A dual-gradient echo 3D-EPI quantitative susceptibility mapping and reconstruction method

Abstract

A dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method is disclosed, relating to magnetic resonance imaging. The method includes: designing a dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence; generating a virtual imaging object, which is generated using a map obtained from T1-weighted image segmentation; generating training samples for a deep neural network, including paired distorted simulated samples and distortion-free target samples; training the deep neural network using the training samples to obtain a trained deep neural network; acquiring data from the actual imaging object using the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence to obtain distorted image signals of the actual imaging object; and using the trained deep neural network to quantitatively reconstruct the magnetic susceptibility of the distorted image signals of the actual imaging object to obtain a distortion-free quantitative magnetic susceptibility image. This method enables rapid and efficient quantitative magnetic susceptibility reconstruction and can correct EPI image distortion.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging (MRI), and in particular to a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method. Background Technology

[0002] Quantitative susceptibility mapping (QSM) is an advanced technique used in magnetic resonance imaging (MRI) to measure the spatial distribution of magnetic susceptibility within an object. Magnetic susceptibility is a physical quantity that describes the change in magnetization intensity of a material under the influence of an external magnetic field. Human brain QSM has significant clinical implications and has been applied to cerebrovascular injuries such as microbleeds and many neurological diseases related to iron deposition, such as Alzheimer's disease, Parkinson's disease, and multiple sclerosis. Traditional QSM utilizes the phase of signals acquired via multi-echo 3D GRE sequences for calculation, involving multiple processing steps including phase unfolding, tissue volume extraction, background phase removal, and dipole inversion. Traditional acquisition methods are time-consuming, requiring 5–12 minutes, which patients may not tolerate. Furthermore, prolonged acquisition increases the possibility of involuntary movements, making the reconstruction results susceptible to motion artifacts. To reduce acquisition time and mitigate the effects of motion, 2D EPI sequences have also been used for QSM reconstruction. However, this requires a trade-off between SNR and resolution; higher layer resolution leads to lower SNR and increases acquisition time. Therefore, 3D EPI sequences have been used for acquisition to achieve efficient and high-quality reconstruction, but issues such as lower SNR and distortion still exist. Thus, a fast QSM reconstruction method that requires no additional image distortion correction is urgently needed. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for quantitative magnetic susceptibility imaging and reconstruction using dual-gradient echo 3D-EPI. This method and system can achieve rapid reconstruction of high-quality QSM images without distortion without additional distortion-free reference scanning or repeated scanning. It utilizes the image distortion in the opposite direction caused by bipolar phase encoding and does not require a multi-step reconstruction process. It employs a deep neural network to perform quantitative magnetic susceptibility reconstruction on dual-gradient echo 3D-EPI signals, thereby achieving distortion-free reconstruction of high-quality QSM images.

[0004] The present invention adopts the following technical solution:

[0005] On one hand, the present invention provides a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method, comprising:

[0006] S1: Design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence;

[0007] S2: Generate a virtual imaging object; the virtual imaging object is generated using the atlas obtained from T1 weighted image segmentation; specifically including:

[0008] S21: Obtain the T1-weighted magnetic resonance image from the public dataset, and use a segmentation tool to segment the T1-weighted image to obtain the spectrum;

[0009] S22: Based on the range of true values ​​of different magnetic resonance image modes, fill different values ​​into the different labels of the obtained map, and then perform reasonable deformation, blurring and other operations to obtain the maps of different magnetic resonance modes T1, T2, PD and χ modes, which serve as the T1, T2, PD and χ maps of virtual imaging objects;

[0010] S23: Obtain the local field from the χ-map of the virtual imaging object through forward evolution;

[0011] S24: Model the background inhomogeneity as the background inhomogeneity field caused by the external magnetic susceptibility source of VOI, and then superimpose it with the local field to obtain the total inhomogeneity field, which is used as the dB0 (main magnetic field inhomogeneity) map of the virtual imaging object.

[0012] S25: Combine T1, T2, PD, dB0 and χ to form a virtual imaging object;

[0013] S26: Repeat steps S21-S25 to generate a sufficient number of virtual imaging objects;

[0014] S3: Generate training samples for the deep neural network; the training samples include paired distorted simulation samples and undistorted target samples; the distorted simulation samples serve as input to the deep neural network training samples, and the undistorted target samples serve as labels for the deep neural network training samples; in the Bloch simulation, a pseudo-3D method is used for fast simulation, that is, the signal of each layer is obtained by simulating layer by layer in 2D, and then all layers are stitched together to obtain 3D samples, specifically including:

[0015] S31: Generate randomly distributed two-dimensional Gaussian noise to simulate the noise present in the actual sampling process of magnetic resonance images;

[0016] S32: In the Bloch simulation, the virtual imaging object is acquired using the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence in the presence of the two-dimensional Gaussian noise, resulting in a distorted simulation signal.

