Method, program and apparatus for generating electrocardiogram by using artificial intelligence

Abstract

Disclosed are a method, a program and an apparatus for generating an electrocardiogram by using artificial intelligence, the method being performed by a computing device, according to one embodiment of the present disclosure. The method may comprise the steps of: acquiring a first limb lead electrocardiogram generated by an electrocardiogram measurement device; inputting the acquired first limb lead electrocardiogram into a pre-trained first neural network model so as to generate a second limb lead electrocardiogram different from the first limb lead electrocardiogram; performing a vector operation on the basis of the first limb lead electrocardiogram and the second limb lead electrocardiogram so as to generate remaining limb lead electrocardiograms that exclude the first limb lead electrocardiogram and the second limb lead electrocardiogram; and inputting the first limb lead electrocardiogram, the second limb lead electrocardiogram, and the limb lead electrocardiogram generated through the vector operation into a pre-trained second neural network model so as to generate a chest lead electrocardiogram.

Description

Method, program and device for generating an electrocardiogram using artificial intelligence

[0001] The present disclosure relates to artificial intelligence technology in the medical field, and more particularly, to a method for generating an electrocardiogram of some leads of a standard 12-lead electrocardiogram based on the electrocardiogram of the remaining leads using artificial intelligence.

[0002]

[0003] An electrocardiogram (ECG) is a common, noninvasive data source for diagnosing cardiovascular disease, providing crucial evidence for conditions that are difficult to identify based solely on symptoms. To accurately measure an ECG in a hospital setting, a device is typically used, attaching three electrodes to the limbs and six to the chest. This process creates a standard 12-lead ECG, comprising six limb leads (Leads I, II, III, aVR, aVL, aVF) and six chest leads (Leads V1-V6). Physicians diagnose cardiovascular disease by analyzing the patterns observed in this 12-lead ECG.

[0004] Recently, health screenings utilizing wearable devices capable of measuring electrocardiograms (ECGs) have gained attention due to their convenience and accessibility. However, these wearable devices typically measure only a single lead, which limits their diagnostic accuracy compared to the standard 12-lead ECG used in hospitals. Therefore, accurately generating the remaining 11 leads using only a single lead obtained from a wearable device has become a significant technological challenge.

[0005] Generating the remaining 11 leads from a single lead is a very challenging problem in practice. In this regard, various attempts have been made to generate all 12 leads, even in environments where only three leads can be measured, such as an ambulance. These existing methods typically reconstructed the remaining leads using two limb leads and one chest lead. However, in the medical field, it is known that while generating the remaining limb (or chest) leads from some limb (or chest) leads is mathematically feasible, generating the chest lead from only limb leads, or only chest leads from only chest leads, is practically difficult. In this context, generating the remaining 11 leads from only a single limb lead measured by a wearable device becomes an even more challenging task.

[0006] To overcome this, research utilizing deep learning-based approaches has been actively conducted recently. While these methods have shown significant progress, several practical challenges still need to be addressed for practical application.

[0007]

[0008] The present disclosure aims to provide a method for generating an electrocardiogram of the remaining leads constituting the standard 12 leads from at least one lead obtained from various devices by combining a deep learning-based approach and a mathematical operation-based approach.

[0009] However, the problems to be solved in this disclosure are not limited to the problems mentioned above, and other problems not mentioned can be clearly understood based on the description below.

[0010]

[0011] According to one embodiment of the present disclosure for realizing the task described above, a method for generating an electrocardiogram using artificial intelligence, performed by a computing device, is disclosed. The method may include the steps of: obtaining a first limb-derived electrocardiogram generated by an electrocardiogram measuring device; inputting the obtained first limb-derived electrocardiogram into a pre-trained first neural network model to generate a second limb-derived electrocardiogram different from the first limb-derived electrocardiogram; performing a vector operation based on the first limb-derived electrocardiogram and the second limb-derived electrocardiogram to generate a limb-derived electrocardiogram remaining excluding the first limb-derived electrocardiogram and the second limb-derived electrocardiogram; and inputting the first limb-derived electrocardiogram, the second limb-derived electrocardiogram, and the limb-derived electrocardiogram generated through the vector operation into a pre-trained second neural network model to generate a chest-derived electrocardiogram.

[0012] Alternatively, the first neural network model and the second neural network model may include a first neural network block that learns the induced pattern of an electrocardiogram; and a second neural network block that generates an electrocardiogram based on the output of the first neural network block.

[0013] Alternatively, the first neural network block can generate a vector representation for the derivation of the input electrocardiogram, and a vector representation for the derivation of the target electrocardiogram to be generated through the second neural network block.

[0014] Alternatively, when there are N input electrocardiograms (N is a natural number greater than or equal to 2), the first neural network block can generate a single vector representation that includes the entire pattern of the derivation of the N input electrocardiograms.

[0015] Alternatively, the second neural network block can generate the target electrocardiogram based on the input electrocardiogram and a vector representation of the derivation of the target electrocardiogram.

[0016] Alternatively, the first neural network block and the second neural network block may perform a first learning process (phase) for generating an electrocardiogram corresponding to ground truth (GT) from a source electrocardiogram, and a second learning process for reconstructing the source electrocardiogram from an electrocardiogram generated through the first learning process.

[0017] Alternatively, the neural network parameters constituting the first neural network block and the second neural network block may be updated based on a combination of multiple loss functions during the entire learning process in which the first learning process and the second learning process are sequentially performed.

[0018] Alternatively, the plurality of loss functions may include a first loss function for minimizing the difference between the ground truth (GT) and the electrocardiogram of the derivation corresponding to the GT; a second loss function for minimizing the difference between the source electrocardiogram and the electrocardiogram reconstructed through the second learning process; and a third loss function for minimizing the difference between the vector representation of the derivation of the source electrocardiogram generated in the first learning process and the vector representation of the derivation of the source electrocardiogram generated based on the electrocardiogram of the derivation corresponding to the GT in the second learning process, and minimizing the difference between the vector representation of the derivation of the electrocardiogram corresponding to the GT generated based on the source electrocardiogram in the first learning process and the vector representation of the derivation of the electrocardiogram corresponding to the GT generated in the second learning process.

