Device for acquiring magnetic resonance image on basis of deep learning model and control method thereof

Abstract

The present disclosure provides a device for acquiring a magnetic resonance image on the basis of deep learning, and a control method thereof. According to an embodiment of the present disclosure, the control method comprises the steps of: acquiring a magnetic resonance signal; applying, to the magnetic resonance signal, at least one of a plurality of elements configured in relation to a quality of a magnetic resonance image corresponding to the magnetic resonance signal, thereby acquiring a training image corresponding to the magnetic resonance image; acquiring a training data set including the magnetic resonance image as label data and the acquired training image as input data matching the label data; and training a neural network model on the basis of the training data set and context data corresponding to the training image.

Description

Device for acquiring magnetic resonance imaging based on a deep learning model and method for controlling the same

[0001] The present disclosure relates to deep learning technology in the medical field, and more specifically, to an apparatus for restoring the quality of magnetic resonance imaging based on a deep learning model and a method for controlling the same.

[0002] X-ray machines, ultrasound diagnostic devices, computed tomography (CT) scanners, and magnetic resonance imaging (MRI) machines are utilized to acquire internal bodily information for observing and diagnosing the internal structure of a patient. Among these, MRI is currently receiving attention for its utility compared to other imaging technologies, particularly because it allows imaging without exposing the patient to radiation or administering contrast agents, and offers high resolution and excellent soft tissue contrast.

[0003] Magnetic Resonance Imaging (MRI) technology faces the problem of requiring a long time to acquire images. Consequently, active research on accelerated imaging techniques has been conducted in this field to reduce imaging time. However, MRI images acquired through accelerated imaging techniques suffer from low resolution or noise, making accurate analysis difficult. In particular, situations can arise where internal bodily information is lost from the MRI.

[0004] Ultimately, solving the dual problem of achieving acceleration in magnetic resonance imaging while simultaneously acquiring high-quality images has been a long-standing challenge in this technical field. Artificial intelligence technology has been proposed as a solution to this. Specifically, this involves restoring the quality of magnetic resonance images acquired through accelerated imaging techniques using artificial intelligence models.

[0005] To achieve this, it is necessary to effectively train artificial intelligence models, and in particular, a training dataset consisting of not only high-quality input data but also corresponding high-quality label data must be secured. However, existing input data only included cases of uniform undersampling (or random undersampling) performed during the acceleration imaging process, which led to performance limitations for artificial intelligence models. In other words, while the degradation of resolution and noise in magnetic resonance imaging during acceleration imaging can occur due to various causes, there is currently no appropriate method to generate a training dataset that includes these.

[0006] The present disclosure is devised in response to the aforementioned background technology and aims to provide an apparatus for acquiring magnetic resonance images based on a deep learning model and a method for controlling the same.

[0007] However, the problems to be solved in this disclosure are not limited to those mentioned above, and other unmentioned problems may be clearly understood based on the description below.

[0008] A control method for an apparatus for acquiring a magnetic resonance image based on a deep learning model, performed by a computing device comprising at least one processor according to one embodiment of the present disclosure for realizing the aforementioned task, comprises the steps of: acquiring a training image corresponding to the magnetic resonance image by applying at least one of a plurality of elements set in relation to the quality of the magnetic resonance image to a magnetic resonance signal corresponding to the magnetic resonance image; acquiring a training data set comprising the magnetic resonance image as label data and the acquired training image as input data matching the label data; and training a neural network model based on the training data set and context data corresponding to the training image, wherein the step of acquiring the training image includes the step of distorting the magnetic resonance signal by applying at least one of the plurality of elements and acquiring the training image based on the distorted magnetic resonance signal.

[0009] Alternatively, the step of acquiring the training images may include the step of repeatedly distorting the magnetic resonance signal by varying at least one of the type of applied element or the number of applied elements, and acquiring a plurality of training images based on a plurality of differently distorted magnetic resonance signals, wherein the plurality of training images may have different qualities corresponding to at least one of the type of applied element or the number of applied elements.

[0010] Alternatively, the plurality of elements may include at least two of Gaussian noise addition, uniform pattern undersampling, random pattern undersampling, Kmax undersampling, elliptical undersampling, and partial Fourier undersampling.

[0011] Alternatively, if the plurality of training images is less than a preset number, the method may include the step of further distorting the magnetic resonance signal by adjusting the frequency range of the Kmax undersampling, and further acquiring the training images based on the additionally distorted magnetic resonance signal.

[0012] Alternatively, if the plurality of training images is less than a preset number, the method includes the step of adjusting at least one sampling factor among the uniform pattern undersampling, the random pattern undersampling, the Kmax undersampling, the elliptical undersampling, and partial Fourier undersampling to further distort the magnetic resonance signal, and further acquiring the training images based on the further distorted magnetic resonance signal.

[0013] Alternatively, if the plurality of training images is less than a preset number, the method includes the step of adjusting the intensity of the Gaussian noise, adding the adjusted Gaussian noise to further distort the magnetic resonance signal, and further acquiring the training images based on the additionally distorted magnetic resonance signal.

[0014] Alternatively, the neural network model may include a dynamic modulation path connected to an intermediate layer of a plurality of layers constituting the neural network model, which extracts feature information of the context data when the context data is input.

[0015] Alternatively, the method may include the step of identifying scan parameters corresponding to the distorted magnetic resonance signal and identifying the identified scan parameters as context data corresponding to the training image.

[0016] Alternatively, the method may include the step of comparing noise between the magnetic resonance signal and the distorted magnetic resonance signal to identify a change in noise, and identifying the identified change in noise as context data corresponding to the training image.

[0017] Alternatively, the step of acquiring the training data set may include, when the magnetic resonance image is three-dimensional data, setting a first slice among a plurality of image slices included in the training image as the first input data; setting at least one slice adjacent to the first slice among a plurality of image slices included in the training image as the second input data; setting a third slice corresponding to the first slice among a plurality of image slices included in the magnetic resonance image as label data; and setting the first input data, the second input data, and the label data as the training data set.

[0018] Alternatively, the method may include the step of performing standardization on the training data set, including at least one of scaling of the size, orientation, pixel spacing, and pixel values ​​of the magnetic resonance image and the training image.

[0019] Alternatively, the method may include the step of setting a plurality of reconstruction scenarios for the magnetic resonance image according to at least one of the type of element applied to the magnetic resonance signal or the number of elements applied, classifying the plurality of training images according to the set plurality of scenarios, and obtaining a sub-training data set corresponding to each scenario.

[0020] A method for acquiring a magnetic resonance image based on deep learning, performed by a computing device comprising at least one processor according to one embodiment of the present disclosure, comprises the steps of acquiring a magnetic resonance image based on an acceleration method and inputting the acquired magnetic resonance image and context data corresponding to the acquired magnetic resonance image into a pre-trained neural network model to restore the quality of the acquired magnetic resonance image, wherein the magnetic resonance image is acquired based on a magnetic resonance signal to which at least one of a plurality of elements set in relation to the quality of the magnetic resonance image according to the acceleration method is applied or noise is added.

[0021] Alternatively, the plurality of elements may include at least one of Gaussian noise addition, uniform pattern undersampling, random pattern undersampling, Kmax undersampling, elliptical undersampling, and partial Fourier undersampling.

[0022] Alternatively, the neural network model may include a dynamic modulation path connected to an intermediate layer of a plurality of layers constituting the neural network model, which extracts feature information of the context data when the context data is input.

[0023] Alternatively, the method may identify scan parameters corresponding to the magnetic resonance signal, identify the identified scan parameters as context data, and input the identified context data into the dynamic modulation path.