[0017] S33: After the layer-by-layer simulation is completed, the simulation signals are spliced ​​to obtain 3D data, and the phase data is taken to obtain the simulation sample with distortion.

[0018] S34: Use the χ-map of the virtual imaging object as the distortion-free target sample;

[0019] S35: Repeat the process from S31 to S34 until a set number of deep neural network training samples are generated;

[0020] S4: Use the training samples to train a deep neural network to obtain a trained deep neural network;

[0021] S5: The dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence is used to acquire data of the actual imaging object to obtain a distorted image of the actual imaging object; wherein the distortion direction of the first echo chain image is opposite to that of the second echo chain image.

[0022] S6: The trained deep neural network is used to reconstruct the distorted image of the actual imaging object using quantitative magnetic susceptibility to obtain a distortion-free quantitative magnetic susceptibility image.

[0023] Preferably, in step S1, the design of a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence specifically includes:

[0024] Phase coding design: Modify the phase coding of the dual-echo-chain dual-gradient echo 3D-EPI sequence so that the phase coding gradient directions of the odd and even echo chains are opposite, resulting in a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence.

[0025] The phase-encoded wraparound design adds a phase-encoded wraparound gradient before the phase encoding of the second echo chain, ensuring that the echo signals of the two echo chains of the dual-echo chain bipolar phase-encoded gradient echo 3D-EPI sequence fill the same position in their respective k-spaces.

[0026] Preferably, in step S23, the local field is obtained from the χ-map of the virtual imaging object through forward evolution, specifically including:

[0027] According to the following formula:

[0028]

[0029]

[0030] The local field phase is obtained through forward evolution. Where B0 is the principal magnetic field strength, FT is the Fourier transform operator, χ(k) and D(k) represent the magnetic susceptibility and magnetic dipole in k-space, respectively, and k is the k-space vector (k x ,k x ,k x );

[0031] Preferably, in step S24, the background non-uniform field caused by the external magnetic susceptibility source of VOI is modeled to obtain the VOI external magnetic field non-uniformity model, as follows:

[0032]

[0033] Among them, the background field inhomogeneity model It can be expressed as the sum of a three-dimensional polynomial function and a three-dimensional Gaussian function; Represents a three-dimensional polynomial function; x, y, and z represent coordinates in three-dimensional space; n x n y , and n z Let r represent the orders of x, y, and z, respectively; p N represents the coefficients of a randomly generated three-dimensional polynomial function; p G represents the highest order of a three-dimensional polynomial function; G represents a three-dimensional Gaussian function; N g The number of three-dimensional Gaussian functions is represented by r; m and r s Let represent the mean and standard deviation of a randomly generated three-dimensional Gaussian function, respectively.

[0034] Then, according to the following formula:

[0035]

[0036] Find the total non-uniform field dB0 (main magnetic field inhomogeneity) map as a virtual imaging object;

[0037] On the other hand, the present invention provides a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction system, comprising:

[0038] The data acquisition module is used to design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence and to acquire data from the actual imaging object using the dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence to obtain a distorted image of the actual imaging object; wherein, the distortion direction of the odd-numbered echo chain image is opposite to that of the even-numbered echo chain image; the dual-gradient echo 3D-EPI sequence is a magnetic resonance imaging sequence that uses radio frequency pulses to generate gradient echo signals and uses EPI technology for signal readout;

[0039] The virtual imaging object generation module is used to obtain a 3D human brain atlas from a public dataset, and then use a parameter map generator to obtain magnetic resonance parameter maps of each modality as virtual imaging objects. The parameter map generator realizes the process of filling different brain regions of the 3D human brain atlas with reasonable range values ​​and performing a certain degree of affine transformation and Gaussian sampling according to the prior distribution of the parameter maps of each modality to obtain the magnetic resonance parameter maps of each modality.