[0019] Alternatively, the first neural network block may include a first encoder that generates vector representations from L input electrocardiograms (where L is a natural number greater than or equal to 1); and a second encoder that receives the vector representations generated by the first encoder and generates vector representations for derivations of the L input electrocardiograms and vector representations for derivations of K target electrocardiograms (where K is a natural number greater than or equal to 1) to be generated through the second neural network block.

[0020] Alternatively, the second neural network block may include a third encoder that generates vector representations from the L input electrocardiograms; and a decoder that receives the vector representations generated by the encoder and the vector representations for the derivations of the K target electrocardiograms generated by the first neural network block as inputs, and generates the K target electrocardiograms.

[0021] Alternatively, the first limb lead ECG may be lead I of the standard limb leads and the second limb lead ECG may be lead II of the standard limb leads.

[0022] According to one embodiment of the present disclosure for achieving the above-described task, a computer program stored in a computer-readable storage medium is disclosed. When the computer program is executed on one or more processors, it performs operations for generating an electrocardiogram using artificial intelligence. At this time, the operations include: an operation of obtaining a first limb-induced electrocardiogram generated by an electrocardiogram measuring device; an operation of inputting the obtained first limb-induced electrocardiogram into a pre-trained first neural network model to generate a second limb-induced electrocardiogram different from the first limb-induced electrocardiogram; an operation of performing a vector operation based on the first limb-induced electrocardiogram and the second limb-induced electrocardiogram to generate a limb-induced electrocardiogram remaining from the first limb-induced electrocardiogram and the second limb-induced electrocardiogram; And it may include an operation of generating a chest-induced electrocardiogram by inputting the first limb-induced electrocardiogram, the second limb-induced electrocardiogram, and the limb-induced electrocardiogram generated through the vector operation into a pre-learned second neural network model.

[0023] According to one embodiment of the present disclosure for achieving the aforementioned task, a computing device for generating an electrocardiogram using artificial intelligence is disclosed. The device may include a processor including at least one core; a memory including program codes executable by the processor; and a network unit for acquiring a first limb-derived electrocardiogram generated by an electrocardiogram measuring device. At this time, the processor may input the obtained first limb-induced electrocardiogram into a pre-trained first neural network model to generate a second limb-induced electrocardiogram different from the first limb-induced electrocardiogram, perform a vector operation based on the first limb-induced electrocardiogram and the second limb-induced electrocardiogram to generate the remaining limb-induced electrocardiograms excluding the first limb-induced electrocardiogram and the second limb-induced electrocardiogram, and input the first limb-induced electrocardiogram, the second limb-induced electrocardiogram, and the limb-induced electrocardiogram generated through the vector operation into a pre-trained second neural network model to generate a chest-induced electrocardiogram.

[0024]

[0025] The present disclosure can accurately reconstruct a standard 12-lead electrocardiogram with high precision and reliability comparable to that used in a hospital setting using only a single lead measurable on a wearable device.

[0026] The present disclosure optimizes the order of induction generation through a framework that combines deep learning and vector operations to generate limb inductions and then generates thoracic inductions based on the limb inductions, thereby producing more accurate and stable results compared to existing methods that utilize only deep learning.

[0027]

[0028] FIG. 1 is a block diagram of a computing device according to one embodiment of the present disclosure.

[0029] FIG. 2 is a conceptual diagram illustrating an electrocardiogram generation process according to one embodiment of the present disclosure.

[0030] FIG. 3 is a block diagram illustrating a learning process of a neural network model according to one embodiment of the present disclosure.

[0031] FIG. 4 is a block diagram showing the structure of a first neural network block according to one embodiment of the present disclosure.

[0032] FIG. 5 is a block diagram showing the structure of a second neural network block according to one embodiment of the present disclosure.

[0033] Figures 6a, 6b and 7 are tables showing evaluation results for the generation accuracy of a framework according to one embodiment of the present disclosure.

[0034] FIG. 8 is a graph showing the results of comparing prediction accuracy based on data generated by a neural network model according to one embodiment of the present disclosure.

[0035] FIG. 9 is a flowchart illustrating a method for generating an electrocardiogram according to one embodiment of the present disclosure.

[0036]

[0037] Below, embodiments of the present disclosure are described in detail with reference to the attached drawings so that those skilled in the art can easily implement the present disclosure. The embodiments presented in this disclosure are provided to enable those skilled in the art to utilize 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 implemented in various different forms and is not limited to the embodiments described below.

[0038] Throughout the specification of this disclosure, identical or similar drawing numbers refer to identical or similar components. Furthermore, for the purpose of clearly describing the disclosure, drawing numbers for parts in the drawings that are not relevant to the description of the disclosure may be omitted.

[0039] The term "or" as used herein is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless otherwise specified herein or clear from context, "X employs A or B" should be understood to mean either of the natural inclusive permutations. For example, unless otherwise specified herein or clear from context, "X employs A or B" can be interpreted to mean either X employs A, X employs B, or X employs both A and B.

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

[0041] The terms "comprises" and / or "comprising" as used herein should be understood to mean the presence of certain features and / or components. However, it should be understood that the terms "comprises" and / or "comprising" do not exclude the presence or addition of one or more other features, other components, and / or combinations thereof.

[0042] Unless otherwise specified in this disclosure or unless the context makes it clear that the singular form is intended to be referred to, the singular should generally be construed to include “one or more.”

[0043] The term "Nth (N is a natural number)" used in the present disclosure can be understood as an expression used to mutually distinguish components of the present disclosure based on a predetermined standard such as a functional perspective, a structural perspective, or convenience of explanation. For example, components performing different functional roles in the present disclosure can be distinguished as a first component or a second component. However, components that are substantially the same within the technical spirit of the present disclosure but must be distinguished for convenience of explanation may also be distinguished as a first component or a second component.

[0044] The term "acquisition" as used in this disclosure may 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.

[0045] Meanwhile, the term "module" or "unit" used in the present disclosure can be understood as a term 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. At this time, the "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, as a narrow concept, a "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 processing process implemented through software execution, or a set of instructions for program execution, etc. In addition, as a broad concept, a "module" or "unit" may refer to the computing device itself that constitutes the system, or an application running on the computing device, etc. However, since the above-described concept is only an example, the concept of “module” or “part” may be defined in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure.

