Apparatus for acquiring magnetic resonance images based on a deep learning model and control method thereof

Abstract

This invention relates to an apparatus and control method for acquiring magnetic resonance imaging (MRI) images based on deep learning. One embodiment of the control method includes the following steps: acquiring an MRI signal; applying at least one of a plurality of elements, set for the quality of an MRI image corresponding to the MRI signal, to the MRI signal to acquire a training image corresponding to the MRI image; acquiring a training dataset that includes the MRI image as label data and the acquired training image as input data matching the label data; and enabling a neural network model to learn based on the training dataset and context data corresponding to the training image. The step of acquiring the training image may include the following steps: applying at least one of the plurality of elements in a K-space domain to distort the MRI signal and acquiring the training image based on the distorted MRI signal.

Description

Technical Field

[0001] This invention relates to a deep learning technology in the medical field, specifically to an apparatus and control method for restoring the quality of magnetic resonance imaging based on a deep learning model. Background Technology

[0002] As a means of obtaining information about a patient's internal body for observation and diagnosis, devices such as X-ray imaging, ultrasound diagnostic equipment, computed tomography (CT) scanners, and magnetic resonance imaging (MRI) are used. Among these, MRI imaging is more noteworthy for its ability to avoid patient exposure to radiation, eliminate the need for contrast agents, and achieve high resolution and excellent soft tissue contrast, making it more practical than other imaging techniques.

[0003] Regarding magnetic resonance imaging (MRI) technology, a significant drawback is the long waiting time required to acquire MRI images. To address this, extensive research has been conducted on accelerated imaging techniques as a solution to shorten MRI image acquisition time. However, MRI images acquired using accelerated imaging techniques are often difficult to analyze accurately due to low resolution or noise. In particular, sometimes MRI images miss internal bodily information about the patient.

[0004] As a result, accelerating the acquisition of magnetic resonance imaging (MRI) images and obtaining high-quality MRI images have become long-standing unsolved problems in this field. Artificial intelligence (AI) technology has then been proposed as a solution to these problems. Specifically, this involves using AI models to restore the quality of MRI images acquired through accelerated imaging techniques.

[0005] Therefore, it is necessary for artificial intelligence models to learn effectively, especially by ensuring high-quality input data and a corresponding high-quality labeled training dataset. However, existing input data only includes cases of uniform undersampling (or random undersampling) performed during accelerated imaging, which leads to the performance limits of artificial intelligence models. That is, resolution degradation and noise in MRI images during accelerated imaging can occur for various reasons, and current technology cannot provide a suitable solution for generating a training dataset that summarizes it. Summary of the Invention

[0006] Technical issues

[0007] The present invention aims to solve the problems of the prior art mentioned above. The purpose of the present invention is to provide a device and control method for acquiring magnetic resonance images based on a deep learning model.

[0008] However, the technical problems to be solved by the present invention are not limited to the aforementioned problems, and other problems not mentioned above can be clearly understood from the following description.

[0009] Technical solution

[0010] To address the aforementioned problems, an embodiment of the present invention provides a control method for an apparatus for acquiring magnetic resonance images based on a deep learning model, executed by a computing device including at least one processor. This method includes the following steps: applying at least one of a plurality of elements set for magnetic resonance image quality to a magnetic resonance signal corresponding to the magnetic resonance image to acquire a training image corresponding to the magnetic resonance image; acquiring a training dataset that includes the magnetic resonance image as label data and the acquired training image as input data matching the label data; and enabling a neural network model to learn based on the training dataset and context data corresponding to the training image. The step of acquiring the training image includes the following steps: applying at least one of the plurality of elements to distort the magnetic resonance signal and acquiring the training image based on the distorted magnetic resonance signal.

[0011] Alternatively, the step of acquiring the training images may include the following steps: changing at least one of the type of the applied element or the number of applied elements, repeatedly distorting the magnetic resonance signal, and acquiring multiple training images based on multiple magnetic resonance signals with different distortions, the multiple training images having different qualities corresponding to at least one of the type of the applied element or the number of applied elements.

[0012] Alternatively, the plurality of elements may include at least two of the following: superposition of Gaussian noise, uniform pattern undersampling, random pattern undersampling, Kmax undersampling, elliptical undersampling, and partial Fourier undersampling.

[0013] Alternatively, the method may include the following steps: if the number of training images is less than a preset number, adjust the frequency range of the Kmax undersampling to further distort the magnetic resonance signal, and further acquire the training images based on the further distorted magnetic resonance signal.

[0014] Alternatively, the method may include the following steps: if the number of training images is less than a preset number, adjust the sampling factor of at least one of the uniform mode undersampling, the random mode undersampling, the Kmax undersampling, the elliptical undersampling, and the partial Fourier undersampling to further distort the magnetic resonance signal, and then further acquire the training images based on the further distorted magnetic resonance signal.

[0015] Alternatively, the method may include the following steps: if the number of training images is less than a preset number, the intensity of the Gaussian noise is adjusted and the adjusted Gaussian noise is superimposed to further distort the magnetic resonance signal, and the training images are further acquired based on the further distorted magnetic resonance signal.

[0016] Alternatively, the neural network model may include a dynamic modulation path that connects to intermediate layers of the multiple layers constituting the neural network model and extracts feature information from the context data when it is input.

[0017] Alternatively, the method includes the following steps: identifying scanning parameters corresponding to the distorted magnetic resonance signal, and identifying the identified scanning parameters as context data corresponding to the training image.

[0018] Alternatively, the method may include the following steps: comparing the noise between the magnetic resonance signal and the distorted magnetic resonance signal, identifying the noise variation, and identifying the identified noise variation as context data corresponding to the training image.

[0019] Alternatively, the steps for obtaining the training dataset include the following: if the magnetic resonance image is three-dimensional data, then the first slice among the multiple image slices contained in the training image is set as the first input data; at least one slice among the multiple image slices contained in the training image adjacent to the first slice is set as the second input data; the third slice among the multiple image slices contained in the magnetic resonance image corresponding to the first slice is set as the label data; and the first input data, the second input data, and the label data are set as the training dataset.

[0020] Alternatively, the method may include the following steps: standardizing the training dataset by including at least one of the following scaling adjustments: size, orientation, pixel spacing, and pixel value of the magnetic resonance image and the training image.

[0021] Alternatively, the method may include the following steps: setting multiple restoration scenarios for the magnetic resonance image based on at least one of the type of elements applicable to the magnetic resonance signal or the number of applicable elements; classifying the multiple training images according to the set multiple scenarios; and obtaining a sub-training dataset corresponding to each scenario.

[0022] An embodiment of the present invention provides a method for acquiring magnetic resonance images based on deep learning, executed by a computing device including at least one processor, comprising the following steps: acquiring magnetic resonance images based on an accelerated imaging method; and inputting the acquired magnetic resonance images and context data corresponding to the acquired magnetic resonance images into a learned neural network model to restore the quality of the acquired magnetic resonance images; wherein the magnetic resonance images are acquired based on the accelerated imaging method, using at least one of a plurality of elements applied to the magnetic resonance image quality settings or a magnetic resonance signal superimposed with noise.

[0023] Alternatively, the plurality of elements may include at least one of the following: superposition of Gaussian noise, uniform pattern undersampling, random pattern undersampling, Kmax undersampling, elliptical undersampling, and partial Fourier undersampling.

[0024] Alternatively, the neural network model may include a dynamic modulation path that connects to intermediate layers of the multiple layers constituting the neural network model and extracts feature information from the context data when it is input.

[0025] Alternatively, the method identifies scanning parameters corresponding to the magnetic resonance signal, recognizes the identified scanning parameters as context data, and inputs the identified context data into the dynamic modulation path.

