Magnetic resonance angiography image reconstruction method

Abstract

The present invention discloses a magnetic resonance vascular image reconstruction method, comprising the following steps: 1) collecting magnetic resonance three-dimensional k-space data related to low-resolution 3D magnetic resonance brain images for calculating coil sensitivity information; 2) using the ESPIRiT method to calculate coil sensitivity for the low-resolution three-dimensional k-space data in step 1) to obtain a sensitivity map; 3) using the CAIPIRINHA fast imaging technology to collect 4DMRA data, and rearranging the k-space data in accordance with the acquisition time sequence; 4) using the downsampled k-space data obtained in step 3) and the sensitivity map obtained in step 2) as input parameters of a 4D magnetic resonance vascular image reconstruction model based on a self-supervised reconstruction model to reconstruct the magnetic resonance vascular image. The present invention solves the problem of poor image quality in conventional magnetic resonance parallel imaging when undersampling is high, and the reconstructed image has high quality and a wide range of applications.

Description

Technical Field

[0001] The invention relates to the field of magnetic resonance image reconstruction, in particular to a non-contrast enhanced four-dimensional magnetic resonance vascular image reconstruction method based on a self-supervised reconstruction model. Background Art

[0002] Magnetic resonance imaging can obtain different tissue signals by selecting different imaging sequences and parameters to provide more diagnostic information. At the same time, it also has the advantages of high soft tissue resolution, multi-directional imaging, and no ionizing radiation. It is currently a routine clinical examination method. The biggest disadvantage of magnetic resonance imaging is that the data acquisition time is long, which leads to slow imaging speed.

[0003] Dynamic magnetic resonance angiography is an emerging and important tool for clinical diagnosis of cerebrovascular diseases. It can simultaneously display vascular structure and dynamic blood flow patterns. Traditional time-resolved four-dimensional magnetic resonance angiography (MR angiography, MRA) based on arterial spin labeling (arterial spin labeling, ASL) adopts a multiphase segmented three-dimensional Cartesian acquisition scheme. The segmented acquisition scheme can achieve higher spatial and temporal resolution. However, when the imaging scan field is large, the acquisition time will be greatly extended and easily affected by the patient's voluntary or involuntary movement, thus hindering the clinical application of ASL-based four-dimensional MRA imaging technology.

[0004] The emergence of multi-channel acquisition technology and parallel imaging image reconstruction algorithms has greatly accelerated the speed of magnetic resonance imaging. Commonly used parallel imaging methods in clinical practice include SENSE (sensitivity encoding), GRAPPA (generalized autocalibrating partially parallel acquisitions), and improved technologies based on these two methods. Parallel imaging is currently the most widely used and mature rapid imaging method in clinical magnetic resonance imaging. Parallel imaging uses the spatial sensitivity information of the coil to recover undersampled data, and it has its own limitations. First, the maximum acceleration factor of parallel imaging is constrained by the number of coils, which is determined by the principle of parallel imaging. Secondly, in the case of high acceleration factor sampling, parallel imaging technology can easily amplify noise, making the g-factor (geometric factor) in the result too large, resulting in large residuals in the final reconstructed result and poor image quality.

[0005] SENSE, proposed by Pruessmann et al. in 1999, is a classic parallel imaging MRI image reconstruction method, which mainly includes four steps: (1) generating coil sensitivity maps; (2) acquiring downsampled k-space data; (3) reconstructing images of part of the field of view from each coil; and (4) expanding / merging the images of part of the field of view by matrix inversion to obtain the final reconstructed image.

[0006] Felix A. Breuer et al. proposed a fast MRI imaging method in 2006 (doi: 10.1002 / mrm.20787), called 2D CAIPIRINHA, which uses a unique set of k-space sampling patterns to reduce pixel aliasing and overlap on reconstructed images. Due to the additional offset applied to the phase encoding gradient table, the data points collected in k-space are also shifted. Using this improved sampling strategy, coil sensitivity changes can be more effectively utilized in multiple dimensions, resulting in more robust parallel imaging reconstructed images.

[0007] In 2014, Martin Uecker et al. proposed a high-quality method for calculating coil sensitivity, ESPIRiT (doi: 10.1002 / mrm.24751).

[0008] CAIPIRINHA is an improved parallel imaging technology, which is also constrained by the principle of parallel imaging. When the acceleration factor is too high, the noise level of the reconstructed image will be greatly increased and the aliasing artifacts will become very serious. Summary of the invention

[0009] In view of this, the present invention aims to solve the problem that the quality of reconstructing non-contrast enhanced four-dimensional magnetic resonance vascular images under high-speed sampling under the above-mentioned existing parallel imaging technology is poor, and proposes a non-contrast enhanced four-dimensional magnetic resonance vascular image reconstruction method based on a self-supervised reconstruction model.

