WO2021119875A1 - Fast magnetic resonance imaging method and apparatus based on neural architecture search - Google Patents

Fast magnetic resonance imaging method and apparatus based on neural architecture search Download PDF

Info

Publication number
WO2021119875A1
WO2021119875A1 PCT/CN2019/125460 CN2019125460W WO2021119875A1 WO 2021119875 A1 WO2021119875 A1 WO 2021119875A1 CN 2019125460 W CN2019125460 W CN 2019125460W WO 2021119875 A1 WO2021119875 A1 WO 2021119875A1
Authority
WO
WIPO (PCT)
Prior art keywords
network model
magnetic resonance
data
neural network
sampling
Prior art date
Application number
PCT/CN2019/125460
Other languages
French (fr)
Chinese (zh)
Inventor
肖韬辉
王珊珊
李程
郑海荣
刘新
梁栋
Original Assignee
中国科学院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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2019/125460 priority Critical patent/WO2021119875A1/en
Publication of WO2021119875A1 publication Critical patent/WO2021119875A1/en

Links

Images

Classifications

    • 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
    • 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

Definitions

  • the present invention relates to the field of image processing, in particular to a method for reconstructing a magnetic resonance image using a neural network algorithm.
  • Magnetic Resonance Imaging as a multi-parameter, multi-contrast imaging technology, is one of the main imaging methods in modern medical imaging. It can reflect various characteristics of tissues such as T1, T2 and proton density. Provide information for the detection and diagnosis of diseases.
  • the basic working principle of magnetic resonance imaging is to use the phenomenon of magnetic resonance, use radio frequency excitation to excite hydrogen protons in the human body, use gradient fields for position encoding, and then use a receiving coil to receive electromagnetic signals with position information, and finally use Fourier transform to reconstruct Image information.
  • MRI requires a longer scan time, which not only brings some discomfort to the patient, but also prone to motion artifacts in the reconstructed image.
  • the long scan time limits the imaging of moving objects by MRI, such as blood flow and heart.
  • the method of accelerating acquisition is restricted by the ability of human nerves to withstand the magnetic field transformation and there is no room for further improvement.
  • deep learning methods have achieved significant results in image recognition and segmentation.
  • DNN deep neural networks
  • the traditional neural network algorithm needs to set the structure of the neural network in advance, and then use the pre-labeled label data to train the neural network, so as to obtain the neural network model that can be used for image processing.
  • how to determine the network structure can only rely on the experience of the algorithm designer to adjust and test the parameters to find a better network structure.
  • tuning parameters is a very difficult thing for deep models. Numerous hyperparameters and network structure parameters will produce explosive combinations, and traditional methods are difficult to find the optimal solution.
  • a lot of early work used evolutionary algorithms represented by genetic algorithms to optimize the hyperparameters and weights of neural networks because the neural network at that time had only a few layers, with more than a dozen neurons in each layer, and there was no complicated network architecture. The parameters are very limited and can be optimized directly.
  • the deep learning model has a complex network structure.
  • the weight parameters are usually in the hundreds of millions to billions, and evolutionary algorithms cannot be optimized at all.
  • the present invention is based on at least one of the above technical problems, and proposes a new method for determining the structure of a neural network model and a method for reconstructing a highly under-sampled magnetic resonance image.
  • the search space of the neural network structure is found to be the best Optimizing the structure, and using the neural network model corresponding to the optimal structure to reconstruct the highly under-sampled magnetic resonance data to obtain the reconstructed magnetic resonance image, which improves the optimization efficiency of the neural network and the effect of under-sampling magnetic resonance image reconstruction.
  • the embodiment of the first aspect of the present invention proposes a method for determining a neural network model for rapid reconstruction of magnetic resonance images, including:
  • S2 construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
  • the reinforcement learning algorithm is an algorithm based on a recurrent neural network model.
  • the recurrent neural network model is a long short-term memory network model (LSTM).
  • the first network model is a convolutional neural network model (CNN)
  • the convolutional neural network model can include different computing units, such as convolutional architectures, rectified linear units (ReLU), batch reduction Batch normalization, skip connections, etc. form the necessary units and extended structures of deep learning networks.
  • the search space includes network topology parameters that characterize these computing units and their connection relationships. Through the network topology parameters, the units and topology mechanisms of the neural network can be characterized, that is, according to a set of network topology parameters in the search space, A neural network model with corresponding structure can be constructed.
  • the step S1 of obtaining sample data and label data for model training further includes:
  • S12 Reconstruct the full-sampling magnetic resonance data to obtain a magnetic resonance image, which is used as tag data for training;
  • different sampling templates and different under-sampling magnifications can be used to generate multiple sets of under-sampling data for training, so as to improve the robustness of the model, and to adapt to different sampled data more widely to realize the image During reconstruction, it has good adaptability to different sampling methods and different sampling magnification data.
  • re-sampling can use a specific sampling template and sampling rate, that is, use a specific sampling method and sampling rate to generate training data, so that the trained model will be specifically suitable for the specific under-sampling Data, but the reconstruction accuracy can be improved.
  • This method is suitable for rapid reconstruction of magnetic resonance data sampled in a specific way. If the sampling method and sampling magnification have been determined during the magnetic resonance examination of the patient, the model can be trained and reconstructed in this way.
  • Another embodiment of the present invention provides a method for rapid reconstruction of magnetic resonance images, including:
  • the neural network model determined in the foregoing embodiment is used to reconstruct the under-sampled magnetic resonance data to obtain a magnetic resonance image of the target object.
  • the sampling method of the under-sampling data in the training data can use the same sampling method used when the magnetic resonance data is sampled on the target object.
  • another embodiment of the present invention provides a neural network model determining device, including:
  • the training data acquisition module is used to acquire sample data and label data for model training
  • the search space module is used to construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
  • a sub-network training module configured to use the sample data and the label data to train the first network model to obtain a trained first network model
  • An error calculation module configured to use test data to test the trained first network model to obtain an error result
  • the controller module is configured to use the reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and to correspond the optimal solution to the trained first network model Determined to be the neural network model used for rapid reconstruction of magnetic resonance images.
  • another embodiment of the present invention provides a computer storage medium storing one or more first instructions, and the one or more first instructions are suitable for being loaded and executed by a processor
  • the model training method in the foregoing embodiment or, the computer storage medium stores one or more second instructions, and the one or more second instructions are suitable for being loaded by the processor and executed in the foregoing embodiment Image processing method.
  • the neural structure search method can be used to automatically generate the network, and the reinforcement learning method can be used to continuously loop iteratively to obtain the optimal result.
  • Fig. 1 shows a method for determining a neural network model according to the first embodiment of the present invention
  • Fig. 2 shows a schematic diagram of a training data acquisition method according to the first embodiment of the present invention
  • Fig. 3 shows a schematic diagram of a neural network structure search method according to the first embodiment of the present invention
  • Fig. 4 shows a schematic diagram of a magnetic resonance image reconstruction method according to the second embodiment of the present invention
  • Fig. 5 shows a schematic block diagram of an apparatus for determining a neural network model according to a third embodiment of the present invention
  • Fig. 6 shows a schematic diagram of a magnetic resonance device according to the fourth embodiment of the present invention.
  • module or unit when a module or unit is referred to as being “on,” “connected to,” or “coupled to” another module or unit, it can be directly on the other module or unit or an intermediate module or unit that may exist, Connected or coupled to other modules or units or intermediate modules or units that may be present. In contrast, when a module or unit is referred to as being “directly on”, “directly connected to” or “directly coupled to” another module or unit, there may be no intervening modules or units.
  • the term “and/or” can include any and all combinations of one or more of the related listed items.
  • the MRI image can be generated by manipulating a virtual space called k-space.
  • k-space used herein may refer to a digital array (matrix) representing the spatial frequency in the MR image.
  • the k-space may be a 2D or 3D Fourier transform of the MR image.
  • the way of manipulating k-space, called k-space sampling, can affect the acquisition time (TA).
  • acquisition time can refer to the time to collect the signal of the entire pulse sequence.
  • the term "acquisition time” may refer to the time from the beginning of filling k-space to collecting the entire k-space data set.
  • acquisition time may refer to the time from the beginning of filling k-space to collecting the entire k-space data set.
  • Cartesian sampling two k-space sampling methods, Cartesian sampling and non-Cartesian sampling, are provided to manipulate k-space.
  • Cartesian sampling the k-space trajectory is a straight line
  • non-Cartesian sampling such as radiation sampling or spiral sampling
  • the k-space trajectory can be longer than the k-space trajectory in Cartesian sampling.
  • Fig. 1 shows a schematic block diagram of a method for determining a neural network model according to an embodiment of the present invention.
  • this embodiment includes the following steps:
  • the neural network of the present invention is used for the reconstruction of magnetic resonance images, and its input is under-sampled magnetic resonance data, and its output is a reconstructed magnetic resonance image.
  • An important prerequisite for the application of neural networks is the need for a training set.
  • the output samples in the training set are generally high-quality, noise-free magnetic resonance images.
  • the high-quality noise-free magnetic resonance image is generally reconstructed from full-sampling or super-full-sampling k-space data.
  • the acquisition of the full-sampling or ultra-full-sampling k-space data requires a long acquisition time.
  • the input data for model training should also be the same under-sampled K-space data as fast imaging.
  • the under-sampled K-space data can be obtained by sub-sampling the full-sampled K-space data. .
  • the step S1 of obtaining sample data and label data for model training further includes:
  • S12 Reconstruct the full-sampling magnetic resonance data to obtain a magnetic resonance image, which is used as tag data for training;
  • different sampling templates and different under-sampling magnifications can be used to generate multiple sets of under-sampling data for training, so as to improve the robustness of the model, and to adapt to different sampled data more widely to realize the image During reconstruction, it has good adaptability to different sampling methods and different sampling magnification data.
  • re-sampling can use a specific sampling template and sampling rate, that is, use a specific sampling method and sampling rate to generate training data, so that the trained model will be specifically suitable for the specific under-sampling Data, but the reconstruction accuracy can be improved.
  • This method is suitable for rapid reconstruction of magnetic resonance data sampled in a specific way. If the sampling method and sampling magnification have been determined during the magnetic resonance examination of the patient, the model can be trained and reconstructed in this way.
  • the full sampling data used for training can be filtered in advance to filter out magnetic resonance images with a signal-to-noise ratio lower than a preset threshold to obtain better results.
  • the image reconstructed from the full-sampled data can be normalized.
