WO2021168745A1 - 一种磁共振成像模型的训练方法及装置 - Google Patents

一种磁共振成像模型的训练方法及装置 Download PDF

Info

Publication number
WO2021168745A1
WO2021168745A1 PCT/CN2020/077011 CN2020077011W WO2021168745A1 WO 2021168745 A1 WO2021168745 A1 WO 2021168745A1 CN 2020077011 W CN2020077011 W CN 2020077011W WO 2021168745 A1 WO2021168745 A1 WO 2021168745A1
Authority
WO
WIPO (PCT)
Prior art keywords
magnetic resonance
resonance image
simulated
sampled
under
Prior art date
Application number
PCT/CN2020/077011
Other languages
English (en)
French (fr)
Inventor
郑海荣
刘新
张娜
胡战利
陈其航
梁栋
杨永峰
Original Assignee
深圳先进技术研究院
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 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Publication of WO2021168745A1 publication Critical patent/WO2021168745A1/zh
Priority to US17/889,189 priority Critical patent/US20230047647A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2210/00—Indexing scheme for image generation or computer graphics
    • G06T2210/41—Medical
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

  • the present invention claims the priority of a Chinese patent application filed with the Chinese Patent Office on February 24, 2020, the application number is 202010113026.2, and the invention title is "a training method and device for a magnetic resonance imaging model". The entire content of the application is approved The citation is incorporated in the present invention.
  • This application relates to the field of computer technology, and in particular to a training method, device, electronic equipment, and computer-readable storage medium for a magnetic resonance imaging model.
  • Magnetic Resonance Imaging (MRI) technology is a technology that uses MRI to reconstruct images of human tissues. It can provide rich image information of human tissues without ionizing damage to the human body. It is a widely used clinical medical examination method in my country.
  • the current way to shorten the image data acquisition time of MRI is mainly to reduce the amount of image data acquisition, such as regular under-sampling based on partial k-space, random under-sampling based on compressed sensing (Compressed Sensing, CS) theory, and non-Cartesian based Radial and Spiral under-sampling of sampled trajectories, etc.
  • the reduction in the amount of image data collection will inevitably reduce the clarity of the image.
  • the deep learning method used mainly learns the mapping relationship from under-sampled image data to full-sampled image data to achieve imaging, but this learning method is not efficient in learning and the quality of the generated image is relatively low.
  • the embodiments of this specification provide a training method, device, electronic device, and computer-readable storage medium for a magnetic resonance imaging model, so as to solve the low learning efficiency of the deep learning method in the prior art and the relatively low quality of the generated image The problem.
  • a training method of a magnetic resonance imaging model includes:
  • the magnetic resonance image data set includes: an under-sampled magnetic resonance image and a full-sampled magnetic resonance image;
  • a ring-shaped deep neural network to be trained; wherein the ring-shaped deep neural network includes neural networks on both sides and two input ports; the two input ports are used to input the under-sampled magnetic resonance image and the full-sampling Magnetic resonance image
  • One side of the neural network of the two sides of the neural network is used to generate a first simulated full-sampling magnetic resonance image according to the under-sampled magnetic resonance image, and a first simulation generated according to the other side of the neural network of the two sides of the neural network
  • the under-sampled magnetic resonance image is used to generate a second simulated full-sampled magnetic resonance image
  • the other side neural network is used to generate the first simulated under-sampled magnetic resonance image according to the full-sampled magnetic resonance image, and according to the one side nerve
  • the first simulated fully-sampled magnetic resonance image generated by the network, and the second simulated under-sampled magnetic resonance image is generated;
  • each simulated magnetic resonance image includes: the first simulated under-sampled magnetic resonance image, the first simulated full-sampled magnetic resonance image, and the second simulated under-sampled magnetic resonance image An image and the second simulated full-sampling magnetic resonance image;
  • the first simulated full-sampling magnetic resonance image and the full-sampling magnetic resonance image are input to a pre-built first simulated magnetic resonance image classification model to obtain whether the first simulated full-sampling magnetic resonance image is The first discrimination result of the simulated magnetic resonance image category;
  • the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image are input to a pre-built second simulated magnetic resonance image category discrimination model to obtain whether the first simulated under-sampled magnetic resonance image is The second discrimination result of the simulated magnetic resonance image category;
  • the network parameters of the circular deep neural network are adjusted to obtain a trained magnetic resonance imaging model.
  • a magnetic resonance image generation method based on the training method of the magnetic resonance imaging model, the magnetic resonance image generation method includes:
  • the under-sampled image data is input into the trained magnetic resonance imaging model to generate a magnetic resonance image.
  • a training device for a magnetic resonance imaging model including:
  • the acquisition module is used to acquire a magnetic resonance image data set; wherein the magnetic resonance image data set includes: an under-sampled magnetic resonance image and a full-sampled magnetic resonance image;
  • the construction module is used to construct a ring-shaped deep neural network to be trained; wherein the ring-shaped deep neural network includes two sides of the neural network and two input ports; the two input ports are used to input the under-sampled magnetic resonance images respectively And the full-sampled magnetic resonance image; one side of the neural network on both sides is used to: generate a first simulated full-sampled magnetic resonance image according to the under-sampled magnetic resonance image, and according to the other side of the neural network on both sides The first simulated under-sampled magnetic resonance image generated by the neural network on one side is used to generate the second simulated full-sampled magnetic resonance image; the other side neural network is used to: generate the first simulated under-sampled magnetic resonance image according to the full-sampled magnetic resonance image A resonance image, generating a second simulated under-sampled magnetic resonance image based on the first simulated fully sampled magnetic resonance image generated by the one-side neural network;
  • the generation module is used to input the under-sampled magnetic resonance image and the full-sampled magnetic resonance image into the to-be-trained circular deep neural network through two input ports of the circular-shaped deep neural network.
  • the first discrimination module is configured to input the first simulated full-sampled magnetic resonance image and the full-sampled magnetic resonance image into a pre-built first simulated magnetic resonance image category discrimination model to obtain a comparison of the first simulated Whether the fully sampled magnetic resonance image is the first judgment result of the simulated magnetic resonance image category;
  • the second discrimination module is used to input the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image into a pre-built second simulated magnetic resonance image category discrimination model to obtain a comparison of the first simulated Whether the under-sampled magnetic resonance image is the second judgment result of the simulated magnetic resonance image category;
  • the parameter adjustment module is configured to perform according to a preset loss function, the first discrimination result, the second discrimination result, the second simulated under-sampled magnetic resonance image, the second simulated full-sampled magnetic resonance image, and the The under-sampled magnetic resonance image and the full-sampled magnetic resonance image are adjusted, and the network parameters of the circular deep neural network are adjusted to obtain a trained magnetic resonance imaging model.
  • a magnetic resonance image generation device based on the training device of the magnetic resonance imaging model, the magnetic resonance image generation device includes:
  • Image data acquisition module for acquiring under-sampled image data
  • the generating module is used to input the under-sampled image data into the trained magnetic resonance imaging model to generate a magnetic resonance image.
  • An electronic device comprising: a memory, a processor, and a computer program stored in the memory and running on the processor, the computer program being executed by the processor to implement any one of the magnetic resonance imaging The steps of the model training method.
  • a computer-readable storage medium having a computer program stored on the computer-readable storage medium, and when the computer program is executed by a processor, realizes the steps of any one of the training methods of the magnetic resonance imaging model.
  • mapping relationship of a single image generation direction for example, an under-sampled image to a full-sampled image
  • the ring-shaped deep neural network constructed in this application during the training process, it not only learns the mapping relationship from the under-sampled magnetic resonance image to the fully-sampled magnetic resonance image, the image generation direction, but also increases the direction opposite to the image generation direction.
  • the neural network on the other side can also learn the mapping relationship from the full-sampled magnetic resonance image to the under-sampled magnetic resonance image in the opposite direction, so that the mapping relationship learned by the neural network on one side can be corrected, so that the neural network on the other side can work as expected.
  • the image generation direction can form a correct mapping, thereby reducing the deviation between the generated MRI image and the actual MRI image, improving the quality of the MRI image generated by the magnetic resonance imaging model, and improving the learning ability and learning efficiency of the neural network.
  • the generated first simulated magnetic resonance image and the first simulated under-sampled magnetic resonance image are also discriminated, and The discrimination result is fed back to the training of the ring deep neural network through the loss function. It is hoped to achieve the purpose of making the simulated magnetic resonance image discrimination model misjudge the simulated magnetic resonance image generated by the circular deep neural network as a non-simulated magnetic resonance image, so that the MRI image generated by the trained magnetic resonance imaging model is as close to the actual MRI as possible Image, thereby further improving the quality of the MRI image generated by the MRI model.
  • FIG. 1 is a schematic flowchart of a training method for a magnetic resonance imaging model provided by an embodiment of this specification
  • FIG. 2 is a schematic diagram of the structure of a ring deep neural network provided by an embodiment of the specification
  • FIG. 3 is a schematic diagram of the structure of a residual block provided by an embodiment of the specification.
  • FIG. 4 is a schematic diagram of the structure of a side neural network provided by an embodiment of the specification.
  • FIG. 5 is a schematic diagram of the structure of the other side of the neural network provided by the embodiment of this specification.
  • Fig. 6 is a schematic structural diagram of a simulated magnetic resonance image classification model provided by an embodiment of the specification.
  • FIG. 7 is a schematic flowchart of a method for generating a magnetic resonance image according to an embodiment of the specification.
  • FIG. 8 is a schematic structural diagram of a training device for a magnetic resonance imaging model provided by an embodiment of this specification.
  • FIG. 9 is a schematic structural diagram of a magnetic resonance image generating device provided by an embodiment of this specification.
  • Fig. 10 is a schematic structural diagram of an electronic device provided by an embodiment of the specification.
  • the adopted deep learning method mainly learns the mapping relationship from under-sampled image data to full-sampled image data to achieve imaging, but this learning method is not efficient in learning, and the quality of the generated magnetic resonance image is relatively low.
  • the embodiments of this specification provide a training method of a magnetic resonance imaging model, which is used to improve the learning efficiency of the neural network and the quality of the generated magnetic resonance image.
  • the execution subject of the method includes, but is not limited to, servers, personal computers, notebook computers, tablet computers, smart phones, and other intelligent electronic devices that can execute predetermined processing procedures such as numerical calculations and/and logic calculations by running predetermined programs and instructions.
  • the server may be a server group composed of a single web server and multiple web servers, or a cloud composed of a large number of computers and web servers based on Cloud Computing.
  • the execution subject of this method is not limited.
  • the schematic flow chart of this method is shown in Figure 1, including the following steps:
  • Step 11 Obtain a magnetic resonance image data set.
  • the magnetic resonance image data set here may include: an under-sampled magnetic resonance image and a full-sampled magnetic resonance image.
  • the full-sampling magnetic resonance image can come from the magnetic resonance image data actually collected by the magnetic resonance imaging device.
  • the under-sampled magnetic resonance image may be magnetic resonance image data formed by extracting partial sampling point data from a full-sampled magnetic resonance image.
  • the extraction of part of the sampling point data may be based on the rule of K-space, or may be randomly selected based on the compressed sensing theory, which is not limited in this application.
  • the full-sampled magnetic resonance image after extracting part of the sampling point data from the full-sampled magnetic resonance image to generate an under-sampled magnetic resonance image, it may also include establishing a matching relationship between the under-sampled magnetic resonance image and the full-sampled magnetic resonance image, so that the magnetic The full-sampled magnetic resonance image and the under-sampled magnetic resonance image in the resonance image data set are matched correspondingly, which facilitates the training of the neural network later.
  • Step 12 Construct a ring-shaped deep neural network to be trained.
  • the deep learning method used in the prior art only learns the mapping relationship of a single image generation direction to achieve imaging, resulting in relatively low quality of generated magnetic resonance images and low learning efficiency of neural networks
  • the constructed circular deep neural network can include two sides of the neural network and two input ports.
  • the two input ports can be used to respectively input an under-sampled magnetic resonance image and a full-sampled magnetic resonance image, where the under-sampled magnetic resonance image and the full-sampled magnetic resonance image may be obtained through step 11.
  • one side of the ring-shaped deep neural network may include: a first down-sampling layer, a first residual network layer , The first upsampling layer; the other side of the neural network may include: a second sampling layer, a second residual network layer, and a second upsampling layer.
  • the first down-sampling layer, the first up-sampling layer, the second up-sampling layer, and the second down-sampling layer may each include at least one convolutional layer, and each convolutional layer in the at least one convolutional layer adopts serial Connection, that is, the output image of the previous convolutional layer can be used as the input image of the next convolutional layer.
  • Each convolutional layer can perform convolution processing on the input image, and each convolutional layer can include at least one convolution kernel, and each convolution kernel is used to indicate a weight matrix for a convolution operation.
  • the above-mentioned input map and output map may both refer to a feature map (feature map).
  • the first residual network layer and the second residual network layer may each include at least one residual block.
  • the schematic diagram of the structure of each residual block may be as shown in FIG. 3, and each residual block may specifically be formed by using two layers of convolutional layers plus jump connections. Among them, the two-layer convolutional layer can be used to extract deep-level image features, and the jump connection is used to pass low-level image features directly backwards. Combining with deep-level image features can improve the learning ability and stability of the neural network. , To avoid the deterioration of the training effect of the deeper neural network due to the increase in depth.
  • Fig. 4 is a schematic diagram of a specific structure of a side neural network of an exemplary deep circular deep neural network given in the embodiment of this specification.
  • the first down-sampling layer and the first up-sampling layer respectively include 3 layers of convolutional layers
  • the first residual network layer may include 9 residual blocks.
  • the number of convolution kernels and the size of the convolution kernels of each convolution layer are also shown.
  • conv, 64 means that there are 64 convolution kernels in this convolution layer
  • 3*3 means that the convolution kernel uses a small convolution kernel with a size of 3*3
  • the number of convolution kernels and the size of the convolution kernels of the remaining convolution layers will not be repeated.
  • the first convolutional layer of the first down-sampling layer can be used to perform convolution processing on the pixel values of the input magnetic resonance image (for example, perform convolution processing on the pixel values of the under-sampled magnetic resonance image) to obtain the
  • the output feature map of the first layer of convolutional layer the output feature map is used as the input image of the next second layer of convolutional layer, and so on, the output image of the second layer of convolutional layer is used as the third layer of convolutional layer Input map.
  • the number of feature maps output by each convolutional layer can be the same as the number of convolution kernels of the convolutional layer, for example, the features output by the first convolutional layer of the first downsampling layer shown in Figure 4
  • the number of maps is 64
  • the number of feature maps output by the second convolutional layer is 12
  • the number of feature maps output by the third convolutional layer is 256.
  • Each feature map output by the first down-sampling layer can represent the local features of the input under-sampled magnetic resonance image, and then through the first residual network layer, the feature map output by the first down-sampling layer can be residual Processing can be understood as performing residual calculation on each feature map representing local features layer by layer through each residual block of the first residual network to further extract deeper image features, making the extracted local features more significant , Improve the learning ability of neural network.
  • the two-layer convolutional layer of the first upsampling layer can merge the feature maps layer by layer after the residual processing of the first residual network layer. For example, 256 feature maps are merged into 128 feature maps, 128 feature maps Merge into 64 feature maps, and correspondingly expand the size of feature maps layer by layer. Then connect the feature maps through the fully connected layer, and output the generated simulated magnetic resonance image. For example, the 64 feature maps combined by the two convolutional layers are input to the fully connected layer to connect the 64 feature maps, and output The first simulated fully sampled magnetic resonance image generated.
  • the network structure of the neural network on the other side of the ring deep neural network can be the same as the network structure of the neural network on the other side.
  • the related description of, I won’t repeat it here.
  • the above-mentioned exemplary circular deep neural network is a specific implementation manner for constructing a circular deep neural network provided by the embodiment of this specification, and does not represent all implementation manners of the embodiment of this specification.
  • the number of convolutional layers, the number of residual blocks, the number of convolution kernels, and the size of the convolution kernel included in the circular deep neural network can be set according to actual needs, which is not limited in this application.
  • Step 13 Input the under-sampled magnetic resonance image and the full-sampled magnetic resonance image into the two sides of the neural network included in the circular deep neural network to be trained through the two input ports of the circular deep neural network to be trained. Generate each simulated magnetic resonance image.
  • the under-sampled magnetic resonance image and the full-sampled magnetic resonance image may be obtained through step 11.
  • the circular deep neural network to be trained here can be constructed through step 12.
  • Each simulated magnetic resonance image here may include: a first simulated under-sampled magnetic resonance image, a first simulated fully-sampled magnetic resonance image, a second simulated under-sampled magnetic resonance image, and a second simulated fully-sampled magnetic resonance image.
  • under-sampled magnetic resonance image input the under-sampled magnetic resonance image from input port 1 to one side of the neural network, and output the first simulated full-sampled magnetic resonance image, and the first simulated full-sampled magnetic resonance image is then input ⁇ 2 is input to the neural network on the other side to generate a second simulated under-sampled magnetic resonance image;
  • the full-sampling magnetic resonance image is input from input port 2 to the other side of the neural network, and the first simulated under-sampled magnetic resonance image is output, and the first simulated under-sampled magnetic resonance image is passed through
  • the input port 1 is input to a neural network on one side to generate a second simulated full-sampling magnetic resonance image.
  • the functions of the two sides of the neural network included in the ring deep neural network can include: one side of the neural network is used to: generate the first simulated full-sampled magnetic resonance image based on the under-sampled magnetic resonance image, and according to the two sides of the neural network
  • the first simulated under-sampled magnetic resonance image generated by the neural network on the other side of the network is used to generate the second simulated fully-sampled magnetic resonance image;
  • the other side of the neural network is used to: generate the first simulated under-sampled magnetic resonance image from the fully-sampled magnetic resonance image
