KR102336003B1 - Apparatus and method for increasing learning data using patch matching - Google Patents

Abstract

An apparatus and method for increasing training data using patch matching are disclosed. A method of increasing training data using patch matching includes: a first step of generating a patch image by dividing an image in a medical image by a patch image generator into a size specified at an arbitrary angle; a second step of classifying by the first determining unit using a clustering technique, and determining a valid patch or an invalid patch through a similarity analysis with a preset valid patch; a second step of determining, by a second determination unit, a valid patch or an invalid patch by analyzing a preset valid patch and a similar pattern through a histogram analysis; a step 2c of the third determining unit analyzing the similarity for each patch using the similarity analysis function and the learned artificial neural network and determining whether the patch is a valid patch or an invalid patch; and a third step in which the effectiveness analyzer rotates the medical image at different angles and divides the image by a specified size to generate a patch, and analyzes the effectiveness through one of steps 2a, 2b, and 2c.

Description

Apparatus and method for increasing learning data using patch matching

An embodiment of the present invention relates to an apparatus and method for increasing training data using patch matching.

In order to develop an AI-based image analysis system, it is essential to secure a large amount of image data. In order to overcome this limitation, a method of learning by dividing a large-capacity image into small pieces, that is, an image fragment, that is, a patch, is widely used. Nevertheless, in the case of special image data such as pathological images for a specific cancer, it is difficult to secure a sufficient number of image data. In addition, in the case of such an image, since the size of the repeated pattern to be analyzed is non-uniformly repeated, it is difficult to secure the accuracy even when learning by dividing the patch.

On the other hand, since a medical image has to undergo different image processing according to its type, there is a disadvantage in that a detection or diagnosis apparatus or an operating system thereof is complicated and large. Most of the conventional lesion detection technologies using medical images require the participation of a user such as a doctor or assist the user in diagnosing a lesion based on a medical image, and remain at the level of assisting the user in diagnosing the lesion or automatically diagnosing the lesion or grading the lesion. The analysis still has limitations.

Recently, by introducing machine learning or artificial intelligence in the medical field, active R&D is in progress. However, in relation to a specific lesion, for example, a brain disease such as a stroke, the results of R&D for automatically diagnosing a lesion and analyzing the grade of the lesion are still insignificant.

An object of the present invention is to provide an apparatus and method for automatically dividing a patch effective for learning in a large-capacity image in which a uniformly repeating pattern exists.

Another object of the present invention is to provide a computer-readable recording medium in which the above-described embodiment of the present invention records a program for implementing an apparatus and method for increasing learning data using patch matching.

A method for increasing learning data using patch matching according to an aspect of the present invention for solving the above technical problem includes: a first step of generating a patch image by a patch image generator dividing an image into a size specified at an arbitrary angle in a medical image; a second step of classifying by the first determining unit using a clustering technique, and determining a valid patch or an invalid patch through a similarity analysis with a preset valid patch; a second step of determining, by a second determination unit, a valid patch or an invalid patch by analyzing a preset valid patch and a similar pattern through a histogram analysis; a step 2c of the third determining unit analyzing the similarity for each patch using the similarity analysis function and the learned artificial neural network and determining whether the patch is a valid patch or an invalid patch; and a third step in which the effectiveness analyzer rotates the medical image at different angles and divides the image by a specified size to generate a patch, and analyzes the effectiveness through one of steps 2a, 2b, and 2c.

In an embodiment, the method for increasing learning data using patch matching may further include, after the third step, terminating the patch generation when the basic operation end unit rotates the medical image to 360 degrees.

In one embodiment, the method for increasing training data using patch matching includes, after the terminating step, further classifying the patches for which a classification unit has been validated in the validity analysis using a clustering technique; and generating, by the first image generator, a plurality of images in the same cluster according to the additional classification to generate one image.

