Deformation-free single-shot planar echo imaging method and device based on deep learning
阅读说明：本技术 基于深度学***面回波成像方法及装置 (Deformation-free single-shot planar echo imaging method and device based on deep learning ) 是由 郭华 胡张选 于 2019-10-31 设计创作，主要内容包括：本发明公开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；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；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。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。(The invention discloses a high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging method and device based on deep learning, wherein the method comprises the following steps: acquiring a first image and a related auxiliary image of a single-time excitation plane echo, and acquiring a second image which is acquired in a multi-time excitation mode and meets a preset condition; performing deep neural network training according to the first image, the related auxiliary image and the second image to obtain network weight parameters so as to generate a deep neural network; and receiving a third image and a related auxiliary image of the single-shot plane echo, and inputting the third image and the related auxiliary image into the deep neural network to generate an imaging result. The method can realize the acquisition of the high-resolution high-signal-to-noise-ratio deformation-free magnetic resonance image under the quick scanning of single excitation, and effectively solves the problems of serious signal-to-noise ratio, serious deformation artifact and low resolution in the existing single-excitation EPI technology and the problem of overlong acquisition time in the multiple-excitation technology.)
1. A high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging method based on deep learning is characterized by comprising the following steps of:
acquiring a first image and a related auxiliary image of a single-time excitation plane echo, and acquiring a second image which is acquired in a multi-time excitation mode and meets a preset condition;
performing deep neural network training according to the first image, the related auxiliary image and the second image to obtain network weight parameters so as to generate a deep neural network; and
and receiving a third image and a related auxiliary image of the single-shot plane echo, and inputting the third image and the related auxiliary image into the deep neural network to generate an imaging result.
2. The method of claim 1, wherein the loss function used in the training process is as follows:
wherein u is an image predicted and generated by the network, and u is*For a high quality image to be actually acquired,for SSIM loss, | u-u*|1Is a loss of the l1 norm,
3. The method of claim 1, wherein the associated auxiliary images are a T2 weighted structure, a T2-FLAIR, a T1 weighted structure, the first and third images are single shot EPI acquisitions of a single phase encoding direction, and the first and third images are diffusion magnetic resonance images, the diffusion magnetic resonance images including images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
4. The method of claim 3, wherein inputting the third image and the associated auxiliary image to the deep neural network generates an imaging result comprising:
block-inputting the third image of a single shot EPI and a corresponding auxiliary image;
and combining the outputs into a complete image according to the blocks to obtain the imaging result.
5. A high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging device based on deep learning is characterized by comprising the following components:
the acquisition module is used for acquiring a first image and a related auxiliary image of a single-time excitation plane echo and acquiring a second image which is acquired in a multi-time excitation mode and meets a preset condition;
the training module is used for carrying out deep neural network training according to the first image, the related auxiliary image and the second image to obtain network weight parameters so as to generate a deep neural network; and
and the imaging module is used for receiving a third image and a related auxiliary image of the single-shot plane echo, inputting the third image and the related auxiliary image into the deep neural network and generating an imaging result.
6. The apparatus of claim 5, wherein the loss function used in the training process is as follows:
wherein u is an image predicted and generated by the network, and u is*For a high quality image to be actually acquired,
7. The apparatus of claim 5, wherein the associated auxiliary images are a T2 weighted structure, a T2-FLAIR, a T1 weighted structure, the first and third images are single shot EPI acquisitions of a single phase encoding direction, and the first and third images are diffusion magnetic resonance images including images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
8. The apparatus of claim 7, wherein the imaging module is further configured to block the third image and the corresponding auxiliary image of the single shot EPI, and to combine the outputs into a complete image on a block-by-block basis to obtain the imaging result.
Technical Field
The invention relates to the technical field of plane echo imaging, in particular to a high-resolution high-signal-to-noise-ratio deformation-free single-shot plane echo imaging method and device based on deep learning.
Background
The rapid acquisition characteristic of EPI (Echo Planar Imaging, Planar Echo acquisition technology) enables the EPI to have the advantages of high Imaging speed, insensitivity to motion and the like, and the EPI can be widely applied in clinic, especially single-time excitation EPIThe acquisition of the whole k-space is completed after the F excitation, and the method has very important value in the application with high requirement on imaging speed, such as diffusion imaging, functional imaging, perfusion imaging, cardiac imaging, real-time imaging and the like. However, EPI acquisition also has its own deficiencies, and longer readout times can be introduced
The combination of single shot EPI with parallel acquisition techniques may reduce the length of the readout window and the ESP (effective echo spacing), reducing the number of echoes
The Point Spread Function (PSF) -EPI collection of the PSF (Point spread function, Point spread function-based encoding) provides an effective mode for solving the problems, the obtained EPI is completely free from deformation and image blurring caused by T2 attenuation, and meanwhile, the collection acceleration of the tipped-CAIPI technology greatly improves the time efficiency of the PSF-EPI. Another type of deformation-free fast imaging method in magnetic resonance imaging includes Fast Spin Echo (FSE) and fast gradient echo (FFE), which acquire multiple spin echoes or gradient echo signals after one excitation, i.e., acquire multiple encoding positions by one excitation, thereby achieving the purpose of accelerating acquisition. However, the above mentioned deformation-free imaging techniques are all based on multiple excitation, which greatly prolongs the acquisition time, and is often difficult to apply in clinical acquisition with high time efficiency requirements, such as diffusion magnetic resonance imaging, functional magnetic resonance imaging, etc.
Disclosure of Invention
The present application is based on the recognition and discovery by the inventors of the following problems:
with the rise of big data analysis and the progress of large-scale computing power, deep learning has been greatly developed and plays a great role in various research fields. The same is true in the field of magnetic resonance imaging, which relates to various aspects of magnetic resonance imaging, including but not limited to image acquisition, reconstruction, recovery, quantitative analysis, high resolution reconstruction, and the like. The relevant study generates a 7T acquired image, for example from a magnetic resonance image acquired at a 3T magnetic field strength; a high resolution T1 weighted image is generated from a low resolution T1 weighted image, and so on. These studies provide new ideas for improving image quality (e.g., improving image resolution, signal-to-noise ratio, reducing image artifacts, etc.) without introducing other problems (e.g., acquisition time enhancement, signal-to-noise ratio reduction, etc.).
The present invention is directed to solving, at least to some extent, one of the technical problems in the related art.
Therefore, one objective of the present invention is to provide a single shot planar echo imaging method with high resolution, high signal-to-noise ratio and no deformation based on deep learning, which can achieve the purpose of obtaining a high-resolution, high-signal-to-noise ratio and no deformation magnetic resonance image under the fast scanning of single shot, and effectively solve the problems of serious signal-to-noise ratio, serious deformation artifact and low resolution existing in the existing single shot EPI technology and the problem of overlong acquisition time existing in the multiple shot technology.
Another objective of the present invention is to provide a single shot planar echo imaging device with high resolution, high signal-to-noise ratio and no deformation based on deep learning.
In order to achieve the above object, an embodiment of the present invention provides a high-resolution, high-signal-to-noise ratio and deformation-free single-shot planar echo imaging method based on deep learning, which includes the following steps: acquiring a first image and a related auxiliary image of a single-time excitation plane echo, and acquiring a second image which is acquired in a multi-time excitation mode and meets a preset condition; performing deep neural network training according to the first image, the related auxiliary image and the second image to obtain network weight parameters so as to generate a deep neural network; and receiving a third image and a related auxiliary image of the single-shot plane echo, and inputting the third image and the related auxiliary image into the deep neural network to generate an imaging result.
The high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging method based on the deep learning combines the deep learning with single-shot EPI acquisition in the single-phase coding direction, and uses a deformation-free EPI image acquired in a multi-shot mode as a standard to perform network learning, so that a high-resolution high-signal-to-noise-ratio deformation-free high-quality image can be acquired only by using the acquisition time of the single-shot EPI in the single-phase coding direction, and meanwhile, the quality of an output image is improved by combining with an auxiliary image strategy.
In addition, the high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging method based on the deep learning according to the above embodiment of the present invention may further have the following additional technical features:
further, in one embodiment of the present invention, the loss function used in the training process is as follows:
wherein u is an image predicted and generated by the network, and u is*For a high quality image to be actually acquired,for SSIM loss, | u-u*|1Is a loss of the l1 norm,
Further, in an embodiment of the invention, the associated auxiliary images are T2 weighted configurations, T2-FLAIR, T1 weighted configurations, etc., the first and third images are acquired with single shot EPI in a single phase encoding direction, and the first and third images are diffusion magnetic resonance images including images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
Further, in an embodiment of the present invention, the inputting the third image and the related auxiliary image into the deep neural network to generate an imaging result includes: block-inputting the third image of a single shot EPI and a corresponding auxiliary image; and combining the outputs into a complete image according to the blocks to obtain the imaging result.
In order to achieve the above object, another embodiment of the present invention provides a high-resolution high-snr single shot echo planar imaging device without distortion based on deep learning, including: the acquisition module is used for acquiring a first image and a related auxiliary image of a single-time excitation plane echo and acquiring a second image which is acquired in a multi-time excitation mode and meets a preset condition; the training module is used for carrying out deep neural network training according to the first image, the related auxiliary image and the second image to obtain network weight parameters so as to generate a deep neural network; and the imaging module is used for receiving a third image and a related auxiliary image of the single-shot plane echo, inputting the third image and the related auxiliary image into the deep neural network and generating an imaging result.
The high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging device based on the deep learning combines the deep learning with single-shot EPI acquisition in the single-phase coding direction, and performs network learning by using a deformation-free EPI image acquired in a multi-shot mode as a standard, so that a high-resolution high-signal-to-noise-ratio deformation-free high-quality image can be acquired only by using the acquisition time of the single-shot EPI in the single-phase coding direction, and meanwhile, the quality of an output image is improved by combining with an auxiliary image strategy.
In addition, the high-resolution high-signal-to-noise ratio deformation-free single-shot planar echo imaging device based on the deep learning according to the above embodiment of the present invention may further have the following additional technical features:
further, in one embodiment of the present invention, the loss function used in the training process is as follows:
wherein u is an image predicted and generated by the network, and u is*For a high quality image to be actually acquired,
Further, in an embodiment of the invention, the associated auxiliary images are T2 weighted configurations, T2-FLAIR, T1 weighted configurations, etc., the first and third images are acquired with single shot EPI in a single phase encoding direction, and the first and third images are diffusion magnetic resonance images including images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
Further, in an embodiment of the present invention, the imaging module is further configured to block the third image of the single shot EPI and the corresponding auxiliary image, and combine the outputs into a complete image by block to obtain the imaging result.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The foregoing and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a flowchart of a high resolution high SNR deformation-free single shot echo planar imaging method based on deep learning according to an embodiment of the present invention;
FIG. 2 is a flowchart of a method for high resolution, high signal-to-noise ratio, deformation-free single shot echo planar imaging based on deep learning according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a deep neural network according to an embodiment of the present invention;
fig. 4 is a schematic structural diagram of a high-resolution high-snr deformation-free single-shot planar echo imaging device based on deep learning according to an embodiment of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.
The method and the device for high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging based on deep learning according to the embodiment of the invention are described below with reference to the accompanying drawings.
FIG. 1 is a flowchart of a method for high resolution, high signal-to-noise ratio, deformation-free single shot planar echo imaging based on deep learning according to an embodiment of the present invention.
As shown in FIG. 1, the high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging method based on the deep learning comprises the following steps:
in step S101, a first image and a related auxiliary image of a single shot echo planar are acquired, and a second image that satisfies a preset condition and is acquired by a multi-shot mode is acquired.
It can be understood that the embodiment of the present invention acquires an image of a single shot EPI and a related auxiliary image, and acquires a high resolution high signal-to-noise ratio deformation-free image in a multi-shot mode. The first image is an image of single-shot EPI, and the second image meeting the preset condition is a high-resolution high-signal-to-noise-ratio deformation-free image.
