A method and system for simultaneous multi-parametric quantitative magnetic resonance imaging based on overlapping echoes

Abstract

The application discloses an overlapping echo-based magnetic resonance multi-parameter simultaneous quantitative imaging method and system. The method comprises the following steps: designing an overlapping echo-based steady-state free precession multi-parameter quantitative imaging sequence; determining sampling parameters of the overlapping echo-based steady-state free precession multi-parameter quantitative imaging sequence; adding the sequence and setting the sampling parameters in an advanced magnetic resonance instrument which meets the performance requirements of the sequence and the sampling parameters, and completing data acquisition to obtain magnetic resonance data; generating training samples of a deep neural network; training the deep neural network by using the training samples to obtain a trained deep neural network; and reconstructing a magnetic resonance multi-parameter image by using the trained deep neural network and k-space data collected by the magnetic resonance imaging sequence. The method can realize simultaneous quantitative imaging of magnetic resonance multi-parameters, and can utilize different echoes and / or different k-space information in the k-space to realize correction of non-ideal factors, including but not limited to artifacts, radio frequency field inhomogeneity, main magnetic field inhomogeneity and the like.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging, and in particular to a method and system for simultaneous quantitative imaging of multiple parameters in magnetic resonance based on overlapping echoes. Background Technology

[0002] Magnetic resonance imaging (MRI) is an imaging technique that reconstructs images by utilizing signals generated when atomic nuclei resonate within a strong magnetic field. MRI can be divided into qualitative parametric-weighted imaging and quantitative parametric imaging. Quantitative parametric imaging (MRI) can directly measure the characteristic parameters of tissues and has been applied to the study of fine structures in the brain, heart, and spinal cord (Zhao B, Lam F, Liang ZP. Model-based MR parameter mapping with sparsity constraints: parameter estimation and performance bounds. IEEE T MedImaging 2014; 33:1832-1844; Garces P, Pereda E, Hernandez-Tamames JA, Del-Pozo F, Maestu F, Pineda-Pardo JA. Multimodal description of whole brain connectivity: a comparison of resting state MEG, fMRI, and DWI. Hum Brain Mapp 2016; 37:20-34). MRI can reduce the influence of factors other than tissue properties on imaging results, such as operator dependence, differences in scanning parameters, spatial variations in the magnetic field, and image scaling (Margaret Cheng HL, Stiko N, Ghugre NR, Wright GA. Practical medical applications of quantitative MR). Relaxometry (J MagnResonImaging 2012; 36:805-824) has important application value.

[0003] Multiparametric quantitative magnetic resonance imaging (MRI) provides more comprehensive insights into the quantitative study of tissue properties, showing great potential in medical diagnosis (Wang L, Gaddam S, Wang N, et al. Multiparametric Mapping Magnetic Resonance Imaging of Pancreatic Disease[J]. Frontiers in Physiology, 2020, 11) and scientific research. However, the long acquisition time required for this process still limits its widespread application, especially for functional / dynamic imaging or motion scenarios. Currently, a series of fast quantitative magnetic resonance imaging methods based on overlapping echoes using EPI readout have been developed (Cai C, Wang C, Zeng Y, et al. Single-shot T2 mapping using overlapping-echo detachment planar imaging and a deep convolutional neural network[J]. Magnetic Resonance in Medicine, 2018, 80(5).;6. Zhang J, Wu J, Chen S, et al. Robust Single-shot T2 Mapping via Multiple Overlapping-Echo Acquisition and Deep Neural Network[J]. IEEE transactions on medical imaging, 2019:1801). However, the acquired magnetic resonance images have shortcomings such as distortion and eddy current artifacts, which increases the difficulty of reconstructing the quantitative magnetic resonance parameter map.

[0004] Balanced steady-state free precession sequences offer a higher signal-to-noise ratio per unit time than other imaging sequences and have been applied to imaging different tissues of the human body, such as the brain, chest cavity, and abdomen. They have high imaging quality and stable performance (Scheffler K, Lehnhardt S. Principles and applications of balanced SSFP techniques[J]. European Radiology, 2003, 13(11): 2409-2418). Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for simultaneous quantitative imaging of multiple parameters of magnetic resonance, which can realize quantitative imaging of multiple parameters of magnetic resonance and can correct for non-ideal factors, including but not limited to artifacts, radio frequency field inhomogeneity, and main magnetic field inhomogeneity.

