KR102097742B1 - System for Searching medical image using artificial intelligence and Driving method thereof - Google Patents

Abstract

The present invention relates to a medical image search system based on artificial intelligence and a driving method thereof. The medical image search system based on artificial intelligence comprises: a data input unit to receive medical image data obtained by photographing at least a portion of a patient′s body; a preprocessing unit to perform preprocessing on the medical image data inputted to the data input unit; a feature map extraction unit to extract a feature map of the medical image data on which the preprocessing is performed by the preprocessing unit; and an image search unit to compare the extracted feature map with a feature map of previously stored image data, and search for image data with a matching degree with the extracted feature map higher than or equal to a reference value among the previously stored image data to output search results. An image is found by a feature map of images to increase search accuracy and improve search speed in comparison to a conventional method of simply searching for an image by a keyword of an image query statement.

Description

System for searching medical image using artificial intelligence and driving method thereof

The present invention relates to an artificial intelligence based medical image search system and a driving method thereof, and more particularly, to an artificial intelligence based medical image search system using the feature map of medical image data and a driving method thereof.

Video search engines, such as Google's image search and Naver image search, use a method of database based on tag values or category values assigned to existing images themselves and then searching based on them.

On the other hand, in modern medicine, medical imaging is a very important tool for effective disease diagnosis and patient treatment. In addition, the development of imaging technology makes it possible to acquire more sophisticated medical imaging data. In exchange for this sophistication, the amount of data is gradually increasing, making it difficult to analyze medical image data depending on human vision. Therefore, in the clinical decision support system and computer-assisted diagnostic technology, automatic analysis of medical images and large-scale medical image search are frequently required.

For example, in Korean Patent Application Publication No. 10-2017-0017614 (2017.02.15), a method and apparatus for calculating disease diagnosis information based on a medical image, detect a region of interest in which an object to be analyzed is photographed, and calculate a coefficient of variation And, it includes the steps of creating a coefficient of variation image and comparing it with a reference sample, and refers to the effect of diagnosing a patient's disease degree by using medical images obtained through CT, MRI, and ultrasound imaging devices. However, since this prior art uses a method of comparing medical images, there is still a need for a method capable of preventing fundamental diagnosis errors of a diagnostic device.

In addition, in Korean Patent Publication No. 10-2017-0047423 (2017.05.08), 'Automatic Tuberculosis Diagnosis Prediction System for CAD-based Digital X-rays' includes a diagnostic unit that determines whether tuberculosis is applied by applying a deep learning algorithm, and is based on CAD. It is emphasized that the diagnostic efficiency is improved by digital X-ray's automatic TB diagnosis prediction system. However, this prior art uses a deep learning algorithm, but has a drawback that still contains errors or limitations in medical diagnosis by the deep learning algorithm.

Meanwhile, with the development of artificial intelligence technologies such as AlphaGo, the development and use of artificial intelligence-based search engines are increasing. Search engine algorithms based on existing artificial intelligence technology are produced based on natural images, not medical images.

In the case of a search engine for medical images, the search is performed by a simple search algorithm based on tag values assigned to medical image data, such as the past natural image search algorithm.

Therefore, the necessity of an artificial intelligence search algorithm specialized for medical images is emerging to secure the accuracy and speed of search in medical images.

KR 1020170047423 A KR 1020170017614 A KR 1018180740000 B KR 1018110280000 B

The present invention is derived from such a technical background, and provides an artificial intelligence-based medical image search system and a driving method capable of providing an artificial intelligence search algorithm specialized for medical images, away from the existing general tag and database search algorithms. Has its purpose.

The present invention for achieving the above object includes the following configuration.

That is, the artificial intelligence-based medical image search system according to an embodiment of the present invention includes a data input unit that receives medical image data captured at least a part of a patient's body, and pre-processes the medical image data received by the data input unit. The feature map extractor extracts a feature map of medical image data pre-processed by the pre-processor and compares the extracted feature map with feature maps of pre-stored image data, and extracts the pre-stored image data. And an image search unit that searches for image data having a degree of matching with the feature map that is greater than or equal to a reference value and outputs a search result.

On the other hand, the driving method of the artificial intelligence-based medical image retrieval system is a data input unit receiving medical image data captured at least a part of the patient's body, and a pre-processing unit performing pre-processing on the medical image data received by the data input unit. The feature map extracting unit extracts a feature map of the medical image data preprocessed by the preprocessing unit, and the image search unit compares the extracted feature map with feature maps of pre-stored image data, and among the pre-stored image data And searching image data having a degree of matching with the extracted feature map that is greater than or equal to a reference value and outputting a search result.

