WO2021082103A1 - 基于深度学习的无变形单次激发平面回波成像方法及装置 - Google Patents

基于深度学习的无变形单次激发平面回波成像方法及装置 Download PDF

Info

Publication number
WO2021082103A1
WO2021082103A1 PCT/CN2019/119490 CN2019119490W WO2021082103A1 WO 2021082103 A1 WO2021082103 A1 WO 2021082103A1 CN 2019119490 W CN2019119490 W CN 2019119490W WO 2021082103 A1 WO2021082103 A1 WO 2021082103A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
images
shot
imaging
loss
Prior art date
Application number
PCT/CN2019/119490
Other languages
English (en)
French (fr)
Inventor
郭华
胡张选
Original Assignee
清华大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 清华大学 filed Critical 清华大学
Publication of WO2021082103A1 publication Critical patent/WO2021082103A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R33/00—Arrangements or instruments for measuring magnetic variables
    • G01R33/20—Arrangements or instruments for measuring magnetic variables involving magnetic resonance
    • G01R33/44—Arrangements or instruments for measuring magnetic variables involving magnetic resonance using nuclear magnetic resonance [NMR]
    • G01R33/48—NMR imaging systems
    • G01R33/54—Signal processing systems, e.g. using pulse sequences ; Generation or control of pulse sequences; Operator console
    • G01R33/56—Image enhancement or correction, e.g. subtraction or averaging techniques, e.g. improvement of signal-to-noise ratio and resolution
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves 
    • A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves  involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging

Definitions

  • the invention relates to the technical field of plane echo imaging, in particular to a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot plane echo imaging method and device.
  • EPI Echo Planar Imaging, Planar Echo Acquisition Technology
  • fast acquisition such as fast imaging speed and insensitivity to motion. It has been widely used in clinical practice, especially single-shot EPI, after a single RF excitation Accomplishing the acquisition of the entire k-space has very important value in applications that require high imaging speed, such as diffusion imaging, functional imaging, perfusion imaging, cardiac imaging, and real-time imaging.
  • EPI acquisition also has its own shortcomings. Longer readout time will introduce The blurring effect caused by attenuation and the lower bandwidth of the phase encoding direction will cause severe image distortion at the junction of different tissues with large magnetic media rates, which will affect the observation of important tissue structures and the results of quantitative analysis.
  • MS-EPI multiple excitation EPI
  • iEPI internal EPI
  • rsEPI readout-segmented EPI
  • PROPELLER-EPI etc. divide the entire k-space acquisition into several parts, which can be Under the condition of noise ratio, the above problems are reduced, but it is still unable to completely eliminate the distortion artifacts unique to EPI.
  • PSF Point spread function, based on point spread function coding
  • EPI acquisition provides an effective way to solve these problems.
  • the resulting EPI has no distortion and no image blur caused by T2* attenuation.
  • tilted - The acquisition acceleration of CAIPI technology greatly improves the time efficiency of PSF-EPI.
  • Another type of fast imaging method without distortion in MRI includes fast spin echo (FSE) and fast gradient echo (FFE).
  • FSE fast spin echo
  • FFE fast gradient echo
  • This type of technology collects multiple images after one excitation.
  • a spin echo or gradient echo signal that is, one excitation and collection of multiple coding positions, to achieve the purpose of accelerating the collection.
  • the deformation-free imaging techniques mentioned above are all based on multiple excitation methods, which greatly extend the acquisition time and are often difficult to apply in clinical acquisitions that require high time efficiency, such as diffusion magnetic resonance imaging, functional magnetic resonance imaging, etc. .
  • the present invention aims to solve one of the technical problems in the related art at least to a certain extent.
  • an object of the present invention is to propose a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging method, which can achieve high resolution under a single-shot fast scan.
  • the distortion-free magnetic resonance image with high signal-to-noise ratio can effectively solve the problems of signal-to-noise ratio, serious distortion artifacts, low resolution in the existing single-shot EPI technology, and the acquisition time that exists in the use of multiple-shot technology. Long question.
  • Another object of the present invention is to provide a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging device.
  • one embodiment of the present invention proposes a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot plane echo imaging method, which includes the following steps: obtaining a single-shot plane echo The first image and related auxiliary images, and the second image that meets preset conditions collected by multiple excitation methods is acquired; deep neural network training is performed according to the first image, related auxiliary images, and the second image to obtain the network Weight parameters to generate a deep neural network; receive a third image and related auxiliary images of a single excitation plane echo, and input the third image and related auxiliary images to the deep neural network to generate imaging results.
  • the deep learning-based high-resolution, high-signal-to-noise ratio non-deformation single-shot planar echo imaging method in the embodiment of the present invention combines deep learning with single-shot EPI acquisition in a single-phase encoding direction, and uses a multiple-shot mode
  • the collected undistorted EPI images are used as the standard for network learning, so that the acquisition time of a single excitation EPI in a single phase encoding direction can be used to obtain high-resolution, high-signal-to-noise ratio and distortion-free high-quality images, and at the same time, combine Auxiliary image strategy to improve the quality of the output image.
  • the deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging method according to the above-mentioned embodiments of the present invention may also have the following additional technical features:
  • the loss function used in the training process is as follows:
  • u is the image predicted by the network
  • u * is the high-quality image actually collected
  • 1 is the l1 norm loss
  • G SSIM is a convolution kernel used to make multiple loss function values at the same scale. It comes from the calculation process of the SSIM index.
  • W 1 , w 2 , and w 3 are various loss functions in the compound loss function The proportion of
  • the related auxiliary images are T2-weighted structured images, T2-FLAIR, T1-weighted structured images, etc.
  • the first image and the third image are single-phase encoding directions.
  • the first image and the third image are diffusion magnetic resonance images, and the diffusion magnetic resonance imaging includes images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions .
  • the inputting the third image and related auxiliary images to the deep neural network to generate an imaging result includes: inputting the third image of the EPI in a single shot in blocks. The image and the corresponding auxiliary image; the output is combined into a complete image by block to obtain the imaging result.
  • another embodiment of the present invention proposes a deep learning-based high-resolution, high-signal-to-noise ratio non-deformation single-shot plane echo imaging device, including: an acquisition module for obtaining a single-shot plane echo Echo the first image and related auxiliary images, and obtain the second image that meets the preset conditions collected by multiple excitation methods; the training module is used to obtain the second image according to the first image and related auxiliary images and the second image Perform deep neural network training to obtain network weight parameters to generate a deep neural network; imaging module for receiving a third image and related auxiliary images of a single excitation plane echo, and inputting the third image and related auxiliary images to The deep neural network generates imaging results.
  • the deep learning-based high-resolution, high-signal-to-noise ratio non-deformation single-excitation planar echo imaging device combines deep learning with single-excitation EPI acquisition in a single-phase encoding direction, and uses multiple excitations.
  • the collected undistorted EPI images are used as the standard for network learning, so that the acquisition time of a single excitation EPI in a single phase encoding direction can be used to obtain high-resolution, high-signal-to-noise ratio and distortion-free high-quality images, and at the same time, combine Auxiliary image strategy to improve the quality of the output image.
  • the deep learning-based high-resolution, high-signal-to-noise ratio non-deformation single-shot planar echo imaging device may also have the following additional technical features:
  • the loss function used in the training process is as follows:
  • u is the image predicted by the network
  • u * is the high-quality image actually collected
  • 1 is the l1 norm loss
  • G SSIM is a convolution kernel used to make multiple loss function values at the same scale. It comes from the calculation process of the SSIM index.
  • W 1 , w 2 , and w 3 are various loss functions in the compound loss function The proportion of
  • the related auxiliary images are T2-weighted structured images, T2-FLAIR, T1-weighted structured images, etc.
  • the first image and the third image are single-phase encoding directions.
  • the first image and the third image are diffusion magnetic resonance images, and the diffusion magnetic resonance imaging includes images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions .
  • the imaging module is further configured to block input the third image of the single-shot EPI and the corresponding auxiliary image, and combine the output into a complete image by block to obtain the Imaging results.
  • Fig. 1 is a flowchart of a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging method according to an embodiment of the present invention
  • FIG. 2 is a flow chart of a method for single-shot planar echo imaging with high resolution and high signal-to-noise ratio without distortion based on deep learning according to an embodiment of the present invention
  • Fig. 3 is a schematic diagram of a deep neural network structure according to an embodiment of the present invention.
  • Fig. 4 is a schematic structural diagram of a deep learning-based high-resolution, high-signal-to-noise ratio non-deformation single-shot planar echo imaging device according to an embodiment of the present invention.
  • Fig. 1 is a flow chart of a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging method according to an embodiment of the present invention.
  • the deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging method includes the following steps:
  • step S101 a first image of a single excitation plane echo and related auxiliary images are acquired, and a second image that meets a preset condition collected in a multiple excitation mode is acquired.
  • the embodiment of the present invention collects an image of a single excitation EPI and related auxiliary images, and collects a high-resolution, high-signal-to-noise ratio and distortion-free image in a multiple excitation mode.
  • the first image is an image that single-shot EPI
  • the second image that meets a preset condition is a high-resolution, high-signal-to-noise ratio and distortion-free image.
  • the related auxiliary images are T2-weighted structured images, T2-FLAIR, T1-weighted structured images, etc.
  • the first image and the third image are single-shot EPI acquisitions in the single-phase encoding direction. Obtained, and the first image and the third image are diffusion magnetic resonance images.
  • the diffusion magnetic resonance imaging includes an image with no diffusion encoding gradient and an image with 6 different diffusion encoding gradient directions.
