KR102208087B1 - Apparatus and method for Auto classification of skin disease - Google Patents

Abstract

Skin diseases such as melanoma are high based on the pixel characteristics of images of skin diseases, the GLCM characteristics and GLRLM characteristics extracted from the gray level co-occurrence matrix (GLCM) and the gray level continuous length matrix (GLRLM) converted from the image. Disclosed are an apparatus, a method, and a recording medium for automatic classification of skin diseases capable of automatic classification with accuracy. According to an exemplary embodiment of the present invention, a method for automatically classifying skin diseases includes: calculating a histogram of brightness values of pixels in an image of a skin lesion, and statistically analyzing the histogram to extract pixel features; Converting the pixel values of the image into a gray-level co-occurrence matrix (GLCM) representing proximity between the pixel values; Extracting GLCM features based on pixel values of the gray level co-occurrence matrix; Converting the pixel values of the image into a gray level run length matrix (GLRLM) representing a continuous length of the pixel values; Extracting a GLRLM feature based on the pixel values of the gray level continuous length matrix; And classifying a skin disease by applying the pixel feature, the GLCM feature, and the GLRLM feature to a machine learning model.

Description

Apparatus and method for Auto classification of skin disease {Apparatus and method for Auto classification of skin disease}

The present invention relates to an apparatus and method for automatic classification of skin diseases, and more particularly, from pixel characteristics of an image photographing skin diseases, a gray level co-occurrence matrix (GLCM) and a gray level continuous length matrix (GLRLM) converted from the image. The present invention relates to a skin disease automatic classification device and a skin disease automatic classification method capable of automatically classifying skin diseases such as melanoma with high accuracy based on the extracted GLCM features and GLRLM features.

Due to the recent change in living environment and the development of medical technology, the lifespan of humans is gradually extending, and for this reason, silver medical technology is in the spotlight. Therefore, early response to diseases caused by decreased immunity has become important in an aging society. In particular, early detection of senile skin diseases is difficult, and if the treatment timing is missed, it may progress to a malignant disease, making treatment difficult. Melanoma (melanoma) is one of the most common diseases of such senile skin diseases, and has a modality similar to that of nevus in the early stage. For this reason, melanoma is difficult to detect early on and is often mistaken for nevus. Melanoma is a disease in which a malignant tumor spreads from the dermal layer of the skin to the muscles, and unlike nevus, symptoms such as vascular malformations, melanin pigment deposition and malformed forms appear in the affected area. Malignant melanoma with the highest degree of malignancy has no subjective symptoms such as itching or pain, and is characterized as a normal black spot.

Abbasi et al. developed ABCD (Asymmetry, Border, Color, Diameter) method, which allows doctors to detect melanoma based on clinical experience with the naked eye. However, this method is a method of visually discriminating the shape of a lesion area, and because subjective judgment of a specialist is involved, there is a disadvantage that even the same image can be diagnosed differently for each specialist. Bowling et al. used a method to reduce possible errors by observing the ABCD method, which was confirmed with the naked eye, in more detail using a dermoscopy. However, the conventional dermatoscope is a method of attaching an optical lens of a microscope to a general camera, and the disease area can be enlarged and viewed, but there is a problem of using the conventional method of reading with the naked eye. Ham et al. attempted to automatically classify melanoma and nevus using SVM (support vector machine) in dermatoscope images, but it was difficult to accurately detect the skin condition, hair, and light source that occur for each patient .

The present invention relates to skin such as melanoma based on the pixel features of an image photographing skin diseases, the GLCM features and GLRLM features extracted from the gray level co-occurrence matrix (GLCM) and the gray level continuous length matrix (GLRLM) converted from the image. It is to provide an apparatus and method for automatic classification of skin diseases, and a recording medium capable of automatically classifying diseases with high accuracy.

The problem to be solved by the present invention is not limited to the problems mentioned above. Other technical problems that are not mentioned will be clearly understood by those of ordinary skill in the art from the following description.

According to an aspect of the present invention, a method for automatically classifying skin diseases includes: calculating a histogram of brightness values of pixels in an image of a skin lesion, and statistically analyzing the histogram to extract pixel features; Converting the pixel values of the image into a gray-level co-occurrence matrix (GLCM) representing proximity between the pixel values; Extracting GLCM features based on pixel values of the gray level co-occurrence matrix; Converting the pixel values of the image into a gray level run length matrix (GLRLM) representing a continuous length of the pixel values; Extracting a GLRLM feature based on the pixel values of the gray level continuous length matrix; And classifying a skin disease by applying the pixel feature, the GLCM feature, and the GLRLM feature to a machine learning model.

The pixel characteristics may include an average of pixel values of the image, a standard deviation, a skew, a curtosis, an entropy, and a root mean square.

The GLCM features are auto-correlation, contrast, correlation, dissimilarity, energy, and homogeneity of pixel values of the gray level co-occurrence matrix. It may include.

The GLRLM feature is SRE (Short Run Emphasis), LRE (Long Run Emphasis), GLNU (Gray Level Non-Uniformity), RP (Run Percentage), RLNU (Run Percentage) extracted based on the pixel values of the gray level matrix. Length Non-Uniformity) and High Gray Level Run Emphasis (HGLRE).