  • a second simulated under-sampled magnetic resonance image is generated according to the first simulated fully sampled magnetic resonance image generated by the neural network on one side of the neural network on both sides.
  • the process of generating each simulated magnetic resonance image is specifically combined with the network structure of the neural network on both sides shown in FIG. Can include:
  • the under-sampled magnetic resonance image input the under-sampled magnetic resonance image from input port 1 to one side of the neural network, extract the first feature map of the under-sampled magnetic resonance image through the first down-sampling layer, and pass the first residual
  • the difference network performs residual processing on the first feature map to obtain the first feature map after the residual processing, and then through the first upsampling layer, the first simulated full-sampling magnetic resonance is generated according to the first feature map after the residual processing Image; the first simulated full-sampling magnetic resonance image is then input to the other side of the neural network through the input port 2, and the fourth feature map of the first simulated full-sampling magnetic resonance image is extracted through the second down-sampling layer, and the fourth feature map of the first simulated full-sampling magnetic resonance image is extracted through the second
  • the residual network layer performs residual processing on the fourth feature map to obtain the fourth feature map after the residual processing, and then through the second up-sampling layer to generate the second simulated under-sampled magnetic field according
  • full-sampled magnetic resonance images input the full-sampled magnetic resonance image from input port 2 to the other side of the neural network, extract the third feature map of the full-sampled magnetic resonance image through the second down-sampling layer, and pass the second
  • the residual network performs residual processing on the third characteristic map to obtain the third characteristic map after the residual processing, and then through the second upsampling layer, according to the third characteristic map after the residual processing, the first simulated under-sampled magnetic field is generated.
  • the first simulated under-sampled magnetic resonance image is then input to a side neural network through input port 1, and the second feature map of the first simulated under-sampled magnetic resonance image is extracted through the first down-sampling layer, and the second feature map of the first simulated under-sampled magnetic resonance image is extracted through the first
  • the residual network layer performs residual processing on the second feature map to obtain the second feature map after the residual processing, and then through the first up-sampling layer to generate the second simulated fully sampled magnetic field according to the second feature map after the residual processing Resonance image.
  • Step 14 Input the first simulated full-sampling magnetic resonance image and the full-sampling magnetic resonance image into the pre-built first simulated magnetic resonance image category discrimination model to obtain whether the first simulated full-sampling magnetic resonance image is a simulated magnetic resonance The first discrimination result of the image category.
  • the first simulated full-sampling magnetic resonance image here may be generated through step 13.
  • the full-sampling magnetic resonance image here may be obtained through step 11.
  • the first simulated magnetic resonance image category discrimination model constructed in advance can be used for the first simulation. Determine whether the sampled magnetic resonance image is a simulated magnetic resonance image category, and obtain the first discrimination result, so that the first discrimination result can be fed back to the training of the ring deep neural network through the preset loss function, that is, the first simulation The deviation between the full-sampled magnetic resonance image and the full-sampled magnetic resonance image is fed back to the training of the circular deep neural network.
  • the pre-built first simulated magnetic resonance image category discrimination model here can be constructed based on Convolutional Neural Networks (CNN), for example, the one shown in Figure 6 can be a The schematic diagram of the structure of the first simulated magnetic resonance image classification model.
  • CNN Convolutional Neural Networks
  • other types of neural networks can also be used for construction, which is not limited in this application.
  • a simulated full-sampling magnetic resonance image is a simulated magnetic resonance image category.
  • the first simulated magnetic resonance image can be adjusted by the fourth loss function. The model parameters of the resonance image category discrimination model, so that the first simulated magnetic resonance image category discrimination model accurately outputs the first discrimination result.
  • the fourth loss function may be:
  • x represents an under-sampled magnetic resonance image
  • y represents a full-sampled magnetic resonance image
  • G sd (x) represents the first simulated full-sampled magnetic resonance image
  • D d represents the first simulated magnetic resonance image classification model
  • D d (G sd (x)) represents the first discrimination result
  • MSE represents the Mean Square Error (MSE) function
  • E x represents the mathematical expectation of the function when the input is an under-sampled magnetic resonance image
  • D d (y) represents the first simulation The classification result of the magnetic resonance image classification model for the classification of the full-sampling magnetic resonance image
  • E y represents the functional mathematical expectation that the input is a full-sampling magnetic resonance image.
  • Adjusting the model parameters according to the fourth loss function may specifically include: calculating the loss value of the discrimination result of the first simulated magnetic resonance image category discrimination model through the fourth loss function, and feeding back the loss value to the first simulated magnetic resonance image by way of back propagation.
  • Resonance image category discrimination model to adjust model parameters.
  • the first simulated full-sampling magnetic resonance image may be generated by a side neural network of the ring deep neural network If yes, it can be considered that the first simulated magnetic resonance image classification model and the circular deep neural network can be trained at the same time.
  • the training effects of the two are opposite, that is, the training of the circular deep neural network is expected to achieve the purpose of making the first simulated magnetic resonance image category discrimination model discriminate the first simulated full-sampling magnetic resonance image as a non-simulated magnetic resonance image category, while training the first
  • the simulated magnetic resonance image category discrimination model may be expected to improve the accuracy of model discrimination, so as to discriminate as far as possible the first simulated full-sampling magnetic resonance image as the simulated magnetic resonance image category.
  • the accuracy of the discrimination of the first simulated magnetic resonance image category discrimination model is improved, that is, the output is increased
  • the accuracy of the first discrimination result is also equivalent to promoting the improvement of the quality of simulated magnetic resonance images generated by the circular deep neural network.
  • the pre-built first simulated magnetic resonance image category discrimination model it is possible to determine whether the generated first simulated full-sampling magnetic resonance image is a simulated magnetic resonance image category, and obtain the first discrimination result, so that Afterwards, the first discrimination result is fed back to the training of the circular deep neural network through the preset loss function, so as to improve the quality of the simulated full-sampling magnetic resonance image generated by the neural network on one side of the circular deep neural network.
  • Step 15 Input the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image to the pre-built second simulated magnetic resonance image category discrimination model to obtain whether the first simulated under-sampled magnetic resonance image is a simulated magnetic resonance The second discrimination result of the image category.
  • the first simulated under-sampled magnetic resonance image may be generated through step 13.
  • the under-sampled magnetic resonance image may be obtained through step 11.
  • the neural network on the other side of the ring-shaped deep neural network can also learn the mapping relationship from the full-sampled magnetic resonance image to the under-sampled magnetic resonance image.
  • it is based on the same as step 14 That is, the deviation between the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image matched with the full-sampled magnetic resonance image is fed back to the training of the ring deep neural network, which can increase the ring depth The quality of the simulated magnetic resonance image generated by the neural network.
  • the pre-built second simulated magnetic resonance image category discrimination model it is possible to determine whether the generated first simulated under-sampled magnetic resonance image is a simulated magnetic resonance image category, and obtain the second discrimination result so as to Afterwards, the second discrimination result is fed back to the training of the circular deep neural network through the preset loss function, that is, the deviation between the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image is fed back to the circular deep neural network Network training.
  • the pre-built second simulated magnetic resonance image category discrimination model may also be constructed based on a convolutional neural network.
  • the second simulated magnetic resonance image category discrimination model may also be used as shown in 6 shows the network structure.
  • other types of neural network constructions can also be used, which is not limited in this application.
  • the second simulated magnetic resonance image category discrimination model in order to make the second simulated magnetic resonance image category discrimination model accurately output the second discrimination result, it can be understood as making the second simulated magnetic resonance image discriminate the first
  • the simulated under-sampled magnetic resonance image is a simulated magnetic resonance image category.
  • the second simulated magnetic resonance image can be adjusted by the fifth loss function The model parameters of the image category discrimination model, so that the second simulated magnetic resonance image category discrimination model accurately outputs the second discrimination result.
  • the fifth loss function Can be:
  • x represents an under-sampled magnetic resonance image
  • y represents a fully-sampled magnetic resonance image
  • G ds (y) represents the first simulated under-sampled magnetic resonance image
  • D s represents the second simulated magnetic resonance image category discrimination model
  • D s (G ds (y)) represents the second discrimination result
  • D s (x) represents the discrimination result of the second simulated magnetic resonance image category discrimination model for the under-sampled magnetic resonance image category
  • E x represents the function mathematics of the under-sampled magnetic resonance image as the input Expectation
  • E y represents the mathematical expectation of the function when the input is a fully sampled magnetic resonance image
  • MSE represents the mean square error function.
  • Adjusting the model parameters according to the fifth loss function may specifically include: calculating the loss value of the discrimination result of the second simulated magnetic resonance image category discrimination model through the fifth loss function, and feeding back the loss value to the second simulated magnetic resonance image by way of back propagation.
  • Resonance image category discrimination model to adjust model parameters.
  • the input data of the second simulated magnetic resonance image classification model includes the first simulated under-sampled magnetic resonance image
  • the first simulated under-sampled magnetic resonance image may be the other side neural network through the ring deep neural network It can be considered that the second simulated magnetic resonance image category discrimination model and the circular deep neural network can be trained at the same time.
  • the training effects of the two are opposite, that is, training the ring-shaped deep neural network is expected to achieve the purpose of making the second simulated magnetic resonance image category discrimination model discriminate the first simulated under-sampled magnetic resonance image as a non-simulated magnetic resonance image category, while training the first
  • the simulated magnetic resonance image category discrimination model may be expected to improve the accuracy of model discrimination, so as to discriminate as far as possible the first simulated under-sampled magnetic resonance image as the simulated magnetic resonance image category.
  • the second discrimination result output by the second magnetic resonance image discrimination model can be fed back to the training of the circular deep neural network through the preset loss function, the accuracy of the discrimination of the second simulated magnetic resonance image category discrimination model is improved, that is, the output is increased
  • the accuracy of the second discrimination result is also equivalent to promoting the improvement of the quality of simulated magnetic resonance images generated by the circular deep neural network.
  • the pre-built second simulated magnetic resonance image category discrimination model it is possible to determine whether the generated first simulated under-sampled magnetic resonance image is a simulated magnetic resonance image category, and obtain the second discrimination result so as to Then, the second discrimination result is fed back to the training of the ring deep neural network through the preset loss function, so as to improve the quality of the simulated under-sampled magnetic resonance image generated by the other side neural network of the ring deep neural network.
  • Step 16 Adjust according to the preset loss function, the first discrimination result, the second discrimination result, the second simulated under-sampled magnetic resonance image, the second simulated full-sampled magnetic resonance image, the under-sampled magnetic resonance image, and the full-sampled magnetic resonance image.
  • the first discrimination result and the second discrimination result here can be obtained through step 14 and step 15 respectively.
  • the second simulated under-sampled magnetic resonance image and the second simulated full-sampled magnetic resonance image may be generated through step 13.
  • the under-sampled magnetic resonance image and the full-sampled magnetic resonance image may be obtained through step 11.
  • the circular deep neural network here can be constructed through step 12.
  • adjusting the network parameters of the circular deep neural network to obtain a trained magnetic resonance imaging model may specifically include:
  • the preset loss function may specifically include:
  • the first loss function is used to determine the first loss value according to the first discrimination result of whether the first simulated full-sampled magnetic resonance image is a simulated magnetic resonance image category;
  • the second loss function is used to determine the second loss value according to the second discrimination result of whether the first simulated under-sampled magnetic resonance image is a simulated magnetic resonance image category;
  • the third loss function is used according to the average absolute error between the second simulated full-sampled magnetic resonance image and the full-sampled magnetic resonance image, and the average absolute error between the second simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image , Determine the third loss value.
  • the first discrimination result, the second discrimination result, the second simulated under-sampled magnetic resonance image, the second simulated full-sampled magnetic resonance image, the under-sampled magnetic resonance image, and the full-sampled magnetic resonance image are respectively substituted into the preset
  • the loss value obtained in the loss function can specifically include:
  • the formulas of the first loss function, the second loss function, and the third loss function may be as follows:
  • Second loss function Can be:
  • x represents the under-sampled magnetic resonance image
  • G sd (x) represents the first simulated full-sampled magnetic resonance image
  • D d represents the first simulated magnetic resonance image classification model
  • D d (G sd (x)) represents the first Discrimination result
  • MSE represents the mean square error function
  • E x represents the mathematical expectation of the function that the input is an under-sampled magnetic resonance image
  • y represents the fully sampled magnetic resonance image
  • G ds (y) represents the first simulated under-sampled magnetic resonance image
  • D s represents the second simulated magnetic resonance image classification model
  • D s (G ds (y)) represents the second Judgment result
  • E y represents the mathematical expectation that the input is a function of a fully sampled magnetic resonance image
  • x cir represents the second simulated under-sampled magnetic resonance image
  • y cir represents the second simulated full-sampled magnetic resonance image
  • ⁇ x cir -x ⁇ 1 represents the difference between the second simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image
  • the average absolute error of ⁇ y cir -y ⁇ 1 represents the average absolute error between the second simulated full magnetic resonance image and the full sampled magnetic resonance image.
  • the training effect of the neural network can be measured according to the size of the loss value of the loss function.
  • the training objective of the circular deep neural network may be to reduce the loss value of the preset loss function as much as possible. Then it can be considered that in the iterative training process of the ring deep neural network, the trained ring deep neural network is obtained when the loss value of the preset loss function is lower than a certain preset value by adjusting the network parameters.
  • the trained magnetic resonance imaging model may be the side of the trained circular deep neural network. Neural Networks.
  • the constructed circular deep neural network during the training process, not only learned the mapping relationship from the under-sampled magnetic resonance image to the fully-sampled magnetic resonance image, the image generation direction, but also the
  • the other side of the neural network with the opposite direction of image generation can also learn the mapping relationship from the full-sampled magnetic resonance image to the under-sampled magnetic resonance image in the opposite direction, so that the mapping relationship learned by the neural network on one side can be corrected, so that the neural network on one side
  • the network can form a correct mapping in the desired image generation direction, thereby reducing the deviation between the generated MRI image and the actual MRI image, improving the quality of the MRI image generated by the magnetic resonance imaging model, and improving the learning ability of the neural network. Learning efficiency.
  • the generated first simulated magnetic resonance image and the first simulated under-sampled magnetic resonance image are also discriminated, and The discrimination result is fed back to the training of the ring deep neural network through the loss function. It is hoped to achieve the purpose of making the simulated magnetic resonance image discrimination model misjudge the simulated magnetic resonance image generated by the circular deep neural network as a non-simulated magnetic resonance image, so that the MRI image generated by the trained magnetic resonance imaging model is as close to the actual MRI as possible Image, thereby further improving the quality of the MRI image generated by the MRI model.
  • the above is the training method of the magnetic resonance imaging model provided in the embodiment of this specification.
  • the embodiment of this specification also provides a specific application scenario of the magnetic resonance imaging model trained by the training method of the magnetic resonance imaging model.
  • the specific application scenario may be a magnetic resonance image generation method provided by an embodiment of this specification, and the execution subject of the method may be a magnetic resonance imaging device.
  • the magnetic resonance image generation method specifically includes the following steps :
  • Step 21 Obtain under-sampled image data.
  • the under-sampled image data here may be image data collected by magnetic resonance imaging equipment using nuclear magnetic resonance phenomena. It is understandable that the under-sampled image data here may be a certain amount of image data collected by a magnetic resonance imaging device in order to increase the image data collection time.
  • under-sampling can be collected through regular under-sampling based on partial k-space, random under-sampling based on Compressed Sensing (CS) theory, and Radial and Spiral under-sampling based on non-Cartesian sampling trajectories.
  • Image data this application is not limited.
  • Step 22 Input the under-sampled image data obtained in step 21 into the trained magnetic resonance imaging model to generate a magnetic resonance image.
  • the trained magnetic resonance imaging model can be obtained by but not limited to the training method of the magnetic resonance imaging model in the embodiment of this specification.
  • the training method of the magnetic resonance imaging model reference may be made to the content shown in the above-mentioned embodiment of the specification, and in order to avoid redundant description, the description is omitted here.
  • the under-sampled image data is input into the trained MRI model, and high-quality MRI images can be generated through the trained MRI model, which is convenient for clinical diagnosis.
  • the resonance imaging model can further improve the quality of the generated MRI image.
  • the device specifically includes:
  • the acquiring module 101 is configured to acquire a magnetic resonance image data set; wherein the magnetic resonance image data set includes: an under-sampled magnetic resonance image and a full-sampled magnetic resonance image;
  • the construction module 102 is used to construct a ring-shaped deep neural network to be trained; wherein the ring-shaped deep neural network includes two sides of the neural network and two input ports; the two input ports are used to input the under-sampled magnetic resonance Image and the full-sampled magnetic resonance image; one side of the neural network on both sides is used to generate a first simulated full-sampled magnetic resonance image according to the under-sampled magnetic resonance image, and according to the neural network on both sides The first simulated under-sampled magnetic resonance image generated by the neural network on the other side is used to generate the second simulated full-sampled magnetic resonance image; the other side neural network is used to: generate the first simulated under-sampling based on the full-sampled magnetic resonance image A magnetic resonance image, generating a second simulated under-sampled magnetic resonance image according to the first simulated fully sampled magnetic resonance image generated by the one-side neural network;