According to the present invention, by finding an image through a feature map of the image itself, an effect capable of increasing the accuracy of the search and improving the search speed is derived than the existing method of simply searching the image with the keyword of the image query.

1 is a block diagram showing the configuration of an artificial intelligence based medical image search system according to an embodiment of the present invention,
2 is a flowchart illustrating a method of driving an AI-based medical image search system according to an embodiment of the present invention.

It should be noted that the technical terms used in the present invention are only used to describe specific embodiments, and are not intended to limit the present invention. In addition, technical terms used in the present invention should be interpreted as meanings generally understood by a person having ordinary knowledge in the technical field to which the present invention belongs, unless defined otherwise. It should not be interpreted as a meaning, or an excessively reduced meaning.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram showing the configuration of an artificial intelligence-based medical image search system according to an embodiment of the present invention.

As can be seen in Figure 1, the AI-based medical image search system according to an embodiment of the present invention includes a data input unit 100, a pre-processing unit 110, a feature map extraction unit 120, an image search unit 130 , Database 140.

The data input unit 100 receives medical image data taken of at least a part of the patient's body.

The data input unit 100 receives medical image data obtained by capturing at least a part of a patient's body from at least one medical device. A medical device is a device that measures and outputs an electrical signal by measuring a patient's affected part signal, and converts the output electrical signal into an image, and includes an ultrasound scanner, an MRI device, and a CT.

For example, the types of images include X-ray, CT (Computed Tomography), Magnetic Resonance Imaging (MRI), US (Ultrasound), PET (Positron Emission Tomography), etc., and at least a part of the body includes a brain and a chest ( Chest, abdomen, lungs, cardiovascular, bone, eyeball, etc. Since there are many types of images for each body part, there can be many categories.

The data input unit 100 may categorize and manage additionally input medical image data for each body part or for each imaging technique.

The output electrical signal may be imaged in a continuous form and transmitted to the data input unit 100 in real time. That is, the data input unit 100 may sequentially receive real-time images from a medical imaging apparatus.

In one aspect, the data input unit 100 receives image data converted to a binary file in a message to memory (DMA) using a message queue method. Accordingly, it is possible to design to offload the system hardware.

In one embodiment, the binary file input to the medical image data input unit 100 may be converted into image data and usable in an AI-based medical image search system.

The pre-processing unit 110 pre-processes the medical image data received through the data input unit 100.

In one aspect, the pre-processing unit 110 constantly adjusts the resolution of the input image data, analyzes the intensity value of the image data, and histograms to highlight a specific area based on the histogram result. The weak and strong parts are separated to clarify the distribution of differences in intensity values of the image data.

Since the image data received by the data input unit 100 is present in the memory, it can be used immediately without additional processing.

In one embodiment, the pre-processing unit 110 separates meta data and image data from the medical image data input to the data input unit 100 and then performs a pre-processing process to maximize the performance of the artificial intelligence system. As described above, it is possible to systematically manage a large amount of medical image data by separating the image data from the metadata and tagging the metadata to store and manage it in a database.

In one embodiment, image data input to the data input unit 100 has different resolutions. Therefore, it is necessary to constantly adjust the resolution of the input image data and standardize it.

In one embodiment, it is possible to standardize the resolution by performing a process of mapping the feature map of the low-resolution image to the most similar area of the feature map of the high-resolution image and converting it to the most similar high-resolution information. .

Then, the intensity value of the image data is analyzed to make a histogram, and based on this, a specific area is emphasized to separate the weak and strong parts. Then, the process of clarifying the distribution of differences in intensity values of image data is performed.

The brightness of light is a noise factor that is a major obstacle in classifying images. This is because when the shadow is generated, the code corresponding to the color of the pixel at the corresponding position changes, that is, it is said to be an image of the same body part. It is an image with completely different pixel values. This technique of minimizing interference by illumination is called intensity normalization.

In one embodiment, noise may be removed using a Gaussian weighted average of the weakest and strongest light intensity. Thereafter, a specific numerical Gaussian filter may be applied to the image data to undergo a process of erasing the characteristics of the individual cases of the patient.

The feature map extraction unit 120 obtains a feature map image including feature points of the image through a deep learning-based extraction algorithm.

In this case, the deep learning extraction algorithm may be a combination of auto-encoder-based extraction algorithms and Resnet, Densenet, and Mobilenet-based neural networks used in deep learning algorithms. However, it is not limited thereto. The feature map extraction unit 120 secures a feature map image at a specific layer of the neural network.