The machine learning model may include a support vector machine (SVM) model.

A method for automatically classifying skin diseases according to an embodiment of the present invention includes: obtaining a dermatoscope image by photographing the skin lesion using a dermoscopy; Converting the dermatoscope image to a binary image, and detecting an object having a size smaller than that of the skin lesion in the binary image; Removing noise including hair and dead skin cells based on the difference in size between the skin lesion and the object; And converting a pixel value of normal skin to 0 in the dermatoscope image to extract a region of interest including only skin lesions.

In the method for automatically classifying skin diseases according to an embodiment of the present invention, based on learning image data including melanoma images and nevus images, the machine learning model is learned by learning the pixel characteristics, the GLCM characteristics, and the GLRLM characteristics. It may further include the step of generating. Classifying the skin disease may include classifying the skin lesion into either melanoma or nevus based on the learned machine learning model.

According to another aspect of the present invention, there is provided a computer-readable recording medium in which a program for executing the method for automatically classifying skin diseases is recorded.

According to another aspect of the present invention, in an apparatus for automatically classifying skin diseases including at least one processor, the at least one processor calculates a histogram of brightness values of pixels in an image photographing a skin lesion, and the histogram To extract pixel features by statistical analysis; Converting the pixel values of the image into a gray-level co-occurrence matrix (GLCM) representing the proximity between the pixel values; Extracting GLCM features based on pixel values of the gray level co-occurrence matrix; Converting pixel values of the image into a gray level run length matrix (GLRLM) representing a continuous length of the pixel values; Extracting GLRLM features based on the pixel values of the gray level continuous length matrix; And applying the pixel feature, the GLCM feature, and the GLRLM feature to a machine learning model to classify skin diseases.

The at least one processor converts a dermatoscope image photographed of the skin lesion into a binary image by dermoscopy; Detecting an object having a size smaller than that of the skin lesion in the binary image; Removing noise including hair and dead skin cells based on the difference in size between the skin lesion and the object; In addition, it may be configured to convert a pixel value of normal skin to 0 in the dermatoscope image to extract a region of interest including only skin lesions.

The at least one processor generates the machine learning model by learning the pixel feature, the GLCM feature, and the GLRLM feature using training image data including melanoma images and nevus images; In addition, it may be configured to classify the skin lesion as either melanoma or nevus based on the machine learning model learned using the training image data.

According to an embodiment of the present invention, based on the pixel features of an image photographing skin diseases, the GLCM features and GLRLM features extracted from the gray level co-occurrence matrix (GLCM) and the gray level continuous length matrix (GLRLM) converted from the image A skin disease automatic classification apparatus, method, and recording medium capable of automatically classifying skin diseases such as melanoma with high accuracy are provided.

The effect of the present invention is not limited to the above-described effects. Effects not mentioned will be clearly understood by those of ordinary skill in the art from the present specification and the accompanying drawings.

1 is a schematic flowchart of a method for automatically classifying skin diseases according to an embodiment of the present invention.
2 is a block diagram of an apparatus for automatically classifying skin diseases according to an embodiment of the present invention.
3 is a flowchart of a method for automatically classifying skin diseases according to an embodiment of the present invention.
4 is a detailed flowchart of step S30 of FIG. 3.
5 is an exemplary diagram of a size distribution diagram of objects extracted from a dermatoscope image.
6 is an exemplary diagram of a process of pre-processing a dermatoscope image according to an embodiment of the present invention.
7 is a block diagram of a skin disease classification unit constituting an automatic skin disease classification apparatus according to an embodiment of the present invention.
8 is an exemplary diagram of a gray level co-occurrence matrix (GLCM) converted from an image according to an embodiment of the present invention.
9 is an exemplary diagram of a gray level continuous length matrix (GLRLM) transformed from an image according to an embodiment of the present invention.
10 is an exemplary diagram of melanoma images (a) and nevus images (b) used for machine learning according to an embodiment of the present invention.
11 is a diagram showing accuracy performance of skin disease classification in a method for automatically classifying skin diseases according to an embodiment of the present invention.

Other advantages and features of the present invention, and a method of achieving them will become apparent with reference to embodiments to be described later in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, and the present invention is only defined by the scope of the claims. Even if not defined, all terms (including technical or scientific terms) used herein have the same meaning as commonly accepted by universal technology in the prior art to which this invention belongs. General descriptions of known configurations may be omitted so as not to obscure the subject matter of the present invention. In the drawings of the present invention, the same reference numerals are used as much as possible for the same or corresponding configurations. In order to help the understanding of the present invention, some configurations in the drawings may be somewhat exaggerated or reduced.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise", "have" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or a combination thereof described in the specification. It is to be understood that the possibility of the presence or addition of other features, numbers, steps, actions, components, parts, or combinations thereof, or any further features, is not excluded in advance.

