Method and system for flow artifact correction of multi-scan magnetic resonance images

Abstract

The application discloses a kind of multi-scan magnetic resonance image flow artifact correction method and system, comprising: in magnetic resonance imaging instrument import magnetic resonance imaging pulse sequence and according to the set sampling parameter to actual imaging object data acquisition, obtain the multi-scan magnetic resonance imaging signal of actual imaging object;Actual imaging object multi-scan magnetic resonance imaging signal is handled, and the image of actual imaging object with flow artifact is obtained;Generation training sample, including paired flow artifact simulation sample and flow artifact-free simulation sample;Using training sample to train deep neural network, obtain trained deep neural network;The image of actual imaging object with flow artifact is input into trained deep neural network and is corrected to flow artifact, and the image without flow artifact is obtained.The application can realize the fast acquisition of high-resolution magnetic resonance image without flow artifact without additional acquisition of navigation echo signal for flow artifact correction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of magnetic resonance imaging, in particular to a method and system for flow artifact correction of multi-scan magnetic resonance images. BACKGROUND

[0002] To improve the resolution of magnetic resonance images, multi-scan technology is introduced into magnetic resonance data acquisition. Multi-scan refers to multiple excitation scans, which segment the k-space signal along the phase encoding direction, and the signal collected each time is re-excited by the same radio frequency pulse. This technology can shorten the sampling echo chain of EPI, avoid the problem of serious reduction of signal-to-noise ratio, and thus realize the acquisition of high-resolution images. However, due to the need to wait for a period of time between two scans, the acquisition time is longer than that of single scan. During data acquisition, the pseudo-periodic motion of free water in cerebrospinal fluid (such as cerebrospinal fluid pulsation affected by respiration and heartbeat) will introduce different phase changes in different scans, resulting in flow artifacts in the final full-sampling image, which reduces its medical application value. SUMMARY

[0003] The present application aims to overcome the defects and deficiencies in the prior art, and provides a method and system for flow artifact correction of multi-scan magnetic resonance images, which does not require additional acquisition of navigation echo signals for flow artifact correction. The deep neural network can be used to correct the flow artifacts of multi-scan magnetic resonance images, and the high-resolution magnetic resonance images without flow artifacts can be quickly obtained.

[0004] The technical solutions of the present application are as follows.

[0005] On the one hand, a method for flow artifact correction of multi-scan magnetic resonance images comprises:

[0006] S1: determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; the multi-scan magnetic resonance imaging pulse sequence is any magnetic resonance imaging sequence that uses radio frequency pulses to generate spin echo signals and uses EPI technology to read out signals;

[0007] S2: importing the multi-scan magnetic resonance imaging pulse sequence into a magnetic resonance imaging instrument and collecting data on an actual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the actual imaging object;

[0008] S3: rearranging, splicing, zero padding, domain transforming and normalizing the multi-scan magnetic resonance imaging signals of the actual imaging object to obtain an image of the actual imaging object with flow artifacts;

[0009] S4: generating training samples of the deep neural network; the training samples include paired simulation samples with flow artifacts and simulation samples without flow artifacts; the simulation samples with flow artifacts are used as inputs of the deep neural network, and the simulation samples without flow artifacts are used as labels of the deep neural network; specifically comprising:

[0010] S41: generating a virtual imaging object;

[0011] S42: compiling the multi-scan magnetic resonance imaging pulse sequence on a simulation platform and simulating data acquisition of the virtual imaging object according to determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the virtual imaging object;

[0012] S43: generating simulation samples with flow artifacts; adding inter-scan phase changes to the multi-scan magnetic resonance imaging signals of the virtual imaging object to simulate flow artifacts caused by cerebrospinal fluid pulsation of an actual imaging object, and performing rearrangement, splicing, zero padding, domain transformation, normalization and noise adding processing on the signals to obtain simulation samples with flow artifacts;

[0013] S44: generating simulation samples without flow artifacts; the simulation samples without flow artifacts are set as magnetic resonance weighted images or magnetic resonance parameter maps according to requirements; if the image to be reconstructed is a magnetic resonance weighted image, the multi-scan magnetic resonance imaging signals of the virtual imaging object are rearranged, spliced, zero-padded, domain-transformed and normalized to obtain the simulation samples without flow artifacts; if the image to be reconstructed is a magnetic resonance parameter map, the virtual imaging object is used as the simulation sample without flow artifacts;

[0014] S45: repeating the processes of S41-S44 until a set amount of deep neural network training samples are generated;

[0015] S5: training the deep neural network using the training samples to obtain a trained deep neural network;

[0016] S6: inputting an image with flow artifacts of an actual imaging object into the trained deep neural network for flow artifact correction to obtain an image without flow artifacts.

[0017] Preferably, in S1, the multi-scan magnetic resonance imaging pulse sequence and the sampling parameters are determined, specifically comprising:

[0018] determining the number n of radio frequency excitation pulses; determining the flip angle size and pulse shape of each radio frequency excitation pulse; determining the flip angle size and pulse shape of the rephasing pulse; determining the echo shift gradient area; determining the pulse sequence scan number, imaging field of view, imaging matrix, number of layers, layer thickness, interlayer spacing, bandwidth and echo time; determining the repetition time of the pulse sequence.