[0046] The term "model" as used herein 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 of a processing process to solve a specific problem. For example, a neural network "model" may refer to the entire system implemented as a neural network that has problem-solving capabilities through learning. In this case, the 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 a combination of multiple neural networks.

[0047] The explanation of the above terms is intended to aid understanding of the present disclosure. Therefore, unless explicitly stated as limiting the contents of the present disclosure, it should be noted that the above terms are not intended to limit the technical ideas of the contents of the present disclosure.

[0048] FIG. 1 is a block diagram of a computing device according to one embodiment of the present disclosure.

[0049] A computing device (100) 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 calculation of data, or 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 shares resources, or may be a client that shares resources through interaction with a server. In addition, the computing device (100) may be a cloud system in which multiple servers and clients interact to comprehensively process data. Since the above description is only one example related to 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.

[0050] Referring to FIG. 1, a computing device (100) according to one embodiment of the present disclosure may include a processor (110), a memory (120), and a network unit (130). However, FIG. 1 is merely an example, and the computing device (100) may include other components for implementing a computing environment. In addition, only some of the disclosed components may be included in the computing device (100).

[0051] The processor (110) according to one embodiment of the present disclosure may be understood as a configuration unit including hardware and / or software for performing computing operations. For example, the processor (110) may read a computer program to perform data processing for machine learning. The processor (110) may process computational processes such as processing input data for machine learning, feature extraction for machine learning, and error calculation based on backpropagation. The 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 type of the processor (110) described above is only one example, and thus, the type of the processor (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.

[0052] The processor (110) can train a neural network model to receive at least one of the standard 12-lead electrocardiograms as input and generate an electrocardiogram having different leads from the input electrocardiogram. The electrocardiogram reflects the characteristics of the heart from various perspectives depending on the direction of the electrodes attached to the body of the subject. Therefore, each lead of the standard 12-lead electrocardiogram exhibits a unique waveform pattern as a result of observing the electrical activity of the heart from different directions. The processor (110) can train a neural network model to generate an electrocardiogram of a target lead by reflecting the induction characteristics of the electrocardiogram. For example, the processor (110) can input an electrocardiogram of a portion of the training data consisting of the standard 12-lead electrocardiogram into the neural network model. The processor (110) can cause the neural network model to identify a specific induction pattern of the electrocardiogram to be generated from the induction pattern of the input electrocardiogram. In addition, the processor (110) can cause the neural network model to generate a target electrocardiogram based on the specific induction pattern identified by the neural network model. The processor (110) can calculate an error using multiple loss functions during this learning process and update the parameters constituting the neural network of the neural network model based on the calculated error. At this time, the neural network model may be a model that receives an electrocardiogram of N leads (N is a natural number greater than or equal to 1) and generates an electrocardiogram of N leads that is different from the input.

[0053] The processor (110) can generate an entire standard 12-lead electrocardiogram from at least one of the standard 12-lead electrocardiograms by combining a pre-trained neural network model and vector operations that generate limb-induced electrocardiograms as described above. For example, the processor (110) can input one of the six limb-induced electrocardiograms into a pre-trained first neural network model to generate one of the remaining five limb-induced electrocardiograms. The processor (110) can perform vector operations based on the limb-induced electrocardiograms input into the first neural network model and the limb-induced electrocardiograms generated through the first neural network model to generate the remaining four limb-induced electrocardiograms. At this time, the vector operations are based on cardiovascular equations utilizing the physical relationships of the limb leads, and can be performed based on at least one of equations derived from Einthoven's triangle principle or Goldberger's formula. The processor (110) can generate six chest-lead electrocardiograms by inputting six limb-lead electrocardiograms generated through vector operations with the first neural network model into a pre-trained second neural network model. At this time, the second neural network model may be an independent model having the same neural network structure as the first neural network model. Through this process, the processor (110) can generate all of the remaining leads constituting the standard 12-lead electrocardiogram from one limb-lead electrocardiogram. Through this efficient combination of deep learning and vector operations, the processor (110) can provide a standard 12-lead electrocardiogram capable of making a diagnosis at the same level and accuracy as a conventional hospital environment even in an environment with limited electrocardiogram measurement, such as a wearable device.

[0054] The memory (120) according to one embodiment of the present disclosure may be understood as a configuration unit including hardware and / or software for storing and managing data processed in the computing device (100). That is, the memory (120) may store any type of data generated or determined by the processor (110) and any type of data received by the network unit (130). For example, the memory (120) may include at least one type of storage medium among a flash memory type, a hard disk type, a multimedia card micro type, a card type memory, a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, and an optical disk. In addition, the memory (120) may also include a database system that controls and manages data in a predetermined system. The type of memory (120) described above is only one example, and thus the type of memory (120) can be configured in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure.

[0055] The memory (120) can structure and organize and manage data, combinations of data, and program codes executable by the processor (110) required for the processor (110) to perform operations. For example, the memory (120) can store medical data received through the network unit (130) described below. The memory (120) can store program codes that operate a neural network model to receive medical data as input and perform learning, program codes that operate a neural network model to receive medical data as input and perform inference according to the purpose of use of the computing device (100), and processed data generated as the program codes are executed.

[0056] The network unit (130) according to one embodiment of the present disclosure may be understood as a component that transmits and receives data through any type of known wired or wireless communication system. For example, the network unit (130) 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), fifth generation mobile communication (5G), ultra wide-band, ZigBee, radio frequency (RF) communication, wireless LAN, wireless fidelity, near field communication (NFC), or Bluetooth. Since the above-described communication systems are only examples, the wired and wireless communication system for data transmission and reception of the network unit (130) may be applied in various ways other than the above-described examples.