[0040] A rapid simulation module is used to generate training samples for a deep neural network. The training samples include paired distorted simulation samples and undistorted target samples. The distorted simulation samples serve as input to the deep neural network training samples, and the undistorted target samples serve as labels for the deep neural network training samples. The pseudo-3D Bloch simulator implements the superposition of 2D layer-by-layer Bloch simulation signals as 3D signal objects, specifically including:

[0041] A Gaussian noise generation unit is used to generate randomly distributed two-dimensional Gaussian noise to simulate the noise present in the actual sampling process of magnetic resonance images;

[0042] The Bloch simulation unit is used to acquire data from the virtual imaging object using the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence in the presence of the two-dimensional Gaussian noise during Bloch simulation, thereby obtaining a distorted simulation signal.

[0043] There is a distortion simulation sample generation unit, which is used to generate distorted simulation samples; after the layer-by-layer simulation is completed, the simulation signals are spliced ​​to obtain 3D data, and the phase data is taken to obtain distorted simulation samples.

[0044] A distortion-free target sample generation unit is used to generate distortion-free target samples; the image of the virtual imaging object is used as the distortion-free target sample.

[0045] The repetitive processing unit repeatedly executes the Gaussian noise generation unit, the Bloch simulation unit, the distorted simulation sample generation unit, and the undistorted target sample generation unit until a set amount of deep neural network training samples are generated.

[0046] The magnetic susceptibility reconstruction module is used to train a deep neural network using the training samples to obtain a trained deep neural network, and then use the trained deep neural network to reconstruct the distorted image of the actual imaging object to obtain a distortion-free quantitative magnetic susceptibility reconstruction result.

[0047] The beneficial effects of this invention are as follows:

[0048] (1) This invention designs a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence to obtain images with opposite distortion directions in a single scan.

[0049] (2) Existing image distortion correction methods cannot be used due to the existence of phase-encoded wraparound gradients; This invention is based on deep neural networks and uses complementary distortion information to correct image distortion.

[0050] (3) The present invention uses atlas filling as a virtual imaging object, which eliminates the need to collect a large amount of real MRI data as the initial data source, thus saving resources;

[0051] (4) The present invention does not require additional reference scans to provide distortion-free information, nor does it require repeated scans to obtain different distortion information, thus improving scanning efficiency;

[0052] (5) By using simulated samples to train a deep neural network, the present invention does not require the collection of a large amount of MRI data as a training set, thus saving resources; the trained deep neural network enables quantitative magnetic susceptibility reconstruction based on 3D-EPI, resulting in faster reconstruction speed and higher image quality. Attached Figure Description

[0053] Figure 1 This is a flowchart of the dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method according to an embodiment of the present invention;

[0054] Figure 2 This is a diagram of a dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence according to an embodiment of the present invention.

[0055] Figure 3 The image shows a quantitative magnetic susceptibility reconstruction image after distortion correction by a trained deep neural network on data acquired by a dual-echo chain bipolar phase-encoded gradient echo 3D-EPI sequence according to an embodiment of the present invention, and a comparison image of the quantitative magnetic susceptibility reconstruction results of data acquired by a traditional multi-echo 3D-GRE sequence using a traditional method (the image shows a certain axial plane, coronal plane and sagittal plane).

[0056] Figure 4 The image shown is a quantitative magnetic susceptibility reconstruction image after distortion correction by a trained deep neural network on the data acquired by the dual-echo-chain bipolar phase-encoded gradient echo 3D-EPI sequence in this embodiment of the invention, and a comparison with the reconstruction result of the single-echo-chain 3D-EPI sequence (the figure shows the axial plane).

[0057] Figure 5 This is a structural block diagram of the dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction system according to an embodiment of the present invention. Detailed Implementation

[0058] The present invention will be further described below through specific embodiments. It should be noted that the specific embodiments described herein are only for the convenience of illustrating and explaining the specific implementation of the present invention, and are not intended to limit the present invention.

[0059] To make the objectives and technical solutions of this invention clearer, the invention will be further described below with reference to the accompanying drawings and examples. It should be understood that the examples described herein are for illustrative purposes only and are not intended to limit the invention.

[0060] See Figure 1 This embodiment provides a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method, including:

[0061] S1: Design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence;

[0062] S2: Generate a virtual imaging object; the virtual imaging object is generated using the atlas obtained from T1 weighted image segmentation;

[0063] S3: Generate training samples for the deep neural network; the training samples include paired distorted simulation samples and undistorted target samples; the distorted simulation samples serve as input to the deep neural network training samples, and the undistorted target samples serve as labels for the deep neural network training samples; in the Bloch simulation, a pseudo-3D method is used for fast simulation, that is, the signal of each layer is obtained by 2D simulation layer by layer, and then all layers are stitched together to obtain 3D samples.