[0006] This invention provides a method for simultaneous quantitative multi-parameter magnetic resonance imaging based on overlapping echoes, comprising the following steps:

[0007] S1: Design a steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes;

[0008] S2: Determine the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes;

[0009] S3: Add the sequence and set the sampling parameters in an advanced magnetic resonance spectrometer that meets the performance requirements of the sequence and sampling parameters, and complete the data acquisition to obtain magnetic resonance data;

[0010] S4: Generate training samples for the deep neural network based on the sequences and parameters described in S1 to S3;

[0011] S5: Train the deep neural network using training samples to obtain a trained deep neural network;

[0012] S6: Using the trained deep neural network and magnetic resonance data, reconstruct a multi-parameter quantitative image of magnetic resonance;

[0013] In step S1, the steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes includes:

[0014] Several preparatory pulses α during the preparation period;

[0015] Within each TR, the flip angle is α. i The radio frequency excitation pulse and the corresponding applied to α i The shift gradient G on the subsequent frequency coding dimension ROi and applied to α i The subsequent shift gradient G on the phase encoding dimension PEi Where i = 1, 2, ..., n, a total of n radio frequency excitation pulses are applied;

[0016] Simultaneously applying the radio frequency excitation pulse, a layer / segment gradient G with a layer selection dimension is applied. cr Choose different Gs cr This invention can achieve both two-dimensional planar acquisition and three-dimensional volume acquisition.

[0017] Prephase gradient G ppAn equal-magnitude, opposite-direction phase-encoding gradient -G is applied along the phase-encoding dimension to change the k-space filling start point for each scan, with a different value for each scan. At the end of each TR period, a phase-encoding gradient -G is applied. pp According to different TRs G pp The order of values ​​can be used to achieve different k-space filling methods, such as filling the k-space in the order of high and low Cartesian sampling and filling the k-space in a linear order.

[0018] This sequence generates echoes by gradient inversion, first utilizing a dephasing gradient G in the frequency coding direction. dp The magnetization intensity of the atomic nuclei is dephased, and then a readout gradient G1 with opposite polarity is used to reconcile the magnetization intensity. The area of ​​the readout gradient is twice the area of ​​the dephase gradient. Without considering the shift gradient, at the midpoint of the readout gradient, the divergent phase is completely compensated to form the echo peak. However, due to the existence of the shift gradient, the time of forming the echo peak is advanced or delayed, which can make the echo separate in the readout direction.

[0019] Within each TR, m readout gradients are applied in the frequency coding direction, and the j-th readout gradient is represented by G. j This means that each readout gradient corresponds to a k-space j, j = 1, 2, ..., m, and there are a total of m k-spaces;

[0020] In G j Then apply the shift gradient G' on the frequency coding dimension. ROj and the shift gradient G' on the phase encoding dimension PEj .

[0021] In step S2, determining the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes specifically includes:

[0022] Determine the preparation period time tpre;

[0023] Determine the number of preparatory pulses during the preparation period, as well as the size, phase, and pulse shape of each preparatory pulse;

[0024] Determine the number n of radio frequency excitation pulses within each TR;

[0025] Determine the flip angle α of each radio frequency excitation pulse. i The magnitude, phase, and pulse shape;

[0026] Determine the time interval δi of each radio frequency excitation pulse;

[0027] Determine the number m of k-spaces;

[0028] Determine the time Δj for applying the readout gradient;

[0029] Determine the phase gradient G on the frequency coding dimension dp The area and orientation of the readout gradients G j Area and orientation;

[0030] Determine the prephase gradient G on different phase encoding dimensions within the TR. pp The area and orientation of the k-space are used to determine the filling method of the k-space;

[0031] Determine the level selection dimension G cr Area and orientation;

[0032] Determine each shift gradient G ROi G PEi G' ROj G' PEj The area, direction, and application time are used to determine the number of overlapping echoes in each k-space and the position of each overlapping echo in the k-space;

[0033] Determine the pulse sequence repetition time TR;

[0034] Determine other parameters such as imaging field of view, imaging matrix, parallel imaging acceleration factor, 2D / 3D acquisition, etc.

[0035] In step S4, training samples for the deep neural network are generated based on the sequences and parameters described in S1 to S3. A set number of deep neural network training samples are simulated and generated based on the characteristics of the object to be tested, forming a training sample set. A specific feasible solution is given, which can be expanded and reduced in practice. The specific solution includes:

[0036] The pulse sequence is input into the magnetic resonance imaging simulation software, and corresponding non-ideal terms are added according to the non-idealities of the real experiment in order to simulate the real situation as much as possible.

[0037] A set number of random templates are generated based on the characteristics of the test object (or templates are generated using an MRI dataset with the characteristics of the test object), while ensuring that the templates are highly complex and can contain all the characteristics of the experimental samples; the number of templates needs to be sufficient to ensure good reconstruction quality;

[0038] The template was simulated and sampled using simulation software to obtain the template's magnetic resonance signal. During the simulation sampling process, considering changes in the real experimental environment, undesirable factors were added to improve the robustness of the network model to undesirable experimental environments. These undesirable factors included excitation pulse angle deviation, shift gradient deviation, and noise.

[0039] The magnetic resonance echo signals of the templates are rearranged into two-dimensional k-space signals, and then two-dimensional Fourier transform is performed to obtain the magnetic resonance images of each template. The magnetic resonance images are then normalized and noise-added.

[0040] The magnetic resonance images of each template and the corresponding template are used to form a training sample, and a set number of training samples are obtained to form a training sample set.

[0041] The training samples mentioned above can be simulated samples, samples created from real samples, or a combination of real and simulated samples.