[0057] The network unit (130) can receive data necessary for the processor (110) to perform calculations through wired or wireless communication with any system or any client, etc. In addition, the network unit (130) can transmit data generated through calculations of the processor (110) through wired or wireless communication with any system or any client, etc. For example, the network unit (130) can receive electrocardiogram data through communication with a cloud server that performs tasks such as databases in a hospital environment, standardization of medical data, clients such as smart watches, or medical computing devices, etc. The network unit (130) can transmit output data of a neural network model, intermediate data derived from the calculation process of the processor (110), processed data, etc. through communication with the aforementioned database, server, client, or computing device, etc.

[0058] FIG. 2 is a conceptual diagram illustrating an electrocardiogram generation process according to one embodiment of the present disclosure.

[0059] A computing device (100) according to one embodiment of the present disclosure can input a first limb-induced electrocardiogram (11) into a pre-learned first neural network model (210) to generate a second limb-induced electrocardiogram (13) that is different from the first limb-induced electrocardiogram (11). The first neural network model (210) can generate the second limb-induced electrocardiogram (13) by converting the induction pattern of the first limb-induced electrocardiogram (11) into an induction pattern in a different direction. For example, when the first limb-induced electrocardiogram (11) is induction I among standard limb inductions, the first neural network model (210) can generate a vector representation representing an induction pattern corresponding to induction II, which is a target induction, based on a vector representation representing the pattern of induction I. At this time, the vector representation can be understood as an embedding vector value that represents a pattern or feature of an electrocardiogram induction as an array of numbers. The first neural network model (210) can generate a second limb-induced electrocardiogram (13) corresponding to lead II by transforming the first limb-induced electrocardiogram (11) based on the generated vector representation.

[0060] The computing device (100) can perform a vector operation (220) based on the first limb-lead electrocardiogram (11) and the second limb-lead electrocardiogram (13) to generate the remaining limb-lead electrocardiograms (15) excluding the first limb-lead electrocardiogram (11) and the second limb-lead electrocardiogram (13). For example, if the first limb-lead electrocardiogram (11) is lead I among the standard limb leads and the second limb-lead electrocardiogram (13) is lead II among the standard limb leads, the computing device (100) can perform the vector operation (220) using the following [Mathematical Formula 1]. The computing device (100) can generate electrocardiograms corresponding to the remaining four limb leads excluding leads I and II among the six limb leads through this vector operation (220).

[0061] [Mathematical Formula 1]

[0062] Induction III = Induction II - Induction I,

[0063] Induction aVR = -(Induction I + Induction II) / 2,

[0064] Lead aVL = (Lead I - Lead III) / 2,

[0065] Lead aVF = (Lead II + Lead III) / 2

[0066] The computing device (100) can input the first limb-lead electrocardiogram (11), the second limb-lead electrocardiogram (13), and the limb-lead electrocardiogram (15) generated through vector operation (220) into a pre-learned second neural network model (230), thereby generating a chest-lead electrocardiogram (16). The second neural network model (230) can generate six chest-lead electrocardiograms (17) by converting the induction patterns of the six limb-lead electrocardiograms (11, 13, 15) into induction patterns in different directions. For example, the second neural network model (230) can generate vector representations representing induction patterns corresponding to leads V1 to V6, respectively, based on vector representations representing the patterns of leads I, II, III, aVR, aVL, and aVF, respectively. At this time, the vector representation can be understood as an embedding vector value that expresses the pattern or characteristic of the electrocardiogram induction as an array of numbers. The second neural network model (230) can generate six thoracic limb-induced electrocardiograms (17) by modifying the first limb-induced electrocardiogram (11) based on the generated vector representations.

[0067] The computing device (100) can generate a standard 12-lead electrocardiogram by generating the remaining limb-lead electrocardiograms and chest-lead electrocardiograms from a single limb-lead electrocardiogram using a hybrid model that combines the strengths of deep learning and vector operations as described above. The step-by-step framework for generating chest leads from limb leads utilizing the hybrid model of the present disclosure can achieve the effect of generating an electrocardiogram that is more accurate and reliable than generating the remaining 11 leads that constitute the standard 12 leads from a single-lead electrocardiogram using only deep learning.

[0068] FIG. 3 is a block diagram illustrating a learning process of a neural network model according to one embodiment of the present disclosure.

[0069] The first neural network model (210) and the second neural network model (230) according to one embodiment of the present disclosure may have a common structure in that they are M-to-M (M is a natural number) models. Referring to FIG. 3, the first neural network model (210) and the second neural network model (230) may include a first neural network block (310) that learns an electrocardiogram induction pattern and a second neural network block (320) that generates an electrocardiogram based on the output of the first neural network block. By utilizing the idea that each induction pattern reflects a specific cardiac point in time, the first neural network block (310) can learn how to transform a pattern from one induction to another. In addition, the second neural network block (320) can generate a target induction from an input induction using the transformation pattern learned by the first neural network block (310).

[0070] The learning process of the neural network model according to one embodiment of the present disclosure can be divided into a first learning phase for generating a target electrocardiogram (30) derived from a source electrocardiogram (20) corresponding to a ground truth (GT) (25), and a second learning phase for reconstructing the source electrocardiogram (20) from the target electrocardiogram (30) generated through the first learning phase. In the first learning phase, the first neural network block (310) can receive the source electrocardiogram (20) as input, and generate a vector representation (40) for the derivation of the source electrocardiogram (20) and a vector representation (45) for the derivation of the target electrocardiogram (30), respectively. In addition, the second neural network block (320) can use the source electrocardiogram (20) and the vector representation (45) for the derivation of the target electrocardiogram (30) as input, and generate the target electrocardiogram (30). In the second learning process, the first neural network block (310) can receive a target electrocardiogram (30) as input and generate a vector representation (50) for the derivation of the source electrocardiogram (20) and a vector representation (55) for the derivation of the target electrocardiogram (30). Then, the second neural network block (320) can use the target electrocardiogram (30) and the vector representation (50) for the derivation of the source electrocardiogram (20) as inputs to generate a reconstructed electrocardiogram (35). The reconstructed electrocardiogram (35) can be understood as a result of the second neural network block (320) restoring the source electrocardiogram (20) from the target electrocardiogram (30).