[0064] S4: Use the training samples to train a deep neural network to obtain a trained deep neural network;

[0065] S5: The dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence is used to acquire data of the actual imaging object to obtain a distorted image of the actual imaging object; wherein the distortion direction of the first echo chain image is opposite to that of the second echo chain image.

[0066] S6: The trained deep neural network is used to reconstruct the distorted image of the actual imaging object using quantitative magnetic susceptibility to obtain a distortion-free quantitative magnetic susceptibility image.

[0067] Specifically, in step S1, the design of the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence includes:

[0068] The phase coding design modifies the phase coding of the dual echo chain dual gradient echo 3D-EPI sequence so that the phase coding gradient directions of the odd and even echo chains are opposite, resulting in a dual echo chain bipolar phase-coded gradient echo 3D-EPI sequence. The purpose is to make the distortion directions of the first echo chain image and the second echo chain image opposite, so as to provide complementary information.

[0069] The phase-encoded back-wrap design incorporates a phase-encoded back-wrap gradient before the phase encoding of the second echo chain, ensuring that the echo signals of the two echo chains in the dual-echo-chain bipolar phase-encoded gradient echo 3D-EPI sequence fill the same position in their respective k-spaces. Specifically, the area of ​​the phase-encoded back-wrap gradient is the sum of the phase-encoded gradients, and its direction is opposite to the phase-encoded direction.

[0070] To reduce the accumulated error over multiple steps, a deep neural network is introduced for single-step magnetization reconstruction.

[0071] See Figure 2 In step S2, the virtual imaging object is generated using the map obtained from the T1 weighted image segmentation as follows.

[0072] S21: Obtain the T1-weighted magnetic resonance image from the public dataset, and use a segmentation tool to segment the T1-weighted image to obtain the spectrum;

[0073] S22: Based on the range of true values ​​of different magnetic resonance image modes, fill different values ​​into the different labels of the obtained map, and then perform reasonable deformation, blurring and other operations to obtain the maps of different magnetic resonance modes T1, T2, PD and χ modes, which serve as the T1, T2, PD and χ maps of virtual imaging objects;

[0074] S23: Obtain the local field from the χ-map of the virtual imaging object through forward evolution;

[0075] S24: Model the background inhomogeneity as the background inhomogeneity field caused by the external magnetic susceptibility source of VOI, and then superimpose it with the local field to obtain the total inhomogeneity field, which is used as the dB0 (main magnetic field inhomogeneity) map of the virtual imaging object.

[0076] S25: Combine T1, T2, PD, dB0 and χ to form a virtual imaging object;

[0077] S26: Repeat steps S21-S25 to generate a sufficient number of virtual imaging objects;

[0078] See Figure 3 In step S3, training samples for the deep neural network are generated. In the Bloch simulation, a pseudo-3D method is used for fast simulation, that is, the signal of each layer is obtained by simulating layer by layer in 2D, and then all layers are stitched together to obtain 3D samples, as follows:

[0079] S31: Generate randomly distributed two-dimensional Gaussian noise to simulate the noise present in the actual sampling process of magnetic resonance images;

[0080] S32: In the Bloch simulation, the virtual imaging object is acquired using the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence in the presence of the two-dimensional Gaussian noise, resulting in a distorted simulation signal.

[0081] S33: After the layer-by-layer simulation is completed, the simulation signals are spliced ​​to obtain 3D data, and the phase data is taken to obtain the simulation sample with distortion.

[0082] S34: Use the image of the virtual imaging object as a distortion-free target sample;

[0083] S35: Repeat steps S31 to S34 until a set amount of deep neural network training samples are generated.

[0084] The following detailed description of the process of dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method through a specific embodiment includes the following steps.

[0085] Step 1: Design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence. (Design reference for dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence) Figure 4 The process involves generating gradient echo signals using radio frequency pulses and gradients, and then reading them out using EPI technology. To achieve image distortion correction using only word scan data, bipolar phase coding and phase-coded gradient wraparound are introduced into the dual-gradient echo 3D-EPI sequence to obtain images with different distortion directions in the phase coding direction. Deep neural networks can utilize these characteristics to correct EPI images.