  • the generating module 103 is configured to input the under-sampled magnetic resonance image and the full-sampled magnetic resonance image to the to-be-trained ring-shaped deep neural network through two input ports of the to-be-trained ring-shaped deep neural network. Included in the neural network on both sides to generate each simulated magnetic resonance image; each simulated magnetic resonance image includes: the first simulated under-sampled magnetic resonance image, the first simulated full-sampled magnetic resonance image, and the first simulated magnetic resonance image 2. a simulated under-sampled magnetic resonance image and the second simulated full-sampled magnetic resonance image;
  • the first discrimination module 104 is configured to input the first simulated full-sampled magnetic resonance image and the full-sampled magnetic resonance image into a pre-built first simulated magnetic resonance image category discrimination model to obtain a comparison of the first simulated magnetic resonance image Whether the simulated full-sampling magnetic resonance image is the first judgment result of the simulated magnetic resonance image category;
  • the second discrimination module 105 is configured to input the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image into a pre-built second simulated magnetic resonance image category discrimination model to obtain a comparison of the first simulated magnetic resonance image. Whether the simulated under-sampling magnetic resonance image is the second judgment result of the simulated magnetic resonance image category;
  • the parameter adjustment module 106 is configured to perform according to a preset loss function, the first discrimination result, the second discrimination result, the second simulated under-sampled magnetic resonance image, the second simulated full-sampled magnetic resonance image, and the The under-sampled magnetic resonance image and the full-sampled magnetic resonance image are adjusted, and the network parameters of the circular deep neural network are adjusted to obtain a trained magnetic resonance imaging model.
  • the specific workflow of the above device embodiment may include: acquiring module 101, acquiring a magnetic resonance image data set, building module 102, constructing a ring-shaped deep neural network to be trained, generating module 103, combining under-sampled magnetic resonance images and full-sampled magnetic resonance images
  • the images are respectively input through the two input ports of the circular deep neural network to be trained into the neural networks on both sides of the circular deep neural network to be trained to generate each simulated magnetic resonance image.
  • the first discriminating module 104 will The first simulated full-sampling magnetic resonance image and the full-sampling magnetic resonance image are input into the pre-built first simulated magnetic resonance image category discrimination model to obtain the first simulated full-sampling magnetic resonance image category.
  • the second discrimination module 105 inputs the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image into the pre-built second simulated magnetic resonance image category discrimination model to obtain a comparison of the first simulated under-sampled magnetic resonance Whether the resonance image is the second discrimination result of the simulated magnetic resonance image category, the parameter adjustment module 106, according to the preset loss function, the first discrimination result, the second discrimination result, the second simulated under-sampling magnetic resonance image, and the second simulated full sampling
  • the network parameters of the circular deep neural network are adjusted to obtain a trained magnetic resonance imaging model.
  • the parameter adjustment module 106 specifically includes:
  • a calculation loss value unit configured to combine the first discrimination result, the second discrimination result, the second simulated under-sampled magnetic resonance image, the second simulated full-sampled magnetic resonance image, and the under-sampled magnetic resonance The image and the fully sampled magnetic resonance image are respectively substituted into a preset loss function to obtain a loss value;
  • the adjustment unit is configured to adjust the network parameters of the circular deep neural network according to the loss value to obtain a trained magnetic resonance imaging model; the magnetic resonance imaging model includes the side of the circular deep neural network Neural Networks.
  • the neural network on one side of the ring-shaped deep neural network in the device includes: a first down-sampling layer, a first residual network layer, and a first up-sampling layer; the other side of the neural network includes: The second downsampling layer, the second residual network layer, and the second upsampling layer;
  • the first down-sampling layer is used to extract a first feature map of the under-sampled magnetic resonance image, and extract a second feature map of the first simulated under-sampled magnetic resonance image;
  • the first residual network layer is used to perform residual processing on the first feature map and the second feature map to obtain a first feature map after residual processing and a second feature map after residual processing ;
  • the first upsampling layer is configured to generate the first simulated full-sampling magnetic resonance image according to the first characteristic map after the residual processing, and to generate the first simulated magnetic resonance image according to the second characteristic map after the residual processing. 2. Simulate full-sampling magnetic resonance images;
  • the second down-sampling layer is used to extract a third feature map of the full-sampled magnetic resonance image, and extract a fourth feature map of the first simulated full-sampling magnetic resonance image;
  • the second residual network layer is used to perform residual processing on the third feature map and the fourth feature map to obtain a third feature map after residual processing and a fourth feature map after residual processing ;
  • the second ascending layer is used to generate the first simulated under-sampled magnetic resonance image according to the third characteristic map after the residual processing, and to generate the first simulated under-sampled magnetic resonance image according to the fourth characteristic map after the residual processing. 2. Simulate under-sampled magnetic resonance images.
  • the preset loss function specifically includes:
  • the first loss function is used to determine the first loss value according to the first discrimination result of whether the first simulated full-sampled magnetic resonance image is a simulated magnetic resonance image category;
  • the second loss function is used to determine the second loss value according to the second discrimination result of whether the first simulated under-sampled magnetic resonance image is a simulated magnetic resonance image category;
  • the third loss function is used according to the average absolute error between the second simulated full-sampled magnetic resonance image and the full-sampled magnetic resonance image, and the average absolute value between the second simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image Error, determine the third loss value.
  • the first loss function for:
  • x represents the under-sampled magnetic resonance image
  • G sd (x) represents the first simulated full-sampled magnetic resonance image
  • D d represents the first simulated magnetic resonance image classification model
  • D d (G sd ( x)) represents the first discrimination result
  • MSE represents the mean square error function
  • E x represents the mathematical expectation that the input is a function of an under-sampled magnetic resonance image
  • y represents the fully sampled magnetic resonance image
  • G ds (y) represents the first simulated under-sampled magnetic resonance image
  • D s represents the second simulated magnetic resonance image classification model
  • D s (G ds ( y)) represents the second discrimination result
  • E y represents the mathematical expectation that the input is a function of a fully sampled magnetic resonance image
  • x cir represents the second simulated under-sampled magnetic resonance image
  • y cir represents the second simulated full-sampled magnetic resonance image
  • ⁇ x cir -x ⁇ 1 represents the second simulated under-sampled magnetic resonance image and the The average absolute error between the under-sampled magnetic resonance images
  • ⁇ y cir -y ⁇ 1 represents the average absolute error between the second simulated full magnetic resonance image and the full-sampled magnetic resonance image.
  • the device further includes:
  • the first model parameter adjustment module is configured to adjust the model parameters of the first simulated magnetic resonance image category discrimination model through a fourth loss function; the fourth loss function for:
  • the second model parameter adjustment module is configured to adjust the model parameters of the second simulated magnetic resonance image category discrimination model through the fifth loss function; the fifth loss function for:
  • D d (y) represents the discrimination result of the first simulated magnetic resonance image classification model for the full-sampling magnetic resonance image
  • D s (x) represents the comparison of the second simulated magnetic resonance image classification discrimination model The discrimination result of the under-sampled magnetic resonance image category.
  • the constructed circular deep neural network during the training process, not only learned the mapping relationship from the under-sampled magnetic resonance image to the fully-sampled magnetic resonance image, the image generation direction, but also the
  • the other side of the neural network with the opposite direction of image generation can also learn the mapping relationship from the full-sampled magnetic resonance image to the under-sampled magnetic resonance image in the opposite direction, so that the mapping relationship learned by the neural network on one side can be corrected, so that the neural network on one side
  • the network can form a correct mapping in the desired image generation direction, thereby reducing the deviation between the generated MRI image and the actual MRI image, improving the quality of the MRI image generated by the magnetic resonance imaging model, and improving the learning ability of the neural network. Learning efficiency.
  • the generated first simulated magnetic resonance image and the first simulated under-sampled magnetic resonance image are also discriminated, and The discrimination result is fed back to the training of the ring deep neural network through the loss function. It is hoped to achieve the purpose of making the simulated magnetic resonance image discrimination model misjudge the simulated magnetic resonance image generated by the circular deep neural network as a non-simulated magnetic resonance image, so that the MRI image generated by the trained magnetic resonance imaging model is as close to the actual MRI as possible Image, thereby further improving the quality of the MRI image generated by the MRI model.
  • an embodiment of this specification also provides a magnetic resonance image generation device. As shown in FIG. 9, the device specifically includes:
  • the image data acquisition block 201 is used to acquire under-sampled image data
  • the generating module 202 is configured to input the under-sampled image data into the trained magnetic resonance imaging model to generate a magnetic resonance image.
  • the specific work flow of the above embodiment of the magnetic resonance image generation apparatus may include: an image data acquisition block 201 to acquire under-sampled image data; a generation module 202 to input the under-sampled image data acquired through the image data acquisition block 201 to the trained Magnetic resonance imaging model to generate magnetic resonance images.
  • the magnetic resonance image generation device in the embodiment of this specification it is possible to generate high-quality MRI images according to the small amount of image data collected, especially the magnetic resonance image obtained by training using the training method of the magnetic resonance imaging model in the embodiment of this specification.
  • the imaging model can further improve the quality of the generated MRI image.
  • the embodiment of this specification also proposes an electronic device. Please refer to FIG. 10 for a schematic diagram.
  • the electronic device includes a processor, and optionally an internal bus, a network interface, and a memory.
  • the memory may include memory, such as high-speed random access memory (Random-Access Memory, RAM), or may also include non-volatile memory (non-volatile memory), such as at least one disk storage.
  • RAM Random-Access Memory
  • non-volatile memory such as at least one disk storage.
  • the electronic device may also include hardware required by other services.
  • the processor, network interface, and memory can be connected to each other through an internal bus.
  • the internal bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect, peripheral component interconnection standard) bus, and an EISA (Extended) bus. Industry Standard Architecture, extended industry standard structure) bus, etc.
  • the bus can be divided into an address bus, a data bus, a control bus, and so on. For ease of presentation, only one bidirectional arrow is used to indicate in FIG. 10, but it does not mean that there is only one bus and one type of bus.
  • the program may include program code, and the program code includes computer operation instructions.
  • the memory may include memory and non-volatile memory, and provide instructions and data to the processor.
  • the processor reads the corresponding computer program from the non-volatile memory to the memory and then runs it to form a training device applying the magnetic resonance imaging model on a logical level.
  • the processor executes the program stored in the memory and is used to perform at least the following operations:
  • the magnetic resonance image data set includes: an under-sampled magnetic resonance image and a full-sampled magnetic resonance image;
  • a ring-shaped deep neural network to be trained; wherein the ring-shaped deep neural network includes neural networks on both sides and two input ports; the two input ports are used to input the under-sampled magnetic resonance image and the full-sampling Magnetic resonance image
  • One side of the neural network of the two sides of the neural network is used to generate a first simulated full-sampling magnetic resonance image according to the under-sampled magnetic resonance image, and a first simulation generated according to the other side of the neural network of the two sides of the neural network
  • the under-sampled magnetic resonance image is used to generate a second simulated full-sampled magnetic resonance image
  • the other side neural network is used to generate the first simulated under-sampled magnetic resonance image according to the full-sampled magnetic resonance image, and according to the one side nerve
  • the first simulated fully-sampled magnetic resonance image generated by the network, and the second simulated under-sampled magnetic resonance image is generated;
  • each simulated magnetic resonance image includes: the first simulated under-sampled magnetic resonance image, the first simulated full-sampled magnetic resonance image, and the second simulated under-sampled magnetic resonance image An image and the second simulated full-sampling magnetic resonance image;
  • the first simulated full-sampling magnetic resonance image and the full-sampling magnetic resonance image are input to a pre-built first simulated magnetic resonance image classification model to obtain whether the first simulated full-sampling magnetic resonance image is The first discrimination result of the simulated magnetic resonance image category;
  • the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image are input to a pre-built second simulated magnetic resonance image category discrimination model to obtain whether the first simulated under-sampled magnetic resonance image is The second discrimination result of the simulated magnetic resonance image category;
  • the network parameters of the circular deep neural network are adjusted to obtain a trained magnetic resonance imaging model.
  • the method performed by the training device for the magnetic resonance imaging model disclosed in the embodiment shown in FIG. 1 of the present application may be applied to a processor, and may be implemented by the processor.
  • the processor may be an integrated circuit chip with signal processing capabilities.
  • the steps of the above method can be completed by the integrated logic circuit of hardware in the processor and instructions in the form of software.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (Fetwork Processor, FP), etc.; it can also be a digital signal processor (Digital Signal Processor, DSP), a dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) and other programmable logic devices, discrete gates and transistor logic devices, discrete hardware components.
  • DSP Digital Signal Processor
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • the methods, steps, and logical block diagrams disclosed in the embodiments of this specification can be implemented and executed.
  • the general-purpose processor may be a microprocessor and the processor may also be any conventional processor or the like.
  • the steps of the method disclosed in the embodiments of this specification can be directly embodied as being executed and completed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor.
  • the software module can be located in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, and registers.
  • the storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.
  • the electronic device can also execute the method executed by the training device of the magnetic resonance imaging model in FIG. 1 and realize the functions of the embodiment of the training device of the magnetic resonance imaging model shown in FIG. 1, which will not be repeated here.
  • the embodiment of this specification also proposes a computer-readable storage medium, the computer-readable storage medium stores one or more programs, the one or more programs include instructions, when the instructions are executed by an electronic device that includes multiple application programs At this time, the electronic device can be made to execute the method executed by the training device of the magnetic resonance imaging model in the embodiment shown in FIG. 1, and at least be used to execute:
  • the magnetic resonance image data set includes: an under-sampled magnetic resonance image and a full-sampled magnetic resonance image;
  • a ring-shaped deep neural network to be trained; wherein the ring-shaped deep neural network includes neural networks on both sides and two input ports; the two input ports are used to input the under-sampled magnetic resonance image and the full-sampling Magnetic resonance image
  • One side of the neural network of the two sides of the neural network is used to generate a first simulated full-sampling magnetic resonance image according to the under-sampled magnetic resonance image, and a first simulation generated according to the other side of the neural network of the two sides of the neural network
  • the under-sampled magnetic resonance image is used to generate a second simulated full-sampled magnetic resonance image
  • the other side neural network is used to generate the first simulated under-sampled magnetic resonance image according to the full-sampled magnetic resonance image, and according to the one side nerve
  • the first simulated fully-sampled magnetic resonance image generated by the network, and the second simulated under-sampled magnetic resonance image is generated;
  • each simulated magnetic resonance image includes: the first simulated under-sampled magnetic resonance image, the first simulated full-sampled magnetic resonance image, and the second simulated under-sampled magnetic resonance image An image and the second simulated full-sampling magnetic resonance image;
  • the first simulated full-sampling magnetic resonance image and the full-sampling magnetic resonance image are input to a pre-built first simulated magnetic resonance image classification model to obtain whether the first simulated full-sampling magnetic resonance image is The first discrimination result of the simulated magnetic resonance image category;
  • the first simulated under-sampled magnetic resonance image and the under-sampled magnetic resonance image are input to a pre-built second simulated magnetic resonance image category discrimination model to obtain whether the first simulated under-sampled magnetic resonance image is The second discrimination result of the simulated magnetic resonance image category;
  • the network parameters of the circular deep neural network are adjusted to obtain a trained magnetic resonance imaging model.
  • this application can be provided as methods, systems, and computer program products. Therefore, this application may adopt the form of a complete hardware embodiment, a complete software embodiment, and an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
  • computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions can also be stored in a computer-readable memory that can direct computers and other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device realizes the functions specified in one process and multiple processes in the flowchart and/and one block and multiple blocks in the block diagram.
  • These computer program instructions can also be loaded on computers and other programmable data processing equipment, so that a series of operation steps are executed on the computer and other programmable equipment to produce computer-implemented processing, which can be executed on the computer and other programmable equipment. Instructions provide steps for implementing functions specified in one flow and multiple flows in the flowchart and/and one block and multiple blocks in the block diagram.
  • the computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
  • processors CPUs
  • input/output interfaces network interfaces
  • memory volatile and non-volatile memory
  • Memory may include non-permanent memory in computer-readable media, random access memory (RAM) and/and non-volatile memory, such as read-only memory (ROM) and flash memory (flash RAM). Memory is an example of computer readable media.
  • RAM random access memory
  • ROM read-only memory
  • flash RAM flash memory
  • Computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be realized by any method and technology.
  • Information can be computer-readable instructions, data structures, program modules, and other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory and other memory technologies, CD-ROM, digital versatile disc (DVD) and other optical storage, Magnetic tape cartridges, magnetic tape magnetic disk storage and other magnetic storage devices and any other non-transmission media can be used to store information that can be accessed by computing devices. According to the definition in this article, computer-readable media does not include transitory media, such as modulated data signals and carrier waves.