[0071] The neural network parameters constituting the first neural network block (310) and the second neural network block (320) can be updated based on a combination of multiple loss functions during the entire learning process in which the first learning process and the second learning process are sequentially performed. Specifically, the first loss function can be used to minimize the difference between the GT (25) and the target electrocardiogram (30) in the first learning process. For example, the first loss function can be a root mean squared error (RMSE)-based loss function that uses the GT (25) and the target electrocardiogram (30) as input variables. The second loss function can be used to minimize the difference between the source electrocardiogram (20) and the reconstructed electrocardiogram (35) in the second learning process. For example, the second loss function can be a root mean squared error-based loss function that uses the source electrocardiogram (20) and the reconstructed electrocardiogram (35) as input variables. The third loss function can be used to minimize the difference between the vector representation (40) for the derivation of the source electrocardiogram (20) generated in the first learning process and the vector representation (50) for the derivation of the source electrocardiogram (20) generated from the target electrocardiogram (30) through the second learning process, and to minimize the difference between the vector representation (45) for the derivation of the target electrocardiogram (30) generated based on the source electrocardiogram (20) in the first learning process and the vector representation (55) for the derivation of the target electrocardiogram (30) generated in the second learning process. For example, the third loss function can be composed of a mean absolute error (MAE)-based loss function that uses vector representations (40, 50) related to the source electrocardiogram (20) as input variables and a mean absolute error-based loss function that uses vector representations (45, 55) related to the target electrocardiogram (30) as input variables. A third loss function like this can induce the induced information (45, 50) converted through the first neural network block (310) to be aligned with the actual information (40, 55) obtained from each induced information.This results in the learned vector representation better reflecting the actual induction pattern, thereby improving the quality and reliability of the target electrocardiogram (30).

[0072] FIG. 4 is a block diagram showing the structure of a first neural network block according to one embodiment of the present disclosure, and FIG. 5 is a block diagram showing the structure of a second neural network block according to one embodiment of the present disclosure.

[0073] Referring to FIG. 4, a first neural network block (310) according to one embodiment of the present disclosure may include a first encoder (311) that generates vector representations from L input electrocardiograms (L is a natural number greater than or equal to 1), and a second encoder (312) that receives the vector representations generated by the first encoder (311) and generates vector representations (63) for the induction of the L input electrocardiograms and vector representations (66, 69) for the induction of K target electrocardiograms (K is a natural number greater than or equal to 1) to be generated through the second neural network block (320). In this case, when two or more electrocardiograms are input, the first neural network block (310) may generate one vector representation (63) that includes the entire patterns of the two or more input electrocardiograms.

[0074] For example, the first neural network block (310) may include a first encoder (311) including a one-dimensional convolutional layer and a second encoder (312) including K+1 multilayer perceptrons. L input electrocardiograms may be converted into a single vector representation through the first encoder (311). Then, the representation may be flattened and input to each of the K+1 multilayer perceptrons to generate a vector representation (63) representing a collective pattern for the induction of the L input electrocardiograms and an individual vector representation (66, 69) for each of the inductions of the K target electrocardiograms. Meanwhile, in FIG. 4, the input electrocardiogram (60) of the first neural network block (310) corresponds to a concept encompassing the first learning process and the second learning process. Accordingly, the input electrocardiogram (60) of the first neural network block (310) may correspond to the source electrocardiogram (20) or target electrocardiogram (30) of FIG. 3.

[0075] Referring to FIG. 5, a second neural network block (320) according to an embodiment of the present disclosure may include a third encoder (321) that generates vector representations from L input electrocardiograms (60), and a decoder that receives the vector representations generated by the third encoder (321) and the vector representations (66, 69) for the induction of K target electrocardiograms (70) generated by the first neural network block (310) as inputs, and generates K target electrocardiograms (70). At this time, in order to solve the problem of the size of the input and the output not being consistent in the first learning process and the second learning process, the second neural network block (320) may duplicate the input or the output to match the number of inductions of the input and the number of inductions of the output.

[0076] For example, the second neural network block (320) may include a third encoder (321) and a decoder (322) including a one-dimensional convolutional layer. L input electrocardiograms may be converted into vector representations of the derivations of the L input electrocardiograms through the third encoder (321). Adaptive instance normalization (AdaIN) may be applied to individual vector representations (66, 69) for each of the derivations of the K target electrocardiograms. The decoder (322) may generate K target electrocardiograms based on the vector representations generated through the third encoder (321) and the patterns injected through the adaptive instance normalization. Meanwhile, in FIG. 4, the input electrocardiogram (60) and the target electrocardiogram (70) of the second neural network block (320) correspond to concepts encompassing the first learning process and the second learning process. Accordingly, the input electrocardiogram (60) of the second neural network block (320) may correspond to the source electrocardiogram (20) or the target electrocardiogram (30) of FIG. 3. In addition, the target electrocardiogram (70) of the second neural network block (320) may correspond to the target electrocardiogram (30) or the reconstructed electrocardiogram (35) of FIG. 3.

[0077] Below, we summarize the results of a performance evaluation conducted to verify the effectiveness of the framework based on the neural network model of the present disclosure, which has the structure described above. The experimental conditions for the performance evaluation are as follows.

[0078] (1) Data set

[0079] We evaluated the performance of the proposed framework using two standard and widely used datasets provided by PhysioNet: PTB-XL and CUSPH. These datasets contain standard 12-lead electrocardiograms measured in hospitals and the types of cardiac diseases diagnosed by physicians based on ECG signals. They also contain demographic data such as patient gender and age. All data consist of 10-second ECG recordings at a sampling frequency of 500 Hz, resulting in a total of 5,000 sampling points per patient. PTB-XL was measured by the Physikalisch-Technische Bundesanstalt and contains information distributed by PhysioNet, as well as general metadata such as ECG recordings, diagnosis names, diagnosis codes, and age, sex, weight, height, and recording date. CUSPH was measured at Chapman University and Shaoxing People's Hospital and includes electrocardiogram records, diagnostic information, and general metadata such as age and gender, as well as information distributed by PhysioNet.