  • normalized preprocessing can also be performed before training.
  • S2 construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
  • the present invention uses a neural structure search method to construct a neural network model for image reconstruction.
  • designing a neural network structure usually requires a lot of structural engineering and technical knowledge. Therefore, Neural Architecture Search (NAS) emerged at the historic moment, and its main task is to automate the process of artificial neural network structure design.
  • NAS Neural Architecture Search
  • NAS Neural Architecture Search
  • the search space describes the set of potentially possible neural network architectures.
  • the search space is specifically designed for applications, such as convolutional network space for computer vision tasks, or recurrent neural network space for language modeling tasks. Therefore, the NAS method is not completely automated, because the design of these search spaces fundamentally relies on a human-designed architecture as a starting point. Even so, there are still many architectural parameters that need to be decided. In fact, the number of potential architectures that need to be considered in these search spaces usually exceeds 10 to the power of 10.
  • the neural network is used for fast magnetic resonance image reconstruction. Therefore, the neural network architecture in the search space is limited to the convolutional neural network.
  • the search space includes different computing units of the convolutional neural network, such as: Convolutional architectures, rectified linear units (ReLU), batch normalization, skip connections, etc. form the necessary units and extended structures for deep learning networks.
  • optimization method The optimization method is used to determine how to browse the search space in order to find a good architecture.
  • the most basic method here is random search, and various adaptive methods are also introduced, such as reinforcement learning, evolutionary search, gradient-based optimization and Bayesian optimization. Although these adaptive methods differ slightly in choosing which architectures to evaluate, they all try to search for network architectures that tend to be more likely to perform well. All these methods have corresponding methods in the context of traditional hyperparameter optimization tasks.
  • the present invention uses a reinforcement learning method to search. Specifically, a recurrent neural network model is used to perform the optimization method. In the present invention, this part of the component is called a controller.
  • Evaluation method The component measures the performance of each structure considered by the optimization method. The simplest, but the most computationally intensive option is to train a network completely.
  • the type of neural network is limited to a convolutional neural network, and a convolutional neural network with an initial structure can be constructed through initial parameters based on the calculation units in the search space and the connection relationship between the calculation units.
  • This step uses the training data obtained in step S1 to train the neural network model in the search space.
  • an initial structure is often required, and then the initial structure is trained using training data until convergence.
  • the convergence formula for training the first network model is as follows:
  • F represents the end-to-end mapping relationship between the sub-networks, F(x m,n ; ⁇ ) gets the output of the network; ⁇ represents the parameters that need to be learned; x m,n represents the input of the network; y m ,n represents the output label of the network.
  • training data For each neural network searched in the neural network structure search, training data needs to be used for training to obtain the trained first network model.
  • This step is used to evaluate different neural network structures in the search space.
  • different evaluation parameters can be used. For example, some application scenarios pay more attention to the calculation performance of the neural network model, so the calculation delay parameter can be used for evaluation; some application scenarios pay more attention to the accuracy of the neural network model processing result, so the error parameter can be selected for evaluation.
  • the reconstruction of magnetic resonance images pays more attention to the accuracy of the reconstruction results. Therefore, only error parameters are selected to evaluate different network structures.
  • the error evaluation uses test data that is different from the training data set.
  • the test data also uses the fully sampled k-space magnetic resonance data, the accurate magnetic resonance image is reconstructed from the fully sampled magnetic resonance data, and the under-sampled data is obtained by sub-sampling as the input data in the test data. .
  • the error calculation can use the mean square error, or other error calculation methods known in the prior art.
  • the test data set also includes a large number of different magnetic resonance images, and finally the test error results are obtained from multiple magnetic resonance images.
  • the part that performs reinforcement learning is generally called the controller.
  • the controller generates a sub-network based on the search space and trains the prepared training samples until convergence, and then tests on the verification set to obtain the corresponding error results , The result is fed back to the controller, and the controller makes corresponding adjustments based on the result to regenerate a sub-network, train, test and feedback again, and repeat the process to get the best result.
  • the adjustment process of the controller here uses reinforcement learning to train.
  • multiple controllers can be set up at the same time and multiple sub-networks can be generated, which can greatly improve the efficiency and performance of neural structure search.
  • Fig. 4 shows a schematic diagram of another embodiment of the present invention.
  • the second embodiment of the present invention provides a method for reconstructing a magnetic resonance image using a neural network model, which specifically includes:
  • the neural network model determined in the foregoing embodiment is used to reconstruct the under-sampled magnetic resonance data to obtain a magnetic resonance image of the target object.
  • the step of obtaining the under-sampled magnetic resonance data of the target object is to perform under-sampling magnetic resonance scanning of the human body through a magnetic resonance device.
  • Commonly used under-sampling methods for rapid magnetic resonance imaging generally include radial trajectories and spiral trajectories.
  • a higher sampling acceleration rate can be used for under-sampling to obtain a faster sampling speed.
  • the neural network model used is in the neural network structure search and the optimal network trained, the network can be used directly for image reconstruction to obtain the output magnetic resonance image.
  • the third embodiment of the present invention provides a neural network model determination device.
  • the model determination device may be a computer program (including program code) running in a terminal.
  • the model training device can execute the model determination method in the first embodiment, which specifically includes:
  • the training data acquisition module is used to acquire sample data and label data for model training
  • the search space module is used to construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
  • a sub-network training module configured to use the sample data and the label data to train the first network model to obtain a trained first network model
  • An error calculation module configured to use test data to test the trained first network model to obtain an error result
  • the controller module is configured to use the reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and to correspond the optimal solution to the trained first network model Determined to be the neural network model used for rapid reconstruction of magnetic resonance images.
  • Each unit in the model training device can be separately or completely combined into one or several other units to form, or some of the units can be further divided into functionally smaller units to form multiple units. The same operation can be achieved without affecting the realization of the technical effects of the embodiments of the present invention.
  • the above-mentioned units are divided based on logical functions. In practical applications, the function of one unit may also be realized by multiple units, or the functions of multiple units may be realized by one unit. In other embodiments of the present invention, the model-based training device may also include other units. In practical applications, these functions may also be implemented with the assistance of other units, and may be implemented by multiple units in cooperation.
  • a general-purpose computing device such as a computer including a central processing unit (CPU), a random access storage medium (RAM), a read-only storage medium (ROM) and other processing elements and storage elements
  • CPU central processing unit
  • RAM random access storage medium
  • ROM read-only storage medium
  • the computer program may be recorded on, for example, a computer-readable recording medium, and loaded into the above-mentioned computing device through the computer-readable recording medium, and run in it.
  • the fourth embodiment of the present invention provides a magnetic resonance device, including:
  • One or more processors are One or more processors;
  • Storage device for storing one or more programs
  • the one or more processors implement the method for rapid magnetic resonance image reconstruction as described in the second embodiment.
  • the device includes a processor 201, a memory 202, an input device 203, and an output device 204; the number of processors 201 in the device can be one or more.
  • one processor 201 is taken as an example; processing in the device
  • the device 201, the memory 202, the input device 203, and the output device 204 may be connected by a bus or other methods. In FIG. 6, the connection by a bus is taken as an example.
  • the memory 202 can be used to store software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the neural network model determination method in the first embodiment of the present invention, or as in the embodiment of the present invention.
  • the processor 201 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 202, that is, realizes the above-mentioned magnetic resonance image reconstruction method.
  • the memory 202 may mainly include a program storage area and a data storage area.
  • the program storage area may store an operating system and an application program required by at least one function; the data storage area may store data created according to the use of the terminal.
  • the memory 202 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices.
  • the memory 202 may further include a memory remotely provided with respect to the processor 201, and these remote memories may be connected to the device through a network. Examples of the foregoing network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
  • the input device 203 can be used to receive input digital or character information, and generate key signal input related to user settings and function control of the device.
  • the output device 204 may include a display device such as a display screen, for example, a display screen of a user terminal.
  • the fifth embodiment of the present invention provides a computer storage medium, the computer storage medium stores one or more first instructions, and the one or more first instructions are suitable for being loaded by a processor and executed in the foregoing embodiments.
  • Model training method or, the computer storage medium stores one or more second instructions, and the one or more second instructions are suitable for being loaded by the processor and executing the neural network determination method in the foregoing embodiment or Image reconstruction method.
  • the program can be stored in a computer-readable storage medium.
  • the storage medium includes read-only Memory (Read-Only Memory, ROM), Random Access Memory (RAM), Programmable Read-only Memory (PROM), Erasable Programmable Read Only Memory, EPROM), One-time Programmable Read-Only Memory (OTPROM), Electronically-Erasable Programmable Read-Only Memory (EEPROM), CD-ROM (Compact Disc) Read-Only Memory, CD-ROM) or other optical disk storage, magnetic disk storage, tape storage, or any other computer-readable medium that can be used to carry or store data.
  • Read-Only Memory ROM
  • RAM Random Access Memory
  • PROM Programmable Read-only Memory
  • EPROM Erasable Programmable Read Only Memory
  • OTPROM One-time Programmable Read-Only Memory
  • EEPROM Electronically-Erasable Programmable Read-Only Memory
  • CD-ROM Compact Disc

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

A fast magnetic resonance imaging method and apparatus based on a neural architecture search. A network is automatically generated in a neural architecture search manner, and the optimal result is obtained by using a reinforced learning manner to constantly perform loop iteration. A network architecture, namely, a sub-network, is obtained in a search space by means of a controller; then, training is performed on a prepared data set by using the sub-network; an error is obtained by performing testing on a validation set, and the error is transferred to the controller; the controller continues to perform optimization to obtain another network architecture; and this process is constantly repeated until an optimal reconstruction result is obtained.

Description

基于神经网络结构搜索的磁共振快速成像方法和装置Magnetic resonance fast imaging method and device based on neural network structure search 技术领域Technical field
本发明涉及图像处理领域，具体而言，涉及一种使用神经网络算法进行磁共振图像重建的方法。The present invention relates to the field of image processing, in particular to a method for reconstructing a magnetic resonance image using a neural network algorithm.
背景技术Background technique
磁共振成像(Magnetic Resonance Imaging，MRI)作为一种多参数、多对比度的成像技术，是现代医疗影像学中主要的成像方式之一，可以反映组织T1、T2和质子密度等多种特性，可为疾病的检出和诊断提供信息。磁共振成像的基本工作原理是利用磁共振现象，采用射频激励激发人体中的氢质子，运用梯度场进行位置编码，随后采用接收线圈接收带位置信息的电磁信号，最终利用傅里叶变换重建出图像信息。Magnetic Resonance Imaging (MRI), as a multi-parameter, multi-contrast imaging technology, is one of the main imaging methods in modern medical imaging. It can reflect various characteristics of tissues such as T1, T2 and proton density. Provide information for the detection and diagnosis of diseases. The basic working principle of magnetic resonance imaging is to use the phenomenon of magnetic resonance, use radio frequency excitation to excite hydrogen protons in the human body, use gradient fields for position encoding, and then use a receiving coil to receive electromagnetic signals with position information, and finally use Fourier transform to reconstruct Image information.
受傅里叶编码方式和奈奎斯特采样定理的限制，磁共振成像需要较长的扫描时间，不但给患者带来一定的不适，而且在重建的图像容易产生运动伪影。同时，过长的扫描时间限制了MRI对运动物体的成像，如血流、心脏等。依靠提高硬件性能，如梯度切换率和磁场强度等，加速采集的方式受制于人类神经对磁场变换的承受能力而无进一步提升的余地。近来，深度学习方法在图像识别，分割等方向取得显著成果，对于磁共振图像扫描时间慢的问题，深度神经网络(DNN，deep neural network)在近期被应用于加速磁共振扫描，以解决磁共振成像扫描速度慢的问题。Limited by the Fourier coding method and the Nyquist sampling theorem, MRI requires a longer scan time, which not only brings some discomfort to the patient, but also prone to motion artifacts in the reconstructed image. At the same time, the long scan time limits the imaging of moving objects by MRI, such as blood flow and heart. Relying on the improvement of hardware performance, such as gradient switching rate and magnetic field strength, the method of accelerating acquisition is restricted by the ability of human nerves to withstand the magnetic field transformation and there is no room for further improvement. Recently, deep learning methods have achieved significant results in image recognition and segmentation. For the problem of slow scanning time of MRI images, deep neural networks (DNN, deep neural network) have recently been applied to accelerate MRI scanning to solve MRI The imaging scan speed is slow.
但是传统神经网络算法需要预先设置好神经网络的结构，然后使用预先标注好的标签数据对神经网络进行训练，从而得到最终可以用于图像处理的神经网络模型。在传统框架下，如何确定网络结构，只能依靠算法设计人员的经验来进行调参和测试，找到较优的网络结构。但调参对于深度模型来说是一项非常困难的事情，众多的超参数和网络结构参数会产生***性的组合，传统方式难以找到最优解。早期很多工作都是用以遗传算法 为代表的进化算法对神经网络的超参数和权重进行优化，因为当时的神经网络只有几层，每层十几个神经元，也不存在复杂的网络架构，参数很有限，可直接进行优化。而深度学习模型一方面有着复杂的网络结构，另一方面权重参数通常都以百万到亿来计，进化算法根本无法优化。However, the traditional neural network algorithm needs to set the structure of the neural network in advance, and then use the pre-labeled label data to train the neural network, so as to obtain the neural network model that can be used for image processing. Under the traditional framework, how to determine the network structure can only rely on the experience of the algorithm designer to adjust and test the parameters to find a better network structure. But tuning parameters is a very difficult thing for deep models. Numerous hyperparameters and network structure parameters will produce explosive combinations, and traditional methods are difficult to find the optimal solution. A lot of early work used evolutionary algorithms represented by genetic algorithms to optimize the hyperparameters and weights of neural networks, because the neural network at that time had only a few layers, with more than a dozen neurons in each layer, and there was no complicated network architecture. The parameters are very limited and can be optimized directly. On the one hand, the deep learning model has a complex network structure. On the other hand, the weight parameters are usually in the hundreds of millions to billions, and evolutionary algorithms cannot be optimized at all.
发明内容Summary of the invention
本发明正是基于上述技术问题至少之一，提出了一种新的神经网络模型结构的确定方法和磁共振高度欠采样图像的重建方法，通过强化学习，在神经网络结构的搜索空间中找到最优结构，并使用该最优结构对应的神经网络模型对高度欠采样磁共振数据进行重建，得到重建的磁共振图像，提高了神经网络的优化效率和欠采样磁共振图像重建的效果。The present invention is based on at least one of the above technical problems, and proposes a new method for determining the structure of a neural network model and a method for reconstructing a highly under-sampled magnetic resonance image. Through reinforcement learning, the search space of the neural network structure is found to be the best Optimizing the structure, and using the neural network model corresponding to the optimal structure to reconstruct the highly under-sampled magnetic resonance data to obtain the reconstructed magnetic resonance image, which improves the optimization efficiency of the neural network and the effect of under-sampling magnetic resonance image reconstruction.
有鉴于此，本发明的第一方面的实施例，提出了一种用于磁共振图像快速重建的神经网络模型的确定方法，包括：In view of this, the embodiment of the first aspect of the present invention proposes a method for determining a neural network model for rapid reconstruction of magnetic resonance images, including:
S1：获取用于模型训练的样本数据和标签数据；S1: Obtain sample data and label data for model training;
S2：基于神经网络模型的拓扑结构相关的网络拓扑参数，构建搜索空间；根据所述搜索空间中的所述网络拓扑参数，建立对应的第一网络模型；S2: construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
S3：使用所述样本数据和所述标签数据对所述第一网络模型进行训练，得到训练好的第一网络模型；S3: Use the sample data and the label data to train the first network model to obtain a trained first network model;
S4：使用测试数据对所述训练好的第一网络模型进行测试，得到误差结果；S4: Use the test data to test the trained first network model to obtain an error result;
S5：使用强化学习算法和所述误差结果，在所述搜索空间中找到所述网络拓扑参数的最优解，将所述最优解对应的所述训练好的第一网络模型确定为所述用于磁共振图像快速重建的神经网络模型。S5: Use a reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and determine the trained first network model corresponding to the optimal solution as the A neural network model for rapid reconstruction of magnetic resonance images.
优选的，所强化学习算法为基于循环神经网络模型的算法。另一个优选实施方式中，所述循环神经网络模型为长短时记忆网络模型(LSTM)。Preferably, the reinforcement learning algorithm is an algorithm based on a recurrent neural network model. In another preferred embodiment, the recurrent neural network model is a long short-term memory network model (LSTM).
本实施方式中，第一网络模型为卷积神经网络模型(CNN)，卷积神经网络模型可以包括不同的计算单元，如：卷积结构(convolutional architectures)、整流线性单元(ReLU)、批量归一化(batch  normalization)、跳跃连接(skip connections)等组成深度学习网络必须的单元及扩展结构。对应的，搜索空间包括表征这些计算单元以及他们之间连接关系的网络拓扑参数，通过网络拓扑参数，可以表征神经网络的单元和拓扑机构，即，根据搜索空间中的一组网络拓扑参数，就可以构建出一个对应结构的神经网络模型。In this embodiment, the first network model is a convolutional neural network model (CNN), and the convolutional neural network model can include different computing units, such as convolutional architectures, rectified linear units (ReLU), batch reduction Batch normalization, skip connections, etc. form the necessary units and extended structures of deep learning networks. Correspondingly, the search space includes network topology parameters that characterize these computing units and their connection relationships. Through the network topology parameters, the units and topology mechanisms of the neural network can be characterized, that is, according to a set of network topology parameters in the search space, A neural network model with corresponding structure can be constructed.
本实施例中，获取用于模型训练的样本数据和标签数据的步骤S1进一步包括：In this embodiment, the step S1 of obtaining sample data and label data for model training further includes:
S11：获取用于模型训练的全采样磁共振数据；S11: Obtain full-sampling magnetic resonance data for model training;
S12：对所述全采样磁共振数据进行重建，得到磁共振图像，作为训练使用的标签数据；S12: Reconstruct the full-sampling magnetic resonance data to obtain a magnetic resonance image, which is used as tag data for training;
S13：对所述全采样磁共振数据进行重采样，得到欠采样数据，作为训练使用的样本数据。S13: Re-sampling the full-sampling magnetic resonance data to obtain under-sampling data as sample data for training.
进行重采样的步骤中，可以使用不同采样模板，不同欠采样倍率生成多组用于训练的欠采样数据，以提高模型后续的鲁棒性，更广泛的适应不同的采样数据，以实现在图像重建时对不同采样方法，不同采样倍率数据均有较好的适应性。In the re-sampling step, different sampling templates and different under-sampling magnifications can be used to generate multiple sets of under-sampling data for training, so as to improve the robustness of the model, and to adapt to different sampled data more widely to realize the image During reconstruction, it has good adaptability to different sampling methods and different sampling magnification data.
与之相反的，为了提高最终的重建效果，重采样可以使用特定的采样模板和采样倍率，即，使用特定采样方法和采样倍率生成训练数据，这样训练出来的模型将特定适用于该特定欠采样数据，但是可以提高重建精度。这种方法适用于特定方式采样的磁共振数据的快速重建。如果在对病人进行磁共振检查时，已经确定了采样方法和采样倍率，则可以使用这种方式训练模型并进行重建。On the contrary, in order to improve the final reconstruction effect, re-sampling can use a specific sampling template and sampling rate, that is, use a specific sampling method and sampling rate to generate training data, so that the trained model will be specifically suitable for the specific under-sampling Data, but the reconstruction accuracy can be improved. This method is suitable for rapid reconstruction of magnetic resonance data sampled in a specific way. If the sampling method and sampling magnification have been determined during the magnetic resonance examination of the patient, the model can be trained and reconstructed in this way.
本发明另一方面的实施例提出一种用于磁共振图像快速重建方法，包括：Another embodiment of the present invention provides a method for rapid reconstruction of magnetic resonance images, including:
获取目标对象的欠采样磁共振数据；Obtain under-sampled magnetic resonance data of the target object;
使用前述实施方式中确定的神经网络模型对所述欠采样磁共振数据进行重建，得到所述目标对象的磁共振图像。The neural network model determined in the foregoing embodiment is used to reconstruct the under-sampled magnetic resonance data to obtain a magnetic resonance image of the target object.
本发明中，为了获得更好的重建效果，训练数据中的欠采样数据的采样方式可以使用对目标对象进行磁共振数据采样时使用的相同的采样方式。In the present invention, in order to obtain a better reconstruction effect, the sampling method of the under-sampling data in the training data can use the same sampling method used when the magnetic resonance data is sampled on the target object.
再一个方面，本发明的又一个实施例提供一种神经网络模型确定装 置，包括：In another aspect, another embodiment of the present invention provides a neural network model determining device, including:
训练数据获取模块，用于获取用于模型训练的样本数据和标签数据；The training data acquisition module is used to acquire sample data and label data for model training;
搜索空间模块，用于基于神经网络模型的拓扑结构相关的网络拓扑参数，构建搜索空间；根据所述搜索空间中的所述网络拓扑参数，建立对应的第一网络模型；The search space module is used to construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
子网络训练模块，用于使用所述样本数据和所述标签数据对所述第一网络模型进行训练，得到训练好的第一网络模型；A sub-network training module, configured to use the sample data and the label data to train the first network model to obtain a trained first network model;
误差计算模块，用于使用测试数据对所述训练好的第一网络模型进行测试，得到误差结果；An error calculation module, configured to use test data to test the trained first network model to obtain an error result;
控制器模块，用于使用强化学习算法和所述误差结果，在所述搜索空间中找到所述网络拓扑参数的最优解，将所述最优解对应的所述训练好的第一网络模型确定为所述用于磁共振图像快速重建的神经网络模型。The controller module is configured to use the reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and to correspond the optimal solution to the trained first network model Determined to be the neural network model used for rapid reconstruction of magnetic resonance images.
再一个方面，本发明的又一个实施例提供一种计算机存储介质，所述计算机存储介质存储有一条或多条第一指令，所述一条或多条第一指令适于由处理器加载并执行前述实施例中的模型训练方法；或者，所述计算机存储介质存储有一条或多条第二指令，所述一条或多条第二指令适于由所述处理器加载并执行前述实施例中的图像处理方法。In yet another aspect, another embodiment of the present invention provides a computer storage medium storing one or more first instructions, and the one or more first instructions are suitable for being loaded and executed by a processor The model training method in the foregoing embodiment; or, the computer storage medium stores one or more second instructions, and the one or more second instructions are suitable for being loaded by the processor and executed in the foregoing embodiment Image processing method.
通过以上技术方案，可以采用神经结构搜索的方式进行自动生成网络，利用强化学习方式不断循环迭代得到最优结果。通过控制器在搜索空间中得到一个网络结构，即子网络，然后用子网络在制作好的数据集上进行训练，在验证集上测试得到误差，并将该误差传给控制器，控制器继续优化得到另一个网络结构，这样不断反复直到得到最佳的重建结果。Through the above technical solutions, the neural structure search method can be used to automatically generate the network, and the reinforcement learning method can be used to continuously loop iteratively to obtain the optimal result. Obtain a network structure in the search space through the controller, that is, the sub-network, and then use the sub-network to train on the prepared data set, test the error on the verification set, and pass the error to the controller, and the controller continues Optimize to get another network structure, and repeat this way until the best reconstruction result is obtained.
附图说明Description of the drawings
图1示出了根据本发明的实施例一的神经网络模型确定方法；Fig. 1 shows a method for determining a neural network model according to the first embodiment of the present invention;
图2示出了根据本发明的实施例一的训练数据获取方法的示意图；Fig. 2 shows a schematic diagram of a training data acquisition method according to the first embodiment of the present invention;
图3示出了根据本发明的实施例一的神经网络结构搜索方法的示意图；Fig. 3 shows a schematic diagram of a neural network structure search method according to the first embodiment of the present invention;
图4示出了根据本发明的实施例二的磁共振图像重建方法的示意图；Fig. 4 shows a schematic diagram of a magnetic resonance image reconstruction method according to the second embodiment of the present invention;
图5示出了根据本发明的实施例三的神经网络模型确定装置的示意框 图；Fig. 5 shows a schematic block diagram of an apparatus for determining a neural network model according to a third embodiment of the present invention;
图6示出了根据本发明的实施例四的磁共振设备的示意图。Fig. 6 shows a schematic diagram of a magnetic resonance device according to the fourth embodiment of the present invention.
具体实施方式Detailed ways
为了能够更清楚地理解本发明的上述目的、特征和优点，下面结合附图和具体实施方式对本发明进行进一步的详细描述。需要说明的是，在不冲突的情况下，本申请的实施例及实施例中的特征可以相互组合。In order to be able to understand the above objectives, features and advantages of the present invention more clearly, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other if there is no conflict.
在下面的描述中阐述了很多具体细节以便于充分理解本发明，但是，本发明还可以采用其他不同于在此描述的方式来实施，因此，本发明的保护范围并不受下面公开的具体实施例的限制。In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Therefore, the scope of protection of the present invention is not covered by the specific implementations disclosed below. Limitations of cases.
应当理解，当模块或单元被称为“在之上”、“连接到”或“耦合到”另一个模块或单元时，可以是直接在其他模块或单元或可以存在的中间模块或单元上、连接或耦合到其他模块或单元或可以存在的中间模块或单元。相反，当模块或单元被称为“直接在之上”、“直接连接到”或“直接耦合到”另一模块或单元时，可能不存在中间模块或单元。在本申请中，术语“和/或”可包括一个或以上相关所列条目的任何和所有组合。It should be understood that when a module or unit is referred to as being "on," "connected to," or "coupled to" another module or unit, it can be directly on the other module or unit or an intermediate module or unit that may exist, Connected or coupled to other modules or units or intermediate modules or units that may be present. In contrast, when a module or unit is referred to as being "directly on", "directly connected to" or "directly coupled to" another module or unit, there may be no intervening modules or units. In this application, the term "and/or" can include any and all combinations of one or more of the related listed items.
本申请中所使用的术语仅用于描述特定的示例性实施例，并不限制本申请的范围。如本申请使用的单数形式“一”、“一个”及“该”可以同样包括复数形式，除非上下文明确提示例外情形。还应当理解，如在本申请说明书中，术语“包括”、“包含”仅提示存在所述特征、整体、步骤、操作、组件和/或部件，但并不排除存在或添加一个或以上其他特征、整体、步骤、操作、组件、部件和/或其组合的情况。The terms used in this application are only used to describe specific exemplary embodiments and do not limit the scope of this application. For example, the singular forms "a", "an" and "the" used in this application may also include plural forms, unless the context clearly indicates exceptions. It should also be understood that, as in the specification of this application, the terms "including" and "including" only indicate the presence of the described features, wholes, steps, operations, components and/or components, but do not exclude the presence or addition of one or more other features , Overall, steps, operations, components, parts, and/or combinations thereof.
本申请一般涉及磁共振成像(MRI)，更具体地，涉及用于MRI中的快速成像的***和方法。可以通过操纵被称为k空间的虚拟空间来生成MRI图像。这里使用的术语“k空间”可以指表示MR图像中的空间频率的数字阵列(矩阵)。在一些实施例中，k空间可以是MR图像的2D或3D傅里叶变换。操纵k空间的方式，被称为k空间采样，可以影响采集时间(TA)。如这里所使用的，术语“采集时间”可以指采集整个脉冲序列的信 号的时间。例如，术语“采集时间”可以指从开始填充k空间到采集整个k空间数据集的时间。传统上，提供两个k空间采样方法，笛卡尔采样和非笛卡尔采样，以操纵k空间。在笛卡尔采样中，k空间轨迹是直线，而在非笛卡尔采样中，例如辐射采样或螺旋采样，k空间轨迹可以比笛卡尔采样中的k空间轨迹更长。This application relates generally to magnetic resonance imaging (MRI), and more specifically, to systems and methods for rapid imaging in MRI. The MRI image can be generated by manipulating a virtual space called k-space. The term "k-space" used herein may refer to a digital array (matrix) representing the spatial frequency in the MR image. In some embodiments, the k-space may be a 2D or 3D Fourier transform of the MR image. The way of manipulating k-space, called k-space sampling, can affect the acquisition time (TA). As used herein, the term "acquisition time" can refer to the time to collect the signal of the entire pulse sequence. For example, the term "acquisition time" may refer to the time from the beginning of filling k-space to collecting the entire k-space data set. Traditionally, two k-space sampling methods, Cartesian sampling and non-Cartesian sampling, are provided to manipulate k-space. In Cartesian sampling, the k-space trajectory is a straight line, while in non-Cartesian sampling, such as radiation sampling or spiral sampling, the k-space trajectory can be longer than the k-space trajectory in Cartesian sampling.
实施例一Example one
图1示出了根据本发明的一个实施例的神经网络模型确定方法的示意框图。Fig. 1 shows a schematic block diagram of a method for determining a neural network model according to an embodiment of the present invention.
如图1所示，根据本发明的一个实施例的神经网络模型确定方法，本实施例包括以下步骤：As shown in Fig. 1, according to the method for determining a neural network model according to an embodiment of the present invention, this embodiment includes the following steps:
S1：获取用于模型训练的样本数据和标签数据；S1: Obtain sample data and label data for model training;
本发明的神经网络用于磁共振图像的重建，其输入为欠采样的磁共振数据，输出为重建的磁共振图像。神经网络应用的一个重要前提是需要训练集，训练集中的输出样本一般为高质量无噪声的磁共振图像。该高质量无噪声的磁共振图像一般由全采样或超全采样的k-空间数据重建得到。该全采样或超全采样的k-空间数据的采集需要消耗较长的采集时间。为了模型能够用于快速磁共振成像，模型训练的输入数据同样应该是和快速成像相同的欠采样K空间数据，可以通过对全采样的K空间数据进行二次采样，得到欠采样的K空间数据。The neural network of the present invention is used for the reconstruction of magnetic resonance images, and its input is under-sampled magnetic resonance data, and its output is a reconstructed magnetic resonance image. An important prerequisite for the application of neural networks is the need for a training set. The output samples in the training set are generally high-quality, noise-free magnetic resonance images. The high-quality noise-free magnetic resonance image is generally reconstructed from full-sampling or super-full-sampling k-space data. The acquisition of the full-sampling or ultra-full-sampling k-space data requires a long acquisition time. In order for the model to be used for fast magnetic resonance imaging, the input data for model training should also be the same under-sampled K-space data as fast imaging. The under-sampled K-space data can be obtained by sub-sampling the full-sampled K-space data. .
如图2所示，本实施例中，获取用于模型训练的样本数据和标签数据的步骤S1进一步包括：As shown in FIG. 2, in this embodiment, the step S1 of obtaining sample data and label data for model training further includes:
S11：获取用于模型训练的全采样磁共振数据；S11: Obtain full-sampling magnetic resonance data for model training;
S12：对所述全采样磁共振数据进行重建，得到磁共振图像，作为训练使用的标签数据；S12: Reconstruct the full-sampling magnetic resonance data to obtain a magnetic resonance image, which is used as tag data for training;
S13：对所述全采样磁共振数据进行重采样，得到欠采样数据，作为训练使用的样本数据。S13: Re-sampling the full-sampling magnetic resonance data to obtain under-sampling data as sample data for training.
进行重采样的步骤中，可以使用不同采样模板，不同欠采样倍率生成多组用于训练的欠采样数据，以提高模型后续的鲁棒性，更广泛的适应不同的采样数据，以实现在图像重建时对不同采样方法，不同采样倍率数据 均有较好的适应性。In the re-sampling step, different sampling templates and different under-sampling magnifications can be used to generate multiple sets of under-sampling data for training, so as to improve the robustness of the model, and to adapt to different sampled data more widely to realize the image During reconstruction, it has good adaptability to different sampling methods and different sampling magnification data.
与之相反的，为了提高最终的重建效果，重采样可以使用特定的采样模板和采样倍率，即，使用特定采样方法和采样倍率生成训练数据，这样训练出来的模型将特定适用于该特定欠采样数据，但是可以提高重建精度。这种方法适用于特定方式采样的磁共振数据的快速重建。如果在对病人进行磁共振检查时，已经确定了采样方法和采样倍率，则可以使用这种方式训练模型并进行重建。On the contrary, in order to improve the final reconstruction effect, re-sampling can use a specific sampling template and sampling rate, that is, use a specific sampling method and sampling rate to generate training data, so that the trained model will be specifically suitable for the specific under-sampling Data, but the reconstruction accuracy can be improved. This method is suitable for rapid reconstruction of magnetic resonance data sampled in a specific way. If the sampling method and sampling magnification have been determined during the magnetic resonance examination of the patient, the model can be trained and reconstructed in this way.
