WO2019164064A1

WO2019164064A1 - System for interpreting medical image through generation of refined artificial intelligence reinforcement learning data, and method therefor

Info

Publication number: WO2019164064A1
Application number: PCT/KR2018/005641
Authority: WO
Inventors: 이병일; 김성현
Original assignee: (주)헬스허브
Priority date: 2018-02-26
Filing date: 2018-05-17
Publication date: 2019-08-29
Also published as: KR102153920B1; KR20190102399A

Abstract

The present invention relates to a system for interpreting a medical image through generation of refined artificial intelligence reinforcement learning data, and a method therefor, and to a system for interpreting a medical image through generation of refined artificial intelligence reinforcement learning data, and a method therefor, the system extracting reinforcement learning data for medical image interpretation from an interpretation text of a medical image interpretation expert so as to utilize same as artificial intelligence learning data, thereby decreasing the calculation costs and the complexity of medical image interpretation using artificial intelligence and improving the accuracy thereof.

Description

Medical image reading system and its method by generating purified artificial intelligence reinforcement learning data

The present invention relates to a medical image reading system and method through the generation of purified artificial intelligence reinforcement learning data, and more particularly, to extract the reinforcement learning data of medical image reading from the medical image reading expert's readings The present invention relates to a medical image reading system and a method for generating artificial intelligence enhanced learning data that can reduce the computational cost and complexity of artificial intelligence medical image reading and improve accuracy.

Among medical devices used to diagnose human diseases, many of them acquire and output medical images. An image obtained by scanning a human body part is processed to read whether a lesion is present.

A doctor's ability is often determined by the ability to read lesions from medical images without missing them. Accurate judgment of medical images is of paramount importance because early detection of pathologies and early detection of pathologies are very important in increasing the likelihood of treating and curing the disease.

However, it is not difficult to find all lesions well in various fields of medical imaging, and it is very difficult to accurately find lesions in medical imaging without gaining a long time experience in a particular field. In addition, the reality is that it is not uncommon for specialists with long experience in a particular field.

Many people have recognized this problem for a long time and tried to solve it. A representative one is the concept of computer-aided diagnosis (CAD). This is done by digitizing the scanned medical image, processing the image with a computer, and then constructing a rule-based or expert system through mathematical modeling based on the characteristics of the object.

The image analysis technology used here is mainly evolving from pattern recognition by image processing to prediction of lesions through machine learning. It extracts features from images, vectorizes images with extracted features, and then uses various machine learning classification techniques. In recent years, deep learning has become a mainstream method.

The main tasks of analyzing medical images include classifying images, detecting objects, segmenting objects, and regulating different images. Convolutional neural networks (CNNs), which are optimized for processing images as inputs, are the most widely used as a means to solve these major problems.

Most of the medical image learning methods belong to the category of supervised learning in which CNN learns the functional relationship between input and correct answers using input data and correct answer data as learning data. However, deep learning-based artificial intelligence technology requires a very large number of learning data, and thus, refining the learning data is a very important issue to cover a large number of learning data. In other words, it is important to secure a large amount of high quality learning data and to ensure accurate read performance.

Therefore, in the present invention, by extracting the reinforcement learning data of the medical image reading from the readings of the medical image reading expert and using it as the artificial intelligence learning data, it is a refined artificial which can lower the calculation cost and complexity of the artificial intelligence medical image reading and improve the accuracy. We present a medical image reading system and its method through the generation of intelligent reinforcement learning data

More specifically, the structure and method of a supervised learning model that generates a normalized form from the read text data of the reading expert, the structure of the refined learning data extracted from multiple valid reads, and the data generation. The present invention provides a medical image reading system and method for improving performance of a conventional CNN by applying purified learning data to a CNN.

Next, the prior art existing in the technical field of the present invention will be briefly described, and then the technical matters to be made differently from the prior art will be described.