[0086] Step 2: Generate a virtual imaging object.

[0087] Step 3: Generate training samples for the deep neural network. To achieve quantitative magnetic susceptibility reconstruction using a dual-gradient echo 3D-EPI sequence, corresponding training samples need to be generated. The training samples for the deep neural network can be found in [reference needed]. Figure 5 . Figure 5 As an example of training samples required for the deep neural network in this invention, for the dual-echo chain bipolar phase-encoded gradient echo 3D-EPI sequence: by using this sequence to collect data on the virtual imaging object in the presence of the VOI external magnetic susceptibility source inhomogeneity model and Gaussian noise in Bloch simulation, a distorted simulation sample is obtained; the distorted simulation sample is used as the input of the deep neural network, and the χ map of the virtual imaging object is used as the label of the deep neural network.

[0088] Step 4: Train the deep neural network. Use the training samples generated in Step 3 to train the deep neural network.

[0089] Step 5: Use the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence designed in Step 1 to acquire data of the actual imaging object and obtain a distorted EPI image of the actual imaging object.

[0090] Step 6: Use the deep neural network trained in Step 3 to perform quantitative magnetization reconstruction on the distorted EPI image acquired in Step 5.

[0091] Images acquired using actual dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequences and quantitative magnetic susceptibility reconstruction results after distortion correction using the dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method of this embodiment can be referenced. Figure 4 .

[0092] In this embodiment, in step 1, the flip angle of the excitation pulse of the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence is 18°, TE1 = 9ms, TE2 = 26ms, and TR = 49ms.

[0093] In step 2, the publicly available dataset used is the IXI multimodal brain MRI dataset, which contains nearly 600 MR images of normal healthy subjects, acquired by 1.5T or 3T MRI scanners, and includes multiple modalities such as T1, T2, PD, MRA, and DWI. In this embodiment, 30 cases of 3D T1-weighted images acquired by the 3T MRI scanner are used.

[0094] In step 2, FreeSurfer v6 is used to segment the 3D T1-weighted image, dividing the human brain into 53 categories to obtain a segmented human brain atlas.

[0095] In step 2, the generation of virtual imaging objects is based on the affine transformation and deformation of the segmented human brain map. In this embodiment, 100 simulated human brain objects are created based on each segmented human brain map.

[0096] In step 2, the values ​​for filling different brain regions in different modalities of the simulated human brain object are based on the prior distribution of the parameter maps for each modality, so that the magnetic resonance parameter maps conform as closely as possible to the parameter distribution of the actual imaged object. A Gaussian mixture model is used to randomly sample the prior distributions of different brain regions to replace the category labels of the segmented human brain atlas, thereby increasing the diversity of the training data. The preferred parameter values ​​are as follows: PD values ​​are uniformly distributed between [0,1], T1 values ​​are uniformly distributed between [500ms, 2500ms], T2 values ​​are uniformly distributed between [20ms, 300ms], and χ values ​​are uniformly distributed between [-0.1ppm, 0.1ppm].

[0097] In step 2, the range of background field non-uniformity is [-200Hz, +200Hz].

[0098] In step 3, the range of Gaussian noise is [0dB, 60dB].

[0099] In step 4, the deep neural network uses 3D U-Net.

[0100] In step 4, the training samples need to be normalized. Since there are two echo chains, each training sample has two image data points. Therefore, the maximum value of the first echo chain image in all samples is selected, and the two echo chain image data points of all samples are divided by this maximum value to achieve normalization.

[0101] In step 4, for the dual-echo chain bipolar phase-encoded gradient echo 3D-EPI sequence, the input of the deep neural network is in the form of phase, that is, the input contains 2 images; the output is a distortion-free quantitative magnetic susceptibility image.

[0102] In step 4, mean squared error is used to constrain the training error of the network. The number of training samples is 216, the blocks are divided into 96×96×96, and the number of iterations is 3000.

[0103] In step 5, the imaging parameters of the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence are: resolution of 1×1×1 mm³, sampling matrix size of 224×224×160, and acceleration using GRAPPA technology with a speedup factor of 2. Data is acquired from the imaged object using the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence, and the acquired images are input into the deep neural network trained in step 4 to obtain a distortion-free image.

