CN111582434B - deep learning system - Google Patents
deep learning system Download PDFInfo
- Publication number
- CN111582434B CN111582434B CN201910404061.7A CN201910404061A CN111582434B CN 111582434 B CN111582434 B CN 111582434B CN 201910404061 A CN201910404061 A CN 201910404061A CN 111582434 B CN111582434 B CN 111582434B
- Authority
- CN
- China
- Prior art keywords
- data
- learning
- deep learning
- neural network
- artificial neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
-
- 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
- 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
- G16H50/70—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
Abstract
A deep learning system is disclosed that has access to an external deep learning system for learning a first artificial neural network model through a communications network. The deep learning system includes: a data classification storage module configured to receive and store source learning data; a learning module configured to learn a second artificial neural network model in the system using the source learning data; an adjustment information transmission module configured to transmit adjustment information including learning-related information and parameter information of the second artificial neural network model to an external deep learning system to request adjustment of the first artificial neural network model, wherein the learning-related information and the parameter information are modified according to a result of the learning; and an adjustment module configured to update parameter information of the second artificial neural network model according to adjustment information transmitted from an external deep learning system.
Description
Technical Field
The present invention relates to an artificial intelligence learning system, and more particularly, to a learning system based on deep learning.
Background
Deep learning algorithms have recently received attention as a machine learning system, which is also an artificial neural network, and whose number of hidden layers is much larger than that of the conventional neural network.
In general, an artificial neural network is composed of three layers in total, i.e., an input layer for receiving an input, a hidden layer for performing actual learning, and an output layer for returning a calculation result, as shown in fig. 1. In an artificial neural network having such a structure, each node means one neuron. Each node finds a weighted sum when data is entered, applies an activation function, and transmits the calculation result to the next node. In this case, a weight is assigned to each node by learning. The result of the previous node (neuron) is weighted and input to the current node (neuron), and a bias is added to the weighted result.
In general, a deep learning process is performed by repeatedly performing a forward propagation process for calculating a deep learning parameter and an objective function through a plurality of hidden layers in order from an input layer to an output layer, and a backward propagation process for modifying weights through the hidden layers in order from the output layer to the input layer in consideration of errors. The weights that are modified during the deep learning process are repeatedly updated until errors are minimized and all computers share the weights and the deep learning parameters.
In other words, since a learning process using learning data is performed in a deep learning algorithm to generate an artificial neural network model having optimized parameters, as the number of pieces of learning data increases, an artificial neural network model capable of more accurate prediction is formed. Thus, the deep learning process requires a large number of pieces of learning data.
However, in the case of a medical diagnostic imaging apparatus for determining whether a disease such as cervical cancer has occurred using a deep learning algorithm, there is a limit in ensuring learning data due to a limit in disclosure of personal information or medical information, and it is difficult to form an artificial neural network model that can be accurately predicted. Particularly in the case of rare diseases, the number of patients is small, it is difficult to share learning data between korea and foreign medical institutions, and a long time is required to accumulate the learning data. Therefore, it is necessary to find a new method for effectively solving these problems.
Further, in order to make a more accurate diagnosis in a medical diagnostic imaging apparatus, it is necessary to accurately classify learning data to be used for learning in addition to acquiring source learning data. That is, unless classification of learning data is accurately and clearly performed, it is inevitable that the accuracy of analyzing the occurrence of a specific disease is lowered. Therefore, for the apparatus for determining whether or not a specific disease has occurred, it is necessary to share the learning data classification standard and share the learning data to be used for learning.
(prior art documents)
(patent document)
Korean patent laid-open No. 10-2017-0083419
Disclosure of Invention
Technical problem
Accordingly, the present invention has been devised in view of the above-mentioned necessity, and a main object of the present invention is to provide a deep learning system capable of changing an artificial neural network model in an environment where only a small amount of source data can be learned, so that it becomes a model capable of performing accurate prediction as an artificial neural network model where a large amount of source data is learned.
Further, another object of the present invention is to provide a deep learning system capable of reducing learning time and overcoming an environment of shortage of learning data by acquiring results of learning of other artificial neural network models and reflecting the results in its own artificial neural network model.
Further, it is another object of the present invention to provide a deep learning system that can obtain the same learning result as that obtained by sharing source learning data without sharing the source learning data between artificial neural network models operated by different operators.