[0042] In step S5, training the deep neural network using training samples to obtain a trained deep neural network specifically includes:

[0043] Determine the network structure of the deep neural network;

[0044] Determine the number of input and output channels for the deep neural network;

[0045] Determine the loss function for training the deep neural network;

[0046] When training a deep neural network, the training sample set is input into the deep neural network in batches for iterative training. Each time the network is trained, the value of the loss function is calculated. Based on this value, the parameter values ​​of the neural network are automatically adjusted to reduce the value of the loss function. The above training is repeated until the value of the loss function no longer decreases, and then the parameters of the deep neural network are saved.

[0047] This invention provides a magnetic resonance multi-parameter simultaneous quantitative imaging system based on overlapping echoes, comprising, in sequence: a pulse sequence design module, a signal acquisition module, a signal processing module, a training sample set generation module, a deep neural network determination module, and a multi-parameter quantitative image reconstruction module;

[0048] The pulse sequence design module is used to design steady-state free precession multi-parameter quantitative imaging sequences based on overlapping echoes; and to determine the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequences based on overlapping echoes.

[0049] The signal acquisition module is used to sample the test object using the pulse sequence under set sampling parameters and a high-performance magnetic resonance spectrometer to obtain the overlapping echo magnetic resonance signal of the test object.

[0050] The signal processing module is used to perform two-dimensional Fourier transform, normalization and other processing on the overlapping echo magnetic resonance signal of the object under test to obtain the magnetic resonance image to be reconstructed of the object under test.

[0051] The training sample set generation module is used to simulate and generate a set amount of deep neural network training samples based on the characteristics of the object to be tested, thus forming a training sample set.

[0052] A deep neural network determination module is used to determine a deep neural network for multi-parameter quantitative image reconstruction; the deep neural network is trained using the training sample set to obtain a trained deep neural network;

[0053] The multi-parameter quantitative image reconstruction module is used to input the overlapping echo magnetic resonance image obtained from the magnetic resonance instrument into the trained deep neural network for reconstruction, so as to obtain a multi-parameter quantitative image of the test object.

[0054] This invention, based on a steady-state free precession magnetic resonance imaging (MRI) sequence with overlapping echoes, can rapidly acquire quantitative information, thus exhibiting strong resistance to motion changes. When used with special organs such as the heart, it can be used in conjunction with a motion monitoring module to further enhance its resistance to motion. This invention can simultaneously acquire multiple MRI k-space data. Different echoes in each k-space have different weights for MRI quantitative parameters. Therefore, this invention can achieve simultaneous quantitative imaging of multiple MRI parameters and can utilize the complementary information from different echoes and / or different k-spaces to correct for imperfections, including but not limited to artifacts, radio frequency field inhomogeneities, and main magnetic field inhomogeneities. This invention enables quantitative imaging of multiple MRI parameters and can correct for imperfections, including but not limited to artifacts, radio frequency field inhomogeneities, and main magnetic field inhomogeneities. This invention achieves rapid simultaneous quantitative imaging of MRI parameters M0, T1, T2, T2*, B1, and ΔB0; the obtained MRI quantitative parameter images have high accuracy, high signal-to-noise ratio, and low distortion; this invention can be applied to different parts and organs of the human body and small animals, exhibiting good stability. Attached Figure Description

[0055] Figure 1 This is a flowchart of the magnetic resonance multi-parameter simultaneous quantitative imaging method of the present invention;

[0056] Figure 2 This is a sequence diagram of a steady-state free precession magnetic resonance multi-parameter simultaneous quantitative imaging method based on overlapping echoes, as described in this invention.

[0057] Figure 3 This is a flowchart of the magnetic resonance multi-parameter simultaneous quantitative imaging system of the present invention;

[0058] Figure 4 This is a magnetic resonance imaging sequence diagram from an embodiment of the present invention;

[0059] Figure 5 The k-space data and magnetic resonance images of the simulated samples in this embodiment of the invention;

[0060] Figure 6 This is a multi-parameter quantitative magnetic resonance image of a simulated sample in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention, including but not limited to: adding common preparation modules such as reversal, saturation, and compression before using the sequence and during the intermediate process of cyclic use of the sequence. These should all be considered as not involving significant creative effort based on the present invention, and are therefore within the scope of protection of this patent.

[0062] Figure 1 This is a flowchart of the magnetic resonance multi-parameter simultaneous quantitative imaging method of the present invention. Figure 1 As shown in the embodiment of the present invention, a method for simultaneous quantitative imaging of multiple parameters based on overlapping echoes in magnetic resonance imaging includes:

[0063] S1: Design a steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes;

[0064] S2: Determine the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes;

[0065] S3: Add the sequence and set the sampling parameters in an advanced magnetic resonance spectrometer that meets the sequence and sampling parameters, and complete the data acquisition to obtain magnetic resonance data;

[0066] S4: Generate training samples for the deep neural network based on the sequences and parameters described in steps S1 to S3;

[0067] S5: The deep neural network is trained using the training samples to obtain a trained deep neural network;

[0068] S6: Using the trained deep neural network and magnetic resonance data (k-space data (or amplitude and / or phase) acquired from the magnetic resonance imaging sequence), reconstruct a multi-parameter quantitative image of magnetic resonance.