First of all, Korean Patent Publication No. 10-2017-0140757 (December 21, 2017) relates to a clinical decision support ensemble system and a clinical decision support method using the same, and the clinical practice of a patient through machine learning received from a plurality of external medical institutions. A system for integrating prediction results and performing ensemble predictions to predict not only the current state of the patient but also the progress of the patient's disease in the future, thereby supporting prompt and accurate clinical decision making of medical personnel; It is about how.

In addition, Korean Patent Publication No. 10-2015-0108701 (2015.09.30.) Relates to a system and method for visualizing anatomical elements in a medical image, and to verify anatomical elements included in the medical image by using anatomical context information. By presenting a system and method for visualizing anatomical elements in a medical image that automatically classifies and user-friendly visualization of the classified anatomical elements.

Meanwhile, Korean Patent Publication No. 10-2015-0098119 (2015.08.27.) Relates to a system and method for removing false-positive lesion candidates in a medical image, and to verify lesion candidates detected in a medical image using anatomical context information. By providing a system and method for removing false-positive lesion candidates in a medical image to remove false-positive lesion candidates.

The prior art integrates clinical prediction results of patients to perform ensemble prediction, to use anatomical context information in medical images, and to remove false positive lesion candidates, or to provide extensive learning data for deep learning as in the present invention. It has not been suggested to reduce the computational cost and complexity and improve the accuracy of the learning model for the purification and the artificial intelligence medical image reading through it.

The present invention was created to solve the above problems, and extracts the reinforcement learning data of the medical image reading from the readings of the medical image reading expert (imagine medicine specialist) and utilizes it as artificial intelligence learning data It is an object of the present invention to provide a medical image reading system and method through refined artificial intelligence enhanced learning data generation that can reduce the calculation cost and complexity of reading and improve accuracy.

In addition, the present invention to improve the performance of the medical image reading system, to improve the learning effect through the image, to identify the presence and location of the lesion as well as to identify the various types of conditions that can appear in the same body parts Another object is to generate a medical report supervised learning model that can normalize texts contained in well-trained radiologist's readings to produce refined learning data.

In addition, the present invention to improve the performance of the medical image reading system, by extracting the refined learning data (refined data) by learning the readings that have been verified through the generated reading record supervision learning model, and together with the medical image image The purpose of the present invention is to provide a structure and generation method of refined learning data with a new structure that can enhance learning effects and identify types of diseases by reinforcement learning.

In addition, the present invention improves the read performance of the existing convolutional neural network and extracts the type of the disease by extracting filters that can identify the type of the condition from the refined learning data having the new data structure and interacting it with the convolutional neural network. The goal is to build a reading-based supervised learning model that can form a converged convolutional neural network with a recognizable new structure.

In order to achieve the above object, the medical image reading system through the generation of purified artificial intelligence reinforcement learning data according to a preferred embodiment of the present invention generates the normalized type of purified learning data extracted from the readings of a medical image reading expert. A reading record supervised learning unit and a learning model generating unit performing machine learning to read the medical image by inputting the learning data purified by the reading record supervising learning unit, wherein the machine learning includes the medical image. The learning data is automatically received by reading the expert's readings, and the input is characterized in that the learning data automatically becomes reinforcement learning data.

The reading record supervised learning unit may include a medical record data loading unit reading data from a file location address of the reading statement, findings, findings, and recommendations by the body parts. A labeling unit which classifies the section into a section including a recommendation, and labels a disease-related word or phrase as a set from the plain text of each section, extracts a disease-related word or phrase from the labeled reading, and extracts the word. Or a feature extractor for extracting a common feature from a phrase, a feature matrix generator for generating a feature matrix by regularizing the extracted features, and mapping the read to a feature matrix, if given. A feature analysis unit for analyzing features, and to generate purified learning data using only the analyzed features. It characterized in that it comprises a data generator.