[0024] A computing device for acquiring a magnetic resonance image based on deep learning according to one embodiment of the present disclosure includes a memory for storing a neural network model and at least one processor for acquiring a training image corresponding to the magnetic resonance image by applying at least one of a plurality of elements set in relation to the quality of the magnetic resonance image corresponding to the magnetic resonance signal to the magnetic resonance signal corresponding to the magnetic resonance signal, acquiring a training data set including the magnetic resonance image as label data and the acquired training image as input data matching the label data, and training the neural network model based on the training data set and context data corresponding to the training image, wherein the at least one processor distorts the magnetic resonance signal by applying at least one of the plurality of elements and acquires the training image based on the distorted magnetic resonance signal.

[0025] According to a method for acquiring magnetic resonance images based on a deep learning model according to one embodiment of the present disclosure, by acquiring training data including various cases of resolution degradation and noise generation occurring during various accelerated shooting processes and training a neural network model based thereon, the performance of the neural network model can be improved to more effectively restore the quality of low-quality magnetic resonance images acquired through accelerated shooting.

[0026] FIG. 1 is an exemplary diagram of a computing device for acquiring magnetic resonance images based on a deep learning model according to one embodiment of the present disclosure.

[0027] FIG. 2 is a block diagram of a computing device for acquiring magnetic resonance images based on a deep learning model according to one embodiment of the present disclosure.

[0028] FIG. 3 is a flowchart of a method for controlling a computing device that acquires magnetic resonance images based on a deep learning model according to one embodiment of the present disclosure.

[0029] FIGS. 4a and 4b are drawings for illustrating a plurality of elements applied to a magnetic resonance signal in the K-Space region according to one embodiment of the present disclosure.

[0030] FIG. 5 is an exemplary diagram showing the acquisition of a plurality of learning images corresponding to magnetic resonance images according to one embodiment of the present disclosure.

[0031] FIG. 6 is an exemplary diagram schematically illustrating the structure of a neural network model according to one embodiment of the present disclosure.

[0032] FIG. 7 is an exemplary diagram illustrating a method for training a neural network model to acquire a three-dimensional magnetic resonance image according to one embodiment of the present disclosure.

[0033] FIG. 8 is a detailed block diagram of a computing device for acquiring magnetic resonance images based on a deep learning model according to another embodiment of the present disclosure.

[0034] Embodiments of the present disclosure are described below with reference to the attached drawings so that those skilled in the art (hereinafter, those skilled in the art) can easily implement them. The embodiments presented in the present disclosure are provided to enable those skilled in the art to use or implement the contents of the present disclosure. Accordingly, various modifications to the embodiments of the present disclosure will be apparent to those skilled in the art. That is, the present disclosure may be embodied in various different forms and is not limited to the embodiments below.

[0035] Throughout the specification of the present disclosure, identical or similar reference numerals refer to identical or similar components. Additionally, to clearly explain the present disclosure, reference numerals in the drawings that are unrelated to the description of the present disclosure may be omitted.

[0036] The term “or” as used in this disclosure is intended to mean an implicit “or” rather than an exclusive “or.” That is, unless otherwise specified in this disclosure or its meaning is not clear from the context, “X uses A or B” should be understood to mean one of the natural implicit substitutions. For example, unless otherwise specified in this disclosure or its meaning is not clear from the context, “X uses A or B” may be interpreted as X using A, X using B, or X using both A and B.

[0037] The term “and / or” as used in this disclosure should be understood to refer to and include all possible combinations of one or more of the enumerated related concepts.

[0038] The terms “comprising” and / or “comprising” as used in this disclosure should be understood to mean the presence of certain features and / or components. However, the terms “comprising” and / or “comprising” should be understood not to exclude the presence or addition of one or more other features, other components and / or combinations thereof.

[0039] Where not otherwise specified in the present disclosure or where it is not clear from the context that the singular form indicates, the singular should generally be interpreted as including “one or more.”

[0040] The term "the N (N is a natural number)" used in this disclosure may be understood as an expression used to distinguish the components of this disclosure from one another according to certain criteria, such as functional perspectives, structural perspectives, or convenience of explanation. For example, components performing different functional roles in this disclosure may be distinguished as a first component or a second component. However, components that are substantially identical within the technical scope of this disclosure but need to be distinguished for the convenience of explanation may also be distinguished as a first component or a second component.

[0041] The term “acquisition” as used in this disclosure can be understood to mean not only receiving data through a wired or wireless communication network with an external device or system, but also generating data in an on-device form.

[0042] Meanwhile, the terms "module" or "unit" used in this disclosure may 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" may be a unit composed of a single element, or a unit expressed as a combination or set of multiple elements. For example, in a narrow sense, "module" or "unit" may refer to a hardware element of a computing device or a set thereof, an application program that performs a specific function of software, a procedure implemented through software execution, or a set of instructions for program execution. Furthermore, in a broad sense, "module" or "unit" may refer to the computing device itself that constitutes the system, or an application executed on the computing device. However, since the above-described concept is merely an example, the concepts of "module" or "part" may be defined in various ways within the scope understandable to those skilled in the art based on the contents of this disclosure.

[0043] As used in this disclosure, the term "model" may be understood as a system implemented using mathematical concepts and language to solve a specific problem, a set of software units to solve a specific problem, or an abstract model regarding a processing process to solve a specific problem. For example, a neural network "model" may refer to an overall system implemented as a neural network that possesses problem-solving capabilities through learning. In this case, the neural network may possess problem-solving capabilities by optimizing parameters connecting nodes or neurons through learning. A neural network "model" may include a single neural network or a set of neural networks composed of multiple neural networks.

[0044] The term “data” as used in this disclosure may include images, signals, etc. The term “image” as used in this disclosure may refer to multidimensional data composed of discrete image elements. In other words, “image” may be understood as a term referring to a digital representation of an object visible to the human eye. For example, “image” may refer to multidimensional data composed of elements corresponding to pixels in a two-dimensional image. “Image” may refer to multidimensional data composed of elements corresponding to voxels in a three-dimensional image.

[0045] As used in this disclosure, the term “image” may refer to multi-dimensional data composed of discrete image elements (e.g., pixels in a two-dimensional image and voxels in a three-dimensional image). For example, an image may include, but is not limited to, medical images acquired by a medical imaging device such as a magnetic resonance imaging device, a computed tomography (CT) device, an ultrasound device, or an X-ray device.

[0046] As used in this disclosure, the term "medical image" is a general concept encompassing all forms of images that include medical knowledge, and may include images acquired through various modalities such as visible light cameras, IR cameras, ultrasound, X-ray, CT, MRI, PET, etc.

[0047] As used in this disclosure, the term "picture archiving and communication system (PACS)" may refer to a system that stores, processes, and transmits medical images in accordance with the Digital Imaging and Communications in Medicine (DICOM) standard. For example, the "picture archiving and communication system" may be linked with digital medical imaging equipment to store medical images, such as magnetic resonance imaging (MRI) and computed tomography (CT) images, in accordance with the Digital Imaging and Communications in Medicine standard. The "picture archiving and communication system" may transmit medical images to terminals inside or outside the hospital via a communication network. At this time, meta information, such as interpretation results and medical records, may be added to the medical images.

[0048] As used in this disclosure, the term "object" refers to a subject of imaging and may include a person, an animal, or a part thereof. For example, an object may include a part of the body (such as an organ or tissue) or a phantom. A phantom refers to a substance having a volume that is very close to the density and effective atomic number of a living organism, and may include a spherical phantom having properties similar to those of the body.

[0049] A Magnetic Resonance Imaging (MRI) system is a system that acquires images of cross-sectional areas of an object by expressing the intensity of a Magnetic Resonance (MR) signal in response to a Radio Frequency (RF) signal generated in a magnetic field of a specific intensity as contrast.

[0050] The MRI system causes the main magnet to form a static magnetic field and aligns the direction of the magnetic dipole moment of specific atomic nuclei of an object located within the static field with the direction of the static field. A gradient field coil applies a gradient signal to the static field to form a gradient field, thereby inducing different resonance frequencies for different parts of the object. An RF coil can irradiate magnetic resonance signals in accordance with the resonance frequency of the area where image acquisition is desired. Additionally, as the gradient field is formed, the RF coil can receive magnetic resonance signals of different resonance frequencies radiated from various parts of the object. The MRI system acquires images by applying image reconstruction techniques to the magnetic resonance signals received through these steps. Furthermore, the MRI system may reconstruct multiple magnetic resonance signals into image data by performing serial or parallel signal processing on multiple magnetic resonance signals received by a multi-channel RF coil.

[0051] The explanation of the foregoing terms is intended to aid in understanding the present disclosure. Accordingly, it should be noted that unless a foregoing term is explicitly stated as a matter limiting the content of the present disclosure, it is not to be used in the sense of limiting the technical concept of the content of the present disclosure.

[0052] FIG. 1 is an exemplary diagram of a computing device (100) for acquiring a magnetic resonance image based on a deep learning model according to one embodiment of the present disclosure.

[0053] A computing device (100) for acquiring magnetic resonance images based on a deep learning model according to one embodiment of the present disclosure may be a hardware device or a part of a hardware device that performs comprehensive processing and computation of data, or it may be a software-based computing environment connected to a communication network. For example, the computing device (100) may be a server that performs intensive data processing functions and is an entity that shares resources, or it may be a client that shares resources through interaction with the server. Additionally, the computing device (100) may be a cloud system that enables multiple servers and clients to interact to comprehensively process data. Since the above description is merely one example regarding the type of computing device (100), the type of computing device (100) may be configured in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure. For example, the computing device (100) may include a smartphone, tablet PC, PC, smart TV, micro server, cloud server, etc., that processes magnetic resonance images or performs processing functions. As another example, the computing device (100) may be a magnetic resonance imaging (MRI) device that directly acquires magnetic resonance images.

[0054] Referring to FIG. 1, a computing device (100) can acquire a training data set (50) for training a neural network model (10). Specifically, the computing device (100) can acquire a plurality of magnetic resonance images (30) or magnetic resonance signals (20) corresponding to the magnetic resonance images (30) acquired by a plurality of other electronic devices (200) that acquire magnetic resonance images (30), and acquire a training data set (50).

[0055] 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 through an inverse Fourier operation on the magnetic resonance signal (20). The computing device (100) may obtain pulse sequence data obtained by each of the multiple electronic devices (200) from the multiple electronic devices (200). Here, the pulse sequence data may include K-Space data collected based on a specific pulse sequence used in the other electronic device (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.

[0056] Here, the magnetic resonance image (30) and the magnetic resonance signal (20) can be transmitted and received by being included in DICOM (Digital Imaging and Communications in Medicine) data. DICOM stands for Digital Imaging and Communications in Medicine, which is a collective term for various standards used for digital image representation and communication in medical devices. DICOM data may primarily contain patient information and media characteristics. For example, the various medical information data included in DICOM data consists of patient-related text information and unprocessed media information collected at the medical site, and there are no specific restrictions on the format. More specifically, DICOM data may include the patient's biological information, image information about the patient or treatment site generated at the medical site (e.g., magnetic resonance image (30)), and information about the device that acquired the image.

[0057] Meanwhile, the computing device (100) can acquire a learning data set (50) based on the acquired magnetic resonance image (30) and / or magnetic resonance signal (20). Specifically, the computing device (100) can acquire a learning image (40) of lower quality than the magnetic resonance image (30) by adjusting the quality of the magnetic resonance image (30) and / or magnetic resonance signal (20). The computing device (100) can degrade the quality of the magnetic resonance image (30) by adjusting at least one of a plurality of elements set in relation to the quality of the magnetic resonance image (30) and / or magnetic resonance signal (20). In particular, the quality of the same magnetic resonance image (30) can be degraded in various ways by selectively combining a plurality of elements. This may be referred to as adjusting the quality of the magnetic resonance image (30) in a plurality of aspects or in a plurality of dimensions. Accordingly, by degrading the quality of the same magnetic resonance image (30) in various ways, the computing device (100) can acquire multiple learning images (40) for the same magnetic resonance image (30).

[0058] Meanwhile, the computing device (100) can train a neural network model (10) based on context data corresponding to each training data included in the training data set (50) together with the acquired training data set (50). Here, the context data may be data explaining the relationship between the magnetic resonance image (30) constituting the training data and the training image (40), and the background of the degraded quality of the magnetic resonance image (30). Through this, the computing device (100) enables the neural network model (10) to accurately learn the relationship between the various acquired training images (40) and the magnetic resonance images (30) corresponding to the training images (40).

[0059] Hereinafter, embodiments of the present disclosure relating thereto will be described in detail with reference to FIGS. 2 to 8.

[0060] FIG. 2 is a block diagram of a computing device (100) for acquiring a magnetic resonance image (30) based on a deep learning model according to one embodiment of the present disclosure. FIG. 3 is a flowchart of a method for controlling a computing device (100) for acquiring a magnetic resonance image (30) based on a deep learning model according to one embodiment of the present disclosure.

[0061] Referring to FIG. 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, since FIG. 2 is merely an example, the computing device (100) may include other configurations for implementing a computing environment. Additionally, only some of the disclosed configurations may be included in the computing device (100).

[0062] A processor (110) according to one embodiment of the present disclosure may be understood as a constituent unit comprising hardware and / or software for performing computing operations. For example, the processor (110) may read a computer program and perform data processing for machine learning. The processor (110) may process computational processes such as processing input data for machine learning, extracting features for machine learning, and calculating errors based on backpropagation. A processor (110) for performing such data processing 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). Since the above-described types of processors (110) are merely examples, the types of processors (110) may be configured in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure.

[0063] The processor (110) is electrically connected to other components of the computing device (100) (i.e., communication interface (120), and memory (130)) to control the overall operation of the computing device (100).

[0064] A memory (130) according to one embodiment of the present disclosure may be understood as a configuration unit comprising hardware and / or software for storing and managing data processed by a computing device (100). That is, the memory (130) may store data of any form generated or determined by a processor (110) and data of any form received by a communication interface (120). For example, the memory (130) may include at least one type of storage medium among a flash memory type (130), a hard disk type, a multimedia card micro type, a card type memory, RAM (random access memory), SRAM (static random access memory), ROM (read-only memory), EEPROM (electrically erasable programmable read-only memory), PROM (programmable read-only memory), magnetic memory, a magnetic disk, and an optical disk. Additionally, the memory (130) may include a database system that controls and manages data in a predetermined system. Since the above-described type of memory (130) is merely an example, the type of memory (130) may be configured in various ways within a range understandable to those skilled in the art based on the contents of this disclosure.

[0065] The memory (130) can structure and organize data, combinations of data, and program code executable on the processor (110) that are necessary for the processor (110) to perform operations. For example, the memory (130) may store a neural network model (10), a training data set (50), and context data. Additionally, the memory (130) may store program code that operates the neural network model (10) to perform training based on the training data set (50) and context data, program code that operates the neural network model (10) to receive a magnetic resonance image (30) as input and perform inference according to the purpose of use of the computing device (100), and processed data generated as the program code is executed.

[0066] A communication interface (120) according to one embodiment of the present disclosure may be understood as a configuration unit that transmits and receives data through any known wired or wireless communication system. For example, the communication interface (120) may perform data transmission and reception using a wired or wireless communication system such as a local area network (LAN), wideband code division multiple access (WCDMA), long term evolution (LTE), wireless broadband internet (WiBro), 5th generation mobile communication (5G), ultrawide-band communication, ZigBee, radio frequency (RF) communication, wireless LAN, wireless fidelity (Wi-Fi), near field communication (NFC), or Bluetooth. Since the communication systems described above are merely examples, wired or wireless communication systems for data transmission and reception of the communication interface (120) may be applied in various ways other than those described above.

[0067] The communication interface (120) can receive data necessary for the processor (110) to perform calculations through wired or wireless communication with any system or any client. Additionally, the communication interface (120) can transmit data generated through the calculations of the processor (110) through wired or wireless communication with any system or any client. For example, the communication interface (120) can receive medical data through communication with a database within a hospital environment, a cloud server performing tasks such as standardization of medical data, or a computing device (100). The communication interface (120) can transmit output data of the neural network model (10), and intermediate data and processed data derived during the calculation process of the processor (110), through communication with the aforementioned database, server, or computing device (100).

[0068] Referring to FIG. 3, according to one embodiment of the present disclosure, a processor (110) acquires a magnetic resonance signal (20) (S310). Here, the magnetic resonance signal (20) may be acquired by the computing device (100) through a magnetic field generated on an object (e.g., a patient), or it may be acquired from a plurality of other electronic devices (200) that acquire a magnetic resonance image (30) through a communication interface (120). For example, the computing device (100) may acquire pulse sequence data acquired by each of the plurality of other electronic devices (200) as the magnetic resonance signal (20). At this time, the pulse sequence data may include K-Space data collected based on a specific pulse sequence used in the other electronic device (200). The pulse sequence data may include two-dimensional pulse sequence data collected in a two-dimensional space or three-dimensional pulse sequence data collected in a three-dimensional space.

[0069] Meanwhile, the processor (110) may receive DICOM (Digital Imaging and Communications in Medicine) data from another electronic device (200) and acquire a magnetic resonance signal (20). Specifically, the processor (110) may extract the magnetic resonance signal (20) included in the DICOM data or extract the magnetic resonance image (30) and then acquire the magnetic resonance signal (20) through a Discrete Fourier Transformation. Since the same description as in FIG. 1 applies to this, a detailed description will be omitted.

[0070] And, the processor (110) can obtain a learning image (40) corresponding to the magnetic resonance image (30) by applying at least one of a plurality of elements set in relation to the quality of the magnetic resonance image (30) corresponding to the magnetic resonance signal (20) to the magnetic resonance signal (20) (S320).

[0071] Multiple elements are set in relation to the quality of the magnetic resonance image (30) corresponding to the magnetic resonance signal (20), and may be elements that influence and determine the quality of the magnetic resonance image (30). For example, the quality of the magnetic resonance image (30) may be evaluated based on the resolution of the magnetic resonance image (30) and the degree of noise contained in the magnetic resonance image (30). That is, high quality of the magnetic resonance image (30) may mean that the resolution of the magnetic resonance image (30) is high and the noise is low. In this case, multiple elements may be elements that determine the resolution of the magnetic resonance image (30) and the degree of noise contained in the magnetic resonance image (30). In particular, the element related to the resolution of the magnetic resonance image (30) may be related to the type of undersampling method for the magnetic resonance signal (20). Undersampling is a technique for acquiring magnetic resonance images by scanning an object but sampling only a portion of the K-space region instead of completely filling the entire area with K-space data. Undersampling may also be referred to as subsampling.

[0072] 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) can degrade the quality of the magnetic resonance image (30) corresponding to the magnetic resonance image (30) by applying at least one of a plurality of elements set in relation to the quality of the magnetic resonance image (30) to the magnetic resonance signal (20). Then, the processor (110) can acquire the degraded magnetic resonance image (30) as a training image (40) used to train the neural network model (10).

[0073] Meanwhile, according to one embodiment of the present disclosure, the processor (110) can obtain a learning image (40) corresponding to a magnetic resonance image (30) by applying at least one of a plurality of elements to a magnetic resonance signal (20) in the K-Space region. Specifically, the processor (110) can distort the magnetic resonance signal (20) in the K-Space region and obtain a learning image (40) based on the distorted magnetic resonance signal (20). As an example, the processor (110) can distort K-Space data by applying at least one of a plurality of elements of K-Space data included in the magnetic resonance image (30) signal in the K-Space region. The processor (110) can distort K-Space data by adding noise to the K-Space data, or by applying an undersampling pattern to the K-Space data and selecting a part of the K-Space data.

[0074] FIGS. 4a and 4b are drawings for explaining a plurality of elements applied to a magnetic resonance signal (20) in the K-Space region according to an embodiment of the present disclosure. FIG. 4a shows a plurality of elements applied to a magnetic resonance signal (20) on a plane of the axis of the phase encoding direction (Ky) and the axis of the slice encoding selection direction (Kz), excluding the frequency encoding direction (Kx).

[0075] Referring to FIG. 4a, a plurality of 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).

[0076] Specifically, applying Gaussian noise to the magnetic resonance signal (20) may involve adding or subtracting Gaussian noise from the K-Space data. That is, the processor (110) may generate random noise (e.g., Gaussian noise) in the K-Space region and subtract it from the K-Space data to distort the magnetic resonance signal. Alternatively, the processor (110) may distort the magnetic resonance signal by subtracting random noise (i.e., Gaussian noise) present in the K-Space data.

[0077] Uniform pattern undersampling (62) is a method of selecting K-Space data at preset intervals in the K-Space region and may include uniform GRAPPA pattern undersampling and uniform CAIPIRINHA pattern undersampling. Random pattern undersampling (63) may be a method of randomly selecting K-Space data in the K-Space region. Additionally, Kmax undersampling (64) may be performed in all encoding directions (Kx, Ky, and Kz) in a manner of selectively omitting high-frequency K-Space data at the edges of the K-Space region. Furthermore, elliptical undersampling (65) may be a method of selecting K-Space data contained in the elliptical region of the K-Space region and omitting the remaining K-Space data, which may be a method of retaining low-frequency data in the center and omitting high-frequency data at the edges. Meanwhile, partial Fourier undersampling (66) may be performed in all encoding directions (Kx, Ky, Kz) in a manner of selecting K-Space data contained in a part of the K-Space region and omitting the remaining K-Space data. By applying various undersampling methods like this, the processor (110) can distort the magnetic resonance signal.

[0078] Meanwhile, when Gaussian noise (61) is added to the first magnetic resonance signal (21) shown in FIG. 4b, and the Kmax undersampling method (64) is applied in all encoding directions, the elliptical undersampling method (65) is applied, and the partial Fourier undersampling method (66) is applied in all encoding directions so that the first magnetic resonance signal (21) is distorted, a second magnetic resonance signal (22) with some K-Space data missing can be obtained. Then, based on the second magnetic resonance signal (22), the processor (110) can obtain a training image (40). However, the order in which the plurality of elements shown in FIG. 4b are applied is merely an example for explaining the present disclosure and is not limited thereto.

[0079] Meanwhile, the processor (110) can acquire a training image (40) based on a distorted magnetic resonance signal (20). Specifically, the processor (110) can acquire a training image (40) corresponding to the distorted magnetic resonance signal (20) by performing an inverse Fourier operation on the distorted magnetic resonance signal (20) (i.e., K-Space data). At this time, the processor (110) can acquire a training image (40) corresponding to the distorted magnetic resonance signal (20) based on a parallel image technique (e.g., Grappa) and another previously trained neural network model.

[0080] FIG. 5 is an exemplary diagram showing the acquisition of a plurality of learning images (40) corresponding to a magnetic resonance image (30) according to one embodiment of the present disclosure.

[0081] Meanwhile, the processor (110) can repeatedly distort the magnetic resonance signal (20) by varying the type and number of 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) may have different qualities corresponding to the type, number, and degree of distortion of the applied elements.

[0082] The processor (110) can acquire a plurality of learning images (40) corresponding to the same magnetic resonance image (30). Specifically, the processor (110) can repeatedly distort the magnetic resonance signal (20) by varying the type, number, and degree of distortion of the elements applied to the magnetic resonance signal (20). Referring to FIG. 5, the processor (110) can acquire a plurality of learning images (40) corresponding to the magnetic resonance image (30) by combining a plurality of elements (first to seventh elements). The processor (110) can obtain a first learning image (40-1) by applying a first element to a magnetic resonance signal (20), obtain a second learning image (40-2) by applying the first and second elements to the magnetic resonance signal (20), obtain a 13th learning image (40-13) by applying the third, sixth, and seventh elements to the magnetic resonance signal (20), and obtain a 127th learning image (40-127) by applying the first to seventh elements to the magnetic resonance signal (20).

[0083] In this way, the processor (110) can obtain various training images (40) by applying a plurality of elements (first to seventh elements) to the magnetic resonance signal (20) or by optionally applying at least two combined elements to the magnetic resonance signal (20). At this time, the plurality of training images (40) obtained for the same magnetic resonance image (30) may have different resolutions and noise depending on the type and number of elements applied to the magnetic resonance signal (20). That is, the processor (110) can obtain training data including various cases in which the quality of the same magnetic resonance image (30) is degraded by setting a plurality of elements related to the quality of the magnetic resonance image (30) for the magnetic resonance signal (20) and applying various combinations of elements to the magnetic resonance signal (20). In addition, the processor (110) can train the neural network model (10) to restore the magnetic resonance image (30) more effectively by obtaining multiple training data that are degraded in quality in various ways for each of the multiple magnetic resonance images.

[0084] Meanwhile, according to one embodiment of the present disclosure, if the number of training images (40) is less than a preset number, the processor (110) can further distort the magnetic resonance signal (20) by adjusting the frequency range of Kmax undersampling, and further acquire training images (40) based on the additionally distorted magnetic resonance signal (20).

[0085] That is, the processor (110) can determine whether the number of training images (40) is greater than or equal to a preset number in order to determine whether a sufficient number of training images (40) for the magnetic resonance image (30) has been secured. Then, if the processor determines that the number of training images (40) is less than the preset number, or if it determines that the number of multiple training images (40) acquired for the same magnetic resonance image (30) is less than the preset number, it can further distort the magnetic resonance signal (20) by applying a different frequency range of Kmax undersampling among the multiple elements. For example, the processor (110) can change the frequency range from a first range to a second range greater than the first range. At this time, it goes without saying that the processor (110) can apply other elements to the magnetic resonance signal (20) along with the Kmax undersampling with the changed frequency range. Kmax undersampling is a process of omitting high-frequency K-Space data located at the edges of the K-Space region, and the processor (110) can further distort the magnetic resonance signal (20) by adjusting the range of the omitted high-frequency K-Space data, thereby enabling the additional acquisition of a training image (40).

[0086] At this time, the processor (110) can determine the similarity between multiple training images (40) and, for multiple training images (40) where the similarity is higher than a preset value, consider them as the same training image (40) to determine the number of training images (40). Meanwhile, for multiple training images (40) where the similarity is higher than a preset value, the processor (110) may apply different elements to the magnetic resonance signal (20) or apply different frequency ranges of Kmax undersampling to reduce the similarity and re-acquire the corresponding training images (40). In particular, the processor (110) can determine the similarity after distorting the magnetic resonance signal through all possible combinations using multiple elements.

[0087] Additionally, if the number of training images (40) is less than a preset number, the processor (110) may further distort the magnetic resonance signal (20) by changing the Gaussian noise intensity, and further acquire training images (40) based on the additionally 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 may further distort the magnetic resonance signal (20) by changing the Gaussian noise intensity applied to the same magnetic resonance image (30), and further acquire training images (40) based on the additionally distorted magnetic resonance signal (20). For example, the processor (110) may acquire additionally distorted magnetic resonance signals (20) by changing the Gaussian noise intensity from a first intensity to a second intensity and applying it to the magnetic resonance signal (20).

[0088] Additionally, if the number of training images (40) is less than a preset number, the processor (110) may correct the undersampling multiplier to further distort the magnetic resonance signal (20) and further acquire training images (40) based on the further distorted magnetic resonance signal (20). Specifically, the processor (110) may further distort the magnetic resonance signal (20) by adjusting at least one of a uniform pattern undersampling multiplier, a random pattern undersampling multiplier, a Kmax undersampling multiplier, an elliptical undersampling multiplier, and a partial Fourier undersampling multiplier, and further acquire training images (40) based on the further distorted magnetic resonance signal (20). For example, the processor (110) may acquire the further distorted magnetic resonance signal (20) by changing the uniform pattern undersampling multiplier from a first value to a second value and applying it to the magnetic resonance signal (20). Through this, the processor (110) can improve the performance of the neural network model (10) by obtaining a wider variety of training images (40) and training the neural network model (10).

[0089] According to one embodiment of the present disclosure, a processor (110) acquires a learning data set (50) that includes a magnetic resonance image (30) as label data and an acquired learning image (40) as input data that matches the label data (S330). At this time, the processor (110) can acquire a plurality of learning data sets (50) by matching a plurality of different input data (learning images (40)) to the same label data (magnetic resonance image (30)).

[0090] Meanwhile, according to one embodiment of the present disclosure, the processor (110) may set a plurality of restoration scenarios for a magnetic resonance image (30) according to the type, number, and degree of distortion of the elements applied to the magnetic resonance signal (20), classify a plurality of training images (40) according to the set plurality of restoration scenarios, and obtain a sub-training data set (50) corresponding to each restoration scenario.

[0091] Specifically, the processor (110) can acquire a plurality of training images (40) corresponding to a plurality of magnetic resonance images (30) by distorting a plurality of magnetic resonance signals (20) corresponding to a plurality of magnetic resonance images (30). At this time, the processor (110) can set a plurality of restoration scenarios for the magnetic resonance images (30) according to the type, number, and degree of distortion of the elements applied to acquire the plurality of training images (40). Here, the restoration scenario may be information describing the process of restoring the quality of the training images (40) by backtracking the magnetic resonance images (30) from the training images (40).

[0092] And, the processor (110) classifies a plurality of training images (40) included in the training data set (50) and a plurality of magnetic resonance images (30) that match the plurality of training images (40) according to each restoration scenario, and can identify the number of sub-training data sets (50) corresponding to each restoration scenario. And, for restoration scenarios where the number of sub-training data sets is less than a preset value, the processor (110) can additionally secure the corresponding sub-training data sets. That is, the processor can additionally receive a magnetic resonance signal (20) from another electronic device (200) through a communication interface (120), and apply at least one element corresponding to the restoration scenario to the additionally received magnetic resonance signal (20) to additionally secure a sub-training data set.

[0093] According to one embodiment of the present disclosure, the processor (110) may perform standardization on a training data set (50) including at least one of adjusting the size, orientation, pixel spacing, and pixel value scale of the magnetic resonance image (30) and the training image (40). Standardization may also be performed through a standardization module connected to the input end of the neural network model (10).

[0094] The processor (110) may perform standardization work on the magnetic resonance images (30) and training images (40) included in the training data set (50) prior to training the neural network model (10). Here, standardization is the matching of the size, orientation, etc. of the multiple magnetic resonance images (30) and the multiple training images (40), and can be performed to effectively train the neural network model (10) and to further enhance the training effect of the neural network model (10).

[0095] For example, the processor (110) can adjust the orientation of a plurality of magnetic resonance images (30) and a plurality of learning images (40) to match. Specifically, the processor (110) can match the orientation of the plurality of magnetic resonance images (30) and a plurality of learning images (40) such that the row direction (or vertical direction) matches the phase encoding direction and the column direction (or horizontal direction) matches the frequency encoding direction. Additionally, the processor (110) can match the shape and size of the plurality of magnetic resonance images (30) and a plurality of learning images (40) by cropping the zero padding area of ​​the plurality of magnetic resonance images (30) and a plurality of learning images (40) in order to adjust the asymmetrically displayed Field of View (FOV) of the plurality of magnetic resonance images (30) and a plurality of learning images (40). For example, the shape of the plurality of magnetic resonance images (30) and a plurality of learning images (40) can be matched as a rectangular shape. Additionally, the processor (110) can adjust the sizes of the plurality of magnetic resonance images (30) and the plurality of training images (40) to match them. Specifically, based on the Lanczos method, the processor (110) can adjust the column size to 1024 when the plurality of magnetic resonance images (30) and the plurality of training images (40) are 2D images (or when the magnetic resonance signal (20) is 2D sequence data), and adjust the column size to 768 when the plurality of magnetic resonance images (30) and the plurality of training images (40) are 3D images (or when the magnetic resonance signal (20) is 3D sequence data) for a 3D pulse sequence, thereby maintaining the image pixel spacing constant. Additionally, the processor (110) can perform normalization operations on the plurality of magnetic resonance images (30) and the plurality of training images (40). Specifically, the processor (110) can adjust the pixel values ​​so that the range of pixel values ​​of the plurality of magnetic resonance images (30) and the plurality of training images (40) matches.

[0096] The above-described standardization work may be performed sequentially, or it may be performed selectively according to a plurality of magnetic resonance images (30) and a plurality of learning images (40). In particular, the processor (110) may selectively perform the standardization work by determining device information that acquired each magnetic resonance signal (20) (or magnetic resonance image (30)) based on DICOM data containing magnetic resonance signals (20) (or magnetic resonance images (30)) or by identifying the size, direction, shape, etc. of the magnetic resonance signals (20) (or magnetic resonance images (30)).

[0097] FIG. 6 is an exemplary diagram schematically showing the structure of a neural network model (10) according to one embodiment of the present disclosure.

[0098] According to one embodiment of the present disclosure, a processor (110) can train a neural network model (10) based on context data (70) corresponding to a training data set (50) and a training image (40) (S540). The processor (110) can train the neural network model (10) by utilizing context data (70) as an auxiliary input along with the training data set (50). For example, the neural network model (10) is a neural network model (10) having a U-Net framework, and the neural network model (10) may include network models such as a Deep Neural Network (DNN), Recurrent Neural Network (RNN), Bidirectional Recurrent Deep Neural Network (BRDNN), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN).

[0099] Context data (70) corresponds to a distorted magnetic resonance signal (20) and thus can correspond to a training image (40). Here, the context data (70) may be data explaining the relationship between the magnetic resonance image (30) constituting the training data and the training image (40), and the background of the degraded quality of the magnetic resonance image (30). Meanwhile, the context data (70) may be included in the training data set (50) together with the training image (40) corresponding to the context data (70).

[0100] According to one embodiment of the present disclosure, a neural network model (10) may include a Dynamic Modulation Pathway (DMP) connected to an intermediate layer of a plurality of layers constituting the neural network model (10), which extracts feature information of context data (70) when context data (70) is input. That is, context data (70) may be input to an intermediate layer of the neural network model (10) through the Dynamic Modulation Pathway (DMP) (12). To extract feature information of context data (70), the Dynamic Modulation Pathway (DMP) (12) may include a fully connected layer and an activation function (e.g., a Relu function). Meanwhile, the extracted feature information of the context may be integrated into a U-Net framework and function as a convolution kernel.

[0101] According to one embodiment of the present disclosure, a processor (110) may identify a scan parameter corresponding to a distorted magnetic resonance signal (20) in the K-Space region and identify the identified scan parameter as context data (70) corresponding to a training image (40). In particular, the processor (110) may identify the context data (70) based on at least one element applied to distort the magnetic resonance signal (20). The processor (110) may identify the value of the scan parameter that has been changed as the magnetic resonance signal (20) is distorted (or as at least one element is applied) among the scan parameters corresponding to the magnetic resonance signal (20), and identify the context data (70) with the value of the identified scan parameter.

[0102] Additionally, according to one embodiment of the present disclosure, the processor (110) can compare noise between the magnetic resonance signal (20) and the distorted magnetic resonance signal (20) in the K-Space region to identify the amount of noise change and identify the identified amount of noise change as context data (70) corresponding to the training image (40). At this time, the processor (110) can determine the amount of noise change based on the amount of Gaussian noise added to the K-Space data in relation to Gaussian noise addition among a plurality of elements.

[0103] Meanwhile, referring to FIG. 6, context data (70) can be input into the dynamic modulation path (DMP) (12) as a one-dimensional matrix. At this time, the context data (70) can be input into the dynamic modulation path (DMP) by converting it into a one-dimensional matrix corresponding to the scan parameters or a one-dimensional matrix corresponding to the amount of noise change, and in particular, it may be a one-dimensional matrix in which the scan parameters and the amount of noise change are combined.

[0104] Meanwhile, according to one embodiment of the present disclosure, based on a plurality of training images (40) of which quality has been degraded in various ways for a plurality of magnetic resonance images (30), the processor (110) can train the neural network model (10) to restore the quality degraded in various ways for the same magnetic resonance image (30), and can obtain a neural network model (10) that has superior restoration ability and is more effective than a conventional neural network model trained only on training data of which quality has been degraded by applying only a single element (e.g., uniform undersampling).

[0105] Meanwhile, according to one embodiment of the present disclosure, the processor (110) can train a neural network model (10) by inputting set restoration scenario information corresponding to the type, number, and degree of distortion of elements applied to the magnetic resonance signal (20) together with a training data set (50) and context data (70). Accordingly, the neural network model (10) can be trained to restore a magnetic resonance image (30) and simultaneously output a restoration scenario corresponding to the restoration process.

[0106] FIG. 7 is an exemplary diagram illustrating a method of training a neural network model (10) to obtain a three-dimensional magnetic resonance image (30) according to one embodiment of the present disclosure.

[0107] According to one embodiment of the present disclosure, when the magnetic resonance image (30) is three-dimensional data, the processor (110) may set a target image slice (hereinafter, first image slice) among a plurality of image slices included in the training image (40) as target input data (hereinafter, first input data). Then, the processor (110) may identify a plurality of image slices (hereinafter, second image slices) adjacent to the first slice among the plurality of image slices included in the training image (40) as a plurality of reference input data (hereinafter, second input data), and set an image slice (hereinafter, third image slice) corresponding to the first slice among the plurality of image slices included in the magnetic resonance image (30) as label data. Then, the processor (110) may set the first input data, the plurality of second input data, and the label data as a training data set (50). At this time, the first input data and the plurality of second input data may be obtained by applying the same element among a plurality of elements.

[0108] Specifically, when the magnetic resonance image (30) is three-dimensional data, the processor (110) can distort the magnetic resonance signal (20) corresponding to the magnetic resonance image (30) by applying at least one of a plurality of elements in a three-dimensional K-Space region to the magnetic resonance signal (20), and acquire a training image (40), which is three-dimensional data corresponding to the magnetic resonance image (30), based on the distorted magnetic resonance signal (20). At this time, the processor (110) can train the neural network model (10) by setting a training data set (50) for each image slice constituting the training image (40). At this time, the processor (110) can set a plurality of other image slices adjacent to each image slice together as a data set.

[0109] For example, referring to FIG. 7, when the processor (110) sets the 7th image slice (40-7) among a plurality of image slices that are sequentially stacked to form a three-dimensional learning image (40) as the first image slice as the first input data, it may set six image slices adjacent to each other (specifically, the 4th, 5th, 6th and 8th, 9th, and 10th image slices (40-4, 40-5, 40-6 and 40-8, 40-9, 40-10)) as the second input data. Additionally, the processor (110) may set the 3rd image slice (i.e., the 7th image slice (30-7)) corresponding to the first image slice among a plurality of image slices that are sequentially stacked to form a three-dimensional magnetic resonance image (30) as label data. In addition, the processor (110) can set the first input data, a plurality of second input data, and context data (70) as a training data set (50). In the process of distorting three-dimensional K-Space data, for example, when performing slice encoding direction Kmax undersampling on a magnetic resonance signal (20), sinc blurring may occur in the slice encoding direction in each image slice, which causes a problem where a specific area of ​​each image slice is blurry or information of each image slice is lost to surrounding slices. To solve this, the processor (110) can input a plurality of other image slices adjacent to each image slice together into the neural network model (10), so that the neural network model (10) obtains information of each image slice from a plurality of other image slices adjacent to it and learns to effectively restore the quality of the first image slice based on this.

[0110] Meanwhile, according to one embodiment of the present disclosure, the neural network model (10) may include a plurality of neural network models (10) that restore the quality of a two-dimensional magnetic resonance image (30) and a three-dimensional magnetic resonance image (30). Specifically, the neural network model (10) may include a first neural network model (10) trained on a training image (40) obtained by distorting a two-dimensional magnetic resonance signal (20), and a second neural network model (10) trained on a plurality of image slices constituting a training image (40) obtained by distorting a three-dimensional magnetic resonance signal (20). In this case, the second neural network model (10) may be trained based on each image slice and a plurality of other image slices adjacent to each image slice, as described above. That is, the processor (110) can distinguish whether the type of acquired magnetic resonance signal (20) (or magnetic resonance image (30)) is two-dimensional or three-dimensional and use it for training a plurality of neural network models (10).

[0111] Meanwhile, according to one embodiment of the present disclosure, when the training of the neural network model (10) is completed, the processor (110) can restore the quality of the magnetic resonance image using the neural network model (10) trained according to the above-described embodiment of the present disclosure.

[0112] In this regard, the processor (110) can acquire a magnetic resonance image based on an acceleration imaging method. To this end, the magnetic resonance image can be acquired based on a magnetic resonance signal to which at least one of a plurality of elements regarding the quality of the resonance image is applied by an undersampling method. Specifically, the processor (110) can acquire only a portion of the K-Space data by using an undersampling method through acceleration imaging, and can acquire a magnetic resonance image by performing an inverse Fourier operation on the acquired portion of the K-Space data. As an example, elements related to the quality of the magnetic resonance image may include a uniform undersampling method, a random undersampling method, a Kmax undersampling method, an elliptical undersampling method, and a partial Fourier undersampling method related to the undersampling method, and may include the application of Gaussian noise. Specifically, in the process of acquiring a magnetic resonance signal by applying an undersampling method, a distorted magnetic resonance signal in which some magnetic resonance signals are omitted may be acquired as the various undersampling methods described above are applied. In addition, magnetic resonance images can be acquired based on a magnetic resonance signal with noise added (or subtracted) depending on the adjustment of scan parameters (i.e., scan parameters) for acceleration imaging.

[0113] Additionally, magnetic resonance images can be obtained by performing an inverse Fourier operation on a distorted magnetic resonance signal. In this case, the quality of the magnetic resonance image may be relatively lower than that of a magnetic resonance image obtained by the oversampling method.

[0114] And, the processor (110) can restore the quality of the acquired magnetic resonance image (30) by inputting the acquired magnetic resonance image (30) and context data (70) corresponding to the acquired magnetic resonance image (30) into the neural network model (10) trained according to the embodiment of the present disclosure described above. Here, the context data (70) may be scan parameters applied during the undersampling process, or may be noise reduction values ​​in the magnetic resonance image (30) input by the user. The processor (110) can input the context data (70) through the dynamic modulation path (DMP) (12) of the pre-trained neural network model (10). And, the processor (110) can acquire the quality-enhanced magnetic resonance image (30) as the output of the pre-trained neural network model (10).

[0115] The processor (110) may perform standardization on the magnetic resonance image (30) before inputting the acquired magnetic resonance image (30) into a pre-trained neural network model. For example, the processor (110) may adjust the orientation of the magnetic resonance image (30) so that the row direction (or vertical direction) of the magnetic resonance image (30) matches the phase encoding direction and the column direction (or horizontal direction) matches the frequency encoding direction. Additionally, the processor (110) may crop the zero padding area of ​​the magnetic resonance image (30) to match the shape and size of the magnetic resonance image (30) in order to adjust the asymmetrically displayed Field of View (FOV) of the magnetic resonance image (30). Additionally, the processor (110) can maintain a constant image pixel spacing by adjusting 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) based on the Lanczos method, and by adjusting 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). Additionally, the processor (110) can perform normalization operations on the pixel values ​​of the magnetic resonance image (30). Meanwhile, standardization can be selectively performed based on device information that acquired the magnetic resonance signal (20) (or magnetic resonance image (30)) included in the DICOM data corresponding to the magnetic resonance image (30), and the size, direction, shape, etc. of the magnetic resonance signal (20) (or magnetic resonance image (30)).

[0116] Meanwhile, the processor (110) can re-align the quality-enhanced (or restored) magnetic resonance image (30) obtained from the previously trained neural network model (10) to match the device information identified from the DICOM data, that is, perform a de-standardization operation.

[0117] According to one embodiment of the present disclosure, the processor (110) may input the magnetic resonance image (30) to the first and second neural network models (10) separately depending on the type of the magnetic resonance image (30) (whether it is two-dimensional or three-dimensional).

[0118] Meanwhile, according to one embodiment of the present disclosure, the processor (110) may obtain restoration scenario information for a magnetic resonance image (30) input from a neural network model (10) and provide it to a user. To this end, the neural network model (30) may be trained to determine the type, number, and degree of distortion of at least one element applied to the input magnetic resonance image (30), or to identify the type, number, and degree of distortion of at least one element adjusted to restore the magnetic resonance image (30), and to identify and output a restoration scenario corresponding to the element. Meanwhile, the processor (110) may induce the user to adjust scan parameters by providing restoration scenario information.

[0119] FIG. 8 is a detailed block diagram of a computing device (800) for acquiring a magnetic resonance image (30) based on a deep learning model according to another embodiment of the present disclosure. For example, the computing device (800) may include one or more processors (810), a communication interface (820), a memory (830), an image processing unit (840), a display (850), a user interface (860), and an output interface (870). The computing device shown in FIG. 8 may be the same device as the computing device (100) shown in FIG. 2, and therefore, detailed descriptions of components that overlap with the configuration shown in FIG. 2 (one or more processors (810), a communication interface (820), and a memory (830)) are omitted.

[0120] The image processing unit (840) can obtain a magnetic resonance image (30) corresponding to the magnetic resonance signal (20) obtained through the scanning unit or the magnetic resonance signal (20) obtained through the communication interface by performing image processing (e.g., inverse Fourier operation, etc.). Alternatively, the image processing unit (840) may restore the quality of the obtained magnetic resonance image (30) using a pre-trained neural network model.

[0121] The display (850) can display various images. Here, the images include both still images and video. In particular, the display (850) can display acquired or restored magnetic resonance images (30), and can also provide information related to the magnetic resonance images (30) (e.g., restoration scenario information, etc.) to a user or object. The display (850) can be implemented as various types of displays (850), such as LCD (Liquid Crystal Display Panel), OLED (Organic Light Emitting Diodes), LCoS (Liquid Crystal on Silicon), DLP (Digital Light Processing), etc. Additionally, the display (850) may also include a driving circuit, a backlight unit, etc., which can be implemented in forms such as a-si TFT, LTPS (low temperature poly silicon) TFT, OTFT (organic TFT), etc.

[0122] Meanwhile, the display (850) may be implemented as a touch screen by combining it with a touch panel, and in this case, the display (850) can perform the function of an input interface that receives user touch input as well as an output interface that outputs an image through the touch screen.

[0123] The user interface (860) can receive control commands regarding the overall operation of the computing device (100) from the user. For example, the user interface (860) can receive information regarding object information, parameter information, scan conditions, pulse sequences, etc. from the user, and in particular, can receive context data (70). To this end, the user interface (860) can be implemented using a keyboard, mouse, microphone, etc.