  • the single-shot EPI image used in the embodiment of the present invention is diffusion magnetic resonance imaging, including an image without diffusion encoding gradient (b0) and an image with 6 different diffusion encoding gradient directions (b1), and the output is A high-resolution, high-signal-to-noise ratio and distortion-free image acquired using the same diffusion coding strategy and using multiple excitation methods (in the embodiment of the present invention, the PSF-EPI method is used for acquisition).
  • images with multiple contrasts such as T2-weighted structure images, etc., are generally scanned.
  • the embodiment of the present invention proposes a strategy of using auxiliary images, that is, deep neural
  • the input of the network includes not only single-shot EPI images, but also collected auxiliary images, such as T2-weighted structure images, T2-FLAIR, T1-weighted structure images, etc. (as shown in Figure 2), forming a composite multi-channel image.
  • auxiliary images such as T2-weighted structure images, T2-FLAIR, T1-weighted structure images, etc. (as shown in Figure 2), forming a composite multi-channel image.
  • step S102 deep neural network training is performed according to the first image, the related auxiliary image, and the second image to obtain network weight parameters to generate a deep neural network.
  • the training stage use the images collected by the single-shot EPI and the related auxiliary images and the high-resolution, high-signal-to-noise-ratio undistorted images collected by the multiple-shot method for deep neural network training to obtain network weight parameters.
  • the deep neural network structure used in the embodiment of the present invention is shown in FIG. 3, and the input image contains 8 channels (respectively the b0 image, 6 b1 images and the auxiliary T2 weighted structure of the single-shot EPI acquisition. Image), the output image contains 7 channels (b0 image and 6 b1 images collected by PSF-EPI respectively).
  • the network uses the commonly used U-net structure, in which the scaling of features is achieved through a stepped convolutional layer and a deconvolutional layer. Each convolutional layer includes a BN layer (Batch Normalization) and an activation layer (ReLU).
  • the image is trained in blocks along the phase encoding direction.
  • the input matrix size is 217*32*8, and the output matrix size is 224*32*8.
  • the size and number of convolution kernels of each layer are not detailed here. Listed.
  • the loss function used in the training process is as follows:
  • u is the image predicted by the network
  • u * is the high-quality image actually collected
  • 1 is the l1 norm loss (mean absolute error, MAE)
  • G SSIM is a convolution kernel used to make multiple loss function values at the same scale. It comes from the calculation process of the SSIM index.
  • W 1 , w 2 , and w 3 are various loss functions in the compound loss function The proportion of
  • the embodiment of the present invention does not impose restrictions on the imaging contrast.
  • it can also be applied to T1 weighting, T2 weighting, T2* weighting, and proton density acquired using this method. (PD) weighting, etc.; in the embodiment of the present invention, the EPI image in the single phase encoding direction is used, and the EPI image in the bidirectional phase encoding direction can also be input at the same time.
  • the present invention does not impose restrictions on this; high resolution and high signal-to-noise ratio
  • the acquisition method of undistorted images is not limited to point spread function-encoded EPI (PSF-EPI), it can also use multi-excitation spin echo acquisition (FSE), or gradient echo acquisition (FFE), etc., or use images
  • Deformation correction algorithms (such as fieldmap correction, topup, or a combination of the two) are used to obtain high-resolution non-deformation images
  • the diffusion preparation sequence of diffusion imaging used in the embodiment of the present invention is PGSE (pulsed gradient spin echo, pulse gradient automatic STE (stimulated echo) diffusion preparation sequence, oscillating gradient spin echo (OGSE) diffusion preparation sequence, double diffusion encoding (DDE) diffusion preparation sequence can also be used Sequence, convex optimized diffusion encoding (CODE) diffusion preparation sequence, etc.; the number of network layers, number of convolution kernels, convolution kernel size, activation mode, and regularization mode used by U-net in the embodiment of the present invention ,
  • the present invention does not impose restrictions on the optimizers (such as Adam, SGD, etc.) used in the training process, and various parameters (such as learning rate, batch-size, etc.); this
  • the network structure used in the invention is not limited to U-net, and ResNet (Residual Networks), GAN (Generative Adversarial Network), etc. and their variants can also be used;
  • the invention does not impose restrictions on the matrix size of the input image and output image, depending on Regarding the collected image resolution, image modality, and block size, etc.;
  • the embodiment of the present invention does not impose restrictions on the selected auxiliary image, and can use T2-weighted structure image, T1-weighted structure image, T2-FLAIR, etc. or Its combination.
  • step S103 a third image and related auxiliary images of a single excitation plane echo are received, and the third image and related auxiliary images are input to the deep neural network to generate an imaging result.
  • test (application) stage input the image collected by a single excitation of EPI and related auxiliary images to the deep neural network obtained in the training stage to generate the corresponding high-resolution, high-signal-to-noise ratio and high-quality image without distortion .
  • inputting the third image and related auxiliary images to the deep neural network to generate the imaging result includes: inputting the third image of a single excitation EPI and the corresponding auxiliary image in blocks; The block combines the output into a complete image to get the imaging result.
  • the EPI image and the corresponding auxiliary image are excited once by block input, and then the output is combined into a complete image by block , That is, the corresponding high-resolution and high-signal-to-noise ratio undistorted images can be obtained.
  • the embodiment of the present invention combines deep learning and single-shot EPI acquisition technology, and the image learning obtained by single-shot EPI acquisition generates high-resolution, high-signal-to-noise ratio acquired by multiple-shot technology acquisition without distortion Image, so as to achieve the purpose of quickly obtaining high-quality magnetic resonance images using only a single excitation EPI acquisition time.
  • deep learning is combined with single-shot EPI acquisition in a single-phase encoding direction, using multiple
  • the distortion-free EPI images acquired by the sub-excitation method are used as the standard for network learning, so that the high-resolution, high-signal-to-noise ratio and distortion-free high-quality images can be obtained using only the acquisition time of the single-shot EPI in the single-phase encoding direction.
  • combined with the strategy of auxiliary images improve the quality of the output image.
  • Fig. 4 is a schematic structural diagram of a deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging device based on an embodiment of the present invention.
  • the deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging device 10 includes: an acquisition module 100, a training module 200, and an imaging module 300.
  • the acquisition module 100 is used to acquire the first image and related auxiliary images of a single excitation plane echo, and to acquire the second image that meets the preset conditions collected by the multiple excitation mode;
  • the training module 200 is used to acquire the first image according to the first image And related auxiliary images and second images for deep neural network training to obtain network weight parameters to generate a deep neural network;
  • the imaging module 300 is used to receive the third image and related auxiliary images of a single excitation plane echo, and combine the third image And related auxiliary images are input to the deep neural network to generate imaging results.
  • the device 10 of the embodiment of the present invention can obtain high-resolution and high-signal-to-noise ratio non-deformation magnetic resonance images under a single-shot rapid scan, effectively solving the existing single-shot EPI technology of signal-to-noise ratio and distortion artifacts.
  • the loss function used in the training process is as follows:
  • u is the image predicted by the network
  • u * is the high-quality image actually collected
  • 1 is the l1 norm loss
  • G SSIM is a convolution kernel used to make multiple loss function values at the same scale. It comes from the calculation process of the SSIM index.
  • W 1 , w 2 , and w 3 are various loss functions in the compound loss function The proportion of
  • the related auxiliary images are T2-weighted structured images, T2-FLAIR, T1-weighted structured images, etc.
  • the first image and the third image are single-shot excitations in the single-phase encoding direction.
  • EPI acquisition is obtained, and the first image and the third image are diffusion magnetic resonance images.
  • the diffusion magnetic resonance imaging includes images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
  • the imaging module 300 is further configured to block input the third image of the single-shot EPI and the corresponding auxiliary image, and combine the output into a complete image by block to obtain the imaging result.
  • the deep learning-based high-resolution, high-signal-to-noise ratio and distortion-free single-shot planar echo imaging device combines deep learning with single-shot EPI acquisition in a single-phase encoding direction, and uses multiple
  • the distortion-free EPI images acquired by the sub-excitation method are used as the standard for network learning, so that the high-resolution, high-signal-to-noise ratio and distortion-free high-quality images can be obtained using only the acquisition time of the single-shot EPI in the single-phase encoding direction.
  • the strategy of auxiliary images improve the quality of the output image.
  • first and second are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with “first” and “second” may explicitly or implicitly include at least one of the features. In the description of the present invention, “plurality” means at least two, such as two, three, etc., unless otherwise specifically defined.

