A fast high-resolution multi-parametric quantitative magnetic resonance imaging method and system

Abstract

The application discloses a kind of fast high-resolution multi-parameter quantitative magnetic resonance imaging method and system, comprising: design fast high-resolution multiple echo imaging pulse sequence;Determine the sampling parameter of pulse sequence;Construct the deep neural network for multi-parameter quantitative magnetic resonance image reconstruction;Generate the training sample of deep neural network;Using training sample, the trained deep neural network is obtained by training deep neural network;Using pulse sequence is scanned under the set sampling parameter to entity imaging object, and the k space data of entity imaging object is obtained;The k space data of the entity imaging object is preprocessed, and the image domain data of the entity imaging object is obtained;The image domain data of the entity imaging object is input into trained deep neural network and is reconstructed, and high-resolution multi-parameter quantitative magnetic resonance image is obtained.The application can realize fast high-resolution multi-parameter quantitative magnetic resonance imaging, and can resist the image distortion caused by main magnetic field inhomogeneity.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of magnetic resonance imaging, in particular to a fast high-resolution multi-parameter quantitative magnetic resonance imaging method and system. BACKGROUND

[0002] Magnetic resonance imaging can non-invasively and non-destructively image biological tissues, and quantitative magnetic resonance imaging has good repeatability, is not sensitive to instrument differences, and has great clinical diagnostic value. Traditional quantitative magnetic resonance imaging requires the acquisition of multiple images with different parameter weighting for fitting, which takes a long time. Moreover, during the long sampling process, the motion of the imaging object will cause image artifacts and other problems. In order to shorten the imaging acquisition time, a series of fast quantitative magnetic resonance imaging methods based on overlapping echo technology have been proposed, including CN108663644B, CN108010100B, CN110807492A, etc. However, the above methods are based on planar echo imaging sampling, so the spatial resolution is limited, and are easily affected by magnetic field inhomogeneity. SUMMARY

[0003] To solve the above problems, the present application provides a fast high-resolution multi-parameter quantitative magnetic resonance imaging method and system, which can realize fast high-resolution multi-parameter quantitative magnetic resonance imaging and has the ability to resist image distortion caused by magnetic field inhomogeneity.

[0004] In one aspect, a fast high-resolution multi-parameter quantitative magnetic resonance imaging method comprises:

[0005] designing a fast high-resolution multi-overlapping echo imaging pulse sequence;

[0006] determining the sampling parameters of the fast high-resolution multi-overlapping echo imaging pulse sequence;

[0007] constructing a deep neural network for multi-parameter quantitative magnetic resonance image reconstruction;

[0008] generating training samples for the deep neural network;

[0009] training the deep neural network using the training samples to obtain a trained deep neural network;

[0010] scanning a physical imaging object using the fast high-resolution multi-overlapping echo imaging pulse sequence under the set sampling parameters to obtain k-space data of the physical imaging object;

[0011] preprocessing the k-space data of the physical imaging object to obtain image domain data of the physical imaging object;

[0012] The image domain data of the physical imaging object is input into a trained deep neural network for reconstruction to obtain a high-resolution multi-parameter quantitative magnetic resonance image of the physical imaging object.

[0013] The fast high-resolution multi-overlap echo imaging pulse sequence includes a signal excitation module and a data acquisition module.

[0014] The signal excitation module includes N time intervals t. n The flip angle is α n RF excitation pulse and corresponding layer selection gradient G ss and echo shift gradient G n Where n = 1, 2, ..., N, N ≥ 2;

[0015] Each radio frequency excitation pulse is associated with a layer selection gradient G in the layer selection direction. ss Combined layer selection;

[0016] Echo shift gradient G n The frequency encoding dimension and the phase encoding dimension are applied after the corresponding radio frequency excitation pulse;

[0017] The data acquisition module includes a pre-phase gradient G. pre Destroy gradient G cr Frequency coding gradient G ro M time intervals are ESP and the flip angle is β. m Refocusing pulse, phase-coded gradient G pe,i,m and dephase-encoded gradient G pe,i,m ′; where i represents the i-th scan, m = 1, 2, ..., M, M ≥ 2;

[0018] Prephase gradient G pre Apply along the frequency coding dimension, with a magnitude of the frequency coding gradient G. ro Half of;

[0019] Destroy gradient G cr Before and after each refocusing pulse, directions are selected along the frequency encoding dimension, phase encoding dimension, and plane, with equal magnitude and the same direction;

[0020] Phase-encoded gradient G pe,i,m Frequency coding gradient G ro and dephase-encoded gradient G pe,i,m After each refocusing pulse, the phase-encoded gradient and the dephase-encoded gradient are applied with equal magnitudes but opposite directions.