[0069] like Figure 2 The image shown is a sequence diagram of a steady-state free precession magnetic resonance multi-parameter simultaneous quantitative imaging method based on overlapping echoes, according to the present invention.

[0070] Where, α i α represents the flip angle of the i-th radio frequency excitation pulse, and in particular, α i Different sizes, phases, and pulse shapes can be selected. δ i Represents α i With α i+1 The time interval between;

[0071] TR represents the repetition time, which is the time interval between α1 and the next α1.

[0072] t pre The preparation period is the time interval between the first preparation pulse α and the first radio frequency excitation pulse α1. During the preparation period, several preparation pulses can be applied, and each preparation pulse can have different magnitudes, phases, and linear shapes.

[0073] Preferably, a set (or one) of preparatory pulses with a flip angle of 1 / 2αi is applied during the preparation period, followed by a set (or one) of radio frequency excitation pulses with a flip angle of αi and a phase that alternates between 0° and 180° at each TR. The time interval between the preparatory pulse 1 / 2αi and the first radio frequency excitation pulse α1 is preferably 1 / 2TR.

[0074] G ROi Indicates the application of α i The subsequent shift gradient on the frequency encoding dimension, G PEi Indicates the application of α i The shift gradient on the subsequent phase encoding dimension, i = 1, 2, ..., n, where n is a positive integer, is applied by a total of n radio frequency excitation pulses;

[0075] G cr The layer / segment gradient of the layer selection dimension is represented by the application of G simultaneously with the application of the radio frequency excitation pulse. cr Choose different Gs cr This invention can achieve both two-dimensional planar acquisition and three-dimensional volume acquisition. Figure 2 The image shown is only a two-dimensional planar acquisition mode;

[0076] G pp This represents the pre-phase gradient applied to the phase encoding dimension to change the k-space filling start point for each scan. The value is different for each scan. At the end of each TR, a phase encoding gradient of equal magnitude but opposite direction, -G, is applied. pp According to different TRs, G pp Depending on the order of the values, this invention can achieve different k-space filling methods, such as filling the k-space in the order of high and low Cartesian sampling and filling the k-space in a linear order. Figure 2 The pattern shown is only the linear order filling of the k-space.

[0077] This sequence generates echoes by gradient inversion, first utilizing a dephasing gradient G in the frequency coding direction. dpDephase the nuclear magnetization, and use a readout gradient G1 with the opposite polarity to rephase the magnetization. The area of the readout gradient is twice that of the dephasing gradient. When the shift gradient is not considered, at the middle time of the readout gradient, the divergent phase is completely compensated to form an echo peak. However, due to the existence of the shift gradient, the time to form the echo peak is advanced or delayed, which can separate the echoes in the readout direction.

[0078] Within each TR, apply m readout gradients in the frequency-encoding direction, G j represents the j-th readout gradient, and each readout gradient corresponds to a k-space kspace j , Δj represents the time interval between the last RF excitation pulse α n and the middle time of the j-th readout gradient G j . j = 1, 2,..., m, and there are m k-spaces in total;

[0079] G' ROj represents the shift gradient applied on the frequency-encoding dimension after G j , and G' PEj represents the shift gradient applied on the phase-encoding dimension after G j .

[0080] In the present invention, the RF pulse takes a small angle (usually less than 90°). After the RF pulse excitation, the transverse magnetization vector M ⊥ and the relatively large longitudinal magnetization vector M z are left. After several cycles, M ⊥ and M z after the RF pulse excitation remain unchanged, that is, the steady-state free-precession state is reached. The condition for forming the steady-state balance is TR < T2, so TR is very short and the imaging speed is very fast.

[0081] By applying shift gradients after the RF excitation pulse and after the readout gradient, several overlapping echoes will be formed in each k-space, and the distribution of the overlapping echoes in the k-space is determined by the magnitudes and directions of the respective shift gradients. Different echoes in each k-space have different magnetic resonance quantitative parameter weights. Using the echo information in each k-space, magnetic resonance multi-parameter simultaneous quantitative imaging is achieved.

[0082] Determine the sampling parameters of the steady-state free-precession and overlapping echo multi-parameter quantitative imaging sequence, specifically including:

[0083] Determine the preparation period time tpre;

[0084] Determine the number of preparation pulses during the preparation period, as well as the magnitudes, phases, and pulse shapes of each preparation pulse;

[0085] Determine the number n of RF excitation pulses within each TR;

[0086] Determine the flip angle α of each radio frequency excitation pulse. i The magnitude, phase, and pulse shape;

[0087] Determine the time interval δi of each radio frequency excitation pulse;

[0088] Determine the number m of k-spaces;

[0089] Determine the time Δj for applying the readout gradient;

[0090] Determine the phase gradient G on the frequency coding dimension dp The area and orientation of the readout gradients G j Area and orientation;

[0091] Determine the prephase gradient G on different phase encoding dimensions within the TR. pp The area and orientation of the k-space are used to determine the filling method of the k-space;

