WO2023048437A1 - 의료 데이터를 기반으로 하는 딥러닝 모델의 학습 및 추론 방법, 프로그램 및 장치 - Google Patents
의료 데이터를 기반으로 하는 딥러닝 모델의 학습 및 추론 방법, 프로그램 및 장치 Download PDFInfo
- Publication number
- WO2023048437A1 WO2023048437A1 PCT/KR2022/013833 KR2022013833W WO2023048437A1 WO 2023048437 A1 WO2023048437 A1 WO 2023048437A1 KR 2022013833 W KR2022013833 W KR 2022013833W WO 2023048437 A1 WO2023048437 A1 WO 2023048437A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- neural network
- network model
- function
- operation function
- medical data
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 67
- 238000013136 deep learning model Methods 0.000 title claims abstract description 30
- 238000012549 training Methods 0.000 title claims abstract description 13
- 230000006870 function Effects 0.000 claims abstract description 257
- 238000003062 neural network model Methods 0.000 claims abstract description 226
- 238000013528 artificial neural network Methods 0.000 claims abstract description 145
- 230000036541 health Effects 0.000 claims description 29
- 238000013527 convolutional neural network Methods 0.000 claims description 19
- 238000004364 calculation method Methods 0.000 claims description 15
- 238000006243 chemical reaction Methods 0.000 claims description 13
- 230000001131 transforming effect Effects 0.000 claims description 8
- 230000000306 recurrent effect Effects 0.000 claims description 7
- 230000002123 temporal effect Effects 0.000 claims description 6
- 238000004590 computer program Methods 0.000 claims description 5
- 230000001939 inductive effect Effects 0.000 description 31
- 230000008569 process Effects 0.000 description 28
- 238000004891 communication Methods 0.000 description 17
- 238000012545 processing Methods 0.000 description 11
- 230000009466 transformation Effects 0.000 description 11
- 230000037213 diet Effects 0.000 description 8
- 235000005911 diet Nutrition 0.000 description 8
- 238000012546 transfer Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000013140 knowledge distillation Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 208000017667 Chronic Disease Diseases 0.000 description 4
- 206010003119 arrhythmia Diseases 0.000 description 4
- 230000006793 arrhythmia Effects 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 3
- 208000028399 Critical Illness Diseases 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000004821 distillation Methods 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/25—Bioelectric electrodes therefor
- A61B5/279—Bioelectric electrodes therefor specially adapted for particular uses
- A61B5/28—Bioelectric electrodes therefor specially adapted for particular uses for electrocardiography [ECG]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
Definitions
- the present disclosure relates to deep learning technology in the medical field, and specifically, to a method for training a deep learning model using inductive bias transfer and using the learned deep learning model in the medical field. .
- DeiT data efficient image transformer
- a convolutional neural network as a transformer's teacher, and adopts a method of distilling the knowledge of the convolutional neural network, which plays an important role in generalization performance, into the transformer through an additional token.
- DeiT assumes that performance can be improved by encoding the inductive bias of the convolutional neural network into a transformer through this distillation method.
- DeiT in order to carry out knowledge distillation, DeiT also requires that the number of data be secured above a certain level. Therefore, DeiT has a problem in that it is difficult to utilize in the medical field where the number of classes of data for knowledge distillation is not sufficiently accumulated.
- ImageNet previously used to evaluate the performance of DieT consists of a thousand classes
- medical data such as electrocardiogram data and electronic health records (EHR) data have tens to tens of classes. contains only Considering that successful knowledge transfer depends on the quality of the teacher's signal, where various signals are important, it should be clear that when using DieT, the inductive bias of the teacher's model successfully transfers on ImageNet, but not on medical data. can
- An object of the present invention is to provide a method for estimating various health conditions using a model.
- the learning method may include learning a first neural network model based on medical data; and by matching a first operation function representing a neural network block included in the learned first neural network model with a second operation function representing a neural network block included in the second neural network model, The method may include training the second neural network model based on the first neural network model.
- the learned first neural network model is obtained by matching a first operation function representing a neural network block included in the learned first neural network model with a second operation function representing a neural network block included in the second neural network model.
- the loss function may include: a first sub-loss function having as an input variable an output of a first operation function corresponding to an n-1th (n is a natural number) neural network block included in the first neural network model; and a second sub-loss function having an output of a second operation function corresponding to the n ⁇ 1 th neural network block included in the second neural network model as an input variable.
- each of the first sub-loss function and the second sub-loss function is an output of a first operation function corresponding to an n-th neural network block included in the first neural network model and an n-th included in the second neural network model. It may be a function for calculating a difference between outputs of the second calculation function corresponding to the neural network block.
- each of the first sub-loss function and the second sub-loss function may be configured to match dimensions of a first operation function corresponding to the n-th neural network block and a second operation function corresponding to the n-th neural network block.
- a conversion function may be included for
- the transform function may include: a first sub-transform function for linearly transforming an input variable of the transform function in a temporal direction; and a second sub-transform function for linearly transforming an input variable of the transform function in a feature dimension.
- the transformation function included in the first sub-loss function may convert the dimension of the output of the first operation function corresponding to the n ⁇ 1 th neural network block into an input of the second operation function corresponding to the n th neural network block. It may be a function for matching the dimension of and matching the dimension of the output of the second operation function corresponding to the n-th neural network block to the dimension of the output of the first operation function corresponding to the n-th neural network block.
- the transformation function included in the second sub-loss function may convert the dimension of the output of the second operation function corresponding to the n ⁇ 1 th neural network block to the input of the first operation function corresponding to the n th neural network block. It may be a function for matching the dimension of and matching the dimension of the output of the first operation function corresponding to the n-th neural network block to the dimension of the output of the second operation function corresponding to the n-th neural network block.
- the loss function further includes a third sub-loss function for calculating a difference between an output of the first neural network model receiving the medical data and an output of the second neural network model receiving the medical data. can do.
- the third sub-loss function may include a conversion function for matching a dimension of an output of the first neural network model receiving the medical data to a dimension of an output of the second neural network model receiving the medical data.
- the learning method may further include performing fine tuning on the second neural network model based on medical data.
- the fine adjustment may be to train the second neural network model while maintaining a weight of the second neural network model close to a weight in a state in which learning based on the learned first neural network model is completed.
- the first neural network model includes at least one of a convolutional neural network and a recurrent neural network
- the second neural network model includes a self-attention based neural network. can do.
- the medical data may include at least one of electrocardiogram data or electronic health records (EHR) data.
- EHR electronic health records
- the reasoning method may include obtaining medical data including at least one of electrocardiogram data and electronic health record data; and estimating a health state of a person based on the medical data by using a second neural network model.
- the second neural network model includes a first operation function corresponding to a neural network block included in the learned first neural network model and a neural network block included in the second neural network model, based on the pre-learned first neural network model. It may be learned through an operation matching the second operation function corresponding to .
- a computer program stored in a computer readable storage medium is disclosed according to an embodiment of the present disclosure for realizing the above object.
- the operations may include learning a first neural network model based on medical data; and by matching a first operation function representing a neural network block included in the learned first neural network model with a second operation function representing a neural network block included in the second neural network model, based on the learned first neural network model.
- An operation of learning the second neural network model may be included.
- a computing device for learning a deep learning model based on medical data includes a processor including at least one core; a memory containing program codes executable by the processor; and a network unit for obtaining medical data.
- the processor learns a first neural network model based on the medical data, and a first operation function representing a neural network block included in the learned first neural network model and a second operation function representing a neural network block included in the second neural network model
- the second neural network model may be trained based on the learned first neural network model.
- the present disclosure provides a learning method for a deep learning model capable of generalizing a student model based on a small amount of data by appropriately utilizing the strong inductive bias of the teacher model, and a deep learning model learned through this learning method. It can provide a method to effectively utilize the learning model in the medical field.
- FIG. 1 is a block diagram of a computing device according to one embodiment of the present disclosure.
- FIG. 2 is a block diagram illustrating a process of learning a deep learning model according to an embodiment of the present disclosure.
- FIG. 3 is a conceptual diagram illustrating a process of calculating a loss function according to an embodiment of the present disclosure.
- FIG. 4 is a graph showing results of comparing and evaluating the performance of a deep learning model trained according to an embodiment of the present disclosure and a conventional solution.
- FIG. 5 is a flowchart illustrating a method for learning a deep learning model according to an embodiment of the present disclosure.
- FIG. 6 is a flowchart illustrating an inference method of a deep learning model according to an embodiment of the present disclosure.
- N N is a natural number
- N a natural number
- components performing different functional roles in the present disclosure may be classified as first components or second components.
- components that are substantially the same within the technical spirit of the present disclosure but should be distinguished for convenience of description may also be classified as first components or second components.
- acquisition used in the present disclosure is understood to mean not only receiving data through a wired/wireless communication network with an external device or system, but also generating data in an on-device form. It can be.
- module refers to a computer-related entity, firmware, software or part thereof, hardware or part thereof , It can be understood as a term referring to an independent functional unit that processes computing resources, such as a combination of software and hardware.
- a “module” or “unit” may be a unit composed of a single element or a unit expressed as a combination or set of a plurality of elements.
- a “module” or “unit” is a hardware element or set thereof of a computing device, an application program that performs a specific function of software, a process implemented through software execution, or a program. It may refer to a set of instructions for execution.
- a “module” or “unit” may refer to a computing device constituting a system or an application executed in the computing device.
- the concept of “module” or “unit” may be defined in various ways within a range understandable by those skilled in the art based on the contents of the present disclosure.
- model used in this disclosure refers to a system implemented using mathematical concepts and language to solve a specific problem, a set of software units to solve a specific problem, or a process to solve a specific problem. It can be understood as an abstract model for a process.
- a neural network “model” may refer to an overall system implemented as a neural network having problem-solving capabilities through learning. At this time, the neural network may have problem solving ability by optimizing parameters connecting nodes or neurons through learning.
- a neural network "model” may include a single neural network or may include a neural network set in which a plurality of neural networks are combined.
- Data used in the present disclosure may include “image”, signals, and the like.
- image used in this disclosure may refer to multidimensional data composed of discrete image elements.
- image can be understood as a term referring to a digital representation of an object that is visible to the human eye.
- image may refer to multidimensional data composed of elements corresponding to pixels in a 2D image.
- Image may refer to multidimensional data composed of elements corresponding to voxels in a 3D image.
- inductive bias used in the present disclosure may be understood as a set of assumptions that enable inductive inference of a deep learning model. In order to solve the error of generalization in which the deep learning model shows appropriate performance only for the given training data, it is necessary for the deep learning model to infer data other than the given training data to get close to an accurate output. Therefore, “inductive bias” can be understood as a set of preconditions that a deep learning model has in the process of learning to predict the output of an input that is not given.
- block used in the present disclosure may be understood as a set of components classified based on various criteria such as type and function. Accordingly, a configuration classified as one “block” may be variously changed according to a criterion.
- a neural network “block” may be understood as a neural network set comprising at least one neural network. In this case, it may be assumed that the neural networks included in the neural network "block” perform the same specific operation.
- operation function used in the present disclosure may be understood as a mathematical expression for a unit that performs a specific function or processes an operation.
- an "operation function" of a neural network block may be understood as a mathematical expression representing a neural network block that processes a specific operation. Therefore, the relationship between the input and output of the neural network block can be expressed as a formula through the "operation function" of the neural network block.
- FIG. 1 is a block diagram of a computing device according to an embodiment of the present disclosure.
- the computing device 100 may be a hardware device or part of a hardware device that performs comprehensive processing and calculation of data, or may be a software-based computing environment connected through a communication network.
- 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 the server.
- the computing device 100 may be a cloud system in which a plurality of servers and clients interact to comprehensively process data. Since the above description is only one example related to the type of the computing device 100, the type of the computing device 100 may be configured in various ways within a range understandable by those skilled in the art based on the contents of the present disclosure.
- a computing device 100 may include a processor 110, a memory 120, and a network unit 130. there is.
- the computing device 100 may include other configurations for implementing a computing environment. Also, only some of the components disclosed above may be included in the computing device 100 .
- the processor 110 may be understood as a structural unit including hardware and/or software for performing computing operations.
- the processor 110 may read a computer program and perform data processing for machine learning.
- the processor 110 may process input data processing for machine learning, feature extraction for machine learning, calculation of an error based on backpropagation, and the like.
- the processor 110 for performing such data processing includes a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), a tensor processing unit (TPU), and on-demand It may include a semiconductor (application specific integrated circuit (ASIC)) or a field programmable gate array (FPGA). Since the above-described type of processor 110 is just one example, the type of processor 110 may be variously configured within a range understandable by those skilled in the art based on the content of the present disclosure.
- the processor 110 may train a neural network model for estimating a health state of a person based on knowledge distillation in which knowledge of a teacher model is transferred to a student model. For example, the processor 110 may train a teacher model to estimate a health condition of a person based on medical data. The processor 110 may train the student model based on the inductive bias generated in the process of learning the teacher model. In order to transfer the inductive bias of the teacher model to the student model during the learning process of the student model, the processor 110 includes each operation function representing a plurality of neural network blocks included in the teacher model and a plurality of neural network blocks included in the student model. Each operation function expressing can be matched.
- the matching of the operation function can be understood as an operation that allows the student model to learn the inductive bias of the teacher model by minimizing the difference between the characteristics interpreted by each neural network block of the teacher model and the characteristics interpreted by each neural network block of the student model.
- the processor 110 can effectively transfer the inductive bias of the teacher model to the student model even with a small amount of medical data. That is, through learning that matches calculation functions, the processor 110 can efficiently learn the student model based on the knowledge of the teacher model even in a state where a sufficient number of data is not secured.
- the processor 110 may ensure generalization performance of the student model.
- the processor 110 may estimate a person's health condition based on medical data using the neural network model generated through the above-described learning process.
- the processor 110 may generate inference data indicating a result of estimating a health condition of a person by inputting medical data to the neural network model learned through the above process.
- the processor 110 may input electrocardiogram data to a student model for which learning has been completed, and predict whether a chronic disease such as arrhythmia will occur or not, and the degree of progression of the chronic disease.
- the processor 110 may input electronic health records (EHR) data to the student model for which learning has been completed, and generate data related to information necessary for managing the patient, such as changes in the patient's heart rate.
- EHR electronic health records
- the type of medical data and the output of the neural network model may be configured in various ways within a range understandable by those skilled in the art based on the contents of the present disclosure.
- the memory 120 may be understood as a unit including hardware and/or software for storing and managing data processed by 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 .
- the memory 120 may include a flash memory type, a hard disk type, a multimedia card micro type, a card type memory, and random access memory (RAM). ), SRAM (static random access memory), ROM (read-only memory), EEPROM (electrically erasable programmable read-only memory), PROM (programmable read-only memory), magnetic memory , a magnetic disk, and an optical disk may include at least one type of storage medium.
- the memory 120 may include a database system that controls and manages data in a predetermined system. Since the above-described type of memory 120 is just one example, the type of memory 120 may be configured in various ways within a range understandable by those skilled in the art based on the contents of the present disclosure.
- the memory 120 may organize and manage data necessary for the processor 110 to perform calculations, data combinations, program codes executable by the processor 110, and the like.
- the memory 120 may store medical data received through the network unit 130 to be described later.
- the memory 120 includes program codes for operating the neural network model to perform learning by receiving medical data, program codes for operating the neural network model to perform inference according to the purpose of use of the computing device 100 by receiving medical data, and Processing data generated as the program code is executed may be stored.
- the network unit 130 may be understood as a unit that transmits and receives data through any type of known wired/wireless communication system.
- the network unit 130 may include a local area network (LAN), wideband code division multiple access (WCDMA), long term evolution (LTE), and WiBro (wireless).
- broadband internet 5th generation mobile communication (5G), ultra wide-band wireless communication, ZigBee, radio frequency (RF) communication, wireless LAN, wireless fidelity ), near field communication (NFC), or data transmission/reception may be performed using a wired/wireless communication system such as Bluetooth. Since the above-described communication systems are only examples, a wired/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.
- the network unit 130 may receive data necessary for the processor 110 to perform an operation through wired/wireless communication with an arbitrary system or an arbitrary client.
- the network unit 130 may transmit data generated through the operation of the processor 110 through wired/wireless communication with an arbitrary system or an arbitrary client.
- the network unit 130 may receive medical data through communication with a database in a hospital environment, a cloud server that performs tasks such as standardization of medical data, or a computing device.
- the network unit 130 may transmit output data of the neural network model, intermediate data derived from the calculation process of the processor 110, and processed data through communication with the aforementioned database, server, or computing device.
- FIG. 2 is a block diagram illustrating a process of learning a deep learning model according to an embodiment of the present disclosure.
- the processor 110 may perform a three-step learning process in order to learn a neural network model for estimating a health condition of a person based on medical data.
- the processor 110 may perform step A of training the first neural network model 200 to have an inductive bias for estimating a health condition of a person based on medical data.
- the processor 110 may perform step B to transfer the inductive bias of the first neural network model 200 reinforced through step A to the second neural network model 300 .
- the processor 110 may perform step C for fine-tuning the second neural network model 300 learned through step B according to the purpose of use.
- step C is not necessarily performed, and may or may not be performed depending on the purpose of using the second neural network model 300 .
- the processor 110 may train the first neural network model 200 to generate an inductive bias to pass to the second neural network model 300 .
- the processor 110 may input the medical data 11 to the first neural network model 200 and generate an output for estimating a health condition of a person.
- the processor 110 may compare the output of the first neural network model 200 with the ground truth (GT) through a loss function to reconstruct parameters such as weights of the neural network.
- the processor 110 may train the first neural network model 200 based on the medical data 11 by repeatedly performing this operation until the loss function converges.
- the first neural network model 200 includes a convolutional neural network
- the first neural network model 200 determines the locality and translation invariance of the medical data 11 through step A.
- the first neural network model 200 may reinforce the inductive bias of the temporal invariance of the medical data 11 through step A. there is.
- the processor 110 configures the plurality of neural network blocks included in the first neural network model 200 so that the second neural network model 300 can learn the inductive bias of the first neural network model 200 reinforced through step A. 210 and the plurality of neural network blocks 220 included in the second neural network model 300 may be matched.
- the processor 110 may respectively input the medical data 15 to the first neural network model 200 and the second neural network model 300 .
- the medical data 15 of step B may be the same as or different from the medical data 11 of step A.
- the processor 100 uses a loss function to output each of the plurality of neural network blocks 210 included in the first neural network model 200.
- the processor 110 sets the plurality of neural network blocks 210 included in the first neural network model 200 and the plurality of neural network blocks 220 included in the second neural network model 300 in pairs. It can be configured to calculate the loss.
- the input and/or dimension between the neural network blocks 210 and 220 to be matched difference may occur.
- the first neural network model 200 includes at least one of a convolutional neural network and a recurrent neural network
- the second neural network model 300 includes a self-attention based neural network
- the first neural network model Due to differences in input and dimensions between the neural network blocks 210 of 200 and the neural network blocks 220 of the second neural network model 300, it is inevitably difficult to calculate the loss.
- the processor 110 may perform an operation for matching inputs and/or dimensions between the neural network blocks 210 and 220 in a process of calculating a difference using a loss function.
- the processor 110 alternately inputs the intermediate output of the first neural network model 200 and the intermediate output of the second neural network model 300 to blocks 210 and 220, respectively, and blocks 210 and 220 respectively.
- the processor 110 can effectively match the neural network blocks 210 and 220 even if the types of neural networks included in each of the first neural network model 200 and the second neural network model 300 are different. there is.
- the processor 110 may calculate a difference between a final output of the first neural network model 200 and a final output of the second neural network model 300 using the loss function.
- the processor 110 repeatedly performs the above-described matching between neural network blocks and matching between final outputs until the loss function converges, thereby transferring the inductive bias of the first neural network model 200 to the second neural network model 300.
- the processor 110 may train the second neural network model 300 based on the first neural network model 200 learned through step A through neural network block matching in step B.
- the processor 110 may perform fine-tuning on the second neural network model 300 having learned the inductive bias of the first neural network model 200 .
- fine tuning can be understood as a learning process of optimizing the second neural network model 300 for a specific task according to the purpose of using the second neural network model 300 .
- the processor 110 may input the medical data 19 to the second neural network model 300 learned through step B to generate an output for estimating a person's health condition.
- the medical data 19 may be electrocardiogram data or electronic health record data depending on the purpose of using the second neural network model 300 .
- the second neural network model 300 may receive electrocardiogram data to estimate chronic diseases such as arrhythmias, or receive electronic health record data to estimate changes in a critically ill patient's condition.
- the processor 110 may compare the output of the second neural network model 300 with the ground truth (GT) through a loss function to reconstruct parameters such as weights of the neural network.
- the processor 110 may retrain the second neural network model 300 based on the medical data 19 by repeatedly performing this operation until the loss function converges.
- FIG. 3 is a conceptual diagram illustrating a process of calculating a loss function according to an embodiment of the present disclosure.
- the first neural network model 200 is a convolutional neural network model or a recurrent neural network model
- the second neural network model 300 is a self-attention based transformer model. do.
- the first neural network model 200 and the second neural network model 300 may be interpreted as a combination of operation functions representing neural network blocks as shown in [Equation 1] below.
- f is the first neural network model 200
- f n is the nth (n is a natural number) neural network block of the first neural network model 200
- g is the second neural network model 300
- g n is the second neural network model Indicates the n-th neural network block of (300).
- l f denotes a classifier of the first neural network model 200
- l g denotes a classifier of the second neural network model 300.
- the processor 110 uses f You can perform an operation that matches n and g n . Matching operations can be largely divided into input matching and dimension matching.
- a problem of input difference may naturally occur in function matching between neural network blocks.
- the input of the second function g 2 of the second neural network model 300 becomes g 1 (x)
- the input of the second function f 2 of the first neural network model 200 becomes f 1 (x). do.
- matching between the two functions is performed by an operation that minimizes ?g 2 (g 1 (x))-f 2 (f 1 (x)) ?.
- this matching operation cannot guarantee that g 2 approximates f 2 .
- the processor 110 alleviates the above-described problem by passing signals to both the computational flows of the first neural network model 200 and the second neural network model 300 . That is, in the present disclosure, the matching between the two functions is ?g 2 (f 1 (x)) - f 2 (f 1 (x)) - + ?g 2 (g 1 (x)) - f 2 (g 1 (x)) can be performed as an operation that minimizes.
- a dimension difference problem may occur due to a function through which an input passes.
- f n ( ⁇ ) representing the n-th neural network block of the first neural network model 200
- the pooling operation and the convolution with stride change the dimension according to the temporal direction, and the convolution operation change It also expands the dimension of features.
- processor 110 may use a transform function h( ⁇ ) that transforms each dimension equally.
- I( ⁇ ) represents a linear transformation in a temporal direction
- W represents a linear transformation in a feature dimension. That is, the transformation function h( ⁇ ) can be expressed as a product of a function for linear transformation of the time direction of an input variable and a function for linear transformation of the feature dimension of the input variable.
- L 1 n may correspond to FIG. 3 (a), and L 2 n may correspond to FIG. 3 (b).
- the total loss function L for learning the inductive bias of the first neural network model 200 which is a convolutional neural network model or a recurrent neural network model
- the second neural network model 300 which is a transformer model
- the weight ⁇ in a state in which learning is completed is an initial weight of learning using a randomly initialized classifier.
- the processor 110 may fine-tune the learned second neural network model 300 by minimizing loss through the same operation as that of the first neural network model 200 having an inductive bias. While fine-tuning the second neural network model 300, the processor 110 maintains the weight ⁇ g close to the initial weight ⁇ f ⁇ g so that the a posteriori bias transferred from the first neural network model 200 is transferred to the second neural network model ( The second neural network model 300 may be normalized so that 300) does not forget. This normalization can be expressed as the following [Equation 6].
- the inductive bias of the first neural network model 200 is efficiently transferred to the second neural network model 300 through the above-described three-step learning in FIG. 2 and the calculation of the loss function in FIG. 300) can be effectively optimized according to the purpose of use in the medical field.
- the second neural network model 300 can be trained to ensure generalization performance even with a small amount of data, and can be easily used for various tasks in the medical field.
- FIG. 4 is a graph showing results of comparing and evaluating the performance of a deep learning model trained according to an embodiment of the present disclosure and a conventional solution.
- FIG. 4(a) shows the performance of the neural network 49 of the present disclosure learned by receiving the inductive bias from the convolutional neural network 41, the DieT 45, and the convolutional neural network 41 based on electrocardiogram data. Shows the comparison result. And, F1 represents the f-1 score, and P-21 represents the physionet 21 score.
- electrocardiogram data it can be seen that the performance of the neural network 49 of the present disclosure is superior to the convolutional neural network 41 that delivered the inductive bias.
- the performance of the neural network 49 of the present disclosure is far superior to that of the DieT 45 based on knowledge distillation.
- electrocardiogram it can be seen that the performance of the DieT 45 is lower than that of the convolutional neural network 41.
- FIG. 4(b) shows the neural network 59 of the present disclosure learned by receiving the inductive bias from the convolutional neural network 51, the DieT 55, and the convolutional neural network 51 based on electronic medical record data. Shows the result of comparing performance. And, P-19 represents the physionet 19 score.
- the performance of the neural network 59 of the present disclosure is superior to the convolutional neural network 51 that delivered the inductive bias.
- the performance of the neural network 49 of the present disclosure is far superior to that of the DieT 45 based on knowledge distillation based on P-19.
- the neural network trained through the learning method of the present disclosure is more suitable for the medical field where sufficient data for learning is not accumulated compared to DieT.
- the neural network trained through the learning method of the present disclosure successfully encodes the inductive bias of the preceding model and shows better performance than the preceding model.
- FIG. 5 is a flowchart illustrating a method for learning a deep learning model according to an embodiment of the present disclosure.
- the computing device 100 may learn a first neural network model based on medical data (S110).
- the computing device 100 may train the first neural network model so that the first neural network model estimates a health condition of a person based on medical data.
- the computing device 100 may receive at least one of electrocardiogram data or electronic health record data as learning data through communication with a database in a hospital environment. there is.
- the computing device 100 may generate at least one of electrocardiogram data or electronic health record data as learning data through communication with an electrocardiogram measuring device in the hospital environment.
- the computing device 100 learns the first neural network model to estimate the occurrence, change trend, risk level, etc. of a person's disease based on at least one of the electrocardiogram data and the electronic health record data obtained by the first neural network model. can make it
- the computing device 100 may train the second neural network model by matching an operation function representing a neural network block between the first neural network model and the second neural network model (S120).
- the computing device 100 may match a first operation function representing a neural network block included in the first neural network model learned through step S110 with a second operation function representing a neural network block included in the second neural network model.
- the computing device 100 uses a loss function for matching at least one of inputs or dimensions between the first operation function and the second operation function, and based on the first neural network model learned through step S110, the second A neural network model can be trained.
- the loss function for matching the first operation function and the second operation function in step S120 is the first operation function corresponding to the n-1th neural network block included in the first neural network model. It may include a first sub loss function having an output as an input variable and a second sub loss function having an output of a second operation function corresponding to the n ⁇ 1 th neural network block included in the second neural network model as an input variable.
- Each of the first sub-loss function and the second sub-loss function is an output of the first operation function corresponding to the n-th neural network block included in the first neural network model and a second sub-loss function corresponding to the n-th neural network block included in the second neural network model.
- the first sub-loss function may be understood as a function for calculating a difference between outputs generated by inputting an output of a previous first calculation function to a current first calculation function and a current second calculation function, respectively.
- the second sub-loss function may be understood as a function for calculating a difference between outputs generated by inputting an output of a previous second operation function to the current first operation function and the current second operation function, respectively.
- each of the first sub-loss function and the second sub-loss function may include a conversion function for matching the dimensions of the first operation function corresponding to the n-th neural network block and the second operation function corresponding to the n-th neural network block.
- the transform function may include a first sub-transform function for linearly transforming an input variable of the transform function in a temporal direction and a second sub-transform function for linearly transforming an input variable of the transform function in a feature dimension.
- the transform function corresponds to h( ⁇ ) in [Equation 2]
- the first sub-transform function corresponds to I( ⁇ ) in [Equation 2]
- the second sub-transform function corresponds to W in [Equation 2]
- the conversion function included in the first sub-loss function matches the dimension of the output of the first operation function corresponding to the n ⁇ 1 th neural network block to the dimension of the input of the second operation function corresponding to the n th neural network block, n It may be a function for matching the dimension of the output of the second operation function corresponding to the n-th neural network block to the dimension of the output of the first operation function corresponding to the n-th neural network block.
- the conversion function included in the first sub-loss function is the first operation function corresponding to the n-1 th neural network block, such as h n-1 f->g.
- the conversion function included in the first sub-loss function converts the dimension of the output of the second operation function corresponding to the n-th neural network block to that of the first operation function corresponding to the n-th neural network block, such as h n g->f . It can be used for operations that transform according to the dimension of the output.
- the transformation function included in the second sub-loss function matches the dimension of the output of the second operation function corresponding to the n ⁇ 1 th neural network block to the dimension of the input of the first operation function corresponding to the n th neural network block, and It may be a function for matching the dimension of the output of the first operation function corresponding to the n-th neural network block to the dimension of the output of the second operation function corresponding to the n-th neural network block.
- the transformation function included in the second sub-loss function is the second operation function corresponding to the n-1th neural network block, such as h n-1 g->f.
- the transformation function included in the second sub-loss function converts the dimension of the output of the first operation function corresponding to the n-th neural network block to that of the second operation function corresponding to the n-th neural network block, such as h n f->g. It can be used for operations that transform according to the dimension of the output.
- the first sub-loss function corresponds to L 1 n in [Equation 3]
- the second sub-loss function corresponds to L 2 n in [Equation 3]. That is, it can be understood that the calculation process of the first sub-loss function corresponds to FIG. 3(a) and the calculation process of the second sub-loss function corresponds to FIG. 3(b).
- the loss function for matching the first operation function and the second operation function in step S120 calculates the difference between the output of the first neural network model receiving the medical data and the output of the second neural network model receiving the medical data.
- the third sub-loss function may include a conversion function for matching a dimension of an output of the first neural network model receiving the medical data to a dimension of an output of the second neural network model receiving the medical data. Similar to the first sub-loss function and the second sub-loss function, if there is a difference in the type of neural network between the first and second neural network models, there will inevitably be a dimensional difference between the data derived from each operation process. To solve the difference problem, the third sub-loss function may include a conversion function. For example, referring to (c) of FIG. 3, the transformation function included in the third sub-loss function converts the dimension of the output of the first operation function corresponding to the final neural network block to the final, such as h n f -> g. It can be used for an operation of converting according to the dimension of the output of the second operation function corresponding to the neural network block.
- the third sub-loss function corresponds to L 3 n in [Equation 4], and the calculation process of the third sub-loss function corresponds to FIG. 3(c).
- the loss function for matching the first operation function and the second operation function may be expressed as a combination of the first sub loss function, the second sub loss function, and the third loss function.
- Equation 5 it is expressed as a simple sum of three sub-loss functions, but the sub-loss functions can be combined based on various operations such as weighted sum and multiplication as well as simple sum.
- the computing device 100 may perform fine-tuning on the second neural network model, which has learned the inductive bias of the first neural network model, through step S120 based on the medical data (S130).
- the fine adjustment may be understood as a process of learning the second neural network model while keeping the weight of the second neural network model close to the weight of the state in which learning based on the first neural network model (step S120) is completed.
- the computing device 100 inputs electrocardiogram data related to predicting arrhythmia to the second neural network model in which step S120 has been completed, thereby generating a second neural network model. 2 You can fine-tune the neural network model.
- the computing device 100 adds the electronic health record of the critically ill patient to the second neural network model in which step S120 has been completed. You can fine-tune the second neural network model by inputting data.
- the fine adjustment of the present disclosure is not limited to the above-described examples, and may be variously performed within a range understandable by those skilled in the art based on the contents of the present disclosure.
- FIG. 6 is a flowchart illustrating an inference method of a deep learning model according to an embodiment of the present disclosure.
- the computing device 100 may obtain medical data including at least one of electrocardiogram data and electronic health record data (S210).
- the computing device 100 may receive at least one of electrocardiogram data or electronic health record data of an inference target through communication with a database in a hospital environment. there is.
- the computing device 100 is a database in a hospital environment
- the computing device 100 may generate at least one of electrocardiogram data or electronic health record data of an inference target through communication with an electrocardiogram measuring device in the hospital environment.
- the computing device 100 may estimate a health condition of a person based on the medical data by using the pretrained second neural network model (S220).
- the computing device 100 inputs at least one of electrocardiogram data or electronic health record data acquired through step S210 into the pre-learned second neural network model to predict whether a person's disease will occur, change trend, risk level, etc.
- the second neural network model includes a first operation function corresponding to a neural network block included in the first neural network model and a second operation function corresponding to a neural network block included in the second neural network model, based on the pretrained first neural network model. It may be learned through an operation matching an operation function. Since the pre-learning of the second neural network model corresponds to the description of FIG. 6 described above, a detailed description thereof will be omitted below.
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Pathology (AREA)
- Epidemiology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Primary Health Care (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Animal Behavior & Ethology (AREA)
- Veterinary Medicine (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Cardiology (AREA)
- Surgery (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Heart & Thoracic Surgery (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
Description
Claims (15)
- 적어도 하나의 프로세서(processor)를 포함하는 컴퓨팅 장치에 의해 수행되는, 의료 데이터를 기반으로 하는 딥러닝 모델의 학습 방법으로서,의료 데이터를 기초로 제 1 신경망 모델을 학습시키는 단계; 및상기 학습된 제 1 신경망 모델에 포함된 신경망 블록(block)을 나타내는 제 1 연산 함수(operation function)와 제 2 신경망 모델에 포함된 신경망 블록을 나타내는 제 2 연산 함수를 매칭시킴으로써, 상기 학습된 제 1 신경망 모델을 기초로 상기 제 2 신경망 모델을 학습시키는 단계;를 포함하는,방법.
- 제 1 항에 있어서,상기 학습된 제 1 신경망 모델에 포함된 신경망 블록을 나타내는 제 1 연산 함수와 제 2 신경망 모델에 포함된 신경망 블록을 나타내는 제 2 연산 함수를 매칭시킴으로써, 상기 학습된 제 1 신경망 모델을 기초로 상기 제 2 신경망 모델을 학습시키는 단계는,상기 제 1 연산 함수와 상기 제 2 연산 함수 간의 입력 혹은 차원 중 적어도 하나를 매칭시키기 위한 손실 함수를 이용하여, 상기 학습된 제 1 신경망 모델을 기초로 상기 제 2 신경망 모델을 학습시키는 단계;를 포함하는,방법.
- 제 2 항에 있어서,상기 손실 함수는,상기 제 1 신경망 모델에 포함된 n-1번째(n은 자연수) 신경망 블록에 대응되는 제 1 연산 함수의 출력을 입력 변수로 하는 제 1 서브 손실 함수; 및상기 제 2 신경망 모델에 포함된 n-1번째 신경망 블록에 대응되는 제 2 연산 함수의 출력을 입력 변수로 하는 제 2 서브 손실 함수;를 포함하되,상기 제 1 서브 손실 함수 및 상기 제 2 서브 손실 함수 각각은,상기 제 1 신경망 모델에 포함된 n번째 신경망 블록에 대응되는 제 1 연산 함수의 출력과 상기 제 2 신경망 모델에 포함된 n번째 신경망 블록에 대응되는 제 2 연산 함수의 출력 간 차이를 연산하기 위한 함수인,방법.
- 제 3 항에 있어서,상기 제 1 서브 손실 함수 및 상기 제 2 서브 손실 함수 각각은,상기 n번째 신경망 블록에 대응되는 제 1 연산 함수와 상기 n번째 신경망 블록에 대응되는 제 2 연산 함수의 차원을 매칭시키기 위한 변환 함수를 포함하는,방법.
- 제 4 항에 있어서,상기 변환 함수는,상기 변환 함수의 입력 변수를 시간적 방향(temporal direction)에서 선형 변환하기 위한 제 1 서브 변환 함수; 및상기 변환 함수의 입력 변수를 특징 차원(feature dimension)에서 선형 변환하기 위한 제 2 서브 변환 함수;를 포함하는,방법.
- 제 4 항에 있어서,상기 제 1 서브 손실 함수에 포함된 변환 함수는,상기 n-1번째 신경망 블록에 대응되는 제 1 연산 함수의 출력의 차원을 상기 n번째 신경망 블록에 대응되는 제 2 연산 함수의 입력의 차원에 매칭시키고,상기 n번째 신경망 블록에 대응되는 제 2 연산 함수의 출력의 차원을 상기 n번째 신경망 블록에 대응되는 제 1 연산 함수의 출력의 차원에 매칭시키기 위한 함수인,방법.
- 제 4 항에 있어서,상기 제 2 서브 손실 함수에 포함된 변환 함수는,상기 n-1번째 신경망 블록에 대응되는 제 2 연산 함수의 출력의 차원을 상기 n번째 신경망 블록에 대응되는 제 1 연산 함수의 입력의 차원에 매칭시키고,상기 n번째 신경망 블록에 대응되는 제 1 연산 함수의 출력의 차원을 상기 n번째 신경망 블록에 대응되는 제 2 연산 함수의 출력의 차원에 매칭시키기 위한 함수인,방법.
- 제 2 항에 있어서,상기 손실 함수는,상기 의료 데이터를 입력 받은 상기 제 1 신경망 모델의 출력과 상기 의료 데이터를 입력 받은 상기 제 2 신경망 모델의 출력 간의 차이를 연산하기 위한 제 3 서브 손실 함수;를 더 포함하는,방법.
- 제 8 항에 있어서,상기 제 3 서브 손실 함수는,상기 의료 데이터를 입력 받은 상기 제 1 신경망 모델의 출력의 차원을 상기 의료 데이터를 입력 받은 상기 제 2 신경망 모델의 출력의 차원에 매칭시키기 위한 변환 함수를 포함하는,방법.
- 제 1 항에 있어서,의료 데이터를 기초로 상기 제 2 신경망 모델에 대한 미세 조정(fine tuning)을 수행하는 단계;를 더 포함하고,상기 미세 조정은,상기 제 2 신경망 모델의 가중치가 상기 학습된 제 1 신경망 모델에 기반한 학습이 완료된 상태의 가중치와 가깝게 유지하면서 상기 제 2 신경망 모델을 학습시키는 것인,방법.
- 제 1 항에 있어서,상기 제 1 신경망 모델은,컨볼루셔널(convolutional) 신경망 혹은 순환(recurrent) 신경망 중 적어도 하나를 포함하고,상기 제 2 신경망 모델은,셀프-어텐션(self-attention) 기반 신경망을 포함하는,방법.
- 제 1 항에 있어서,상기 의료 데이터는,심전도 데이터 혹은 전자 건강 기록(EHR: electronic health records) 데이터 중 적어도 하나를 포함하는,방법.
- 적어도 하나의 프로세서를 포함하는 컴퓨팅 장치에 의해 수행되는, 의료 데이터를 기반으로 하는 딥러닝 모델의 추론 방법으로서,심전도 데이터 혹은 전자 건강 기록 데이터(EHR: electronic health records) 중 적어도 하나를 포함하는 의료 데이터를 획득하는 단계; 및제 2 신경망 모델을 사용하여, 상기 의료 데이터를 기초로 사람의 건강 상태를 추정하는 단계;를 포함하되,상기 제 2 신경망 모델은,사전 학습된 제 1 신경망 모델을 기초로, 상기 학습된 제 1 신경망 모델에 포함된 신경망 블록에 대응되는 제 1 연산 함수와 상기 제 2 신경망 모델에 포함된 신경망 블록에 대응되는 제 2 연산 함수를 매칭시키는 연산을 통해 학습된 것인,방법.
- 컴퓨터 판독가능 저장 매체 저장된 컴퓨터 프로그램(program)으로서, 상기 컴퓨터 프로그램은 하나 이상의 프로세서(processor)에서 실행되는 경우, 의료 데이터를 기반으로 하는 딥러닝 모델의 학습을 위한 동작들을 수행하도록 하며,상기 동작들은,의료 데이터를 기초로 제 1 신경망 모델을 학습시키는 동작; 및상기 학습된 제 1 신경망 모델에 포함된 신경망 블록(block)을 나타내는 제 1 연산 함수(operation function)와 제 2 신경망 모델에 포함된 신경망 블록을 나타내는 제 2 연산 함수를 매칭시킴으로써, 상기 학습된 제 1 신경망 모델을 기초로 상기 제 2 신경망 모델을 학습시키는 동작;을 포함하는,컴퓨터 프로그램.
- 의료 데이터를 기반으로 하는 딥러닝 모델의 학습을 위한 컴퓨팅 장치로서,적어도 하나의 코어(core)를 포함하는 프로세서(processor);상기 프로세서에서 실행 가능한 프로그램 코드(code)들을 포함하는 메모리(memory); 및의료 데이터를 획득하기 위한 네트워크부(network unit);를 포함하고,상기 프로세서는,의료 데이터를 기초로 제 1 신경망 모델을 학습시키고,상기 학습된 제 1 신경망 모델에 포함된 신경망 블록(block)을 나타내는 제 1 연산 함수(operation function)와 제 2 신경망 모델에 포함된 신경망 블록을 나타내는 제 2 연산 함수를 매칭시킴으로써, 상기 학습된 제 1 신경망 모델을 기초로 상기 제 2 신경망 모델을 학습시키는,장치.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP22873142.8A EP4407633A1 (en) | 2021-09-25 | 2022-09-15 | Method, program, and apparatus for training and inferring deep learning model on basis of medical data |
CN202280063391.4A CN118020112A (zh) | 2021-09-25 | 2022-09-15 | 以医疗数据为基础的深度学习模型的学习及推论方法、程序及装置 |
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR20210126784 | 2021-09-25 | ||
KR10-2021-0126784 | 2021-09-25 | ||
KR1020220113975A KR20230044123A (ko) | 2021-09-25 | 2022-09-08 | 의료 데이터를 기반으로 하는 딥러닝 모델의 학습 및 추론 방법, 프로그램 및 장치 |
KR10-2022-0113975 | 2022-09-08 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023048437A1 true WO2023048437A1 (ko) | 2023-03-30 |
Family
ID=85720885
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/KR2022/013833 WO2023048437A1 (ko) | 2021-09-25 | 2022-09-15 | 의료 데이터를 기반으로 하는 딥러닝 모델의 학습 및 추론 방법, 프로그램 및 장치 |
Country Status (2)
Country | Link |
---|---|
EP (1) | EP4407633A1 (ko) |
WO (1) | WO2023048437A1 (ko) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117253611A (zh) * | 2023-09-25 | 2023-12-19 | 四川大学 | 基于多模态知识蒸馏的癌症智能早期筛查方法及系统 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20200045128A (ko) * | 2018-10-22 | 2020-05-04 | 삼성전자주식회사 | 모델 학습 방법 및 장치, 및 데이터 인식 방법 |
KR20200052453A (ko) * | 2018-10-31 | 2020-05-15 | 삼성에스디에스 주식회사 | 딥러닝 모델 학습 장치 및 방법 |
KR102232138B1 (ko) * | 2020-11-17 | 2021-03-25 | (주)에이아이매틱스 | 지식 증류 기반 신경망 아키텍처 탐색 방법 |
KR20210035381A (ko) * | 2019-09-23 | 2021-04-01 | 삼성에스디에스 주식회사 | 의료 진단 방법 및 장치 |
KR20210068713A (ko) * | 2019-12-02 | 2021-06-10 | 주식회사 피디젠 | 딥러닝 기반 다중의료데이터를 통한 질병의 진행 예측 분석 시스템 |
-
2022
- 2022-09-15 EP EP22873142.8A patent/EP4407633A1/en active Pending
- 2022-09-15 WO PCT/KR2022/013833 patent/WO2023048437A1/ko active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20200045128A (ko) * | 2018-10-22 | 2020-05-04 | 삼성전자주식회사 | 모델 학습 방법 및 장치, 및 데이터 인식 방법 |
KR20200052453A (ko) * | 2018-10-31 | 2020-05-15 | 삼성에스디에스 주식회사 | 딥러닝 모델 학습 장치 및 방법 |
KR20210035381A (ko) * | 2019-09-23 | 2021-04-01 | 삼성에스디에스 주식회사 | 의료 진단 방법 및 장치 |
KR20210068713A (ko) * | 2019-12-02 | 2021-06-10 | 주식회사 피디젠 | 딥러닝 기반 다중의료데이터를 통한 질병의 진행 예측 분석 시스템 |
KR102232138B1 (ko) * | 2020-11-17 | 2021-03-25 | (주)에이아이매틱스 | 지식 증류 기반 신경망 아키텍처 탐색 방법 |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117253611A (zh) * | 2023-09-25 | 2023-12-19 | 四川大学 | 基于多模态知识蒸馏的癌症智能早期筛查方法及系统 |
CN117253611B (zh) * | 2023-09-25 | 2024-04-30 | 四川大学 | 基于多模态知识蒸馏的癌症智能早期筛查方法及系统 |
Also Published As
Publication number | Publication date |
---|---|
EP4407633A1 (en) | 2024-07-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Sivapalan et al. | ANNet: A lightweight neural network for ECG anomaly detection in IoT edge sensors | |
US20200027556A1 (en) | Systems and Methods for Medical Topic Discovery Based on Large-Scale Machine Learning | |
JP2023060820A (ja) | 機械学習モデル・スケーリングのためのディープ・ニューラル・ネットワーク最適化システム | |
Cai et al. | Hierarchical domain adaptation projective dictionary pair learning model for EEG classification in IoMT systems | |
WO2022164230A1 (ko) | 심전도 신호 기반의 만성질환 예측 방법 | |
WO2023048437A1 (ko) | 의료 데이터를 기반으로 하는 딥러닝 모델의 학습 및 추론 방법, 프로그램 및 장치 | |
Jiang et al. | Distributed deep learning optimized system over the cloud and smart phone devices | |
Xie et al. | Fast and peer-to-peer vital signal learning system for cloud-based healthcare | |
CN111160443B (zh) | 一种基于深度多任务学习的活动和用户识别方法 | |
WO2023249402A1 (ko) | 딥러닝 모델의 학습을 위한 의료 데이터의 처리 방법, 프로그램 및 장치 | |
CN114358250A (zh) | 数据处理方法、装置、计算机设备、介质及程序产品 | |
Agarwal et al. | Edge optimized and personalized lifelogging framework using ensembled metaheuristic algorithms | |
WO2023229384A1 (ko) | 학습 데이터 생성 방법, 컴퓨터 프로그램 및 장치 | |
Zeghlache et al. | Longitudinal self-supervised learning using neural ordinary differential equation | |
WO2024039043A1 (ko) | 학습 데이터 생성 방법, 컴퓨터 프로그램 및 장치 | |
WO2023177151A1 (ko) | 3차원 의료 데이터의 처리 방법, 프로그램 및 장치 | |
WO2022092670A1 (ko) | 뇌 피질 영역의 두께를 분석하는 방법 | |
WO2024014821A1 (ko) | 심전도 기반 신경망 모델의 학습 방법, 프로그램 및 장치 | |
Khan et al. | Data heterogeneity mitigation in healthcare robotic systems leveraging the Nelder–Mead method | |
KR20230044123A (ko) | 의료 데이터를 기반으로 하는 딥러닝 모델의 학습 및 추론 방법, 프로그램 및 장치 | |
WO2023249372A1 (ko) | 딥러닝 모델의 학습을 위한 의료 데이터의 처리 방법, 프로그램 및 장치 | |
WO2023249411A1 (ko) | 데이터 처리 방법, 컴퓨터 프로그램 및 장치 | |
WO2024039218A1 (ko) | 생체신호 데이터의 증강 방법, 컴퓨터 프로그램 및 장치 | |
WO2024053828A1 (ko) | 인공 신경망 기반 해상도 향상 방법, 프로그램 및 장치 | |
WO2024019523A1 (ko) | 심전도를 이용한 건강 상태의 예측 방법, 프로그램 및 장치 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22873142 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202447022110 Country of ref document: IN |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022873142 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2022873142 Country of ref document: EP Effective date: 20240425 |