Further, it is another object of the present invention to provide a deep learning system capable of accurately determining whether a specific disease has occurred by sharing learning data classification criteria while acquiring learning results obtained by other artificial neural network models and reflecting them in their own artificial neural network models.
Technical proposal
According to an embodiment of the present invention, in order to achieve the above object, there is disclosed a deep learning system accessible through a communication network to an external deep learning system for learning a first artificial neural network model, the deep learning system including: a data classification storage module configured to receive and store source learning data; a learning module configured to learn a second artificial neural network model in the system using the source learning data; an adjustment information transmission module configured to transmit adjustment information including learning-related information and parameter information of the second artificial neural network model to an external deep learning system to request adjustment of the first artificial neural network model, wherein the learning-related information and the parameter information are modified according to a result of the learning; and an adjustment module configured to update parameter information of the second artificial neural network model according to adjustment information transmitted from an external deep learning system.
The deep learning system may also include a diagnostic module configured to diagnose whether a particular disease has occurred based on the first artificial neural network model optimized by repeated operation of the learning module and the adjustment module.
The data classification storage module may include: a data receiving unit configured to receive image data of an inspection target from an external device as source learning data and store the received image data; and a data classifying unit configured to classify and store the stored image data of the inspection target according to a plurality of multi-stage classification criteria, and the data classifying unit classifies and stores the image data of the inspection target using at least two or more of a first-stage classification criterion (which is a classification criterion of a color), a second-stage classification criterion (which is a classification criterion of a size of the inspection target), and a third-stage classification criterion (which is a classification criterion of a combination of a color and a form).
The adjustment information transmission module may preferentially transmit, as the adjustment information, the hyper-parameters required to perform synchronization on the initial learning state of the first artificial neural network model.
In some cases, the adjustment information transmitted by the adjustment information transmission module or the external deep learning system may constitute an artificial neural network model and include weight values and bias values of each layer modified according to the learning result.
The learning-related information may include the number of learned source learning data pieces in the source learning data to be learned.
The data classification unit also generates additional source learning data to adjust a balance of the number of source learning data classified and stored based on the classification criteria.
Advantageous effects
According to the above-described aspects, since the parameters of the artificial neural network model obtained by learning by the first deep learning system are transmitted to the second deep learning system and used to update the parameters of the artificial neural network model of the second deep learning system, the same effects as those obtained by learning the same learning data can be obtained without sharing the source learning data. Accordingly, the present invention can overcome problems caused by exposure of personal information or medical information in advance, and can also shorten the time required to optimize an artificial neural network model due to restrictions on sharing personal information or medical information.
Furthermore, the present invention is advantageous in that since the artificial neural network model is partially self-optimized using parameters obtained from learning results generated from an external system, it is practically impossible to recover or infer source learning data, thereby minimizing damage caused by information exposure.
In particular, the present invention can achieve the same effect as that obtained by learning the same learning data without sharing the source learning data, and therefore, even if the artificial neural network model is in an environment where only a small amount of source data can be learned, the artificial neural network model can be changed to an artificial neural network model where a large amount of source learning data is learned.
Also, according to still another embodiment, the present invention is advantageous in that since the learning data classification standard and the parameters modified according to the learning result are shared between different deep learning systems while accurately classifying the learning data for learning in advance and learning on the basis of the classification standard, the diagnosis standard can be shared to accurately diagnose whether or not a specific disease has occurred.
Drawings
Fig. 1 is a diagram showing a configuration of a general artificial neural network.
Fig. 2 is a diagram showing connections between deep learning systems according to an embodiment of the present invention.
Fig. 3 is a diagram showing a configuration of a deep learning system according to an embodiment of the present invention.
Fig. 4 is a diagram for further describing a deep learning process according to an embodiment of the present invention.
Fig. 5 is a diagram showing a multi-stage classification flow of image data of an inspection target according to another embodiment of the present invention.
Fig. 6 is a diagram for describing in detail a multi-level classification criterion of image data of an inspection target according to still another embodiment of the present invention.
Detailed Description
Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, if it is determined that related well-known functions or elements may unnecessarily obscure the gist of the present invention, a detailed description thereof will be omitted.