[0010] The technical solution of the present invention is to provide a non-contrast enhanced four-dimensional magnetic resonance vascular image reconstruction method based on a self-supervised reconstruction model, comprising the following steps:

[0011] 1) Collecting 3D MRI k-space data related to low-resolution 3D MRI brain images for calculating coil sensitivity information;

[0012] 2) using the ESPIRiT method to calculate coil sensitivity on the low-resolution three-dimensional k-space data in step 1) to obtain a sensitivity map;

[0013] 3) Using CAIPIRINHA fast imaging technology to acquire 4D MRA data and rearrange the k-space data in the order of acquisition time;

[0014] 4) Using the downsampled k-space data obtained in step 3) and the sensitivity map obtained in step 2) as input parameters of the improved SENSE magnetic resonance vascular image reconstruction model to achieve magnetic resonance vascular image reconstruction, thereby obtaining the CAIPI-SENSE magnetic resonance vascular image reconstruction model as follows:

[0015]

[0016] Where M is the downsampling matrix, F is the fast Fourier transform operation, C is the coil sensitivity information, are the markers and reference images to be reconstructed, is the undersampled three-dimensional k-space data, T represents the adaptive sparse transform operator, and μ is the weight factor.

[0017] Compared with the prior art, the present invention has the following advantages: (1) The reconstructed image denoising effect is good. Compared with the traditional SENSE or GRAPPA parallel imaging methods, the present invention has higher accuracy, which is specifically manifested as lower normalized root-mean-square-error (NRMSE) and higher structural similarity (SSIM); (2) The imaging speed is fast, and high-quality images can still be obtained under high acceleration factors; (3) The application range is wide, and it can be used for dynamic imaging of the human heart, three-dimensional dynamic imaging of the liver, etc., and can also be used for small animal imaging, mainly for four-dimensional MRA dynamic magnetic resonance imaging, and can also be used for dynamic CT imaging, etc.

[0018] As an improvement, the magnetic resonance image reconstruction method proposed in the present invention is an improvement of the traditional SENSE method on the reconstruction of non-contrast enhanced four-dimensional MRA images. Specifically, a four-dimensional dynamic video block matching model (video block matching 4D, vBM4D) (M. Maggioni, G. Boracchi, A. Foi, K. Egiazarian, "Video Denoising, Deblocking and Enhancement Through Separable 4-D Nonlocal Spatiotemporal Transforms", IEEE Trans. Image Proc., vol. 21, no. 9, Sep. 2012. doi: 10.1109 / TIP. 2012.2199324) is introduced into the SENSE image reconstruction, and the CAIPIRINHA method is used for k-space data downsampling. Therefore, this method is referred to as vBM4D-CAIPI-SENSE. The reconstruction model is a self-supervised model, which mines self-supervised information from MRI data for supervised learning and training, such as prior information such as Gaussian noise distribution of MRI images.

[0019] In step 4), a self-supervised four-dimensional dynamic video block matching model learning mechanism is introduced into the SENSE magnetic resonance vascular image reconstruction model, and the parallel imaging is further denoised and reconstructed by using the Gaussian noise distribution prior condition of the MRI image. The CAIPI-SENSE magnetic resonance vascular image reconstruction model is equivalently transformed into the vBM4D-CAIPI-SE NSE magnetic resonance vascular image reconstruction model. The reconstruction model of vBM4D-CAIPI-SENSE can be equivalently converted to:

[0020]

[0021]

[0022] Where M is the downsampling matrix, F is the fast Fourier transform operation, C is the coil sensitivity information, are the label and control images to be reconstructed, yes The estimated image of is the undersampled 3D k-space data, T t Represents the vBM4D adaptive sparse transform operator in each iteration process, is the result image obtained at the tth iteration, λ 1 and λ 2is a weight factor. The present invention uses an iterative method of convex set projection to solve the magnetic resonance image reconstruction model.

[0023] As an improvement, in step 3), 2D CAIPIRINHA fast imaging technology is used to acquire four-dimensional MRA data.

[0024] As an improvement, the reconstruction model obtained in step 4) is selected according to different experimental data to select the appropriate model parameter λ 1 and λ 2 Perform iterative reconstruction until the optimal reconstructed image is obtained.

[0025] As an improvement, for the model parameter λ 1 and λ 2 If there is a reference image, the NRMSE between the reconstruction result and the reference image shall prevail. The smaller the NRMSE, the better. The model parameter corresponding to the reconstructed image with the minimum NRMSE is the optimal λ. 1 and λ 2 ,If there is no standard reference image, the radiologist's score shall prevail, and the model parameters corresponding to the set of reconstruction results with the highest score shall be the optimal parameter combination. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the principle of the magnetic resonance vascular image reconstruction method of the present invention;

[0027] Figure 2 3-frame reconstructed image comparison of SENSE, CAIPI and vBM4D-CAIPI-SENSE methods under 4-fold undersampling in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments, but the present invention is not limited to these embodiments. The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details.