[0104] To evaluate a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method Figure 3 The figure shows the quantitative magnetic susceptibility reconstruction results of images acquired by a dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence using a deep neural network, as described in this invention. Simultaneously, this embodiment provides quantitative magnetic susceptibility reconstruction results obtained from images acquired by conventional multi-echo 3D-GRE sequences using traditional methods as a reference. The traditional methods used in this embodiment are the background field removal method V-SHARP and the dipole inversion method iLSQR. Specifically, a) is the quantitative magnetic susceptibility reconstruction result obtained from images acquired by multi-echo 3D-GRE sequences using traditional methods, and b) is the quantitative magnetic susceptibility reconstruction result obtained from data acquired by a dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence after processing by a trained deep neural network. The results show that this invention can achieve results similar to those of conventional multi-echo 3D-GRE quantitative magnetic susceptibility reconstruction methods. Furthermore, as indicated by the arrows in the figure, the quantitative magnetic susceptibility reconstruction results obtained by this invention can effectively reduce the generation of star artifacts compared to traditional methods. The present invention requires only about 27 seconds to collect data each time, which is much faster than the 5 to 10 minutes of collection time of traditional reconstruction methods.

[0105] Figure 4The quantitative magnetic susceptibility reconstruction results provided in this embodiment are obtained by processing data acquired using single-echo chain and dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequences through a trained deep neural network. In this case, a) represents the quantitative magnetic susceptibility reconstruction result obtained by processing data acquired using a single-echo chain gradient echo 3D-EPI sequence through a trained deep neural network, and b) represents the quantitative magnetic susceptibility reconstruction result obtained by processing data acquired using a dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence through a trained deep neural network. A comparison of a) and b) shows that the quantitative magnetic susceptibility results obtained using the single-echo chain gradient echo 3D-EPI sequence exhibit significant distortion in areas such as the top of the head and the optic nerve, affecting the reconstruction results. In contrast, this invention can effectively correct the distortion in the EPI image and clearly reconstruct image details.

[0106] See Figure 5 This embodiment also discloses a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction system, including:

[0107] The data acquisition module is used to design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence and to acquire data from the actual imaging object using the dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence to obtain a distorted image of the actual imaging object; wherein, the distortion direction of the odd-numbered echo chain image is opposite to that of the even-numbered echo chain image; the dual-gradient echo 3D-EPI sequence is a magnetic resonance imaging (MRI) sequence that uses radio frequency pulses to generate gradient echo signals and uses EPI technology for signal readout.

[0108] The virtual imaging object generation module is used to obtain a 3D human brain atlas from a public dataset, and then use a parameter map generator to obtain magnetic resonance parameter maps of each modality as virtual imaging objects. The parameter map generator realizes the process of filling different brain regions of the 3D human brain atlas with reasonable range values ​​and performing a certain degree of affine transformation and Gaussian sampling according to the prior distribution of the parameter maps of each modality to obtain the magnetic resonance parameter maps of each modality.

[0109] A rapid simulation module is used to generate training samples for a deep neural network. It simulates real imaging conditions by introducing Gaussian noise and uses the Bloch equation to simulate virtual imaging objects to obtain distorted simulation signals. The training samples include paired distorted simulation samples and undistorted target samples. The distorted simulation samples serve as input to the deep neural network training samples, and the undistorted target samples serve as labels for the deep neural network training samples. The pseudo-3D Bloch simulator implements 3D signal objects by superimposing signals from 2D layer-by-layer Bloch simulations.

[0110] Specifically, it includes:

[0111] A Gaussian noise generation unit is used to generate randomly distributed two-dimensional Gaussian noise to simulate the noise present in the actual sampling process of magnetic resonance images;

[0112] The Bloch simulation unit is used to acquire data from the virtual imaging object using the dual-echo chain bipolar phase-coded gradient echo 3D-EPI sequence in the presence of the two-dimensional Gaussian noise during Bloch simulation, thereby obtaining a distorted simulation signal.

[0113] There is a distortion simulation sample generation unit, which is used to generate distorted simulation samples; after the layer-by-layer simulation is completed, the simulation signals are spliced ​​to obtain 3D data, and the phase data is taken to obtain distorted simulation samples.

[0114] A distortion-free target sample generation unit is used to generate distortion-free target samples; the image of the virtual imaging object is used as the distortion-free target sample.

[0115] The repetitive processing unit repeatedly executes the Gaussian noise generation unit, the Bloch simulation unit, the distorted simulation sample generation unit, and the undistorted target sample generation unit until a set amount of deep neural network training samples are generated.