The medical record data loading unit reads the position of the file, the total number of files, the length of the file, or a combination thereof from the file position address of the read original data given as an input value, and loads the file into the system memory. The data loaded in the memory is labeled for each read section by body part, classified into respective sets, and rearranged on the system memory, and the feature extracting unit includes the plain text of each section, SNOMED-CT, and ICD. Selecting only the plain text in comparison with the standard medical term data set including -11, LOINC, and KCD-10, and selecting the word type, description form, description frequency, or the like from the medical term extracted from the medical term extract section. Combination analysis to characterize terms related to the presence or absence of lesions, to characterize terms associated with the location of lesions, Extracts features of terms related to descriptions, features of terms representing types of conditions, or a combination thereof, and the feature matrix generator maps the features extracted from the feature extractor into data sets that are newly input as plain text. Generates a feature matrix that can be compared and analyzed for terms of similar or identical meaning, and the feature analyzer, when the unpurified original read is input, maps it to the feature matrix in the plain text that describes the read. Extracting, analyzing, and classifying the presence or absence of a lesion, the location of the lesion, the symptoms, and the type of the disease, and the purified data generation unit generates the learning data purified from the data extracted, analyzed, and classified by the feature analysis unit. It is characterized by.

The learning model generation unit is trained by a converged convolutional neural network, and the converged convolutional neural network continuously performs refined learning data from readings of the medical image reading expert. By receiving the updated input and reflecting it in the generation of the learning model, deep learning machine learning is performed through the fusion convolutional neural network to reduce the calculation amount and improve the accuracy to improve the overall performance.

In addition, the converged convolutional neural network is characterized in that the weight is updated by reverse propagation.

Meanwhile, in the medical image reading method through the generation of purified artificial intelligence reinforcement learning data according to another embodiment of the present invention, read-record instructional learning that generates purified learning data in a normalized form extracted from a reading of a medical image reading expert (supervised learning) and a learning model generation step of performing a machine learning to read the medical image with the input of the training data purified in the read recording instruction learning step, the machine learning is a reading of the medical image reading expert By receiving continuous input of the refined learning data from the input, the learning data is characterized in that the reinforcement learning data automatically.

The read history supervised learning step may include loading medical record data reading data from a file location address of the read statement, finding the read statement by body parts, findings, and conclusions. A labeling step of classifying into sections containing recommendations, labeling disease-related words or phrases as a set from the plain text of each section, and medical terminology extracting disease-related words or phrases from the labeled readings. Extracting a feature, extracting a common feature from the extracted word or phrase, generating a feature matrix by regularizing the extracted features, and generating a feature matrix; A feature analysis step of analyzing features by mapping to a feature matrix, and only the analyzed features And a refinement data generation step of generating refined learning data.

The medical record data loading step reads the file location, the total number of files, the length of the file, or a combination thereof from the file location address of the read original data given as an input value, and loads the file into the system memory. The data loaded in the system memory is labeled by each reading section for each body part, classified into respective sets and rearranged on the system memory, and the medical term extraction step includes the plain text of each section. Only the plaintext text is selectively extracted in comparison with a standard medical term data set including SNOMED-CT, ICD-11, LOINC, and KCD-10, and the feature extraction step is performed from the medical terminology extracted by the medical terminology extracting unit. Analyze word types, narrative forms, frequency of descriptions, or a combination thereof to determine the characteristics of a term related to the presence or absence of a lesion, Extracting features of terms related to the location of the sides, features of terms related to the description of symptoms, features of terms representing the types of symptoms, or a combination thereof, wherein the feature matrix generating step includes extracting the features extracted from the feature extractor. Create a feature matrix that can compare and analyze whether the terms have similar or identical meanings by mapping them with newly input plaintext text as a set, and the feature analysis step is performed when an unrefined original readout is inputted. Extracting, analyzing, and classifying the presence or absence of a lesion, the location of the lesion, the symptom, the type of the disease, and the like in the plain text that describes the reading by mapping it to a feature matrix, and the purified data generation step is performed by the feature analysis unit. And generating refined learning data from the classified data.