[0026] A computing device for acquiring magnetic resonance images based on deep learning according to an embodiment of the present invention includes: a memory for storing a neural network model; and at least one processor, which applies at least one of a plurality of elements set for the quality of the magnetic resonance image corresponding to the magnetic resonance signal to the magnetic resonance signal corresponding to the magnetic resonance image, to acquire a training image corresponding to the magnetic resonance image, acquire a training dataset that includes the magnetic resonance image as label data and the acquired training image as input data matching the label data, and enables the neural network model to learn based on the training dataset and context data corresponding to the training image; the at least one processor applies at least one of the plurality of elements to distort the magnetic resonance signal, and acquires the training image based on the distorted magnetic resonance signal.

[0027] The effects of the invention

[0028] According to an embodiment of the present invention, a method for acquiring magnetic resonance images based on a deep learning model ensures that learning data including various resolution degradation and noise occurrence cases that occur during various accelerated imaging processes are used as the basis for learning the neural network model. Therefore, the performance of the neural network model can be improved and the quality of low-quality magnetic resonance images acquired through accelerated imaging can be restored more effectively. Attached Figure Description

[0029] Figure 1 This is an example diagram of a computing device for acquiring magnetic resonance images based on a deep learning model, according to an embodiment of the present invention.

[0030] Figure 2 This is a block diagram of a computing device for acquiring magnetic resonance images based on a deep learning model, according to an embodiment of the present invention.

[0031] Figure 3 This is a sequence diagram of a method for controlling a computing device that acquires magnetic resonance images based on a deep learning model, according to an embodiment of the present invention.

[0032] Figure 4a and Figure 4b Several elements used to illustrate an embodiment of the present invention applicable to magnetic resonance signals in the K-space domain.

[0033] Figure 5 This is an example diagram illustrating the acquisition of multiple training images corresponding to magnetic resonance images according to an embodiment of the present invention.

[0034] Figure 6 This is an illustrative diagram that schematically shows the structure of a neural network model according to an embodiment of the present invention.

[0035] Figure 7This is an illustrative diagram showing an embodiment of the present invention of a method for learning a neural network model to obtain three-dimensional magnetic resonance images.

[0036] Figure 8 This is a block diagram of a computing device for acquiring magnetic resonance images based on a deep learning model, according to another embodiment of the present invention. Detailed Implementation

[0037] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings in order to enable those skilled in the art to readily implement the present invention. The embodiments disclosed herein are intended to enable those skilled in the art to utilize or implement the content of the present invention. Therefore, various modifications to the embodiments of the present invention will be apparent to those skilled in the art. That is, the present invention can be implemented in various different forms and is not limited to the following embodiments.

[0038] Throughout this specification, the same or similar reference numerals denote the same or similar elements. Furthermore, reference numerals for parts unrelated to the description of this invention may be omitted from the drawings for the purpose of clearly illustrating the invention.

[0039] The term "or" as used in this invention does not imply exclusive "or" but rather implied "or". That is, unless otherwise specified in this invention or its meaning is ambiguous in the context of the sentence, "X uses A or B" should be understood to mean one of the natural implied substitutions. For example, if not otherwise specified in this invention or its meaning is ambiguous in the context of the sentence, "X uses A or B" can be interpreted as one of the following: X uses A, X uses B, or X uses both A and B.

[0040] The term "and / or" as used in this invention should be understood to refer to or include all possible combinations of more than one of the listed related concepts.

[0041] The terms "comprising" and / or "having" as used in this invention should be understood to mean the presence of a specific feature and / or element. However, the terms "comprising" and / or "having" should be understood to not exclude the presence or addition of more than one other feature, other element, and / or combination thereof.

[0042] Unless otherwise specified in this invention or where the singular form is not clearly indicated in the context of the sentence, the singular should generally be interpreted as including "one or more".

[0043] The term "Nth (N is a natural number)" as used in this invention can be understood as a way of distinguishing the elements of the invention from each other according to a pre-established basis such as a functional viewpoint, a structural viewpoint, or for ease of explanation. For example, elements that perform different functions in this invention can be distinguished as first elements or second elements. However, elements that are substantially the same in technical spirit of this invention but need to be distinguished for the sake of explanation can also be distinguished as first elements or second elements.

[0044] The term "acquisition" as used in this invention can be understood not only as receiving data via a wired or wireless communication network with an external device or system, but also as generating data in an on-device manner.

[0045] On the other hand, the terms "module" or "unit" used in this invention can be understood as referring to an independent functional unit that processes computing resources, such as a computer-related entity, firmware, software or a part thereof, hardware or a part thereof, or a combination of software and hardware. In this case, "module" or "unit" can be a unit composed of a single element, or a unit represented by a combination or set of multiple elements. For example, as a protocol concept, "module" or "unit" can refer to hardware elements of a computing device or a set thereof, an application program that performs a specific function of software, a process implemented by executing software, or a set of instructions for executing a program. Moreover, as a broad concept, "module" or "unit" can refer to the computing device itself that makes up the system, or the application program executed on the computing device. However, the foregoing concepts are merely illustrative, and the concept of "module" or "unit" can be defined in various ways within the scope understandable to those skilled in the art, based on the content of this invention.

[0046] As used in this invention, the term "model" can be understood as a system implemented using mathematical concepts and language to solve a specific problem, a collection of software units used to solve a specific problem, or an abstract model of a process used to solve a specific problem. For example, a neural network "model" can refer to the entire system implemented by a neural network that has acquired the ability to solve problems through learning. In this case, the neural network acquires the ability to solve problems by optimizing the parameters of the connecting nodes or neurons through learning. A neural network "model" can include a single neural network or a collection of neural networks composed of multiple neural networks.

[0047] The term "data" as used in this invention may include images, signals, etc. The term "image" as used in this invention can refer to multidimensional data composed of multiple discrete image elements. That is, the term "image" can be understood as a digital representation of an object that can be seen by the human eye. For example, "image" can refer to multidimensional data in a two-dimensional image composed of elements equivalent to pixels. "Image" can refer to multidimensional data in a three-dimensional image composed of elements equivalent to voxels.

[0048] As used in this invention, "image" can mean multi-dimensional data consisting of multiple discrete image elements (e.g., multiple pixels in a two-dimensional image and multiple voxels in a three-dimensional image). For example, an image may include medical images acquired by medical imaging devices such as magnetic resonance imaging devices, computed tomography (CT) devices, ultrasound imaging devices, or X-ray imaging devices, but the invention is not limited thereto.

[0049] The term "medical image" as used in this invention is a general concept encompassing all forms of images in medical knowledge, including images acquired through various modalities such as visible light cameras, IR cameras, ultrasound, X-rays, CT, MRI, and PET (positron emission tomography).

[0050] The term "picture archiving and communication system" as used in this invention refers to a system that stores, processes, and transmits medical images according to the DICOM (digital imaging and communications in medicine) standard. For example, a "medical image storage and transmission system" can be linked with digital medical imaging equipment and store medical images such as magnetic resonance imaging (MRI) and computed tomography (CT) images according to the DICOM standard. The "medical image storage and transmission system" can transmit medical images to terminals inside and outside the hospital via a communication network. At this time, meta-information such as interpretation results and treatment records can be added to the medical images.

[0051] As used in this invention, the term "object" refers to the subject of the photograph, which may include a person, an animal, or a part thereof. For example, an object may include a part of the body (organ, etc.) or a phantom. A phantom means a substance with a density and equivalent atomic number very close to that of a living organism, and may include a sphere phantom with properties similar to those of a human body.

[0052] Magnetic Resonance Imaging (MRI) is a system that acquires images of a target tomographic region by presenting the intensity of a radio frequency (RF) signal generated on a magnetic field of a specific intensity using a contrast-based approach.