[0116] The magnetic susceptibility reconstruction module is used to train a deep neural network using the training samples to obtain a trained deep neural network, and then use the trained deep neural network to reconstruct the distorted image of the actual imaging object to obtain a distortion-free quantitative magnetic susceptibility reconstruction result.

[0117] A specific implementation of a dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction system is described in this embodiment, which is the same as the dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method, and will not be repeated here. This invention combines advanced imaging sequence design, simulation technology, and deep learning methods to achieve high-precision quantitative magnetic susceptibility imaging. This is of great significance for promoting clinical applications and basic research, especially in fields such as brain science research and pathological diagnosis.

[0118] It should be understood that those skilled in the art can make improvements and modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method, characterized in that, include: S1: Design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence; S2: Generate a virtual imaging object; The virtual imaging object is generated using a map obtained from T1-weighted image segmentation; specifically, it includes: S21: Obtain the T1-weighted magnetic resonance image from the public dataset, and use a segmentation tool to segment the T1-weighted image to obtain the spectrum; S22: Based on the range of true values ​​of different magnetic resonance image modes, fill different labels of the obtained map with different values, and then perform reasonable deformation and blurring to obtain maps of different magnetic resonance modes T1, T2, PD, and χ modes, which serve as T1, T2, PD, and χ maps of virtual imaging objects; S23: Obtain the local field from the χ-map of the virtual imaging object through forward evolution; S24: Model the background non-uniformity as the background non-uniform field caused by the external magnetic susceptibility source of VOI, and then superimpose it with the local field to obtain the total non-uniform field, which is used as the dB0 map of the virtual imaging object. S25: Combine T1, T2, PD, dB0 and χ to form a virtual imaging object; S26: Repeat steps S21 to S25 to generate a sufficient number of virtual imaging objects; S3: Generate training samples for the deep neural network; the training samples include paired distorted simulation samples and undistorted target samples; the distorted simulation samples serve as input to the deep neural network training samples, and the undistorted target samples serve as labels for the deep neural network training samples; in the Bloch simulation, a pseudo-3D method is used for fast simulation, that is, the signal of each layer is obtained by simulating layer by layer in 2D, and then all layers are stitched together to obtain 3D samples, specifically including: S31: Generate randomly distributed two-dimensional Gaussian noise to simulate the noise present in the actual sampling process of magnetic resonance images; S32: In Bloch simulation, a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence is used to acquire data from the virtual imaging object in the presence of two-dimensional Gaussian noise, resulting in a distorted simulation signal. S33: After the layer-by-layer simulation is completed, the simulation signals are spliced ​​to obtain 3D data, and the phase data is taken to obtain the simulation sample with distortion. S34: Use the χ-map of the virtual imaging object as the distortion-free target sample; S35: Repeat the process from S31 to S34 until a set number of deep neural network training samples are generated; S4: Use the training samples to train a deep neural network to obtain a trained deep neural network; S5: A dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence is used to acquire data from the actual imaging object, resulting in a distorted image of the actual imaging object; wherein, the distortion direction of the first echo-chain image is opposite to that of the second echo-chain image. S6: A trained deep neural network is used to reconstruct the quantitative magnetic susceptibility of the distorted image of the actual imaging object to obtain a distortion-free quantitative magnetic susceptibility image.

2. The dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method according to claim 1, characterized in that, In step S1, a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence is designed, specifically including: Phase coding design: Modify the phase coding of the dual-echo-chain dual-gradient echo 3D-EPI sequence so that the phase coding gradient directions of the odd and even echo chains are opposite, resulting in a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence. The phase-encoded wraparound design adds a phase-encoded wraparound gradient before the phase encoding of the second echo chain, ensuring that the echo signals of the two echo chains of the dual-echo chain bipolar phase-encoded gradient echo 3D-EPI sequence fill the same position in their respective k-spaces.

3. The dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method according to claim 1, characterized in that, In step S23, the local field is obtained from the χ-map of the virtual imaging object through forward evolution, specifically including: According to the following formula: The local field phase is obtained through forward evolution. Where B0 is the principal magnetic field strength, FT is the Fourier transform operator, χ(k) and D(k) represent the magnetic susceptibility and magnetic dipole in k-space, respectively, and k is the k-space vector (k x ,k y ,k z ).