[0021] Preferably, determining the sampling parameters for the fast high-resolution multi-overlapping echo imaging pulse sequence specifically includes:

[0022] determining the number N of radio frequency excitation pulses and the flip angle a of each radio frequency excitation pulse n ;

[0023] determining the composition of the echo signal, thereby determining the time interval t between each radio frequency pulse n ;

[0024] determining the position of each excitation signal in k-space, thereby determining the magnitude of the echo shift gradient G after each radio frequency excitation pulse n ;

[0025] determining the number M of refocusing pulses and the flip angle b of each refocusing pulse m ;

[0026] determining the phase encoding manner, thereby determining the direction and magnitude of each phase encoding gradient and dephasing gradient

[0027] determining the number of pulse sequence scans, the imaging field of view, the imaging matrix, the sampling frequency, the time interval ESP between refocusing pulses, the sampling acceleration multiple and the pulse sequence repetition time TR, thereby determining the magnitude of various gradients in the pulse sequence.

[0028] Preferably, the deep neural network for multi-parameter quantitative magnetic resonance image reconstruction is constructed, specifically comprising:

[0029] constructing the network structure of the deep neural network

[0030] constructing the number of input channels and the number of output channels of the deep neural network

[0031] constructing the loss function of the deep neural network.

[0032] Preferably, the training sample of the deep neural network is generated, specifically comprising:

[0033] According to the characteristics of the entity imaging object, simulated proton density map, T1 quantitative map, T2 quantitative map and T2 * quantitative map are made to constitute a virtual imaging object

[0034] Based on the non-ideal factors existing in the actual scan, a corresponding non-ideal factor model is established

[0035] The fast high-resolution multiple echo imaging pulse sequence and the non-ideal factor model are used to perform Bloch simulation on the virtual imaging object to obtain k-space data of the virtual imaging object

[0036] The k-space data of the virtual imaging object is separated into odd and even rows of data, and filled into two k-spaces respectively to obtain the k-space data of the virtual imaging object, and inverse Fourier transform is performed to obtain the image domain data of the virtual imaging object, which together with the virtual imaging object and the non-ideal factor model constitutes a training sample.

[0037] The above steps are repeated to generate a set number of training samples.

[0038] Preferably, the k-space data of the entity imaging object is preprocessed to obtain the image domain data of the entity imaging object, specifically including:

[0039] The acquired k-space odd and even row data is separated and filled into two k-spaces according to the phase encoding mode to obtain the k-space data of the entity imaging object;

[0040] If the sampling acceleration multiple is greater than 1, the two k-space data of the entity imaging object are reconstructed in parallel;

[0041] Inverse Fourier transform is performed on the two k-space data respectively to obtain the corresponding image domain data of the entity imaging object.

[0042] On the other hand, a fast high-resolution multi-parameter quantitative magnetic resonance imaging system comprises:

[0043] A pulse sequence design module for designing a fast high-resolution multi-echo imaging pulse sequence;

[0044] A sampling parameter determination module for determining the sampling parameters of the fast high-resolution multi-echo imaging pulse sequence;

[0045] A deep neural network construction module for constructing a deep neural network for multi-parameter quantitative magnetic resonance image reconstruction;

[0046] A training sample generation module for generating training samples for the deep neural network;

[0047] A deep neural network training module for training the deep neural network using the training samples to obtain a trained deep neural network;

[0048] A k-space data acquisition module for scanning an entity imaging object under a set of sampling parameters using the fast high-resolution multi-echo imaging pulse sequence to obtain k-space data of the entity imaging object;

[0049] An image domain data acquisition module for preprocessing the k-space data of the entity imaging object to obtain image domain data of the entity imaging object;

[0050] The multi-parameter quantitative magnetic resonance image acquisition module is used to input the image domain data of the entity imaging object into a trained deep neural network for reconstruction, thereby obtaining a high-resolution multi-parameter quantitative magnetic resonance image of the entity imaging object.

[0051] The fast high-resolution multi-overlap echo imaging pulse sequence includes a signal excitation module and a data acquisition module.