[0053] The MRI system uses a main magnet to create a static magnetic field and aligns the magnetic dipole moments of specific atomic nuclei of the target within this field with the direction of the static magnetic field. A gradient magnetic field coil applies a gradient signal to the static magnetic field, creating a gradient magnetic field that induces different resonance frequencies for different parts of the target. An RF coil irradiates the magnetic resonance signal according to the resonance frequency of the area to be imaged. Furthermore, the RF coil can receive multiple magnetic resonance signals radiated from multiple parts of the target at different resonance frequencies as the gradient magnetic field is formed. The MRI system uses image reconstruction techniques to acquire images from the multiple magnetic resonance signals received in this process. Moreover, the MRI system performs serial or parallel signal processing on the multiple magnetic resonance signals received by the multi-channel RF coil to reconfigure the multiple magnetic resonance signals into image data.

[0054] The foregoing description of terms is intended to aid in understanding the present invention. Therefore, unless explicitly stated otherwise as limiting the scope of the invention, these terms are not used in a way that restricts the technical spirit of the invention.

[0055] Figure 1 This is an example diagram of a computing device 100 for acquiring magnetic resonance images based on a deep learning model, according to an embodiment of the present invention.

[0056] A computing device 100 for acquiring magnetic resonance images based on a deep learning model, according to an embodiment of the present invention, can be a hardware device or part of a hardware device that performs comprehensive data processing and computation, or it can be a computing environment based on software connected by a communication network. For example, the computing device 100 can be a server that performs intensive data processing functions and shares resources, or it can be a client that shares resources through interaction with the server. Moreover, the computing device 100 can also be a cloud system that comprehensively processes data through the interaction of multiple servers and multiple clients. The foregoing is merely an example related to the type of computing device 100, and the type of computing device 100 can be constructed in various ways within the scope of understanding of those skilled in the art based on the content of the present invention. As an example, the computing device 100 may include smartphones, tablet PCs, personal computers, smart TVs, microservers, cloud servers, etc., that process or perform processing functions on magnetic resonance images. As another example, the computing device 100 may also be a magnetic resonance imaging (MRI) device that directly acquires magnetic resonance images.

[0057] Please see Figure 1 The computing device 100 can acquire a training dataset 50 for the neural network model 10 to learn. Specifically, the computing device 100 can acquire multiple magnetic resonance images 30 or magnetic resonance signals 20 corresponding to the magnetic resonance images 30 acquired by the multiple other electronic devices 200 that acquire the magnetic resonance images 30, and acquire the training dataset 50.

[0058] For example, the magnetic resonance signal 20 may be K-space data, and the magnetic resonance image 30 may be a two-dimensional or three-dimensional image obtained by performing an inverse Fourier transform operation on the magnetic resonance signal 20. The computing device 100 may acquire pulse sequence data acquired by each of the other electronic devices 200. Here, the pulse sequence data may include K-space data collected based on a specific pulse sequence used by the other electronic devices 200. The pulse sequence data may include two-dimensional pulse sequence data collected in two-dimensional space or three-dimensional pulse sequence data collected in three-dimensional space.

[0059] Here, the magnetic resonance image 30 and the magnetic resonance signal 20 can be transmitted and received after being included in Digital Imaging and Communications in Medicine (DICOM) data. DICOM stands for Digital Imaging and Communications in Medicine, a general term encompassing various standards used in medical devices for digital image representation and communication. DICOM data can primarily include patient information and media characteristics. For example, the various medical information data contained in DICOM data includes textual information about the patient collected at the medical site and unprocessed media information; the present invention does not particularly limit its format. More specifically, DICOM data may include the patient's biometric information, image information of the patient or treatment site generated at the medical site (e.g., magnetic resonance image 30), and information about the device used to acquire the image.

[0060] On the other hand, the computing device 100 can acquire a training dataset 50 based on the acquired magnetic resonance image 30 and / or magnetic resonance signal 20. Specifically, the computing device 100 can adjust the quality of the magnetic resonance image 30 and / or magnetic resonance signal 20 to acquire a training image 40 with a lower quality than the magnetic resonance image 30. The computing device 100 adjusts at least one of multiple factors that set the quality of the magnetic resonance image 30 and / or magnetic resonance signal 20 to reduce the quality of the magnetic resonance image 30. In particular, selectively combining multiple factors can reduce the quality of the same magnetic resonance image 30 in multiple ways. This can be regarded as adjusting the quality of the magnetic resonance image 30 from multiple aspects or multiple dimensions. In this way, by reducing the quality of the same magnetic resonance image 30 in multiple ways, the computing device 100 can acquire multiple training images 40 for the same magnetic resonance image 30.

[0061] On the other hand, the computing device 100 can enable the neural network model 10 to learn based on the acquired training dataset 50 and the context data of each learning data contained in the corresponding training dataset 50. Here, the context data may be data describing the relationship between the magnetic resonance image 30 that makes up the learning data and the training image 40, as well as the background of the quality degradation of the magnetic resonance image 30. In this way, the computing device 100 enables the neural network model 10 to correctly learn the relationship between the diverse acquired training images 40 and the magnetic resonance images 30 corresponding to the training images 40.

[0062] The following combination Figures 2 to 8 Detailed description of embodiments of the present invention relating thereto.

[0063] Figure 2This is a block diagram of a computing device 100 for acquiring magnetic resonance images 30 based on a deep learning model, according to an embodiment of the present invention. Figure 3 This is a sequence diagram of a method for controlling a computing device 100 that acquires magnetic resonance images 30 based on a deep learning model, according to an embodiment of the present invention.

[0064] Please see Figure 2 The computing device 100 includes at least one processor 110 (hereinafter referred to as "processor"), a communication interface 120, and a memory 130. However, Figure 2 This is merely an example; the computing device 100 may also include other elements for implementing the computing environment. Furthermore, the computing device 100 may also include only a portion of the disclosed elements.

[0065] The processor 110 of one embodiment of the present invention can be understood as comprising hardware and / or software components for performing computations. For example, the processor 110 can read a computer program and perform data processing for machine learning. The processor 110 can process computational processes such as processing input data for machine learning, extracting features for machine learning, and calculating errors based on backpropagation. The processor 110 for performing data processing as described above may include a central processing unit (CPU), a general-purpose graphics processing unit (GPGPU), a tensor processing unit (TPU), an application-specific integrated circuit (ASIC), or a field-programmable gate array (FPGA), etc. The aforementioned types of processors 110 are merely illustrative, and various configurations of the processor 110 can be made within the scope of understanding for those skilled in the art based on the content of this invention.

[0066] The processor 110 is electrically connected to other elements of the computing device 100 (i.e., the communication interface 120 and the memory 130) to control the overall operation of the computing device 100.

[0067] The memory 130 of one embodiment of the present invention can be understood as comprising hardware and / or software components for storing and managing data processed by the computing device 100. That is, the memory 130 can store data of any form generated or determined by the processor 110 and data of any form received by the communication interface 120. For example, the memory 130 may include at least one of the following storage media: flash memory, hard disk, multimedia card micro, card memory, random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, magnetic disk, and optical disk. Furthermore, the memory 130 may also include a database system that controls and manages data according to a preset system. The aforementioned types of memory 130 are merely examples, and the types of memory 130 can be defined in various ways within the scope that can be understood by those skilled in the art based on the content of this invention.

[0068] The memory 130 can manage the data required by the processor 110 for computation, combinations of data, and executable code of the processor 110 in a structured and organized manner. For example, the memory 130 can store the neural network model 10, the training dataset 50, and context data. Moreover, the memory 130 can store the code that drives the neural network model 10 to learn based on the training dataset 50 and context data, the code that enables the neural network model 10 to make inferences according to the purpose of use of the computing device 100 after receiving the magnetic resonance image 30 as input, and the processing data generated by executing the code.

[0069] The communication interface 120 of one embodiment of the present invention can be understood as a component unit for transmitting and receiving data through any form of known wired and wireless communication system. For example, the communication interface 120 can use wired and wireless communication systems such as local area network (LAN), wideband code division multiple access (WCDMA), long term evolution (LTE), wireless broadband internet (WiBro), 5G, ultra-wideband, ZigBee, radio frequency (RF) communication, wireless LAN, wireless fidelity, near field communication (NFC), or Bluetooth for data transmission and reception. The foregoing communication systems are merely examples, and various wired and wireless communication systems for data transmission and reception of the communication interface 120 can also be applied beyond the foregoing examples.