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Medical Informatics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • General Physics & Mathematics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Pathology (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Multimedia (AREA)
  • Radiology & Medical Imaging (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Databases & Information Systems (AREA)
  • Signal Processing (AREA)
  • Psychiatry (AREA)
  • Physiology (AREA)
  • Fuzzy Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

一种磁共振成像模型的训练方法、装置、电子设备以及计算机可读存储介质。该方法包括：获取磁共振图像数据集（11）；构建待训练的环形深度神经网络（12）；将欠采样磁共振图像和全采样磁共振图像，分别输入至待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像（13）；将第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型（14），以得到对第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；根据预设损失函数，调整环形深度神经网络的网络参数，以得到训练好的磁共振成像模型（16）。

Description

一种磁共振成像模型的训练方法及装置
交叉引用
本发明要求在2020年02月24日提交中国专利局、申请号为202010113026.2、发明名称为“一种磁共振成像模型的训练方法及装置”的中国专利申请的优先权，该申请的全部内容通过引用结合在本发明中。
技术领域
本申请涉及计算机技术领域，尤其涉及一种磁共振成像模型的训练方法、装置、电子设备以及计算机可读存储介质。
背景技术
磁共振成像(Magnetic Resonance Imaging，MRI)技术是利用核磁共振现象重建出人体组织图像的技术，可以提供丰富的人体组织图像信息且对人体没有电离损伤，是我国广泛应用的临床医学检查手段。
然而，由于MRI技术需要较长的图像数据采集时间，使得成像速度较慢，在成像过程中会因为被检查者的生理性运动而在生成的图像中引入较多伪影，降低图像质量，影响临床诊断。
要减少图像中伪影的出现，需要提高成像速度，也即缩短MRI的图像数据采集时间。目前缩短MRI的图像数据采集时间的方式，主要为减少图像数据的采集量，比如基于部分k空间的规则欠采样、基于压缩感知(Compressed Sensing，CS)理论的随机欠采样、以及基于非笛卡尔采样轨迹的Radial和Spiral欠采样等。然而，图像数据采集量的减少，势必会降低图像清晰程度。
近几年，将卷积神经网络、U-net卷积神经网络或残差卷积神经网络等深度学习方法应用至快速磁共振成像领域中，可以在采集少量图像数据的情况下，利用训练好的神经网络，快速重建出高质量MRI图像，是一种很有应用潜力的快速磁共振图像生成方法。
目前，采用的深度学习方法主要学习从欠采样图像数据到全采样图像数 据映射关系实现成像，但这种学习方式学习效率不高，生成的图像质量也相对较低。
发明内容
本说明书实施例提供一种磁共振成像模型的训练方法、装置、电子设备以及计算机可读存储介质，以解决现有技术中采用深度学习方式的学习效率不高，生成的图像质量也相对较低的问题。
本说明书实施例采用下述技术方案：
一种磁共振成像模型的训练方法，包括：
获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；
所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样磁共振图像；
将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
一种基于所述的磁共振成像模型的训练方法的磁共振图像生成方法，所述磁共振图像生成方法包括：
获取欠采样图像数据；
将所述欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
一种磁共振成像模型的训练装置，包括：
获取模块，用于获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
构建模块，用于构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
生成模块，用于将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样 磁共振图像；
第一判别模块，用于将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
第二判别模块，用于将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
参数调整模块，用于根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
一种基于所述的磁共振成像模型的训练装置的磁共振图像生成装置，所述磁共振图像生成装置包括：
图像数据获取模块，用于获取欠采样图像数据；
生成模块，用于将所述欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
一种电子设备，包括：存储器、处理器及存储在所在存储器上并可在所述处理器上运行的计算机程序，所述计算机程序被所述处理器执行时实现任一项所述磁共振成像模型的训练方法的步骤。
一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现任一项所述磁共振成像模型的训练方法的步骤。
本说明书实施例采用的上述至少一个技术方案能够达到以下有益效果：
由于仅学习单一图像生成方向(例如，欠采样图像至全采样图像)的映射关系，可能会导致生成的MRI图像与实际MRI图像存在较大偏差，且神经网络的学习效率也不高。本申请通过构建的环形深度神经网络，在训练的过程中，不仅学习了由欠采样磁共振图像到全采样磁共振图像这一图像生成方向的映射关系，由于增加了与该图像生成方向相反的另一侧神经网络，还 可以学习由全采样磁共振图像到欠采样磁共振图像这一相反方向的映射关系，从而可以矫正一侧神经网络学习的映射关系，使得一侧神经网络能在期望的图像生成方向上可以形成正确的映射，从而降低生成的MRI图像与实际MRI图像之间存在的偏差，提高磁共振成像模型生成MRI图像的质量，以及提高神经网络的学习能力和学习效率。
另一方面，通过第一模拟磁共振图像类别判别模型和第二模拟磁共振图像类别判别模型，还对生成的第一模拟全采样磁共振图像和第一模拟欠采样磁共振图像进行判别，并将判别结果通过损失函数反馈至环形深度神经网络的训练中。期望达到使模拟磁共振图像判别模型对通过环形深度神经网络生成的模拟磁共振图像误判为非模拟磁共振图像的目的，以使得训练得到的磁共振成像模型生成的MRI图像尽可能贴近实际MRI图像，从而进一步的提高磁共振成像模型生成MRI图像的质量。
附图说明
此处所说明的附图用来提供对本申请的进一步理解，构成本申请的一部分，本申请的示意性实施例及其说明用于解释本申请，并不构成对本申请的不当限定。在附图中：
图1为本说明书实施例提供的磁共振成像模型的训练方法的流程示意图；
图2为本说明书实施例提供的环形深度神经网络的结构示意图；
图3为本说明书实施例提供的残差块的结构示意图；
图4为本说明书实施例提供的一侧神经网络的结构示意图；
图5为本说明书实施例提供的另一侧神经网络的结构示意图；
图6为本说明书实施例提供的模拟磁共振图像类别判别模型的结构示意图；
图7为本说明书实施例提供的磁共振图像生成方法的流程示意图；
图8为本说明书实施例提供的磁共振成像模型的训练装置的结构示意图；
图9为本说明书实施例提供的磁共振图像生成装置的结构示意图；
图10为本说明书实施例提供的电子设备的结构示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合本申请具体实施例及相应的附图对本申请技术方案进行清楚、完整地描述。显然，所描述的实施例仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本申请保护的范围。
以下结合附图，详细说明本申请各实施例提供的技术方案。
近几年，将卷积神经网络、U-net卷积神经网络或残差卷积神经网络等深度学习方法应用至快速磁共振成像领域中，可以在采集少量图像数据的情况下，利用训练好的神经网络，快速重建出高质量MRI图像，是一种很有应用潜力的快速磁共振图像生成方法。
目前，采用的深度学习方法主要学习从欠采样图像数据到全采样图像数据映射关系实现成像，但这种学习方式学习效率不高，生成的磁共振图像质量也相对较低。
为解决上述技术问题，本说明书实施例提供了一种磁共振成像模型的训练方法，用于提高神经网络的学习效率和生成磁共振图像的质量。该方法的执行主体包括但不限于服务器、个人电脑、笔记本电脑、平板电脑、智能手机等可以通过运行预定程序和指令来执行数值计算和/和逻辑计算等预定处理过程的智能电子设备。其中，所述服务器可以是单个网络服务器和者多个网络服务器组成的服务器组和基于云计算(Cloud Computing)的由大量计算机和网络服务器构成的云。在本说明书实施例中，对该方法的执行主体不做限定。该方法的流程示意图如图1所示，包括下述步骤：
步骤11：获取磁共振图像数据集。
这里的磁共振图像数据集可以包括：欠采样磁共振图像和全采样磁共振图像。
在实际应用中，全采样磁共振图像可以来自于磁共振成像设备实际采集 的磁共振图像数据。欠采样磁共振图像可以是从全采样磁共振图像中抽取部分采样点数据形成的磁共振图像数据。其中，抽取部分采样点数据可以是基于K空间的规则抽取，也可以是基于压缩感知理论的随机抽取，对此本申请不做限制。
在实际应用中，在从全采样磁共振图像中抽取部分采样点数据生成欠采样磁共振图像后，还可以包括建立欠采样磁共振图像与该全采样磁共振图像之间的匹配关系，使得磁共振图像数据集中的全采样磁共振图像和欠采样磁共振图像是对应匹配的，方便之后神经网络的训练。
步骤12：构建待训练的环形深度神经网络。
考虑到现有技术中采用的深度学习方法仅学习单一图像生成方向的映射关系实现成像，使得生成磁共振图像质量相对较低以及神经网络的学习效率不高的问题，在本说明书实施例中，通过构建待训练的环形深度神经网络，以解决上述技术问题。其中，构建的环形深度神经网络可以包括两侧神经网络以及两个输入口。
在实际应用中，所述两个输入口可以用于分别输入欠采样磁共振图像和全采样磁共振图像，这里的欠采样磁共振图像和全采样磁共振图像可以是通过步骤11获取的。
在本说明书一个或多个实施例中，如图2所示的环形深度神经网络的结构示意图，该环形深度神经网络的一侧神经网络可以包括：第一降采样层、第一残差网络层、第一升采样层；另一侧神经网络可以包括：第二将采样层、第二残差网络层、第二升采样层。
其中，第一降采样层、第一升采样层、第二升采样层和第二降采样层可以分别包含至少一层卷积层，该至少一层卷积层中各个卷积层采用串行连接，也即，上一个卷积层的输出图可以作为下一个卷积层的输入图。每层卷积层可以对输入图进行卷积处理，每层卷积层中可以包括至少一个卷积核，每个卷积核用于指示一次卷积操作时的权重矩阵。需要说明的是，上述输入图和输出图可以均指特征图(feature map)。
第一残差网络层和第二残差网络层可以分别包含至少一个残差块。其中， 每个残差块的结构示意图可以如图3所示，每个残差块具体可以采用两层卷积层加上跳跃连接构成。其中，两层卷积层可以用于提取深层次的图像特征，跳跃连接用于将低层次的图像特征直接向后传递，与深层次的图像特征结合，可以提升神经网络的学习能力和稳定性，避免由于深度增加而导致较深的神经网络的训练效果产生退化。
如图4所示的是本说明书实施例给出的一实例性的深度环形深度神经网络的一侧神经网络的具体结构示意图。
在该一侧神经网络中，第一降采样层和第一升采样层分别包含3层卷积层，第一残差网络层可以包含9个残差块。在图4中，还示出了各个卷积层的卷积核个数和卷积核尺寸，例如，“conv,64”表示这个卷积层中设置有64个卷积核，“3*3”表示卷积核采用尺寸为3*3的小型卷积核，其余各个卷积层的卷积核个数以及卷积核尺寸不再赘述。
该第一降采样层的第一层卷积层，可以用于对输入的磁共振图像的像素值进行卷积处理(例如，对欠采样磁共振图像的像素值进行卷积处理)，得到该第一层卷积层输出的特征图，该输出的特征图作为接下来的第二层卷积层的输入图，以此类推，第二层卷积层的输出图作为第三层卷积层的输入图。
在实际应用中，每个卷积层输出的特征图数量可以与该卷积层的卷积核数量相同，例如，图4所示的第一降采样层的第一层卷积层输出的特征图的数量为64，第二层卷积层输出的特征图数量为128，第三层卷积层输出的特征图的数量为256，随着每层卷积层的卷积核数量的增加一倍，提取图像特征的深度增加一倍，每层卷积层输出的特征图的尺寸缩小一半。
经第一降采样层输出的每个特征图，可以代表输入的欠采样磁共振图像的局部特征，那么再经第一残差网络层，可以对第一降采样层输出的特征图进行残差处理，可以理解为，通过第一残差网络的每个残差块逐层对每个代表局部特征的特征图进行残差计算，以进一步提取更深层的图像特征，使得提取的局部特征更加显著，提高神经网络的学习能力。
第一升采样层的两层卷积层，可以将经第一残差网络层进行残差处理后的特征图逐层合并，例如，256个特征图合并为128个特征图，128个特征图 合并为64个特征图，以及相应的逐层扩大特征图尺寸。再通过全连接层对各特征图进行连接，输出生成的模拟磁共振图像，例如，将经两层卷积层合并后的64个特征图输入至全连接层以连接这64个特征图，输出生成的第一模拟全采样磁共振图像。
对于环形深度神经网络的另一侧神经网络的网络结构，可以与该一侧神经网络的网络结构相同，例如图5所述的另一侧神经网络的具体结构示意图，其中卷积层数、卷积核个数以及卷积核大小等可参照图4所示的一侧神经网络的解释说明，在此不做赘述。对于另一侧神经网络的第二降采样层、第二残差网络层以及第二升采样层的功能可以分别参照上述对第一降采样层、第一残差网络层以及第一升采样层的相关描述，在此不做赘述。
需要注意的是，上述实例性的环形深度神经网络是本说明书实施例提供的构建环形深度神经网络的一种具体实施方式，不代表本说明书实施例全部的实施方式。对于环形深度神经网络中包含的卷积层数、残差块个数、卷积核个数以及卷积核尺寸等可以根据实际需求设置，对此本申请不做限制。
步骤13：将欠采样磁共振图像和全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像。
这里的欠采样磁共振图像和全采样磁共振图像可以是通过步骤11获取的。这里的待训练的环形深度神经网络可以的通过步骤12构建的。
这里的各模拟磁共振图像可以包括：第一模拟欠采样磁共振图像、第一模拟全采样磁共振图像、第二模拟欠采样磁共振图像以及第二模拟全采样磁共振图像。
为了方便理解生成各模拟磁共振图像的过程，这里结合图2所示的环形深度神经网络做解释说明：
对于欠采样磁共振图像来说：将欠采样磁共振图像，从输入口1输入至一侧神经网络中，输出第一模拟全采样磁共振图像，该第一模拟全采样磁共振图像再经输入口2输入至另一侧神经网络中，生成第二模拟欠采样磁共振图像；
对于全采样磁共振图像来说：将全采样磁共振图像，从输入口2输入至另一侧神经网络中，输出第一模拟欠采样磁共振图像，该第一模拟欠采样磁共振图像再经输入口1输入至一侧神经网络中，生成第二模拟全采样磁共振图像。
由此，可以理解环形深度神经网络包含的两侧神经网络分别所起的作用可以包括：一侧神经网络用于：根据欠采样磁共振图像生成第一模拟全采样磁共振图像，根据两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据全采样磁共振图像生成第一模拟欠采样磁共振图像，根据两侧神经网络的一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像。