[0019] Preferably, in the S41, the virtual imaging object is generated, specifically comprising:

[0020] The virtual imaging object includes a T2 map, a T1 map and a PD map, which are calculated from a registered magnetic resonance T1 weighted image T1w, a magnetic resonance T2 weighted image T2w and a magnetic resonance PD weighted image PDw in a common data set by a magnetic resonance signal formula, and the analytical form of the magnetic resonance signal formula is expressed as follows:

[0021]

[0022] Wherein, SI represents the signal intensity of the magnetic resonance image; PD represents the proton density; TE represents the echo time of the magnetic resonance image acquisition; T2 represents the transverse relaxation time; TR represents the repetition time of the magnetic resonance image acquisition; T1 represents the longitudinal relaxation time;

[0023] For the magnetic resonance PD weighted image, according to the following magnetic resonance signal formula:

[0024] PD≈SI PDw

[0025] The PD map is obtained, and after normalization, it is used as the PD map of the virtual imaging object; wherein, SI PDw represents the signal intensity of the magnetic resonance PD weighted image;

[0026] For the magnetic resonance T2 weighted image, according to the following magnetic resonance signal formula:

[0027]

[0028] The T2 map is obtained, and after scaling within a certain fluctuation range according to the T2 value range of the actual imaging object, it is used as the T2 map of the virtual imaging object; wherein, SI T2w represents the signal intensity of the magnetic resonance T2 weighted image;

[0029] For the magnetic resonance T1 weighted image, according to the following magnetic resonance signal formula:

[0030]

[0031] The T1 map is obtained, and after scaling within a certain fluctuation range according to the T1 value range of the actual imaging object, it is used as the T1 map of the virtual imaging object; wherein, SI T1w represents the signal intensity of the magnetic resonance T1 weighted image.

[0032] Preferably, in S43, inter-scan phase changes are added to the multi-scan magnetic resonance imaging signals of the virtual imaging object to simulate the flow artifacts of the actual imaging object caused by the pulsation of cerebrospinal fluid, and the signals are rearranged, spliced, zero-filled, domain-transformed, normalized, and noise-added to obtain a simulation sample with flow artifacts, specifically including:

[0033] The multi-scan magnetic resonance imaging signals of the virtual imaging object are rearranged and spliced into complete two-dimensional k-space signals, and then inverse two-dimensional Fourier transform is performed to obtain a magnetic resonance weighted image of the virtual imaging object.

[0034] The cerebrospinal fluid region causing the flow artifacts is extracted using the T1 map of the virtual imaging object, and a two-dimensional polynomial function is used to model the phase changes between different scans caused by the pulsation of cerebrospinal fluid, with the expression as follows:

[0035]

[0036] wherein the simulated phase change map randmap m is expressed as a two-dimensional polynomial function; subscript m represents the mth scan; x and y represent the coordinates of the two-dimensional plane; n x and n y represent the orders of x and y, respectively; r p represents the coefficients of the two-dimensional polynomial function; N p represents the highest order of the two-dimensional polynomial function; and the extreme value of the simulated phase change map is adjusted according to the actual situation.

[0037] The magnetic resonance weighted image of the virtual imaging object is multiplied by the simulated phase change map to obtain a magnetic resonance weighted image with phase changes for the corresponding scan number, which is converted to the k-space domain by two-dimensional Fourier transform, and the data of each scan with phase changes in the k-space is extracted according to the multi-scan acquisition mode, rearranged and spliced into complete two-dimensional k-space, zero-filled, inverse two-dimensional Fourier transformed to the image domain, normalized, and noise-added to obtain a simulation sample with flow artifacts.

[0038] Preferably, in S5, the training sample is used to train the deep neural network to obtain a trained deep neural network, specifically including:

[0039] The network structure, parameters, and loss function of the deep neural network are determined; when training the deep neural network, the training sample set is input into the deep neural network in batches for iterative training, the value of the loss function is calculated, the parameter value of the deep neural network is automatically adjusted according to the value to reduce the value of the loss function, the above training is repeated until the loss function converges, and the parameter of the deep neural network is saved.

[0040] In another aspect, a flow artifact correction system for multi-scan magnetic resonance images includes:

[0041] a sequence design module for determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; the multi-scan magnetic resonance imaging pulse sequence being any magnetic resonance imaging sequence that includes using radio frequency pulses to generate spin echo signals and using EPI technique for signal readout;

[0042] a signal acquisition module for importing the multi-scan magnetic resonance imaging pulse sequence into a magnetic resonance imaging instrument and acquiring data from an actual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the actual imaging object;

[0043] a signal processing module for performing reordering, splicing, zero padding, domain transformation, and normalization processing on the multi-scan magnetic resonance imaging signals of the actual imaging object to obtain an image of the actual imaging object with flow artifacts;

[0044] a training sample generation module for generating training samples for a deep neural network; the training samples including paired simulated samples with flow artifacts and simulated samples without flow artifacts; the simulated samples with flow artifacts being used as inputs for the deep neural network and the simulated samples without flow artifacts being used as labels for the deep neural network; specifically including:

[0045] a virtual imaging object generation unit for generating a virtual imaging object; the virtual imaging object including a T2 map, a T1 map, and a PD map, which are obtained from a common data set of registered magnetic resonance T1 weighted images, magnetic resonance T2 weighted images, and magnetic resonance PD weighted images through a magnetic resonance signal formula;

[0046] a simulated signal generation unit for writing the multi-scan magnetic resonance imaging pulse sequence on a simulation platform and performing simulated data acquisition on the virtual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the virtual imaging object;

[0047] a simulated sample with flow artifacts generation unit for generating simulated samples with flow artifacts; the simulated samples with flow artifacts being obtained by adding inter-scan phase changes to the multi-scan magnetic resonance imaging signals of the virtual imaging object to simulate flow artifacts caused by cerebrospinal fluid pulsation in an actual imaging object, and performing reordering, splicing, zero padding, domain transformation, normalization, and noise adding processing on the signals;

[0048] The no-flowing-artifact sample generation unit is configured to generate a simulated sample without flowing artifacts; the simulated sample without flowing artifacts is set as a magnetic resonance weighted image or a magnetic resonance parameter map according to requirements; if the image to be reconstructed is a magnetic resonance weighted image, the multi-scan magnetic resonance imaging signals of the virtual imaging object are rearranged, spliced, zero-filled, domain-transformed and normalized to serve as the simulated sample without flowing artifacts; if the image to be reconstructed is a magnetic resonance parameter map, the virtual imaging object is used as the simulated sample without flowing artifacts;

[0049] The repeating processing unit is configured to repeatedly execute the virtual imaging object generation unit, the simulated signal generation unit, the flowing-artifact sample generation unit and the no-flowing-artifact sample generation unit until a set amount of deep neural network training samples are generated.