[0124] The output interface (870) can output information acquired by the computing device (100) to the outside. To this end, the output interface (870) can be implemented as a speaker, etc. The speaker can output a voice message related to the restoration scenario of the magnetic resonance image (30).

[0125] Meanwhile, according to one embodiment of the present disclosure, the computing device (100) may directly scan an object (e.g., a patient) to obtain a magnetic resonance image. To this end, the computing device (100) may further include a scanning unit, and the scanning unit includes a static magnetic field unit, a gradient magnetic field unit, and an RF coil unit.

[0126] The scanning unit may be implemented in a form that allows an object (e.g., a patient) to be inserted into the empty internal space of the scanning unit. To this end, the scanning unit may further include a table. The scanning unit may form a static magnetic field and a gradient magnetic field in the internal space and irradiate an RF signal. Specifically, the static magnetic field unit may form a static magnetic field to align the direction of the magnetic dipole moments of the atomic nuclei contained in the object with the direction of the static magnetic field. To this end, the static magnetic field unit may be implemented as a permanent magnet or as a superconducting magnet using a cooling coil.

[0127] The gradient magnetic field unit can form a gradient magnetic field by applying a gradient to a static magnetic field according to a control signal from a processor. The gradient magnetic field unit includes X, Y, and Z coils that form gradient magnetic fields in the mutually orthogonal X, Y, and Z axis directions, and generates a gradient signal according to the shooting position so that resonance frequencies can be induced differently for each part of the object.

[0128] The RF coil unit can irradiate an RF signal (e.g., an RF pulse sequence) to a target according to a control signal from a processor. Additionally, the RF coil unit can receive a magnetic resonance signal (20) (MR signal) emitted from the target. The RF coil unit can transmit an RF signal with a frequency equal to the frequency of the precession motion toward a precessing atomic nucleus to the target, stop transmitting the RF signal, and receive the magnetic resonance signal (20) emitted from the target.

[0129] The RF coil section may be implemented as a transmitting RF coil that generates electromagnetic waves having a radio frequency corresponding to the type of atomic nucleus and a receiving RF coil that receives electromagnetic waves radiated from the atomic nucleus, respectively, or as a single RF transmitting and receiving coil having both transmitting and receiving functions.

[0130] The various embodiments of the present disclosure described above may be combined with additional embodiments and modified to the extent understandable to those skilled in the art in light of the detailed description above. The embodiments of the present disclosure are illustrative in all respects and should be understood as not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form. Accordingly, all modifications or variations derived from the meaning, scope, and equivalents of the claims of the present disclosure should be interpreted as being included within the scope of the present disclosure.

Claims

1. A method for acquiring a deep learning-based magnetic resonance image, performed by a computing device comprising at least one processor, wherein A step of obtaining a learning image corresponding to the magnetic resonance image by applying at least one of a plurality of elements set in relation to the quality of the magnetic resonance image to a magnetic resonance signal corresponding to the magnetic resonance image; A step of acquiring a training data set comprising the magnetic resonance image as label data and the acquired training image as input data matching the label data; and The method includes the step of training a neural network model based on the above training dataset and context data corresponding to the above training image; The step of acquiring the above-mentioned learning video is, A step comprising: distorting the magnetic resonance signal by applying at least one of the plurality of elements above, and acquiring the training image based on the distorted magnetic resonance signal; method.

2. In Paragraph 1, The step of acquiring the above-mentioned learning video is, The method includes the step of repeatedly distorting the magnetic resonance signal by making at least one of the type of applied element or the number of applied elements different, and acquiring a plurality of training images based on a plurality of differently distorted magnetic resonance signals. The above plurality of training videos are, Characterized by having different qualities corresponding to at least one of the type of applied element or the number of applied elements. method.

3. In Paragraph 2, The above plurality of elements are, Addition of Gaussian noise, including at least two of Uniform Pattern Under Sampling, Random Pattern Under Sampling, Kmax Under Sampling, Elliptical Under Sampling, and Partial Fourier Under Sampling, method.

4. In Paragraph 3, If the plurality of training images is less than a preset number, a step of adjusting at least one sampling factor among the uniform pattern undersampling, the random pattern undersampling, the Kmax undersampling, the elliptical undersampling, and partial Fourier undersampling to further distort the magnetic resonance signal, and further acquiring the training images based on the further distorted magnetic resonance signal; including, method.

5. In Paragraph 3, If the plurality of training images is less than a preset number, the method comprises the steps of: adjusting the intensity of the Gaussian noise; adding the adjusted Gaussian noise to further distort the magnetic resonance signal; and further acquiring the training images based on the further distorted magnetic resonance signal. including, method.

6. In Paragraph 1, The above neural network model is, A dynamic modulation path connected to an intermediate layer of a plurality of layers constituting the neural network model, which extracts feature information of the context data when the context data is input, method.

7. In Paragraph 2, A step of identifying scan parameters corresponding to the above-mentioned distorted magnetic resonance signal, and identifying the identified scan parameters as context data corresponding to the above-mentioned training image; including, method.

8. In Paragraph 1, A step of comparing noise between the magnetic resonance signal and the distorted magnetic resonance signal to identify a change in noise, and identifying the identified change in noise as context data corresponding to the training image; including, method.

9. In Paragraph 1, The step of acquiring the above training data set is, If the magnetic resonance image is three-dimensional data, a step of setting a first slice among a plurality of image slices included in the training image as the first input data; A step of setting at least one slice adjacent to the first slice among a plurality of image slices included in the training video as second input data; A step of setting a third slice corresponding to the first slice among a plurality of image slices included in the magnetic resonance image as label data; and A step of setting the first input data, the second input data, and the label data as a training data set; including, method.

10. In Paragraph 1, A step of performing standardization on the above training data set, including at least one of scaling of the size, orientation, pixel spacing, and pixel values ​​of the magnetic resonance image and the training image; including, method.

11. In Paragraph 2, A step of setting a plurality of reconstruction scenarios for the magnetic resonance image according to at least one of the type of element applied to the magnetic resonance signal or the number of elements applied, classifying the plurality of training images according to the set plurality of scenarios, and obtaining a sub-training data set corresponding to each scenario; including, method.

12. A method for acquiring a deep learning-based magnetic resonance image, performed by a computing device comprising at least one processor, wherein A step of acquiring a magnetic resonance image based on an acceleration imaging method; and The method includes the step of inputting the acquired magnetic resonance image and context data corresponding to the acquired magnetic resonance image into a pre-trained neural network model to restore the quality of the acquired magnetic resonance image; The above magnetic resonance image is Based on the above acceleration imaging method, at least one of a plurality of elements regarding the quality of the magnetic resonance image is applied, or acquired based on a magnetic resonance signal to which noise has been added, method.

13. In Paragraph 12, The above neural network model is, A dynamic modulation path connected to an intermediate layer of a plurality of layers constituting the neural network model, which extracts feature information of the context data when the context data is input, method.

14. In Paragraph 13, A step of identifying scan parameters corresponding to the magnetic resonance signal and identifying the identified scan parameters as context data; and A step of inputting the identified context data into the dynamic modulation path; including, method.

15. In a computing device for acquiring magnetic resonance images based on deep learning, Memory for storing neural network models; and At least one processor that applies at least one of a plurality of elements set in relation to the quality of a magnetic resonance image corresponding to an acquired magnetic resonance signal to a magnetic resonance signal corresponding to said magnetic resonance image to acquire a training image corresponding to said magnetic resonance image, acquires a training data set including said magnetic resonance image as label data and said acquired training image as input data matching said label data, and trains said neural network model based on said training data set and context data corresponding to said training image. The above-mentioned at least one processor is, Distorting the magnetic resonance signal by applying at least one of the above plurality of elements, and acquiring the learning image based on the distorted magnetic resonance signal, Computing device.

Images