An EPI phase correction method based on deep learning
Through the EPI phase correction method based on deep learning, positive and negative echoes are separated from K space and data recovery is used by neural networks, which solves the artifact problem in EPI images, and achieves efficient image recovery and acceleration, avoiding the traditional methods. The problem of artifact residue.
Patent Information
- Application Number
- CN202111262206.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-28
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-10-28
AI Technical Summary
In the prior art, in dealing with the Nyquist artifact problem due to positive and negative echo phase difference in plane echo imaging (EPI), especially after the acceleration multiple increases during high parallel imaging, traditional methods cannot effectively eliminate image artifacts. , especially for 3 times or 4 times under-recovered data, a large number of artifacts remain in the reconstruction image.
Using the EPI phase correction method based on deep learning, the positive and negative echoes are separated from the K space, and the data is recovered using neural networks such as convolutional neural networks (CNNs). It is iteratively optimized through training and testing data until the difference is minimal and the output is no The image of artifacts realizes the merge and recovery of images.
It can effectively recover low- and high-power (3 or 4 times) under-acquisition data, eliminate image artifacts, and does not require additional reference data acquisition, save scanning time, and significantly improve the quality of EPI images.
Smart Images

Figure CN114041776B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of medical imaging, and in particular relates to a magnetic resonance imaging reconstruction method. Background Art
[0002] Echo-planar imaging (EPI) is an important magnetic resonance imaging technique that is widely used in brain function, neuroscience, and tumor imaging. Compared with other magnetic resonance imaging techniques, its advantages are fast acquisition speed, insensitivity to motion, and the ability to capture magnetic resonance signals with high temporal resolution. However, since EPI uses an acquisition method with alternating positive and negative readout gradient polarity, coupled with hardware eddy currents, magnetic field inhomogeneity and other factors, Nyquist artifacts caused by phase mismatch between positive and negative echoes will inevitably appear on the image. These artifacts will be superimposed on the image area, making the image unusable for diagnosis and post-processing. This is also the most common artifact in applications based on EPI imaging technology such as brain function imaging, diffusion-weighted imaging, diffusion tensor imaging, and traditional perfusion imaging. In the absence of accelerated acquisition, this artifact manifests as ghost images that move half of the FOV up and down, respectively, also known as N / 2-Ghost artifacts. As the acceleration factor increases during parallel imaging, this artifact will also increase. When the acceleration factor is 2, the number of ghost images also increases to twice the original.
[0003] At present, there are many technologies for EPI phase correction. In the early days, a linear correction method based on reference line scanning was proposed. 1 This method assumes that the phase difference between the positive and negative echoes is linear. In order to obtain the phase difference between the positive and negative echoes and compensate for it, several echoes without phase encoding gradients are pre-scanned before image acquisition. The zero-order and first-order terms in the phase difference between the positive and negative echoes are calculated by linear fitting, and the calculated results are used to correct the phase differences of other positive and negative echoes. However, if the system field strength fluctuates slightly during the scanning process, or the subject moves slightly, the phase difference of the pre-scan echo will be unequal to the phase difference of the imaging part echo, and the imaging artifacts cannot be completely eliminated. In order not to rely on the pre-scan reference line, an entropy minimization method has been proposed. 2 This method is similar to the linear correction method, the difference is that this method does not require a reference line to calculate the phase difference, but it iteratively calculates the entropy of the image and searches for the corresponding phase difference when the image entropy is the smallest. For fully sampled images, this method can eliminate most artifacts without a reference line, but when the image is undersampled, the accurate phase difference parameters cannot be found due to the curling artifacts, so it is not effective for conventional acceleration or scanning with increased acceleration factors. In recent years, some methods based on parallel reconstruction have been proposed. This method cleverly converts the phase correction problem of positive and negative echoes into the problem of undersampled recovery. 3,4. Since artifacts come from the phase difference between positive and negative echoes, this type of method separates positive and negative echoes, that is, a K space containing alternating positive and negative echoes is converted into two K spaces with the same polarity of echoes, which is equivalent to undersampling the original K space by 2 times. After separation, there is no phase difference between echoes in the K space. Then, a parallel reconstruction algorithm such as SENSE or GRAPPA is used to restore data of these two K spaces to obtain an image without artifacts, and finally the amplitudes of these two images are merged. Although this can circumvent the problem of phase difference, separating positive and negative echoes is equivalent to increasing the undersampling multiple of the K space to twice the original, while traditional parallel reconstruction methods such as SENSE or GRAPPA can only effectively restore 2-fold or 3-fold undersampling data, and 4-fold undersampling is already the limit. For EPI data with 3-fold or 4-fold undersampling, each K space is equivalent to 6-fold or 8-fold undersampling after separation of positive and negative echoes. Traditional parallel reconstruction methods cannot completely restore these data, and the reconstructed image has a large number of residual Nyquist artifacts. In addition, both SENSE and GRAPPA require additional reference data.