[0029] A magnetic resonance vascular image reconstruction method of the present invention comprises the following steps:

[0030] 1) Collecting 3D MRI k-space data related to low-resolution 3D MRI brain images for calculating coil sensitivity information;

[0031] 2) using the ESPIRiT method to calculate coil sensitivity on the low-resolution three-dimensional k-space data in step 1) to obtain a sensitivity map;

[0032] 3) using CAIPIRINHA fast imaging technology to acquire downsampled k-space data corresponding to four-dimensional magnetic resonance angiography, and rearrange the downsampled k-space data in chronological order of acquisition;

[0033] 4) Using the downsampled k-space data obtained in step 3) and the sensitivity map obtained in step 2) as input parameters of the improved SENSE magnetic resonance vascular image reconstruction model to achieve magnetic resonance vascular image reconstruction, thereby obtaining the CAIPI-SENSE magnetic resonance vascular image reconstruction model as follows:

[0034]

[0035] Where M is the downsampling matrix, F is the fast Fourier transform operation, C is the coil sensitivity information, are the markers and reference images to be reconstructed, is the undersampled three-dimensional k-space data, T represents the adaptive sparse transform operator, and μ is the weight factor.

[0036] The present invention uses normalized root-mean-square-error (NRMSE) and structure similarity (SSIM) to evaluate the quality of magnetic resonance images reconstructed by the proposed vBM4D-CAIPI-SENSE method and the current mainstream magnetic resonance image reconstruction method.

[0037] NRMSE = ||I ref -I|| F / ||I ref || F

[0038] Among them I ref is the standard reference image, and I is the image reconstructed by traditional SENSE, CAIPI or vBM4D-CAIPI-SENSE methods. || || F Refers to the F norm of the image. The smaller the value of NRMSE, the closer the reconstructed image is to the original reference image, and the higher the reconstruction accuracy. Conversely, the greater the difference, the lower the accuracy of the reconstruction algorithm.

[0039] SSIM is an indicator used to measure the similarity of the subtle structures of two images. This indicator is consistent with the evaluation of human vision. We use the reference image as a benchmark to calculate the similarity between each reconstructed image and it. The SSIM calculation formula is as follows:

[0040]

[0041] Similar to the previous formula, Iref μ represents the standard reference image, and I represents the reconstructed image. ref and μ I I ref and the average signal strength value of I, and I ref and the variance of I, isI ref and the covariance of I. k 1 and k 2 These are two control variables to prevent the denominator from being equal to 0. The larger the SSIM value, the closer the reconstructed image is to the original reference image, and the higher the reconstruction accuracy. Conversely, the greater the difference, the lower the accuracy of the reconstruction algorithm.

[0042] Figure 1 The figure shows the principle schematic diagram of the vBM4D-CAIPI-SENSE magnetic resonance vascular image reconstruction method of the present invention. The vBM4D-CAIPI-SENSE reconstruction model is used to reconstruct the undersampled four-dimensional MRA data. Because the noise distribution model of MRA is known, it can be used as prior knowledge to perform self-supervised vBM4D-CAIPI-SENSE image denoising and image reconstruction.

[0043] In step 4), a self-supervised four-dimensional dynamic video block matching model learning mechanism is introduced into the SENSE magnetic resonance vascular image reconstruction model, and the parallel imaging is further denoised and reconstructed by using the Gaussian noise distribution prior condition of the MRI image. The CAIPI-SENSE magnetic resonance vascular image reconstruction model is equivalently transformed into the vBM4D-CAIPI-SE NSE magnetic resonance vascular image reconstruction model. The reconstruction model of the vBM4D-CAIPI-SENSE is:

[0044]

[0045]

[0046] Where M is the downsampling matrix, F is the fast Fourier transform operation, C is the coil sensitivity information, are the label and control images to be reconstructed, yes The estimated image of is the undersampled 3D k-space data, T t Represents the vBM4D adaptive sparse transform operator in each iteration process, is the result image obtained at the tth iteration, λ 1 and λ 2is a weight factor. The present invention uses an iterative method of convex set projection to solve the magnetic resonance image reconstruction model.

[0047] Figure 2 The first embodiment of the present invention is shown in FIG. 4D MRA images are acquired on a 3T MRI device using a 20-channel head and neck coil. The sequence used is a flow-sensitive alternating inversion recovery (FAIR) pulse sequence. The imaging parameters mainly include: matrix size 144×192×28, pulse repetition time TR 100ms, echo time TE 1.8ms, flip angle 12°, layer thickness 1.5mm, number of layers 32, 2x oversampling in the readout direction, and spatial resolution 1.15×1.15×1.5mm 3 .

