Method for generating training data, computer program, and apparatus

Abstract

A method performed by a computing device including at least one processor according to one embodiment of the present disclosure includes the steps of analyzing signal strength and noise characteristics of at least one input image, determining whether there are image pairs among the at least one input image whose signal strength and noise characteristics correspond, and generating training images and labeled images according to the determination results based on a predetermined noise reduction goal.

Description

Technical Field

[0001] The present disclosure relates to a learning data generation method, a computer program, and an apparatus, and specifically to a learning data generation method, a computer program, and an apparatus using medical images.

Background Art

[0002] Generally, a medical imaging device is a device that acquires a patient's body information and provides an image. Examples of medical imaging devices include an X-ray imaging device, an ultrasonic diagnostic device, a computed tomography (CT) device, a magnetic resonance imaging (MRI) device, and the like.

[0003] A magnetic resonance imaging device is equipment that requires a considerably long imaging time. Therefore, an accelerated imaging technology for shortening the imaging time of magnetic resonance imaging occupies a very important part in the medical industry. In order to use the accelerated magnetic resonance imaging in a medical field, it must include all information about the imaging subject and minimize noise that affects the interpretation of the information. In response to such a need, recently, a technology for restoring a low-quality magnetic resonance image obtained by accelerated imaging based on artificial intelligence to a high-quality state without accelerated imaging has been developed.

[0004] In order to restore an accelerated magnetic resonance image based on artificial intelligence to a high quality, it is necessary to effectively train an artificial intelligence model. For this purpose, not only high-quality input data but also corresponding high-quality label data must be ensured. However, even if an accelerated magnetic resonance image is ensured at a certain level as input data, there is a problem that there are almost no realistic high-quality restored images corresponding thereto. Therefore, it is necessary to ensure learning data that is basically required for constructing an artificial intelligence model.

Summary of the Invention

Problems to be Solved by the Invention

[0005] This disclosure aims to solve the problems of the prior art described above and relates to a method for generating training data, a computer program, and an apparatus for generating training images and label images based on the characteristics of the input image and the goal of noise reduction.

[0006] However, the technical challenges that this embodiment aims to address are not limited to those mentioned above; other technical challenges may also exist. [Means for solving the problem]

[0007] An embodiment of the present disclosure for achieving the aforementioned problems is disclosed, which provides a method for generating training data performed by a computing device. The method is characterized by including the steps of: analyzing the signal intensity and noise characteristics of at least one input image; determining whether there is an image pair among the at least one input image whose signal intensity and noise characteristics correspond; and generating a training image and a label image based on the determination result according to a predetermined noise reduction target.

[0008] As an alternative, the step of determining whether the aforementioned image pair exists is characterized by determining that a plurality of images having the same signal intensity and being noise-independent from one another constitute the aforementioned image pair.

[0009] As an alternative, the step of generating the learning image and the label image is characterized in that, if the image pair exists, the first image included in the image pair is determined to be the learning image, and the first image and the second image are combined according to the noise reduction target to generate the label image.

[0010] As an alternative, the step of generating the training image and the label image is characterized by generating the label image such that the first noise of the training image and the second noise of the label image are noise-dependent to the noise reduction target.

[0011] As an alternative, the step of generating the training image and the label image is characterized in that, if the image pair does not exist, a first image from among the at least one input image is determined to be the label image, and the training image is generated based on the first image and the noise reduction target.

[0012] As an alternative, the step of generating the learning image and the label image is characterized in that the signal intensity of the learning image and the signal intensity of the first image are the same, and the learning image is generated such that the first noise of the first image and the second noise of the learning image are noise-dependent according to the noise reduction target.

[0013] Alternatively, the learning image and the label image are input into an artificial neural network model that improves the quality of medical images.

[0014] An embodiment of the present disclosure for achieving the aforementioned problems is disclosed, which is a method for generating training data performed by a computing device. The method includes the steps of setting a noise reduction target, generating a training image based on a first image, and generating a label image by combining the first image and the second image according to the noise reduction target, wherein the first image and the second image have the same signal intensity and are noise-independent of each other.

[0015] As an alternative, the step of generating the label image is characterized by generating the label image using the following mathematical formula.