[0080] (2) Competitive techniques

[0081] The effectiveness of the framework of this disclosure was verified by comparing it with the following three deep learning techniques. ECG-LSTM is a generative model that regards ECG as time-series data and generates desired inductions in a sequence-to-sequence manner using Long Short-Term Memory (LSTM), a representative model for time-series learning. ECG-GAN is a model that uses a conditional GAN ​​(Generic Adversarial Network) to generate desired inductions from given inductions. ECG-T2T is a generative model based on an image-to-image transformation model that generates multiple target inductions simultaneously.

[0082] The results of the performance evaluation conducted by conducting an experiment based on the above-described setting conditions are as follows.

[0083] (1) Given induction I, comparison of the generation accuracy of the remaining 11 standard inductions (II, III, aVR, aVL, aVF, V1-V6)

[0084] Figures 6A and 6B summarize the accuracy when generating the remaining 11 standard leads based on Lead I. It can be seen that the framework of the present disclosure significantly outperforms all competing techniques in almost all leads. Based on the PTB-XL dataset, it showed an average improvement of 17%, 30%, 21%, and 29% in RMSE, DTW, PCC, and KLD, respectively. Based on the CUSPH dataset, it showed an improvement of 17%, 17%, 20%, and 36%, respectively. In particular, for the thoracic leads (V1-V6), the framework of the present disclosure showed an average accuracy improvement of 11%, 28%, 7%, and 19% in RMSE, DTW, PCC, and KLD, respectively. Even in the worst case, the framework of the present disclosure outperforms the best-performing ECG T2T technique among the competing techniques in terms of RMSE and DTW. This demonstrates the clear superiority of the framework of the present disclosure.

[0085] (2) Comparison of R-peak generation accuracy

[0086] ECGs are often distorted by noise such as power line interference, myokymia, and baseline wandering. Traditional metrics used as performance evaluation metrics, such as RMSE, DTW, PCC, and KLD, may not accurately reflect performance in these noisy environments. Therefore, we evaluated the accuracy of R-peak generation, a key reference point in ECGs. Specifically, we measured the accuracy of the R-peak location and amplitude (size).

[0087] Inductions V1 and V5 were selected for evaluation. Induction V1 is known to be the most difficult to generate because it is located perpendicular to induction I, whereas induction V5 is considered relatively easy to generate because it is measured in the direction most similar to induction I.

[0088] Figure 7 summarizes the R-peak generation accuracy of the framework of the present disclosure and competing techniques, evaluating them based on amplitude gap and position error. Experimental results show that the framework of the present disclosure consistently outperforms all competing techniques in both metrics. This demonstrates the framework's ability to more precisely capture a key feature (R-peak) in ECG signals.

[0089] (3) Comparison of disease diagnosis accuracy: myocardial infarction and arrhythmia

[0090] We compared the diagnostic accuracy of myocardial infarction and arrhythmia using 12-lead electrocardiograms generated by the framework of this disclosure and competing techniques. To ensure a fair comparison, disease prediction accuracy was evaluated based on both the original 12-lead electrocardiogram measured in a hospital and the 12-lead electrocardiograms generated from lead I by each method. The performance metrics used were the area under the curve (AUROC) and the area under the precision-recall curve (AUPRC). The disease prediction model used was ResNet18, an image classification model.

[0091] Figure 8 shows the accuracy of diagnosing myocardial infarction and arrhythmia based on the original ECG measured in a hospital and the ECG generated in Lead I by three competing techniques and the framework of the present disclosure. For myocardial infarction, the framework of the present disclosure outperformed all three competing techniques, with an accuracy lower by only 0.63% in AUROC and 0.11% in AUPRC compared to the original ECG. For arrhythmia, the framework of the present disclosure showed a slight accuracy difference of 0.71% in AUROC and 1.55% in AUPRC compared to the original ECG, still outperforming the competing techniques.

[0092] These results suggest that the framework of this disclosure can nearly reproduce the disease information of an original 12-lead electrocardiogram measured in a hospital setting. In other words, the framework of this disclosure can generate an electrocardiogram with diagnostic capabilities comparable to a full 12-lead electrocardiogram using just a single lead, demonstrating its high practicality and clinical utility.

[0093] (4) Similarity with actual hospital measured electrocardiogram signal

[0094] The framework of this disclosure consistently demonstrated results closer to the original electrocardiogram than competing techniques throughout the experiment. The framework of this disclosure generates a QRS complex that most closely resembles the original electrocardiogram. The QRS complex is the most important feature of the electrocardiogram and consists of the R wave, the highest part of the waveform, and the Q and S waves on either side of it, containing key information about the cardiac electrical signal. Even in leads (e.g., leads III and aVL) that competing techniques struggle to generate, the framework of this disclosure maintains sufficiently high tracking accuracy. Furthermore, the framework of this disclosure accurately tracks the ECG waveform even in diseased samples, contributing to further improved disease diagnosis accuracy. These results demonstrate that the framework of this disclosure goes beyond simply generating similar signals and possesses the ability to reproduce precise features of the ECG that are useful for actual diagnosis.

[0095] (5) Robustness to data sets

[0096] To verify the robustness of the framework of this disclosure against external datasets, we conducted experiments using different training and testing datasets. Specifically, the neural network model of this disclosure was trained on the CUSPH dataset and then tested on the PTB-XL dataset. The results were compared with those obtained when training and testing were performed on the same dataset. The main results are summarized in [Table 1].

[0097] TrainingTestRMSEDTWPCCKLDCUSPHCUSPH0.60611.8290.6384.06E-04PTB-XL0.47611.2160.4892.87E-04

[0098] Despite the change in test dataset, there was no significant difference in accuracy compared to training and testing on the same dataset. This consistency in performance demonstrates the strong generalization ability of the framework of this disclosure across datasets. This suggests that the well-trained neural network model of this disclosure can be effectively used in new environments without retraining. This means that the framework of this disclosure is a highly reliable and applicable tool that can be used stably in real-world environments, demonstrating its potential for practical use in the field of electrocardiogram analysis.

[0099] (6) Effectiveness of vector operations for limb-induced generation of the framework of the present disclosure

[0100] To verify the effectiveness of the vector operation and derivation generation steps within the framework of the present disclosure, we compared the performance of the framework of the present disclosure with two modified versions. Version 1 is a method that generates all remaining 11 derivations from derivation I at once using only the neural network model of the present disclosure. Version 2: A method that omits vector operations for limb derivation generation, generates five limb derivations from a single limb derivation using the neural network model of the present disclosure, and then generates six thoracic derivations from the six limb derivations.