Landscapes

  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • General Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Radiology & Medical Imaging (AREA)
  • Surgery (AREA)
  • Biophysics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Pathology (AREA)
  • Animal Behavior & Ethology (AREA)
  • Biomedical Technology (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Signal Processing (AREA)
  • Condensed Matter Physics & Semiconductors (AREA)
  • General Physics & Mathematics (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

一种基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像方法及装置，其中，方法包括：获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次激发方式所采集的满足预设条件的第二图像（S101）；根据第一图像及相关辅助图像和第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络（S102）；接收单次激发平面回波的第三图像及相关辅助图像，将第三图像及相关辅助图像输入至深度神经网络，生成成像结果（S103）。该方法可以实现在单次激发的快速扫描下获得高分辨率高信噪比的无变形磁共振图像，有效解决现有单次激发EPI技术中存在的信噪比、变形伪影严重、分辨率低的问题以及使用多次激发技术中存在的采集时间过长的问题。

Description

基于深度学习的无变形单次激发平面回波成像方法及装置
相关申请的交叉引用
本申请要求清华大学于2019年10月31日提交的、发明名称为“基于深度学习的无变形单次激发平面回波成像方法及装置”的、中国专利申请号“201911053423.9”的优先权。
技术领域
本发明涉及平面回波成像技术领域，特别涉及一种基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法及装置。
背景技术
EPI(Echo Planar Imaging，平面回波采集技术)的快速采集特点使其具有成像速度快，对运动不敏感等优势，在临床中得到了广泛应用，尤其是单次激发EPI，在一次RF激发后完成整个k空间的采集，在对成像速度要求高的应用中具有非常重要的价值，例如扩散成像、功能成像、灌注成像、心脏成像以及实时成像等。然而，EPI采集也有它本身的不足，较长的读出时间会引入
Figure PCTCN2019119490-appb-000001
衰减造成的模糊效应，相位编码方向的较低带宽会导致在磁介质率相差较大的不同组织交界处产生严重的图像变形，从而影响重要组织结构的观察以及量化分析的结果。
单次激发EPI与并行采集技术的结合可以减少读出窗的长度以及ESP(effective echo spacing，有效回波间隔)，减少
Figure PCTCN2019119490-appb-000002
模糊效应以及图像变形，但是依然受限于加速倍数，同时会降低信噪比。MS-EPI(多次激发EPI)技术，例如iEPI(interleaved EPI)，rsEPI(readout-segmented EPI，读出分段EPI)，PROPELLER-EPI等将整个k空间采集分为若干部分，可以在保持信噪比的条件下减少上述问题，但是依然无法完全消除EPI特有的变形伪影。
PSF(Point spread function，基于点扩散函数编码)的EPI采集(PSF-EPI)为解决这些问题提供了一种有效方式，所得到的EPI完全无变形无T2*衰减引起的图像模糊，同时，tilted-CAIPI技术的采集加速大大提高了PSF-EPI的时间效率。磁共振成像中另一类无变形的快速成像方法包括快速自旋回波成像(fast spin echo，FSE)、快速梯度回波成像(fast field echo，FFE)，这一类技术通过一次激发之后采集多个自旋回波或者梯度回波信号，即一次激发采集多个编码位置，达到加速采集的目的。然而，上述提及的无变形成像技术都是基于多次激发方式，大大延长了采集时间，在对时间效率要求较高的临床采集中往往难以应 用，例如扩散磁共振成像、功能磁共振成像等。
发明内容
本申请是基于发明人对以下问题的认识和发现做出的：
受益于大数据分析的兴起以及大规模计算能力的进步，深度学习得到了极大发展，并在各种研究领域中发挥了巨大作用。在磁共振成像领域同样如此，涉及到磁共振成像的各个方面，包括但不限于图像采集、重建、恢复、量化分析、高分辨率重建等。相关研究例如由3T磁场强度所采集的磁共振图像生成得到7T所采集的图像；由低分辨率的T1加权图像生成高分辨率的T1加权图像等。这些研究为提高图像质量(例如提高图像分辨率、信噪比，减小图像伪影等)而不引入其他问题(例如采集时间增强、信噪比降低等)提供了新的思路。
本发明旨在至少在一定程度上解决相关技术中的技术问题之一。
为此，本发明的一个目的在于提出一种基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法，该方法可以实现在单次激发的快速扫描下获得高分辨率高信噪比的无变形磁共振图像，有效解决现有单次激发EPI技术中存在的信噪比、变形伪影严重、分辨率低的问题以及使用多次激发技术中存在的采集时间过长的问题。
本发明的另一个目的在于提出一种基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置。
为达到上述目的，本发明一方面实施例提出了一种基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法，包括以下步骤：获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次激发方式所采集的满足预设条件的第二图像；根据所述第一图像及相关辅助图像和所述第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络；接收单次激发平面回波的第三图像及相关辅助图像，将所述第三图像及相关辅助图像输入至所述深度神经网络，生成成像结果。
本发明实施例的基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像方法，将深度学习与单相位编码方向的单次激发EPI采集相结合，使用多次激发方式采集得到的无变形EPI图像作为标准进行网络学习，从而可以仅使用单相位编码方向的单次激发EPI的采集时间即可获得高分辨率高信噪比无变形的高质量图像，同时，结合辅助图像的策略，提高输出图像的质量。
另外，根据本发明上述实施例的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法还可以具有以下附加的技术特征：
进一步地，在本发明的一个实施例中，训练过程中所使用的损失函数如下：
Figure PCTCN2019119490-appb-000003
其中，u为网络所预测生成的图像，u *为实际采集的高质量图像，
Figure PCTCN2019119490-appb-000004
为SSIM损失，|u-u *| 1为l1范数损失，
Figure PCTCN2019119490-appb-000005
为一阶梯度损失，
Figure PCTCN2019119490-appb-000006
为二阶梯度损失，G SSIM为使多种损失函数值在同一尺度所使用的卷积核，来自于SSIM指数的计算过程，w 1，w 2，w 3为各种损失函数在复合损失函数中所占的比重。
