized by the method proposed in this invention can effectively improve the accuracy of magnetic resonance fingerprint imaging. Detailed Implementation

[0033] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0034] The magnetic resonance fingerprinting process mainly consists of the following three steps:

[0035] (1) Given a magnetic resonance fingerprint pulse sequence and its parameters, the magnetic resonance fingerprint is obtained by running the pulse sequence through a magnetic resonance scanning device.

[0036] Resonant fingerprint data X;

[0037] (2) Based on the pulse sequence and its parameters, a fingerprint signal dictionary D is generated using computer simulation;

[0038] (3) Match the fingerprint data X with the fingerprint signal dictionary D to reconstruct the quantitative image of magnetic resonance parameters.

[0039] The purpose of this invention is to overcome the problems of unclear correlation between pulse sequence and quantitative imaging accuracy in step (1) above, and the complexity of solving pulse sequence parameters. It proposes a deep learning-based magnetic resonance fingerprint sequence parameter optimization design method, which can efficiently optimize and generate pulse sequence parameters and significantly improve the parameter estimation accuracy of magnetic resonance fingerprint imaging without increasing other overhead.

[0040] Figure 1 A schematic flowchart illustrating a deep learning-based magnetic resonance fingerprint sequence parameter optimization design method according to an embodiment of the present invention is shown. Figure 1 As shown, the specific implementation steps of the present invention include:

[0041] Step 1: Construct a pulse sequence parameter optimization model based on the magnetic resonance fingerprinting model. This model includes two parts: the dictionary separability optimal term and the parameter estimation optimal term.

[0042] According to an embodiment of the present invention, the acquisition of magnetic resonance fingerprint data is represented as follows:

[0043]

[0044] in, N represents the acquired undersampled frequency domain (k-space) data. c N represents the number of coils. s This indicates the number of k-space data points acquired by a single coil. This represents a linear degradation operator that incorporates coil sensitivity and undersampling. Let represent the distortion-free frequency-domain magnetic resonance fingerprint data to be recovered, and ... X i，: (1≤i≤N x N y The response signal of an organization at a specific location can be represented as: [2] :

[0045] Xi，: =ρ i B(η i ,θ)i=1,···,N x N y

[0046] in, Let B(·) denote the proton density (PD), B(·) denote the Bloch equation, and η represent the proton density (PD). i This represents a parameter vector consisting of multiple tissue parameters, typically including longitudinal relaxation time T1, lateral relaxation time T2, etc., while θ represents the parameters of the magnetic resonance fingerprint sequence, including repetition time (TR), echo time (TE), and flip angle (FA), etc. Similarly, the pre-constructed tissue fingerprint dictionary in magnetic resonance fingerprinting can be represented as:

[0047] D = [d m，: ], d m，: =B(η) m ,θ)

[0048] Where D represents the constructed organizational fingerprint dictionary, and its entries d m，: m = 1, ..., M represents the tissue parameter η m The theoretical response signal is given below, and the response parameter lookup table is LUT = [η]. m The dictionary entries are recorded, showing the correspondence between them and their corresponding organizational parameters. Therefore, the magnetic resonance fingerprint parameter matching process can be represented as:

[0049]

[0050]

[0051] Where k i This represents the index of the best-matching entry in the dictionary. and This indicates the estimated organizational parameters.

[0052] The constructed magnetic resonance fingerprint pulse sequence parameter optimization model consists of two parts: the optimal term for separability and the optimal term for parameter estimation.

[0053] 1) Optimal term of separability

[0054] Based on the parameter matching model described above, the separability between tissue fingerprint signals can be improved by minimizing the correlation between tissue fingerprint dictionary entries:

[0055]

[0056] If the dictionary is normalized entry by entry, the above formula can be improved in efficiency using matrix operations:

[0057]

[0058] in Let I represent the dictionary after entry-by-entry normalization, and let I represent the unit diagonal matrix. Equation (7) applies the same optimization weights to all entries. This invention introduces a weight focusing matrix to improve the ability of the generated sequence to distinguish fingerprint signals of tissues with similar parameters:

[0059]

[0060] Where W represents the weight focusing matrix, which is updated during optimization iterations:

[0061]

[0062] in, This represents the tissue fingerprint dictionary generated based on the optimized pulse sequence parameters after the previous iteration. The weighted focusing matrix applies higher weight coefficients to tissues with similar parameters, which can improve the ability of the optimized pulse sequence to distinguish tissues with similar parameters.

[0063] 2) Optimal Parameter Estimation Term

[0064] The parameter estimation process for magnetic resonance fingerprinting can be rewritten using matrix operations as follows:

[0065] X = RD

[0066] in Let represent the matching matrix, where each row is a sparse vector. If all data is normalized, matrix operations can be used to solve for the matching matrix to improve efficiency.

[0067]

[0068] in and The operator argmax represents the normalized data. dim＝2 (·) denotes finding the index of the maximum value along the second dimension of the matrix. The matching matrix solution process described above simplifies the parameter estimation process by ignoring the proton density, thus improving the solution efficiency. However, the operator argmax... dim＝2 (·) Since it is not differentiable, it is impossible to directly correlate quantitative imaging accuracy with pulse sequence parameters. Based on a further simplification of the magnetic resonance fingerprint parameter estimation model, this invention proposes a differentiable optimal parameter estimation term to improve the quantitative imaging performance of pulse sequences under the influence of noise interference and data undersampling, which can be expressed as:

[0069]

[0070] in The linear degradation operator, based on the data acquisition model, introduces undersampling artifacts and noise. Compared to undersampled and noisy data, fully sampled, distortion-free data achieves the highest accuracy in pattern matching. Therefore, the optimization scheme described in the above equation aims to minimize the impact of undersampled artifacts and noise on quantitative imaging accuracy.

[0071] Combining the optimal terms for dictionary separability and parameter estimation, the constructed parameter optimization model for magnetic resonance fingerprint pulse sequences can be expressed as:

[0072]

[0073] Here, λ>0 represents an adjustable hyperparameter used to balance the optimization terms of the two parts. The separability optimal term increases the separability of response signals from different tissues, improving the accuracy of pattern matching and parameter-based imaging; the parameter estimation optimal term improves the quantitative imaging performance of pulse sequences under noise interference and data undersampling. The proposed pulse sequence parameter optimization model correlates the pulse sequence with quantitative imaging accuracy and fully considers the impact of imaging system defects on imaging accuracy, thus improving the imaging performance of the optimized pulse sequence.

[0074] Step 2: Construct a neural network based on the physical characteristics of magnetic resonance fingerprinting pulse sequences;

[0075] The proposed pulse sequence parameter generation network structure is attached. Figure 2 To fully utilize the correlation information between pulse sequence contexts, this invention proposes a magnetic resonance fingerprint pulse sequence parameter generation network (PSG-Net) based on recurrent gated neural units. The proposed network consists of a series of pulse sequence parameter generation units (PSG Cells) connected in series, with the number of PSG Cells matching the length of the pulse sequence to be generated. The proposed PSG Cells are designed based on gated recurrent neural units (GRU cells) to capture prior features related to the context within the pulse sequence.

[0076] In addition, studies have shown that... [3] Smooth changes in sequence parameters (especially the flip angle FA) result in a smoother evolution of the tissue response signal, which helps improve the robustness of parameter estimation to undersampling and noise. Therefore, this invention introduces a smoothing module into the proposed PSG Cell to ensure smooth parameter changes in the pulse sequence parameters generated by the network. Simultaneously, the smoothing module with a residual-like structure can further improve the utilization of prior information implied by the contextual relevance of the pulse sequence and effectively reduce the learning difficulty of the neural network. The parameters (Ψ) contained in the smoothing module itself are also learnable and can be gradually adjusted as the network is optimized, allowing for finer adjustments to the generated sequence parameters. The proposed PSG Cell can be expressed as:

[0077]

[0078]

[0079] Ψ = [Ψ FA Ψ TR ]

[0080] in This represents the i-th parameter vector of the input pulse sequence. Let represent the i-th parameter vector generated by the optimization of the i-th proposed PSG Cell. The learnable smoothing parameter Ψ is derived from Ψ. FA ∈[0, ψ FA ] and Ψ TR ∈[0, ψ TR It consists of two parts, which apply smoothing constraints to the flip angle FA and the repetition time TR, respectively. The smoothness of the generated sequence can also be effectively adjusted by setting the range of values ​​for the smoothing module parameters. The features are priors extracted by the GRU Cell, and the parameters of the GRU Cell are used... express. This represents the features extracted from the input sequence parameters by the PSG Cell (before the smoothing module).

[0081] Sequence optimization problems lack the real-sample reference values ​​required by traditional deep learning techniques. The proposed PSG-Net's unique design structure fully considers the characteristics of pulse sequences, significantly reducing the difficulty of sequence optimization. Therefore, the proposed network can optimize and generate magnetic resonance fingerprinting pulse sequences in an unsupervised learning mode based on the proposed cost function.

[0082] Step 3: Construct the neural network input: Use a Gaussian random sequence with the same scale as the pulse sequence as the network input;

[0083] According to an embodiment of the present invention, the input to the pulse sequence parameter generation network is This represents Gaussian white noise with the same scale as the pulse sequence to be generated.

[0084] Step 4: Set sequence parameters to optimize convergence conditions;

[0085] Based on the pulse sequence parameter optimization model proposed in Step 1, the relative change in the performance of the sequence parameters evaluated by the optimization model is defined as: ∈ n =cost n / cost n-1 Let the convergence condition of the iteration be ∈ n <1e -5 .

[0086] Step 5: Input the input sequence into the network and optimize the magnetic resonance fingerprint pulse sequence parameters;

[0087] According to an embodiment of the present invention, the network constructed in step two is used to generate optimized pulse sequence parameters:

[0088]

[0089] in, f represents the input sequence constructed in step three. net (·) represents the network constructed in step two. This represents the optimized pulse sequence parameters generated by the network. These represent the parameters of the network.

[0090] Step 6: Evaluate the pulse sequence parameters after network optimization using the optimization model defined in Step 1;

[0091] The pulse sequence parameters generated by the network are evaluated according to the pulse sequence parameter optimization model defined in step one.

[0092] Step 7: Determine whether the iterative convergence condition has been met. If yes, output the final optimized magnetic resonance fingerprint pulse sequence parameters; otherwise, correct the network parameters according to the backpropagation of the optimization model and return to Step 5.

[0093] According to an embodiment of the present invention, in order to quantitatively analyze the effects of the present invention, the NMSE index is used to analyze the experimental results. The quantitative analysis formula for the NMSE index is as follows:

[0094]

[0095] Where, θ and These represent the reference parameter plot and the estimated parameter plot, respectively.

[0096] The method of this invention is compared with the sequence parameters originally used.

[0097] Figure 3 This diagram illustrates the comparison between the optimized magnetic resonance fingerprint pulse sequence parameters generated using the method proposed in this invention and the original sequence parameters. "Original" represents the original sequence parameters, and "Ours" represents the optimized sequence parameters obtained using this invention. The sequence length is 400.

[0098] Figure 4 This is a schematic diagram illustrating the correlation between the tissue fingerprint dictionaries generated using the optimized sequence parameters and the original sequence parameters of this invention. The correlation was visualized to intuitively show that the darker the color, the lower the correlation. The optimized sequence parameters of this invention result in the lowest correlation of tissue fingerprint signals, which also means the optimal separability.

[0099] Figure 5 Experimental results demonstrating the quantitative imaging performance verification using the optimized sequence parameters of this invention and the original sequence are presented, showcasing imaging results for three parameter maps under the same conditions: longitudinal relaxation time T1, lateral relaxation time T2, and proton density PD. The experimental results show that the pulse sequence optimized by the method proposed in this invention can effectively improve the accuracy of magnetic resonance fingerprint imaging.

[0100] The following documents are cited in this invention:

[0101] [1]Ma D,Gulani V,Seiberlich N,et al.Magnetic resonance fingerprinting[J].Nature,2013,495(7440):187-192.

[0102] [2]Davies M,Puy G,Vandergheynst P,et al.A compressed sensingframework for magnetic resonance fingerprinting[J].Siam journal on imagingsciences,2014,7(4):2623-2656.

[0103] [3] Zhao B, Haldar JP, Liao C, et al. Optimal experiment design formagnetic resonance fingerprinting: Cramér-Rao bound meets spin dynamics [J]. IEEE transactions on medical imaging, 2018, 38(3): 844-861.

Claims

1. A deep learning-based method for optimizing magnetic resonance fingerprint sequence parameters, characterized in that, The method includes the following steps: Step 1: Construct a pulse sequence parameter optimization model based on the magnetic resonance fingerprinting model. This model includes two parts: the dictionary separability optimal term and the parameter estimation optimal term. Step 2: Construct a neural network based on the physical characteristics of magnetic resonance fingerprinting pulse sequences; Step 3: Construct the neural network input: Use a Gaussian random sequence with the same scale as the pulse sequence as the network input; Step 4: Set sequence parameters to optimize convergence conditions; Step 5: Input the input sequence into the network and optimize the magnetic resonance fingerprint pulse sequence parameters; Step 6: Evaluate the pulse sequence parameters after network optimization using the optimization model defined in Step 1; Step 7: Determine whether the iterative convergence condition has been met. If yes, output the final optimized magnetic resonance fingerprint pulse sequence parameters; otherwise, correct the network parameters according to the backpropagation of the optimization model, and return to Step 5 to continue iterating. Step 1: Constructing a pulse sequence parameter optimization model that correlates pulse sequence with quantitative imaging accuracy: in, This represents the weight focusing matrix. This represents an organizational fingerprint dictionary that has been normalized item by item. Represents a unit diagonal matrix. This indicates adjustable hyperparameters. This represents the normalized magnetic resonance fingerprint data. This represents a linear degradation operator that introduces undersampling artifacts and noise.

2. The method for optimizing magnetic resonance fingerprint sequence parameters based on deep learning according to claim 1, characterized in that, The neural network constructed based on the physical characteristics of magnetic resonance fingerprinting pulse sequences in step two is constructed as follows: a sequence parameter generation unit is constructed based on a recurrent gated neural unit and a smoothing constraint module. Several pulse sequence parameter generation units are connected in series to form a pulse sequence parameter generation network. The constructed network can make full use of the correlation information between the pulse sequence context and make the generated pulse sequence parameters meet the physical requirements of magnetic resonance pulse excitation.

3. The method for optimizing magnetic resonance fingerprint sequence parameters based on deep learning according to claim 2, characterized in that, The input to the pulse sequence parameter generation network in step three is , representing Gaussian white noise with the same scale as the pulse sequence to be generated.

4. The method for optimizing magnetic resonance fingerprint sequence parameters based on deep learning according to claim 1, characterized in that, Step four: Based on the proposed pulse sequence parameter optimization model, define the relative change in sequence parameter performance as follows: Set the iterative convergence condition as follows: .

5. The method for optimizing magnetic resonance fingerprint sequence parameters based on deep learning according to claim 1, characterized in that, Step 5 involves inputting the constructed input sequence into the network, which then generates optimized pulse sequence parameters through the constructed sequence parameter generation network. in, This represents the input sequence to be constructed. This represents the constructed pulse sequence parameter generation network. This represents the optimized pulse sequence parameters generated by the network. Represents the parameters of the network.

6. The method for optimizing magnetic resonance fingerprint sequence parameters based on deep learning according to claim 1, characterized in that, Step 6: Evaluate the pulse sequence parameters generated by the network according to the constructed pulse sequence parameter optimization model.

7. The method for optimizing magnetic resonance fingerprint sequence parameters based on deep learning according to claim 1, characterized in that, Step 7: Based on the evaluation results of the generated optimized sequence parameters, determine whether the set iterative convergence condition has been met. If yes, stop the iteration and output the generated pulse sequence parameters; otherwise, return to step 5, continue the iteration, and generate new pulse sequence parameters.

Images