As an example, a low-quality image can be improved to a high-quality image by using a Generative Adversarial Network (GAN) together with an auto-encoder-based generator.

Auto-Encoder trains the neural network to minimize the error of the first data and the result of encoding-decoding. It converts the multi-dimensional input data into low-dimensional code and then converts the low-dimensional code into multi-dimensional data that was first input. It is a way to find. This is an unlabeled unsupervised learning method.

The Generative Adversarial Network (GAN) includes two artificial intelligence networks. It includes a generator that generates data and a discriminator that determines whether the data generated by the generator matches the actual one.

That is, a low-quality image can be improved to a high-definition image through a process of generating a fake image from random noise and comparing it with a real image to learn.

In addition, the structure of ResNet or DenseNet, a neural network in the field of image recognition, was created through a unique design philosophy and delicate coordination.

Neural networks ResNet or DenseNet are connected by a 'skip connection' connection method. While the internal connection in the existing network was only connected between adjacent components, the ResNet or DenseNet neural network uses a method that skips several steps as well as the connection between adjacent components to improve performance. Can be improved.

In one embodiment, the feature map is an image in which a location where a neural network is determined to be a characteristic portion of the image has a large value, and a portion where the neural network is determined has a value close to 0, and is displayed in color. Accordingly, the feature map image can be secured from the image in which the rest of the image is removed in black except for the important portion.

The feature map image secured by the feature map extraction unit 120 is stored in the database 140 as a feature map for the category. And it is used to evaluate the degree of consistency with the pre-built database information through a post-processing process.

Image retrieval unit 130 by using the characteristic map image extracted in the feature map extracting unit 120 determines the degree of match between the images.

The image retrieval unit 130 determines that the images are matched by extracting only the feature-activated portion through the feature map extraction unit 120 and comparing the portion using an image processing algorithm such as Otsu Thresholding. have.

For example, the image search unit 130 separates the objects contained in the image in order to obtain information from the image by a binarization technique such as Otsu Thresholding.

Binarization converts pixels to 0 and 1 (or 255) according to the gray value of the image. Therefore, the object contained in the image can be separated from the background through the binarization process. However, at this time, there is a problem of separating the axel from the gray value.

Various algorithms for setting a threshold value for binarizing an image are known, and in one embodiment, a criterion for setting a threshold can be determined by an algorithm of Otsu using a statistical method.

At this time, if the Otsu algorithm is used, the middle value is obtained from the two histograms, so that the threshold can be more accurately determined.

The image retrieval unit 130 is a clustering-based machine learning algorithm that determines the clustering of features and calculates the correspondence as a score. Then, image data having a score greater than or equal to the set score is extracted.

In one embodiment, clustering image data may be clustered by various criteria such as body parts or patient lesions, or sex, age, and imaging techniques. However, the present invention is not limited thereto, and various modifications are possible.

Through this, the image search unit 130 may extract the most similar image by matching the corresponding information in the related database that has been established.

Thereafter, according to an additional aspect, the result output unit may output a group of images similar to the corresponding similar image obtained from the image search unit on the screen.

The image search unit 130 compares the feature map of the image data extracted by the feature map extraction unit 120 with the feature map of the databased image, which is judged to be available for learning, and then similarity score Search for image data with a certain value or more.

At this time, the similarity score, which is a criterion for extraction as a search result, can be arbitrarily set. The similarity score can be set to a high value to increase accuracy, and the similarity score can be set to a low value to increase the amount of data according to the search result even if the accuracy is somewhat lower.

Accordingly, a similarity difference range of image data for a region having sufficient learning data for each body region may be set to a small value. For example, it can be appropriately judged as learning data only when the similarity is 90% or more, and in the case of medical image data of a body part where the amount of learning data is insufficient, the similarity can be adjusted to 80% to secure the amount of learning data.

Deep learning is a technique developed from a methodology called artificial neural network among many machine learning methods. Deep learning is also divided into various techniques. There is a Convolution Neural Network (CNN) specialized in the analysis of image data.

In one embodiment, the convolutional neural network (CNN) is a type of feedforward deep neural network that is trained using a backpropagation process, inspired by the visual processing structure of the cat brain discovered by Hubel and Torsten. It was designed to receive.

The convolutional neural network is designed by introducing three concepts: filtering for the region of the retina, sharing of weights and bias, and pooling. Typically, a convolutional neural network is designed using a convolution layer, a pooling layer, and a fully-connected layer.