In the learning model generation step, learning is performed by a converged convolutional neural network, and the converged convolutional neural network continuously performs the refined learning data from readings of the medical image reading expert. By inputting the data into the learning model and reflecting it in the generation of the learning model, deep learning machine learning is performed through the fusion convolutional neural network to reduce the calculation amount and improve the accuracy to improve the overall performance.

In addition, the learning model generation step, characterized in that the weight is updated by the backward propagation.

As described above, according to the medical image reading system and the method through the generation of purified artificial intelligence reinforcement learning data of the present invention, the user uses the convolutional neural network to determine the presence of the lesion, the location of the lesion, and the pathology of the medical image. Conventional convolution by fusing the output value of the convolutional neural network through analysis of pixel information of the medical image and the output value through analysis of the read record of the same body region calculated as a learning result in the read record supervised learning model. It is possible to obtain more accurate readings than medical image readings using neural networks or to reduce computational complexity of convolutional neural networks, and to predict the types of complications that were not known through conventional convolutional neural networks. .

1 is a view showing the concept of a conventional medical image reading system.

2 is a view showing the concept of a medical image reading system and method through the generation of purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

3 is a view for explaining the configuration of a medical image reading system through the generation of purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

FIG. 4 is a block diagram illustrating a configuration of a medical report supervised learning model for generating purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

4 is a block diagram showing the configuration of a medical report supervised learning unit for generating purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

5 is a view showing a process of generating purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

6 is a conceptual diagram illustrating a configuration of a converged convolutional neural network (CCNN) through generation of purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

7 is a flowchart for generating a feature matrix of a read / write map learning model according to an embodiment of the present invention.

FIG. 8 is a flowchart of extracting feature values associated with a disease from an arbitrary reading in a reading recording supervising model according to an embodiment of the present invention.

9 is a flowchart of a medical image reading process through generation of purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

Hereinafter, with reference to the accompanying drawings will be described in detail a preferred embodiment of the medical image reading system and method through purified artificial intelligence enhanced learning data generation of the present invention. Like reference numerals in the drawings denote like elements. In addition, specific structural to functional descriptions of the embodiments of the present invention are only illustrated for the purpose of describing the embodiments according to the present invention, and unless otherwise defined, all terms used herein including technical or scientific terms These have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art, and are not construed in ideal or excessively formal meanings unless expressly defined herein. It is preferable not to.

1 is a view showing the concept of a conventional medical image reading system.

As shown in Figure 1, the conventional medical image reading system is a system that helps the doctor to determine the presence of the lesion through the medical image, using the medical image stored in the database of the local, hospital, medical institutions, etc. When a learning model was created and a new medical image was inputted, the input medical image was applied to the learning model to predict a lesion from the reading result.

In this case, since a learning model is generated from a number of medical images, a lot of time is required for learning, and there is a problem of inferior accuracy. Therefore, in order to overcome this problem, the present invention proposes a structure for performing reinforcement learning to upgrade the learning model with readings which are reading results of medical images of specialists.

As shown in FIG. 2, a specialist configures a reading by reading a medical image, and uses the generated reading to improve the learning result of the learning model of the medical image reading system. In particular, the present invention serves to improve the learning performance and reduce the complexity in the process of generating a learning model for learning a medical image using the reading result of the reading.

Hereinafter, a detailed configuration of a medical image reading system according to an embodiment of the present invention will be discussed.

As shown in FIG. 3, the medical image reading system 10 according to an exemplary embodiment of the present invention may include a readout instruction learning unit 100, a medical image learning unit 200, a medical image database 300, and a reading database ( 400) and the like. The medical image database 300 and the reading database 400 may be configured in one database.

The radiology specialist reads the medical image 300 and then writes a readout 400 of the image. The medical image is input to the medical image learning unit 200 to perform machine learning. In addition, the reading 400 is input to the reading guidance learning unit 100 to extract the characteristics of the reading and provide it to the medical image learning unit 200 to improve the learning performance (calculation amount, complexity) for the medical image.