For reference, the term "module" used herein is defined as a combination of functions and structures of hardware for realizing the technical spirit of the present invention and software required to drive the hardware. That is, a "module" refers to a predetermined code and a logic unit of hardware resources for executing the code, and it is apparent to those skilled in the art that it does not necessarily mean physically connected code or one kind of hardware.
Hereinafter, a deep learning system according to an embodiment of the present invention will be described with reference to the drawings.
First, fig. 2 shows connections between the deep learning systems 100, 200, and 300 according to an embodiment of the present invention.
As shown in fig. 2, the deep learning systems 100, 200, and 300 according to the embodiments of the present invention transmit and receive required information to and from each other through a communication network. It is assumed that each of the deep learning systems 100, 200, and 300 may include the same internal artificial neural network model.
Each of the deep learning systems 100, 200, and 300 transmits and receives information to and from a deep learning system designated by an administrator, and updates its own parameters using parameters obtained by learning by other deep learning systems to optimize an internal artificial neural network model.
The configuration of the deep learning systems 100, 200, and 300 shown in fig. 2 will be described below with reference to fig. 3.
Fig. 3 shows a block diagram of a deep learning system 100 according to an embodiment of the invention. As an example, the deep learning system 100 including the Convolutional Neural Network (CNN) based artificial neural network model 150 includes: a data classification storage module 110 configured to receive source learning data and store in memory; a learning module 120 configured to train an artificial neural network model 150 in the system using the source learning data; an adjustment information transmission module 130 configured to transmit adjustment information including learning-related information and parameter information of the artificial neural network model 150 modified using the learning result to the external deep learning system 200 or 300 to request the artificial neural network model in the adjustment system; and an adjustment module 140 configured to update parameter information of the artificial neural network model 150 according to the adjustment information transmitted from the external deep learning system 200 or 300.
According to an implementation, in addition to the elements described above, the deep learning system 100 according to an embodiment of the present invention may further include a diagnosis module 160 configured to diagnose whether a specific disease has occurred based on the artificial neural network model 150 optimized through the repeated operations of the learning module 120 and the adjustment module 140.
The source learning data is data transmitted through a separate storage device or a separate data providing server in which the source learning data is stored, and through an image capturing device connectable to the deep learning system 100. When the deep learning system 100 according to an embodiment of the present invention is used as a part of a medical diagnostic imaging apparatus, the source learning data may be image data of an object to be examined (hereinafter referred to as an examination target) captured for diagnosing a specific disease.
Meanwhile, the adjustment information transmission module 130 preferentially transmits, as adjustment information, super parameters required to perform synchronization with respect to the initial learning state of the artificial neural network model included in the external deep learning system 200 or 300, and then transmits, as adjustment information, weight values W1, W2, W3. of each layer modified according to the learning result and the bias value. The adjustment information transmission module 130 is a control module for controlling the overall operation of the deep learning system 100, and can perform control of authentication and interoperability with another system.
The above-described learning-related information may be the number of learned source learning data pieces in the source learning data to be learned, the ratio of the learned data to the data to be learned, or the number of learned source learning data pieces and the number of times of performing learning. It is important to have the artificial neural network model of the partner-side deep learning system 200 or 300 learn in the same manner as learning in one deep learning system 100.
The deep learning systems 100, 200, and 300 according to another embodiment of the present invention may be re-implemented by modifying the data classification storage module 110 among the above elements. That is, the data classification storage module 110 may include: a data receiving unit configured to receive image data of an inspection target from an external device as source learning data, and store the received image data in a memory; and a data classifying unit configured to classify and store the stored image data of the inspection target according to a plurality of multi-stage classification criteria.
In this case, the data classifying unit may classify and store the image data of the inspection target using at least two or more of a first-stage classification criterion (which is a classification criterion of a color), a second-stage classification criterion (which is a classification criterion of an inspection target size), and a third-stage classification criterion (which is a classification criterion of a combination of a color and a shape).
The first-stage classification standard of the data classification unit includes color values for identifying one or more of an acetic acid reaction image, a rugo solution reaction image, a green filter image, and a normal image as classification standard values.