[0070] The communication interface 120 can receive data required for computation by the processor 110 via wired or wireless communication with any system or client. Furthermore, the communication interface 120 can transmit data generated by the processor 110's computation via wired or wireless communication with any system or client. For example, the communication interface 120 can receive medical data via communication with databases within a hospital environment, cloud servers performing tasks such as medical data standardization, or computing devices 100. The communication interface 120 can also transmit the output data of the neural network model 10, as well as intermediate and processed data derived during the processor 110's computation, via communication with the aforementioned databases, servers, or computing devices 100.

[0071] Please see Figure 3According to one embodiment of the present invention, the processor 110 acquires the magnetic resonance signal 20 (step S310). Here, the computing device 100 can acquire the magnetic resonance signal 20 via a magnetic field generated for a target (e.g., a patient), or it can acquire it via a communication interface 120 from multiple other electronic devices 200 that acquire the magnetic resonance image 30. For example, the computing device 100 can acquire the magnetic resonance signal 20 from the pulse sequence data acquired by each of the multiple other electronic devices 200. In this case, the pulse sequence data may include K-space data collected based on a specific pulse sequence used by the other electronic devices 200. The pulse sequence data may include two-dimensional pulse sequence data collected in two-dimensional space or three-dimensional pulse sequence data collected in three-dimensional space.

[0072] On the other hand, the processor 110 can also acquire the magnetic resonance signal 20 after receiving digital imaging and communications in medicine (DICOM) data from other electronic devices 200. Specifically, after extracting the magnetic resonance signal 20 contained in the digital imaging and communications data or extracting the magnetic resonance image 30, the processor 110 can acquire the magnetic resonance signal 20 through a discrete Fourier transform. This can also be applied to... Figure 1 The description is omitted here.

[0073] Then, the processor 110 can apply at least one of a plurality of elements of quality setting for the magnetic resonance image 30 corresponding to the magnetic resonance signal 20 to the magnetic resonance signal 20 to obtain a training image 40 corresponding to the magnetic resonance image 30 (step S320).

[0074] Several factors are set for the quality of the magnetic resonance image 30 corresponding to the magnetic resonance signal 20, and these factors can influence and determine the quality of the magnetic resonance image 30. For example, the quality of the magnetic resonance image 30 can be evaluated based on the resolution of the magnetic resonance image 30 and the level of noise contained in the magnetic resonance image 30. That is, high quality of the magnetic resonance image 30 can mean that the magnetic resonance image 30 has high resolution and low noise. In this case, several factors can be factors that determine the resolution of the magnetic resonance image 30 and the level of noise contained in the magnetic resonance image 30. In particular, factors related to the resolution of the magnetic resonance image 30 can be factors concerning the type of undersampling method for the magnetic resonance signal 20. Undersampling is a technique for acquiring magnetic resonance images by scanning the target but not completely filling the entire K-space region with K-space data, but rather by sampling locally. Undersampling can also be called subsampling.

[0075] The processor 110 can apply at least one of a plurality of elements to the magnetic resonance image 30 signal. That is, the processor 110 applies at least one of a plurality of elements that set the quality of the magnetic resonance image 30 to the magnetic resonance signal 20, thereby degrading the quality of the corresponding magnetic resonance image 30. Then, the processor 110 can acquire the degraded magnetic resonance image 30 as a training image 40 for the neural network model 10 to learn.

[0076] On the other hand, according to one embodiment of the present invention, the processor 110 can apply at least one of a plurality of elements to the magnetic resonance signal 20 in the K-space domain to obtain a training image 40 corresponding to the magnetic resonance image 30. Specifically, the processor 110 can distort the magnetic resonance signal 20 in the K-space domain and obtain the training image 40 based on the distorted magnetic resonance signal 20. As an example, the processor 110 can apply at least one of a plurality of elements to the K-space data contained in the magnetic resonance image 30 signal in the K-space domain to distort the K-space data. The processor 110 can also distort the K-space data by superimposing noise on the K-space data, or by applying an undersampling pattern to select a portion of the K-space data to distort the K-space data.

[0077] Figure 4a and Figure 4b Several elements used to illustrate an embodiment of the present invention applicable to a magnetic resonance signal 20 in the K-space domain. Figure 4a Several elements applicable to the magnetic resonance signal 20 are shown on the plane of the phase encoding direction axis Ky (other than the frequency encoding direction Kx) and the slice encoding selection direction axis Kz.

[0078] Please see Figure 4a The multiple elements may include at least one of Gaussian noise 61, uniform pattern undersampling 62, random pattern undersampling 63, Kmax undersampling 64, elliptical undersampling 65, and partial Fourier undersampling 66.

[0079] Specifically, applying Gaussian noise to the magnetic resonance signal 20 can refer to superimposing Gaussian noise on the K-space data or reducing Gaussian noise. That is, the processor 110 can generate random noise (e.g., Gaussian noise) in the K-space domain and then reduce the noise in the K-space data to distort the magnetic resonance signal. Alternatively, the processor 110 can reduce the random noise (i.e., Gaussian noise) present in the K-space data to distort the magnetic resonance signal.

[0080] Uniform mode undersampling 62 is a method of selecting K-space data in the K-space domain at preset intervals, which may include uniform generalized automatic calibration partial parallel sampling (GRAPPA) mode undersampling and uniform controlled aliasing in parallel imaging results in higher acceleration (CAIPIRINHA) mode undersampling. Random mode undersampling 63 can be a method of randomly selecting K-space data in the K-space domain. Moreover, Kmax undersampling 64 can be performed in all encoding directions (Kx, Ky, and Kz) by selectively omitting high-frequency K-space data located at the edges of the K-space domain. Moreover, elliptic undersampling 65 selects K-space data contained in an elliptical region of the K-space domain and omits the remaining K-space data, which can be a method of maintaining low-frequency data in the center and omitting high-frequency data at the edges. On the other hand, partial Fourier undersampling 66 is a method of selecting K-space data contained in a part of the K-space domain and omitting the remaining K-space data, which can be performed in all encoding directions Kx, Ky, and Kz. The processor 110 can use various undersampling methods as described above to distort the magnetic resonance signal.

[0081] On the other hand, Figure 4b The first magnetic resonance signal 21 shown is superimposed with Gaussian noise 61, and after applying Kmax undersampling 64 in all encoding directions, elliptic undersampling 65 is applied, and partial Fourier undersampling 66 is applied in all encoding directions, causing the first magnetic resonance signal 21 to be distorted. This allows the acquisition of a portion of the K-space data that was missed, a second magnetic resonance signal 22. Then, the processor 110 can acquire the training image 40 based on the second magnetic resonance signal 22. However, Figure 4b The order in which the various elements are applied is merely an example to illustrate the invention, and the invention is not limited thereto.

[0082] On the other hand, the processor 110 can acquire a training image 40 based on the distorted magnetic resonance signal 20. Specifically, the processor 110 can perform an inverse Fourier transform operation on the distorted magnetic resonance signal 20 (i.e., K-space data) to acquire a training image 40 corresponding to the distorted magnetic resonance signal 20. At this time, the processor 110 can acquire the training image 40 corresponding to the distorted magnetic resonance signal 20 based on a parallel imaging method (e.g., the generalized automatic calibration partially parallel sampling technique (Grappa)) and another learned neural network model.

[0083] Figure 5 This is an example diagram illustrating the acquisition of multiple training images 40 corresponding to a magnetic resonance image 30 according to an embodiment of the present invention.

[0084] On the other hand, the processor 110 can repeatedly distort the magnetic resonance signal 20 by changing the type and number of the applied elements and acquire multiple training images 40 based on the multiple magnetic resonance signals 20 that are distorted differently. At this time, the multiple training images 40 can have different qualities corresponding to the type, number, and degree of distortion of the applied elements.