[0004] References
[0005] 1. A. Jesmanowicz, EC Wong, JS Hyde, Phase correction for EPI using internal reference lines., in “Proc., SMRM, 12th Annual Meeting, NewYork, 1993,” p. 1239.
[0006] 2. Bammer R, Keeling SL, Augustin M, Pruessmann KP, Wolf R, Stollberger R, Hartung HP, Fazekas F. Improved diffusion-weighted single-shotecho-planar imaging (EPI) in stroke using sensitivity encoding (SENSE). MagnReson Med 2001; 46: 548– 554.
[0007] 3. Xie VB, Lyu M, Liu Y, Feng Y, Wu EX. Robust EPI Nyquist ghostremoval by incorporating phase error correction with sensitivity encoding(PEC‐SENSE). Magnetic Resonance in Medicine. 2018;79:943-951.
[0008] 4. Patent number: CN 106574954 A, 2017.04.19. Summary of the invention
[0009] The purpose of the present invention is to propose a new EPI phase correction method based on deep learning to overcome the troubles caused by updating the display menu in the above-mentioned prior art, restore the under-sampled K space, and thus eliminate EPI image artifacts.
[0010] The EPI phase correction method based on deep learning proposed in the present invention has the following specific steps:
[0011] 1.1.1.1 Collection and separation of EPI data
[0012] EPI uses alternating positive and negative readout gradient polarities for data acquisition, such as Figure 1 , the trajectory after the collected echoes are filled into the K space in sequence, such as Figure 2 . Due to the phase difference between adjacent echoes, the reconstructed image has serious artifacts. The present invention first separates the positive and negative echoes from the K space and converts them into two K spaces with the same polarity. The separated positive and negative K space signals can be expressed by the equation:
[0013] , (1)
[0014] , (2)
[0015] in, represents the artifact-free image intensity, x, y represent the K-space position along the readout and phase encoding directions in K-space, respectively, and m, n represent the corresponding image domain positions; Represent the acquisition matrices in the readout direction and phase encoding direction, Indicates the phase difference between positive and negative echoes; and For under-sampled K-space signals, directly performing Fourier transform on the two K-spaces to obtain the reconstructed image will show curling artifacts due to under-sampling.
[0016] (ii) Eliminate the curling artifacts in the above images;
[0017] In order to eliminate the curling artifacts in the above images, the present invention uses deep learning technology based on neural networks to restore these artifact-bearing images respectively by learning the artifact-free images reconstructed when full sampling is performed, and finally the restored images are merged into one image. The detailed steps are described as follows:
[0018] (1) Selecting a neural network. In deep learning, there are many types of neural networks that can be used for data recovery, including convolutional neural networks (CNN), U-net convolutional neural networks, and generative adversarial networks (GAN). Many of these networks can effectively recover high-magnification (10-20 times) undersampled data. The main function of a good network can be expressed as ;
[0019] (2) Collect data. Collect a set (several or dozens of cases) of EPI raw data for network training. These data can be fully sampled or undersampled at low magnification. In order to improve the generalization of the network and be compatible with test data of different contrasts and geometric deformations, different TE values can be set when collecting training data to obtain training sets of different contrasts and geometric deformations.
[0020] (3) Calculate the gold standard image. In order to obtain a gold standard image without artifacts, the data collected in step 2 is phase corrected using traditional phase correction methods, such as linear phase correction, entropy minimization, etc., to eliminate artifacts in the image and save it as the gold standard.