When the data classification storage module 110 is a system capable of classifying according to the data classification criteria, the adjustment information transmission module 130 preferably includes at least two or more of the first-level classification criteria, the second-level classification criteria, and the third-level classification criteria in the adjustment information, and transmits the adjustment information to the external deep learning system 200 or 300 to synchronize the interoperable systems.
While the operation of the deep learning system 100, 200, or 300 including the above elements will be further described, a process in which the first deep learning system 100 and the second deep learning system 200 interoperate with each other to optimize the artificial neural network model will be described. For convenience of description, the artificial neural network model included in the first deep learning system 100 will be referred to as a second artificial neural network model, and the artificial neural network model included in the second deep learning system 200 will be referred to as a first artificial neural network model.
Fig. 4 is a diagram for additionally describing a deep learning process according to an embodiment of the present invention.
In fig. 4, it is assumed that the second deep learning system 200 has a smaller number of source learning data pieces than the first deep learning system 100, and has a different operating system. Further, it is assumed that each of the first deep learning system 100 and the second deep learning system 200 has a storage unit in which source learning data to be learned is stored.
Under these assumptions, the system administrator accesses the first deep learning system 100 and sets up the deep learning system to interoperate (S10). The setting may be performed by inputting the connection information and the authentication information of the second deep learning system 200.
With this arrangement, the adjustment information transmission module 130 operable as the control module of the first deep learning system 100 can access the second deep learning system 200 and perform a mutual system authentication process (S20).
When the mutual system authentication is normally completed, the first deep learning system 100 and the second deep learning system 200 can transmit and receive required information. In this case, the adjustment information transmission module 130 sets the super parameter as the adjustment information to perform synchronization on the initial learning state of the second artificial neural network model of the second deep learning system 200, and preferentially transmits the adjustment information to the second deep learning system 200 (S30). When the super parameter is received as the adjustment information, the adjustment module of the second deep learning system 200 updates the parameter associated with the first artificial neural network model to perform synchronization with the second artificial neural network model.
When the initial learning states of the two deep learning systems are synchronized with each other through the transmission and reception of the super parameters, the learning module 120 of the first deep learning system 100 trains the second artificial neural network model 150 in the system using the source learning data stored in the memory (S40). The weight values W1, W2, W3. of the layers constituting the second artificial neural network model 150 are modified by learning.
When the learning module 120 learns a predetermined number of source learning data segments, the adjustment information transmission module 130 transmits adjustment information including learning-related information (the number of learned data segments in the source learning data) and parameter information (weight value and bias value) of the second artificial neural network model modified according to the learning result to the second deep learning system 200 to request adjustment of the first artificial neural network model (S50).
The adjustment model included in the second deep learning system 200 updates parameter information (i.e., weight values) of the first artificial neural network model in the second deep learning system 200 according to the adjustment information transmitted from the external first deep learning system 100 (S60). The second deep learning system 200 may also obtain the effect of learning the source learning data of the first deep learning system 100 by updating the weight values of the first artificial neural network model included in the external second deep learning system 200 with the weight values modified by the first deep learning system 100 during one learning process.
After updating the parameters of the first artificial neural network model using the adjustment information transmitted from the outside, the learning module of the second deep learning system 200 also learns its own source learning data (S70), and performs the learning according to the learning-related information transmitted from the first deep learning system 100.
When learning is completed based on the learning-related information, parameters of the first artificial neural network model are also modified. By transmitting the adjustment information including the modified weight value and the bias value to the first deep learning system 100 by the adjustment information transmission module 130 included in the second deep learning system 200 (S80), the adjustment module 140 included in the first deep learning system 100 updates the parameter information of the second artificial neural network model 150 based on the received adjustment information (S90). Accordingly, the first deep learning system 100 can also obtain the effect of learning the source data of the second deep learning system 200.
Meanwhile, the first learning system 100 and the second deep learning system 200 repeat the above-described adjustment process until a specific performance is obtained, that is, until the artificial neural network model is optimized (S100), thereby optimizing the first artificial neural network model or the second artificial neural network model.
As described above, by transmitting the parameters obtained by learning its own source learning data to the partner-side deep learning system to update the parameters of the artificial neural network model, the same effects as those obtained by learning the same learning data can be obtained without sharing the source learning data. Accordingly, it is advantageously possible to solve the problem caused by the exposure of the personal information or the medical information in advance, and also to shorten the time required to optimize the artificial neural network model due to the limitation of sharing the personal information or the medical information.