进一步地，在本发明的一个实施例中，所述相关辅助图像为T2加权结构像，T2-FLAIR，T1加权结构像等，所述第一图像和所述第三图像为单相位编码方向的单次激发EPI采集得到，且所述第一图像和所述第三图像为扩散磁共振图像，所述扩散磁共振成像包含无扩散编码梯度的图像及具有6个不同扩散编码梯度方向的图像。
进一步地，在本发明的一个实施例中，所述将所述第三图像及相关辅助图像输入至所述深度神经网络，生成成像结果，包括：分块输入单次激发EPI的所述第三图像以及相应的辅助图像；按块将输出组合成完整图像，得到所述成像结果。
为达到上述目的，本发明另一方面实施例提出了一种基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像装置，包括：获取模块，用于获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次激发方式所采集的满足预设条件的第二图像；训练模块，用于根据所述第一图像及相关辅助图像和所述第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络；成像模块，用于接收单次激发平面回波的第三图像及相关辅助图像，将所述第三图像及相关辅助图像输入至所述深度神经网络，生成成像结果。
本发明实施例的基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像装置，将深度学习与单相位编码方向的单次激发EPI采集相结合，使用多次激发方式采集得到的无变形EPI图像作为标准进行网络学习，从而可以仅使用单相位编码方向的单次激发EPI的采集时间即可获得高分辨率高信噪比无变形的高质量图像，同时，结合辅助图像的策略，提高输出图像的质量。
另外，根据本发明上述实施例的基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像装置还可以具有以下附加的技术特征：
进一步地，在本发明的一个实施例中，训练过程中所使用的损失函数如下：
Figure PCTCN2019119490-appb-000007
其中，u为网络所预测生成的图像，u *为实际采集的高质量图像，
Figure PCTCN2019119490-appb-000008
为SSIM损失，|u-u *| 1为l1范数损失，
Figure PCTCN2019119490-appb-000009
为一阶梯度损失，
Figure PCTCN2019119490-appb-000010
为二阶梯度损失，G SSIM为使多种损失函数值在同一尺度所使用的卷积核，来自于SSIM指数的计算过程，w 1，w 2，w 3为各种损失函数在复合损失函数中所占的比重。
进一步地，在本发明的一个实施例中，所述相关辅助图像为T2加权结构像，T2-FLAIR，T1加权结构像等，所述第一图像和所述第三图像为单相位编码方向的单次激发EPI采集得到，且所述第一图像和所述第三图像为扩散磁共振图像，所述扩散磁共振成像包含无扩散编码梯度的图像及具有6个不同扩散编码梯度方向的图像。
进一步地，在本发明的一个实施例中，所述成像模块进一步用于分块输入单次激发EPI的所述第三图像以及相应的辅助图像，按块将输出组合成完整图像，得到所述成像结果。
本发明附加的方面和优点将在下面的描述中部分给出，部分将从下面的描述中变得明显，或通过本发明的实践了解到。
附图说明
本发明上述的和/或附加的方面和优点从下面结合附图对实施例的描述中将变得明显和容易理解，其中：
图1为根据本发明实施例的基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像方法的流程图；
图2为根据本发明一个实施例的基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像方法的流程图；
图3为根据本发明实施例的深度神经网络结构示意图；
图4为根据本发明实施例的基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像装置的结构示意图。
具体实施方式
下面详细描述本发明的实施例，所述实施例的示例在附图中示出，其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的，旨在用于解释本发明，而不能理解为对本发明的限制。
下面参照附图描述根据本发明实施例提出的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法及装置，首先将参照附图描述根据本发明实施例提出的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法。
图1是本发明一个实施例的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法的流程图。
如图1所示，该基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法包括以下步骤：
在步骤S101中，获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次 激发方式所采集的满足预设条件的第二图像。
可以理解的是，本发明实施例采集单次激发EPI的图像以及相关辅助图像，并采集多次激发方式的高分辨率高信噪比无变形图像。其中，第一图像为单次激发EPI的图像，满足预设条件的第二图像为高分辨率高信噪比无变形图像。
进一步地，在本发明的一个实施例中，相关辅助图像为T2加权结构像，T2-FLAIR，T1加权结构像等，第一图像和第三图像为单相位编码方向的单次激发EPI采集得到，且第一图像和第三图像为扩散磁共振图像，扩散磁共振成像包含无扩散编码梯度的图像及具有6个不同扩散编码梯度方向的图像。
具体而言，本发明实施例中所使用的单次激发EPI图像为扩散磁共振成像，包含无扩散编码梯度(b0)的图像以及具有6个不同扩散编码梯度方向(b1)的图像，输出为使用同一扩散编码策略并采用多次激发方式采集的高分辨率高信噪比无变形图像(本发明实施例中使用PSF-EPI方式进行采集)。同时，在常规扫描中，除了扩散磁共振图像外，一般会扫描多种对比度的图像，例如T2加权结构像等，为充分利用这些数据，本发明实施例提出使用辅助图像的策略，即深度神经网络的输入不仅仅包括单次激发EPI图像，还包括采集的辅助图像，例如T2加权结构像、T2-FLAIR、T1加权结构像等(如图2所示)，构成复合多通道图像。使用这一策略，可以大大提高神经网络的学习效果，提升输出图像的质量。
在步骤S102中，根据第一图像及相关辅助图像和第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络。
可以理解的是，训练阶段：使用单次激发EPI所采集的图像以及相关辅助图像与多次激发方式所采集的高分辨率高信噪比无变形图像进行深度神经网络训练，得到网络权重参数。
具体而言，本发明实施例中所使用的深度神经网络结构如图3所示，输入图像包含8个通道(分别为单次激发EPI采集的b0图、6个b1图以及辅助的T2加权结构像)，输出图像包含7个通道(分别为PSF-EPI采集的b0图、6个b1图)。网络使用了常用的U-net结构，其中特征的尺度缩放通过带步长的卷积层以及反卷积层实现。每个卷积层均包含了BN层(Batch Normalization)以及激活层(ReLU)。
训练过程中图像沿相位编码方向进行了分块训练，输入矩阵大小为217*32*8，输出矩阵大小为224*32*8，每层的卷积核大小及卷积核数目此处不详细列出。训练过程中所使用的损失函数如下：
Figure PCTCN2019119490-appb-000011
其中，u为网络所预测生成的图像，u *为实际采集的高质量图像，
Figure PCTCN2019119490-appb-000012
为SSIM(structural similarity index)损失，|u-u *| 1为l1范数损失(mean absolute error,MAE)，