Therefore, in the medical image reading system 10 of the present invention, the medical image learning unit 200 which learns the medical image by inputting the medical image 30 is extracted from the readout of the medical image 300 read by a specialist. By using the readout guidance learning unit 100 provides the purified learning data required by the medical image learning unit 200, thereby improving the learning performance of the medical image,

In this process, the readings of the medical images of the radiologist and specialist are gradually advanced and reflected, thereby improving the learning performance of the medical images.

Hereinafter, the configuration of the readout guidance learning unit 100 and the medical image learning unit 200 of the medical image will be described in detail.

As shown in FIG. 4, the read record map learning unit 100 discovers the medical record data loading unit 111 that reads data from the file location address of the read statement, and the read document for each body part ( A labeling processor 112 for classifying into sections including Findings, Conclusions, and Recommendations, and labeling a disease-related word or phrase from a plain text of each section into a set, the disease in the labeled readings The feature extractor 113 extracts a related word or phrase, extracts a common feature from the extracted word or phrase, and generates a feature matrix by regularizing the extracted features. ). The resulting feature matrix is stored in a database and can be used to analyze features by mapping incoming readings to the feature matrix.

In addition, the reading history map learning unit 100 maps the readings to a feature matrix through a map learning, given a random reading, and analyzes the features by the map learning feature analysis unit 120 and the analyzed features only. Further comprising a refined data generation unit 130 for generating purified learning data.

Here, the medical record data loading unit 111 reads the position of the file, the total number of files, the length of the file, or a combination thereof from the file position address of the original reading data given as an input value, and loads it into the system memory or the auxiliary memory. The results of the study will be used for reading, reading and teaching.

The labeling processor 112 labels the data loaded in the system memory or the auxiliary memory for each read section for each body part, classifies them into sets, and rearranges the data in the system memory or the auxiliary memory.

The feature extractor 113 selectively extracts only the plaintext text by comparing the plaintext text of each section with a standard medical term data set including SNOMED-CT, ICD-11, LOINC, and KCD-10. Analyze the word type, description form, frequency of description, or a combination thereof from the medical terms extracted by the extraction unit, and the characteristics of the terms related to the presence of the lesion, the characteristics of the terms related to the location of the lesion, and the characteristics of the terms related to the description of symptoms. Extract features or combinations of terms that indicate the type of condition.

The feature matrix generator 114 compares and analyzes whether the feature extracted by the feature extractor 130 is a similar or identical meaning by mapping the features extracted from the feature extractor 130 to the newly input plaintext text. Matrix).

Next, the supervised learning feature analysis unit 120, when the unrefined original readout is input, applies it to the supervised learning model, maps it to the feature matrix, and indicates the presence or absence of the lesion in the plaintext text describing the readout. Extract, analyze and classify symptoms, types of symptoms, and more.

The refined data generation unit 130 generates the refined learning data from the data extracted, analyzed, and classified by the supervised feature analysis unit 120. The refined learning data corresponds to additional information for improving performance of reading a medical image.

In particular, the supervised learning feature analysis unit 120 performs supervised learning, first by learning a readout of any newly input medical image based on a well-defined feature matrix for the medical image, and the corresponding readout is characterized by Determine if you have a matrix.

To this end, the newly read readings on any medical images are extracted from disease-related words or phrases through the Natural Language Processor (NLP), and inputted into the supervised learning model. Is extracted.

If any new reading is inputted, it is classified by comparing with the existing feature matrix. At this time, since the feature matrix must be extracted from the new reading, the medical record data loading unit 111, the labeling processing unit 112, and the feature extraction unit A method of performing the function of the reference feature matrix extractor 110 including the process of the 113 and the feature matrix generator 114 may be performed as it is, or may be inputted into a map learning model and classified.

Preferably both methods can be used in the present invention.