For neural network learning, input and output of all neurons from the first hidden layer to the output layer are calculated step-by-step through the feedforward process, and then the error of the final output stage, δL = ∂C / ∂zL, is calculated. Function, z is the input of the output neuron, L is the final neuron layer), and the error of the final output stage is propagated in reverse order from the output stage to the input stage through the reverse propagation process, and the error of each hidden layer is calculated step by step. And estimating the characteristics of each hidden layer, i.e., weights and bias values.

The connection from the input layer to the hidden layer is called a feature map. The artificial intelligence learning medical image data retrieval system according to an embodiment retrieves and provides medical image data having a similarity between the feature maps of a reference value or higher.

The medical image retrieval system based on artificial intelligence technology has an effect of providing the base data for making a more reliable diagnosis in the medical diagnosis field.

For example, in the case of a conventional medical device, the process of inferring the type of outbreak for treatment or obtaining a diagnosis result by viewing various lesion patterns on the head MRI (Magnetic Resonance Imaging) image of a patient with cerebral infarction is a variety of lesion patterns that occur. It was actually impossible.

However, in the case of artificial intelligence-based medical devices, it is possible to classify the type of cerebral infarction based on the training data by learning the cerebral infarction pattern in an artificial neural network using medical big data related to cerebral infarction.

In this case, the image search unit 130 may search and extract the medical image data related to cerebral infarction of another patient similar to the medical image data of the patient and provide a screen.

The medical image search system according to an embodiment extracts and provides similar medical image data based on the feature map of the medical image data of a specific patient to perform diagnosis directly, or provides a useful reference for a doctor to diagnose. The role as a means is all possible.

The database 140 stores various information necessary for system management and operation, from search results to storage of feature map images used to determine similarity. That is, the database 140 stores programs and files necessary for controlling and driving the overall operation of the medical image search system according to an embodiment.

In addition, for artificial intelligence, deep learning methods store previously learned learning image data, and feature maps of the learning image data are stored.

In one embodiment, the database 140 is distributed, classified, and stored in a pre-conventional hierarchical path in a system as an npz or pkl file based on a Python language based on a user's option selection.

It is implemented in a database structure for management of medical image data and feature map information stored in the database 140.

For example, it is possible to match and store tag information, feature maps, and learning medical image files of the medical image data according to the user's option selection.

In one embodiment, the database 140 additionally stores and manages the tag values separated from the medical image data, thereby improving the efficiency of data management and viewing the monitoring effect of the system.

In addition, in one embodiment, the database 140 stores the search results of the image data of which the medical image data and the feature map correspondence are higher than the reference value. The database 140 may further store feature image data information used for similarity determination. However, the present invention is not limited thereto, and is responsible for storing information necessary for managing and driving the system for various purposes.

2 is a flowchart illustrating a method of driving an AI-based medical image search system according to an embodiment of the present invention.

First, in the driving method of the AI-based medical image search system, the data input unit receives medical image data obtained by photographing at least a part of the patient's body (S200).

In the step of receiving data, medical image data obtained by capturing at least a part of a patient's body is received from at least one medical device.

At this time, the medical device is a device that measures a patient's affected area signal, converts it into an electrical signal, and converts the output electrical signal into an image, and includes an ultrasound scanner, an MRI device, and a CT. The output electrical signal may be imaged in a continuous form and input in real time.

In addition, in one aspect, the step of receiving data receives image data converted to a binary file as a message to memory (DMA) using a message queue method. Accordingly, it is possible to design to offload the system hardware.

Then, the pre-processing unit performs pre-processing on the medical image data received through the data input unit (S210).

In one aspect, the pre-processing step is to adjust the resolution of the input image data constantly, analyze the intensity (intensity) value of the image data, and histogram to highlight a specific area based on the histogramized results to increase the intensity. It is characterized by separating a weak part and a strong part to clarify the distribution of the difference between the intensity values of the image data.

In one embodiment, image data input to the data input unit has different resolutions. Therefore, it is necessary to constantly adjust the resolution of the input image data and standardize it.

Then, the intensity value of the image data is analyzed to make a histogram, and based on this, a specific area is emphasized to separate the weak and strong parts. Then, the process of clarifying the distribution of differences in intensity values of image data is performed.

The brightness of light is a noise factor that is a major obstacle in classifying images. This is because when the shadow is generated, the code corresponding to the color of the pixel at the corresponding position changes, that is, it is said to be an image of the same body part. It is an image with completely different pixel values. This technique of minimizing interference by illumination is called intensity normalization.