Furthermore, the present invention is advantageous in that since the artificial neural network model is partially optimized using parameters obtained from learning results of an external system, it is practically impossible to recover or infer source learning data, thereby minimizing damage caused by information exposure.
In particular, the present invention can achieve the same effect as that obtained by learning the same learning data without sharing the source learning data, and therefore, even if the artificial neural network model is in an environment where only a small amount of source data can be learned, the artificial neural network model can be changed to an artificial neural network model where a large amount of source learning data is learned.
In the above-described embodiment, the process in which the two deep learning systems 100 and 200 interoperate with each other to optimize the artificial neural network model without sharing the source learning data has been described.
A deep learning system according to still another embodiment configured to perform learning after accurately classifying learning data for learning to optimize an artificial neural network model to improve accuracy of analyzing occurrence of a specific disease while the two deep learning systems 100 and 200 interoperate with each other to optimize the artificial neural network model will be described below. The deep learning system may be implemented by dividing the data classification storage module 110 shown in fig. 3 into a data receiving unit and a data classification unit. In the following embodiments, data classification criteria required for accurately diagnosing whether cervical cancer has occurred will be described as an example as an embodiment.
First, the data receiving unit is configured to receive image data of an inspection target corresponding to source learning data from an external device such as an image capturing device and store the received image data, and the data classifying unit is configured to classify and store the image data of the inspection target according to a plurality of multi-level classification criteria. In this case, the data classifying unit may classify the image data of the inspection target using at least two or more of a first-stage classification criterion (which is a classification criterion of color), a second-stage classification criterion (which is a classification criterion of inspection target size), a third-stage classification criterion (which is a classification criterion of a combination of color and shape), and a fourth-stage classification criterion (which is a classification criterion of exposure and focus).
In some cases, the data classifying unit may classify the image data of the inspection target first according to a first-stage classification criterion (which is a classification criterion of color), classify the image data of the inspection target classified by the first stage according to a second-stage classification criterion (which is a classification criterion of the size of the inspection target, for example, the size of cervix), and classify the image data of the inspection target according to a third-stage classification criterion (which is a classification criterion of a combination of color and shape in the image data of the inspection target classified by the second stage).
The first-stage classification standard may include color values for identifying one or more of an acetic acid reaction image, a rugo solution reaction image, a green filter image, and a normal image as classification standard values.
The third level classification criteria may include a combination of colors and shapes for identifying one or more of blood, mucus, rings, colposcopes, treatment scars, and surgical tools in the image data of the examination object.
For example, blood mainly appears as red color flowing down from the center of the cervix, mucus mainly appears as pale yellow color flowing down from the center of the cervix, and a ring is mainly located in the center of the cervix and usually and obviously appears as a boomerang-shaped line. Colposcopes and other surgical tools are displayed in colors other than pink (silver, blue, etc.) that indicate the cervix. Therefore, as described above, by using a combination of the color and morphology of each foreign substance, it is possible to classify the foreign substances affecting the cervix.
Meanwhile, the data classifying unit may separately classify images that are not classified according to the above three classification criteria (i.e., images in which no lesions are identified) according to the exposure degree and the focus classification criteria. For example, the classification (fourth classification) may be performed using a characteristic that the histogram is extremely biased to one side when the exposure is insufficient or excessive, or using a characteristic that an edge is not detected or the color contrast is not sufficiently high when the focus is out of range. During the above-described first to fourth classification processes, classification may be performed using a deep learning technique.
In addition, the data classifying unit may also generate additional source learning data (image data of the inspection target) to adjust the balance of the number of image data of the inspection target, that is, source learning data classified and stored based on the classification criterion. This is to prevent excessive learning from being performed only on cervical cancer image data of a specific type, or normal learning from not being performed on images of a specific type (or category).
A process of classifying image data of the cervix as image data of an inspection target to optimize the artificial neural network model will be described in detail with reference to fig. 5 and 6.
Fig. 5 shows a multi-stage classification flow of image data of an inspection object according to another embodiment of the present invention, and fig. 6 is a diagram for describing in detail a multi-stage classification criterion of image data of an inspection object according to still another embodiment of the present invention.
