Rapid acquisition of cerebral vascular images: scanning and reconstruction methods and magnetic resonance imaging systems

Abstract

A scanning and reconstruction method and magnetic resonance imaging system for rapidly acquiring cerebral vascular images are proposed. This method utilizes Poisson undersampling K-space trajectories to significantly improve the scanning speed of TOF sequences. At the same time, it employs a deep learning network model to better and faster recover Poisson undersampling data images. Ultimately, it obtains human cerebral vascular images in a shorter time that are comparable to those of fully sampled images, thus overcoming the shortcomings of traditional TOF sequences, such as slow scanning speed and susceptibility to motion-induced image interference.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance imaging, specifically relating to a scanning and reconstruction method and a magnetic resonance imaging system for rapidly acquiring images of cerebral blood vessels. Background Technology

[0002] Common angiography techniques in magnetic resonance angiography (MRA) include TOF (Time of Flight), CEMRA (Contrast Enhanced MRA), and PC (Phase-Contrast).

[0003] CEMRA requires the injection of contrast agents and has precise requirements for scanning time. Time-of-flight (TOF) is the most common cerebrovascular imaging technique in clinical practice.

[0004] Time-of-flight (TOF) technology is based on the inflow enhancement effect of blood flow and typically uses short transverse velocity (TR) scanning sequences to acquire cerebral vascular data. Traditional TOF technology acquires images using Cartesian sampling, which has drawbacks such as long scan times and susceptibility to motion-induced image distortion.

[0005] Therefore, it is crucial to design a brain vascular image scanning and reconstruction method that offers fast scanning speed and high image quality. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a scanning and reconstruction method for rapidly acquiring images of cerebral blood vessels.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a scanning and reconstruction method for rapidly acquiring cerebral vascular images, specifically including the following steps: Step S1, in the Ky-Kz plane of the 3D TOF sequence, variable density K-space trajectory data is acquired using Poisson variable density K-space, and undersampling acceleration coefficients in the Y and Z directions are obtained respectively. Based on the undersampling acceleration coefficients, the Poisson undersampling trajectory distribution of K-space variable density is calculated. At the same time, the coding gradient values ​​of the phase direction and layer direction are determined according to the coordinate position of the Poisson point in the plane of Ky-Kz. By applying the corresponding gradient values ​​on the physical axes of the Y and Z axes, the effect of Poisson variable density acquisition is achieved; Step S2, the Poisson undersampling data acquired by Poisson variable density is recovered using a pre-trained deep learning network to obtain multiple segments of cerebral vascular images; Step S3, the multiple uniform cerebral vascular images are stitched together to obtain a three-dimensional cerebral vascular image; Step S4, the three-dimensional cerebral vascular image is subjected to maximum density projection (MIP) in a magnetic resonance imaging system, and background tissue is cropped and removed. The final 3D cerebral vascular image of the head is obtained by 360° projection in three directions.

[0008] In some implementations, in step S2, the deep learning network is trained by: collecting a large amount of fully sampled cerebral vascular data in the early stage; extracting fully sampled cerebral vascular data according to the Poisson undersampling spectrum when the actual Poisson undersampling is performed; forming multiple sets of fully sampled images and images extracted from the Poisson sampling spectrum; and using a large number of fully sampled and undersampled image sets to train the model-based deep learning model to obtain a trained deep learning model.

[0009] In some implementations, to give the trained neural network more coil uniformity information, coil uniformity correction is performed on the fully sampled cerebral vascular images before training.

[0010] In some implementations, uniformity correction requires coil sensitivity information, necessitating a calibration pre-scan using both a volume coil and an imaging coil to acquire the coil sensitivity spectrum for each channel at low resolution. , ,in, For the pre-scanning imaging coil image, For the pre-scanned volume coil image, For the coil channel unit, trilinear interpolation is then used to calculate the sensitive spectrum information of the corresponding layer in TOF3D. Finally, the RAIN correction algorithm is used to obtain a uniform original image. The formula for coil uniformity correction of all acquired cerebral vascular images before training is as follows: ,in, This indicates an image without uniformity correction, and CH represents the number of channels in the receiving coil.

[0011] In some implementations, the model-based convolutional neural network formula is as follows: Where A = SF; F is the 3D discrete Fourier transform, S is the sampling spectrum formed by the Poisson sampling trajectory, b is the actual undersampled K-space data, Nw(x) is a regularization term, and is a noise and artifact estimator trained by a neural network, expressed by the parameter w learned by the neural network.