[0052] The signal excitation module includes N time intervals t. n The flip angle is α n RF excitation pulse and corresponding layer selection gradient G ss and echo shift gradient G n Where n = 1, 2, ..., N, N ≥ 2;

[0053] Each radio frequency excitation pulse is associated with a layer selection gradient G in the layer selection direction. ss Combined layer selection;

[0054] Echo shift gradient G n The frequency encoding dimension and the phase encoding dimension are applied after the corresponding radio frequency excitation pulse;

[0055] The data acquisition module includes a pre-phase gradient G. pre Destroy gradient G cr Frequency coding gradient G ro M time intervals are ESP and the flip angle is β. m Refocusing pulse, phase-coded gradient G pe,i,m and dephase-encoded gradient G pe,i,m ′; where i represents the i-th scan, m = 1, 2, ..., M, M ≥ 2;

[0056] Prephase gradient G pre Apply along the frequency coding dimension, with a magnitude of the frequency coding gradient G. ro Half of;

[0057] Destroy gradient G cr Before and after each refocusing pulse, directions are selected along the frequency encoding dimension, phase encoding dimension, and plane, with equal magnitude and the same direction;

[0058] Phase-encoded gradient G pe,i,m Frequency coding gradient G ro and dephase-encoded gradient G pe,i,m After each refocusing pulse, the phase-encoded gradient and the dephase-encoded gradient are applied with equal magnitudes but opposite directions.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The application discloses a rapid high-resolution multi-parameter quantitative magnetic resonance imaging method and system. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 A method flowchart of an embodiment of the application;

[0062] Figure 2 A rapid high-resolution multi-echo imaging pulse sequence diagram in an embodiment of the application;

[0063] Figure 3 A flowchart of generating a deep neural network training sample in an embodiment of the application;

[0064] Figure 4 A human brain multi-parameter quantitative magnetic resonance image obtained in an embodiment of the application;

[0065] Figure 5 A system structure block diagram of an embodiment of the application.

[0066] The application will be further described in detail below in combination with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0067] The application will be further described in detail below in combination with the accompanying drawings and specific embodiments.

[0068] Reference Figure 1 As shown in the accompanying drawings, the application is a rapid high-resolution multi-parameter quantitative magnetic resonance imaging method, comprising:

[0069] S1, designing a rapid high-resolution multi-echo imaging pulse sequence;

[0070] S2, determining sampling parameters of the rapid high-resolution multi-echo imaging pulse sequence;

[0071] S3, constructing a deep neural network for multi-parameter quantitative magnetic resonance image reconstruction;

[0072] S4, generating training samples of the deep neural network;

[0073] S5, training the deep neural network using the training samples to obtain a trained deep neural network;

[0074] S6, scanning a physical imaging object using the rapid high-resolution multi-echo imaging pulse sequence under the set sampling parameters to obtain k-space data of the physical imaging object;

[0075] S7, preprocess the k-space data of the entity imaging object to obtain the image domain data of the entity imaging object;

[0076] S8, the image domain data of the entity imaging object is input into the trained deep neural network for reconstruction to obtain a high-resolution multi-parameter quantitative magnetic resonance image of the entity imaging object.

[0077] The following is a detailed description of each of the above steps.

[0078] S1 involves designing the fast high-resolution multi-overlap echo imaging pulse sequence, i.e., designing a fast high-resolution multi-parameter quantitative magnetic resonance imaging pulse sequence.

[0079] For details, see Figure 2 As shown, the fast high-resolution multi-overlap echo imaging pulse sequence of this embodiment includes a signal excitation module and a data acquisition module, where RF represents radio frequency pulse, PE represents phase encoding direction, RO represents frequency encoding direction, and SS represents slice selection direction.

[0080] The signal excitation module includes N time intervals t. n The flip angle is α n RF excitation pulse and corresponding layer selection gradient G ss and echo shift gradient G n where n = 1, 2, ..., N, N ≥ 2;

[0081] Each radio frequency excitation pulse is associated with a layer selection gradient G in the layer selection direction. ss Combined layer selection;

[0082] Echo shift gradient G n The frequency encoding dimension and the phase encoding dimension are applied after the corresponding radio frequency excitation pulse;

[0083] The data acquisition module includes a pre-phase gradient G. pre M time intervals are ESP and the flip angle is β. m refocusing pulse, disrupting gradient G cr Phase encoding gradient G pe,i,m Frequency coding gradient G ro and dephase-encoded gradient G pe,i,m The ' group is composed of ', where i represents the i-th scan, m = 1, 2, ..., M, M ≥ 2;