在本说明书实施例中，为进一步理解两侧神经网络生成各模拟磁共振图像的过程，结合图2所示的两侧神经网络的网络结构，具体来说生成各模拟磁共振图像的过程，具体可以包括：
对于欠采样磁共振图像来说：将欠采样磁共振图像，从输入口1输入至一侧神经网络中，通过第一降采样层提取欠采样磁共振图像的第一特征图，通过第一残差网络对第一特征图进行残差处理，得到残差处理后的第一特征图，再通过第一升采样层，根据该残差处理后的第一特征图生成第一模拟全采样磁共振图像；该第一模拟全采样磁共振图像再经输入口2输入至另一侧神经网络中，通过第二降采样层提取该第一模拟全采样磁共振图像的第四特征图，通过第二残差网络层，对第四特征图进行残差处理，得到残差处理后的第四特征图，再通过第二升采样层根据残差处理后的第四特征图生成第二模拟欠采样磁共振图像。
对于全采样磁共振图像来说：将全采样磁共振图像，从输入口2输入至另一侧神经网络中，通过第二降采样层提取全采样磁共振图像的第三特征图，通过第二残差网络对第三特征图进行残差处理，得到残差处理后的第三特征图，再通过第二升采样层，根据该残差处理后的第三特征图生成第一模拟欠采样磁共振图像；该第一模拟欠采样磁共振图像再经输入口1输入至一侧神 经网络中，通过第一降采样层提取该第一模拟欠采样磁共振图像的第二特征图，通过第一残差网络层，对第二特征图进行残差处理，得到残差处理后的第二特征图，再通过第一升采样层根据残差处理后的第二特征图生成第二模拟全采样磁共振图像。
步骤14：将第一模拟全采样磁共振图像和全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果。
这里的第一模拟全采样磁共振图像可以是通过步骤13生成的。这里的全采样磁共振图像可以是通过步骤11获取的。
在实际应用中，考虑到在训练过程中，根据欠采样磁共振图像生成的第一模拟全采样磁共振图像，和与欠采样磁共振图像相匹配的全采样磁共振图像之间可能存在偏差，若能将这种偏差反馈至环形深度神经网络的训练中，可以进一步提高环形深度神经网络生成磁共振图像的质量。
但在通过步骤12构建的环形深度神经网络中无法衡量这种偏差的大小，所以在本说明书实施例中，通过预先构建的第一模拟磁共振图像类别判别模型，可以对生成的第一模拟全采样磁共振图像是否为模拟磁共振图像类别进行判别，得到第一判别结果，以便于之后将第一判别结果通过预设损失函数反馈至环形深度神经网络的训练中，也即，将第一模拟全采样磁共振图像与全采样磁共振图像之间存在的偏差反馈至环形深度神经网络的训练中。
在本说明书实施例中，这里的预先构建的第一模拟磁共振图像类别判别模型，可以是基于卷积神经网络(Convolutional Neural Networks,CNN)构建的，例如，图6所示的可以是一种第一模拟磁共振图像类别判别模型的结构示意图。当然也可以采用其它类型的神经网络构建，对此本申请不做限制。
在实际应用中，在环形深度神经网络的训练过程中，为了使得第一模拟磁共振图像类别判别模型可以准确输出第一判别结果，可以理解为，使得第一模拟磁共振图像尽可能判别出第一模拟全采样磁共振图像为模拟磁共振图像类别，在本说明书一个或多个实施例中，对预先构建的第一模拟磁共振图像类别判别模型，可以通过第四损失函数调整第一模拟磁共振图像类别判别 模型的模型参数，以使得第一模拟磁共振图像类别判别模型准确输出第一判别结果。
所述第四损失函数可以为：
Figure PCTCN2020077011-appb-000001
其中，x表示欠采样磁共振图像，y表示全采样磁共振图像，G s-d(x)表示第一模拟全采样磁共振图像，D d表示第一模拟磁共振图像类别判别模型，D d(G s-d(x))表示第一判别结果，MSE表示均方误差(Mean Square Error，MSE)函数,E x表示输入为欠采样磁共振图像时的函数数学期望；D d(y)表示第一模拟磁共振图像类别判别模型对全采样磁共振图像类别的判别结果；E y表示输入为全采样磁共振图像的函数数学期望。
根据第四损失函数调整模型参数可以具体包括：通过第四损失函数计算第一模拟磁共振图像类别判别模型的判别结果的损失值，通过反向传播的方式将该损失值反馈至第一模拟磁共振图像类别判别模型中以调整模型参数。
在实际应用中，由于第一模拟磁共振图像类别判别模型的输入数据包含第一模拟全采样磁共振图像，而第一模拟全采样磁共振图像可以是通过环形深度神经网络的一侧神经网络生成的，则可以认为第一模拟磁共振图像类别判别模型和环形深度神经网络可以是同时训练的。两者的训练效果相反，即，训练环形深度神经网络期望达到，使第一模拟磁共振图像类别判别模型判别第一模拟全采样磁共振图像为非模拟磁共振图像类别的目的，而训练第一模拟磁共振图像类别判别模型，可以是期望提高模型判别的准确度，以尽可能判别出第一模拟全采样磁共振图像为模拟磁共振图像类别。由于第一磁共振图像判别模型输出的第一判别结果可以通过预设损失函数反馈至环形深度神经网络的训练中，那么提高第一模拟磁共振图像类别判别模型判别的准确度，也即提高输出的第一判别结果的准确度，也相当于促进提高环形深度神经网络生成模拟磁共振图像的质量。
在本说明书实施例中，通过预先构建的第一模拟磁共振图像类别判别模型，可以对生成的第一模拟全采样磁共振图像是否为模拟磁共振图像类别进行判别，得到第一判别结果，以便于之后将第一判别结果通过预设损失函数 反馈至环形深度神经网络的训练中，从而提高环形深度神经网络的一侧神经网络生成模拟全采样磁共振图像的质量。
步骤15：将第一模拟欠采样磁共振图像和欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果。
这里的第一模拟欠采样磁共振图像可以是通过步骤13生成的。这里的欠采样磁共振图像可以是通过步骤11获取的。
在实际应用中，通过构建的环形深度神经网络的另一侧神经网络还可以学习由全采样磁共振图像到欠采样磁共振图像的映射关系，为了提高神经网络的学习效果，基于与步骤14相同的考虑，也即，将第一模拟欠采样磁共振图像和与全采样磁共振图像相匹配的欠采样磁共振图像之间存在的偏差，反馈至环形深度神经网络的训练中，可以提高环形深度神经网络生成的模拟磁共振图像的质量。
在本说明书实施例中，通过预先构建的第二模拟磁共振图像类别判别模型，可以对生成的第一模拟欠采样磁共振图像是否为模拟磁共振图像类别进行判别，得到第二判别结果，以便于之后将第二判别结果通过预设损失函数反馈至环形深度神经网络的训练中，也即，将第一模拟欠采样磁共振图像与欠采样磁共振图像之间存在的偏差反馈至环形深度神经网络的训练中。
在本说明书实施例中，这里的预先构建的第二模拟磁共振图像类别判别模型，也可以是基于卷积神经网络构建的，例如，第二模拟磁共振图像类别判别模型也可以是采用如图6所示的网络结构。当然，也可以采用其它类型的神经网络构建，对此本申请不做限制。
在实际应用中，在环形深度神经网络的训练过程中，为了使得第二模拟磁共振图像类别判别模型准确输出第二判别结果，可以理解为，使得第二模拟磁共振图像尽可能判别出第一模拟欠采样磁共振图像为模拟磁共振图像类别，在本说明书一个或多个实施例中，对预先构建的第二模拟磁共振图像类别判别模型，可以通过第五损失函数调整第二模拟磁共振图像类别判别模型的模型参数，以使得第二模拟磁共振图像类别判别模型准确输出第二判别结 果。
所述第五损失函数
Figure PCTCN2020077011-appb-000002
可以为：
Figure PCTCN2020077011-appb-000003
其中，x表示欠采样磁共振图像，y表示全采样磁共振图像，G d-s(y)表示第一模拟欠采样磁共振图像，D s表示第二模拟磁共振图像类别判别模型，D s(G d-s(y))表示第二判别结果，D s(x)表示第二模拟磁共振图像类别判别模型对欠采样磁共振图像类别的判别结果，E x表示输入为欠采样磁共振图像的函数数学期望，E y表示输入为全采样磁共振图像时的函数数学期望，MSE表示均方误差函数。
根据第五损失函数调整模型参数可以具体包括：通过第五损失函数计算第二模拟磁共振图像类别判别模型的判别结果的损失值，通过反向传播的方式将该损失值反馈至第二模拟磁共振图像类别判别模型中以调整模型参数。
在实际应用中，由于第二模拟磁共振图像类别判别模型的输入数据包含第一模拟欠采样磁共振图像，而第一模拟欠采样磁共振图像可以是通过环形深度神经网络的另一侧神经网络生成的，可以认为第二模拟磁共振图像类别判别模型和环形深度神经网络可以是同时训练的。两者的训练效果相反，即，训练环形深度神经网络期望达到，使第二模拟磁共振图像类别判别模型判别第一模拟欠采样磁共振图像为非模拟磁共振图像类别的目的，而训练第一模拟磁共振图像类别判别模型，可以是期望提高模型判别的准确度，以尽可能判别出第一模拟欠采样磁共振图像为模拟磁共振图像类别。由于第二磁共振图像判别模型输出的第二判别结果可以通过预设损失函数反馈至环形深度神经网络的训练中，那么提高第二模拟磁共振图像类别判别模型判别的准确度，也即提高输出的第二判别结果的准确度，也相当于促进提高环形深度神经网络生成模拟磁共振图像的质量。
在本说明书实施例中，通过预先构建的第二模拟磁共振图像类别判别模型，可以对生成的第一模拟欠采样磁共振图像是否为模拟磁共振图像类别进行判别，得到第二判别结果，以便于之后将第二判别结果通过预设损失函数反馈至环形深度神经网络的训练中，从而提高环形深度神经网络的另一侧神 经网络生成模拟欠采样磁共振图像的质量。
步骤16：根据预设损失函数、第一判别结果、第二判别结果、第二模拟欠采样磁共振图像、第二模拟全采样磁共振图像、欠采样磁共振图像和全采样磁共振图像，调整环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
这里的第一判别结果和第二判别结果可以是分别通过步骤14和步骤15得到的。这里的第二模拟欠采样磁共振图像和第二模拟全采样磁共振图像可以是通过步骤13生成的。这里的欠采样磁共振图像和全采样磁共振图像可以是通过步骤11获取的。这里的环形深度神经网络可以是通过步骤12构建的。
在实际应用中，调整神经网络的网络参数，可以通过损失函数计算神经网络输出结果的损失值，再通过反向传播的方式将损失值反馈至神经网络中以调整网络参数，则在本说明书一个或多个实施例中，调整环形深度神经网络的网络参数，以得到训练好的磁共振成像模型，可以具体包括：
将第一判别结果、第二判别结果、第二模拟欠采样磁共振图像、第二模拟全采样磁共振图像、欠采样磁共振图像和全采样磁共振图像分别代入至预设损失函数中得到损失值；根据该损失值，调整环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
在本说明书一个或多个实施例中，预设损失函数可以具体包括：
第一损失函数，用于根据对第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果，确定第一损失值；
第二损失函数，用于根据对第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果，确定第二损失值；
第三损失函数，用于根据第二模拟全采样磁共振图像与全采样磁共振图像之间的平均绝对误差，以及第二模拟欠采样磁共振图像与欠采样磁共振图像之间的平均绝对误差，确定第三损失值。
在实际应用中，将第一判别结果、第二判别结果、第二模拟欠采样磁共振图像、第二模拟全采样磁共振图像、欠采样磁共振图像和全采样磁共振图像分别代入至预设损失函数中得到损失值，可以具体包括：
将第一判别结果代入至第一损失函数得到第一损失值；
将第二判别结果代入至第二损失函数中得到第二损失值；
将第二模拟全采样磁共振图像、全采样磁共振图像、第二模拟欠采样磁共振图像与欠采样磁共振图像代入至第三损失函数中得到第三损失值。
在本说明书一个或多个实施例中，所述第一损失函数，第二损失函数以及第三损失函数的公式可以分别如下所示：
第一损失函数
Figure PCTCN2020077011-appb-000004
可以为：
Figure PCTCN2020077011-appb-000005
第二损失函数
Figure PCTCN2020077011-appb-000006
可以为：
Figure PCTCN2020077011-appb-000007
第三损失函数
Figure PCTCN2020077011-appb-000008
可以为：
Figure PCTCN2020077011-appb-000009
其中，x表示欠采样磁共振图像，G s-d(x)表示第一模拟全采样磁共振图像，D d表示第一模拟磁共振图像类别判别模型，D d(G s-d(x))表示第一判别结果，MSE表示均方误差函数,E x表示输入为欠采样磁共振图像的函数数学期望；
其中，y表示全采样磁共振图像，G d-s(y)表示第一模拟欠采样磁共振图像，D s表示第二模拟磁共振图像类别判别模型，D s(G d-s(y))表示第二判别结果，E y表示输入为全采样磁共振图像的函数数学期望；
其中，x cir表示第二模拟欠采样磁共振图像，y cir表示第二模拟全采样磁共振图像，‖x cir-x‖ 1表示第二模拟欠采样磁共振图像和欠采样磁共振图像之间的平均绝对误差，‖y cir-y‖ 1表示第二模拟全磁共振图像和全采样磁共振图像之间的平均绝对误差。
在实际应用中，可以根据损失函数的损失值的大小，衡量神经网络的训练效果。那么在本说明书实施例中，环形深度神经网络的训练目标可以是尽可能减小预设损失函数的损失值。那么可以认为在环形深度神经网络的迭代训练过程中，通过调整网络参数，使得预设损失函数的损失值低于某预设值时，得到训练好的环形深度神经网络。
在实际应用中，生成MRI图像通常是由欠采样图像数据生成全采样图像 数据，则在本说明书实施例中，训练好的磁共振成像模型可以是训练好的环形深度神经网络的所述一侧神经网络。
在本说明书实施例中，通过构建的环形深度神经网络，在训练的过程中，不仅学习了由欠采样磁共振图像到全采样磁共振图像这一图像生成方向的映射关系，由于增加了与该图像生成方向相反的另一侧神经网络，还可以学习由全采样磁共振图像到欠采样磁共振图像这一相反方向的映射关系，从而可以矫正一侧神经网络学习的映射关系，使得一侧神经网络能在期望的图像生成方向上可以形成正确的映射，从而降低生成的MRI图像与实际MRI图像之间存在的偏差，提高磁共振成像模型生成MRI图像的质量，以及提高神经网络的学习能力和学习效率。
另一方面，通过第一模拟磁共振图像类别判别模型和第二模拟磁共振图像类别判别模型，还对生成的第一模拟全采样磁共振图像和第一模拟欠采样磁共振图像进行判别，并将判别结果通过损失函数反馈至环形深度神经网络的训练中。期望达到使模拟磁共振图像判别模型对通过环形深度神经网络生成的模拟磁共振图像误判为非模拟磁共振图像的目的，以使得训练得到的磁共振成像模型生成的MRI图像尽可能贴近实际MRI图像，从而进一步的提高磁共振成像模型生成MRI图像的质量。
以上为本说明书实施例所提供的磁共振成像模型的训练方法，本说明书实施例还提供一种通过该磁共振成像模型的训练方法训练得到的磁共振成像模型的一种具体应用场景。该一种具体应用场景可以是本说明书实施例提供的一种磁共振图像生成方法，该方法的执行主体可以是磁共振成像设备，如图7所示，该磁共振图像生成方法具体包括以下步骤：
步骤21，获取欠采样图像数据。
在实际应用中，这里的欠采样图像数据可以是通过磁共振成像设备，利用核磁共振现象采集的图像数据。可以理解的是，这里的欠采样图像数据可以是为了提高图像数据采集时间，通过磁共振成像设备采集的一定数量的图像数据。
在实际应用中，可以通过基于部分k空间的规则欠采样、基于压缩感知(Compressed Sensing，CS)理论的随机欠采样、以及基于非笛卡尔采样轨迹的Radial和Spiral欠采样等方法，采集欠采样图像数据，对此本申请不做限制。
步骤22，将通过步骤21获取的欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
在本说明书实施例中，训练好的磁共振成像模型可以但不限于采用上述本说明书实施例中的磁共振成像模型的训练方法训练得到。其中，针对磁共振成像模型的训练方法的相关描述可参照上述本说明书实施例中示出的内容，为避免赘述，此处不再说明。
在实际应用中，将欠采样图像数据输入至训练好的磁共振成像模型中，可以通过训练好的磁共振成像模型生成高质量的MRI图像，便于临床诊断。
通过本说明书实施例中磁共振图像生成方法，可以实现根据采集的欠采样图像数据，生成高质量的MRI图像，尤其是采用上述本说明书实施例中的磁共振成像模型的训练方法训练得到的磁共振成像模型，相较于现有技术可以进一步提高生成的MRI图像的质量。