In one embodiment, the noise may be removed using a Gaussian weighted average of the weakest and strongest light intensity. Thereafter, a gaussian filter of a specific value may be applied to the image data to perform an operation process of extinguishing the characteristics of the individual cases of the patient.

Thereafter, the feature map extracting unit extracts a feature map of the medical image data that has been preprocessed by the preprocessing unit (S220).

In one aspect, the step of extracting the feature map is characterized in that the feature map of the image is extracted by a convolutional neural network (CNN).

The deep learning extraction algorithm may be a combination of auto-encoder based extraction algorithms and Resnet, Densenet, and Mobilenet based neural networks commonly used in deep learning algorithms. The feature map extractor obtains a feature map image at a specific layer of the neural network.

As an example, a low-quality image can be improved to a high-quality image by using a Generative Adversarial Network (GAN) together with an auto-encoder-based generator.

Auto-Encoder trains the neural network to minimize the error of the first data and the result of encoding-decoding. It converts the multi-dimensional input data into low-dimensional code and then converts the low-dimensional code into multi-dimensional data that was first input. It is a way to find. This is an unlabeled unsupervised learning method.

The Generative Adversarial Network (GAN) includes two artificial intelligence networks. It includes a generator that generates data and a discriminator that determines whether the data generated by the generator matches the actual one.

That is, a low-quality image can be improved to a high-definition image through a process of generating a fake image from random noise and comparing it with a real image to learn.

Neural networks ResNet or DenseNet are connected by a 'skip connection' connection method. While the internal connection in the existing network was only connected between adjacent components, the ResNet or DenseNet neural network uses a method that skips several steps as well as the connection between adjacent components to improve performance. Can be improved.

In one embodiment, the feature map is an image in which a location where a neural network is determined to be a characteristic portion of the image has a large value, and a portion where the neural network is determined has a value close to 0, and is displayed in color. Accordingly, the feature map image can be secured from the image in which the rest of the image is removed in black except for the important portion.

Then, the image search unit compares the extracted feature map with the feature maps of pre-stored image data (S230), and searches for image data having a degree of matching with the extracted feature maps among the pre-stored image data (S240). Output (S250).

At this time, the step of outputting the search result is characterized in that the degree of consistency is determined by determining the clustering of the feature maps using a clustering-based machine learning algorithm.

In one embodiment, clustering image data may be clustered by various criteria such as body parts or patient lesions, or sex, age, and imaging techniques. However, it is not limited thereto, and is interpreted to cover various modifications.

Through this, the image search unit can extract the most similar image data by matching the information in the related database that has been established.

The image retrieval unit compares the feature map of the image data extracted from the feature map extractor with the feature map of the databased image that is judged to be available for learning first, and then compares the image data with a similarity score of a certain value or more Search.

At this time, the similarity score, which is a criterion for extraction as a search result, can be arbitrarily set. The similarity score can be set to a high value to increase accuracy, and the similarity score can be set to a low value to increase the amount of data according to the search result even if the accuracy is somewhat lower.

Accordingly, a similarity difference range of image data for a region having sufficient learning data for each body region may be set to a small value. For example, it is appropriately judged as learning data only when the similarity is 90% or more, and in the case of medical image data of a body part where the amount of learning data is insufficient, the similarity can be adjusted to 80% to secure the amount of learning data.

For example, the connection from the input layer to the hidden layer in a convolutional neural network is called a feature map. The artificial intelligence learning medical image data retrieval system according to an embodiment retrieves and provides medical image data having a similarity between the feature maps of a reference value or higher.

The medical image retrieval system based on artificial intelligence technology has an effect of providing the base data for making a more reliable diagnosis in the medical diagnosis field.

For example, in the case of a conventional medical device, the process of inferring the type of outbreak for treatment or obtaining a diagnosis result by viewing various lesion patterns on the head MRI (Magnetic Resonance Imaging) image of a patient with cerebral infarction is a variety of lesion patterns that occur. It was actually impossible.

However, in the case of artificial intelligence-based medical devices, it is possible to classify the type of cerebral infarction based on the training data by learning the cerebral infarction pattern in an artificial neural network using medical big data related to cerebral infarction.

In this case, the image search unit may search and extract medical image data related to cerebral infarction of another patient similar to the medical image data of the patient and provide the screen.

The medical image search system according to an embodiment extracts and provides similar medical image data based on the feature map of the medical image data of a specific patient to perform diagnosis directly, or provides a useful reference for a doctor to diagnose. The role as a means is all possible.