Figure PCTCN2019119490-appb-000013
为 一阶梯度损失(gradient loss)，
Figure PCTCN2019119490-appb-000014
为二阶梯度损失，G SSIM为使多种损失函数值在同一尺度所使用的卷积核，来自于SSIM指数的计算过程，w 1，w 2，w 3为各种损失函数在复合损失函数中所占的比重。
本发明实施例对成像对比度不加限制，除实施例中应用于单次激发EPI采集的扩散磁共振成像外，也可以应用于使用该方法采集的T1加权、T2加权、T2*加权、质子密度(PD)加权等；本发明实施例中使用了单相位编码方向的EPI图像，也可以同时输入双方向相位编码方向的EPI图像，本发明对此不加限制；高分辨率高信噪比无变形图像的采集方式不局限于点扩散函数编码的EPI(PSF-EPI)，也可以使用多次激发的自旋回波采集(FSE)，或者，梯度回波采集(FFE)等，或者使用图像变形矫正算法(如fieldmap矫正，topup，或者两者相结合)来获得高分辨率无变形图像；本发明实施例中所使用的扩散成像的扩散准备序列为PGSE(pulsed gradient spin echo，脉冲梯度自旋回波)，也可以使用STE(stimulated echo，受激回波)扩散准备序列，振荡梯度自旋回波(oscillating gradient spin echo，OGSE)扩散准备序列，双重扩散编码(double diffusion encoding，DDE)扩散准备序列，凸优化扩散编码(convex optimized diffusion encoding，CODE)扩散准备序列等；本发明实施例对U-net所使用的网络层数、卷积核数目、卷积核大小、激活方式、正则化方式、激活方式、损失函数等均不加限制；本发明对训练过程所使用的优化器(如Adam、SGD等)，各种参数(如学习率、batch-size等)等亦不加限制；本发明中所使用的网络结构不限制于U-net，亦可以使用ResNet(Residual Networks)、GAN(Generative Adversarial Network)等及其变种；本发明对输入图像及输出图像的矩阵大小不加限制，取决于所采集的图像分辨率、图像模态及分块的大小等；本发明实施例对所选择使用的辅助图像不加限制，可以使用T2加权结构像、T1加权结构像、T2-FLAIR等或其组合。
在步骤S103中，接收单次激发平面回波的第三图像及相关辅助图像，将第三图像及相关辅助图像输入至深度神经网络，生成成像结果。
可以理解的是，在测试(应用)阶段：输入单次激发EPI所采集的图像及相关辅助图像至训练阶段得到的深度神经网络，生成相应的高分辨率高信噪比无变形的高质量图像。
进一步地，在本发明的一个实施例中，将第三图像及相关辅助图像输入至深度神经网络，生成成像结果，包括：分块输入单次激发EPI的第三图像以及相应的辅助图像；按块将输出组合成完整图像，得到成像结果。
可以理解的是，在经过训练阶段得到网络权重参数之后，与训练过程类似(如图2所示)，分块输入单次激发EPI图像以及相应的辅助图像，之后按块将输出组合成完整图像，即可以获得相应的高分辨率高信噪比的无变形图像。
综上，本发明实施例将深度学习与单次激发EPI采集技术相结合，由单次激发EPI采 集所得到的图像学习生成由多次激发技术采集所得到的高分辨率高信噪比无变形图像，从而达到仅使用单次激发EPI采集所需时间快速获得高质量的磁共振图像的目的。
根据本发明实施例提出的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法，将深度学习与单相位编码方向的单次激发EPI采集相结合，使用多次激发方式采集得到的无变形EPI图像作为标准进行网络学习，从而可以仅使用单相位编码方向的单次激发EPI的采集时间即可获得高分辨率高信噪比无变形的高质量图像，同时，结合辅助图像的策略，提高输出图像的质量。
其次参照附图描述根据本发明实施例提出的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置。
图4是本发明一个实施例的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置的结构示意图。
如图4所示，该基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置10包括：获取模块100、训练模块200和成像模块300。
其中，获取模块100用于获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次激发方式所采集的满足预设条件的第二图像；训练模块200用于根据第一图像及相关辅助图像和第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络；成像模块300用于接收单次激发平面回波的第三图像及相关辅助图像，将第三图像及相关辅助图像输入至深度神经网络，生成成像结果。本发明实施例的装置10可以实现在单次激发的快速扫描下获得高分辨率高信噪比的无变形磁共振图像，有效解决现有单次激发EPI技术中存在的信噪比、变形伪影严重、分辨率低的问题以及使用多次激发技术中存在的采集时间过长的问题。
进一步地，在本发明的一个实施例中，训练过程中所使用的损失函数如下：
Figure PCTCN2019119490-appb-000015
其中，u为网络所预测生成的图像，u *为实际采集的高质量图像，
Figure PCTCN2019119490-appb-000016
为SSIM损失，|u-u *| 1为l1范数损失，
Figure PCTCN2019119490-appb-000017
为一阶梯度损失，
Figure PCTCN2019119490-appb-000018
为二阶梯度损失，G SSIM为使多种损失函数值在同一尺度所使用的卷积核，来自于SSIM指数的计算过程，w 1，w 2，w 3为各种损失函数在复合损失函数中所占的比重。
进一步地，在本发明的一个实施例中，所述相关辅助图像为T2加权结构像，T2-FLAIR，T1加权结构像等，第一图像和第三图像为单相位编码方向的单次激发EPI采集得到，且第一图像和第三图像为扩散磁共振图像，扩散磁共振成像包含无扩散编码梯度的图像及具有6个不同扩散编码梯度方向的图像。
进一步地，在本发明的一个实施例中，成像模块300进一步用于分块输入单次激发EPI 的第三图像以及相应的辅助图像，按块将输出组合成完整图像，得到成像结果。
需要说明的是，前述对基于深度学习的高分辨率高信噪比无变形单次激发平面回波成像方法实施例的解释说明也适用于该实施例的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置，此处不再赘述。
根据本发明实施例提出的基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置，将深度学习与单相位编码方向的单次激发EPI采集相结合，使用多次激发方式采集得到的无变形EPI图像作为标准进行网络学习，从而可以仅使用单相位编码方向的单次激发EPI的采集时间即可获得高分辨率高信噪比无变形的高质量图像，同时，结合辅助图像的策略，提高输出图像的质量。
此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。在本发明的描述中，“多个”的含义是至少两个，例如两个，三个等，除非另有明确具体的限定。
在本说明书的描述中，参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本发明的至少一个实施例或示例中。在本说明书中，对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外，在不相互矛盾的情况下，本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
尽管上面已经示出和描述了本发明的实施例，可以理解的是，上述实施例是示例性的，不能理解为对本发明的限制，本领域的普通技术人员在本发明的范围内可以对上述实施例进行变化、修改、替换和变型。