[0084] Prephase gradient G pre Apply along the frequency coding dimension, with a size equal to half of the frequency coding gradient;

[0085] Destroy gradient G crBefore and after each refocusing pulse, directions are selected along the frequency encoding dimension, phase encoding dimension, and plane, with equal magnitude and the same direction;

[0086] Phase-encoded gradient G pe,i,m Frequency coding gradient G ro and dephase-encoded gradient G pe,i,m After each refocusing pulse, the phase-encoded gradient and the dephase-encoded gradient are applied, with equal magnitudes and opposite directions.

[0087] S2 determines the sampling parameters of the fast high-resolution multi-overlap echo imaging pulse sequence, as follows.

[0088] S21, determine the number N of radio frequency excitation pulses and the flip angle α of each radio frequency excitation pulse. n .

[0089] In this embodiment, the number of radio frequency excitation pulses N = 4, and the flip angle α of each radio frequency excitation pulse is... n =30°;

[0090] S22, determine the composition of the echo signal, thereby determining the time interval t between each radio frequency pulse. n .

[0091] The signal generated by the fourth RF excitation pulse is acquired after a refocusing pulse. To compensate for the effects of the main magnetic field inhomogeneity and generate a T2-weighted signal, t4 = 0.5ESP is set. The signal generated by the second RF excitation pulse is partially refocused after the third RF excitation pulse. To obtain a T2-weighted signal, t2 = t3 is set, so that it refocuses at the time point of the fourth RF excitation pulse. To ensure that the first RF excitation pulse generates a T2-weighted signal... * Since the weighted signal has a certain difference in echo time from the third RF excitation pulse, and the signal attenuation time is not too long, t1 should be set within a suitable range. In this embodiment, t1 = t2 = t3 = 6ms.

[0092] S23, determine the position of each echo signal in k-space, thereby determining the echo shift gradient G after each RF excitation pulse. n The proportion of size.

[0093] In this embodiment, the echo shift gradients in the frequency coding direction are G1 = -0.1875G ro G2 = 0.125G ro G3 = -0.1875G ro G4 = 0.125G ro The echo shift gradients in the phase encoding direction are G1 = 0.5G PE G2 = 0.5G PE G3 = 0.5GPE G4 = -1.5G PE , where G ro is the frequency encoding gradient, G PE is the phase encoding gradient G pe,i,m , and G m is the maximum value in G

[0094] S24, determine the number M of refocusing pulses and the flip angle β of each refocusing pulse m .

[0095] In this embodiment, the number M of refocusing pulses is 16, and the flip angle β of each refocusing pulse is 180°.

[0096] S25, determine the phase encoding mode, thereby determining the direction and size of each phase encoding and dephasing.

[0097] In this embodiment, the phase encoding mode adopts staggered phase encoding.

[0098] S26, determine the number of pulse sequence scans, the imaging field of view, the imaging matrix, the sampling frequency, the time interval ESP between refocusing pulses, the sampling acceleration multiple, and the pulse sequence repetition time TR, thereby determining the numerical value of each gradient in the pulse sequence.

[0099] In this embodiment, the number of pulse sequence scans is 16, the imaging field of view is 220 mm x 220 mm, the imaging matrix is 256 x 256, the sampling frequency is 62.464 kHz, the time interval ESP between refocusing pulses is 8.84 ms, the sampling acceleration multiple is 2, and the pulse sequence repetition time TR is 4000 ms.

[0100] S3, constructing a deep neural network for multi-parameter quantitative magnetic resonance image reconstruction, specifically as follows.

[0101] S31, determining the network structure of the deep neural network.

[0102] Preferably, in this embodiment, the network structure adopted is a U-shaped neural network, which includes 4 down-sampling layers and 4 up-sampling layers; each down-sampling layer includes two convolutional layers with a convolution kernel size of 3 x 3, a step of 1, and padding of 1, and a pooling layer with a convolution kernel size of 2 x 2 and a step of 2, each convolutional layer being activated using a ReLU activation function; each up-sampling layer includes a deconvolutional layer with a convolution kernel size of 3 x 3 and a step of 2, and two convolutional layers with a convolution kernel size of 3 x 3, a step of 1, and padding of 1, each convolutional layer being activated using a ReLU activation function.