[0085] Processor 110 can acquire multiple training images 40 corresponding to the same magnetic resonance image 30. Specifically, processor 110 can change the type, number, and degree of distortion of the elements applied to the magnetic resonance signal 20, thereby causing the magnetic resonance signal 20 to be repeatedly distorted. Please refer to [link to relevant documentation]. Figure 5 The processor 110 can combine multiple elements (first to seventh elements) to obtain multiple training images 40 corresponding to the magnetic resonance image 30. The processor 110 can apply the first element to the magnetic resonance signal 20 to obtain the first training image 40-1, apply the first and second elements to the magnetic resonance signal 20 to obtain the second training image 40-2, apply the third, sixth, and seventh elements to the magnetic resonance signal 20 to obtain the thirteenth training image 40-13, and apply the first to seventh elements to the magnetic resonance signal 20 to obtain the one hundred and twenty-seventh training image 40-127.

[0086] As previously described, the processor 110 can apply multiple elements (first to seventh elements) to the magnetic resonance signal 20 or selectively combine at least two elements to the magnetic resonance signal 20 to acquire various training images 40. In this case, the multiple training images 40 acquired for the same magnetic resonance image 30 can have different resolutions and noise levels depending on the type and number of elements applied to the magnetic resonance signal 20. That is, the processor 110 sets multiple elements related to the quality of the magnetic resonance image 30 for the magnetic resonance signal 20 and combines various elements to apply to the magnetic resonance signal 20 to obtain learning data including various cases of quality degradation of the same magnetic resonance image 30. Furthermore, the processor 110 ensures that the multiple learning data with varying degrees of quality degradation of the multiple magnetic resonance images enable the neural network model 10 to learn in a way that more effectively restores the magnetic resonance image 30.

[0087] On the other hand, according to one embodiment of the present invention, when the number of training images 40 is less than a preset number, the processor 110 adjusts the frequency range of Kmax undersampling to further distort the magnetic resonance signal 20 and can further acquire training images 40 based on the further distorted magnetic resonance signal 20.

[0088] That is, the processor 110 can determine whether the number of training images 40 is greater than a preset number in order to determine whether sufficient training images 40 for the magnetic resonance image 30 have been ensured. Then, if the processor 110 determines that the number of training images 40 is less than the preset number, it determines that the number of multiple training images 40 acquired for the same magnetic resonance image 30 is less than the preset number. In this case, the processor 110 can further distort the magnetic resonance signal 20 by applying different Kmax undersampling frequency range values ​​to multiple elements. For example, the processor 110 can change the frequency range from a first range to a second range that is larger than the first range. At this time, the processor 110 can of course apply the Kmax undersampling with the changed frequency range along with other elements to the magnetic resonance signal 20. Kmax undersampling omits high-frequency K-space data located at the edge of the K-space domain. The processor 110 adjusts the range of the omitted high-frequency K-space data to further distort the magnetic resonance signal 20, thereby enabling the acquisition of more training images 40.

[0089] At this time, the processor 110 judges the similarity between multiple training images 40 and considers multiple training images 40 with a similarity of more than a preset value as the same training image 40, and determines the number of training images 40. On the other hand, the processor 110 can also reduce the similarity of multiple training images 40 with a similarity of more than a preset value by applying different elements to the magnetic resonance signal 20 or applying different Kmax undersampling frequency ranges, and then re-acquire the corresponding training images 40. In particular, the processor 110 can use multiple elements to distort the magnetic resonance signal through all conceivable combinations and then judge the similarity.

[0090] Furthermore, the processor 110 can also change the Gaussian noise intensity to further distort the magnetic resonance signal 20 when the number of training images 40 is less than a preset number, and then acquire more training images 40 based on the further distorted magnetic resonance signal 20. Specifically, if the processor 110 determines that the number of training images 40 is less than a preset number, it can change the intensity of the Gaussian noise applied to the same magnetic resonance image 30 to further distort the magnetic resonance signal 20, and then acquire more training images 40 based on the further distorted magnetic resonance signal 20. For example, the processor 110 can change the intensity of the Gaussian noise from a first intensity to a second intensity and then apply it to the magnetic resonance signal 20 to obtain a further distorted magnetic resonance signal 20.

[0091] Furthermore, the processor 110 can also modify the undersampling factor to further distort the magnetic resonance signal 20 when the number of training images 40 is less than a preset number, and further acquire training images 40 based on the further distorted magnetic resonance signal 20. Specifically, the processor 110 can also adjust at least one of the uniform mode undersampling factor, random mode undersampling factor, Kmax undersampling factor, elliptic undersampling factor, and partial Fourier undersampling factor to further distort the magnetic resonance signal 20, and further acquire training images 40 based on the further distorted magnetic resonance signal 20. For example, the processor 110 can change the uniform mode undersampling factor from a first value to a second value and apply it to the magnetic resonance signal 20 to acquire the further distorted magnetic resonance signal 20. In this way, the processor 110 can ensure a more diverse range of training images 40 and enable the neural network model 10 to learn, thereby improving the performance of the neural network model 10.

[0092] According to an embodiment of the present invention, the processor 110 acquires a training dataset 50 that includes magnetic resonance images 30 as label data and acquired training images 40 as input data matching the label data (step S330). At this time, the processor 110 can acquire multiple training datasets 50 by matching multiple different input data (training images 40) for the same label data (magnetic resonance images 30).

[0093] On the other hand, according to an embodiment of the present invention, the processor 110 can set multiple restoration scenes for the magnetic resonance image 30 according to the type, quantity and degree of distortion of the elements applicable to the magnetic resonance signal 20, and after classifying the multiple training images 40 according to the multiple restoration scenes, obtain the sub-training dataset 50 corresponding to each restoration scene.

[0094] Specifically, the processor 110 can distort multiple magnetic resonance signals 20 corresponding to multiple magnetic resonance images 30 to acquire multiple training images 40 corresponding to the multiple magnetic resonance images 30. At this time, the processor 110 can set multiple restoration scenes of the magnetic resonance images 30 according to the type, quantity, and degree of distortion of the elements applicable to acquiring the multiple training images 40. Here, the restoration scene can be information describing the process of restoring the quality of the training images 40 by reverse tracking the magnetic resonance images 30.

[0095] Then, the processor 110 can classify the multiple training images 40 contained in the training dataset 50 and the multiple magnetic resonance images 30 matched with the multiple training images 40 according to the reconstruction scene, and identify the number of sub-training datasets 50 corresponding to each reconstruction scene. Then, the processor 110 can further secure the sub-training dataset for reconstruction scenes where the number of sub-training datasets is lower than a predetermined value. That is, the processor 110 can further receive magnetic resonance signals 20 from other electronic devices 200 through the communication interface 120, and further secure the sub-training dataset by applying at least one element corresponding to the reconstruction scene to the further received magnetic resonance signals 20.

[0096] According to one embodiment of the present invention, the processor 110 can perform standardization of the training dataset 50, including scaling adjustments of at least one of the size, orientation, pixel pitch, and pixel values ​​of the magnetic resonance image 30 and the training image 40. Standardization can also be performed via a standardization module connected to the input of the neural network model 10.

[0097] The processor 110 can perform a standardization operation on the magnetic resonance images 30 and training images 40 contained in the training dataset 50 before enabling the neural network model 10 to learn. Here, standardization refers to making the size, orientation, etc. of the multiple magnetic resonance images 30 and the multiple training images 40 consistent, which can be performed to enable the neural network model 10 to learn effectively and further improve the learning effect of the neural network model 10.