[0012] In some implementations, a neural network is used to train the image Dw to remove noise artifacts, as shown in the following formula: Dw is the denoised and artifact-free image of image x after being processed by a pre-trained neural network. When the sampled image x contains a lot of noise and artifacts, the value of the regularization term Nw is large, and vice versa. Where Jn is the Jacobian matrix, Xn + ΔX = X, a denoising and artifact removal neural network is pre-trained, and the denoising and artifact removal image Dw is obtained after N iterations. Substituting this into the model-based convolutional neural network formula, the following formula can be obtained: Further transformed The denoised image is obtained by using a pre-trained denoising model, and the consistency of the data is trained by using the conjugate gradient iteration method to form a data consistency constraint term, thereby accelerating the speed of iteration convergence.

[0013] In some implementations, in step S1, the Kx direction in the 3DTOF sequence is acquired using a Cartesian acquisition method.

[0014] In some implementations, in step S3, multiple uniform cerebral blood vessel images are stitched together, and the connection points of the multiple cerebral blood vessels are processed using venetian blind filtering.

[0015] Another technical problem to be solved by the present invention is to provide a magnetic resonance imaging system.

[0016] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a magnetic resonance imaging system, wherein the magnetic resonance imaging system adopts a scanning and reconstruction method for rapidly acquiring cerebral vascular images as described in any of the above embodiments.

[0017] The scope of this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this application.

[0018] Due to the application of the above technical solutions, the present invention has the following advantages compared with the prior art: The present invention provides a scanning and reconstruction method and magnetic resonance imaging system for rapidly acquiring cerebral vascular images. It utilizes Poisson undersampling K-space trajectory to greatly improve the scanning speed of TOF sequence. At the same time, it uses deep learning network model to better and faster recover Poisson undersampling data images. Finally, it obtains human cerebral vascular images in a shorter time that are no less than those of fully sampled images, solving the shortcomings of traditional TOF sequence scanning speed and image susceptibility to motion. Attached Figure Description

[0019] Appendix Figure 1 This is a schematic diagram of the method flow of the present invention;

[0020] Appendix Figure 2 This is a schematic diagram of the Poisson variable density acquisition trajectory of the present invention;

[0021] Appendix Figure 3 This is a schematic diagram of the Poisson variable density sampling sequence of the present invention;

[0022] Appendix Figure 4 This is a schematic diagram of the TOF cerebral vascular neural network training process of the present invention;

[0023] Appendix Figure 5This is a comparison image of conventional 3DTOF head blood vessel MIP images and 3DTOF blood vessel recovery by Poisson variable density trajectory neural network according to the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings.

[0025] As attached Figure 1 As shown, this invention provides a method for rapidly acquiring and reconstructing cerebral vascular images, with the following specific steps:

[0026] Step S1: In this embodiment, the X, Y, and Z directions of the K-space correspond to the readout direction, phase direction, and layer direction, respectively. In the Ky-Kz plane of the 3DTOF sequence, the variable density K-space trajectory data is collected using the Poisson variable density K-space, and the undersampling acceleration coefficients in the Y and Z directions are obtained respectively. The Cartesian method is still used in the readout direction.

[0027] Based on the undersampling acceleration coefficient, the Poisson undersampling trajectory distribution with K-space variable density is calculated. Simultaneously, the coding gradient values ​​in the phase and layer directions are determined based on the coordinates of the Ky-Kz points in the plane. By applying the corresponding gradient values ​​on the Y and Z physical axes (i.e., on the phase and layer directions), the effect of Poisson variable density acquisition is achieved. Figure 2 A schematic diagram of the Poisson variable density sampling trajectory in the Ky-Kz plane is given.

[0028] Step S2: Use a pre-trained deep learning network to recover the Poisson undersampled data acquired by Poisson variable density acquisition, and obtain multi-segment cerebral blood vessel images.

[0029] A large amount of fully sampled cerebral vascular data was collected in the initial stage. Based on the Poisson undersampling spectrum during actual Poisson undersampling, fully sampled cerebral vascular data was extracted to form multiple sets of fully sampled images and images extracted from the Poisson sampling spectrum. Based on the obtained sets of fully sampled and undersampled images, a model-based deep learning network was trained to obtain a TOF3D cerebral vascular neural network model, i.e., a pre-trained deep learning network. This network requires less training data and training time compared to traditional networks. Moreover, this model can achieve good data recovery results with a smaller network.

[0030] The formula for a model-based convolutional neural network is shown below: In this equation, A = SF; F is the 3D Discrete Fourier Transform; S is the sampling spectrum formed by the Poisson sampling trajectory; Nw(x) is a regularization term; b is the actual undersampled K-space data collected; and is a noise and artifact estimator trained by a neural network, represented by the parameters w learned by the neural network. According to the formula, this deep learning network has two parts. The first term is the traditional neural network term for noise reduction and artifact removal. The second term is the training term to ensure data consistency.