[0103] S32, determining the input channel number and the output channel number of the deep neural network.

[0104] Preferably, in the present embodiment, the input of the deep neural network is the real part and the imaginary part of the two image domain data, so the number of input channels is 4; the output of the deep neural network is the proton density map, the T2 quantitative map, the T2 * quantitative map or the dB0 map, so the number of output channels is 1.

[0105] S33: Determine the loss function of the deep neural network.

[0106] In the present embodiment, the loss function adopted is the mean absolute error, that is, the average of the absolute value of the difference between the network output and the label:

[0107]

[0108] Wherein, L represents the value of the loss function, ∑ represents summation, P represents the number of training samples in the training sample set of the deep neural network, |·| represents taking the absolute value, f() represents the mapping relationship of the deep neural network, x j represents the input in the jth training sample, y j represents the label in the jth training sample.

[0109] Referring to Figure 3 , the S4, the training sample of the deep neural network is generated, and the specific process is as follows.

[0110] The following takes the human brain as an example.

[0111] S41, according to the characteristics of the human brain, the simulated proton density map, T1 quantitative map, T2 quantitative map, T2 * quantitative map, and a virtual imaging object is composed.

[0112] In the present embodiment, the virtual imaging object is made by using the magnetic resonance imaging public data set IXI, and each object of the data set contains T1 weighted map, T2 weighted map, T2 * weighted map and proton density weighted map. The object of the data set is converted into a virtual imaging object, which specifically includes:

[0113] S411, the proton density weighted map is taken as the proton density map of the virtual imaging object;

[0114] S412, the T1 weighted map is expressed as S T1w ∝PD×[1-exp(-TR / T1)], wherein S T1w represents the T1 weighted map, ∝ represents proportional to, PD represents the proton density map, and T1 represents the T1 quantitative map. Assuming that the TR of the T1 weighted map is TR0, the T1 quantitative map of the virtual imaging object is T1 =-TR0 / ln(1-S T1w / PD);

[0115] S413, the T2 weighted image is represented as S T2w = PD x exp(-TE / T2), wherein S T2w represents the T2 weighted image, TE represents the echo time, and T2 represents the T2 quantitative image. Assuming that the TE of the T2 weighted image is TE0, the T2 quantitative image of the virtual imaging object is T2 = -TE0 / ln(S T2w / PD);

[0116] S414, the T2 * weighted image is represented as S T2*w = PD x exp(-TE / T2 * ), wherein S T2*w represents the T2 * weighted image, T2 * represents the T2 * quantitative image. Assuming that the TE of the T2 * weighted image is TE0, the T2 * quantitative image of the virtual imaging object is T2 * = -TE0 / ln(S T2*w / PD);

[0117] The proton density image, the T1 quantitative image, the T2 quantitative image, and the T2 * quantitative image are packaged into a virtual imaging object.

[0118] S42, a corresponding non-ideal factor model is established based on the non-ideal factors existing in the actual scanning, which is specifically as follows.

[0119] In this embodiment, three non-ideal factors are considered, including the inhomogeneity of the main magnetic field B0, the inhomogeneity of the radio frequency field B1, and the non-ideality of the echo shift gradient.

[0120] S421, the inhomogeneity of the main magnetic field B0 is modeled to obtain a main magnetic field deviation model dB0:

[0121] dB0 = amp0 x {a0X + b0Y + c0XY + e0X 2 + e0Y 2 + f0XY 2 + g0X 2 Y + h0exp[(X-i0) 2 + (Y-j0) 2}

[0122] Wherein, amp0 controls the uneven degree of the main magnetic field, subject to uniform distribution between [0, 150]; a0, b0, c0, d0, e0, f0, g0, f0 subject to uniform distribution between [-0.5, 0.5]; i0, j0 subject to uniform distribution between [0.1, 0.9]; X, Y are the coordinate values of the zero padding model generated by dB0 according to the image size of the virtual imaging object after normalization, the range is [-1, 1].

[0123] S422, the unevenness of the radio frequency field B1 is modeled to obtain the radio frequency field model B1:

[0124] B1 = amp1 × {a1X + b1Y + c1XY + d1X 2 + e1Y 2 + f1exp[(X-i1) 2 + (Y-j1) 2 ]} + g1 + 1

[0125] Wherein, amp1 represents the uneven degree of the radio frequency field, subject to uniform distribution between [0, 0.6]; a1, b1 subject to Gaussian distribution with mean 0 and variance 0.5; c1, d1, e1, f1 subject to uniform distribution between [-0.5, 0.5]; g1 subject to uniform distribution between [-0.05, 0.05]; i1, j1 subject to uniform distribution between [0.1, 0.9]; X, Y are the coordinate values of the zero padding model generated by B1 according to the image size of the virtual imaging object after normalization, the range is [-1, 1].