[0021] (4) Train the network, such as Figure 3 First, the artifact-free gold standard image is converted from the image domain to the K space. The converted K space is fully sampled. The fully sampled K space is separated according to the positive and negative echoes according to the acceleration multiple requirements to generate two single-polarity K spaces. The missing echo positions are filled with 0 to keep the size of the K space unchanged. Then, the separated K space is converted to the image domain. Finally, the gold standard image and the positive echo (or negative echo) image after positive and negative separation are fed into the network at the same time. , the network is trained, and the positive and negative echo images are trained twice. The training process is to compare the difference between the training results and the gold standard through continuous iteration until the difference is minimized and the training is terminated to output the trained network parameters The number of images required for training can be determined according to the requirements of the network;
[0022] (5) Test network, such as Figure 3As shown in the dashed box, the real collected data is used as the test data, which includes full-collection data, 2x accelerated collection data, and 4x accelerated collection data. These data are tested separately. First, these data are separated into positive and negative echoes, and the positive echo K-space Fourier transform is used to obtain the image After Fourier transformation of the negative echo, we get , feed these images into the trained network model respectively, and the undersampling method of the separated images is required to be consistent with the network model. The image can be obtained after network recovery ,image The image can be obtained after network recovery .
[0023] The process can be expressed as:
[0024] , (3)
[0025] , (4)
[0026] Finally, the restored image and By combining the squares and square roots, we get an image, which is the final test result.
[0027] The positive and progressive effects of the present invention are: not only can it well restore low-magnification undersampled data, which can replace the traditional method, but it can also achieve good results for high-magnification (3 or 4 times) undersampled data, thereby realizing EPI acquisition acceleration and image artifact elimination. In addition, there is no need to collect additional reference lines, which can save scanning time.
[0028] The method of eliminating image artifacts generated in the single-shot echo planar imaging (EPI) process of the present invention can be used in magnetic resonance technologies such as brain functional imaging, diffusion weighted imaging, diffusion tensor imaging, and traditional perfusion imaging based on the EPI acquisition mode. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 Schematic diagram of the EPI positive and negative gradient alternating acquisition method used in the present invention.
[0030] Figure 2 It is the sampling trajectory of alternating positive and negative echoes in the K-space of EPI used in the present invention.
[0031] Figure 3 This is a flow chart of network training and network testing used in the present invention.
[0032] Figure 4 This is a diagram of the network architecture used in the embodiments of the present invention.
[0033] Figure 5 This is a specific example of the results achieved using the method of the present invention in Implementation 1.
[0034] Figure 6 This is a specific example of the results achieved using the method of the present invention in Implementation 2. DETAILED DESCRIPTION
[0035] The present invention will be further described below in conjunction with the embodiments and drawings.
[0036] Embodiment 1,
[0037] refer to Figure 1 and Figure 2 , showing the EPI acquisition method and K-space filling trajectory used in the present invention. In this embodiment, DWI images based on the EPI sequence are used. A total of 52 data were collected for network training, including 30 healthy subjects and 22 subjects with brain diseases. The system used is Siemens 7T Terra, 32-channel head coil, acquisition parameters: TR=11700ms, FOV=220*220 mm², matrix size = 158*158, b=0 / 1000 s / mm², number of layers = 80 layers, layer thickness = 1.4mm, including three TE values, TE=84, 61 and 50ms. The gold standard image was reconstructed using the traditional parallel reconstruction method.
[0038] The neural network used is a convolutional neural network, and the network architecture is referenced Figure 4 , the network consists of ten modules connected in series, each module contains a convolutional neural network layer and a data fidelity layer, each convolutional neural network layer contains 5 sublayers, and the first four sublayers have rectified linear units. The kernel size of each layer is 3 3.
[0039] In this embodiment, the test data used DWI data from healthy subjects. In order to verify the performance of this method, the acceleration factor was set to 4, which exceeded the acceleration multiple that the traditional method could handle. TE = 50ms, and other parameters were consistent with the training data. The results were compared with other traditional methods including linear correction method, entropy minimization method, and PEC-SENSE correction method based on parallel reconstruction. Figure 5 ,From the figure, we can see that for the 4-fold accelerated data, the ,traditional method cannot completely eliminate the image artifacts, while the ,presentation achieves a good effect.