In one embodiment, in the method of increasing training data using patch matching, after generating the single image, in the generating step, a second image generating unit selects the image generated by the first image generating unit more than one image. The method may further include generating an image of a new size by cropping it to a smaller size.

In an embodiment, the method may further include the step of re-determining validity by analyzing, by the validity analysis unit, a ratio in which the image of the new size overlaps with a lesion area of an existing effective patch image.

In one embodiment, the data augmentation unit generates a patch while rotating the image generated in the new size from an arbitrary angle to 360 degrees, and analyzing the validity in any one of the steps 2a to 2c to augment the learning data. may further include.

According to another aspect of the present invention, there is provided an apparatus for increasing training data using patch matching, comprising: a patch image generator configured to generate a patch image by dividing an image into a size specified at an arbitrary angle in a medical image; a first judging unit that classifies through a clustering technique and determines whether the patch is valid or invalid through a similarity analysis with a preset valid patch; a second determination unit that analyzes a preset valid patch and a similar pattern through histogram analysis to determine a valid patch or an invalid patch; a third judging unit that analyzes the similarity for each patch using the similarity analysis function and the learned artificial neural network and determines whether the patch is valid or invalid; and a validity analysis unit that rotates the medical image at different angles, divides the image into a specified size, generates a patch, and analyzes effectiveness through one of the first to third determination units.

In one embodiment, the apparatus for increasing learning data using patch matching further includes a basic operation terminating unit for terminating patch generation when the rotational angle of the medical image becomes 360 degrees after the validity analysis of the validity analysis unit.

In one embodiment, the apparatus for increasing training data using patch matching includes: a classification unit for further classifying patches that have been validated in the validity analysis by a clustering technique after the end operation of the basic task end unit; and a first image generator configured to generate one image by connecting a plurality of images in the same cluster according to additional classification.

In an embodiment, the apparatus for increasing training data using patch matching generates a second image for generating an image of a new size by cropping one image generated by the first image generating unit to a size smaller than the single image It may include more wealth.

In an embodiment, the validity analysis unit may re-determine validity by analyzing a ratio in which the image of the new size overlaps with a lesion area of an existing effective patch image.

In one embodiment, the image generated in the new size is rotated from an arbitrary angle to 360 degrees while generating a patch, and further comprising a data augmentation unit for augmenting the learning data by analyzing the validity with any one of the 2a to 2c above. do.

An apparatus according to another aspect of the present invention for solving the above technical problem includes a computer-readable recording medium recording a program for implementing the method for increasing learning data using the patch matching of any one of the above embodiments.

In the case of using the apparatus and method for increasing learning data using the patch matching described above, it is possible to automatically divide an effective patch for learning in a large-capacity image in which a non-uniformly repeating pattern exists, and to increase the effective patch, It has the advantage of effectively processing medical images by increasing the training data.

1 is a flowchart of a method of increasing training data using patch matching according to an embodiment of the present invention.
FIG. 2 is a view for explaining an additional conversion process in the method of FIG. 1 .
FIG. 3 is a diagram illustrating a process of obtaining a patch in a specific ROI of a pathological image that can be employed in a process of increasing training data in the method of FIG. 1 .
4 is a structural diagram of a convolutional neural network employable in the method and apparatus of the present embodiment.
5 is a block diagram of an apparatus for increasing training data using patch matching according to another embodiment of the present invention.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in this specification are only exemplified for the purpose of explaining the embodiments according to the concept of the present invention, and the embodiments according to the concept of the present invention are It may be implemented in various forms and is not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention may have various changes and may have various forms, the embodiments will be illustrated in the drawings and described in detail herein. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes all changes, equivalents, or substitutes included in the spirit and scope of the present invention.

The terms used herein are used only to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In this specification, terms such as "comprises" or "having" are intended to designate that the features, numbers, steps, operations, components, parts, or combinations thereof described herein exist, but one or more other features. It should be understood that it does not preclude the possibility of the presence or addition of numbers, steps, operations, components, parts, or combinations thereof.