Referring to fig. 5, first, a data receiving unit included in the data classification storage module 110 receives image data of a cervix from an external device such as an image capturing device, and stores the received image data (S200).
The data classification unit classifies and stores one or more unclassified image data segments of the cervix according to a plurality of multi-level classification criteria (S210).
For example, the data classification unit preferentially performs first classification on unclassified image data of the cervix according to a first-level classification criterion (which is a classification criterion of color).
For the first classification, the data classification unit may classify the acetic acid reaction image, the rugo solution reaction image, the green filter image, and the normal image by using a color value for identifying the image as a classification standard value.
In detail, since white spots appear in the cervix of the acetic acid reaction image, the white spots can be distinguished from the cervix and vagina represented in pink. Since brown or dark orange appears in the rugo solution reaction image and green appears completely in the green filter image, unclassified image data of the cervix can be classified by using a color value indicating an image feature as a classification standard value.
When the first classification is completed, the data classification unit performs a second classification according to a second-stage classification criterion, which is a classification criterion for the size of the cervix in the image data classified for the first time.
The cervix is a 500-won coin-sized circle, usually located in the center of the image. Thus, for the size of the cervix in the image, the data classification unit may classify the image data a second time as being only an image of the enlarged cervix, an image of the entire cervix, an image of the cervix occupying about 80%, an image of the cervix occupying about 50%, and an image including the cervix, colposcope, and other parts, etc.
Subsequently, the data classifying unit classifies the foreign substances affecting the cervix for the third time according to a third-stage classification criterion, which is a classification criterion for a combination of the color and the shape in the cervical image data classified for the second time.
As described above, blood mainly appears as red color flowing downward from the center of the cervix, mucus mainly appears as pale yellow color flowing downward from the center of the cervix, and a ring is mainly located at the center of the cervix and usually and remarkably appears as a line in the shape of a boomerang. Colposcopes and other surgical tools are displayed in colors other than pink (silver, blue, etc.) that indicate the cervix. Therefore, as described above, by using a combination of the color and morphology of each foreign substance, it is possible to classify the foreign substances affecting the cervix.
In some cases, the data classification unit may perform a fourth classification on the third classified image based on the exposure and focus.
As described above, the image data classified according to the multi-level classification standard is stored in the memory on the basis of the classification standard.
When the classification of unclassified image data is completed, the first deep learning system 100 and the second deep learning system 200 optimize their own artificial neural network model through the process shown in fig. 4. In this case, in order to apply the multi-level classification standards of the image data employed by the single deep learning system in the same manner, the adjustment information transmission module 130 preferably includes at least two or more of the first-level classification standard, the second-level classification standard, and the third-level classification standard in the adjustment information, and transmits the adjustment information to the partner-side deep learning system.
In S220, which is not described in fig. 5, learning is performed on image data classified and stored according to the classification standard, and in S40 of fig. 4, learning is performed.
As described above, by a different deep learning system that shares the learning data classification criteria and the parameters modified according to the learning result while accurately classifying the learning data to be used for learning in advance and performing learning based on the classification criteria, even an artificial neural network model in an environment where only a small amount of source data can be learned can be changed to an artificial neural network model in which a large amount of source learning data is learned, and an artificial neural network model established by learning data created according to a plurality of classification criteria can diagnose whether a specific disease has occurred more accurately than in a conventional system.
Although the invention has been described with reference to the embodiments shown in the drawings, it will be understood by those skilled in the art that the embodiments are merely illustrative of the invention, and various modifications and equivalents may be made without departing from the spirit and scope of the invention. Accordingly, the technical scope of the present invention should be defined by the following claims.