[0050] The network training module is configured to train a deep neural network using the training samples to obtain a trained deep neural network.

[0051] The flowing-artifact correction module is configured to input an image of an actual imaging object with flowing artifacts into the trained deep neural network for flowing-artifact correction to obtain an image without flowing artifacts.

[0052] Compared with the prior art, the present application has the following beneficial effects:

[0053] (1) The present application can realize flowing-artifact correction of images acquired by any multi-scan magnetic resonance imaging sequence using radio frequency pulse to generate spin echo signals and EPI technology for signal readout.

[0054] (2) The present application does not need to correct flowing artifacts by means of additional acquired navigation echo signals for flowing-artifact correction, thereby reducing the time cost of data acquisition.

[0055] (3) The present application uses simulated samples to make a training set of a deep neural network, without the need to acquire a large amount of paired magnetic resonance data as a training set, thereby saving time and resources; the trained deep neural network is used to realize flowing-artifact correction of images, which is faster and has higher image quality.

[0056] (4) The present application can not only correct flowing artifacts of magnetic resonance weighted images, but also synchronize flowing-artifact correction with magnetic resonance parameter quantification, thereby avoiding a cumbersome iterative solving process or exponential fitting quantification process. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings described below only show some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0058] Figure 1 Flowchart of the flow artifact correction method for the multi-scan magnetic resonance image of the embodiment of the present application.

[0059] Figure 2 Pulse sequence diagram of the multi-scan magnetic resonance imaging of the embodiment of the present application; wherein (a) represents a multi-scan spin echo EPI sequence, and (b) represents a multi-scan multi-shot echo EPI sequence.

[0060] Figure 3 Sampling trajectory of the multi-scan magnetic resonance imaging signal in the embodiment of the present application.

[0061] Figure 4 Image with flow artifact acquired by the multi-scan spin echo EPI sequence and the image thereof after flow artifact correction by the trained deep neural network in the embodiment of the present application; wherein (a) is the image acquired by the multi-scan spin echo EPI sequence, (b) is the image after flow artifact correction of (a), and (c) is the image acquired by the fast spin echo sequence.

[0062] Figure 5 Image with flow artifact acquired by the multi-scan multi-shot echo EPI sequence and the image thereof after flow artifact correction by the trained deep neural network in the embodiment of the present application; wherein (a) is the real part image of the multi-scan multi-shot echo image, (b) is the imaginary part image of the multi-scan multi-shot echo image, (c) is the T2 image of the multi-scan multi-shot echo image without flow artifact correction, (d) is the T2 image of the multi-scan multi-shot echo image after flow artifact correction, and (e) is the T2 image fitted from the traditional spin echo image.

[0063] Figure 6 Structural block diagram of the flow artifact correction system for the multi-scan magnetic resonance image of the embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present application more clear, the following will further describe the present application in combination with the drawings. Obviously, the described embodiments only show some of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0065] Referring to Figure 1 As shown in the embodiments, a flow artifact correction method for multi-scan magnetic resonance images is disclosed, comprising:

[0066] S1: determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; the multi-scan magnetic resonance imaging pulse sequence is any magnetic resonance imaging sequence containing the generation of spin echo signals by radio frequency pulses and the signal reading by EPI technology;

[0067] S2: importing the multi-scan magnetic resonance imaging pulse sequence into a magnetic resonance imaging instrument and collecting data of an actual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the actual imaging object;

[0068] S3: rearranging, splicing, zero padding, domain transforming, and normalizing the multi-scan magnetic resonance imaging signals of the actual imaging object to obtain an image of the actual imaging object with flow artifacts;

[0069] S4: generating training samples of a deep neural network; the training samples include paired simulation samples with flow artifacts and simulation samples without flow artifacts; the simulation sample with flow artifacts is used as the input of the deep neural network, and the simulation sample without flow artifacts is used as the label of the deep neural network;

[0070] S5: training the deep neural network using the training samples to obtain a trained deep neural network;

[0071] S6: inputting the image of the actual imaging object with flow artifacts into the trained deep neural network for flow artifact correction to obtain an image without flow artifacts.

[0072] Specifically, in S1, the multi-scan magnetic resonance imaging pulse sequence and its sampling parameters are determined, specifically including:

[0073] determining the number n of radio frequency excitation pulses; determining the flip angle size and pulse shape of each radio frequency excitation pulse; determining the flip angle size and pulse shape of the refocusing pulse; determining the echo shift gradient area; determining the pulse sequence scan number, imaging field of view, imaging matrix, acquisition layer number, layer thickness, layer spacing, bandwidth, and echo time; determining the repetition time of the pulse sequence.