[0040] Embodiment 2,
[0041] The network model used in this embodiment is the same as that in embodiment 1, except that the test data used is from a patient with moyamoya disease. DWI data with 1x, 2x, and 4x acceleration were collected respectively. Figure 6 As can be seen from the figure, for an image with lesions, no matter it is 1x, 2x or 4x, the lesions can be well displayed, and there is no obvious residual artifact in the graphics. The present invention still has a good effect on graphics with lesions.
Claims
1. A deep learning-based EPI phase correction method, characterized in that: The specific steps are as follows: 1.1.1.1 Collection and separation of EPI data EPI uses alternating positive and negative readout gradient polarities for data acquisition. The acquired echoes are sequentially filled into the trajectory of the K space. The positive and negative echoes are separated from the K space to form two K spaces with the same polarity. The separated positive and negative K space signals are expressed by the equation: , (1) , (2) in, represents the artifact-free image intensity, x, y represent the K-space position along the readout and phase encoding directions in K-space, respectively, and m, n represent the corresponding image domain positions; Represent the acquisition matrices in the readout direction and phase encoding direction, Indicates the phase difference between positive and negative echoes; and For under-sampled K-space signals, the reconstructed image obtained by directly performing Fourier transform on the two K-spaces will have curling artifacts caused by under-sampled signals. (ii) Eliminate the curling artifacts in the above images; Using deep learning technology based on neural networks, by learning the artifact-free images reconstructed at full sampling, these artifact-bearing images are restored respectively, and finally the restored images are merged into one image; the specific process is as follows: (1) Select a neural network and express the selected network main function as ; (2) Collecting data: Collect a set of EPI raw data for network training. These data are fully sampled or undersampled at low magnification. In order to improve the generalization of the network and be compatible with test data of different contrasts and geometric deformations, different TE values are set when collecting training data to obtain training sets of different contrasts and geometric deformations. (3) Calculate the gold standard image, use the traditional phase correction method on the data collected in step 2 to eliminate artifacts in the image, and save it as the gold standard; (4) Training the network: First, the artifact-free gold standard image is converted from the image domain to the K space. The converted K space is fully sampled. The fully sampled K space is separated according to the positive and negative echoes according to the acceleration multiple requirements to generate two single-polarity K spaces. The missing echo positions are filled with 0 to keep the size of the K space unchanged. Then, the separated K space is converted to the image domain. Finally, the gold standard image and the positive or negative echo image after positive and negative separation are fed into the network at the same time. , the network is trained, and the positive and negative echo images are trained twice; the training process is repeated by comparing the difference between the training results and the gold standard until the difference is minimized, and the training is terminated and the trained network parameters are output ; The number of images required for training depends on the requirements of the network; (5) Test network: The real collected data are used as test data, including full-collection data, 2x accelerated collection data, and 4x accelerated collection data. These data are tested separately. First, these data are separated into positive and negative echoes, and the positive echo K-space Fourier transform is used to obtain the image. After Fourier transformation of the negative echo, we get , feed these images into the trained network model respectively, and the undersampling method of the separated images is required to be consistent with the network model; image The image is recovered through the network ,image The image is recovered through the network ; The process is expressed in the formula: , (3) , (4) Finally, the restored image and By combining the squares and square roots, we get an image, which is the final test result.
2. The EPI phase correction method based on deep learning according to claim 1, characterized in that: The neural network is selected from a convolutional neural network, a U-net convolutional neural network, and an adversarial neural network, and is used to recover data that is 10-20 times undersampled.
3. The EPI phase correction method based on deep learning according to claim 1, characterized in that: The neural network adopts a convolutional neural network, which is composed of ten modules connected in series. Each module includes a convolutional neural network layer and a data fidelity layer. Each convolutional neural network layer contains 5 sublayers, and the first four sublayers have rectified linear units; the kernel size of each layer is 3 3.
Citation Information
Patent Citations
Inside-out echo-planar imaging method for shortening echo time
CN102890255A
Navigation magnetic resonance diffusion imaging method and device based on multiple excitations
CN106443533A
Deformation-free single-excitation plane echo imaging method and device based on deep learning
CN110895320A
Multiple-shot planar echo magnetic resonance imaging method based on neural network
CN112763958A
Navigator-less segmented diffusion weighted imaging enabled by multiplexed sensitivity-encoded imaging with inherent phase correction
US20140002078A1