Hereinafter, a process of generating refined learning data by receiving text data for learning will be described.

As shown in FIG. 5, first, a reading (R _k ) of a radiologist is input to generate a set of extracting only the contents corresponding to finding (F) among readings, and a conclusion of the readings (Conc. Create a set that extracts only the content of C), and create and label a set that extracts only the content of Recommendation (R) from the readings.These sets need to be distinguished by body part (BP). .

Extract the feature from this label. Where FM ₁ ,. , FM _x, etc., is a set of feature metric extracted from the F, C, and R sets of labeled readings. In other words, we extract features and then build metrics for them. For example, the feature metric of the Xth feature is mapped to A, which converges or dominates all k synonyms, expressions, and vocabularies representing the same disease or condition from a ₁ to a _k , and the entire feature matrix. (Feature Metrix) consists of each metric indicating disease name, location expression, and severity. The generated result is RR _x and outputs data whose X-th row data R _x is refined through a feature matrix.

6 is a conceptual diagram illustrating a configuration of a convergent convolutional neural network (CCNN) through generation of purified artificial intelligence enhanced learning data according to an embodiment of the present invention.

As shown in FIG. 6, by applying purified artificial intelligence reinforcement learning data to a conventional convolutional neural network, a fused convolutional neural network can be constructed. For example, from the plain text describing the reading in the refined AI data, the presence or absence of the lesion, the location of the lesion, the symptom, and the type of the disease are extracted, analyzed, and classified to intensively predict the parts to be predicted in the convolutional stage. The other part does not add a relatively large amount of computational complexity. That is, by precisely learning about the area where the lesion exists and learning differently according to the symptoms or the type of the lesion, the complexity is reduced and the learning performance is improved as compared to the same intensity learning for all the input images. .

If a random reading is input, the reading is decoded for the medical image, and then the medical records are analyzed according to the supervised learning model. The information on the presence or absence of the lesion, the location of the lesion, the symptom, and the type of the symptom is extracted. And classify and construct a convolutional neural network fused with supervised learning models.

The fused convolutional neural network has a plurality of convolutional layers, and each time, a customized convolution is performed by using the result of supervised learning. In this way, the present invention can create a learning model with further improved performance. Using the medical images and the readings of the medical images, the radiologists read the readings, label them by findings, conclusions, and recommendations for each body part, extract the features, generate and store the feature matrix, and save the results. The feature matrix is also extracted from the medical image readings and then mapped with the stored feature matrix to derive refined learning data. Segmenting this feature matrix can extract information about the presence or absence of the lesion, the location of the lesion, the symptoms, and the type of symptom. Through this, it is possible to construct a converged CNN and perform learning.

In addition, the convergence convolutional network according to the present invention is configured such that the weight is updated by reverse propagation reflecting the evaluation result of the specialist in reverse. In updating the weight of the parameter in each step of configuring the CNN, the read result of the output unit is propagated to the hidden layer and the convolutional layer by inverting the weight to be corrected to enable more accurate reading.

As shown in FIG. 7, the feature matrix generation process of the supervised learning model first loads the dicom metadata and the readout of the medical image image from the network or the local storage (S110). Next, a body part field included in the dicom metadata of the loaded medical image image is extracted (S120). Subsequently, plain texts are inserted into each set by labeling the readout of the medical image by findings, conclusions, and recommendations (S130).

Subsequently, if there is an additional medical image (S140), the process of S110 to S130 is repeated, or a standard medical term data set is loaded (S150). In addition, a word or vocabulary related to a disease is extracted by sequentially mapping each element of the corresponding set and the standard medical term data set according to labeled findings, conclusions, and recommendations (S160).

Features are extracted from the extracted words and vocabularies through word type, description form, and description frequency analysis (S170). Finally, a feature matrix including the extracted features as an element is generated (S180).

As shown in FIG. 8, the process of extracting a disease-related feature value from an arbitrary reading according to an embodiment of the present invention first loads the dicom metadata and reading of the medical image image from a network or a local storage ( S210). Subsequently, a body part field included in the dicom metadata of the medical image image is extracted (S220).

The plain text is extracted for each section by labeling the readout of the medical image image by findings, conclusions, and recommendations (S230). The extracted plain text is mapped to elements of the feature matrix of the same body part (S240). As a result of mapping the feature matrix, data of text in which similar or identical terms or lexical expressions exist is extracted (S250).

The extracted data can be analyzed to extract information on the presence or absence of lesions, the location of the lesions, symptoms, and types of symptoms. Through this, it is possible to construct a converged CNN and perform learning.

Referring to FIG. 9, the medical image reading process through the generation of purified artificial intelligence reinforcement learning data according to an embodiment of the present invention may be performed to first generate purified learning data in a normalized form extracted from a reading of a medical image reading expert. Record supervised learning is performed (S310).

Subsequently, machine learning is performed to read the medical image with input of the learning data purified in the read / write instruction learning step (S320).

By continuously updating and receiving refined learning data from the readings of the medical image reading expert and reflecting it in the generation of the learning model, deep learning machine learning is performed through the converged convolutional neural network to reduce the calculation amount and improve the accuracy. Improve the performance (S330).

According to the present invention, when a user reads the presence or absence of a lesion of the medical image image, the location of the lesion, the type of the disease, etc. using the convolutional neural network, the user outputs the output value of the convolutional neural network by analyzing pixel information of the medical image image. Convergence of output values through analysis of reading records for the same body parts calculated as learning results in reading records supervised learning model to obtain accurate readings than medical image readings using conventional convolutional neural networks, or computational complexity of convolutional neural networks The complexity can be lowered and even a kind of disease unknown through the conventional convolutional neural network can be predicted.

Claims

A read recording supervised learning unit for generating purified learning data in a normalized form extracted from a reading of a medical image reading expert; And

And a learning model generation unit configured to perform machine learning to read the medical image by inputting the learning data purified by the read / write guidance learning unit.

The machine learning continuously receives the refined learning data from the readout of the medical image reading expert and receives the input, whereby the learning data is automatically enhanced reinforcement learning data generation. Medical Image Reading System.
The method according to claim 1,

The read record supervised learning unit,

A medical record data loading unit for reading data from a file position address of the read statement;

Classifying the readings into sections including findings, conclusions, and recommendations by body part, and labeling disease-related words or phrases as a set from the plain text of each section. A labeling processor;

A feature extracting unit extracting a disease-related word or phrase from the labeled reading and extracting a common feature from the extracted word or phrase;

A feature matrix generator for generating a feature matrix by regularizing the extracted features;

A feature analyzer for analyzing a feature by mapping the read sentence to a feature matrix when an arbitrary read sentence is given; And

And a refined data generation unit for generating purified learning data using only the analyzed features.
The method according to claim 2,

The medical record data loading unit reads the location of the file, the total number of files, the length of the file, or a combination thereof from the file location address of the original reading data given as an input value, and loads the file into the system memory.

The labeling processor may label the data loaded in the system memory for each read section for each body part, classify them into sets, and rearrange the data in the system memory.

The feature extractor selectively extracts only the plaintext text by comparing the plaintext text of each section with a standard medical term data set including SNOMED-CT, ICD-11, LOINC, and KCD-10. Analyze word types, descriptive forms, frequency of descriptions, or a combination thereof from medical terms to determine the characteristics of the term associated with the presence of the lesion, the characteristics of the term associated with the location of the lesion, the characteristics of the term associated with the description of the symptoms, and the type of condition. Extract features or combinations of terms that appear;

The feature matrix generator generates a feature matrix capable of comparing and analyzing whether the terms have similar or identical meanings by mapping the features extracted by the feature extractor as data sets to the newly input plain text.