[0016]

number

[0017] An embodiment of the present disclosure for achieving the aforementioned problems is disclosed, which is a method for generating training data performed by a computing device. The method includes the steps of setting a noise reduction target, generating a label image based on a first image, and generating a training image based on the first image and the noise reduction target, wherein the training image and the label image have the same signal intensity and are noise-dependent on each other to the extent corresponding to the noise reduction target. As an alternative, the step of generating the learning image is characterized by including a step of generating a second noise in which the signal intensity of the first noise of the first image is the same as that of the first noise and is noise-independent of the first noise, and a step of generating the learning image by combining the first image and the second noise according to the noise reduction target. As an alternative, the step of generating the learning image by combining the first image and the second noise according to the noise reduction target is characterized by generating the learning image using the following mathematical formula.

[0018]

number

[0019] According to one embodiment of the present disclosure for realizing the above-described problems, a learning data generation device is disclosed. The device includes a memory that stores a predetermined noise reduction target and at least one input image, and when there is a second image in the at least one input image whose signal intensity is the same as that of the first image and the noise of which is independent of each other, the first image is determined as a learning image, the first image and the second image are combined according to the noise reduction target to generate the label image, and when the second image does not exist, the first image is determined as the label image, and a processor that generates the learning image based on the first image and the noise reduction target.

[0020] As an alternative, the learning image and the label image have the same signal intensity and are only noise-dependent on each other corresponding to the noise reduction target.

[0021] According to one embodiment of the present disclosure for realizing the above-described problems, a computer program stored in a computer-readable storage medium is disclosed. When the computer program is executed by one or more processors, the following operations are performed. Here, the operations include an operation of analyzing the signal intensity and noise characteristics of at least one input image, an operation of determining whether there is an image pair in the at least one input image whose signal intensity and noise characteristics correspond to each other, and an operation of generating a learning image and a label image input to the artificial neural network model based on the determination result and the noise reduction target set in the artificial neural network model.

Advantages of the Invention

[0022] According to the problem-solving means of the present disclosure described above, the present disclosure can finely adjust the degree of noise removal according to the user's requirements, so that it is possible to prevent the loss of image information due to excessive noise removal.

[0023] Also, according to the problem-solving means of the present disclosure described above, the present disclosure discloses various methods for generating learning images and label images based on input image characteristics, so that the time and cost for providing learning data can be reduced.

Brief Description of the Drawings

[0024] [Figure 1] It is a block diagram of a computing device according to an embodiment of the present disclosure. [Figure 2] It is a block diagram of a learning data generation device according to an embodiment of the present disclosure. [Figure 3] It is a flowchart of a learning data generation method according to an embodiment of the present disclosure. [Figure 4] It is a flowchart showing the detailed operation of step S130 in FIG. 3 according to an embodiment of the present disclosure. [Figure 5] It is a flowchart showing the detailed operation of step S130 in FIG. 3 according to an embodiment of the present disclosure. [Figure 6] It is a block diagram of a learning data generation device according to an embodiment of the present disclosure. <00s0113>

Modes for Carrying Out the Invention

[0025] Hereinafter, embodiments of the present disclosure will be described in detail so that those having ordinary knowledge in the technical field of the present disclosure (hereinafter referred to as those skilled in the art) can easily implement them with reference to the accompanying drawings. The embodiments presented in the present disclosure are provided so that those skilled in the art can use or implement the content of the present disclosure. Therefore, various modifications to the embodiments of the present disclosure will be apparent to those skilled in the art. That is, the present disclosure can be embodied in various different forms and is not limited to the following embodiments. [[ID=3s]]

[0026] Throughout the specification of the present disclosure, the same or similar reference numerals refer to the same or similar components. Also, for the purpose of clearly explaining the present disclosure, the reference numerals of the parts not related to the description of the present disclosure can be omitted from the drawings.

[0027] As used in this disclosure, the term "or" is intended to mean an implicational "or" rather than an exclusive "or". That is, wherever not otherwise specified or where its meaning is not clear from the context, "X uses A or B" should be understood to mean one of the natural implicational substitutions. For example, wherever not otherwise specified or where its meaning is not clear from the context, "X uses A or B" can be interpreted as X using A, X using B, or X using both A and B.

[0028] The terms "and / or" as used in this disclosure should be understood to include all possible combinations of one or more of the related concepts listed.

[0029] The terms “contains” and / or “contains” as used in this disclosure should be understood to mean the presence of certain features and / or components. However, the terms “contains” and / or “contains” should be understood not to exclude the presence or addition of one or more other features, other components and / or combinations thereof.

[0030] Wherever the context does not clearly indicate otherwise or singular form in this disclosure, singular should generally be interpreted as including "one or more."