Claims (7)
1. A deep learning system accessible via a communications network to an external deep learning system for learning a first artificial neural network model, the deep learning system comprising:
a data classification storage module configured to receive and store source learning data, wherein the data classification storage module comprises: a data receiving unit configured to receive image data of an inspection target from an external device as source learning data and store the received image data; and a data classifying unit configured to classify and store the stored image data of the inspection target according to a plurality of multi-stage classification criteria, wherein the data classifying unit further generates additional source learning data to adjust a balance of the number of source learning data classified and stored based on the classification criteria;
a learning module configured to learn a second artificial neural network model in the deep learning system using the source learning data;
an adjustment information transmission module configured to transmit adjustment information including learning-related information and parameter information of the second artificial neural network model to the external deep learning system including the same first artificial neural network model as the second artificial neural network model to request adjustment of the first artificial neural network model, wherein the learning-related information and the parameter information are modified according to a result of the learning; and
an adjustment module configured to update the parameter information of the second artificial neural network model according to adjustment information transmitted from the external deep learning system that updated parameters of the first artificial neural network model,
wherein the data classifying unit performs a first-stage classification of the image data of the inspection target according to a first-stage classification criterion which is a classification criterion for a color, performs a second-stage classification of the image data of the inspection target subjected to the first-stage classification according to a second-stage classification criterion which is a classification criterion for a size of the inspection target, and performs a third-stage classification of the image data of the inspection target subjected to the second-stage classification according to a third-stage classification criterion which is a classification criterion for a combination of a color and a shape in the image data of the inspection target subjected to the second-stage classification.
2. The deep learning system of claim 1, further comprising a diagnostic module configured to diagnose whether a particular disease has occurred based on the first artificial neural network model optimized by repeated operation of the learning module and the adjustment module.
3. The deep learning system according to claim 1 or 2, wherein the adjustment information transmission module preferentially transmits, as the adjustment information, super parameters required to perform synchronization on an initial learning state of the first artificial neural network model.
4. The deep learning system according to claim 1 or 2, wherein the parameter information included in the adjustment information transmitted by the adjustment information transmission module or the external deep learning system constitutes an artificial neural network model, and includes a weight value and a bias value of each layer modified according to a learning result.
5. The deep learning system according to claim 1 or 2, wherein the learning-related information includes any one of a ratio of learning data to source learning data to be learned and a number of learned source learning data pieces.
6. The deep learning system according to claim 1 or 2, wherein the first-stage classification criterion of the data classification unit includes color values for identifying one or more of an acetic acid reaction image, a rugo solution reaction image, a green filter image, and a normal image as classification criterion values.
7. The deep learning system of claim 1 or 2, wherein the adjustment information transmission module includes at least two or more of the first level classification criterion, the second level classification criterion, and the third level classification criterion in the adjustment information, and transmits the adjustment information to the external deep learning system.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR1020190018615A KR102391817B1 (en) | 2019-02-18 | 2019-02-18 | Deep learning system |
| KR10-2019-0018615 | 2019-02-18 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111582434A CN111582434A (en) | 2020-08-25 |
| CN111582434B true CN111582434B (en) | 2023-10-17 |
Family
ID=72110758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910404061.7A Active CN111582434B (en) | 2019-02-18 | 2019-05-15 | deep learning system |
Country Status (3)
| Country | Link |
|---|---|
| KR (1) | KR102391817B1 (en) |
| CN (1) | CN111582434B (en) |
| WO (1) | WO2020171321A1 (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113204614B (en) * | 2021-04-29 | 2023-10-17 | 北京百度网讯科技有限公司 | Model training method, method for optimizing training data set and device thereof |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2001026026A2 (en) * | 1999-10-04 | 2001-04-12 | University Of Florida | Local diagnostic and remote learning neural networks for medical diagnosis |