The feature analyzing unit extracts, analyzes, and classifies the presence or absence of the lesion, the location of the lesion, the symptom, the type of the disease, and the like in the plain text that describes the reading by mapping the unrefined original readout into a feature matrix. ,

The purification data generation unit, the medical image reading system through the generation of purified artificial intelligence enhanced learning data, characterized in that for generating the learning data purified from the data extracted, analyzed and classified by the feature analysis unit.
The method according to claim 1,

The learning model generation unit is trained by a converged convolutional neural network, and the converged convolutional neural network continuously performs refined learning data from readings of the medical image reading expert. Refined artificial intelligence reinforcement learning data characterized by improving the overall performance by reducing the amount of calculation and improving accuracy by performing deep learning machine learning through the converged convolutional neural network. Medical image reading system through generation.
The method according to claim 4,

The learning model generation unit is a medical image reading system through the generation of purified artificial intelligence enhanced learning data, characterized in that the weight is updated by the reverse propagation.
A read record supervised learning step of generating purified learning data in a normalized form extracted from a reading of a medical image reading expert; And

And a learning model generation step of performing a machine learning to read the medical image with input of the learning data purified in the read / write instruction learning step.

The machine learning continuously receives the refined learning data from the readout of the medical image reading expert and receives the input, whereby the learning data is automatically enhanced reinforcement learning data generation. How to read medical images.
The method according to claim 6,

The read recording supervised learning step,

A medical record data loading step of reading data from a file location address of the read statement;

Classifying the readings into sections including findings, conclusions, and recommendations by body part, and labeling disease-related words or phrases as a set from the plain text of each section. A labeling process step;

A medical terminology extracting step of extracting a disease-related word or phrase from the labeled reading;

A feature extraction step of extracting a common feature from the extracted word or phrase;

Generating a feature matrix by regularizing the extracted features;

A feature analysis step of mapping a read to a feature matrix to analyze a feature, if any read is given; And

And a purified data generation step of generating purified learning data using only the analyzed features.
The method according to claim 7,

The medical record data loading step reads the location of the file, the total number of files, the length of the file, or a combination thereof from the file location address of the original reading data given as an input value, and loads them into the system memory.

The labeling process may include labeling the data loaded in the system memory for each read section by body part, classifying the data into respective sets, and rearranging the data in the system memory.

The medical term extraction step is to selectively extract only the plaintext text compared to the plain text of each section and the standard medical term data set including SNOMED-CT, ICD-11, LOINC, KCD-10,

The feature extracting step may be performed by analyzing a word type, a description form, a description frequency, or a combination thereof from the medical related terms extracted by the medical terminology extracting unit. Extract features, features of the term associated with the description of symptoms, features of the term indicating the type of condition, or a combination thereof;

The feature matrix generation step may generate a feature matrix capable of comparing and analyzing whether the terms have similar or identical meanings by mapping the features extracted by the feature extractor as data sets to the newly input plain text.

In the feature analysis step, when an unpurified original readout is input, it is mapped to a feature matrix to extract, analyze, and classify the presence or absence of the lesion, the location of the lesion, the symptom, and the type of the disease in the plaintext text describing the readout. ,

Wherein the step of generating the purified data, the method for reading a medical image through the generation of purified artificial intelligence reinforcement learning data, characterized in that for generating the training data purified by the extracted, analyzed and classified data in the feature analysis unit.
The method according to claim 6,

In the learning model generation step, learning is performed by a converged convolutional neural network, and the converged convolutional neural network continuously performs the refined learning data from readings of the medical image reading expert. Refined artificial intelligence reinforcement learning, characterized by improving the overall performance by performing deep learning machine learning through the converged convolutional neural network, reducing the calculation amount and improving the accuracy by receiving the input and updating it to the generation of the learning model. Medical image reading method through data generation.
The method according to claim 9,

The learning model generation step, the medical image reading method through the generation of purified artificial intelligence enhanced learning data, characterized in that the weight is updated by the reverse propagation.