[0031] The term “nth (where n is a natural number)” as used in this disclosure can be understood as an expression used to distinguish components of this disclosure from one another based on predetermined criteria such as functional, structural, or ease of explanation. For example, components in this disclosure that perform different functional roles may be distinguished as either a first component or a second component. However, components that are substantially identical within the technical concept of this disclosure but must be distinguished for ease of explanation may also be distinguished as either a first component or a second component.

[0032] On the other hand, the terms "module" or "unit" as used in this disclosure 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. Here, a "module" or "unit" may be a unit composed of a single element, or it may be a unit expressed as a combination or set of multiple elements. For example, as a concept, a "module" or "unit" may refer to a hardware element or set thereof of a computing device, an application program that performs a specific function of software, a processing procedure embodied by the execution of software, or a set of instructions for executing a program. Furthermore, as a broader concept, a "module" or "unit" may refer to the computing device itself that constitutes a system, or an application executed on a computing device. However, the above concepts are merely examples, and the concepts of "module" or "unit" can be defined in various ways to the extent that a person skilled in the art can understand them based on the content of this disclosure.

[0033] As used in this disclosure, the term "model" can be understood as a system embodied using mathematical concepts and language to solve a particular problem, a set of software units to solve a particular problem, or an abstract model of a processing step to solve a particular problem. For example, a neural network "model" can refer to any system embodied as a neural network that has problem-solving capabilities through learning. Here, a neural network can have problem-solving capabilities by optimizing the parameters connecting nodes or neurons through learning. A neural network "model" may include a single neural network or a set of neural networks that are combinations of multiple neural networks.

[0034] As used in this disclosure, the term “image” may mean 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, a medical image obtained by a medical imaging device such as a magnetic resonance imaging system, computed tomography (CT) scanner, ultrasound scanner, or X-ray scanner.

[0035] 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, a "picture archiving and communication system" can work in conjunction 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. A "picture archiving and communication system" can transmit medical images to terminals inside and outside the hospital via a communication network. Meta information, such as reading results and medical records, can be added to the medical images.

[0036] As used in this disclosure, the term “object” refers to the subject of a photograph and may include a human being, an animal, or a part thereof. For example, an object may include a part of a body (such as an organ or tissue) or a phantom. A phantom means a substance having a volume 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 a body.

[0037] A magnetic resonance imaging (MRI) system is a system that obtains an image of a cross-sectional area of ​​an object by representing the intensity of a magnetic resonance (MR) signal in response to an RF (Radio Frequency) signal generated by a magnetic field of a specific intensity, in contrast to light and dark.

[0038] An MRI system uses a main magnet to form a static magnetic field, aligning the magnetic dipole moment direction of specific atomic nuclei of an object located within the static magnetic field. Gradient coils apply a gradient signal to the static magnetic field to create a gradient, thereby inducing different resonance frequencies for different parts of the object. RF coils can irradiate magnetic resonance signals to match the resonance frequency of the desired area for image acquisition. Furthermore, the RF coils can receive magnetic resonance signals with different resonance frequencies radiated from many parts of the object due to the formation of the gradient magnetic field. The MRI system acquires an image by applying image reconstruction techniques to the magnetic resonance signals received through these steps. The MRI system can also reconstruct multiple magnetic resonance signals into image data by performing serial or parallel signal processing on multiple magnetic resonance signals received by multi-channel RF coils.

[0039] The explanations of the terms provided above are intended to aid in understanding this disclosure. Therefore, unless the terms are explicitly stated to limit the content of this disclosure, it should be noted that they are not intended to limit the technical ideas of this disclosure.

[0040] Figure 1 is a block diagram of a computing device according to one embodiment of the present disclosure.

[0041] A computing device 100 according to one embodiment of this disclosure may be a hardware device or part of a hardware device that performs comprehensive data processing and calculations, or it may be a software-based computing environment connected via a communication network. For example, the computing device 100 may be a server that performs intensive data processing functions and is the entity that shares resources, or it may be a client that shares resources through interaction with a server. Furthermore, the computing device 100 may be a cloud system that enables multiple servers and clients to interact with each other to comprehensively process data. The above description is merely an example related to the type of computing device 100, and the type of computing device 100 can be configured in a variety of ways within the scope that can be understood by a person skilled in the art based on the content of this disclosure.