以上为本说明书实施例所提供的磁共振成像模型的训练方法，以及基于该磁共振成像模型的训练方法的磁共振图像生成方法。在本说明书实施例中，基于与磁共振成像模型的训练方法相同的发明构思，还提供了相应的磁共振成像模型的训练装置。如图8所示，该装置具体包括：
获取模块101，用于获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
构建模块102，用于构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共 振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
生成模块103，用于将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样磁共振图像；
第一判别模块104，用于将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
第二判别模块105，用于将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
参数调整模块106，用于根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
上述装置实施例的具体工作流程可以包括：获取模块101，获取磁共振图像数据集，构建模块102，构建待训练的环形深度神经网络，生成模块103，将欠采样磁共振图像和全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像，第一判别模块104，将第一模拟全采样磁共振图像和全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对第一模拟全采样磁共振图像是否为模拟磁共振图像类 别的第一判别结果，第二判别模块105，将第一模拟欠采样磁共振图像和欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果，参数调整模块106，根据预设损失函数、第一判别结果、第二判别结果、第二模拟欠采样磁共振图像、第二模拟全采样磁共振图像、欠采样磁共振图像和全采样磁共振图像，调整环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
在一种实施方式中，所述参数调整模块106，具体包括：
计算损失值单元，用于将所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像分别代入至预设损失函数中得到损失值；
调整单元，用于根据所述损失值，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型；所述磁共振成像模型包括所述环形深度神经网络的所述一侧神经网络。
在一种实施方式中，所述装置中的环形深度神经网络的一侧神经网络包括：第一降采样层、第一残差网络层、第一升采样层；另一侧神经网络包括：第二降采样层、第二残差网络层、第二升采样层；
所述第一降采样层，用于提取所述欠采样磁共振图像的第一特征图，提取所述第一模拟欠采样磁共振图像的第二特征图；
所述第一残差网络层，用于对所述第一特征图和所述第二特征图进行残差处理，得到残差处理后的第一特征图和残差处理后的第二特征图；
所述第一升采样层，用于根据所述残差处理后的第一特征图生成所述第一模拟全采样磁共振图像，根据所述残差处理后的第二特征图生成所述第二模拟全采样磁共振图像；
所述第二降采样层，用于提取所述全采样磁共振图像的第三特征图，提取所述第一模拟全采样磁共振图像的第四特征图；
所述第二残差网络层，用于对所述第三特征图和所述第四特征图进行残差处理，得到残差处理后的第三特征图和残差处理后的第四特征图；
所述第二升样层，用于根据所述残差处理后的第三特征图生成所述第一模拟欠采样磁共振图像，根据所述残差处理后的第四特征图生成所述第二模拟欠采样磁共振图像。
在一种实施方式中，所述预设损失函数具体包括：
第一损失函数，用于根据对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果，确定第一损失值；
第二损失函数，用于根据对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果，确定第二损失值；
第三损失函数，用于根据第二模拟全采样磁共振图像与全采样磁共振图像之间的平均绝对误差，以及第二模拟欠采样磁共振图像与欠采样磁共振图像之间的平均绝对值误差，确定第三损失值。
在一种实施方式中，所述第一损失函数
Figure PCTCN2020077011-appb-000010
为：
Figure PCTCN2020077011-appb-000011
所述第二损失函数
Figure PCTCN2020077011-appb-000012
为：
Figure PCTCN2020077011-appb-000013
所述第三损失函数
Figure PCTCN2020077011-appb-000014
为：
Figure PCTCN2020077011-appb-000015
其中，x表示所述欠采样磁共振图像，G s-d(x)表示所述第一模拟全采样磁共振图像，D d表示所述第一模拟磁共振图像类别判别模型，D d(G s-d(x))表示所述第一判别结果，MSE表示均方误差函数,E x表示输入为欠采样磁共振图像的函数数学期望；
其中，y表示所述全采样磁共振图像，G d-s(y)表示所述第一模拟欠采样磁共振图像，D s表示所述第二模拟磁共振图像类别判别模型，D s(G d-s(y))表示所述第二判别结果，E y表示输入为全采样磁共振图像的函数数学期望；
其中，x cir表示所述第二模拟欠采样磁共振图像，y cir表示所述第二模拟全采样磁共振图像，‖x cir-x‖ 1表示所述第二模拟欠采样磁共振图像和所述欠采样磁共振图像之间的平均绝对误差，‖y cir-y‖ 1表示所述第二模拟全磁共振图像和所述全采样磁共振图像之间的平均绝对误差。
在一种实施方式中，所述装置，还包括：
第一模型参数调整模块，用于通过第四损失函数，调整所述第一模拟磁共振图像类别判别模型的模型参数；所述第四损失函数
Figure PCTCN2020077011-appb-000016
为：
Figure PCTCN2020077011-appb-000017
第二模型参数调整模块，用于通过第五损失函数，调整第二模拟磁共振图像类别判别模型的模型参数；所述第五损失函数
Figure PCTCN2020077011-appb-000018
为：
Figure PCTCN2020077011-appb-000019
其中，D d(y)表示所述第一模拟磁共振图像类别判别模型对所述全采样磁共振图像类别的判别结果；D s(x)表示所述第二模拟磁共振图像类别判别模型对所述欠采样磁共振图像类别的判别结果。
在本说明书实施例中，通过构建的环形深度神经网络，在训练的过程中，不仅学习了由欠采样磁共振图像到全采样磁共振图像这一图像生成方向的映射关系，由于增加了与该图像生成方向相反的另一侧神经网络，还可以学习由全采样磁共振图像到欠采样磁共振图像这一相反方向的映射关系，从而可以矫正一侧神经网络学习的映射关系，使得一侧神经网络能在期望的图像生成方向上可以形成正确的映射，从而降低生成的MRI图像与实际MRI图像之间存在的偏差，提高磁共振成像模型生成MRI图像的质量，以及提高神经网络的学习能力和学习效率。
另一方面，通过第一模拟磁共振图像类别判别模型和第二模拟磁共振图像类别判别模型，还对生成的第一模拟全采样磁共振图像和第一模拟欠采样磁共振图像进行判别，并将判别结果通过损失函数反馈至环形深度神经网络的训练中。期望达到使模拟磁共振图像判别模型对通过环形深度神经网络生成的模拟磁共振图像误判为非模拟磁共振图像的目的，以使得训练得到的磁共振成像模型生成的MRI图像尽可能贴近实际MRI图像，从而进一步的提高磁共振成像模型生成MRI图像的质量。
基于与上述磁共振图像生成方法相同的发明构思，本说明书实施例还提供一种磁共振图像生成装置，如图9所示，该装置具体包括：
图像数据获取块201，用于获取欠采样图像数据；
生成模块202，用于将所述欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
上述磁共振图像生成装置实施例的具体工作流程可以包括：图像数据获取块201，获取欠采样图像数据；生成模块202，将通过图像数据获取块201获取到的欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
通过本说明书实施例中磁共振图像生成装置，可以实现根据采集的少量图像数据，生成高质量的MRI图像，尤其是采用上述本说明书实施例中的磁共振成像模型的训练方法训练得到的磁共振成像模型，相较于现有技术可以进一步提高生成的MRI图像的质量。
本说明书实施例还提出了一种电子设备，示意图请参考图10，在硬件层面，该电子设备包括处理器，可选地还包括内部总线、网络接口、存储器。其中，存储器可能包含内存，例如高速随机存取存储器(Random-Access Memory，RAM)，也可能还包括非易失性存储器(non-volatile memory)，例如至少1个磁盘存储器等。当然，该电子设备还可能包括其他业务所需要的硬件。
处理器、网络接口和存储器可以通过内部总线相互连接，该内部总线可以是ISA(Industry Standard Architecture，工业标准体系结构)总线、PCI(Peripheral Component Interconnect，外设部件互连标准)总线和EISA(Extended Industry Standard Architecture，扩展工业标准结构)总线等。所述总线可以分为地址总线、数据总线、控制总线等。为便于表示，图10中仅用一个双向箭头表示，但并不表示仅有一根总线和一种类型的总线。
存储器，用于存放程序。具体地，程序可以包括程序代码，所述程序代码包括计算机操作指令。存储器可以包括内存和非易失性存储器，并向处理器提供指令和数据。
处理器从非易失性存储器中读取对应的计算机程序到内存中然后运行， 在逻辑层面上形成应用磁共振成像模型的训练装置。处理器，执行存储器所存放的程序，并至少用于执行以下操作：
获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；
所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样磁共振图像；
将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
上述如本申请图1所示实施例揭示的磁共振成像模型的训练装置执行的方法可以应用于处理器中，和者由处理器实现。处理器可能是一种集成电路芯片，具有信号的处理能力。在实现过程中，上述方法的各步骤可以通过处理器中的硬件的集成逻辑电路和者软件形式的指令完成。上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Fetwork Processor，FP)等；还可以是数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field－Programmable Gate Array，FPGA)和者其他可编程逻辑器件、分立门和者晶体管逻辑器件、分立硬件组件。可以实现和者执行本说明书实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器和者该处理器也可以是任何常规的处理器等。结合本说明书实施例所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，和者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器和者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器，处理器读取存储器中的信息，结合其硬件完成上述方法的步骤。
该电子设备还可执行图1中磁共振成像模型的训练装置执行的方法，并实现磁共振成像模型的训练装置在图1所示实施例的功能，本说明书实施例在此不再赘述。
本说明书实施例还提出了一种计算机可读存储介质，该计算机可读存储介质存储一个和多个程序，该一个和多个程序包括指令，该指令当被包括多个应用程序的电子设备执行时，能够使该电子设备执行图1所示实施例中磁共振成像模型的训练装置执行的方法，并至少用于执行：
获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；
所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样磁共振图像；
将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、和计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、和结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个和多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本说明书实施例的方法、设备(系统)、和计算机程序产品的流程图和/和方框图来描述的。应理解可由计算机程序指令实现流程 图和/和方框图中的每一流程和/和方框、以及流程图和/和方框图中的流程和/和方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机和其它可编程数据处理设备的处理器以产生一个机器，使得通过计算机和其它可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程和多个流程和/和方框图一个方框和多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机和其它可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程和多个流程和/和方框图一个方框和多个方框中指定的功能。
这些计算机程序指令也可装载到计算机和其它可编程数据处理设备上，使得在计算机和其它可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机和其它可编程设备上执行的指令提供用于实现在流程图一个流程和多个流程和/和方框图一个方框和多个方框中指定的功能的步骤。
在一个典型的配置中，计算设备包括一个和多个处理器(CPU)、输入/输出接口、网络接口和内存。
内存可能包括计算机可读介质中的非永久性存储器，随机存取存储器(RAM)和/和非易失性内存等形式，如只读存储器(ROM)和闪存(flash RAM)。内存是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法和技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块和其它数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其它类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体和其它内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)和其它光学存储、磁盒式磁带，磁带磁磁盘存储和其它磁性存储设备和任何其它非传输介质，可用于存储可以被计算设备访问的信息。按照本文中的界定，计算机可读介质不包括暂存电脑可 读媒体(transitory media)，如调制的数据信号和载波。
还需要说明的是，术语“包括”、“包含”和者其任何其它变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品和者设备不仅包括那些要素，而且还包括没有明确列出的其它要素，和者是还包括为这种过程、方法、商品和者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括要素的过程、方法、商品和者设备中还存在另外的相同要素。
以上仅为本申请的实施例而已，并不用于限制本申请。对于本领域技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原理之内所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (10)