4. The dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction method according to claim 1, characterized in that, In step S24, the background inhomogeneity field caused by the external magnetic susceptibility source of VOI is modeled to obtain the inhomogeneity model of the external magnetic field of VOI, as follows: Among them, the background field inhomogeneity model It can be expressed as the sum of a three-dimensional polynomial function and a three-dimensional Gaussian function; Represents a three-dimensional polynomial function; x, y, and z represent coordinates in three-dimensional space; n x n y , and n z Let r represent the orders of x, y, and z, respectively; p N represents the coefficients of a randomly generated three-dimensional polynomial function; p G represents the highest order of a three-dimensional polynomial function; G represents a three-dimensional Gaussian function; N g The number of three-dimensional Gaussian functions is represented by r; m and r s Let represent the mean and standard deviation of a randomly generated three-dimensional Gaussian function, respectively. Then, according to the following formula: Find the total non-uniform field dB0 map as a virtual imaging object.

5. A dual-gradient echo 3D-EPI quantitative magnetic susceptibility imaging and reconstruction system, characterized in that, include: The data acquisition module is used to design a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence and to acquire data from the actual imaging object using the dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence to obtain a distorted image of the actual imaging object; wherein, the distortion direction of the odd-numbered echo chain image is opposite to that of the even-numbered echo chain image; the dual-gradient echo 3D-EPI sequence is a magnetic resonance imaging sequence that uses radio frequency pulses to generate gradient echo signals and uses EPI technology for signal readout; The virtual imaging object generation module is used to obtain a 3D human brain atlas from a public dataset, and then use a parameter map generator to obtain magnetic resonance parameter maps of each modality as virtual imaging objects. The parameter map generator realizes the process of filling different brain regions of the 3D human brain atlas with reasonable range values ​​and performing a certain degree of affine transformation and Gaussian sampling according to the prior distribution of the parameter maps of each modality to obtain the magnetic resonance parameter maps of each modality. Virtual imaging objects are generated using atlases obtained from T1-weighted image segmentation, specifically including: 1) Obtain T1-weighted magnetic resonance images from public datasets, and use segmentation tools to segment the T1-weighted images to obtain the spectra; 2) Based on the range of true values ​​of different magnetic resonance image modes, fill different values ​​into the different labels of the obtained map, and then perform reasonable deformation and blurring to obtain the maps of different magnetic resonance modes T1, T2, PD, and χ modes, which serve as the T1, T2, PD, and χ maps of the virtual imaging object; 3) Obtain the local field from the χ-map of the virtual imaging object through forward evolution; 4) The background non-uniformity is modeled as the background non-uniform field caused by the external magnetic susceptibility source of VOI, and then superimposed with the local field to obtain the total non-uniform field, which is used as the dB0 map of the virtual imaging object. 5) Combine T1, T2, PD, dB0, and χ to form a virtual imaging object; 6) Repeat steps 1) to 5) to generate a sufficient number of virtual imaging objects; A rapid simulation module is used to generate training samples for a deep neural network; the training samples include paired distorted simulation samples and undistorted target samples; the distorted simulation samples serve as input to the deep neural network training samples, and the undistorted target samples serve as labels for the deep neural network training samples; a pseudo-3D Bloch simulator uses superimposed 2D layer-by-layer Bloch simulation signals as 3D signal objects, specifically including: A Gaussian noise generation unit is used to generate randomly distributed two-dimensional Gaussian noise to simulate the noise present in the actual sampling process of magnetic resonance images; The Bloch simulation unit is used to acquire data from the virtual imaging object in Bloch simulation using a dual-echo-chain bipolar phase-coded gradient echo 3D-EPI sequence in the presence of two-dimensional Gaussian noise, thereby obtaining a distorted simulation signal. There is a distortion simulation sample generation unit, which is used to generate distorted simulation samples; after the layer-by-layer simulation is completed, the simulation signals are spliced ​​to obtain 3D data, and the phase data is taken to obtain distorted simulation samples. The distortion-free target sample generation unit is used to generate distortion-free target samples; the image of the virtual imaging object is used as the distortion-free target sample. The repetitive processing unit repeatedly executes the Gaussian noise generation unit, the Bloch simulation unit, the distorted simulation sample generation unit, and the undistorted target sample generation unit until a set amount of deep neural network training samples are generated. The magnetic susceptibility reconstruction module is used to train a deep neural network using the training samples to obtain a trained deep neural network, and then use the trained deep neural network to reconstruct the distorted image of the actual imaging object to obtain a distortion-free quantitative magnetic susceptibility reconstruction result.

Images