[0092] Determine the level selection dimension G cr Area and orientation;

[0093] Determine each shift gradient G ROi G PEi G' ROj G' PEj The area, direction, and application time are used to determine the number of overlapping echoes in each k-space and the position of each overlapping echo in the k-space;

[0094] Determine the pulse sequence repetition time TR;

[0095] Determine other parameters such as imaging field of view, imaging matrix, parallel imaging acceleration factor, and 2D / 3D acquisition;

[0096] Before determining the deep neural network, this invention considers that deep neural networks require a large number of training samples, but real training samples are difficult to obtain. Therefore, a simulation method is used to generate the training samples required for the deep neural network. The specific steps are as follows:

[0097] The pulse sequence is input into the magnetic resonance imaging simulation software, and corresponding non-ideal terms are added according to the non-idealities of the real experiment in order to simulate the real situation as much as possible.

[0098] Generate a set number of random templates based on the characteristics of the test object (or generate templates using an MRI dataset with the characteristics of the test object); the complexity of the templates should match that of the test object and should be able to include all the characteristics of the experimental samples; the number of templates needs to be sufficient to ensure good reconstruction quality.

[0099] The template was simulated and sampled using simulation software to obtain the template's magnetic resonance signal. During the simulation sampling process, considering the changes in the real experimental environment, unstable factors were added to improve the robustness of the network model to the unsatisfactory experimental environment. The unstable factors included excitation pulse angle deviation, shift gradient deviation, and noise.

[0100] The magnetic resonance echo signals of the templates are rearranged into two-dimensional k-space signals, and then two-dimensional Fourier transform is performed to obtain the magnetic resonance images of each template. The magnetic resonance images are then normalized and noise-added.

[0101] The magnetic resonance images of each template and the corresponding template are used to form a training sample, resulting in a set number of training samples, which constitute the training sample set.

[0102] The deep neural network is trained using the training samples to obtain a trained deep neural network, specifically including:

[0103] Determine the network structure of the deep neural network;

[0104] Determine the number of input and output channels for the deep neural network;

[0105] Determine the loss function for training the deep neural network.

[0106] When training a deep neural network, the training sample set is input into the deep neural network in batches for iterative training. Each time the network is trained, the value of the loss function is calculated. Based on this value, the parameter values ​​of the neural network are automatically adjusted to reduce the value of the loss function. The above training is repeated until the value of the loss function no longer decreases, and then the parameters of the deep neural network are saved.

[0107] Using the trained deep neural network and the k-space data acquired by the magnetic resonance imaging sequence, a multi-parameter magnetic resonance image is reconstructed, specifically including:

[0108] The magnetic resonance imaging sequence is used to acquire data from the actual imaging object to obtain the k-space data of the actual imaging object;

[0109] Fourier transform is performed on the k-space data of the actual imaging object to obtain the magnetic resonance image of the actual imaging object.

[0110] The magnetic resonance image of the actual imaging object is reconstructed based on the trained deep neural network to obtain a quantitative multi-parameter magnetic resonance image.

[0111] Figure 3 This is a schematic diagram of the magnetic resonance multi-parameter simultaneous quantitative imaging system of the present invention. Figure 3 As shown, the magnetic resonance multi-parameter simultaneous quantitative imaging system based on overlapping echoes includes:

[0112] The pulse sequence design module is used to design steady-state free precession multi-parameter quantitative imaging sequences based on overlapping echoes; and to determine the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequences based on overlapping echoes.

[0113] The signal acquisition module is used to sample the test object using the pulse sequence under set sampling parameters and a high-performance magnetic resonance spectrometer to obtain the overlapping echo magnetic resonance signal of the test object.

[0114] The signal processing module is used to perform two-dimensional Fourier transform, normalization and other processing on the overlapping echo magnetic resonance signal of the object under test to obtain the magnetic resonance image to be reconstructed of the object under test.

[0115] The training sample set generation module is used to simulate and generate a set amount of deep neural network training samples based on the characteristics of the object to be tested, thus forming a training sample set.

[0116] A deep neural network determination module is used to determine a deep neural network for multi-parameter quantitative image reconstruction; the deep neural network is trained using the training sample set to obtain a trained deep neural network;

[0117] The multi-parameter quantitative image reconstruction module is used to input the overlapping echo magnetic resonance image obtained from the magnetic resonance instrument into the trained deep neural network for reconstruction, so as to obtain a multi-parameter quantitative image of the test object.

[0118] The following is a specific embodiment.

[0119] Step 1: Design a steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes and determine the sequence sampling parameters. Reference for magnetic resonance imaging sequence design. Figure 4 .

[0120] During the preparation period, two preparatory pulses α1 / 2 and α2 / 2 are applied, both of which are sinc pulses with a magnitude of 10°. Subsequently, two radio frequency excitation pulses α1 and α2 are applied within each TR, both of which are sinc pulses with a magnitude of 20°. Within each TR, the phases of α1 and α2 alternate between 0° and 180°.

[0121] The time interval δ1 between α1 and α2 is 3.64 ms.