[0098] As an example, the processor 110 can adjust the orientation of the multiple magnetic resonance images 30 and the multiple training images 40. Specifically, the processor 110 can align the orientation of the multiple magnetic resonance images 30 and the multiple training images 40 in such a way that the row direction (or vertical direction) and the phase encoding direction are aligned, while the column direction (or horizontal direction) and the frequency encoding direction are aligned. Furthermore, in order to adjust the field of view (FOV) of the multiple magnetic resonance images 30 and the multiple training images 40, which are presented asymmetrically, the processor 110 clips the zero-padding areas of the multiple magnetic resonance images 30 and the multiple training images 40 to make their shapes and sizes consistent. As an example, the multiple magnetic resonance images 30 and the multiple training images 40 can be made to have a consistent rectangular shape. Moreover, the processor 110 can adjust the size of the multiple magnetic resonance images 30 and the multiple training images 40 to make them consistent. Specifically, based on the Lanczos method, the processor 110 can adjust the column size to 1024 when the multiple magnetic resonance images 30 and the multiple training images 40 are two-dimensional images (or when the magnetic resonance signal 20 is two-dimensional sequence data), and adjust it to 768 if it is a three-dimensional pulse sequence, so as to maintain a constant image pixel spacing. Furthermore, the processor 110 can perform normalization operations on the multiple magnetic resonance images 30 and the multiple training images 40. Specifically, the processor 110 can adjust the pixel values ​​to make the range of pixel values ​​of the multiple magnetic resonance images 30 and the multiple training images 40 consistent.

[0099] The aforementioned standardized operations can be performed sequentially, or selectively based on multiple magnetic resonance images 30 and multiple training images 40. In particular, the processor 110 can selectively perform standardized operations based on medical digital imaging and communication (DICOM) data containing magnetic resonance signals 20 (or magnetic resonance images 30), after determining the device information of each magnetic resonance signal 20 (or magnetic resonance image 30) or understanding the size, orientation, shape, etc. of the magnetic resonance signal 20 (or magnetic resonance image 30).

[0100] Figure 6 This is an illustrative diagram that schematically shows the structure of a neural network model 10 according to an embodiment of the present invention.

[0101] According to one embodiment of the present invention, the processor 110 can enable the neural network model 10 to learn based on the training dataset 50 and the context data 70 corresponding to the training image 40 (step S540). The processor 110 can enable the neural network model 10 to learn using the training dataset 50 and the context data 70 as auxiliary input. As an example, the neural network model 10 is a neural network model 10 with a U-Net framework. The neural network model 10 may include network models such as deep neural networks (DNNs), recurrent neural networks (RNNs), bidirectional recurrent deep neural networks (BRDNNs), multilayer perceptrons (MLPs), and convolutional neural networks (CNNs).

[0102] Context data 70 corresponds to the distorted magnetic resonance signal 20, and therefore can correspond to the training image 40. Here, context data 70 can be data that explains the relationship between the magnetic resonance image 30 that makes up the training data and the training image 40, as well as the background of the degraded quality of the magnetic resonance image 30. On the other hand, context data 70 can also be included in the training dataset 50 along with the training image 40 corresponding to context data 70.

[0103] According to one embodiment of the present invention, the neural network model 10 may include a Dynamic Modulation Path (DMP) that connects to the intermediate layers of the multiple layers constituting the neural network model 10 and extracts feature information of the context data 70 when it is input. That is, the context data 70 can be input to the intermediate layers of the neural network model 10 through the dynamic modulation path 12. To extract feature information from the context data 70, the dynamic modulation path 12 may include a fully connected layer and an activation function (e.g., a ReLU function). Alternatively, the extracted context feature information can be integrated into the U-Net framework and function as a convolutional kernel.

[0104] According to one embodiment of the present invention, the processor 110 can identify the scanning parameters of a distorted magnetic resonance signal 20 in the K-space domain and recognize the identified scanning parameters as context data 70 for a training image 40. In particular, the processor 110 can recognize the context data 70 based on at least one element applicable when the magnetic resonance signal 20 is distorted. The processor 110 can identify the values ​​of the scanning parameters corresponding to the magnetic resonance signal 20 that change due to the distortion of the magnetic resonance signal 20 (or the application of at least one element) and recognize the context data 70 based on the identified values ​​of the scanning parameters.

[0105] Furthermore, according to one embodiment of the present invention, the processor 110 identifies the noise variation by comparing the noise between the magnetic resonance signal 20 and the distorted magnetic resonance signal 20 in the K-space domain, and can identify the identified noise variation as context data 70 corresponding to the training image 40. At this time, regarding the superposition of Gaussian noise in multiple elements, the processor 110 can determine the noise variation based on the amount of Gaussian noise superimposed on the K-space data.

[0106] On the other hand, please see Figure 6 The context data 70 can be input into the dynamic modulation path 12 in a one-dimensional row and column configuration. At this time, the context data 70 can be transformed into a one-dimensional row and column configuration corresponding to the aforementioned scanning parameters or a one-dimensional row and column configuration corresponding to the noise variation before being input into the dynamic modulation path (DMP). In particular, it can also be a one-dimensional row and column configuration combining the scanning parameters and the noise variation.

[0107] On the other hand, according to an embodiment of the present invention, based on multiple training images 40 whose quality has been reduced in various ways for multiple magnetic resonance images 30, the processor 110 enables the neural network model 10 to learn so that the same magnetic resonance image 30 can also recover the quality reduced in various ways. The processor 110 can obtain a neural network model 10 with excellent recovery ability, such as existing neural network models, which have learned by applying only a single element (e.g., uniform undersampling) to reduce the quality of the training data.

[0108] On the other hand, according to one embodiment of the present invention, the processor 110 can learn by inputting a training dataset 50, context data 70, and set restoration scene information corresponding to the type, quantity, and distortion level of elements applicable to the magnetic resonance signal 20. In this way, the neural network model 10 can learn to output a restoration scene corresponding to the restoration process while restoring the magnetic resonance image 30.

[0109] Figure 7 This is an illustrative diagram showing a method for enabling a neural network model 10 to learn and acquire a three-dimensional magnetic resonance image 30 according to an embodiment of the present invention.

[0110] According to an embodiment of the present invention, when the magnetic resonance image 30 is three-dimensional data, the processor 110 can set a target image slice (hereinafter referred to as a first image slice) among multiple image slices contained in the training image 40 as target input data (hereinafter referred to as first input data). Then, the processor 110 identifies multiple image slices (hereinafter referred to as second image slices) adjacent to the first slice among the multiple image slices contained in the training image 40 as multiple reference input data (hereinafter referred to as second input data), and sets the image slice corresponding to the first slice among the multiple image slices contained in the magnetic resonance image 30 (hereinafter referred to as third image slice) as label data. Then, the processor 110 can set the first input data, multiple second input data, and label data as training dataset 50. At this time, the first input data and multiple second input data can be obtained by applying the same element among multiple elements.

[0111] Specifically, if the magnetic resonance image 30 is three-dimensional data, the processor 110 distorts the magnetic resonance signal 20 corresponding to the magnetic resonance image 30 by applying at least one of multiple elements in the three-dimensional K-space domain to the magnetic resonance signal 20, and obtains a training image 40 as the three-dimensional data of the corresponding magnetic resonance image 30 based on the distorted magnetic resonance signal 20. At this time, the processor 110 can set up a training dataset 50 according to each image slice that makes up the training image 40 and then enable the neural network model 10 to learn. At this time, the processor 110 can also set up multiple other image slices adjacent to each image slice together as a dataset.

[0112] For example, please see Figure 7If the processor 110 sets the seventh image slice 40-7, which is composed of multiple image slices stacked sequentially to form the three-dimensional training image 40, as the first image slice and sets it as the first input data, it can set the six adjacent image slices (specifically, the fourth image slice 40-4, the fifth image slice 40-5, the sixth image slice 40-6, the eighth image slice 40-8, the ninth image slice 40-9, and the tenth image slice 40-10) as the second input data. Then, the processor 110 can set the third image slice (i.e., the seventh image slice 30-7) corresponding to the first image slice in the multiple image slices stacked sequentially to form the three-dimensional magnetic resonance image 30 as label data. Then, the processor 110 can set the first input data, multiple second input data, and context data 70 as the training dataset 50. In the process of distorting three-dimensional K-space data, for example, by undersampling the Kmax direction of the magnetic resonance signal 20, each image slice may undergo Sinc Blurring in the slice encoding direction. This can cause localized blurring of each image slice or loss of information from each image slice to surrounding slices. To solve these problems, the processor 110 can input each image slice and multiple adjacent image slices into the neural network model 10, enabling the neural network model 10 to obtain information from the multiple adjacent image slices and effectively learn to restore the quality of the first image slice based on this information.