[0074] Specifically, in S4, the training samples of the deep neural network are generated, specifically as follows:

[0075] S41: generating a virtual imaging object;

[0076] Specifically, the virtual imaging object includes a T2 map, a T1 map and a PD map, which are calculated from a magnetic resonance T1 weighted image (T1w), a magnetic resonance T2 weighted image (T2w) and a magnetic resonance PD weighted image (PDw) in a common data set by a magnetic resonance signal formula. The analytical form of the magnetic resonance signal formula is expressed as follows:

[0077]

[0078] wherein, SI represents the signal intensity of the magnetic resonance image, PD represents the proton density; TE represents the echo time of the magnetic resonance image acquisition; T2 represents the transverse relaxation time; TR represents the repetition time of the magnetic resonance image acquisition; T1 represents the longitudinal relaxation time;

[0079] For the magnetic resonance PD weighted image, according to the following magnetic resonance signal formula:

[0080] PD≈SI PDw

[0081] The PD map is obtained, and after normalization, it is used as the PD map of the virtual imaging object; wherein, SI PDw represents the signal intensity of the magnetic resonance PD weighted image;

[0082] For the magnetic resonance T2 weighted image, according to the following magnetic resonance signal formula:

[0083]

[0084] The T2 map is obtained, and after scaling within a certain fluctuation range according to the T2 value range of the actual imaging object, it is used as the T2 map of the virtual imaging object; wherein, SI T2w represents the signal intensity of the magnetic resonance T2 weighted image;

[0085] For the magnetic resonance T1 weighted image, according to the following magnetic resonance signal formula:

[0086]

[0087] The T1 map is obtained, and after scaling within a certain fluctuation range according to the T1 value range of the actual imaging object, it is used as the T1 map of the virtual imaging object; wherein, SI T1w represents the signal intensity of the magnetic resonance T1 weighted image;

[0088] S42: compiling the multi-scan magnetic resonance imaging pulse sequence on the simulation platform and performing simulated data acquisition on the virtual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the virtual imaging object;

[0089] S43: generating a simulation sample with flow artifacts; adding inter-scan phase changes to the multi-scan magnetic resonance imaging signals of the virtual imaging object to simulate flow artifacts caused by cerebrospinal fluid pulsation, and performing rearrangement, splicing, zero padding, domain transformation, normalization and noise adding on the signals to obtain a simulation sample with flow artifacts;

[0090] Specifically, the multi-scan magnetic resonance imaging signals of the virtual imaging object are rearranged and spliced into complete two-dimensional k-space signals, and then two-dimensional inverse Fourier transform is performed to obtain a magnetic resonance weighted image of the virtual imaging object;

[0091] The cerebrospinal fluid region causing flow artifacts is extracted using the T1 map of the virtual imaging object, and a two-dimensional polynomial function is used to model the phase changes between different scans caused by cerebrospinal fluid pulsation, and the expression is as follows:

[0092]

[0093] Wherein, the simulated phase change map randmap m is expressed as a two-dimensional polynomial function; subscript m represents the mth scan; x and y represent the coordinates of the two-dimensional plane; n x and n y represent the order of x and y, respectively; r p represents the coefficient of the two-dimensional polynomial function; N p represents the highest order of the two-dimensional polynomial function; the extreme value of the simulated phase change map is adjusted according to the actual situation;

[0094] The magnetic resonance weighted image of the virtual imaging object is multiplied by the simulated phase change map to obtain a magnetic resonance weighted image with phase change for the corresponding scan number, and two-dimensional Fourier transform is performed to convert to the k-space domain. According to the acquisition mode of multi-scan, the data of each scan with phase change in the k-space is extracted, rearranged and spliced into complete two-dimensional k-space, and after zero padding, two-dimensional inverse Fourier transform is performed to convert to the image domain. After normalization and noise adding, a simulation sample with flow artifacts is obtained;

[0095] S44: generating a simulation sample without flow artifacts; the simulation sample without flow artifacts can be set as a magnetic resonance weighted image or a magnetic resonance parameter map according to requirements; if the image to be reconstructed is a magnetic resonance weighted image, the multi-scan magnetic resonance imaging signals of the virtual imaging object are rearranged, spliced, zero padded, domain transformed and normalized to obtain the simulation sample without flow artifacts; if the image to be reconstructed is a magnetic resonance parameter map, the virtual imaging object is used as a simulation sample without flow artifacts;

[0096] S45: repeating the processes of S41-S44 until a set amount of deep neural network training samples are generated;

[0097] Specifically, in the S5, the training sample is used to train the deep neural network to obtain a trained deep neural network, and the training specifically includes:

[0098] determining the network structure of the deep neural network; determining the parameters of the deep neural network; and determining the loss function for training the deep neural network. When training the deep neural network, the training sample set is input into the deep neural network in batches for iterative training, the value of the loss function is calculated, the parameter value of the deep neural network is automatically adjusted according to the value to reduce the value of the loss function, and the above training is repeated until the loss function converges, and the parameter of the deep neural network is saved.

[0099] The process of the flow artifact correction method for multi-scan magnetic resonance images will be described in detail through a specific embodiment as follows, including the following steps.

[0100] Step 1: determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; in the embodiment, the multi-scan magnetic resonance imaging pulse sequence is referred to as Figure 2 , in which two multi-scan magnetic resonance imaging pulse sequences using EPI technology for signal readout are shown, (a) represents a multi-scan spin echo EPI sequence, and (b) represents a multi-scan multi-shot echo EPI sequence.

[0101] Step 2: using the multi-scan magnetic resonance imaging pulse sequence and its sampling parameters determined in step 1 to collect data of an actual imaging object to obtain multi-scan magnetic resonance imaging signals of the actual imaging object.

[0102] Step 3: rearranging, splicing, zero padding, domain transforming, and normalizing the multi-scan magnetic resonance imaging signals of the actual imaging object to obtain an image of the actual imaging object with flow artifacts; taking two scans as an example, first, the data of the first scan and the data of the second scan are rearranged in an interlaced arrangement, Figure 3 shows the rearranged multi-scan magnetic resonance imaging signals (each point represents a k x data, and the gray and black solid circles represent actual collected data, and the hollow circles represent uncollected data), and then the rearranged data of different scans are spliced together to obtain complete k-space data.