[0031] To incorporate more coil uniformity information into the trained neural network, coil uniformity correction is performed on the fully sampled cerebral vascular images before training. By introducing channel-sensitive spectrum information, the trained network can recover more uniform cerebral vascular images.

[0032] Uniformity correction requires coil sensitivity information, necessitating a pre-scan using both a volume coil and an imaging coil to acquire the coil sensitivity spectra for each channel at low resolution. , ,in, For the pre-scanning imaging coil image, The image is a pre-scanned volume coil image, where ic represents the coil channel unit. Then, trilinear interpolation is used to calculate the sensitivity spectrum information of the corresponding layer in the TOF3D algorithm. Finally, the RAIN correction algorithm is used to obtain a uniform original image. The formula for coil uniformity correction of all acquired cerebral vascular images before training is as follows: ,in, This indicates an image without uniformity correction, and CH represents the number of channels in the receiving coil.

[0033] Then, Cubic interpolation is used to interpolate the low-resolution sensitive spectrum into a spectrum of the same size as the cerebral vascular image resolution. The coil sensitivity information of each channel is calculated, and the coil sensitivity information is used to correct the fully sampled cerebral vascular image to obtain multiple sets of more uniform fully sampled cerebral vascular images.

[0034] The obtained fully sampled cerebral vascular images, after coil uniformity correction, are used for training. The training process is as follows: Figure 4 As shown.

[0035] The image Dw, which is used to remove noise artifacts, is trained using a neural network, as shown in the following formula: Dw is the denoised and artifact-free image of image x after being processed by a pre-trained neural network. When the sampled image x contains a lot of noise and artifacts, the value of the regularization term Nw is large, and vice versa. Where Jn is the Jacobian matrix, A denoising and artifact removal neural network is pre-trained, and the denoising and artifact removal image Dw is obtained after N iterations. Substituting this into the model-based convolutional neural network formula, the following formula can be obtained: Further transformed The denoised image is obtained by using a pre-trained denoising model, and the consistency of the data is trained by using the conjugate gradient iteration method to form a data consistency constraint term, thereby accelerating the speed of iteration convergence.

[0036] As shown in the network flow diagram Figure 4 As shown, after correcting multiple sets of fully acquired data using coil sensitivity, undersampled data is obtained using Poisson variable density spectrum sampling. The undersampled data is then used in the model for multiple iterations with the fully sampled image as the target. Through multiple training iterations, a TOF3D cerebral vascular neural network with uniformity correction effect is obtained.

[0037] In practical applications, a pre-designed Time-of-Flight (TOF) sequence with a Poisson distribution sampling trajectory is used to rapidly acquire images of human brain blood vessels. After rearranging, zero-filling, and performing a 3D Fourier transform on the undersampled data, the data is input into a pre-trained TOF3D brain blood vessel neural network for processing, resulting in high-quality multi-segment brain blood vessel images.

[0038] Step S3: The multiple uniform cerebral vascular images are stitched together to obtain a three-dimensional cerebral vascular image. To reduce blood flow saturation, the 3D cerebral vascular images are acquired in a multi-slab overlapping manner. The conventional method is to discard the extra slab portions directly, but this method will produce Venetian blind artifacts. Therefore, Venetian blind filtering is required to eliminate the non-uniformity at the stitching of blocks. In this embodiment, the Venetian blind filtering method is used. The image after Venetian blind correction has smoother stitching at the blocks.

[0039] Step S4: Perform maximum density projection (MIP) on the three-dimensional cerebral vascular image in the magnetic resonance imaging system, and crop and remove background tissue. Perform 360° projection in three directions to obtain the final 3D cerebral vascular image of the head.

[0040] This method employs a Poisson variable density undersampling trajectory to acquire cerebral vascular data, effectively reducing the amount of data scanned and significantly improving scanning speed. It utilizes a TOF deep learning neural network with uniformity correction to recover undersampling data, greatly improving the quality and effect of image reconstruction. Furthermore, venetian blind artifact correction significantly enhances the continuity of stitching between blocks. These methods greatly improve the speed of cerebral vascular scanning and reconstruction, avoiding the drawbacks of conventional methods such as excessively long scanning times and significant motion-related effects, thus achieving the effect of rapidly obtaining high-quality cerebral vascular images.