The terms used in this specification are as follows.

T2-weighted imaging: A technique obtained from magnetic resonance imaging (MRI) in a specific pulse sequence or an image obtained by this technique. The image is mainly of internal tissues of the body. Provides structural information.

FLAIR (Fluid attenuated inversion recovery): A signal acquisition technique using a magnetic imaging device that weakens the signal of the cerebrospinal fluid with a long inversion time and echo time, making it easier to detect lesions that are easy to miss on T2-weighted images or images acquired with this technique refers to FLAIR may be referred to as fluid damping reversal recovery.

DWI (Diffusion weighted imaging): mainly refers to diffusion-weighted images obtained from magnetic resonance imaging, and provides information on the extent and whether water molecules in cell tissues diffuse in a specific direction.

Perfusion weighted imaging (PWI): Refers to a perfusion-weighted image (simply, perfusion image) obtained from magnetic resonance imaging, and informs the change in concentration of the injected contrast agent over time.

Penumbra: A semi-shaded area in an image caused by an ischemic event or embolism. It refers to an area that is viable when hypoxic cell death occurs due to a local decrease in oxygen transport function, or when appropriate treatment within several hours.

ADC (Apparent Diffusion Coefficient): As the apparent diffusion coefficient obtained from magnetic resonance imaging, it provides information on diffusion-interfering factors in tissues inside the human body.

Arterial phase (AP): A period of specific perfusion obtained from magnetic resonance imaging, indicating the time when the injected contrast agent passes through the arterial part.

Capillary phase (CP): This is the period of specific perfusion obtained from magnetic resonance imaging, and represents the period when the injected contrast agent passes through the capillary section over time.

Venous phase (VP): This is the period of specific perfusion obtained from magnetic resonance imaging, and represents the time when the injected contrast agent passes through the venous section over time.

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

1 is a block diagram of an apparatus for increasing training data using patch matching according to an embodiment of the present invention. FIG. 2 is a diagram for explaining an additional conversion process in the method of FIG. 1 . And FIG. 3 is a diagram illustrating a process of obtaining a patch from a specific ROI of a pathological image that can be employed in a process of increasing training data in the apparatus of FIG. 1 .

Referring to FIG. 1 , in the method for increasing training data using patch matching according to the present embodiment (hereinafter, simply referred to as 'method of increasing training data'), patches are generated in multiples of an arbitrarily designated size in a medical image, and a multiple of an arbitrarily designated size. It includes a step (first step) of generating a patch image for terminating the patch generation when the angle becomes 360 degrees (S10, S12, S14).

Next, the method of increasing learning data includes a step of classifying through a clustering technique and determining through a preset valid patch and similarity analysis (step 2a) (S21, S31, S32).

Step 2a will be described in detail as follows.

The designated effective patch is divided into K groups according to the number of classes of lesions. Here, K is the number of types of lesions in the medical image. For example, K indicates that the number of classes of lesions is 4 if the annotation of the medical image is set to end-stage cancer, middle-stage cancer, early cancer, and normal.

And, the distance between two adjacent patches is defined as follows [Equation 1].

In the above [Equation 1], A(i, j, k) means a pixel vector value of the three-dimensional A patch. m denotes the width of the patch, and n denotes the height of the patch.

For example, assuming there is a patch A, if there are at least s points within a certain distance (e) from the patch A, the patch is called a core patch.

로 고려되는 집합의 범위를 고정한다. E와 S는 K개의 그룹과 겹쳐지지 않도록 사용자가 미리 결정하여 설정할 수 있다.here,

Fixes the scope of the set to be considered as E and S can be set by the user in advance so as not to overlap with K groups.

A patch A^ having a maximum number of patches within a predetermined distance from patch A may be obtained by changing a predetermined distance e according to each group within the K groups. The maximum number of patches (A^) is expressed as [Equation 2] below.