[0048] In Example 1, the k-space data is undersampled by 4 times, λ 1 and λ 2 Take 0.06 and 0.01 respectively. It can be seen from the NRMSE results that the difference between the vBM4D-CAIPI-SENSE method and the standard reference image is smaller than that between the traditional SENSE method and the CAIPI method. The NRMSE of SENSE, CAIPI and vBM4D-CAIPI-SENSE are 42.98%, 37.52% and 27.55% respectively. The SSIM of SENSE, CAIPI and vBM4D-CAIPI-SENSE are 0.7602, 0.7784 and 0.9431 respectively. It can be seen that the similarity between vBM4D-CAIPI-SENSE and the reference image is the highest. It can also be seen from the figure that the vBM4D-CAIPI-SENSE method proposed in the present invention has a cleaner background, a higher signal-to-noise ratio and better image quality than the traditional SENSE and CAIPI methods.

[0049] The above only describes the preferred embodiments of the present invention, but it should not be understood as limiting the claims. The present invention is not limited to the above embodiments, and its specific structure is allowed to be changed. In short, all changes made within the protection scope of the independent claims of the present invention are within the protection scope of the present invention.

Claims

1. A magnetic resonance vascular image reconstruction method, comprising the following steps: 1) Collecting 3D MRI k-space data related to low-resolution 3D MRI brain images for calculating coil sensitivity information; 2) using the ESPIRiT method to calculate coil sensitivity on the low-resolution three-dimensional k-space data in step 1) to obtain a sensitivity map; 3) using CAIPIRINHA fast imaging technology to acquire downsampled k-space data corresponding to four-dimensional magnetic resonance angiography, and rearrange the downsampled k-space data in chronological order of acquisition; 4) Using the downsampled k-space data obtained in step 3) and the sensitivity map obtained in step 2) as input parameters of the improved SENSE magnetic resonance vascular image reconstruction model to achieve magnetic resonance vascular image reconstruction, thereby obtaining the CAIPI-SENSE magnetic resonance vascular image reconstruction model as follows: Where M is the downsampling matrix, F is the fast Fourier transform operation, C is the coil sensitivity information, are the marker and reference images to be reconstructed, is the undersampled three-dimensional k-space data, T represents the adaptive sparse transform operator, and μ is the weight factor.

2. The magnetic resonance vascular image reconstruction method according to claim 1, characterized in that: In step 4), a self-supervised four-dimensional dynamic video block matching model learning mechanism is introduced into the SENSE magnetic resonance vascular image reconstruction model, and the parallel imaging is further denoised and reconstructed by using the Gaussian noise distribution prior condition of the MRI image. The CAIPI-SENSE magnetic resonance vascular image reconstruction model is equivalently transformed into the vBM4D-CAIPI-SENSE magnetic resonance vascular image reconstruction model: Where M is the downsampling matrix, F is the fast Fourier transform operation, C is the coil sensitivity information, are the marker and reference images to be reconstructed, yes The estimated image of is the undersampled 3D k-space data, T t represents the adaptive sparse transform operator of the four-dimensional dynamic video block matching model in each iteration process, is the result image obtained at the tth iteration, λ1 and λ2 are weight factors, and the convex set projection iterative method is used to solve the magnetic resonance image reconstruction model.

3. The magnetic resonance vascular image reconstruction method according to claim 1 or 2, characterized in that: In step 3), the 2DCAIPIRINHA fast imaging technology is used to collect the corresponding downsampled k-space data of four-dimensional magnetic resonance angiography.

4. The magnetic resonance vascular image reconstruction method according to claim 3, characterized in that: The reconstruction model obtained in step 4) is iteratively reconstructed by selecting appropriate model parameters λ1 and λ2 according to different experimental data until the optimal reconstructed image is obtained.

5. The magnetic resonance image reconstruction method according to claim 4, characterized in that: For the selection of model parameters λ1 and λ2, if there is a reference image, the normalized root mean square error between the reconstruction result and the reference image shall be used as the basis. The smaller the normalized root mean square error, the better. The model parameters corresponding to the reconstructed image with the minimum normalized root mean square error are the optimal λ1 and λ2. If there is no standard reference image, the score of the radiologist shall be used as the basis. The model parameters corresponding to the set of reconstruction results with the best score shall be the optimal parameter combination.

6. The magnetic resonance image reconstruction method according to claim 4, characterized in that: For the selection of model parameters λ1 and λ2, if there is a reference image, the structural similarity between the reconstruction result and the reference image shall be used as the basis. The greater the structural similarity, the better. The model parameters corresponding to the reconstructed image with the maximum structural similarity are the optimal λ1 and λ2. If there is no standard reference image, the score of the radiologist shall be used as the basis. The model parameters corresponding to the set of reconstruction results with the best score shall be the optimal parameter combination.

Images