[0101] [Table 2] summarizes the performance of the present disclosure framework and its two variants based on RMSE, DTW, PCC, and KLD metrics.

[0102] RMSEDTWPCCKLDVersion 10.1335.4620.7074.04E-06Version 20.1315.3220.7083.50E-06Framework of this disclosure0.1275.2340.7223.42E-06

[0103] The framework of this disclosure showed significantly higher accuracy than the two variants in all metrics. In particular, compared to version 1, the framework of this disclosure showed improvements of 4.32% in RMSE, 4.18% in DTW, 2.19% in PCC, and 15.3% in KLD. Compared to version 2, the framework of this disclosure showed improvements of 3.36% in RMSE, 1.64% in DTW, 2.06% in PCC, and 2.29% in KLD. These results clearly demonstrate that vector operations in the framework of this disclosure contribute to substantial performance improvements in limb-guided generation, and emphasize the effectiveness of the sophisticated design of the framework of this disclosure, including the complex guidance generation step. In other words, the combination of vector operations and learning yields significantly better results than simple end-to-end learning, which is experimentally supported.

[0104] (7) Accuracy according to the type of limb guidance generated through the neural network model of the present disclosure

[0105] We conducted an analysis to determine which limb guidance was most effective. [Table 3] summarizes the accuracy of the framework of this disclosure based on RMSE, DTW, PCC, and KLD criteria, depending on which limb guidance was generated.

[0106] RMSEDTWPCCKLDInduction II0.1275.2150.7223.42E-06Induction III0.2038.0050.5198.94E-06Induction aVR0.2589.8960.5289.59E-06Induction aVL0.2108.5740.5289.59E-06Induction aVF0.2138.5970.5229.68E-06

[0107] Lead II showed the highest accuracy. This supports the claim that generating Lead II using the neural network model of this disclosure is the most appropriate method, both medically and technically. In fact, it is well known in the medical community that Lead II generally contains the most cardiac information, and this result experimentally confirms this assertion.

[0108] (8) Effectiveness of the loss function of the present disclosure

[0109] The neural network model of this disclosure is trained using a loss function composed of a combination of three loss functions. To evaluate the effectiveness of this loss function, we compared the performance of the framework of this disclosure with two variants. Version 1 uses only the first and third loss functions, without the second loss function. Version 2 uses only the first and second loss functions, without the third loss function.

[0110] Additionally, to verify whether techniques commonly used in general image generation tasks are also useful for ECG generation, we conducted the following additional experiments. Version 3 added a discriminator and adversarial loss. Version 4 added random noise to the vector representation to increase generation diversity.

[0111] [Table 4] summarizes the results of comparing the performance of the framework of this disclosure and the four modified versions above based on RMSE, DTW, PCC, and KLD criteria.

[0112] RMSEDTWPCCKLDVersion 10.1445.8050.6604.87E-06Version 20.1435.6890.6704.65E-06Version 30.1295.2890.7183.58E-06Version 40.1275.2470.7213.60E-06Framework of this disclosure0.1275.2340.7223.40E-06

[0113] Version 1 and Version 2 show performance degradation of 10.9% and 8.69%, respectively, based on the DTW criterion, which emphasizes that the second and third loss functions are essential for generating 12-lead ECGs. Versions 3 and 4, on the other hand, cause performance degradation, which suggests that common techniques that are effective for image generation are not suitable for generating precise biosignals such as ECGs. This experiment clearly demonstrates that the customized loss function of the present disclosure is very effective for ECG generation, and that simply borrowing an image generation technique does not help improve the quality of ECG generation. In other words, it can be seen that the loss function of the present disclosure that considers the physiological characteristics of ECG signals is experimentally proven to be one of the core features of the present disclosure.

[0114] The performance evaluation results described above demonstrate that an approach solely utilizing deep learning has limitations in generating a full 12-lead ECG from a single lead, and that utilizing vector operations for limb lead generation, as in the framework of the present disclosure, is more effective. Furthermore, the method of first generating limb leads and then generating chest leads based on them, as in the framework of the present disclosure, demonstrates higher accuracy than generating all leads simultaneously.

[0115] FIG. 9 is a flowchart illustrating a method for generating an electrocardiogram according to one embodiment of the present disclosure.

[0116] Referring to FIG. 9, a computing device (100) according to one embodiment of the present disclosure can acquire a first limb lead electrocardiogram generated by an electrocardiogram measuring device (S100). For example, the computing device (100) can acquire a lead I electrocardiogram measured by a wearable device, such as a smartwatch equipped with electrodes. In this case, acquisition of the lead I electrocardiogram can be performed through wired or wireless communication between the computing device (100) and the wearable device.

[0117] The computing device (100) can input the first limb-induced electrocardiogram obtained through S100 into a pre-trained first neural network model to generate a second limb-induced electrocardiogram that is different from the first limb-induced electrocardiogram (S200). For example, the computing device (100) can input the lead I electrocardiogram into the first neural network model to generate a vector representation of the pattern of lead II based on the pattern of lead I. In addition, the computing device (100) can generate the lead II electrocardiogram from the lead I electrocardiogram based on the vector representation of the pattern of lead II through the first neural network model.

[0118] The computing device (100) can perform vector operations based on the first limb-lead electrocardiogram and the second limb-lead electrocardiogram to generate the remaining limb-lead electrocardiograms excluding the first limb-lead electrocardiogram and the second limb-lead electrocardiogram (S300). For example, the computing device (100) can generate each of leads III, aVR, aVL, and aVF using a cardiovascular equation that uses lead I and lead II electrocardiograms as input variables.

[0119] The computing device (100) can input the first limb-induced electrocardiogram, the second limb-induced electrocardiogram, and the limb-induced electrocardiogram generated through vector operations into a pre-trained second neural network model to generate a chest-induced electrocardiogram (S400). For example, the computing device (100) can input lead I obtained through S100, lead II generated through the first neural network model, leads III, aVR, aVL, and aVF generated through vector operations into the second neural network model to generate a total of six chest-induced electrocardiograms from V1 to V6.

[0120] The various embodiments of the present disclosure described above can be combined with additional embodiments and modified within the scope understood by those skilled in the art in light of the detailed description above. It should be understood that the embodiments of the present disclosure are illustrative in all respects and not restrictive. For example, each component described as a single component may be implemented in a distributed manner, and likewise, components described as distributed may be implemented in a combined manner. Accordingly, all changes or modifications derived from the meaning, scope, and equivalent concepts of the claims of the present disclosure should be construed as being included within the scope of the present disclosure.

Claims

1. A method for generating an electrocardiogram using artificial intelligence, performed by a computing device including at least one processor, A step of obtaining a first limb lead electrocardiogram generated by an electrocardiogram measuring device; A step of inputting the obtained first limb-induced electrocardiogram into a pre-trained first neural network model to generate a second limb-induced electrocardiogram different from the first limb-induced electrocardiogram; A step of performing a vector operation based on the first limb-induced electrocardiogram and the second limb-induced electrocardiogram to generate the remaining limb-induced electrocardiograms excluding the first limb-induced electrocardiogram and the second limb-induced electrocardiogram; and A step of generating a chest-lead electrocardiogram by inputting the first limb-induced electrocardiogram, the second limb-induced electrocardiogram, and the limb-lead electrocardiogram generated through the vector operation into a pre-trained second neural network model; including, method.

2. In paragraph 1, The above first neural network model and the above second neural network model, A first neural network block that learns the induced pattern of the electrocardiogram; and A second neural network block that generates an electrocardiogram based on the output of the first neural network block; including, method.

3. In paragraph 2, The above first neural network block is, Generating a vector representation for the derivation of the input electrocardiogram, and a vector representation for the derivation of the target electrocardiogram to be generated through the second neural network block. method.

4. In paragraph 3, If the above input electrocardiograms are N (N is a natural number greater than or equal to 2), The above first neural network block is, Generating a single vector representation containing the entire pattern for the derivation of the N input electrocardiograms, method.

5. In paragraph 3, The above second neural network block is, Generating the target electrocardiogram based on the vector representation of the input electrocardiogram and the derivation of the target electrocardiogram, method.

6. In paragraph 5, The above first neural network block and the above second neural network block, A first learning process (phase) for generating an electrocardiogram corresponding to ground truth (GT) from a source electrocardiogram and a second learning process for reconstructing the source electrocardiogram from the electrocardiogram generated through the first learning process are performed. method.

7. In paragraph 6, The neural network parameters constituting the first neural network block and the second neural network block are: In the entire learning process in which the first learning process and the second learning process are sequentially performed, the update is based on a combination of multiple loss functions. method.

8. In paragraph 7, The above multiple loss functions are, A first loss function to minimize the difference between the ground truth (GT) and the electrocardiogram corresponding to the GT; A second loss function for minimizing the difference between the source electrocardiogram and the electrocardiogram reconstructed through the second learning process; and A third loss function for minimizing the difference between the vector representation of the derivation of the source electrocardiogram generated in the first learning process and the vector representation of the derivation of the source electrocardiogram generated based on the derivation of the electrocardiogram corresponding to the GT in the second learning process, and for minimizing the difference between the vector representation of the derivation of the electrocardiogram corresponding to the GT generated based on the source electrocardiogram in the first learning process and the vector representation of the derivation of the electrocardiogram corresponding to the GT generated in the second learning process; including, method.

9. In paragraph 2, The above first neural network block is, A first encoder that generates a vector representation from L input electrocardiograms (where L is a natural number greater than or equal to 1); and A second encoder that receives the vector representation generated from the first encoder as input, and generates a vector representation for the derivation of the L input electrocardiograms, and a vector representation for the derivation of K target electrocardiograms (K is a natural number greater than or equal to 1) to be generated through the second neural network block; including, method.

10. In paragraph 9, The above second neural network block is, a third encoder for generating a vector representation from the L input electrocardiograms; and A decoder that receives as input the vector representation generated by the third encoder and the vector representation for the induction of the K target electrocardiograms generated by the first neural network block, and generates the K target electrocardiograms; including, method.

11. In paragraph 1, The above first limb lead ECG is lead I of the standard limb leads, The above second limb lead ECG is lead II among the standard limb leads. method.

12. A computer program stored in a computer-readable storage medium, wherein the computer program, when executed on one or more processors, performs operations for generating an electrocardiogram using artificial intelligence. The above actions are, An act of obtaining a first limb lead electrocardiogram generated by an electrocardiogram measuring device; An operation of inputting the obtained first limb-induced electrocardiogram into a pre-trained first neural network model to generate a second limb-induced electrocardiogram that is different from the first limb-induced electrocardiogram; An operation of performing a vector operation based on the first limb-induced electrocardiogram and the second limb-induced electrocardiogram to generate the remaining limb-induced electrocardiograms excluding the first limb-induced electrocardiogram and the second limb-induced electrocardiogram; and An operation of generating a chest-lead electrocardiogram by inputting the first limb-lead electrocardiogram, the second limb-lead electrocardiogram, and the limb-lead electrocardiogram generated through the vector operation into a pre-trained second neural network model; including, Computer program.

13. A computing device for generating an electrocardiogram using artificial intelligence, A processor comprising at least one core; A memory containing program codes executable by the processor; and A network unit for obtaining a first limb lead electrocardiogram generated from an electrocardiogram measuring device; Including, The above processor, By inputting the first limb-induced electrocardiogram obtained above into the pre-trained first neural network model, a second limb-induced electrocardiogram different from the first limb-induced electrocardiogram is generated, By performing a vector operation based on the first limb-induced electrocardiogram and the second limb-induced electrocardiogram, the remaining limb-induced electrocardiograms excluding the first limb-induced electrocardiogram and the second limb-induced electrocardiogram are generated. By inputting the first limb-induced electrocardiogram, the second limb-induced electrocardiogram, and the limb-induced electrocardiogram generated through the vector operation into the pre-trained second neural network model, a chest-induced electrocardiogram is generated. device.

Images