[0042] Referring to Figure 1, a computing device 100 according to one embodiment of the present disclosure may include a processor 110, memory 120, and a network unit 130. However, since Figure 1 is merely an example, the computing device 100 may include other configurations to embody a computer environment. Furthermore, the computing device 100 may include only a portion of the disclosed configurations.

[0043] A processor 110 according to one embodiment of the present disclosure can be understood as a component unit including hardware and / or software for performing computing operations. For example, the processor 110 can read a computer program and perform data processing for machine learning. The processor 110 can handle computational processes such as processing input data for machine learning, feature extraction for machine learning, and error calculation 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). The types of processors 110 described above are merely examples, and the types of processors 110 can be configured in a variety of ways within the scope understandable to those skilled in the art based on the content of the present disclosure.

[0044] According to one embodiment of this disclosure, the processor 110 can generate training data to be input to an artificial neural network model. The artificial neural network model is trained by the training data and can output low-quality images as high-quality images. For example, the artificial neural network model can generate high-quality images by removing noise from the input low-quality images.

[0045] An artificial neural network model may include at least one neural network. The neural network may include, but is not limited to, network models such as DNN (Deep Neural Network), RNN (Recurrent Neural Network), BRDNN (Bidirectional Recurrent Deep Neural Network), MLP (Multilayer Perceptron), and CNN (Convolutional Neural Network).

[0046] The image could be a medical image, such as a magnetic resonance imaging (MSR) image. Here, the MRS image could be obtained using accelerated imaging. Accelerated imaging can be understood as an imaging technique that reduces imaging time by increasing the acceleration factor compared to general imaging. The acceleration factor is a term used in parallel imaging techniques and can be understood as the value obtained by dividing the number of signal lines fully sampled in the k-space domain by the number of signal lines sampled by imaging.

[0047] The processor 110 can generate training data in various ways depending on the number of input images. In this specification, training data refers to pairs of training images and their corresponding label images.

[0048] The input image is a medical image and may be magnetic resonance imaging or k-space data. The training data may include training images and corresponding label images. The label images may be of higher quality than the training images. The training images and label images may be noise-independent by a predetermined value and noise-dependent by a predetermined value. In this case, the artificial neural network model trained on the training images and label images can be trained to remove noise corresponding to a predetermined value.

[0049] The processor 110 can either directly set a noise reduction target for the artificial neural network model or receive a noise reduction target setting. The processor 110 can generate training data based on the noise reduction target.

[0050] The processor 110 analyzes the image characteristics of at least one input image. Specifically, it analyzes the number of input images, signal strength, and noise characteristics. The processor 110 determines whether there is an image pair among the at least one input image whose signal strength and noise characteristics correspond. In this specification, an image pair means, but is not limited to, two images.

[0051] On the other hand, if the noise characteristics correspond but the signal strengths are different, the processor 110 can additionally perform a preprocessing step to ensure that the signal strengths of the corresponding input images are the same.

[0052] The processor 110 generates training images and label images based on the judgment results and noise reduction targets.

[0053] In one embodiment, if an image pair exists among at least one input image, the processor 110 combines the first and second images included in the image pair to generate a training image and a label image. If the noise reduction target is set to reduce noise by x, the processor 110 determines weight values ​​corresponding to the first and second images, respectively, based on x, and combines the first and second images based on the weight values.

[0054] In one embodiment, if there is only one input image or if no image pairs exist among the input images, the processor 110 generates a training image and a label image using one image, i.e., the first image. Specifically, the processor 110 generates noise that is noise-independent from the noise of the first image, and can generate training data using the first image, the generated noise, and the noise reduction target.

[0055] On the other hand, the processor 110 can also perform the operation of determining the existence of an image pair without performing the operation itself.

[0056] According to embodiments of this disclosure, the processor 110 can finely adjust the degree of noise reduction according to the user's requirements, thereby preventing the loss of image information due to excessive noise reduction. Furthermore, since various methods for generating training images and label images based on input image characteristics are disclosed, the time and cost required to provide training data can be reduced.

[0057] According to embodiments of this disclosure, the processor 110 can be trained to train an artificial neural network model to improve the quality of the image. The image may be, for example, a magnetic resonance image, which may be acquired by accelerated imaging. Accelerated imaging can be understood as an imaging technique that reduces imaging time by increasing the acceleration factor compared to general imaging. The acceleration factor is a term used in parallel imaging techniques and can be understood as the value obtained by dividing the number of signal lines fully sampled in the K-space domain by the number of signal lines sampled by imaging. For example, an acceleration factor of 2 can be understood as acquiring half the number of signal lines compared to the number of fully sampled signal lines when sampling the magnetic resonance signal in the phase encoding direction to acquire lines. Therefore, as the acceleration factor increases, the imaging time of the magnetic resonance image can be reduced proportionally. That is, by increasing the acceleration factor when imaging a magnetic resonance image, accelerated imaging with a reduced imaging time for the magnetic resonance image can be realized.