[0122] The preparation period tpre = 9.25 ms

[0123] Repetition time TR = 18.5ms

[0124] The imaging field of view (FOV) is 0.22m x 0.22m.

[0125] The imaging matrix size is 128*128

[0126] Apply three readout gradients within each TR. The time intervals between the mid - times of each readout gradient and α2 are respectively: Δ1 = 3.64 ms, Δ2 = 7.28 ms, Δ3 = 10.92 ms

[0127] Assume that the area of the readout gradient during sampling is 1 unit area in the frequency - encoding direction; assume that for all TRs, twice the area of the maximum value of the absolute value in the phase - encoding gradient Gpp is 1 unit area in the phase - encoding direction, that is, the area of max(abs(S Gpp ))*2 is 1 unit area in the phase - encoding direction. Apply shift gradients in the frequency - encoding dimension and the phase - encoding dimension after each radio - frequency excitation pulse and readout gradient. The relative areas of each shift gradient are respectively: (GRO1, GPE1) = (-0.22, 0.22), (GRO2, GPE2) = (0.11, -0.11), (G’RO1, G’PE1) = (-0.11, 0.11), (G’RO2, G’PE2) = (0.22, -0.22), (G’RO3, G’PE3) = (-0.31, -0.09).

[0128] Gcr represents the slice / segment gradient in the slice - selection dimension, and Gcr is applied simultaneously with the radio - frequency excitation pulse. By selecting different Gcr modes, the present invention can achieve two - dimensional planar acquisition and three - dimensional volume acquisition, Figure 4 where Gcr is selected as the two - dimensional planar acquisition mode;

[0129] Gpp represents the pre - phase gradient applied in the phase - encoding dimension, which is used to change the starting point of k - space filling for each scan. The value of Gpp is different for each scan. At the end of each TR, a phase - encoding gradient - Gpp with the same magnitude and opposite direction is applied. According to the order of the values of Gpp within different TRs, the present invention can achieve high - low order filling of k - space and linear order filling of k - space in Cartesian sampling. Figure 4 where Gpp is selected as the linear order filling k - space mode;

[0130] The radio - frequency pulse in the present invention takes a small angle. After the radio - frequency pulse excitation, a transverse magnetization vector M⊥ and a relatively large longitudinal magnetization vector Mz are left. After several cycles, M⊥ and Mz after the radio - frequency pulse excitation remain unchanged, that is, it reaches the steady - state free - precession state. The condition for forming a steady - state balance is TR < T₂, so TR is very short and the imaging speed is very fast.

[0131] By applying a shift gradient after the radio frequency excitation pulse and after the readout gradient, several overlapping echoes are formed in each k-space. The distribution of each overlapping echo in the k-space is determined by the magnitude and direction of each shift gradient. Different echoes in each k-space have different weights for the quantitative magnetic resonance parameters. By utilizing the echo information in each k-space, this invention can achieve simultaneous quantitative imaging of multiple magnetic resonance parameters.

[0132] Step 2: Generate training samples for the deep neural network;

[0133] The pulse sequence is input into the magnetic resonance imaging simulation software, and corresponding non-ideal terms are added according to the non-idealities of the real experiment to simulate the real situation as much as possible.

[0134] A set number of random templates are generated based on the characteristics of the test object; the complexity of the templates is matched with that of the test object and should be able to contain all the characteristics of the experimental sample; the number of templates is sufficient to ensure good reconstruction quality.

[0135] The template was simulated and sampled using simulation software to obtain the template's magnetic resonance signal. During the simulation sampling process, considering the changes in the real experimental environment, unstable factors were added to improve the robustness of the network model to the unsatisfactory experimental environment. The unstable factors included excitation pulse angle deviation, shift gradient deviation, and noise.

[0136] The magnetic resonance echo signals of the templates are rearranged into two-dimensional k-space signals, and two-dimensional Fourier transform is performed to obtain the magnetic resonance images of each template. The magnetic resonance images are then normalized and noise-added.

[0137] The magnetic resonance images of each template and the corresponding template are used to form a training sample, resulting in a set number of training samples, which constitute the training sample set.

[0138] Step 3: Use the training samples to train the deep neural network to obtain a trained deep neural network;

[0139] The network structure of the deep neural network is determined. Since U-Net has good performance in nonlinear fitting and feature extraction for image-based tasks, this embodiment selects the U-Net deep neural network, but other deep neural network structures can also be selected.

[0140] Determine the number of input and output channels for the deep neural network; such as Figure 4 As shown, the sequence described in this embodiment acquires a total of 3 k-space data points, corresponding to 3 magnetic resonance images. The real and imaginary parts of each magnetic resonance image are used as inputs to the deep neural network, so the number of input channels of the deep neural network is 6. This embodiment reconstructs 6 quantitative magnetic resonance parameter maps: M0, T1, T2, T2*, B1, and ΔB0, so the number of output channels of the deep neural network is 6.

[0141] Determine the loss function for training the deep neural network. Use the pixel-wise mean squared error as the loss function, as described below:

[0142]

[0143] Where n(·) is the output of the deep neural network, M represents the batch size, Xi represents the network input data in the i-th training sample, and Yi represents the network label data in the i-th training sample. W and b represent the parameters of the deep neural network.

[0144] The learning rate is adjusted using an exponential decay method: the initial learning rate is set to 0.00005, and the learning rate is reduced once after a set number of training iterations are reached.

[0145] The training sample set is randomly divided into two parts: 80% for training and 20% for validation.

[0146] When training a deep neural network, the training sample set is input into the deep neural network in batches for iterative training. Each time the network is trained, the value of the loss function is calculated. Based on this value, the parameter values ​​of the neural network are automatically adjusted to reduce the value of the loss function. The above training is repeated until the value of the loss function no longer decreases, and the parameters W and b of the deep neural network are saved.

[0147] Step 4: Reconstruct the multi-parameter magnetic resonance image using the trained deep neural network and the k-space data acquired by the magnetic resonance imaging sequence.

[0148] The magnetic resonance imaging sequence is used to acquire data of the imaging object to obtain the k-space data of the imaging object;

[0149] A two-dimensional Fourier transform is performed on the k-space data of the imaging object to obtain the magnetic resonance image of the imaging object.

[0150] Figure 5 This includes k-space data and magnetic resonance images of a human brain simulation sample; each k-space contains several overlapping echoes, and the position and signal strength of each overlapping echo are determined by... Figure 4 The shift gradients in the magnetic resonance imaging sequence shown are determined by these gradients. Figure 5 As shown, the magnetic resonance image corresponding to the k-space containing overlapping echoes is a striped aliased image.

[0151] The magnetic resonance image of the simulated human brain sample is input into the deep neural network trained in step 3 for reconstruction, resulting in... Figure 6 The image shown is a multi-parameter quantitative image from magnetic resonance imaging. Figure 6The images displayed are of six quantitative magnetic resonance imaging (MRI) parameters: M0, T1, T2, T2*, B1, and ΔB0, from the simulated human brain sample. The first row of images shows the MRI MRI MRI MRI MRI MRI MRI images reconstructed by a deep neural network; the second row shows the corresponding label images; and the third row shows the absolute error values ​​between the reconstruction results from the deep neural network and the corresponding label images. Specifically, the absolute error value image for B1 is magnified 10 times. The error graphs show that the MRI MRI MRI MRI MRI MRI MRI MRI images reconstructed by the deep neural network have accurate quantitative values ​​and high imaging precision.

[0152] Specific examples are used in this invention to illustrate the principles and implementation methods of the invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the invention.

Claims

1. A method for simultaneous quantitative multi-parameter magnetic resonance imaging based on overlapping echoes, characterized in that... Includes the following steps: S1: Design of a steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes: Several preparatory pulses α during the preparation period; Within each TR, the flip angle is α. i The radio frequency excitation pulse and the corresponding applied to α i The shift gradient G on the subsequent frequency coding dimension ROi and applied to α i The subsequent shift gradient G on the phase encoding dimension PEi Where i = 1, 2, ..., n, a total of n radio frequency excitation pulses are applied; Simultaneously applying the radio frequency excitation pulse, a layer / segment gradient G with a layer selection dimension is applied. cr Choose different Gs cr The mode enables both two-dimensional planar acquisition and three-dimensional volume acquisition. Prephase gradient G pp An equal-magnitude, opposite-direction phase-encoding gradient -G is applied along the phase-encoding dimension to change the k-space filling start point for each scan, with a different value for each scan. At the end of each TR period, a phase-encoding gradient -G is applied. pp According to different TRs G pp The order of values ​​determines the different ways to fill the k-space, such as filling the k-space in the order of high and low Cartesian sampling and filling the k-space in a linear order. This sequence generates echoes by gradient inversion, first utilizing a dephasing gradient G in the frequency coding direction. dp The magnetization intensity of the atomic nuclei is dephased, and then a readout gradient G1 with opposite polarity is used to reconcile the magnetization intensity. The area of ​​the readout gradient is twice the area of ​​the dephase gradient. Without considering the shift gradient, at the midpoint of the readout gradient, the divergent phase is completely compensated to form the echo peak. However, due to the existence of the shift gradient, the time of forming the echo peak is advanced or delayed, so that the echo is separated in the readout direction. Within each TR, m readout gradients are applied in the frequency coding direction, and the j-th readout gradient is represented by G. j This means that each readout gradient corresponds to a k-space j, j = 1,2,…,m, and there are a total of m k-spaces; In G j Then apply the shift gradient G' on the frequency coding dimension. ROj and the shift gradient G' on the phase encoding dimension Pej ; S2: Determine the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes; S3: Add the sequence and set the sampling parameters in an advanced magnetic resonance spectrometer that meets the performance requirements of the sequence and sampling parameters, and complete the data acquisition to obtain magnetic resonance data; S4: Generate training samples for the deep neural network based on the sequences and parameters described in S1~S3; S5: Train the deep neural network using training samples to obtain a trained deep neural network; S6: Using a trained deep neural network and magnetic resonance data, reconstruct a multi-parameter quantitative image of magnetic resonance.