[0103] Step 4: generating training samples of a deep neural network. Specifically, it includes:

[0104] Step 41: generating a virtual imaging object;

[0105] Step 42: writing the multi-scan magnetic resonance imaging pulse sequence on a simulation platform and simulating data collection of the virtual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the virtual imaging object;

[0106] Step 43: generating a simulation sample with flow artifacts; adding inter-scan phase variations to the multi-scan magnetic resonance imaging signals of the virtual imaging object to simulate flow artifacts caused by cerebrospinal fluid pulsation of the actual imaging object, and performing reordering, stitching, zero padding, domain transformation, normalization and noise adding to the signals to obtain a simulation sample with flow artifacts;

[0107] Step 44: generating a simulation sample without flow artifacts; the simulation sample without flow artifacts can be set as a magnetic resonance weighted image or a magnetic resonance parameter map according to requirements; if the image to be reconstructed is a magnetic resonance weighted image, the multi-scan magnetic resonance imaging signals of the virtual imaging object are reordered, stitched, zero-padded, domain-transformed and normalized to obtain the simulation sample without flow artifacts; if the image to be reconstructed is a magnetic resonance parameter map, the virtual imaging object is used as the simulation sample without flow artifacts; for different multi-scan magnetic resonance imaging pulse sequences, corresponding simulation samples without flow artifacts need to be generated. For multi-scan spin echo EPI sequence: the image to be reconstructed is a magnetic resonance weighted image without flow artifacts, so the multi-scan magnetic resonance imaging signals of the virtual imaging object are reordered, stitched, zero-padded, domain-transformed and normalized to obtain the simulation sample without flow artifacts. For multi-scan multi-shot echo EPI sequence: the image to be reconstructed is a T2 map, so the virtual imaging object is used as the simulation sample without flow artifacts;

[0108] Step 45: repeating the process of steps 41-44 until a set amount of deep neural network training samples are generated.

[0109] Step 5: training the deep neural network; the deep neural network is trained using the training samples generated in step 4.

[0110] Step 6: using the deep neural network trained in step 5 to correct the flow artifacts of the actual imaging object in step 3.

[0111] In the embodiment, in step 1, the sequence diagram of the multi-scan spin echo EPI sequence is as shown in Figure 2 (a), wherein N shot is the number of scans; α is the radio frequency excitation pulse; β is the rephasing pulse; G ro is the pre-phase gradient of the frequency encoding dimension; G pe is the pre-phase gradient of the phase encoding dimension; G cr is the destruction gradient; G preThe pre-phase encoding gradient is denoted by TE; echo time is denoted by TR; and the repetition time of the pulse sequence between two scans is denoted by TR. The sampling parameters of the multi-scan spin echo EPI sequence are as follows: the number of RF excitation pulses is 1; the RF excitation pulse is a Sinc pulse with a flip angle of 30°; the refocusing pulse is a Sinc pulse with a flip angle of 180°; the number of pulse sequence scans is 4; the imaging field of view is 220mm×220mm; the imaging matrix is ​​256×260; the number of acquisition layers is 21; the layer thickness is 3mm; the interlayer interval is 1mm; the bandwidth is 1220Hz / Px; the ratio of the pre-phase gradient area in the frequency encoding dimension to the area of ​​the first readout gradient is -1 / 2; the ratio of the pre-phase gradient area in the phase encoding dimension to the total area of ​​the blip gradient is -1 / 2; TE is 74.0ms; and TR is 5000ms.

[0112] In this embodiment, the sequence diagram of the multi-scan multi-overlap echo EPI sequence in step 1 is as follows: Figure 2 (b), where N shot α is the number of scans; β is the radio frequency excitation pulse; G is the refocusing pulse; ro1 G ro2 G ro3 G ro4 These are the four shift gradients of the frequency coding dimension; G pe1 G pe2 G pe3 G pe4 These are the four shift gradients of the phase encoding dimension; G cr To destroy the gradient; G pre The pre-phase encoding gradient is defined as follows: TE1, TE2, TE3, and TE4 are the four echo times; TR is the repetition time of the pulse sequence between two scans. The sampling parameters for the multi-scan, multi-overlap echo EPI sequence are: 4 RF excitation pulses; all RF excitation pulses are Sinc pulses with a 30° flip angle; the refocusing pulse is a Sinc pulse with a 180° flip angle; the number of pulse sequence scans is 4; the imaging field of view is 220mm × 220mm; the imaging matrix is ​​256 × 260; the number of acquisition layers is 21; the layer thickness is 3mm; the interlayer interval is 1mm; the bandwidth is 1220Hz / Px; and the frequency encoding dimension is four. The ratios of the shift gradient area to the first readout gradient area are -74 / 256, 74 / 256, -74 / 256, and -93 / 256, respectively; the ratios of the four shift gradient areas to the total blip gradient area in the phase encoding dimension are -64 / 260, -88 / 260, -24 / 260, and -11 / 260, respectively; the four TEs are 12.4ms, 25.9ms, 75.0ms, and 110.7ms, respectively; and the TR is 5000ms.

[0113] In step 3, the multi-scan magnetic resonance imaging signals of the actual imaging object are rearranged and spliced into a two-dimensional k-space signal (matrix size 256*260), and after zero padding (matrix size 512*512), two-dimensional inverse Fourier transform is performed to convert the signal to the image domain, and after normalization, an image with flow artifacts of the actual imaging object is obtained.

[0114] In step 4, the inter-scan phase change added to the multi-scan magnetic resonance imaging signals of the virtual imaging object is determined to be randomly obtained within the range of [-π / 5, π / 5] through phase analysis of the actual imaging object with flow artifacts image.

[0115] In step 5, the deep neural network adopts U-Net. The Adam optimizer is used, and the learning rate is adjusted in an exponentially decaying manner: the initial learning rate is 0.0001, and the learning rate is reduced by 20% every 80000 iterations.