[0058] If necessary, the processor 110 can generate a user interface that provides an environment for interaction with the user of the computing device 100 or any client user. For example, the processor 110 can generate a user interface for receiving an input image and a noise reduction target for an artificial neural network model.

[0059] The processor 110 can generate a user interface that embodies functions such as outputting, modifying, changing, or adding data based on external input signals applied by the user. The role of the user interface described above is merely an example, and the role of the user interface can be defined in various ways within the scope understandable to those skilled in the art based on the content of this disclosure.

[0060] A memory 120 according to one embodiment of the present disclosure can be understood as a component unit including hardware and / or software for storing and managing data processed by the computing device 100. That is, the memory 120 can store any form of data generated or determined by the processor 110 and any form of data received by the network unit 130. For example, the memory 120 may include at least one type of storage medium from among flash memory type, hard disk type, multimedia card micro type, 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, magnetic disk, or optical disk. The memory 120 may also include a database system for controlling and managing data in a predetermined manner. The types of memory 120 described above are merely examples, and the types of memory 120 can be configured in a variety of ways within the scope understandable to those skilled in the art based on the content of the present disclosure.

[0061] Memory 120 can structure and organize and manage data, data combinations, and program code (code) that can be executed by the processor 110, which are necessary for the processor 110 to perform calculations. For example, memory 120 can store input images received via the network unit 130, which will be described later. Memory 120 can also store training images and label images generated from the input images by the processor 110. Memory 120 can store noise reduction targets set for the artificial neural network model. Memory 120 can also store program code that causes the processor 110 to generate training data.

[0062] A network unit 130 according to one embodiment of this disclosure can be understood as a component that transmits and receives data via any known wired wireless communication system. For example, the network unit 130 can transmit and receive data using a wired wireless communication system such as a local area network (LAN), wideband code division multiple access (WCDMA), LTE (long term evolution), WiBro (wireless broadband internet), 5th generation mobile communication (5G), ultrawide-band wireless communication (ultrawide-band), ZigBee, radio frequency (RF) communication, wireless LAN (wireless LAN), Wi-Fi (wireless fidelity), near field communication (NFC), or Bluetooth®. The above-mentioned communication systems are merely examples, and a wide variety of wired wireless communication systems for data transmission and reception of the network unit 130 are applicable beyond the examples given above.

[0063] The network unit 130 can receive data necessary for the processor 110 to perform calculations via wired or wireless communication with any system or client. The network unit 130 can also transmit data generated by the processor 110's calculations via wired or wireless communication with any system or client. For example, the network unit 130 can receive medical images through communication with a medical image storage and transmission system, a cloud server that performs tasks such as improving the resolution and standardizing medical images, or a computing device 100. The network unit 130 can also transmit improved images output from artificial intelligence models and mathematical models through communication with the aforementioned systems, servers, or computing devices 100.

[0064] If necessary, the network unit 130 can provide a user interface generated by the processor 110 to any system or client via wired or wireless communication with any system or client. For example, the network unit 130 can provide a user interface that visualizes data processed by the processor 110 to the aforementioned system or device by communicating with a medical image storage and transmission system or a computing device 100 including a display. The network unit 130 can receive external input signals applied by the user via the user interface from the aforementioned system or device and transmit them to the processor 110. Here, the processor 110 can operate to implement functions such as data output, modification, change, or addition based on the external input signals transmitted from the network unit 130. On the other hand, even without communication via the network unit 130, the system or device that receives the user interface can operate to implement functions such as data output, modification, change, or addition in response to external input signals applied by the user.

[0065] Figure 2 is a block diagram of a learning data generation device according to one embodiment of the present disclosure.

[0066] Referring to Figure 2, the training data generator 10 receives at least one input image 210 and generates training data including a training image 310 and a label image 320. The training data generator 10 can receive a noise reduction target set for the artificial neural network model 400. The artificial neural network model 400 is trained using the training data to improve low-quality images to high-quality images by the noise reduction target.

[0067] The learning data generation device 10 can determine whether there are image pairs 211 and 212 among at least one input image 210 that have corresponding signal strength and noise characteristics. Here, the timing of receiving at least one input image 210 may differ from each other. Furthermore, the criteria for determining image pairs 211 and 212 may be diverse, including not only signal strength and noise characteristics, but also the video acquisition method and the video shooting environment.

[0068] If there are image pairs 211 and 212, i.e., a first image 211 and a second image 212 whose image characteristics correspond to each other, the training data generation device 10 combines the first image 211 and the second image 212 based on the noise reduction target to generate a training image 310 and a label image 320.

[0069] If image pairs 211 and 212 do not exist, or if there is only one input image 210, the training data generation device 10 generates a training image 310 and a label image 320 using the noise reduction target and the first image 211.

[0070] The training data generation device 10 can provide the artificial neural network model 400 with training data including the generated training image 310 and label image 320.

[0071] On the other hand, the configuration shown in Figure 1 is illustrative, and the training data generation device 10 and the artificial neural network model 400 may be included in different systems. Here, the training data generation device 10 and the artificial neural network model 400 can send and receive training data via a network.

[0072] Figure 3 is a flowchart of a training data generation method according to one embodiment of the present disclosure.

[0073] Referring to Figure 3, the learning data generation device may be an embodiment of the learning data generation device 10 in Figure 1 or the learning data generation device 20 in Figure 6.

[0074] Prior to step S110, the training data generation device can set a noise reduction target for the artificial neural network model. For example, it can receive input on the noise reduction target from the user.

[0075] The training data generation device can analyze the signal intensity and noise characteristics of at least one input image (S110). For example, the training data generation device analyzes the signal intensity and noise characteristics of each of at least one input image. Alternatively, it can analyze whether two or more input images are noise-independent of each other.

[0076] The training data generation device can determine whether there is an image pair among at least one input image in which the signal strength and noise characteristics correspond (S120). The first and second images included in the image pair have identical signal strengths and are noise-independent.

[0077] The learning data generation device can generate a learning image and a label image based on the judgment result, according to a predetermined noise reduction target (S130).

[0078] Specifically, if an image pair exists, the training data generation device can combine the first and second images contained in the image pair to generate a training image and a label image. This will be explained in detail later based on Figure 4.

[0079] Alternatively, if no image pairs exist, the training data generator can generate training images and label images using a single image, i.e., the first image. This will be explained in detail later based on Figure 5.

[0080] Here, the training data generation device can generate training images and label images to correspond to the noise reduction target. Furthermore, the training images and label images are included in the training data input to an artificial neural network model that improves the quality of medical images.

[0081] According to the embodiments of this disclosure, the training data generation device can generate training data in different ways depending on the characteristics of the input image, thereby reducing the time and cost required to provide training data that satisfies specific conditions, and thus enabling rapid training of the artificial neural network model.

[0082] Figure 4 is a flowchart showing the detailed operation of step S130 in Figure 3 according to one embodiment of the present disclosure.

[0083] Referring to Figure 4, the training data generation device may be an embodiment of the training data generation device 10 in Figure 1 or the training data generation device 20 in Figure 6. Hereinafter, the first image and the second image may refer to images that are noise-independent of each other and have the same signal intensity.

[0084] The training data generator can set a noise reduction target to be applied to the artificial neural network model (S210). Exemplarily, the noise reduction target may be a real number between 0 and 1.

[0085] The training data generation device can generate a training image based on the first image (S220). For example, the training data generation device can determine the first image to be a training image.

[0086] The training data generator can generate a label image based on the first image, the second image, and a noise reduction target (S230). Specifically, the training data generator can generate a label image by combining the first image and the second image based on the noise reduction target x. Here, the noise reduction target x means a value set so that the noise is reduced by x times or the noise-to-signal ratio (SNR) is increased by 1 / x times.

[0087] In other words, the training image and label image are generated by the following equation 3.

[0088]

number

[0089] In Math 3, T represents the training image 310, L represents the label image 320, Input1 represents the first image 310, Input2 represents the second image 320, and x represents the noise reduction target.

[0090] On the other hand, the order of steps S220 and S230 can be changed, and each step can be executed in parallel.

[0091] On the other hand, the method for generating the training image and label image by combining the first and second images is not limited to this. That is, there can be various methods for ensuring that the signal intensities of the training image and label image are the same, and for setting the noise independence ratio or noise dependency ratio between the training image and label image based on the noise reduction target x.

[0092] Figure 5 is a flowchart showing the detailed operation of step S130 in Figure 3 according to one embodiment of the present disclosure.

[0093] Referring to Figure 5, the learning data generation device may be an embodiment of the learning data generation device 10 in Figure 1 or the learning data generation device 20 in Figure 6.

[0094] The training data generation device can measure the magnitude of the noise contained in the first image, i.e., the magnitude of the first noise, before or after step S310. The magnitude of the first noise may represent the standard deviation of the pixel values ​​contained in the first image. For example, if the first image is a magnetic resonance image, the background of the first image is segmented, and the magnitude of the first noise represents the standard deviation of the pixel values ​​of the segmented background. For example, the first image may be a k-space domain image, i.e., k-space data. Here, the magnitude of the first noise represents the standard deviation of the pixel values ​​located at a certain distance from the center of the first image.

[0095] The training data generation device can generate a label image based on the first image (S320). For example, the training data generation device can determine the first image to be the label image.

[0096] The training data generation device can generate a training image based on the first image and the noise reduction target (S330). Specifically, the training data generation device can generate a second noise that has the same magnitude as the first noise and is noise-independent of the first noise. The second noise can be random noise, which can mean noise based on a complex Gaussian distribution, a Rice distribution, or a noncentral chi distribution. Specifically, if the first image is k-space data, the second noise is generated based on a complex Gaussian distribution. If the first image is a magnetic resonance image, the second noise is generated based on a Rice distribution or a noncentral chi distribution.

[0097] The training data generation device can generate training images such that the training image and the label image are mutually dependent on noise, based only on the first image, second noise, and noise reduction target. That is, the training image and label image are generated by the following equation 4.

[0098]

number

[0099] In Math 4, T represents the training image 310, L represents the label image 320, Input1 represents the first image 310, x represents the noise reduction target, and n2 represents the second noise.

[0100] On the other hand, the order of steps S320 and S330 can be changed, and each step can be executed in parallel.

[0101] On the other hand, additional preprocessing operations can be performed on the first image. For example, operations such as transforming the domain of the first image can be performed.

[0102] Subsequently, the training data generation device can provide the training images and label images to the artificial neural network model. The artificial neural network model can then be trained to output high-quality medical images based on low-quality medical images using the training images and label images.

[0103] Figure 6 is a block diagram of a learning data generation device according to one embodiment of the present disclosure.

[0104] Referring to Figure 6, the training data generation device 20 is similar to the training data generation device 10 in Figure 2, so the common points are omitted. The training image 310 and label image 320 may be magnetic resonance images on the image domain. At least one input image 510 may be an image in the k-space domain, i.e., k-space data.

[0105] The learning data generation device 20 can generate k-space learning images 610 and k-space label images 620 in which the signal strengths are the same and the noise reduction target is the only noise dependency relationship between them.

[0106] The learning data generation device 20 can determine whether an image pair 511, 512 exists among at least one input image 500.

[0107] If the first image 511 and the second image 512 are present in the image pair 511 and 512, the training data generation device 20 combines the first image 511 and the second image 512 based on the noise reduction target to generate a k-space training image 610 and a k-space label image 620. The specific details are the same as the method described above in Figure 4, so the explanation will be omitted below.

[0108] If no image pair 510 exists or if the input image is a single first image 511, the training data generation device 10 generates a k-space training image 610 and a k-space label image 620 using the first image 511 according to the noise reduction target. The specific details are the same as the method described above in Figure 5, so the explanation will be omitted below.

[0109] On the other hand, the learning data generation device 20 can skip the step of determining whether image pairs 511 and 512 exist and generate a k-space learning image 610 and a k-space label image 620 using two images, or generate a k-space learning image 610 and a k-space label image 620 using one image.

[0110] The learning data generation device 20 can perform Fourier transforms on the k-space learning image 610 and the k-space label image 620, respectively, to generate a learning image 310 and a label image 320, which can then be provided to the artificial neural network model 400.

[0111] The descriptions of this disclosure set forth herein are illustrative, and a person with ordinary skill in the art to which this disclosure pertains will understand that it can be easily adapted to other specific forms without altering the technical idea or essential features of this disclosure. Therefore, the embodiments of this disclosure described above should be understood to be illustrative and not limiting in all respects. For example, each component described as a single type can be implemented in a distributed manner, and similarly, components described as distributed can be implemented in a combined form.

[0112] The scope of this disclosure is determined by the claims set forth below rather than by the detailed description above, and all forms of modification or alteration derived from the meaning, scope, and equivalent concepts of the claims should be interpreted as being included within the scope of this disclosure.

Claims

1. A method for generating training data, performed by a computing device including at least one processor, The steps include analyzing the signal strength and noise characteristics of at least one input image, The step of determining whether there is an image pair (pair) among the at least one input image in which the signal strength and noise characteristics correspond, The step includes generating a training image and a label image based on the judgment result according to a predetermined noise reduction target, The steps of generating the aforementioned learning image and the aforementioned label image are as follows: If it is determined that the aforementioned image pair exists, The first image included in the aforementioned image pair is determined to be the training image. The first image and the second image included in the image pair are combined according to the noise reduction target to generate the label image. If it is determined that the aforementioned image pair does not exist, The first image among the at least one input image is determined to be the label image. A method characterized by generating the learning image based on the first image and the noise reduction target.

2. The step of determining whether the aforementioned image pair exists is: The method according to claim 1, characterized in that a plurality of images having the same signal intensity and being noise-independent of each other are determined as the image pair.

3. The steps of generating the aforementioned learning image and the aforementioned label image are as follows: The method according to claim 2, characterized in that, if it is determined that the aforementioned image pair exists, the label image is generated such that the first noise of the learning image and the second noise of the label image are noise dependent only to the noise reduction target.

4. The steps of generating the aforementioned learning image and the aforementioned label image are as follows: If it is determined that the aforementioned image pair does not exist, The signal intensity of the learning image and the signal intensity of the first image are made to be the same. The method according to claim 2, characterized in that the learning image is generated such that the first noise of the first image and the second noise of the learning image are in a noise dependency relationship according to the noise reduction target.

5. The method according to claim 2, characterized in that the learning image and the label image are input into an artificial neural network model that improves the quality of medical images.

6. The step of generating the learning image and the label image is: The method according to claim 1, characterized in that if it is determined that the aforementioned image pair exists, the label image is generated by the following mathematical formula. [Math 1] L = the aforementioned label image x = Noise reduction target Input1 = First image Input2 = the second image

7. The steps of generating the aforementioned learning image and the aforementioned label image are as follows: If it is determined that the aforementioned image pair does not exist, A step of generating a second noise in which the first noise and signal intensity of the first image are identical and the first noise and the second noise are noise-independent of each other, A step of generating the learning image by combining the first image and the second noise according to the noise reduction target, The method according to claim 1, characterized by including

8. The method according to claim 7, wherein the step of generating the learning image by combining the first image and the second noise according to the noise reduction target is characterized in that the learning image is generated by the following mathematical formula. [Math 2] T = The aforementioned learning image Input1 = First image n2 = the second noise x = Noise reduction target

9. A learning data generation device, A memory that stores a predetermined noise reduction target and at least one input image, Equipped with a processor, The processor determines whether there is an image pair (pair) among the at least one input image whose signal strength and noise characteristics correspond. If the processor determines that an image pair exists because, among the at least one input image, there exists a second image whose signal intensity is the same as that of the first image and which is noise-independent of the first image, it determines the first image to be the learning image and generates a label image by combining the first image and the second image according to the noise reduction target. The apparatus is characterized in that, when the processor determines that the image pair does not exist because the second image does not exist, it determines the first image to be the label image and generates the learning image based on the first image and the noise reduction target.

10. The apparatus according to claim 9, characterized in that the learning image and the label image have the same signal intensity and are noise-dependent to each other to the extent that they correspond to the noise reduction target.

11. A computer program stored on a computer-readable storage medium, which, when executed on one or more processors, performs the following actions: The aforementioned operation is, The operation involves analyzing the signal strength and noise characteristics of at least one input image, The operation of determining whether there is an image pair (pair) among the at least one input image in which the signal strength and noise characteristics correspond, This includes the operation of generating training images and label images to be input to the artificial neural network model based on the judgment result and the noise reduction target set for the artificial neural network model, The steps of generating the aforementioned learning image and the aforementioned label image are as follows: If it is determined that the aforementioned image pair exists, The first image included in the aforementioned image pair is determined to be the training image. The first image and the second image included in the image pair are combined according to the noise reduction target to generate the label image. If it is determined that the aforementioned image pair does not exist, The first image among the at least one input image is determined to be the label image. A computer program characterized by generating a learning image based on the first image and the noise reduction target.

Images