传统算法对于高度欠采样的k空间数据的重建往往无法得到较好的效果，现有的用于高度欠采样磁共振图像重建的深度神经网络模型，如如MoDL、ADMM-Net、AUTOMAP、U-Net、VN-Net等神经网络模型，都是事先构建好的具有特定拓扑结构的网络，一般对于磁共振成像的部位、加速倍率、采样方式均有特定要求，适应性和扩展性差。Traditional algorithms often fail to achieve good results in the reconstruction of highly undersampled k-space data. Existing deep neural network models for highly undersampled magnetic resonance image reconstruction, such as MoDL, ADMM-Net, AUTOMAP, U- Neural network models such as Net and VN-Net are all pre-built networks with specific topological structures. Generally, they have specific requirements for the location, acceleration magnification, and sampling method of MRI, and they have poor adaptability and scalability.
此外，对于训练使用的全采样数据，可以预先进行筛选，滤除信噪比低于预设阈值的磁共振图像，以获得更好效果。In addition, the full sampling data used for training can be filtered in advance to filter out magnetic resonance images with a signal-to-noise ratio lower than a preset threshold to obtain better results.
进行训练前，可以对全采样数据重建的图像进行归一化处理。对于重采样得到的欠采样数据，也可以在训练前进行归一化的预处理。Before training, the image reconstructed from the full-sampled data can be normalized. For the under-sampled data obtained by resampling, normalized preprocessing can also be performed before training.
S2：基于神经网络模型的拓扑结构相关的网络拓扑参数，构建搜索空间；根据所述搜索空间中的所述网络拓扑参数，建立对应的第一网络模型；S2: construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
本发明使用神经结构搜索方法来构建用于图像重建神经网络模型，针对于不同的任务，设计神经网络结构通常需要大量的结构工程和技能知识。因此，神经网络结构搜索(Neural Architecture Search，NAS)应运而生，其主要任务就是把人工设计神经网络结构的过程自动化。The present invention uses a neural structure search method to construct a neural network model for image reconstruction. For different tasks, designing a neural network structure usually requires a lot of structural engineering and technical knowledge. Therefore, Neural Architecture Search (NAS) emerged at the historic moment, and its main task is to automate the process of artificial neural network structure design.
神经网络结构搜索(Neural Architecture Search，NAS)的三个主要组成部分是：The three main components of Neural Architecture Search (NAS) are:
1.搜索空间。搜索空间描述了潜在可能的神经网络架构集合。搜索空间是针对应用而专门设计的，典型的如针对计算机视觉任务的卷积网络空间，或针对语言建模任务的递归神经网络空间。因此，NAS方法并非完全自动化，因为这些搜索空间的设计从根本上依赖于人为设计的架构作为起点。即便如此，仍然存在许多架构参数需要决策。实际上，在这些搜索空间中需要考虑的潜在架构的数量通常超过10的10次方。1. Search the space. The search space describes the set of potentially possible neural network architectures. The search space is specifically designed for applications, such as convolutional network space for computer vision tasks, or recurrent neural network space for language modeling tasks. Therefore, the NAS method is not completely automated, because the design of these search spaces fundamentally relies on a human-designed architecture as a starting point. Even so, there are still many architectural parameters that need to be decided. In fact, the number of potential architectures that need to be considered in these search spaces usually exceeds 10 to the power of 10.
对于本发明，神经网络用于快速磁共振图像重建，因此，搜索空间中的神经网络架构限定为卷积神经网络，相应的，搜索空间中包括卷积神经网络的不同的计算单元，如：卷积结构(convolutional architectures)、整流线性单元(ReLU)、批量归一化(batch normalization)、跳跃连接(skip connections)等组成深度学习网络必须的单元及扩展结构。For the present invention, the neural network is used for fast magnetic resonance image reconstruction. Therefore, the neural network architecture in the search space is limited to the convolutional neural network. Correspondingly, the search space includes different computing units of the convolutional neural network, such as: Convolutional architectures, rectified linear units (ReLU), batch normalization, skip connections, etc. form the necessary units and extended structures for deep learning networks.
2.优化方法。优化方法用于确定如何浏览搜索空间以便找到一个好的架构。这里最基本的方法是随机搜索，同时还引入了各种自适应方法，例如强化学习，进化搜索，基于梯度的优化和贝叶斯优化。虽然这些自适应方法在选择评估哪些架构上存在些许不同，但它们都试图搜索倾向于更可能表现良好的网络架构。所有这些方法都具有在传统超参数优化任务的情境下的对应方法。2. Optimization method. The optimization method is used to determine how to browse the search space in order to find a good architecture. The most basic method here is random search, and various adaptive methods are also introduced, such as reinforcement learning, evolutionary search, gradient-based optimization and Bayesian optimization. Although these adaptive methods differ slightly in choosing which architectures to evaluate, they all try to search for network architectures that tend to be more likely to perform well. All these methods have corresponding methods in the context of traditional hyperparameter optimization tasks.
本发明使用强化学习方法进行搜索，具体的，使用一个循环神经网络模型来执行优化方法，本发明中，将这部分组件称为控制器。The present invention uses a reinforcement learning method to search. Specifically, a recurrent neural network model is used to perform the optimization method. In the present invention, this part of the component is called a controller.
3.评估方法。该组件测量优化方法考虑的每种结构的表现。最简单，但计算量最大的选择是完整的训练一个网络。3. Evaluation method. The component measures the performance of each structure considered by the optimization method. The simplest, but the most computationally intensive option is to train a network completely.
本步骤中，将神经网络的类型限定为卷积神经网络，根据搜索空间中的计算单元以及计算单元之间的连接关系等参数，通过初始参数可以构建一个初始结构的卷积神经网络。In this step, the type of neural network is limited to a convolutional neural network, and a convolutional neural network with an initial structure can be constructed through initial parameters based on the calculation units in the search space and the connection relationship between the calculation units.
S3：使用所述样本数据和所述标签数据对所述第一网络模型进行训练，得到训练好的第一网络模型；S3: Use the sample data and the label data to train the first network model to obtain a trained first network model;
本步骤使用步骤S1中获取的训练数据对搜索空间中的神经网络模型进行训练。神经网络结构搜索进行时，往往需要一个初始结构，然后使用训练数据对该初始结构进行训练，直到收敛。This step uses the training data obtained in step S1 to train the neural network model in the search space. When the neural network structure search is performed, an initial structure is often required, and then the initial structure is trained using training data until convergence.
对于第一网络模型进行训练的收敛公式如下：The convergence formula for training the first network model is as follows:
Figure PCTCN2019125460-appb-000001
Figure PCTCN2019125460-appb-000001
式中：F表示子网络端到端之间的映射关系，F(x m,n；Θ)得到的是网络的输出；Θ表示需要学习的参数；x m,n表示网络的输入；y m,n表示网络的输出标签。 Where: F represents the end-to-end mapping relationship between the sub-networks, F(x m,n ; Θ) gets the output of the network; Θ represents the parameters that need to be learned; x m,n represents the input of the network; y m ,n represents the output label of the network.
对于在神经网络结构搜索中每个搜索到的神经网络，均需要使用训练 数据进行训练，得到训练好的第一网络模型。For each neural network searched in the neural network structure search, training data needs to be used for training to obtain the trained first network model.
S4：使用测试数据对所述训练好的第一网络模型进行测试，得到误差结果；S4: Use the test data to test the trained first network model to obtain an error result;
本步骤用于对搜索空间中不同的神经网络结构进行评估。对于不同的应用场景，可以使用不同的评估参数。例如，有些应用场景更关注神经网络模型的计算性能，因此可以使用计算延迟参数进行评估；有些应用场景更关注神经网络模型处理结果的精确度，因此可以选择误差参数进行评估。具体到本发明，磁共振图像重建更关注重建结果的精度，因此，仅选择误差参数来对不同的网络结构进行评估。This step is used to evaluate different neural network structures in the search space. For different application scenarios, different evaluation parameters can be used. For example, some application scenarios pay more attention to the calculation performance of the neural network model, so the calculation delay parameter can be used for evaluation; some application scenarios pay more attention to the accuracy of the neural network model processing result, so the error parameter can be selected for evaluation. Particularly in the present invention, the reconstruction of magnetic resonance images pays more attention to the accuracy of the reconstruction results. Therefore, only error parameters are selected to evaluate different network structures.
误差评估使用和训练数据集不同的测试数据。和训练数据的产生方法相同，测试数据同样使用全采样的k空间磁共振数据，由全采样磁共振数据重建得到准确的磁共振图像，通过二次采样得到欠采样数据作为测试数据中的输入数据。将欠采样数据输入训练好的第一网络模型，得到输出图像，将输出图像和由全采样数据重建的准确图像进行对比，得到重建图像的误差。误差计算可以使用均方差，或其他现有技术中公知的误差计算方法。测试数据集同样包括大量不同的磁共振图像，最终由多幅磁共振图像得到测试误差结果。The error evaluation uses test data that is different from the training data set. The same as the training data generation method, the test data also uses the fully sampled k-space magnetic resonance data, the accurate magnetic resonance image is reconstructed from the fully sampled magnetic resonance data, and the under-sampled data is obtained by sub-sampling as the input data in the test data. . Input the under-sampled data into the trained first network model to obtain the output image, and compare the output image with the accurate image reconstructed from the full-sampled data to obtain the error of the reconstructed image. The error calculation can use the mean square error, or other error calculation methods known in the prior art. The test data set also includes a large number of different magnetic resonance images, and finally the test error results are obtained from multiple magnetic resonance images.
S5：使用强化学习算法和所述误差结果，在所述搜索空间中找到所述网络拓扑参数的最优解，将所述最优解对应的所述训练好的第一网络模型确定为所述用于磁共振图像快速重建的神经网络模型。S5: Use a reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and determine the trained first network model corresponding to the optimal solution as the A neural network model for rapid reconstruction of magnetic resonance images.
如图3所示，一般将执行强化学习的部分称为控制器，控制器基于搜索空间生成子网络并对制作好的训练样本进行训练直至收敛，之后在验证集上进行测试得到相应的误差结果，该结果被反馈至控制器，控制器根据结果做出相应的调整重新生成一个子网络，再次训练、测试及反馈，如此反复得到最佳结果。这里控制器的调整过程采用强化学习来训练。此外，如果条件允许采用分布式并行训练，可以同时设置多个控制器并生成多个子网络，能大大提高神经结构搜索的效率与性能。As shown in Figure 3, the part that performs reinforcement learning is generally called the controller. The controller generates a sub-network based on the search space and trains the prepared training samples until convergence, and then tests on the verification set to obtain the corresponding error results , The result is fed back to the controller, and the controller makes corresponding adjustments based on the result to regenerate a sub-network, train, test and feedback again, and repeat the process to get the best result. The adjustment process of the controller here uses reinforcement learning to train. In addition, if conditions permit the use of distributed parallel training, multiple controllers can be set up at the same time and multiple sub-networks can be generated, which can greatly improve the efficiency and performance of neural structure search.
实施例二Example two
图4示出了本发明的另一个实施例的示意图。Fig. 4 shows a schematic diagram of another embodiment of the present invention.
如图4所示，本发明的第二实施例提供了一种使用神经网络模型进行 磁共振图像重建方法，具体包括：As shown in Figure 4, the second embodiment of the present invention provides a method for reconstructing a magnetic resonance image using a neural network model, which specifically includes:
获取目标对象的欠采样磁共振数据；Obtain under-sampled magnetic resonance data of the target object;
使用前述实施例中确定的神经网络模型对所述欠采样磁共振数据进行重建，得到所述目标对象的磁共振图像。The neural network model determined in the foregoing embodiment is used to reconstruct the under-sampled magnetic resonance data to obtain a magnetic resonance image of the target object.
获取目标对象的欠采样磁共振数据的步骤通过磁共振设备对人体进行欠采样的磁共振扫描，常用的磁共振快速成像一般使用的欠采样方法包括径向轨迹和螺旋轨迹。本发明中，可以使用更高的采样加速倍率进行欠采样，以得到更快的采样速度。The step of obtaining the under-sampled magnetic resonance data of the target object is to perform under-sampling magnetic resonance scanning of the human body through a magnetic resonance device. Commonly used under-sampling methods for rapid magnetic resonance imaging generally include radial trajectories and spiral trajectories. In the present invention, a higher sampling acceleration rate can be used for under-sampling to obtain a faster sampling speed.
进行图像重建时，由于使用的神经网络模型是在神经网络结构搜索中以及训练好的最优网络，所以可以直接使用该网络进行图像重建，得到输出的磁共振图像。When performing image reconstruction, since the neural network model used is in the neural network structure search and the optimal network trained, the network can be used directly for image reconstruction to obtain the output magnetic resonance image.
实施例三Example three
如附图5所示，本发明的实施例三提供一种神经网络模型确定装置，模型确定装置可以是运行于终端中的一个计算机程序(包括程序代码)。该模型训练装置可以执行实施例一中的模型确定方法，具体包括：As shown in FIG. 5, the third embodiment of the present invention provides a neural network model determination device. The model determination device may be a computer program (including program code) running in a terminal. The model training device can execute the model determination method in the first embodiment, which specifically includes:
训练数据获取模块，用于获取用于模型训练的样本数据和标签数据；The training data acquisition module is used to acquire sample data and label data for model training;
搜索空间模块，用于基于神经网络模型的拓扑结构相关的网络拓扑参数，构建搜索空间；根据所述搜索空间中的所述网络拓扑参数，建立对应的第一网络模型；The search space module is used to construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
子网络训练模块，用于使用所述样本数据和所述标签数据对所述第一网络模型进行训练，得到训练好的第一网络模型；A sub-network training module, configured to use the sample data and the label data to train the first network model to obtain a trained first network model;
误差计算模块，用于使用测试数据对所述训练好的第一网络模型进行测试，得到误差结果；An error calculation module, configured to use test data to test the trained first network model to obtain an error result;
控制器模块，用于使用强化学习算法和所述误差结果，在所述搜索空间中找到所述网络拓扑参数的最优解，将所述最优解对应的所述训练好的第一网络模型确定为所述用于磁共振图像快速重建的神经网络模型。The controller module is configured to use the reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and to correspond the optimal solution to the trained first network model Determined to be the neural network model used for rapid reconstruction of magnetic resonance images.
模型训练装置中的各个单元可以分别或全部合并为一个或若干个另外的单元来构成，或者其中的某个(些)单元还可以再拆分为功能上更小的多个单元来构成，这可以实现同样的操作，而不影响本发明的实施例的技术 效果的实现。上述单元是基于逻辑功能划分的，在实际应用中，一个单元的功能也可以由多个单元来实现，或者多个单元的功能由一个单元实现。在本发明的其它实施例中，基于模型训练装置也可以包括其它单元，在实际应用中，这些功能也可以由其它单元协助实现，并且可以由多个单元协作实现。Each unit in the model training device can be separately or completely combined into one or several other units to form, or some of the units can be further divided into functionally smaller units to form multiple units. The same operation can be achieved without affecting the realization of the technical effects of the embodiments of the present invention. The above-mentioned units are divided based on logical functions. In practical applications, the function of one unit may also be realized by multiple units, or the functions of multiple units may be realized by one unit. In other embodiments of the present invention, the model-based training device may also include other units. In practical applications, these functions may also be implemented with the assistance of other units, and may be implemented by multiple units in cooperation.
根据本发明的另一个实施例，可以通过在包括中央处理单元(CPU)、随机存取存储介质(RAM)、只读存储介质(ROM)等处理元件和存储元件的例如计算机的通用计算设备上运行能够执行实施例一中相应方法所涉及的各步骤的计算机程序(包括程序代码)，来构造如附图5中所示的模型训练装置设备，以及来实现本发明实施例的神经网络模型确定方法。所述计算机程序可以记载于例如计算机可读记录介质上，并通过计算机可读记录介质装载于上述计算设备中，并在其中运行。According to another embodiment of the present invention, a general-purpose computing device such as a computer including a central processing unit (CPU), a random access storage medium (RAM), a read-only storage medium (ROM) and other processing elements and storage elements Run a computer program (including program code) that can execute the steps involved in the corresponding method in the first embodiment to construct the model training device as shown in FIG. 5, and to implement the neural network model determination of the embodiment of the present invention method. The computer program may be recorded on, for example, a computer-readable recording medium, and loaded into the above-mentioned computing device through the computer-readable recording medium, and run in it.
实施例四Example four
本发明的实施例四提供一种磁共振设备，包括：The fourth embodiment of the present invention provides a magnetic resonance device, including:
一个或多个处理器；One or more processors;
存储装置，用于存储一个或多个程序；Storage device for storing one or more programs;
当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如实施例二中所述的磁共振快速图像重建方法。When the one or more programs are executed by the one or more processors, the one or more processors implement the method for rapid magnetic resonance image reconstruction as described in the second embodiment.
图6中，设备包括处理器201、存储器202、输入装置203以及输出装置204；设备中处理器201的数量可以是一个或多个，图6中以一个处理器201为例；设备中的处理器201、存储器202、输入装置203以及输出装置204可以通过总线或其他方式连接，图6中以通过总线连接为例。In FIG. 6, the device includes a processor 201, a memory 202, an input device 203, and an output device 204; the number of processors 201 in the device can be one or more. In FIG. 6, one processor 201 is taken as an example; processing in the device The device 201, the memory 202, the input device 203, and the output device 204 may be connected by a bus or other methods. In FIG. 6, the connection by a bus is taken as an example.
存储器202作为一种计算机可读存储介质，可用于存储软件程序、计算机可执行程序以及模块，如本发明实施例一中的神经网络模型确定方法对应的程序指令/模块，或者如本发明实施例二中的磁共振图像重建算法对应的程序指令/模块。处理器201通过运行存储在存储器202中的软件程序、指令以及模块，从而执行设备的各种功能应用以及数据处理，即实现上述的磁共振图像重建方法。As a computer-readable storage medium, the memory 202 can be used to store software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the neural network model determination method in the first embodiment of the present invention, or as in the embodiment of the present invention. Program instructions/modules corresponding to the magnetic resonance image reconstruction algorithm in the second. The processor 201 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 202, that is, realizes the above-mentioned magnetic resonance image reconstruction method.
存储器202可主要包括存储程序区和存储数据区，其中，存储程序区 可存储操作***、至少一个功能所需的应用程序；存储数据区可存储根据终端的使用所创建的数据等。此外，存储器202可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他非易失性固态存储器件。在一些实例中，存储器202可进一步包括相对于处理器201远程设置的存储器，这些远程存储器可以通过网络连接至设备。上述网络的示例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。The memory 202 may mainly include a program storage area and a data storage area. The program storage area may store an operating system and an application program required by at least one function; the data storage area may store data created according to the use of the terminal. In addition, the memory 202 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some examples, the memory 202 may further include a memory remotely provided with respect to the processor 201, and these remote memories may be connected to the device through a network. Examples of the foregoing network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
输入装置203可用于接收输入的数字或字符信息，以及产生与设备的用户设置以及功能控制有关的键信号输入。The input device 203 can be used to receive input digital or character information, and generate key signal input related to user settings and function control of the device.
输出装置204可包括显示屏等显示设备，例如，用户终端的显示屏。The output device 204 may include a display device such as a display screen, for example, a display screen of a user terminal.
实施例五Example five
本发明的实施例五提供一种计算机存储介质，所述计算机存储介质存储有一条或多条第一指令，所述一条或多条第一指令适于由处理器加载并执行前述实施例中的模型训练方法；或者，所述计算机存储介质存储有一条或多条第二指令，所述一条或多条第二指令适于由所述处理器加载并执行前述实施例中的神经网络确定方法或图像重建方法。The fifth embodiment of the present invention provides a computer storage medium, the computer storage medium stores one or more first instructions, and the one or more first instructions are suitable for being loaded by a processor and executed in the foregoing embodiments. Model training method; or, the computer storage medium stores one or more second instructions, and the one or more second instructions are suitable for being loaded by the processor and executing the neural network determination method in the foregoing embodiment or Image reconstruction method.
本发明各实施例方法中的步骤可根据实际需要进行顺序调整、合并和删减。The steps in the method of each embodiment of the present invention can be adjusted, merged, and deleted in order according to actual needs.
本发明各实施例装置中的单元可根据实际需要进行合并、划分和删减。The units in the devices in each embodiment of the present invention can be combined, divided, and deleted according to actual needs.
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序可以存储于一计算机可读存储介质中，存储介质包括只读存储器(Read-Only Memory，ROM)、随机存储器(Random Access Memory，RAM)、可编程只读存储器(Programmable Read-only Memory，PROM)、可擦除可编程只读存储器(Erasable Programmable Read Only Memory，EPROM)、一次可编程只读存储器(One-time Programmable Read-Only Memory，OTPROM)、电子抹除式可复写只读存储器(Electrically-Erasable Programmable Read-Only Memory，EEPROM)、只读光盘(Compact Disc  Read-Only Memory，CD-ROM)或其他光盘存储器、磁盘存储器、磁带存储器、或者能够用于携带或存储数据的计算机可读的任何其他介质。A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above-mentioned embodiments can be completed by a program instructing relevant hardware. The program can be stored in a computer-readable storage medium. The storage medium includes read-only Memory (Read-Only Memory, ROM), Random Access Memory (RAM), Programmable Read-only Memory (PROM), Erasable Programmable Read Only Memory, EPROM), One-time Programmable Read-Only Memory (OTPROM), Electronically-Erasable Programmable Read-Only Memory (EEPROM), CD-ROM (Compact Disc) Read-Only Memory, CD-ROM) or other optical disk storage, magnetic disk storage, tape storage, or any other computer-readable medium that can be used to carry or store data.
以上结合附图详细说明了本发明的技术方案，以上所述仅为本发明的优选实施例而已，并不用于限制本发明，对于本领域的技术人员来说，本发明可以有各种更改和变化。凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。The technical solutions of the present invention are described in detail above with reference to the accompanying drawings. The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and changes. Variety. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (10)

  1. 一种用于磁共振图像快速重建的神经网络模型的确定方法，其特征在于，包括：A method for determining a neural network model for rapid reconstruction of magnetic resonance images, which is characterized in that it comprises:
    S1：获取用于模型训练的样本数据和标签数据；S1: Obtain sample data and label data for model training;
    S2：基于神经网络模型的拓扑结构相关的网络拓扑参数，构建搜索空间；根据所述搜索空间中的所述网络拓扑参数，建立对应的第一网络模型；S2: construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
    S3：使用所述样本数据和所述标签数据对所述第一网络模型进行训练，得到训练好的第一网络模型；S3: Use the sample data and the label data to train the first network model to obtain a trained first network model;
    S4：使用测试数据对所述训练好的第一网络模型进行测试，得到误差结果；S4: Use the test data to test the trained first network model to obtain an error result;
    S5：使用强化学习算法和所述误差结果，在所述搜索空间中找到所述网络拓扑参数的最优解，将所述最优解对应的所述训练好的第一网络模型确定为所述用于磁共振图像快速重建的神经网络模型。S5: Use a reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and determine the trained first network model corresponding to the optimal solution as the A neural network model for rapid reconstruction of magnetic resonance images.
  2. 如权利要求1所述的方法，其特征在于：The method of claim 1, wherein:
    所强化学习算法为基于循环神经网络模型的算法。The reinforcement learning algorithm is an algorithm based on a recurrent neural network model.
  3. 如权利要求1所述的方法，其特征在于，所述第一网络模型为卷积神经网络模型(CNN)。The method of claim 1, wherein the first network model is a convolutional neural network model (CNN).
  4. 如权利要求3所述的方法，其特征在于，所述网络拓扑参数中的计算单元包括：卷积结构、整流线性单元(ReLU)、批量归一化(batch normalization)和跳跃连接。The method according to claim 3, wherein the calculation unit in the network topology parameter includes: a convolution structure, a rectified linear unit (ReLU), batch normalization, and skip connection.
  5. 如权利要求1所述的方法，其特征在于，所述步骤S1进一步包括：The method according to claim 1, wherein the step S1 further comprises:
    S11：获取用于模型训练的全采样磁共振数据；S11: Obtain full-sampling magnetic resonance data for model training;
    S12：对所述全采样磁共振数据进行重建，得到磁共振图像，作为训练使用的标签数据；S12: Reconstruct the full-sampling magnetic resonance data to obtain a magnetic resonance image, which is used as tag data for training;
    S13：对所述全采样磁共振数据进行重采样，得到欠采样数据，作为训练使用的样本数据。S13: Re-sampling the full-sampling magnetic resonance data to obtain under-sampling data as sample data for training.
  6. 如权利要求2所述的方法，其特征在于，所述循环神经网络模型为 长短时记忆网络模型(LSTM)。The method of claim 2, wherein the recurrent neural network model is a long short-term memory network model (LSTM).
  7. 一种磁共振图像快速重建方法，其特征在于，包括：A method for rapid reconstruction of magnetic resonance images, which is characterized in that it comprises:
    获取目标对象的欠采样磁共振数据；Obtain under-sampled magnetic resonance data of the target object;
    使用所述权利要求1-6之一中确定的神经网络模型对所述欠采样磁共振数据进行重建，得到磁共振图像。The neural network model determined in one of the claims 1-6 is used to reconstruct the under-sampled magnetic resonance data to obtain a magnetic resonance image.
  8. 一种神经网络模型的确定装置，其特征在于，包括：A device for determining a neural network model is characterized in that it comprises:
    训练数据获取模块，用于获取用于模型训练的样本数据和标签数据；The training data acquisition module is used to acquire sample data and label data for model training;
    搜索空间模块，用于基于神经网络模型的拓扑结构相关的网络拓扑参数，构建搜索空间；根据所述搜索空间中的所述网络拓扑参数，建立对应的第一网络模型；The search space module is used to construct a search space based on network topology parameters related to the topology of the neural network model; establish a corresponding first network model according to the network topology parameters in the search space;
    子网络训练模块，用于使用所述样本数据和所述标签数据对所述第一网络模型进行训练，得到训练好的第一网络模型；A sub-network training module, configured to use the sample data and the label data to train the first network model to obtain a trained first network model;
    误差计算模块，用于使用测试数据对所述训练好的第一网络模型进行测试，得到误差结果；An error calculation module, configured to use test data to test the trained first network model to obtain an error result;
    控制器模块，用于使用强化学习算法和所述误差结果，在所述搜索空间中找到所述网络拓扑参数的最优解，将所述最优解对应的所述训练好的第一网络模型确定为所述用于磁共振图像快速重建的神经网络模型。The controller module is configured to use the reinforcement learning algorithm and the error result to find the optimal solution of the network topology parameter in the search space, and to correspond the optimal solution to the trained first network model Determined to be the neural network model used for rapid reconstruction of magnetic resonance images.
  9. 一种磁共振图像快速重建装置，包括：A rapid magnetic resonance image reconstruction device, including:
    数据采集模块，用于获取目标对象的欠采样磁共振数据；The data acquisition module is used to acquire the under-sampled magnetic resonance data of the target object;
    图像重建模块，使用所述权利要求1-6之一中确定的神经网络模型对所述欠采样磁共振数据进行重建，得到磁共振图像。The image reconstruction module uses the neural network model determined in one of the claims 1-6 to reconstruct the under-sampled magnetic resonance data to obtain a magnetic resonance image.
  10. 一种包含计算机可执行指令的存储介质，其特征在于，所述计算机可执行指令在由计算机处理器执行时用于执行如权利要求1-7中任一所述的方法。A storage medium containing computer-executable instructions, wherein the computer-executable instructions are used to execute the method according to any one of claims 1-7 when the computer-executable instructions are executed by a computer processor.
PCT/CN2019/125460 2019-12-15 2019-12-15 Fast magnetic resonance imaging method and apparatus based on neural architecture search WO2021119875A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/125460 WO2021119875A1 (en) 2019-12-15 2019-12-15 Fast magnetic resonance imaging method and apparatus based on neural architecture search

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/125460 WO2021119875A1 (en) 2019-12-15 2019-12-15 Fast magnetic resonance imaging method and apparatus based on neural architecture search

Publications (1)

Publication Number Publication Date
WO2021119875A1 true WO2021119875A1 (en) 2021-06-24

Family

ID=76476509

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/125460 WO2021119875A1 (en) 2019-12-15 2019-12-15 Fast magnetic resonance imaging method and apparatus based on neural architecture search

Country Status (1)

Country Link
WO (1) WO2021119875A1 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113722902A (en) * 2021-08-23 2021-11-30 西安电子科技大学 Neural network-based optimal fit parameter estimation method for antenna of shaped reflector
CN115760777A (en) * 2022-11-21 2023-03-07 脉得智能科技(无锡)有限公司 Hashimoto's thyroiditis diagnostic system based on neural network structure search
CN116628457A (en) * 2023-07-26 2023-08-22 武汉华康世纪医疗股份有限公司 Harmful gas detection method and device in operation of magnetic resonance equipment

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120099774A1 (en) * 2010-10-21 2012-04-26 Mehmet Akcakaya Method For Image Reconstruction Using Low-Dimensional-Structure Self-Learning and Thresholding
CN109325985A (en) * 2018-09-18 2019-02-12 上海联影智能医疗科技有限公司 MR image reconstruction method, apparatus and computer readable storage medium
CN109712208A (en) * 2018-12-13 2019-05-03 深圳先进技术研究院 Big visual field magnetic resonance imaging image rebuilding method and device based on deep learning
CN110570486A (en) * 2019-08-23 2019-12-13 清华大学深圳研究生院 Under-sampling nuclear magnetic resonance image reconstruction method based on deep learning

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120099774A1 (en) * 2010-10-21 2012-04-26 Mehmet Akcakaya Method For Image Reconstruction Using Low-Dimensional-Structure Self-Learning and Thresholding
CN109325985A (en) * 2018-09-18 2019-02-12 上海联影智能医疗科技有限公司 MR image reconstruction method, apparatus and computer readable storage medium
CN109712208A (en) * 2018-12-13 2019-05-03 深圳先进技术研究院 Big visual field magnetic resonance imaging image rebuilding method and device based on deep learning
CN110570486A (en) * 2019-08-23 2019-12-13 清华大学深圳研究生院 Under-sampling nuclear magnetic resonance image reconstruction method based on deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
XIAO TAOHUI, GUO JIAN;ZHAO TAO;WANG SHANSHAN;LIANG DONG: "Fast magnetic resonance imaging with deep learning and design of undersampling trajectory", JOURNAL OF IMAGE AND GRAPHICS, ZHONGGUO TUXIANG TUXING XUEHUI, CN, vol. 23, no. 2, 1 January 2018 (2018-01-01), CN, pages 194 - 208, XP055822195, ISSN: 1006-8961, DOI: 10.11834/jig.170274 *
XIAO TAOHUI: "Research on Fast Magnetic Resonance Imaging Based on Deep Learning", BASIC SCIENCES, CHINA MASTER’S THESES FULL-TEXT DATABASE, 1 June 2018 (2018-06-01), XP055822194 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113722902A (en) * 2021-08-23 2021-11-30 西安电子科技大学 Neural network-based optimal fit parameter estimation method for antenna of shaped reflector
CN115760777A (en) * 2022-11-21 2023-03-07 脉得智能科技(无锡)有限公司 Hashimoto's thyroiditis diagnostic system based on neural network structure search
CN115760777B (en) * 2022-11-21 2024-04-30 脉得智能科技(无锡)有限公司 Hashimoto thyroiditis diagnosis system based on neural network structure search
CN116628457A (en) * 2023-07-26 2023-08-22 武汉华康世纪医疗股份有限公司 Harmful gas detection method and device in operation of magnetic resonance equipment
CN116628457B (en) * 2023-07-26 2023-09-29 武汉华康世纪医疗股份有限公司 Harmful gas detection method and device in operation of magnetic resonance equipment

Similar Documents

Publication Publication Date Title
WO2021119875A1 (en) Fast magnetic resonance imaging method and apparatus based on neural architecture search
CN107610194B (en) Magnetic resonance image super-resolution reconstruction method based on multi-scale fusion CNN
CN107680657B (en) Medical scanner self-learning optimized clinical protocol and image acquisition
JP2023169278A (en) System and method for improving magnetic resonance imaging by using deep learning
US11023785B2 (en) Sparse MRI data collection and classification using machine learning
EP3785231A1 (en) Image enhancement using generative adversarial networks
US11250543B2 (en) Medical imaging using neural networks
CN110163260A (en) Image-recognizing method, device, equipment and storage medium based on residual error network
US10878311B2 (en) Image quality-guided magnetic resonance imaging configuration
US10627470B2 (en) System and method for learning based magnetic resonance fingerprinting
CN111870245B (en) Cross-contrast-guided ultra-fast nuclear magnetic resonance imaging deep learning method
CN114450599B (en) Maxwell Wei Binghang imaging
US10871535B2 (en) Magnetic resonance fingerprinting optimization in magnetic resonance imaging
CN112767504A (en) System and method for image reconstruction
CN110992440B (en) Weak supervision magnetic resonance rapid imaging method and device
US20200327379A1 (en) Fastestimator healthcare ai framework
CN111063000B (en) Magnetic resonance rapid imaging method and device based on neural network structure search
US10527695B2 (en) Systems and methods for efficient magnetic resonance fingerprinting scheduling
CN110189302A (en) Brain image analysis method, computer equipment and readable storage medium storing program for executing
CN111815558A (en) Medical image processing system, method and computer storage medium
Huang et al. Enhanced MRI reconstruction network using neural architecture search
CN112990359B (en) Image data processing method, device, computer and storage medium
WO2021184195A1 (en) Medical image reconstruction method, and medical image reconstruction network training method and apparatus
WO2021114098A1 (en) Weakly supervised fast magnetic resonance imaging method and apparatus
WO2024093083A1 (en) Magnetic resonance weighted image synthesis method and apparatus based on variational autoencoder

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: 19956891

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19956891

Country of ref document: EP

Kind code of ref document: A1

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 1205A DATED 20.01.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 19956891

Country of ref document: EP

Kind code of ref document: A1