[0116] In step 5, for multi-scan spin echo EPI sequence: the input of the deep neural network adopts the form of amplitude, that is, the input is 1 magnetic resonance weighted image with flow artifacts, and the output is 1 magnetic resonance weighted image without flow artifacts; the number of training samples is 3000 (10% for test set); the loss function of the deep neural network uses the pixel-based mean square error.

[0117] In step 5, for multi-scan multi-echo EPI sequence: the input of the deep neural network adopts the form of real part and imaginary part separation, that is, the input is 2 multi-echo images with flow artifacts; the output is 1 T2 image without flow artifacts; the number of training samples is 5800 (10% for test set), half of which do not add phase change; the loss function I1 is as follows:

[0118]

[0119] Where N is the number of training samples in a batch, f is the nonlinear mapping represented by the network, W and b are the parameters of each convolution kernel, and x. k represents the kth simulated sample with flow artifacts, y k represents the kth simulated sample without flow artifacts, y ch represents an inverse normalization matrix, and · represents Hadamard product. In order to reduce noise while preserving edges, a regularization term is added to the objective function, where y mask is the edge of the label image extracted using the Canny operator, the threshold value of edge extraction is σ=0.1, and λ represents the regularization term parameter, represents the gradient operator.y ch as follows:

[0120]

[0121] where a is the threshold of the inverse normalization matrix truncation. In the training, a = 0.05s, and λ = 2 x 10 -6 .

[0122] In order to evaluate the flow artifact correction method of the multi-scan magnetic resonance image of the present application, the present embodiment shows the image with flow artifacts acquired by the multi-scan spin echo EPI sequence and the image after the flow artifact correction by the trained deep neural network in Figure 4 . At the same time, the present embodiment provides the image acquired by the fast spin echo sequence as a reference of the texture structure in the brain. Among them, (a) is the image acquired by the multi-scan spin echo EPI sequence, (b) is the image after the flow artifact correction of (a), and (c) is the image acquired by the fast spin echo sequence. The arrow in the figure points to the flow artifact. By comparing (b) and (c), it can be seen that the present application can well correct the flow artifact of the image and clearly reconstruct the image details.

[0123] Figure 5 For the present embodiment, the T2 map of the image with flow artifacts acquired by the multi-scan multi-echo EPI sequence and the T2 map obtained after the flow artifact correction by the trained deep neural network. At the same time, the present embodiment provides the T2 map obtained by fitting the image acquired by the traditional spin echo sequence as a reference. Among them, (a) and (b) are the real part and imaginary part images of the multi-scan multi-echo image, (c) is the T2 map obtained without flow artifact correction of the multi-scan multi-echo image, (d) is the T2 map obtained after the flow artifact correction of the multi-scan multi-echo image, and (e) is the T2 map obtained by fitting the traditional spin echo image. The arrow in the figure points to the flow artifact. By comparing (c) and (d) with (e), it can be seen that the present application can well correct the flow artifact of the image and clearly reconstruct the image details, and obtain a result similar to the reference map. The present application not only can correct the flow artifact of the magnetic resonance weighted image, but also can synchronize the flow artifact correction with the magnetic resonance parameter quantification, avoiding the cumbersome iterative solving process or exponential fitting quantification process, and realizing the fast acquisition of the high-resolution magnetic resonance image without flow artifact.

[0124] Referring to Figure 6 , the present embodiment further discloses a flow artifact correction system for multi-scan magnetic resonance images, comprising:

[0125] a sequence design module 61 for determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; the multi-scan magnetic resonance imaging pulse sequence is any magnetic resonance imaging sequence containing a spin echo signal generated by a radio frequency pulse and a signal readout by an EPI technology;

[0126] The signal acquisition module 62 is configured to introduce the multi-scan magnetic resonance imaging pulse sequence into the magnetic resonance imaging device and perform data acquisition on the actual imaging object according to the determined sampling parameters, so as to obtain multi-scan magnetic resonance imaging signals of the actual imaging object.

[0127] The signal processing module 63 is configured to perform rearrangement, splicing, zero padding, domain transformation and normalization processing on the multi-scan magnetic resonance imaging signals of the actual imaging object, so as to obtain an image of the actual imaging object with flow artifacts.

[0128] The training sample generation module 64 is configured to generate training samples of the deep neural network; the training samples include paired simulation samples with flow artifacts and simulation samples without flow artifacts; the simulation sample with flow artifacts is used as the input of the deep neural network, and the simulation sample without flow artifacts is used as the label of the deep neural network; and the training sample generation module 64 specifically includes:

[0129] The virtual imaging object generation unit 641 is configured to generate a virtual imaging object; the virtual imaging object includes a T2 image, a T1 image and a PD image, which are obtained from a magnetic resonance T1 weighted image, a magnetic resonance T2 weighted image and a magnetic resonance PD weighted image in a common data set by a magnetic resonance signal formula;

[0130] The simulation signal generation unit 642 is configured to compile the multi-scan magnetic resonance imaging pulse sequence on a simulation platform and perform simulation data acquisition on the virtual imaging object according to the determined sampling parameters, so as to obtain multi-scan magnetic resonance imaging signals of the virtual imaging object;

[0131] The simulation sample with flow artifacts generation unit 643 is configured to generate a simulation sample with flow artifacts; the multi-scan magnetic resonance imaging signals of the virtual imaging object are added with inter-scan phase changes to simulate flow artifacts caused by the pulsation of cerebrospinal fluid, and the signals are subjected to rearrangement, splicing, zero padding, domain transformation, normalization and noise adding processing, so as to obtain the simulation sample with flow artifacts;

[0132] The simulation sample without flow artifacts generation unit 642 is configured to generate a simulation sample without flow artifacts; the simulation sample without flow artifacts can be set as a magnetic resonance weighted image or a magnetic resonance parameter image according to requirements; if the image to be reconstructed is a magnetic resonance weighted image, the multi-scan magnetic resonance imaging signals of the virtual imaging object are subjected to rearrangement, splicing, zero padding, domain transformation and normalization processing, and then the processed signals are used as the simulation sample without flow artifacts; if the image to be reconstructed is a magnetic resonance parameter image, the virtual imaging object is used as the simulation sample without flow artifacts;

[0133] The repeated processing unit 645 is configured to repeatedly execute the virtual imaging object generation unit, the simulation signal generation unit, the simulation sample with flow artifacts generation unit and the simulation sample without flow artifacts generation unit until a set amount of training samples of the deep neural network are generated.

[0134] a network training module 65, configured to train the deep neural network by using the training samples, to obtain a trained deep neural network;

[0135] a flow artifact correction module 66, configured to input an image of the actual imaging object with flow artifacts into the trained deep neural network for flow artifact correction, to obtain an image without flow artifacts.

[0136] A specific implementation of a flow artifact correction system for multi-scan magnetic resonance images is similar to the method for flow artifact correction of multi-scan magnetic resonance images, and the specific implementation is not repeated here.

[0137] The principles and operation of the present application are explained in the above specific embodiments. These examples are intended to provide a clear understanding of the framework for the reader to grasp the core ideas and operational points of the present application. However, it should be clear that these examples are not a limitation on the scope of application, and various forms of improvement and innovation based on the core ideas of the present application should be included within the scope of protection of the present application for those skilled in the art.

Claims

1. A method of flow artifact correction of a multi-scan magnetic resonance image, characterized in that, The method comprises the following steps: S1: determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; the multi-scan magnetic resonance imaging pulse sequence is any magnetic resonance imaging sequence comprising generating spin echo signals by using radio frequency pulses and reading out signals by using EPI technology; S2: introducing the multi-scan magnetic resonance imaging pulse sequence into a magnetic resonance imaging instrument and collecting data of an actual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the actual imaging object; S3: performing rearrangement, splicing, zero padding, domain transformation and normalization processing on the multi-scan magnetic resonance imaging signals of the actual imaging object to obtain an image of the actual imaging object with flow artifacts; S4: generating training samples of a deep neural network; the training samples comprise paired simulation samples with flow artifacts and simulation samples without flow artifacts; the simulation sample with flow artifacts is used as the input of the deep neural network, and the simulation sample without flow artifacts is used as the label of the deep neural network; specifically comprising: S41: generating a virtual imaging object; S42: compiling the multi-scan magnetic resonance imaging pulse sequence on a simulation platform and collecting simulated data of the virtual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the virtual imaging object; S43: generating a simulation sample with flow artifacts; adding inter-scan phase changes to the multi-scan magnetic resonance imaging signals of the virtual imaging object to simulate flow artifacts caused by the pulsation of cerebrospinal fluid of the actual imaging object, and performing rearrangement, splicing, zero padding, domain transformation, normalization and noise adding processing on the signals to obtain the simulation sample with flow artifacts; S44: generating a simulation sample without flow artifacts; the simulation sample without flow artifacts is set as a magnetic resonance weighted image or a magnetic resonance parameter map according to requirements; if the image to be reconstructed is a magnetic resonance weighted image, the multi-scan magnetic resonance imaging signals of the virtual imaging object are rearranged, spliced, zero padded, domain transformed and normalized to obtain the simulation sample without flow artifacts; if the image to be reconstructed is a magnetic resonance parameter map, the virtual imaging object is used as the simulation sample without flow artifacts; S45: repeating the processes of S41-S44 until a set amount of training samples of the deep neural network are generated; S5: training the deep neural network by using the training samples to obtain a trained deep neural network; S6: inputting the image of the actual imaging object with flow artifacts into the trained deep neural network for flow artifact correction to obtain an image without flow artifacts.

2. The flow artifact correction method of multiple scan magnetic resonance images according to claim 1, characterized in that, In S1, the multi-scan magnetic resonance imaging pulse sequence and its sampling parameters are determined, specifically comprising: determining the number n of radio frequency excitation pulses; determining the flip angle size and pulse shape of each radio frequency excitation pulse; determining the flip angle size and pulse shape of the rephasing pulse; determining the echo shift gradient area; determining the scan number, imaging field of view, imaging matrix, number of layers, layer thickness, interlayer spacing, bandwidth and echo time of the pulse sequence; determining the repetition time of the pulse sequence.

3. The flow artifact correction method of multiple scan magnetic resonance images according to claim 1, characterized in that, In S41, the virtual imaging object is generated, specifically comprising: The virtual imaging object includes a T2 image, a T1 image and a PD image, which are calculated by a magnetic resonance T1 weighted image T1w, a magnetic resonance T2 weighted image T2w and a magnetic resonance PD weighted image PDw registered in a common data set, and a magnetic resonance signal formula, the analytical form of the magnetic resonance signal formula is expressed as follows: Wherein, SI represents the signal intensity of the magnetic resonance image; PD represents the proton density; TE represents the echo time of the magnetic resonance image acquisition; T2 represents the transverse relaxation time; TR represents the repetition time of the magnetic resonance image acquisition; T1 represents the longitudinal relaxation time; For the magnetic resonance PD weighted image, according to the following magnetic resonance signal formula: PD ~ SI PDw The PD map is obtained, which is normalized to serve as the PD map of the virtual imaging object; wherein SI PDw represents the signal intensity of the magnetic resonance PD-weighted image; For the magnetic resonance T2 weighted image, according to the following magnetic resonance signal formula: The T2 map is obtained, which is scaled according to the T2 value range of the actual imaging object within a certain fluctuation range to serve as the T2 map of the virtual imaging object; wherein, SI T2w represents the signal intensity of the magnetic resonance T2 weighted image; For the magnetic resonance T1 weighted image, according to the following magnetic resonance signal formula: A T1 map is obtained, which is scaled according to the range of T1 values of the actual imaged object within a certain fluctuation range to serve as the T1 map of the virtual imaging object; wherein SI T1w represents the signal intensity of the magnetic resonance T1 -weighted image.

4. The flow artifact correction method of multiple scan magnetic resonance images of claim 1, wherein, In the S43, the inter-scan phase change is added to the multi-scan magnetic resonance imaging signal of the virtual imaging object to simulate the flow artifact caused by the cerebrospinal fluid pulsation of the actual imaging object, and the signal is rearranged, spliced, zero-filled, domain transformed, normalized and noise-added to obtain a simulation sample with flow artifact, specifically including: The multi-scan magnetic resonance imaging signal of the virtual imaging object is rearranged and spliced into complete two-dimensional k-space signal, and then inverse two-dimensional Fourier transform is performed to obtain the magnetic resonance weighted image of the virtual imaging object; The cerebrospinal fluid region causing the flow artifact is extracted from the T1 image of the virtual imaging object, and a two-dimensional polynomial function is used to model the phase change between different scans caused by the cerebrospinal fluid pulsation, and the expression is as follows: wherein the simulated phase change map randmap m is represented as a two-dimensional polynomial function; subscript m represents the mth scan; x and y represent the coordinates of a two-dimensional plane; n x and n y represent the order of x and y, respectively; r p represents the coefficients of the two-dimensional polynomial function; N p represents the highest order of the two-dimensional polynomial function; the extreme value of the simulated phase change map is adjusted according to actual conditions; The magnetic resonance weighted image of the virtual imaging object is multiplied by the simulated phase change map to obtain the magnetic resonance weighted image with phase change of the corresponding scan number, which is converted to k-space domain by two-dimensional Fourier transform, and the data of each scan with phase change in k-space is extracted according to the multi-scan acquisition mode, rearranged and spliced into complete two-dimensional k-space, and then zero-filled, inverse two-dimensional Fourier transformed to the image domain, normalized and noise-added to obtain a simulation sample with flow artifact.

5. The flow artifact correction method of multiple scan magnetic resonance images of claim 1, wherein, In the S5, the training sample is used to train the deep neural network to obtain a trained deep neural network, specifically including: The network structure, parameters and loss function of the deep neural network are determined, and when the deep neural network is trained, the training sample set is input into the deep neural network for iterative training, the value of the loss function is calculated, the parameter value of the deep neural network is automatically adjusted according to the value to reduce the value of the loss function, and the above training is repeated until the loss function converges, and the deep neural network parameters are saved.

6. A system for flow artifact correction of multiple-scan magnetic resonance images, characterized in that Including: A sequence design module for determining a multi-scan magnetic resonance imaging pulse sequence and its sampling parameters; the multi-scan magnetic resonance imaging pulse sequence is any magnetic resonance imaging sequence containing a radio frequency pulse to generate a spin echo signal and an EPI technology to read out the signal; A signal acquisition module for importing the multi-scan magnetic resonance imaging pulse sequence into a magnetic resonance imaging instrument and acquiring data of an actual imaging object according to the determined sampling parameters to obtain multi-scan magnetic resonance imaging signals of the actual imaging object; The signal processing module is configured to perform rearrangement, splicing, zero padding, domain transformation and normalization processing on the multi-scan magnetic resonance imaging signals of the actual imaging object to obtain an image of the actual imaging object with flow artifacts. The training sample generation module is configured to generate training samples of the deep neural network, wherein the training samples include paired simulation samples with flow artifacts and simulation samples without flow artifacts, the simulation samples with flow artifacts are used as inputs of the deep neural network, and the simulation samples without flow artifacts are used as labels of the deep neural network. The virtual imaging object generation unit is configured to generate a virtual imaging object, wherein the virtual imaging object includes a T2 image, a T1 image and a PD image, and the virtual imaging object is obtained by using a common data set including a magnetic resonance T1 weighted image, a magnetic resonance T2 weighted image and a magnetic resonance PD weighted image, and a magnetic resonance signal formula. The simulation signal generation unit is configured to compile the multi-scan magnetic resonance imaging pulse sequence on a simulation platform, perform simulation data acquisition on the virtual imaging object according to determined sampling parameters, and obtain multi-scan magnetic resonance imaging signals of the virtual imaging object. The simulation sample generation unit is configured to generate simulation samples with flow artifacts, wherein the multi-scan magnetic resonance imaging signals of the virtual imaging object are added with inter-scan phase changes to simulate flow artifacts caused by the pulsation of cerebrospinal fluid in the actual imaging object, and the signals are subjected to rearrangement, splicing, zero padding, domain transformation, normalization and noise adding processing to obtain the simulation samples with flow artifacts. The simulation sample generation unit is configured to generate simulation samples without flow artifacts, wherein the simulation samples without flow artifacts are set as magnetic resonance weighted images or magnetic resonance parameter images according to requirements, if the images to be reconstructed are the magnetic resonance weighted images, the multi-scan magnetic resonance imaging signals of the virtual imaging object are subjected to rearrangement, splicing, zero padding, domain transformation and normalization processing to obtain the simulation samples without flow artifacts, and if the images to be reconstructed are the magnetic resonance parameter images, the virtual imaging object is used as the simulation samples without flow artifacts. The repeating processing unit is configured to repeatedly execute the virtual imaging object generation unit, the simulation signal generation unit, the simulation sample generation unit with flow artifacts and the simulation sample generation unit without flow artifacts until a set amount of training samples of the deep neural network are generated. The network training module is configured to train the deep neural network by using the training samples to obtain a trained deep neural network. The flow artifact correction module is configured to input the image of the actual imaging object with flow artifacts into the trained deep neural network to correct the flow artifacts and obtain an image without flow artifacts.

Images