이하이면, 유효한 패치로 판정할 수 있다. 유효성 분석에서 유사도 분석 대상이 되는 패치들 사이의 거리가

이하가 아니면 유효하지 않은 패치(비유효패치)로 판정할 수 있다.Based on the core patch found for each K group in the validity analysis, the distance between the patches to be analyzed for similarity is

If it is below, it can be determined as a valid patch. In the validity analysis, the distance between the patches to be analyzed for similarity

If it is not below, it can be determined as an invalid patch (invalid patch).

In addition, the method of increasing the learning data of the present embodiment may include a step (step 2b) of analyzing and determining a preset effective patch and similar pattern through a histogram analysis after the first step (S22, S31, S32). Step 2b may be performed in parallel with step 2a.

Step 2b will be described in detail as follows.

에 대해 mean(A(i,j,k))가 유효한 패치들의 (max(B(i,j,k))의 최솟값에 대해 기준 설정 배율보다 적으면 유효하다고 판정한다. 기준 설정 배율은 0.9배일 수 있다. 이러한 기준 설정 배율은 실험에 의해 최적화된 것이다.fixed for RGB channels

It is judged as valid if mean(A(i,j,k)) is less than the reference-setting magnification for the minimum value of (max(B(i,j,k)) of valid patches. The reference-setting magnification is 0.9 times. This reference setting magnification is optimized by experimentation.

If step 2b is expressed as an equation, it can be expressed as in the following [Equation 3].

In addition, the method of increasing the learning data of the present embodiment may include a step of analyzing and determining the degree of similarity for each patch (step 2c) using a similarity analysis function after the first step (S23, S31, S32). Step 2c ( S23 ) may be performed in parallel with step 2a ( S21 ) and step 2b ( S22 ).

Step 2c will be described in detail as follows.

의 차원 3mn의 정규분포를

라고 하자. 그리고, 유효하지 않는 패치의 그룹에 대한 차원 3mn의 정규 분포를

라고 하자.Valid patches for each K group

Normal distribution of dimension 3mn of

let's say Then, we obtain a normal distribution of dimension 3mn for the group of invalid patches.

let's say

라고 하고, Z에 들어갈 확률을

라고 하면

이 된다.The probability that the pixel data of a given patch will fit into group K is

, and the probability of entering Z is

If you say

becomes this

At this time, the following [Equation 4] is solved for each K group by applying Bayes' rule.

의 K+1개의 방정식을 연립해서 풀어서 각

값을 구한 뒤 가장 높은 값으로 대상 패치의 유효그룹을 결정한다.

가 가장 클 경우 대상패치는 유효하지 않은 패치로 분류되고,

가 가장 클 경우 K 그룹중 I번째 그룹에 속하게 된다.In [Equation 4] above

By solving a system of K+1 equations of

After finding the value, the effective group of the target patch is determined with the highest value.

is the largest, the target patch is classified as an invalid patch,

If is the largest, it belongs to the I-th group among K groups.

In addition, the method of increasing the learning data of this embodiment can analyze the effectiveness through all of the second step 2a, step 2b and step 2c (S21, S22, S23, S31, S32).

In addition, in the method of increasing the learning data of this embodiment, after steps 2a, 2b, and 2c, patches that have not been validated (invalid patches) are added to the normal distribution for the RGB channel of the valid pathological image. It may include the step of transforming and generating (S33) (refer to FIG. 2). Through this, the increased number of valid patches can be classified (S40).

In addition, the method of increasing the training data of the present embodiment may re-analyze the validity of the additionally transformed and generated patches through all of the steps 2a, 2b, and 2c.

The process of processing the additional transformation and generation steps using the normal distribution of effective patches is as follows.

First, patches that do not receive valid validation are normalized.

Then, using a convolution operator, the patch that has received a valid definition is converted into a normal distribution for the RGB channel. This is expressed as [Equation 5].