[0126] S423, the non-idealness of the echo shift gradient is modeled to obtain the echo shift gradient model G n :

[0127] G n = (1 + a2) × G n,ideal

[0128] Wherein, a2 represents the non-ideal degree of the shift gradient, subject to uniform distribution [-0.01, 0.01]; G n represents the size of the n-th echo shift gradient actually applied, G n,ideal represents the size of the n-th echo shift gradient, n = 1, 2, 3, 4.

[0129] S43, the fast high-resolution multiple echo imaging pulse sequence and the non-ideal factor model in step S42 are used to perform Bloch simulation on the virtual imaging object to obtain the k-space data of the virtual imaging object;

[0130] S44, separate the k-space data of the virtual imaging object into odd and even rows, fill them into two k-spaces respectively to obtain two k-space data of the virtual imaging object, and perform inverse Fourier transform to obtain two image domain data of the virtual imaging object, which together with the virtual imaging object and the non-ideal factor model form a training sample;

[0131] S45, repeat steps S41-S44 to generate a set number of training samples.

[0132] S5, training the deep neural network using the training samples to obtain a trained deep neural network, which is specifically as follows.

[0133] The selected deep neural network is trained in batches using the training samples, the loss function of the deep neural network is calculated each time, the parameters of the deep neural network are updated by back propagation, and the iteration is updated until the deep neural network converges.

[0134] In this embodiment, the number of training samples of the deep neural network is 2000, of which 1600 are used as the training set of the deep neural network and 400 are used as the test set of the deep neural network, the number of training samples in each batch is 8, and the network converges after 200,000 iterations to obtain a trained deep neural network model.

[0135] S6, specifically, importing the compiled fast high-resolution multiple echo imaging pulse sequence into the magnetic resonance imaging instrument, setting the sampling parameters, and scanning the human brain according to the imaging standard operation steps to obtain the k-space data of the human brain.

[0136] S7, specifically, preprocessing the k-space data of the human brain to obtain the image domain data of the human brain, including:

[0137] Separate the collected k-space odd and even row data, fill them into two k-spaces according to the phase encoding mode to obtain the k-space data of the human brain;

[0138] Parallel reconstruction of the two k-space data of the human brain;

[0139] Inverse Fourier transform of the two k-space data to obtain the corresponding image domain data of the human brain.

[0140] S8, inputting the image domain data of the entity imaging object into the trained deep neural network for reconstruction to obtain a high-resolution multi-parameter quantitative magnetic resonance image of the entity imaging object. The results of the embodiment are shown in Figure 4 (a) is the real part and the imaginary part of the two image domain data input into the deep neural network; (b) is the reconstructed T2 quantitative map, T2 *quantitative maps, proton density maps and dB0 maps, without visible image distortions; (c) T2 quantitative maps, T2 * quantitative maps, proton density maps and dB0 maps.

[0141] Referring to Figure 5 As an implementation of the method shown in the above figures, the embodiment provides an embodiment of a fast high-resolution multi-parameter quantitative magnetic resonance imaging system, which corresponds to the method embodiment shown in the above figures, and can be applied to various electronic devices. Figure 1 As an implementation of the method shown in the above figures, the embodiment provides an embodiment of a fast high-resolution multi-parameter quantitative magnetic resonance imaging system, which corresponds to the method embodiment shown in the above figures, and can be applied to various electronic devices.

[0142] Specifically, a fast high-resolution multi-parameter quantitative magnetic resonance imaging system comprises:

[0143] A pulse sequence design module 501 is configured to design a fast high-resolution multi-echo imaging pulse sequence.

[0144] A sampling parameter determination module 502 is configured to determine sampling parameters of the fast high-resolution multi-echo imaging pulse sequence.

[0145] A deep neural network construction module 503 is configured to construct a deep neural network for multi-parameter quantitative magnetic resonance image reconstruction.

[0146] A training sample generation module 504 is configured to generate training samples of the deep neural network.