Claims (8)

  1. 一种基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像方法，其特征在于，包括以下步骤：
    获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次激发方式所采集的满足预设条件的第二图像；
    根据所述第一图像及相关辅助图像和所述第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络；以及
    接收单次激发平面回波的第三图像及相关辅助图像，将所述第三图像及相关辅助图像输入至所述深度神经网络，生成成像结果。
  2. 根据权利要求1所述的方法，其特征在于，训练过程中所使用的损失函数如下：
    Figure PCTCN2019119490-appb-100001
    其中，u为网络所预测生成的图像，u *为实际采集的高质量图像，
    Figure PCTCN2019119490-appb-100002
    为SSIM损失，|u-u *| 1为l1范数损失，
    Figure PCTCN2019119490-appb-100003
    为一阶梯度损失，
    Figure PCTCN2019119490-appb-100004
    为二阶梯度损失，G SSIM为使多种损失函数值在同一尺度所使用的卷积核，来自于SSIM指数的计算过程，w 1，w 2，w 3为各种损失函数在复合损失函数中所占的比重。
  3. 根据权利要求1所述的方法，其特征在于，所述相关辅助图像为T2加权结构像、T2-FLAIR、T1加权结构像，第一图像和所述第三图像为单相位编码方向的单次激发EPI采集得到，且所述第一图像和所述第三图像为扩散磁共振图像，所述扩散磁共振成像包含无扩散编码梯度的图像及具有6个不同扩散编码梯度方向的图像。
  4. 根据权利要求3所述的方法，其特征在于，所述将所述第三图像及相关辅助图像输入至所述深度神经网络，生成成像结果，包括：
    分块输入单次激发EPI的所述第三图像以及相应的辅助图像；
    按块将输出组合成完整图像，得到所述成像结果。
  5. 一种基于深度学习的高分辨率高信噪比无变形的单次激发平面回波成像装置，其特征在于，包括：
    获取模块，用于获取单次激发平面回波的第一图像及相关辅助图像，并且获取多次激发方式所采集的满足预设条件的第二图像；
    训练模块，用于根据所述第一图像及相关辅助图像和所述第二图像进行深度神经网络训练，得到网络权重参数，以生成深度神经网络；以及
    成像模块，用于接收单次激发平面回波的第三图像及相关辅助图像，将所述第三图像及相关辅助图像输入至所述深度神经网络，生成成像结果。
  6. 根据权利要求5所述的装置，其特征在于，训练过程中所使用的损失函数如下：
    Figure PCTCN2019119490-appb-100005
    其中，u为网络所预测生成的图像，u *为实际采集的高质量图像，
    Figure PCTCN2019119490-appb-100006
    为SSIM损失，|u-u *| 1为l1范数损失，
    Figure PCTCN2019119490-appb-100007
    为一阶梯度损失，
    Figure PCTCN2019119490-appb-100008
    为二阶梯度损失，G SSIM为使多种损失函数值在同一尺度所使用的卷积核，来自于SSIM指数的计算过程，w 1，w 2，w 3为各种损失函数在复合损失函数中所占的比重。
  7. 根据权利要求5所述的装置，其特征在于，所述相关辅助图像为T2加权结构像、T2-FLAIR、T1加权结构像，所述第一图像和所述第三图像为单相位编码方向的单次激发EPI采集得到，且所述第一图像和所述第三图像为扩散磁共振图像，所述扩散磁共振成像包含无扩散编码梯度的图像及具有6个不同扩散编码梯度方向的图像。
  8. 根据权利要求7所述的装置，其特征在于，所述成像模块进一步用于分块输入单次激发EPI的所述第三图像以及相应的辅助图像，按块将输出组合成完整图像，得到所述成像结果。
PCT/CN2019/119490 2019-10-31 2019-11-19 基于深度学习的无变形单次激发平面回波成像方法及装置 WO2021082103A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911053423.9 2019-10-31
CN201911053423.9A CN110895320B (zh) 2019-10-31 2019-10-31 基于深度学习的无变形单次激发平面回波成像方法及装置

Publications (1)

Publication Number Publication Date
WO2021082103A1 true WO2021082103A1 (zh) 2021-05-06

Family

ID=69787439

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/119490 WO2021082103A1 (zh) 2019-10-31 2019-11-19 基于深度学习的无变形单次激发平面回波成像方法及装置

Country Status (2)

Country Link
CN (1) CN110895320B (zh)
WO (1) WO2021082103A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113808017A (zh) * 2021-08-24 2021-12-17 北京理工大学 基于照明优化的结构光成像方法及装置
CN115494439A (zh) * 2022-11-08 2022-12-20 中遥天地(北京)信息技术有限公司 一种基于深度学习的时空编码图像校正方法
CN117011409A (zh) * 2023-08-10 2023-11-07 厦门大学 多部位物理智能高清扩散磁共振数据生成方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112763958B (zh) * 2020-12-10 2022-06-21 复旦大学 一种基于神经网络的多次激发平面回波磁共振成像方法
CN113281690B (zh) * 2021-05-18 2022-08-12 上海联影医疗科技股份有限公司 一种磁共振成像方法和系统
CN113855235B (zh) * 2021-08-02 2024-06-14 应葵 用于肝脏部位的微波热消融手术中磁共振导航方法及装置
CN118096920B (zh) * 2024-03-07 2024-09-06 复旦大学 基于深度学习和线圈灵敏度的腹部器官成像方法及装置