After this, the database stores various information necessary for system management and operation, from the search results to the storage of the feature map images used to determine similarity. That is, the database stores programs and files necessary for controlling and driving the overall operation of the medical image search system according to an embodiment.

In addition, the database stores learning image data previously learned in a deep learning method for artificial intelligence, and stores a feature map of the learning image data.

In one embodiment, the database is distributed, classified, and stored in a pre-conventional hierarchical path in a system as a npz or pkl file based on a Python language based on a user's option selection. The database is implemented in a database structure for management of stored medical image data and image feature map information.

The above-described method may be implemented as an application or in the form of program instructions that can be executed through various computer components to be recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, or the like alone or in combination.

The program instructions recorded on the computer-readable recording medium are specially designed and configured for the present invention, and may be known and usable by those skilled in the computer software field.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs, DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.

Examples of program instructions include not only machine language codes produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the present invention, and vice versa.

Although described above with reference to embodiments, those skilled in the art understand that various modifications and changes can be made to the present invention without departing from the spirit and scope of the present invention as set forth in the claims below. Will be able to.

100: data input unit 110: pre-processing unit
120: feature map extraction unit 130: image search unit
140: database

Claims (10)

A data input unit that receives medical image data captured at least a part of the patient's body;
A pre-processing unit which performs pre-processing on the medical image data received by the data input unit;
A feature map extracting unit for extracting a feature map of the medical image data preprocessed by the preprocessing unit; And
It includes; an image search unit for comparing the extracted feature map with feature maps of pre-stored image data, and searching for image data having a degree of matching with the extracted feature map that is greater than or equal to a reference value among pre-stored image data, and outputting a search result. Characterized by,
The feature map extractor extracts a feature map of the image by a convolutional neural network (CNN),
The image search unit compares the feature map of the image data extracted from the feature map extraction unit with the feature map of the database image determined to be available for learning, and the image data having a similarity score equal to or greater than a certain value Search for, the artificial intelligence-based medical image search system, characterized in that the similarity score can be arbitrarily set as a basis for extraction as a search result.
delete According to claim 1,
The image search unit,
An artificial intelligence-based medical image search system characterized by grasping consistency by determining the clustering of the feature maps using a clustering-based machine learning algorithm.
According to claim 1, The data input unit,
AI-based medical image retrieval system, characterized in that the medical image data converted to a binary file is received by a message queue method.
According to claim 1,
The pre-processing unit,
The resolution of the input image data is constantly adjusted, the intensity value of the image data is analyzed, and a histogram is used to highlight a specific region based on the histogram to separate the weak and strong portions. The artificial intelligence-based medical image retrieval system, characterized in that to clarify the difference distribution of the intensity (intensity) value of the image data.
In the driving method of the artificial intelligence-based medical image search system,
Receiving, by the data input unit, medical image data of at least a part of the patient's body;
A pre-processing unit performing pre-processing on the medical image data received through the data input unit;
A feature map extracting unit extracting a feature map of medical image data pre-processed by the pre-processing unit; And
An image search unit comparing the extracted feature map with feature maps of pre-stored image data, and searching for image data having a degree of matching with the extracted feature map of the pre-stored image data having a reference value or higher and outputting a search result; Including,
The step of extracting the feature map extracts the feature map of the image by a convolutional neural network (CNN),
In the step of outputting the search result, it is determined that the feature map of the image data extracted from the feature map extraction unit is possible for learning, and the similarity score is constant by comparing the feature map of the databaseized image. A method of driving an artificial intelligence-based medical image search system, wherein image similarity scores that are more than numerical values are searched, and the similarity score that is a reference for extraction as a search result can be arbitrarily set.
delete The method of claim 6,
The step of outputting the search result,
A method of driving an artificial intelligence-based medical image retrieval system, characterized by determining the degree of congruence by determining the clustering of the feature maps using a clustering-based machine learning algorithm.
The method of claim 6,
The step of receiving the data,
A method of driving an artificial intelligence-based medical image search system, characterized in that the medical image data converted to a binary file is received by a message queue method.
The method of claim 6,
The pre-processing step,
The resolution of the input image data is constantly adjusted, the intensity value of the image data is analyzed, and a histogram is used to highlight a specific region based on the histogram to separate the weak and strong portions. A method of driving an artificial intelligence-based medical image retrieval system, characterized in that the difference distribution of intensity values of the image data is clarified.

Images