2. The method for simultaneous quantitative imaging of multiple parameters based on overlapping echoes in magnetic resonance imaging as described in claim 1, characterized in that... In step S2, determining the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequence based on overlapping echoes specifically includes: Determine the preparation period time tpre; Determine the number of preparatory pulses during the preparation period, as well as the size, phase, and pulse shape of each preparatory pulse; Determine the number n of radio frequency excitation pulses within each TR; Determine the flip angle α of each radio frequency excitation pulse. i The magnitude, phase, and pulse shape; Determine the time interval δi of each radio frequency excitation pulse; Determine the number m of k-spaces; Determine the time Δj for applying the readout gradient; Determine the phase gradient G on the frequency coding dimension dp The area and orientation of the readout gradients G j Area and orientation; Determine the prephase gradient G on different phase encoding dimensions within the TR. pp The area and orientation of the k-space are used to determine the filling method of the k-space; Determine the level selection dimension G cr Area and orientation; Determine each shift gradient G ROi G PEi G' ROj G' PEj The area, direction, and application time are used to determine the number of overlapping echoes in each k-space and the position of each overlapping echo in the k-space; Determine the pulse sequence repetition time TR; Determine the imaging field of view, imaging matrix, parallel imaging acceleration factor, and other parameters for 2D / 3D acquisition.

3. The method for simultaneous quantitative imaging of multiple parameters based on overlapping echoes in magnetic resonance imaging as described in claim 1, characterized in that... In step S4, the training samples for the deep neural network are generated based on the sequences and parameters described in S1-S3. This is achieved using a training sample set generation module, and the specific scheme includes: The pulse sequence is input into the magnetic resonance imaging simulation software, and corresponding non-ideal terms are added according to the non-idealities of the real experiment to simulate the real situation; Generate a set number of random templates based on the characteristics of the test object, or generate templates using MRI datasets with the characteristics of the test object, ensuring that the templates are highly complex and contain all the features of the experimental samples; the number of templates is sufficient to ensure good reconstruction quality; The template was simulated and sampled using simulation software to obtain the magnetic resonance signal of the template. During the simulation sampling process, considering the changes in the real experimental environment, undesirable factors were added to improve the robustness of the network model to undesirable experimental environments. The undesirable factors included excitation pulse angle deviation, shift gradient deviation, and noise. The magnetic resonance echo signals of the templates are rearranged into two-dimensional k-space signals, and then two-dimensional Fourier transform is performed to obtain the magnetic resonance images of each template. The magnetic resonance images are then normalized and noise-added. The magnetic resonance images of each template and the corresponding template are used to form a training sample, and a set number of training samples are obtained to form a training sample set. The training samples above are either simulated samples, created from real collected samples, or a combination of real and simulated samples.

4. The method for simultaneous quantitative imaging of multiple parameters based on overlapping echoes in magnetic resonance imaging as described in claim 1, characterized in that... In step S5, training the deep neural network using training samples to obtain a trained deep neural network specifically includes: Determine the network structure of the deep neural network; Determine the number of input and output channels for the deep neural network; Determine the loss function for training the deep neural network; When training a deep neural network, the training sample set is input into the deep neural network in batches for iterative training. Each time the network is trained, the value of the loss function is calculated. Based on this value, the parameter values ​​of the neural network are automatically adjusted to reduce the value of the loss function. The above training is repeated until the value of the loss function no longer decreases, and then the parameters of the deep neural network are saved.

5. A magnetic resonance multi-parameter simultaneous quantitative imaging system based on overlapping echoes, characterized in that... The magnetic resonance multi-parameter simultaneous quantitative imaging method based on overlapping echoes as described in claim 1 is provided, wherein the system sequentially includes a pulse sequence design module, a signal acquisition module, a signal processing module, a training sample set generation module, a deep neural network determination module, and a multi-parameter quantitative image reconstruction module; The pulse sequence design module is used to design steady-state free precession multi-parameter quantitative imaging sequences based on overlapping echoes; and to determine the sampling parameters of the steady-state free precession multi-parameter quantitative imaging sequences based on overlapping echoes. The signal acquisition module is used to sample the test object using the pulse sequence under set sampling parameters and a high-performance magnetic resonance spectrometer to obtain the overlapping echo magnetic resonance signal of the test object. The signal processing module is used to perform two-dimensional Fourier transform and normalization processing on the overlapping echo magnetic resonance signal of the object under test to obtain the magnetic resonance image to be reconstructed of the object under test. The training sample set generation module is used to simulate and generate a set amount of deep neural network training samples based on the characteristics of the object to be tested, thus forming a training sample set. A deep neural network determination module is used to determine the deep neural network for multi-parameter quantitative image reconstruction; The deep neural network is trained using the training sample set to obtain a trained deep neural network; The multi-parameter quantitative image reconstruction module is used to input the overlapping echo magnetic resonance image obtained from the magnetic resonance instrument into the trained deep neural network for reconstruction, thereby obtaining a multi-parameter quantitative image of the analyte.

Images