  1. 一种磁共振成像模型的训练方法，其特征在于，包括：
    获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
    构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；
    所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；所述另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
    将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样磁共振图像；
    将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
    将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
    根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
  2. 如权利要求1所述的方法，其特征在于，所述根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型，具体包括：
    将所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像分别代入至预设损失函数中得到损失值；
    根据所述损失值，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
  3. 如权利要求1所述的方法，其特征在于，所述一侧神经网络包括：第一降采样层、第一残差网络层、第一升采样层；所述另一侧神经网络包括：第二降采样层、第二残差网络层、第二升采样层；
    所述第一降采样层，用于提取所述欠采样磁共振图像的第一特征图，提取所述第一模拟欠采样磁共振图像的第二特征图；
    所述第一残差网络层，用于对所述第一特征图和所述第二特征图进行残差处理，得到残差处理后的第一特征图和残差处理后的第二特征图；
    所述第一升采样层，用于根据所述残差处理后的第一特征图生成所述第一模拟全采样磁共振图像，根据所述残差处理后的第二特征图生成所述第二模拟全采样磁共振图像；
    所述第二降采样层，用于提取所述全采样磁共振图像的第三特征图，提取所述第一模拟全采样磁共振图像的第四特征图；
    所述第二残差网络层，用于对所述第三特征图和第四特征图进行残差处理，得到残差处理后的第三特征图和残差处理后的第四特征图；
    所述第二升样层，用于根据所述残差处理后的第三特征图生成所述第一模拟欠采样磁共振图像，根据所述残差处理后的第四特征图生成所述第二模拟欠采样磁共振图像。
  4. 如权利要求1所述的方法，其特征在于，所述预设损失函数具体包括：
    第一损失函数，用于根据对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果，确定第一损失值；
    第二损失函数，用于根据对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果，确定第二损失值；
    第三损失函数，用于根据所述第二模拟全采样磁共振图像与所述全采样磁共振图像之间的平均绝对值误差，以及所述第二模拟欠采样磁共振图像与所述欠采样磁共振图像之间的平均绝对值误差，确定第三损失值。
  5. 如权利要求4所述的方法，其特征在于，
    所述第一损失函数为
    Figure PCTCN2020077011-appb-100001
    所述第二损失函数为
    Figure PCTCN2020077011-appb-100002
    所述第三损失函数为
    Figure PCTCN2020077011-appb-100003
    其中，x表示所述欠采样磁共振图像，G s-d(x)表示所述第一模拟全采样磁共振图像，D d表示所述第一模拟磁共振图像类别判别模型，D d(G s-d(x))表示所述第一判别结果，MSE表示均方误差函数,E x表示输入为欠采样磁共振图像的函数数学期望；
    其中，y表示所述全采样磁共振图像，G d-s(y)表示所述第一模拟欠采样磁共振图像，D s表示所述第二模拟磁共振图像类别判别模型，D s(G d-s(y))表示所述第二判别结果，E y表示输入为全采样磁共振图像的函数数学期望；
    其中，x cir表示所述第二模拟欠采样磁共振图像，y cir表示所述第二模拟全采样磁共振图像，‖x cir-x‖ 1表示所述第二模拟欠采样磁共振图像和所述欠采样磁共振图像之间的平均绝对误差，‖y cir-y‖ 1表示所述第二模拟全采样磁共振图像和所述全采样磁共振图像之间的平均绝对误差。
  6. 如权利要求1所述的方法，其特征在于，所述方法还包括：
    通过第四损失函数，调整所述第一模拟磁共振图像类别判别模型的模型参数；
    通过第五损失函数，调整所述第二模拟磁共振图像类别判别模型的模型参数；
    所述第四损失函数为
    Figure PCTCN2020077011-appb-100004
    Figure PCTCN2020077011-appb-100005
    所述第五损失函数为
    Figure PCTCN2020077011-appb-100006
    Figure PCTCN2020077011-appb-100007
    其中，D d(y)表示所述第一模拟磁共振图像类别判别模型对所述全采样磁共振图像类别的判别结果；D s(x)表示所述第二模拟磁共振图像类别判别模型对所述欠采样磁共振图像类别的判别结果。
  7. 一种基于权利要求1所述的磁共振成像模型的训练方法的磁共振图像生成方法，其特征在于，所述磁共振图像生成方法包括：
    获取欠采样图像数据；
    将所述欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
  8. 一种磁共振成像模型的训练装置，其特征在于，包括：
    获取模块，用于获取磁共振图像数据集；其中，所述磁共振图像数据集包括：欠采样磁共振图像和全采样磁共振图像；
    构建模块，用于构建待训练的环形深度神经网络；其中，所述环形深度神经网络包括两侧神经网络以及两个输入口；所述两个输入口用于分别输入所述欠采样磁共振图像和所述全采样磁共振图像；所述两侧神经网络的一侧神经网络用于：根据所述欠采样磁共振图像生成第一模拟全采样磁共振图像，根据所述两侧神经网络的另一侧神经网络生成的第一模拟欠采样磁共振图像，生成第二模拟全采样磁共振图像；另一侧神经网络用于：根据所述全采样磁共振图像生成所述第一模拟欠采样磁共振图像，根据所述一侧神经网络生成的第一模拟全采样磁共振图像，生成第二模拟欠采样磁共振图像；
    生成模块，用于将所述欠采样磁共振图像和所述全采样磁共振图像，分别通过待训练的环形深度神经网络的两个输入口，分别输入至所述待训练的环形深度神经网络包含的两侧神经网络中，以生成各模拟磁共振图像；所述各模拟磁共振图像包括：所述第一模拟欠采样磁共振图像、所述第一模拟全采样磁共振图像、所述第二模拟欠采样磁共振图像以及所述第二模拟全采样磁共振图像；
    第一判别模块，用于将所述第一模拟全采样磁共振图像和所述全采样磁共振图像，输入至预先构建的第一模拟磁共振图像类别判别模型，以得到对所述第一模拟全采样磁共振图像是否为模拟磁共振图像类别的第一判别结果；
    第二判别模块，用于将所述第一模拟欠采样磁共振图像和所述欠采样磁共振图像，输入至预先构建的第二模拟磁共振图像类别判别模型，以得到对所述第一模拟欠采样磁共振图像是否为模拟磁共振图像类别的第二判别结果；
    参数调整模块，用于根据预设损失函数、所述第一判别结果、所述第二判别结果、所述第二模拟欠采样磁共振图像、所述第二模拟全采样磁共振图像、所述欠采样磁共振图像和所述全采样磁共振图像，调整所述环形深度神经网络的网络参数，以得到训练好的磁共振成像模型。
  9. 一种基于权利要求8所述的磁共振成像模型的训练装置的磁共振成像装置，其特征在于，所述磁共振成像装置包括：
    图像数据获取模块，用于获取欠采样图像数据；
    生成模块，用于将所述欠采样图像数据输入至训练好的磁共振成像模型中，以生成磁共振图像。
  10. 一种电子设备，其特征在于，包括：存储器、处理器及存储在所在存储器上并可在所述处理器上运行的计算机程序，所述计算机程序被所述处理器执行时实现如权利要求1至7中任一项所述的磁共振成像模型的训练方法的步骤。
PCT/CN2020/077011 2020-02-24 2020-02-27 一种磁共振成像模型的训练方法及装置 WO2021168745A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/889,189 US20230047647A1 (en) 2020-02-24 2022-08-16 Methods and apparatuses for training magnetic resonance imaging model