| WO2017043680A1 (en) * | 2015-09-11 | 2017-03-16 | 주식회사 루닛 | Artificial neural-network distributed learning system and method for protecting personal information of medical data |
| CN108182427A (en) * | 2018-01-30 | 2018-06-19 | 电子科技大学 | A kind of face identification method based on deep learning model and transfer learning |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2008533615A (en) * | 2005-03-14 | 2008-08-21 | エル ターラー、ステフエン | Neural network development and data analysis tools |
| CN105894087A (en) * | 2015-01-26 | 2016-08-24 | 华为技术有限公司 | System and method for training parameter set in neural network |
| KR20170083419A (en) | 2016-01-08 | 2017-07-18 | 마우키스튜디오 주식회사 | Deep learning model training method using many unlabeled training data and deep learning system performing the same |
| JP2020503604A (en) * | 2016-12-05 | 2020-01-30 | アビギロン コーポレイションＡｖｉｇｉｌｏｎ Ｃｏｒｐｏｒａｔｉｏｎ | Appearance search system and method |
| US11182676B2 (en) * | 2017-08-04 | 2021-11-23 | International Business Machines Corporation | Cooperative neural network deep reinforcement learning with partial input assistance |
| KR101841222B1 (en) * | 2017-08-11 | 2018-03-22 | 주식회사 뷰노 | Method for generating prediction results for early prediction of fatal symptoms of a subject and apparatus using the same |
-
2019
- 2019-02-18 KR KR1020190018615A patent/KR102391817B1/en active IP Right Grant
- 2019-05-15 CN CN201910404061.7A patent/CN111582434B/en active Active
- 2019-08-19 WO PCT/KR2019/010480 patent/WO2020171321A1/en active Application Filing
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2001026026A2 (en) * | 1999-10-04 | 2001-04-12 | University Of Florida | Local diagnostic and remote learning neural networks for medical diagnosis |
| WO2017043680A1 (en) * | 2015-09-11 | 2017-03-16 | 주식회사 루닛 | Artificial neural-network distributed learning system and method for protecting personal information of medical data |
| CN108182427A (en) * | 2018-01-30 | 2018-06-19 | 电子科技大学 | A kind of face identification method based on deep learning model and transfer learning |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020171321A1 (en) | 2020-08-27 |
| KR102391817B1 (en) | 2022-04-29 |
| KR20200100388A (en) | 2020-08-26 |
| CN111582434A (en) | 2020-08-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10580530B2 (en) | Diagnosis assistance system and control method thereof | |
| Bilal et al. | Diabetic retinopathy detection and classification using mixed models for a disease grading database | |
| KR102345026B1 (en) | Cloud server and diagnostic assistant systems based on cloud server | |
| AU2014271202B2 (en) | A system and method for remote medical diagnosis | |
| US20180260954A1 (en) | Method and apparatus for providing medical information service on basis of disease model | |
| KR102333670B1 (en) | Diagnostic auxiliary image providing device based on eye image | |
| WO2021180244A1 (en) | Disease risk prediction system, method and apparatus, device and medium | |
| KR20190132832A (en) | Method and apparatus for predicting amyloid positive or negative based on deep learning | |
| KR20200139606A (en) | Cervical cancer diagnosis system | |
| CN111582434B (en) | deep learning system | |
| CN114782394A (en) | Cataract postoperative vision prediction system based on multi-mode fusion network | |
| Bhuiyan et al. | An artificial-intelligence-and telemedicine-based screening tool to identify glaucoma suspects from color fundus imaging | |
| Gong et al. | Application of deep learning for diagnosing, classifying, and treating age-related macular degeneration | |
| JP6468576B1 (en) | Image diagnostic system for fertilized egg, image diagnostic program for fertilized egg, and method for creating classifier for image diagnosis of fertilized egg. | |
| KR20050043869A (en) | Developing a computer aided diagnostic system on breast cancer using adaptive neuro-fuzzy inference system | |
| KR102036052B1 (en) | Artificial intelligence-based apparatus that discriminates and converts medical image conformity of non-standardized skin image | |
| Lo et al. | Data Homogeneity Effect in Deep Learning-Based Prediction of Type 1 Diabetic Retinopathy | |
| CN111402246A (en) | Eye ground image classification method based on combined network | |
| CN116563224A (en) | Image histology placenta implantation prediction method and device based on depth semantic features | |
| JP7346600B2 (en) | Cervical cancer automatic diagnosis system | |
| WO2022205780A1 (en) | Method and apparatus for classifying eye examination data on basis of cross-modal relationship inference | |
| CN115035133A (en) | Model training method, image segmentation method and related device | |
| CN113191413A (en) | Prostate multimode MR image classification method and system based on foveal residual error network | |
| Biswas et al. | Estimating Risk Levels and Epidemiology of Diabetic Retinopathy using Transfer Learning | |
| KR20200018360A (en) | Cervical learning data generation system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20200825 Assignee: Shenzhen baozhiplatinum Intelligent Medical Technology Co.,Ltd. Assignor: Korea Treasure Platinum Co.,Ltd. Contract record no.: X2021990000544 Denomination of invention: Deep learning system License type: Common License Record date: 20210903 |
|
| EE01 | Entry into force of recordation of patent licensing contract | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |