CN-115455609-B - Magnetic resonance fingerprint sequence parameter optimization design method based on deep learning

CN115455609BCN 115455609 BCN115455609 BCN 115455609BCN-115455609-B

Abstract

A magnetic resonance fingerprint sequence parameter optimization design method based on deep learning belongs to the field of magnetic resonance fingerprint imaging and is used for solving the problems that the association relation between a pulse sequence and quantitative imaging precision is unknown and the pulse sequence parameter solving is complex. The method comprises the technical key points of constructing a pulse sequence parameter optimization model related to pulse sequence and quantitative imaging precision, constructing a pulse sequence parameter generation network based on a magnetic resonance pulse excitation physical principle, adopting a Gaussian random sequence with the same scale as the pulse sequence as input of the network, setting iteration convergence conditions according to the proposed pulse sequence parameter optimization model, inputting the constructed input sequence into the network, generating optimized pulse sequence parameters through the constructed sequence parameter generation network, evaluating the pulse sequence parameters generated by the network according to the constructed pulse sequence parameter optimization model, judging whether the set iteration convergence conditions are met or not according to the generated optimized sequence parameter evaluation result, and stopping iteration and outputting the generated pulse sequence parameters if the set iteration convergence conditions are met. The invention can be used for the optimal design of the magnetic resonance fingerprint pulse sequence parameters, and can greatly improve the parameter estimation precision of magnetic resonance fingerprint imaging under the condition of not increasing other expenses.

Inventors

HU YUE
LI PENG

Assignees

哈尔滨工业大学

Dates

Publication Date: 20260505
Application Date: 20221006

Claims (7)

1. The magnetic resonance fingerprint sequence parameter optimization design method based on deep learning is characterized by comprising the following steps of: step one, constructing a pulse sequence parameter optimization model according to a magnetic resonance fingerprint imaging model, wherein the model comprises a dictionary partitionability optimal term and a parameter estimation optimal term; step two, constructing a neural network based on physical characteristics of a magnetic resonance fingerprint imaging pulse sequence; Constructing a neural network input, namely adopting a Gaussian random sequence with the same scale as the pulse sequence as the input of the network; setting sequence parameter optimization convergence conditions; inputting the input sequence into a network, and optimizing parameters of the magnetic resonance fingerprint pulse sequence; step six, evaluating pulse sequence parameters after network optimization by utilizing the optimization model defined in the step one; Step seven, judging whether an iteration convergence condition is reached, if so, outputting the finally optimized magnetic resonance fingerprint pulse sequence parameters, otherwise, reversely propagating and correcting network parameters according to an optimization model, returning to the step five, and continuing iteration; Step one, constructing a pulse sequence parameter optimization model related to the pulse sequence and the quantitative imaging precision, wherein the pulse sequence parameter optimization model comprises the following steps: Wherein, the Representing a weight focus matrix of the image, Representing an organized fingerprint dictionary normalized by item, Representing the unit diagonal matrix of the cell, Representing the super-parameters that can be adjusted, Representing the normalized magnetic resonance fingerprint data, Representing a linear degeneration operator, introducing undersampling artifacts and noise.
2. The magnetic resonance fingerprint sequence parameter optimization design method based on deep learning as claimed in claim 1, wherein the neural network is constructed based on physical characteristics of a magnetic resonance fingerprint imaging pulse sequence in the second step, the construction process is specifically that a sequence parameter generating unit is constructed based on a cyclic gating neural unit and a smoothness constraint module, a pulse sequence parameter generating network is formed by connecting a plurality of pulse sequence parameter generating units in series, the constructed network can fully utilize correlation information among pulse sequence contexts, and the generated pulse sequence parameters meet physical requirements of magnetic resonance pulse excitation.
3. The optimal design method for the parameters of the magnetic resonance fingerprint sequence based on the deep learning as set forth in claim 2, wherein the input of the pulse sequence parameter generating network in the third step is Representing gaussian white noise of the same scale as the pulse sequence to be generated.
4. The method for optimizing design of magnetic resonance fingerprint sequence parameters based on deep learning as set forth in claim 1, wherein the defining the relative variation of sequence parameter performance according to the proposed pulse sequence parameter optimization model is as follows: setting the iteration convergence condition as 。
5. The method for optimizing design of magnetic resonance fingerprint sequence parameters based on deep learning as set forth in claim 1, wherein the step five is to input the constructed input sequence into a network, and the constructed sequence parameter generating network generates optimized pulse sequence parameters: Wherein, the The input sequence of the construct is represented, Representing the constructed pulse sequence parameter generation network, Representing the optimized pulse sequence parameters generated by the network, Representing parameters of the network.
6. The magnetic resonance fingerprint sequence parameter optimization design method based on deep learning as claimed in claim 1, wherein step six evaluates the pulse sequence parameters generated by the network according to the constructed pulse sequence parameter optimization model.
7. The magnetic resonance fingerprint sequence parameter optimization design method based on deep learning as claimed in claim 1, wherein step seven judges whether the set iteration convergence condition is reached according to the generated optimized sequence parameter evaluation result, if yes, the iteration is stopped and the generated pulse sequence parameter is output, otherwise, the step five is returned, the iteration is continued, and a new pulse sequence parameter is generated.

Description

Magnetic resonance fingerprint sequence parameter optimization design method based on deep learning Technical Field The invention relates to the technical field of magnetic resonance fingerprint imaging, in particular to a magnetic resonance fingerprint sequence parameter optimization design method based on deep learning. Background The magnetic resonance quantitative imaging performs quantitative imaging by measuring tissue parameters such as Proton Density (PD), spin lattice relaxation time (T1) and spin relaxation time (T2), so that more accurate tissue characteristic information can be provided, subjectivity of diagnosis is reduced, and accurate disease diagnosis and tracking are realized. The magnetic resonance fingerprint imaging technology [1] is used as a brand new magnetic resonance quantitative imaging scheme, and the defects of long imaging time, difficult multi-parameter imaging and the like in the traditional quantitative imaging scheme are overcome through a brand new designed data acquisition and post-processing scheme, so that a foundation is laid for further clinical application of the quantitative imaging technology. In the magnetic resonance fingerprint imaging technology, a pulse scanning sequence which changes randomly or pseudo-randomly is often adopted, so that the response signals of various tissues of a human body have unique response signal evolution, and are often called tissue magnetic resonance fingerprint signals. Meanwhile, a fingerprint dictionary containing theoretical magnetic resonance signals of all possible human tissues is constructed according to a Bloch model excited by the magnetic resonance signals, and then the collected tissue magnetic resonance fingerprint signals are matched with entries in the fingerprint dictionary based on a pattern matching method, so that simultaneous quantitative imaging of multiple tissue parameters is realized. The pulse sequence directly influences the evolution of the tissue magnetic resonance fingerprint signal and plays a decisive role in imaging quality, but the influence mechanism of the pulse sequence on the quantitative imaging effect is not clear. The existing method generates a pulse sequence according to experience, and lacks a theoretical basis of a correlation model between the pulse sequence and quantitative imaging precision. In addition, in order to ensure the precision of pattern matching, fingerprint signals of longer time frames need to be acquired during magnetic resonance fingerprint imaging, so that the pulse sequence length is longer (> 1000), the number of parameters is more, and the complexity of solving the parameters of the magnetic resonance fingerprint pulse sequence is high. Therefore, a correlation model of pulse sequences and quantitative imaging precision is required to be constructed, and a magnetic resonance fingerprint sequence parameter optimization design method with high calculation efficiency is designed based on the correlation model. Disclosure of Invention The invention provides a magnetic resonance fingerprint sequence parameter optimization design method based on deep learning, which is used for solving the problems that the association relation between a pulse sequence and quantitative imaging precision is unknown and the pulse sequence parameter solving is complex. The invention adopts the technical scheme for solving the problems: A magnetic resonance fingerprint sequence parameter optimization design method based on deep learning comprises the following steps: step one, constructing a pulse sequence parameter optimization model according to a magnetic resonance fingerprint imaging model, wherein the model comprises a dictionary partitionability optimal term and a parameter estimation optimal term; step two, constructing a neural network based on physical characteristics of a magnetic resonance fingerprint imaging pulse sequence; Constructing a neural network input, namely adopting a Gaussian random sequence with the same scale as the pulse sequence as the input of the network; setting sequence parameter optimization convergence conditions; inputting the input sequence into a network, and optimizing parameters of the magnetic resonance fingerprint pulse sequence; step six, evaluating pulse sequence parameters after network optimization by utilizing the optimization model defined in the step one; And step seven, judging whether an iteration convergence condition is reached, if so, outputting the finally optimized magnetic resonance fingerprint pulse sequence parameters, otherwise, correcting the network parameters according to the back propagation of the optimization model, and returning to the step five to continue iteration. Further, the pulse sequence parameter optimization model constructed in the first step is as follows: Wherein W represents a weight focus matrix, Represents an organized fingerprint dictionary normalized by item, I represents a unit diagonal matrix, lambda >0 represent