[0147] A deep neural network training module 505 is configured to train the deep neural network using the training samples to obtain a trained deep neural network.

[0148] A k-space data acquisition module 506 is configured to scan a physical imaging object using the fast high-resolution multi-echo imaging pulse sequence under the set sampling parameters to obtain k-space data of the physical imaging object.

[0149] An image domain data acquisition module 507 is configured to preprocess the k-space data of the physical imaging object to obtain image domain data of the physical imaging object.

[0150] A multi-parameter quantitative magnetic resonance image acquisition module 508 is configured to input the image domain data of the physical imaging object into the trained deep neural network for reconstruction to obtain a high-resolution multi-parameter quantitative magnetic resonance image of the physical imaging object.

[0151] A specific implementation of a fast high-resolution multi-parameter quantitative magnetic resonance imaging system is the same as a fast high-resolution multi-parameter quantitative magnetic resonance imaging method, and the embodiment will not be repeated.

[0152] The above examples are only for illustrating the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made according to the spirit and essence of the present application shall be covered within the protection scope of the present application.

Claims

1. A rapid, high-resolution, multi-parameter quantitative magnetic resonance imaging method, characterized in that, include: Design a fast, high-resolution multi-overlapping echo imaging pulse sequence; Determine the sampling parameters for a fast, high-resolution multi-overlap echo imaging pulse sequence; Construct a deep neural network for multi-parameter quantitative magnetic resonance image reconstruction; Generate training samples for deep neural networks; The deep neural network is trained using training samples to obtain a trained deep neural network. The k-space data of the physical imaging object is obtained by scanning the object with a fast high-resolution multi-overlapping echo imaging pulse sequence under set sampling parameters. Preprocess the k-space data of the entity imaging object to obtain the image domain data of the entity imaging object; The image domain data of the physical imaging object is input into a trained deep neural network for reconstruction to obtain a high-resolution multi-parameter quantitative magnetic resonance image of the physical imaging object. The fast high-resolution multi-overlap echo imaging pulse sequence includes a signal excitation module and a data acquisition module. The signal excitation module includes N time intervals t. n The flip angle is α n RF excitation pulse and corresponding layer selection gradient G ss and echo shift gradient G n Where n = 1, 2, ..., N, N ≥ 2; Each radio frequency excitation pulse is associated with a layer selection gradient G in the layer selection direction. ss Combined layer selection; Echo shift gradient G n The frequency encoding dimension and the phase encoding dimension are applied after the corresponding radio frequency excitation pulse; The data acquisition module includes a pre-phase gradient G. pre Destroy gradient G cr Frequency coding gradient G ro M time intervals are ESP and the flip angle is β. m Refocusing pulse, phase-coded gradient G pe,i,m and dephase-encoded gradient G pe,i,m ′; where i represents the i-th scan, m = 1, 2, ..., M, M ≥ 2; Prephase gradient G pre Apply along the frequency coding dimension, with a magnitude of the frequency coding gradient G. ro Half of; Destroy gradient G cr Before and after each refocusing pulse, directions are selected along the frequency encoding dimension, phase encoding dimension, and plane, with equal magnitude and the same direction; Phase-encoded gradient G pe,i,m Frequency coding gradient G ro and dephase-encoded gradient G pe,i,m After each refocusing pulse, the phase-encoded gradient and the dephase-encoded gradient are applied with equal magnitudes but opposite directions.

2. The rapid high-resolution multi-parameter quantitative magnetic resonance imaging method according to claim 1, characterized in that, The determination of the sampling parameters for the fast high-resolution multi-overlap echo imaging pulse sequence specifically includes: Determine the number N of radio frequency excitation pulses and the flip angle α of each radio frequency excitation pulse. n ; The composition of the echo signal is determined, thereby determining the time interval t between each radio frequency pulse. n ; By determining the position of each excitation signal in k-space, the echo shift gradient G after each RF excitation pulse can be determined. n The proportion; Determine the number M of refocusing pulses and the flip angle β of each refocusing pulse. m ; The phase encoding method is determined, thereby determining the direction and magnitude of each phase encoding gradient and dephase encoding gradient; The numerical values ​​of various gradients in the pulse sequence are determined by determining the number of pulse sequence scans, imaging field of view, imaging matrix, sampling frequency, time interval between refocusing pulses (ESP), sampling acceleration factor, and pulse sequence repetition time (TR).