[0113] On the other hand, according to an embodiment of the present invention, the neural network model 10 may include multiple neural network models 10 for restoring the quality of the two-dimensional magnetic resonance image 30 and the three-dimensional magnetic resonance image 30. Specifically, the neural network model 10 may include a first neural network model 10 and a second neural network model 10. The first neural network model 10 is trained on a training image 40 obtained by distorting the two-dimensional magnetic resonance signal 20, and the second neural network model 10 is trained on multiple image slices, which are multiple image slices constituting the training image 40 obtained by distorting the three-dimensional magnetic resonance signal 20. In this case, the second neural network model 10 may be a neural network model that has been trained based on each image slice and multiple other image slices adjacent to each image slice, as described above. That is, the processor 110 can distinguish whether the acquired magnetic resonance signal 20 (or magnetic resonance image 30) is two-dimensional or three-dimensional and then use it for the learning of multiple neural network models 10.

[0114] On the other hand, according to an embodiment of the present invention, the processor 110 can restore the quality of the magnetic resonance image using the neural network model 10 that has been learned according to the aforementioned embodiment of the present invention after the neural network model 10 has been learned.

[0115] Relatedly, the processor 110 can acquire magnetic resonance images based on accelerated imaging. To this end, magnetic resonance images can be acquired based on a magnetic resonance signal that incorporates at least one of several factors related to the quality of the resonance image using an undersampling method. Specifically, the processor 110 can acquire only a portion of the K-space data using an undersampling method achieved through accelerated imaging, and perform an inverse Fourier transform operation on the acquired portion of the K-space data to acquire a magnetic resonance image. As an example, factors related to the quality of the magnetic resonance image may include uniform undersampling, random undersampling, Kmax undersampling, elliptic undersampling, and partial Fourier undersampling methods, and may include the application of Gaussian noise. Specifically, during the acquisition of the magnetic resonance signal using the applied sampling method, distorted magnetic resonance signals that omit a portion of the magnetic resonance signal may be acquired due to the application of the aforementioned various undersampling methods. Furthermore, magnetic resonance images can be acquired based on a magnetic resonance signal that has been superimposed (or noise-reduced) by adjusting the scan parameters (i.e., scan parameters) used for accelerated imaging.

[0116] Then, an inverse Fourier transform of the distorted magnetic resonance signal can be performed to obtain a magnetic resonance image. In this case, the quality of the magnetic resonance image may be relatively worse than that obtained by oversampling.

[0117] Then, the processor 110 inputs the acquired magnetic resonance image 30 and the corresponding context data 70 to the neural network model 10, which has been learned according to the aforementioned embodiments of the present invention, to restore the quality of the acquired magnetic resonance image 30. Here, the context data 70 may be scanning parameters applicable during undersampling, or it may be a noise reduction value in the magnetic resonance image 30 input by the user. The processor 110 can receive the input of the context data 70 through the dynamic modulation path 12 of the learned neural network model 10. Then, the processor 110 can acquire the improved quality magnetic resonance image 30 based on the output of the learned neural network model 10.

[0118] Before inputting the acquired magnetic resonance image 30 into the learned neural network model, the processor 110 can perform normalization on the magnetic resonance image 30. For example, the processor 110 can adjust the orientation of the magnetic resonance image 30 so that its row direction (or vertical direction) aligns with the phase encoding direction, and its column direction (or horizontal direction) aligns with the frequency encoding direction. Furthermore, to adjust the asymmetric field of view (FOV) of the magnetic resonance image 30, the processor 110 can crop the zero-padding areas of the magnetic resonance image 30 to make its shape and size consistent. Moreover, based on the Lanczos approach, the processor 110 can adjust the column size to 1024 when the magnetic resonance image 30 is a two-dimensional image (or when the magnetic resonance signal 20 is two-dimensional sequence data), and adjust the column size to 768 when the magnetic resonance image 30 is a three-dimensional image (or when the magnetic resonance signal 20 is three-dimensional sequence data) to maintain a constant pixel pitch. Furthermore, the processor 110 can perform a normalization operation on the pixel values ​​of the magnetic resonance image 30. On the other hand, normalization can be selectively performed based on device information of the magnetic resonance signal 20 (or magnetic resonance image 30) contained in the medical digital imaging and communication (DICOM) data of the corresponding magnetic resonance image 30, as well as the size, orientation, shape, etc. of the magnetic resonance signal 20 (or magnetic resonance image 30).

[0119] On the other hand, the processor 110 can readjust the improved (or restored) magnetic resonance image 30 obtained from the learned neural network model 10 according to the device information obtained from medical digital imaging and communication (DICOM) data, that is, it can perform reverse standardization operations.

[0120] According to an embodiment of the present invention, the processor 110 can also input the magnetic resonance image 30 into the first neural network model 10 and the second neural network model 10 in a differentiated manner according to the type of the magnetic resonance image 30 (whether it is two-dimensional or three-dimensional).

[0121] On the other hand, according to one embodiment of the present invention, the processor 110 may also acquire and provide restoration scene information for the magnetic resonance image 30 input from the neural network model 10 to the user. For this purpose, the neural network model 30 may learn to determine the type, number, and degree of distortion of at least one element applicable to the input magnetic resonance image 30, or identify the type, number, and degree of distortion of at least one element adjusted for restoring the magnetic resonance image 30, and output the restoration scene corresponding to that element. Furthermore, the processor 110 provides restoration scene information to guide the user to adjust scanning parameters.

[0122] Figure 8 This is a block diagram of a computing device 800 that acquires magnetic resonance images 30 based on a deep learning model, according to another embodiment of the present invention. For example, the computing device 800 may include one or more processors 810, communication interfaces 820, memory 830, image processing units 840, displays 850, user interfaces 860, and output interfaces 870. Figure 8 The computing device shown can be and Figure 2 The same device as the computing device 100 shown will be omitted for the purposes of calculation and calculation. Figure 2 Detailed description of the repeated elements shown (more than one processor 810, communication interface 820, memory 830).

[0123] The image processing unit 840 can perform image processing (e.g., inverse Fourier transform operation, etc.) on the magnetic resonance signal 20 acquired by the scanning unit or the magnetic resonance signal 20 acquired by the communication interface to acquire a magnetic resonance image 30 corresponding to the magnetic resonance signal 20. Alternatively, the image processing unit 840 can also use a learned neural network model to restore the quality of the acquired magnetic resonance image 30.

[0124] The display 850 can display various images. These images include still images and videos. Specifically, the display 850 can display the acquired or restored magnetic resonance image 30, and in addition, can provide information related to the magnetic resonance image 30 (e.g., restored scene information) to the user or target. The display 850 can be implemented in various forms, such as a liquid crystal display panel (LCD), an organic light-emitting diode (OLED), a liquid crystal on silicon (LCoS), or a digital light processing (DLP) display. Furthermore, the display 850 may also include driving circuits, backlight units, etc., implemented in forms such as amorphous silicon thin-film transistors (a-si TFTs), low-temperature polysilicon thin-film transistors (LTPS), and organic thin-film transistors (OTFTs).

[0125] On the other hand, the display 850 can also be combined with a touch panel to function as a touch screen. In this case, the display 850 can not only perform the output interface function of outputting images through the touch screen, but also the input interface function of receiving touch input from the user.

[0126] The user interface 860 can receive control commands from the user regarding the overall operation of the computing device 100. For example, the user interface 860 can receive input from the user regarding target information, parameter information, scanning conditions, pulse sequences, etc., and in particular, can receive input of context data 70. For this purpose, the user interface 860 can be implemented using a keyboard, mouse, microphone, etc.