In [Equation 5] above, N nonvalid means a Gaussian distribution for the RGB channel of a patch that has not been validated. Since there are 3 channels, R, G, and B, a total of 3 N valid means the normal distribution for the RGB channels of the valid patch.

The convolution operator means the following Equation (6).

According to [Equation 5] and [Equation 6] above, the Gaussian distribution of the patch not receiving the valid patch by the convolution operator is converted into a normal distribution for the RGB channel of the valid patch. Therefore, in this embodiment, an invalid patch is converted into a valid patch using this.

In addition, the method of increasing the learning data of this embodiment generates an image of a new size by generating a single image by using the valid patches obtained through the above process, and then cutting the generated image at a predetermined smaller magnification than the single image. It may further include the step of

In addition, the method of increasing the learning data of this embodiment may further include the step of re-determining validity by generating an image of a new size, and then analyzing a ratio in which the image of the new size overlaps with the lesion area of an existing effective patch image. .

In addition, the method of increasing the learning data of this embodiment generates a patch while rotating the image created in the new size by 360 degrees at an arbitrary angle after generating an image of a new size, and any one of the steps 2a to 2c above. It may further include the step of augmenting the learning data by analyzing the validity of one.

4 is a structural diagram of a convolutional neural network that can be employed in the artificial neural network learning method and apparatus using the operation of the present embodiment.

The artificial neural network learning apparatus using the operation according to the present embodiment may use a deep learning architecture as an artificial intelligence neural network. Deep learning architecture uses convolutional neural network (CNN), deconvolution, skip connection, etc. for medical image signals to learn motions according to motion-based labeling, and according to the learning results, It may be implemented to process a series of processes that generate an external signal or generate a specific noise.

As an example, the deep learning architecture may have a form including a convolutional network, a deconvolutional network, and a shortcut. As shown in Fig. 3, the deep learning architecture stacks a 3x3 color convolution layer and an activation layer (ReLU) in order to extract local features of a medical image (X), and strides a 2x2 size filter. By applying (stride) 1, the operation of the convolution block connected to the next lower depth level is repeated 4 times, and then a 2x2 deconvolution layer and an activation layer (ReLU) are applied. After connecting to the next higher depth level, the operation of the inverse convolution block stacking the 3x3 size color convolution layer and the activation layer is repeated 4 times, where each level of the convolution network including the operation of the convolution block of each level It can be made to copy and contate the image of the convolution block of the level with the convolution result of the corresponding level of the inverse convolution network of the same level, and perform a convolution operation in each block.

A convolutional block in a convolutional network and a deconvolutional network can be implemented as a combination of conv-ReLU-conv layers. And, the output of the deep learning architecture may be obtained through a classifier connected to a convolutional network or a deconvolutional network, but is not limited thereto. The classifier can be used to extract local features from an image using a fully connectivity network (FCN) technique.

In addition, the deep learning architecture may be implemented to additionally use an inception module or a multi filter pathway in the convolution block according to implementation. Different filters in the inception module or multi-filter path may include a 1x1 filter.

For reference, when the input image has 32 horizontal, 32 vertical, and RGB channels in the deep learning architecture, the size of the input image X corresponding to the medical image may be [32x32x3]. In a deep learning architecture's convolutional neural network (CNN), a convolutional (CONV) layer is connected to some region of the input image, and it can be designed to compute the dot product of the connected region and its weight. .

Here, the rectified linear unit (ReLU) layer is an activation function applied to each element, such as max(0,x). The ReLU layer does not change the size of the volume. The POOLING layer may output a reduced volume by performing downsampling or subsampling on a dimension expressed by (horizontal, vertical).

In addition, the fully-connected (FC) layer may output a volume having a size of, for example, [1x1x10] by calculating class scores. In this case, ten numbers correspond to class scores for ten categories. The preconnect layer connects all elements of the previous volume. There, some layers may have parameters and some layers may not. CONV/FC layers may include not only input volume but also weight and bias as an activation function. Meanwhile, the ReLU/POOLING layers are fixed functions, and the parameters of the CONV/FC layer may be learned by gradient descent so that the class score for each image is the same as the label of the corresponding image.