[0041] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A rapid scanning and reconstruction method for acquiring cerebral vascular images, characterized in that: Specifically, the following steps are included: Step S1: In the Ky-Kz plane of the 3DTOF sequence, Poisson variable density K-space trajectory data is acquired using the K-space, and the undersampling acceleration coefficients in the Y and Z directions are obtained respectively. Based on the undersampling acceleration coefficients, the Poisson undersampling trajectory distribution of the K-space variable density is calculated. At the same time, the coding gradient values ​​in the phase direction and layer direction are determined based on the coordinate position of the Poisson point in the Ky-Kz plane. By applying the corresponding gradient values ​​on the physical axes of the Y and Z axes, the effect of Poisson variable density acquisition is achieved. Step S2: Use a pre-trained deep learning network to recover the Poisson undersampled data acquired by Poisson variable density acquisition to obtain multi-segment cerebral blood vessel images; Step S3: stitch together multiple uniform cerebral vascular images to obtain a three-dimensional cerebral vascular image; Step S4: Perform maximum density projection (MIP) on the three-dimensional cerebral vascular image in the magnetic resonance imaging system, and crop and remove background tissue. Perform 360° projection in three directions to obtain the final 3D cerebral vascular image of the head.

2. The scanning and reconstruction method for rapidly acquiring cerebral vascular images according to claim 1, characterized in that: In step S2, the deep learning network is trained as follows: a large amount of fully sampled cerebral vascular data is collected in the early stage. Based on the Poisson undersampling spectrum when the actual Poisson undersampling is performed, the fully sampled cerebral vascular data is extracted to form multiple sets of fully sampled images and images extracted from the Poisson sampling spectrum. The model-based deep learning model is trained using a large number of fully sampled and undersampled image sets to obtain a trained deep learning model.

3. The method for rapidly acquiring and reconstructing cerebral vascular images according to claim 2, characterized in that: To ensure that the trained neural network has more information on coil uniformity, coil uniformity correction is performed on the fully sampled cerebral vascular images before training.

4. The scanning and reconstruction method for rapidly acquiring cerebral vascular images according to claim 3, characterized in that: Uniformity correction requires coil sensitivity information, necessitating a pre-scan using both a volume coil and an imaging coil to acquire the coil sensitivity spectra for each channel at low resolution. , ,in, For the pre-scan imaging coil image, The image is a pre-scanned volume coil image, where ic represents the coil channel unit. Then, trilinear interpolation is used to calculate the sensitivity spectrum information of the corresponding layer in the TOF3D algorithm. Finally, the RAIN correction algorithm is used to obtain a uniform original image. The formula for coil uniformity correction of all acquired cerebral vascular images before training is as follows: ,in, This indicates an image without uniformity correction, and CH represents the number of channels in the receiving coil.

5. The scanning and reconstruction method for rapidly acquiring cerebral vascular images according to claim 2, characterized in that: The formula for a model-based convolutional neural network is shown below: Where A = SF; F is the 3D discrete Fourier transform, S is the sampling spectrum formed by the Poisson sampling trajectory, b is the actual undersampled K-space data, Nw(x) is a regularization term, and is a noise and artifact estimator trained by a neural network, expressed by the parameters w learned by the neural network.

6. The scanning and reconstruction method for rapidly acquiring cerebral vascular images according to claim 5, characterized in that: The image Dw, which is used to remove noise artifacts, is trained using a neural network, as shown in the following formula: Dw is the denoised and artifact-free image of image x after being processed by a pre-trained neural network. When the sampled image x contains a lot of noise and artifacts, the value of the regularization term Nw is large, and vice versa. Where Jn is the Jacobian matrix, A denoising and artifact removal neural network is pre-trained, and the denoising and artifact removal image Dw is obtained after N iterations. Substituting this into the model-based convolutional neural network formula, the following formula can be obtained: Further transformed The denoised image is obtained by using a pre-trained denoising model, and the consistency of the data is trained by using the conjugate gradient iteration method to form a data consistency constraint term, thereby accelerating the speed of iteration convergence.

7. The method for rapidly acquiring and reconstructing cerebral vascular images according to claim 1, characterized in that: In step S1, the Kx direction in the 3DTOF sequence is acquired using a Cartesian acquisition method.

8. The method for rapidly acquiring and reconstructing cerebral vascular images according to claim 1, characterized in that: In step S3, multiple uniform cerebral blood vessel images are stitched together, and the connection points of the multiple cerebral blood vessels are processed using venetian blind filtering.

9. A magnetic resonance imaging system, characterized in that, The magnetic resonance imaging system employs a scanning and reconstruction method for rapidly acquiring cerebral vascular images, as described in any one of claims 1-8.

Images