[0127] The output interface 870 can output the information acquired by the computing device 100 to the outside. For this purpose, the output interface 870 can be implemented as a speaker or the like. The speaker can output voice information related to the restored scene of the magnetic resonance image 30.

[0128] On the other hand, according to one embodiment of the present invention, the computing device 100 can also directly scan a target (e.g., a patient) to acquire magnetic resonance images. For this purpose, the computing device 100 may further include a scanning unit, which includes a static magnetic field unit, a gradient magnetic field unit, and an RF coil unit.

[0129] The scanning unit can be implemented in a form where the target (e.g., a patient) can be inserted into an empty internal space of the scanning unit. For this purpose, the scanning unit may also include a scanning table. The scanning unit can generate a static magnetic field and a gradient magnetic field within its internal space and irradiate an RF signal. Specifically, the static magnetic field unit can generate a static magnetic field that aligns the directions of the magnetic dipole moments of the multiple atomic nuclei contained in the target with the direction of the static magnetic field. For this purpose, the static magnetic field unit can be implemented with a permanent magnet or with a superconducting magnet utilizing a cooling coil.

[0130] The gradient magnetic field unit can apply a gradient to the static magnetic field according to the control signal of the processor to form a gradient magnetic field. The gradient magnetic field unit includes an X coil, a Y coil, and a Z coil that form gradient magnetic fields in mutually perpendicular X-axis, Y-axis, and Z-axis directions. It generates gradient signals according to the shooting position in order to induce different resonant frequencies according to the location of each target.

[0131] The RF coil unit can irradiate an RF signal (e.g., an RF pulse sequence) onto the target according to the control signal of the processor. Furthermore, the RF coil unit can receive the magnetic resonance signal 20 (MR signal) emitted from the target. After transmitting an RF signal with the same frequency as the precession motion to the target towards the precession nucleus, the RF coil unit stops transmitting the RF signal and can receive the magnetic resonance signal 20 emitted from the target.

[0132] The RF coil unit can be implemented as a transmitting RF coil that generates electromagnetic waves with radio frequencies corresponding to the types of atomic nuclei and a receiving RF coil that receives electromagnetic waves radiated by atomic nuclei, or it can be implemented as a single RF transceiver coil that has both transmitting and receiving functions.

[0133] The various embodiments of the invention described above can be combined with other different embodiments and can be modified within the scope understandable to those skilled in the art based on the detailed description above. The embodiments of the invention are illustrative in all respects and should be interpreted as not limiting. For example, elements described in a single integral form can be implemented separately, and similarly, elements described in a separate form can be implemented in combination. Therefore, all modifications and variations derived from the meaning, scope, and equivalent concepts of the claims of this invention should be interpreted as falling within the scope of this invention.

Claims

1. A method for acquiring magnetic resonance images based on deep learning, executed by a computing device including at least one processor, characterized in that, Includes the following steps: At least one of a plurality of elements set for magnetic resonance image quality is applied to the magnetic resonance signal corresponding to the magnetic resonance image to obtain a training image corresponding to the magnetic resonance image. Obtain a training dataset that includes the magnetic resonance images as label data and the acquired training images as input data matching the label data; and Based on the training dataset and the corresponding context data of the training images, the neural network model learns... The steps of acquiring the training image include the following steps: applying at least one of the plurality of elements to distort the magnetic resonance signal, and acquiring the training image based on the distorted magnetic resonance signal.

2. The method according to claim 1, characterized in that, The steps of acquiring the training images include the following: changing at least one of the type or number of the applied elements, repeatedly distorting the magnetic resonance signal, and acquiring multiple training images based on multiple magnetic resonance signals with different distortions. The plurality of training images have distinct qualities corresponding to at least one of the types or numbers of the applied elements.

3. The method according to claim 2, characterized in that, The multiple elements include at least two of the following: superposition of Gaussian noise, uniform mode undersampling, random mode undersampling, Kmax undersampling, elliptic undersampling, and partial Fourier undersampling.

4. The method according to claim 3, characterized in that, Includes the following steps: If the number of training images is less than a preset number, the sampling factor of at least one of the uniform mode undersampling, the random mode undersampling, the Kmax undersampling, the elliptical undersampling, and the partial Fourier undersampling is adjusted to further distort the magnetic resonance signal, and the training images are further acquired based on the further distorted magnetic resonance signal.

5. The method according to claim 3, characterized in that, Includes the following steps: If the number of training images is less than a preset number, the intensity of the Gaussian noise is adjusted and the adjusted Gaussian noise is superimposed to further distort the magnetic resonance signal, and the training images are further acquired based on the further distorted magnetic resonance signal.

6. The method according to claim 1, characterized in that, The neural network model includes a dynamic modulation path that connects to intermediate layers of the multiple layers constituting the neural network model and extracts feature information of the context data when context data is input.

7. The method according to claim 2, characterized in that, Includes the following steps: Identify the scanning parameters corresponding to the distorted magnetic resonance signal, and use the identified scanning parameters as context data corresponding to the training image.

8. The method according to claim 1, characterized in that, Includes the following steps: The noise between the magnetic resonance signal and the distorted magnetic resonance signal is compared, the noise variation is identified, and the identified noise variation is used as context data corresponding to the training image.

9. The method according to claim 1, characterized in that, The steps to obtain the training dataset include the following: If the magnetic resonance image is three-dimensional data, then the first slice among the multiple image slices contained in the training image is set as the first input data; At least one slice adjacent to the first slice among the multiple image slices contained in the training image is set as the second input data; The third slice, corresponding to the first slice, among the multiple image slices contained in the magnetic resonance image is set as the label data; as well as The first input data, the second input data, and the label data are set as the training dataset.

10. The method according to claim 1, characterized in that, Includes the following steps: The training dataset is standardized by scaling at least one of the following: the size, orientation, pixel spacing, and pixel value of the magnetic resonance image and the training image.

11. The method according to claim 2, characterized in that, Includes the following steps: Based on at least one of the types of elements applicable to the magnetic resonance signal or the number of applicable elements, multiple restoration scenarios are set for the magnetic resonance image, and the multiple training images are classified according to the set multiple scenarios to obtain a sub-training dataset corresponding to each scenario.

12. A method for acquiring magnetic resonance images based on deep learning, executed by a computing device including at least one processor, comprising the following steps: Acquiring magnetic resonance images based on accelerated imaging methods; and The acquired magnetic resonance images and their corresponding context data are input into a learned neural network model to restore the quality of the acquired magnetic resonance images. The magnetic resonance image is acquired based on the accelerated imaging method and on a magnetic resonance signal that incorporates at least one of a plurality of factors related to the quality of the magnetic resonance image or is superimposed with noise.

13. The method according to claim 12, characterized in that, The neural network model includes a dynamic modulation path that connects to intermediate layers of the multiple layers constituting the neural network model and extracts feature information of the context data when context data is input.

14. The method according to claim 13, characterized in that, Includes the following steps: Identify the scanning parameters corresponding to the magnetic resonance signal, and use the identified scanning parameters as the context data; and The identified context data is input into the dynamic modulation path.

15. A computing device that acquires magnetic resonance images based on deep learning, characterized in that, include: Memory, storing neural network models; and At least one processor takes at least one of a plurality of elements set for the quality of a magnetic resonance image corresponding to an acquired magnetic resonance signal and applies them to the magnetic resonance signal corresponding to the acquired magnetic resonance signal to obtain a training image corresponding to the magnetic resonance image; obtains a training dataset that includes the acquired training image as input data matching the label data and uses the magnetic resonance image as label data; and enables the neural network model to learn based on the training dataset and context data corresponding to the training image. The at least one processor uses at least one of the plurality of elements to distort the magnetic resonance signal and acquires the training image based on the distorted magnetic resonance signal.

Images