5 is a block diagram of an apparatus for increasing training data using patch matching according to another embodiment of the present invention.

Referring to FIG. 5 , the apparatus for increasing learning data using patch matching according to the present embodiment includes a memory 52 for storing a program, and a controller 51 connected to the memory 52 to execute a program. In addition, it may further include a communication unit 53 connected to the network for transmitting and receiving signals and data such as medical image signals.

A program or software module may be loaded in the memory 52 . The software module includes a patch image generation unit 521 , a first determination unit 522 , a second determination unit 523 , a third determination unit 524 , a validity analysis unit 525 , and a basic operation end unit 526 . , a classification unit 527 , a first image generation unit 528a , a second image generation unit 528b , and a data augmentation unit 529 may be included.

The patch image generator 521 generates a patch image by dividing an image into a size specified at an arbitrary angle in the medical image.

The first determination unit 522 classifies through a clustering technique, and determines through a similarity analysis with a preset valid patch.

The second determination unit 523 determines by analyzing a preset effective patch and a similar pattern through histogram analysis.

The third determining unit 524 analyzes and determines the degree of similarity for each patch using the similarity analysis function and the learned artificial neural network.

The effectiveness analyzer 525 rotates the medical image at different angles, divides the image into a specified size, generates a patch, and analyzes effectiveness through one of the first to third determining units.

The basic operation end unit 526 ends the patch generation when the rotated angle of the medical image becomes 360 degrees after the validity analysis of the validity analysis unit.

The classification unit 527 further classifies the patches that have been validated in the validity analysis by the clustering technique after the end operation of the basic work end unit.

The first image generator 528a generates one image by connecting a plurality of images in the same cluster according to additional classification.

The second image generating unit 528b crops one image generated by the image generating unit to a size smaller than the single image to generate an image of a new size.

The data augmentation unit 529 generates a patch while rotating the image generated in a new size from an arbitrary angle to 360 degrees, and analyzes the validity in any one of steps 2a to 2c above to augment the learning data.

The above-described first to third determination units, validity analysis unit, classification unit, etc. may be configured to learn by an artificial neural network. The artificial neural network learning apparatus including such a learning unit may be mounted on a computing device or implemented to include the computing device.

The control unit 51 may correspond to a microprocessor, an arithmetic processing unit or a control unit, may include a central processing unit (CPU) or a core, and execute a program or software module stored in the memory 52 to It may be implemented to execute the method of the present embodiment. For example, as shown in FIG. 3 , a patch may be obtained from a specific region of interest in a pathological image and data may be augmented by analyzing the patch.

The communication unit 53 may include at least one sub communication system for an intranet, the Internet, a wired network, a wireless network, a satellite network, or a combination thereof. The communication unit 53 may be connected to an MRI apparatus or a database to receive or read a medical image signal, and may be implemented to transmit a processed result to a display device or an external device.

The artificial intelligence employable in the above-described device may have a form connected to the classification system as a separate preprocessor, recognizer, learner, or a combination thereof as an independent functional unit or component unit, or may have a form mounted on each component unit. . Artificial intelligence can be designed to include machine learning, artificial neural networks, or deep learning. Deep learning may include convolutional neural networks and the like. Such artificial intelligence or an artificial intelligence system including the same may be referred to as a learning means or the like.

Although the present invention has been described with reference to the embodiments with reference to the drawings, these are merely exemplary, and it is apparent that various modifications and equivalent other embodiments are possible therefrom by those skilled in the art. Accordingly, the true technical protection scope of the present invention should be defined by the technical spirit of the appended claims.

Claims (12)

A first step of generating a patch image by dividing the image by a patch image generator into a size specified at an arbitrary angle in the medical image;
a second step of classifying by the first determining unit using a clustering technique and determining whether the patch is valid or invalid through a similarity analysis with a preset valid patch;
a second step of determining, by a second determination unit, a valid patch or an invalid patch by analyzing a preset valid patch and a similar pattern through a histogram analysis;
2c step of analyzing the similarity for each patch by using the similarity analysis function and the learned artificial neural network by the third determining unit and determining whether the patch is valid or invalid; and
Including a third step of generating a patch by rotating the medical image by a validity analysis unit at a different angle and dividing the image by a specified size, and analyzing the effectiveness through one of the steps 2a, 2b, and 2c, patch matching A method of increasing training data using The method according to claim 1,
terminating the patch generation when the rotation angle of the medical image by the basic operation end unit reaches 360 degrees after the third step;
after the terminating step, further classifying the patches for which the classification unit has been validated in the validity analysis using a clustering technique; and
The method of increasing learning data using patch matching further comprising the step of the first image generator generating one image by connecting a plurality of images in the same cluster according to the additional classification. 3. The method according to claim 2,
After generating the single image,
The method of increasing training data using patch matching further comprising: a second image generating unit cropping one image generated by the first image generating unit to a size smaller than the single image to generate an image of a new size. 4. The method according to claim 3,
The method of increasing learning data using patch matching further comprising the step of re-determining validity by analyzing, by the validity analysis unit, the overlapping ratio of the image of the new size with the lesion area of the existing effective patch image. 4. The method according to claim 3,
Further comprising the step of augmenting the learning data by generating a patch while rotating the image generated in the new size by the data augmentation unit from an arbitrary angle to 360 degrees and analyzing the validity in any one of the steps 2a to 2c, A method of increasing training data using patch matching. The method according to claim 1,
generating, by the validity analysis unit, additionally converting the invalid patch obtained from the validity analysis into a normal distribution for the RGB channel of the pathological image; and
Patch matching further comprising the step of re-analyzing the validity of the patches generated by the additional conversion by the validity analysis unit through all of the steps 2a to 2c to convert the invalid patch into a valid patch A method of increasing training data using a patch image generator for generating a patch image by dividing an image into a size specified at an arbitrary angle in a medical image;
a first judging unit that classifies through a clustering technique and determines whether the patch is valid or invalid through a similarity analysis with a preset valid patch;
a second determination unit that analyzes a preset valid patch and a similar pattern through histogram analysis to determine a valid patch or an invalid patch;
a third determination unit that analyzes the similarity for each patch using the similarity analysis function and the learned artificial neural network and determines whether the patch is valid or invalid; and
An apparatus for increasing learning data using patch matching, including a validity analyzer that rotates the medical image at different angles and divides the image into a specified size to generate a patch and analyzes the effectiveness through one of the first to third determination units . 8. The method of claim 7,
After the validity analysis of the validity analysis unit, the apparatus for increasing the learning data using patch matching further comprises a basic operation terminating unit for terminating the patch generation when the rotated angle of the medical image becomes 360 degrees. 9. The method of claim 8,
a classification unit for further classifying patches that have been validated in the validity analysis by a clustering technique after the end operation of the basic work end unit; and
The apparatus for increasing learning data using patch matching, further comprising a first image generator configured to generate one image by connecting a plurality of images in the same cluster according to additional classification. 10. The method of claim 9,
The apparatus for increasing learning data using patch matching, further comprising a second image generator configured to generate an image of a new size by cropping one image generated by the first image generator to a size smaller than the one image. 11. The method of claim 10,
The effectiveness analyzer analyzes the ratio in which the image of the new size overlaps with the lesion area of the existing valid patch image to determine the validity again, an apparatus for increasing learning data using patch matching. 11. The method of claim 10,
Further comprising a data augmentation unit that generates a patch while rotating the image generated in the new size from an arbitrary angle to 360 degrees and analyzes the validity with any one of the first, second, and third determination units to augment the learning data, Training data augmentation device using patch matching.

Images