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010113026.2 2020-02-24
CN202010113026.2A CN111358430B (zh) 2020-02-24 2020-02-24 一种磁共振成像模型的训练方法及装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/889,189 Continuation US20230047647A1 (en) 2020-02-24 2022-08-16 Methods and apparatuses for training magnetic resonance imaging model

Publications (1)

Publication Number Publication Date
WO2021168745A1 true WO2021168745A1 (zh) 2021-09-02

Family

ID=71199306

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077011 WO2021168745A1 (zh) 2020-02-24 2020-02-27 一种磁共振成像模型的训练方法及装置

Country Status (3)

Country Link
US (1) US20230047647A1 (zh)
CN (1) CN111358430B (zh)
WO (1) WO2021168745A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11460528B2 (en) * 2020-07-23 2022-10-04 Shanghai United Imaging Intelligence Co., Ltd. MRI reconstruction with image domain optimization
CN112924913B (zh) * 2021-02-02 2022-08-12 厦门大学 一种时空编码磁共振成像超分辨率重建方法及系统
CN117710514B (zh) * 2024-02-06 2024-06-11 中国科学院深圳先进技术研究院 动态磁共振成像方法、模型训练方法、装置、设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109325985A (zh) * 2018-09-18 2019-02-12 上海联影智能医疗科技有限公司 磁共振图像重建方法、装置和计算机可读存储介质
WO2019194865A1 (en) * 2018-04-02 2019-10-10 12 Sigma Technologies Enhanced convolutional neural network for image segmentation
CN110728732A (zh) * 2019-10-12 2020-01-24 深圳先进技术研究院 一种图像重建方法、装置、设备及介质
WO2020028382A1 (en) * 2018-07-30 2020-02-06 Memorial Sloan Kettering Cancer Center Multi-modal, multi-resolution deep learning neural networks for segmentation, outcomes prediction and longitudinal response monitoring to immunotherapy and radiotherapy
US20200058126A1 (en) * 2018-08-17 2020-02-20 12 Sigma Technologies Image segmentation and object detection using fully convolutional neural network

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106560848B (zh) * 2016-10-09 2021-05-11 辽宁工程技术大学 模拟生物双向认知能力的新型神经网络模型及训练方法
US10595727B2 (en) * 2018-01-25 2020-03-24 Siemens Healthcare Gmbh Machine learning-based segmentation for cardiac medical imaging
CN109544652B (zh) * 2018-10-18 2024-01-05 上海威豪医疗科技有限公司 基于深度生成对抗神经网络的核磁共振多加权成像方法
CN109523584B (zh) * 2018-10-26 2021-04-20 上海联影医疗科技股份有限公司 图像处理方法、装置、多模态成像系统、存储介质及设备
CN109712208B (zh) * 2018-12-13 2023-01-24 深圳先进技术研究院 基于深度学习的大视野磁共振扫描图像重建方法和装置
CN109658352B (zh) * 2018-12-14 2021-09-14 深圳市商汤科技有限公司 图像信息的优化方法及装置、电子设备和存储介质
CN109919831B (zh) * 2019-02-13 2023-08-25 广州视源电子科技股份有限公司 一种用于将视网膜眼底图像在不同图像域中迁移的方法、电子设备及计算机可读存储介质
CN109993809B (zh) * 2019-03-18 2023-04-07 杭州电子科技大学 基于残差U-net卷积神经网络的快速磁共振成像方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019194865A1 (en) * 2018-04-02 2019-10-10 12 Sigma Technologies Enhanced convolutional neural network for image segmentation
WO2020028382A1 (en) * 2018-07-30 2020-02-06 Memorial Sloan Kettering Cancer Center Multi-modal, multi-resolution deep learning neural networks for segmentation, outcomes prediction and longitudinal response monitoring to immunotherapy and radiotherapy
US20200058126A1 (en) * 2018-08-17 2020-02-20 12 Sigma Technologies Image segmentation and object detection using fully convolutional neural network
CN109325985A (zh) * 2018-09-18 2019-02-12 上海联影智能医疗科技有限公司 磁共振图像重建方法、装置和计算机可读存储介质
CN110728732A (zh) * 2019-10-12 2020-01-24 深圳先进技术研究院 一种图像重建方法、装置、设备及介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JING CHENG; HAIFENG WANG; YANJIE ZHU; QIEGEN LIU; QIYANG ZHANG; TING SU; JIANWEI CHEN; YONGSHUAI GE; ZHANLI HU; XIN LIU; HAIRONG Z: "Model-based Deep Medical Imaging: the roadmap of generalizing iterative reconstruction model using deep learning", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 19 June 2019 (2019-06-19), 201 Olin Library Cornell University Ithaca, NY 14853, XP081494204 *

Also Published As

Publication number Publication date
US20230047647A1 (en) 2023-02-16
CN111358430A (zh) 2020-07-03
CN111358430B (zh) 2021-03-09

Similar Documents

Publication Publication Date Title
WO2021168745A1 (zh) 一种磁共振成像模型的训练方法及装置
Xu et al. Locate globally, segment locally: A progressive architecture with knowledge review network for salient object detection
US11798132B2 (en) Image inpainting method and apparatus, computer device, and storage medium
WO2021253316A1 (zh) 图像降噪模型的训练方法、装置、电子设备和存储介质
CN111369440B (zh) 模型训练、图像超分辨处理方法、装置、终端及存储介质
WO2016124103A1 (zh) 一种图片检测方法及设备
CN111860398B (zh) 遥感图像目标检测方法、系统及终端设备
CN111429452A (zh) 基于UNet卷积神经网络的膀胱超声图像分割方法及装置
Sander et al. Automatic segmentation with detection of local segmentation failures in cardiac MRI
WO2022067874A1 (zh) 图像数据增强网络的训练方法及其训练装置、存储介质
CN111899185A (zh) 图像降噪模型的训练方法、装置、电子设备和存储介质
CN107633522A (zh) 基于局部相似性活动轮廓模型的脑部图像分割方法和系统
CN111429421A (zh) 模型生成方法、医学图像分割方法、装置、设备及介质
CN111179366A (zh) 基于解剖结构差异先验的低剂量图像重建方法和系统
WO2023071154A1 (zh) 图像分割方法及相关模型的训练方法和装置、设备
Chen et al. Shape registration with learned deformations for 3D shape reconstruction from sparse and incomplete point clouds
WO2024045442A1 (zh) 图像矫正模型的训练方法、图像矫正方法、设备及存储介质
WO2019196718A1 (zh) 元素图像生成方法、装置及系统
CN111915626B (zh) 心脏超声图像的心室区域自动分割方法、装置及存储介质
CN112541900A (zh) 基于卷积神经网络的检测方法、装置、计算机设备及存储介质
WO2024051018A1 (zh) 一种pet参数图像的增强方法、装置、设备及存储介质
WO2023109086A1 (zh) 文字识别方法、装置、设备及存储介质
US20220405916A1 (en) Method for detecting the presence of pneumonia area in medical images of patients, detecting system, and electronic device employing method
Cao et al. Orienting raw point sets by global contraction and visibility voting
US20230084240A1 (en) Neural-network-based mapping of potential leakage pathways of subsurface carbon dioxide storage

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: 20920828

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205N DATED 10.01.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20920828

Country of ref document: EP

Kind code of ref document: A1