Further, in an embodiment of the present invention, the associated auxiliary images are T2 weighted structure, T2-FLAIR, T1 weighted structure, etc., the first image and the third image are acquired by single shot EPI in a single phase encoding direction, and the first image and the third image are diffusion magnetic resonance images, which include images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
Specifically, the single-shot EPI image used in the embodiment of the present invention is diffusion magnetic resonance imaging, and includes an image without a diffusion encoding gradient (b0) and an image with 6 different diffusion encoding gradient directions (b1), and the output is a high-resolution high-snr (signal to noise ratio) deformation-free image acquired by using the same diffusion encoding strategy and using a multi-shot mode (acquired by using a PSF-EPI mode in the embodiment of the present invention). Meanwhile, in a conventional scan, images with various contrasts, such as T2 weighted structure images and the like, are generally scanned besides diffusion magnetic resonance images, and in order to fully utilize the data, the embodiment of the present invention proposes a strategy of using auxiliary images, that is, the input of the deep neural network includes not only single-shot EPI images but also acquired auxiliary images, such as T2 weighted structure images, T2-FLAIR, T1 weighted structure images and the like (as shown in fig. 2), to form a composite multi-channel image. By using the strategy, the learning effect of the neural network can be greatly improved, and the quality of an output image is improved.
In step S102, a deep neural network training is performed according to the first image, the related auxiliary image, and the second image to obtain a network weight parameter, so as to generate a deep neural network.
It is understood that the training phase: and carrying out deep neural network training by using the image acquired by single-excitation EPI, the related auxiliary image and the high-resolution high-signal-to-noise-ratio deformation-free image acquired by a multi-excitation mode to obtain network weight parameters.
Specifically, as shown in fig. 3, the deep neural network structure used in the embodiment of the present invention includes 8 channels (b0, 6 b1 and auxiliary T2 weighted structural images for single shot EPI acquisition) in the input image, and 7 channels (b0, 6 b1 for PSF-EPI acquisition). The network uses a commonly used U-net structure, where scaling of features is achieved by convolutional layers with step size and deconvolution layers. Each convolutional layer contains a BN layer (batch normalization) and an active layer (ReLU).
In the training process, the image is subjected to block training along the phase encoding direction, the input matrix size is 217 × 32 × 8, the output matrix size is 224 × 32 × 8, and the convolution kernel size and the convolution kernel number of each layer are not listed in detail here. The loss function used during training is as follows:
wherein u is an image predicted and generated by the network, and u is*For a high quality image to be actually acquired,
The embodiment of the invention does not limit the imaging contrast, and besides the embodiment of the invention is applied to diffusion magnetic resonance imaging acquired by single-shot EPI, the embodiment of the invention can also be applied to T1 weighting, T2 weighting, T2 × weighting, Proton Density (PD) weighting and the like acquired by using the method; in the embodiment of the invention, the EPI image in the single-phase encoding direction is used, and the EPI image in the two-direction phase encoding direction can also be input at the same time, which is not limited by the invention; the acquisition mode of the high-resolution high-signal-to-noise-ratio deformation-free image is not limited to point spread function encoded EPI (PSF-EPI), and a high-resolution deformation-free image can be obtained by using multi-shot spin echo acquisition (FSE), gradient echo acquisition (FFE) and the like, or using an image deformation correction algorithm (such as fieldmap correction, topup, or a combination of the two); the diffusion preparation sequence for diffusion imaging used in the embodiment of the present invention is PGSE (pulsed gradient spin echo), and an STE (stimulated echo) diffusion preparation sequence, an Oscillating Gradient Spin Echo (OGSE) diffusion preparation sequence, a Double Diffusion Encoding (DDE) diffusion preparation sequence, a Convex Optimized Diffusion Encoding (CODE) diffusion preparation sequence, and the like may also be used; the embodiment of the invention does not limit the number of network layers, the number of convolution kernels, the size of the convolution kernels, the activation mode, the regularization mode, the activation mode, the loss function and the like used by the U-net; the invention does not limit the optimizers (such as Adam, SGD and the like) used in the training process, various parameters (such as learning rate, batch-size and the like) and the like; the network structure used in the present invention is not limited to U-net, and ResNet (reactive networks), GAN (generic adaptive networks), and the like and their variants can be used; the invention does not limit the matrix size of the input image and the output image, and depends on the resolution of the collected image, the image mode, the size of the blocks and the like; the selected auxiliary image is not limited by the embodiment of the invention, and a T2 weighted structure, a T1 weighted structure, a T2-FLAIR and the like or a combination thereof can be used.
In step S103, the third image and the related auxiliary image of the single shot echo planar are received, and the third image and the related auxiliary image are input to the deep neural network, so as to generate an imaging result.
It will be appreciated that, during the testing (application) phase: and inputting the image acquired by single-shot excitation of the EPI and the related auxiliary image to a deep neural network obtained in a training stage to generate a corresponding high-resolution high-signal-to-noise-ratio deformation-free high-quality image.
Further, in an embodiment of the present invention, the third image and the related auxiliary image are input to a deep neural network, and the imaging result is generated, including: inputting a third image of the single shot EPI and a corresponding auxiliary image in a blocking mode; and combining the outputs into a complete image according to the blocks to obtain an imaging result.
It will be appreciated that after the network weight parameters are obtained through the training phase, similar to the training process (as shown in fig. 2), the single shot EPI images and the corresponding auxiliary images are input in blocks, and then the outputs are combined into a complete image in blocks, i.e., a corresponding high resolution high signal-to-noise ratio deformation-free image can be obtained.
In summary, the embodiment of the present invention combines the deep learning with the single-shot EPI acquisition technique, and learns the image acquired by the single-shot EPI acquisition to generate the high-resolution, high-snr, and deformation-free image acquired by the multi-shot EPI acquisition technique, thereby achieving the purpose of rapidly acquiring the high-quality magnetic resonance image by using only the time required for the single-shot EPI acquisition.
According to the high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging method based on the deep learning, the deep learning is combined with the single-shot EPI acquisition in the single-phase coding direction, the deformation-free EPI image acquired in a multi-shot mode is used as a standard for network learning, so that the high-resolution high-signal-to-noise-ratio deformation-free high-quality image can be acquired only by using the acquisition time of the single-shot EPI in the single-phase coding direction, and meanwhile, the quality of an output image is improved by combining with an auxiliary image strategy.
The invention further provides a high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging device based on deep learning, which is provided by the embodiment of the invention and is described with reference to the accompanying drawings.
Fig. 4 is a schematic structural diagram of a high-resolution high-snr deformation-free single-shot planar echo imaging device based on deep learning according to an embodiment of the present invention.
As shown in fig. 4, the depth learning-based high-resolution high-snr deformation-free single-shot planar
The acquiring
Further, in one embodiment of the present invention, the loss function used in the training process is as follows:
wherein u is an image predicted and generated by the network, and u is*For a high quality image to be actually acquired,
Further, in an embodiment of the present invention, the related auxiliary images are T2 weighted structure, T2-FLAIR, T1 weighted structure, etc., the first image and the third image are acquired by single shot EPI in a single phase encoding direction, and the first image and the third image are diffusion magnetic resonance images, which include images without diffusion encoding gradients and images with 6 different diffusion encoding gradient directions.
Further, in an embodiment of the present invention, the
It should be noted that the foregoing explanation of the embodiment of the high-resolution high-snr non-deformation single-shot planar echo imaging method based on deep learning is also applicable to the high-resolution high-snr non-deformation single-shot planar echo imaging device based on deep learning of this embodiment, and details are not repeated here.
According to the depth learning-based high-resolution high-signal-to-noise-ratio deformation-free single-shot planar echo imaging device provided by the embodiment of the invention, the depth learning is combined with the single-shot EPI acquisition in the single-phase coding direction, and the deformation-free EPI image acquired by a multi-shot mode is used as a standard for network learning, so that the high-resolution high-signal-to-noise-ratio deformation-free high-quality image can be acquired only by using the acquisition time of the single-shot EPI in the single-phase coding direction, and meanwhile, the quality of the output image is improved by combining with the strategy of an auxiliary image.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.
- 上一篇：一种医用注射器针头装配设备
- 下一篇：一种基于录波文件基准通道的二次设备时标对齐方法