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070247151A1 (en) * 2006-04-20 2007-10-25 Yuval Zur Method and apparatus of mr imaging with two dimensional phase and magnitude correction
CN109544652A (zh) * 2018-10-18 2019-03-29 江苏大学 基于深度生成对抗神经网络的核磁共振多加权成像方法
CN109597012A (zh) * 2018-12-24 2019-04-09 厦门大学 一种基于残差网络的单扫描时空编码成像重建方法
CN109696647A (zh) * 2019-02-21 2019-04-30 奥泰医疗系统有限责任公司 三维多次激发扩散加权成像的k空间采集方法及重建方法
CN110095742A (zh) * 2019-05-13 2019-08-06 上海东软医疗科技有限公司 一种基于神经网络的平面回波成像方法和装置
CN110244246A (zh) * 2019-07-03 2019-09-17 上海联影医疗科技有限公司 磁共振成像方法、装置、计算机设备和存储介质
CN110346743A (zh) * 2019-07-22 2019-10-18 上海东软医疗科技有限公司 一种磁共振弥散加权成像方法和装置

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
RU2698997C1 (ru) * 2016-09-06 2019-09-02 Электа, Инк. Нейронная сеть для генерации синтетических медицинских изображений
US11454690B2 (en) * 2017-12-08 2022-09-27 Rensselaer Polytechnic Institute Synergized pulsing-imaging network (SPIN)

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070247151A1 (en) * 2006-04-20 2007-10-25 Yuval Zur Method and apparatus of mr imaging with two dimensional phase and magnitude correction
CN109544652A (zh) * 2018-10-18 2019-03-29 江苏大学 基于深度生成对抗神经网络的核磁共振多加权成像方法
CN109597012A (zh) * 2018-12-24 2019-04-09 厦门大学 一种基于残差网络的单扫描时空编码成像重建方法
CN109696647A (zh) * 2019-02-21 2019-04-30 奥泰医疗系统有限责任公司 三维多次激发扩散加权成像的k空间采集方法及重建方法
CN110095742A (zh) * 2019-05-13 2019-08-06 上海东软医疗科技有限公司 一种基于神经网络的平面回波成像方法和装置
CN110244246A (zh) * 2019-07-03 2019-09-17 上海联影医疗科技有限公司 磁共振成像方法、装置、计算机设备和存储介质
CN110346743A (zh) * 2019-07-22 2019-10-18 上海东软医疗科技有限公司 一种磁共振弥散加权成像方法和装置

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113808017A (zh) * 2021-08-24 2021-12-17 北京理工大学 基于照明优化的结构光成像方法及装置
CN115494439A (zh) * 2022-11-08 2022-12-20 中遥天地(北京)信息技术有限公司 一种基于深度学习的时空编码图像校正方法
CN117011409A (zh) * 2023-08-10 2023-11-07 厦门大学 多部位物理智能高清扩散磁共振数据生成方法
CN117011409B (zh) * 2023-08-10 2024-05-10 厦门大学 多部位物理智能高清扩散磁共振数据生成方法

Also Published As

Publication number Publication date
CN110895320A (zh) 2020-03-20
CN110895320B (zh) 2021-12-24

Similar Documents

Publication Publication Date Title
WO2021082103A1 (zh) 基于深度学习的无变形单次激发平面回波成像方法及装置
US9229082B2 (en) Method for diffusion magnetic resonance imaging
DE102008007048B4 (de) Dynamische Verzeichnungskorrektur bei EPI-Messungen in der medizinischen Magnet-Resonanz-Bildgebung
JP5127841B2 (ja) 磁気共鳴イメージング装置及び磁化率強調画像撮影方法
US10262385B2 (en) Method and apparatus for acquiring magnetic resonance data and entering the data into k-space
CN105232045A (zh) 基于双回波的单扫描定量磁共振扩散成像方法
WO2022236850A1 (zh) 磁共振图像的重建方法、计算机设备及存储介质
WO2020037814A1 (zh) 基于多板块同时激发的等体素磁共振扩散成像方法及装置
WO2022236851A1 (zh) 卷褶视野磁共振图像的重建方法、计算机设备及存储介质
WO2022213666A1 (zh) 结合k空间和图像空间重建的成像方法和装置
US9880247B2 (en) System and method for magnetic resonance imaging using highly accelerated projection imaging
Rieseberg et al. Diffusion tensor imaging using partial Fourier STEAM MRI with projection onto convex subsets reconstruction
WO2016021603A1 (ja) 磁気共鳴イメージング装置および磁気共鳴イメージング方法
WO2022236823A1 (zh) 卷褶视野磁共振图像的重建方法、计算机设备及存储介质
JP5278914B2 (ja) 磁気共鳴イメージング装置及び磁化率強調画像撮影方法
JP2008154990A (ja) 非線形位相補正方法、記録媒体および磁気共鳴装置
CN117495992A (zh) 一种欠采样核磁共振图像的重建方法
CN117233674A (zh) 基于单激发buda采集和联合约束重建的mr成像方法及系统
WO2019181006A1 (ja) 磁気共鳴撮影装置、感度分布算出方法、及び、感度分布算出プログラム
US11119173B2 (en) Dynamic imaging based on echo planar imaging sequence
Matsuo et al. Feasibility study of super-resolution deep learning-based reconstruction using k-space data in brain diffusion-weighted images
Xue et al. Imaging transformer for MRI denoising with the SNR unit training: enabling generalization across field-strengths, imaging contrasts, and anatomy
CN113920211A (zh) 一种基于深度学习的快速磁敏感加权成像方法
EP4189416A1 (en) Method and device for acquiring and reconstructing a sequence of diffusion-weighted magnetic resonance images covering a volume
Dang et al. Joint MR sequence optimization beats pure neural network approaches for spin-echo MRI super-resolution

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19951020

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19951020

Country of ref document: EP

Kind code of ref document: A1