3. The rapid high-resolution multi-parameter quantitative magnetic resonance imaging method according to claim 1, characterized in that, The construction of the deep neural network for multi-parameter quantitative magnetic resonance image reconstruction specifically includes: Constructing the network structure of a deep neural network; Construct the number of input channels and output channels for a deep neural network; Construct the loss function for a deep neural network.

4. The rapid high-resolution multi-parameter quantitative magnetic resonance imaging method according to claim 1, characterized in that, The training samples for generating the deep neural network specifically include: Based on the characteristics of the physical imaging object, simulated proton density maps, T1 quantitative maps, T2 quantitative maps, and T2 maps are generated. * Quantitative images are used to form a virtual imaging object; Establish a corresponding non-ideal factor model based on the non-ideal factors present in actual scanning; The virtual imaging object is simulated using the fast high-resolution multi-overlapping echo imaging pulse sequence and the non-ideal factor model to obtain the k-space data of the virtual imaging object. The k-space data of the virtual imaging object is separated into odd and even rows and filled into two k-spaces respectively to obtain the k-space data of the virtual imaging object. The inverse Fourier transform is then performed to obtain the image domain data of the virtual imaging object, which together with the virtual imaging object and the non-ideal factor model constitute a training sample. Repeat the above steps to generate the set number of training samples.

5. The rapid high-resolution multi-parameter quantitative magnetic resonance imaging method according to claim 1, characterized in that, The preprocessing of the k-space data of the entity imaging object to obtain the image domain data of the entity imaging object specifically includes: The collected k-space odd and even row data are separated and filled into two k-spaces according to the phase encoding method to obtain the k-space data of the entity imaging object. If the sampling acceleration factor is greater than 1, the two k-space data of the entity imaging object are reconstructed in parallel. Perform inverse Fourier transform on the two k-space data respectively to obtain the image domain data of the corresponding entity imaging object.

6. A rapid, high-resolution, multi-parameter quantitative magnetic resonance imaging system, characterized in that, include: The pulse sequence design module is used to design fast, high-resolution, multi-overlapping echo imaging pulse sequences. The sampling parameter determination module is used to determine the sampling parameters of the fast high-resolution multi-overlap echo imaging pulse sequence; A deep neural network building module for constructing deep neural networks for multi-parameter quantitative magnetic resonance image reconstruction; The training sample generation module is used to generate training samples for deep neural networks. The deep neural network training module is used to train a deep neural network using training samples to obtain a trained deep neural network. The k-space data acquisition module is used to scan the entity imaging object using a fast high-resolution multi-overlapping echo imaging pulse sequence under set sampling parameters to obtain the k-space data of the entity imaging object. The image domain data acquisition module is used to preprocess the k-space data of the entity imaging object to obtain the image domain data of the entity imaging object. The multi-parameter quantitative magnetic resonance image acquisition module is used to input the image domain data of the entity imaging object into a trained deep neural network for reconstruction, thereby obtaining a high-resolution multi-parameter quantitative magnetic resonance image of the entity imaging object. The fast high-resolution multi-overlap echo imaging pulse sequence includes a signal excitation module and a data acquisition module. The signal excitation module includes N time intervals t. n The flip angle is α n RF excitation pulse and corresponding layer selection gradient G ss and echo shift gradient G n Where n = 1, 2, ..., N, N ≥ 2; Each radio frequency excitation pulse is associated with a layer selection gradient G in the layer selection direction. ss Combined layer selection; Echo shift gradient G n The frequency encoding dimension and the phase encoding dimension are applied after the corresponding radio frequency excitation pulse; The data acquisition module includes a pre-phase gradient G. pre Destroy gradient G cr Frequency coding gradient G ro M time intervals are ESP and the flip angle is β. m Refocusing pulse, phase-coded gradient G pe,i,m and dephase-encoded gradient G pe,i,m ′; where i represents the i-th scan, m = 1, 2, ..., M, M ≥ 2; Prephase gradient G pre Apply along the frequency coding dimension, with a magnitude of the frequency coding gradient G. ro Half of; Destroy gradient G cr Before and after each refocusing pulse, directions are selected along the frequency encoding dimension, phase encoding dimension, and plane, with equal magnitude and the same direction; Phase-encoded gradient G pe,i,m Frequency coding gradient G ro and dephase-encoded gradient G pe,i,m After each refocusing pulse, the phase-encoded gradient and the dephase-encoded gradient are applied with equal magnitudes but opposite directions.

Images