The'~ unit' used throughout this specification is a unit that processes at least one function or operation, and may mean, for example, a hardware component such as software, FPGA, or ASIC. However,'~ part' is not limited to software or hardware. The'~ unit' may be configured to be in an addressable storage medium, or may be configured to reproduce one or more processors.

As an example,'~ unit' refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, procedures, and subs. Routines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The components and functions provided by the'~ unit' may be performed separately by a plurality of elements and the'~ units', or may be integrated with other additional elements.

In the automatic classification method for skin diseases according to an embodiment of the present invention, features of skin diseases are extracted by image processing, and skin diseases such as melanoma are objectively and automatically classified through an algorithm of a machine learning technique based on the extracted features. can do. 1 is a schematic flowchart of a method for automatically classifying skin diseases according to an embodiment of the present invention. Referring to FIG. 1, first, a skin lesion is photographed by a dermoscopy to obtain a dermoscopy image (S1), and after preprocessing the dermatoscope image (S2), a region of interest (ROI) is Extract (S3). Then, image features related to skin diseases such as melanoma are extracted from the region of interest of the image (S4 to S5).

Image features include pixel features extracted by first-order statistical analysis using information of the pixel itself, gray-level co-occurrence matrix (GLCM) and gray level run length matrix (GLLM) converted from image to matrix form. It may include GLCM features and GLRLM features extracted from. The extracted features (pixel features, GLCM features, and GLRLM features) may be applied to a machine learning discrimination algorithm such as a support vector machine and used to classify skin diseases. According to an embodiment of the present invention, in particular, it is possible to accurately and automatically classify nevus and melanoma.

2 is a block diagram of an apparatus for automatically classifying skin diseases according to an embodiment of the present invention. Referring to FIG. 2, the apparatus 100 for automatically classifying skin diseases according to an embodiment of the present invention includes a control unit 110, a learning unit 120, an image acquisition unit 130, an image preprocessor 140, and a skin disease classification. A unit 150 and a storage unit 160 may be included. The control unit 110 includes at least one processor, and controls the learning unit 120, the image acquisition unit 130, the image preprocessing unit 140, the skin disease classification unit 150, and the storage unit 160 to Execute a function (program) for automatic disease classification.

3 is a flowchart of a method for automatically classifying skin diseases according to an embodiment of the present invention. 2 and 3, the learning unit 120 may learn features of a skin disease by using learning image data including melanoma images and nevus images (S10). In an embodiment, the learning unit 120 may extract features appearing in melanoma by using an image of an authorized skin disease database and generate a machine learning model for classifying melanoma and nevus. The learning unit 120 may calculate a histogram from pixel values of an image of the skin disease database, and statistically analyze the histogram to extract pixel features. In addition, the learning unit 120 may convert the image of the skin disease database into GLCM and GLRLM, respectively, and then extract GLCM features and GLRLM features from the GLCM and GLRLM. The learning unit 120 may generate a machine learning model by learning the extracted pixel features, GLCM features, and GLRLM features. The machine learning model generated by the learning unit 120 may be stored in the storage unit 160 to be used for classification of skin diseases in the future.

The image acquisition unit 130 may acquire an image by photographing a skin lesion (S20). In an embodiment, the image acquisition unit 130 may include a dermoscopy for capturing a skin lesion suspected of melanoma to obtain a dermoscopy image. The image acquired by the image acquisition unit 130 may be stored in the storage unit 160.

For dermatoscope images, the focus of the image and the location of the disease may appear differently within the image for each specialist. In addition, various noises excluding diseases are generated in the image due to the patient's skin condition and hair. Such noise may cause an error in extracting features from an image. In order to prevent this problem, the image preprocessor 140 converts the dermatoscope image into a binary image, and then removes noise such as hair and dead skin cells (S30), and the region of interest including only skin lesions in the dermatoscope image Can be extracted. The image preprocessed by the image preprocessor 140 may be stored in the storage unit 160.

4 is a detailed flowchart of step S30 of FIG. 3. 2 and 4, the image preprocessor 140 converts a dermatoscope image into a binary image (S32). In an embodiment, a set of objects having similar values may be extracted from a histogram of pixel values of a dermatoscope image using the Otsu technique or the like, and a threshold value for maximizing variance between divided regions may be created. The total variance can be expressed as the sum of the variance within the class and the variance between classes, and can be expressed as Equations (1) to (3) below.

[Equations (1) to (3)]

는 클래스 내의 분산,

는 클래스들 간 분산,

는 클래스

에 화소가 포함될 확률의 가중치,

는 클래스의 평균값이다. 이진 영상에는 피부 질환 분류의 정확성을 떨어뜨리는 불필요한 정보인 잡음이 제거되지 않고 포함될 수 있다. 잡음은 주로 환자의 모발이나 피부 각질 등의 피부 상태에 의하여 발생하게 된다. 이러한 잡음은 피부 병변 부위의 특징을 추출하는데 오류를 발생시키며, 정확도를 저하시키는 문제점을 발생시킨다.In formulas (1) to (3),

Is the variance within the class,

Is the variance between classes,

The class

The weight of the probability that a pixel is included in

Is the average value of the class. The binary image can be included without removing noise, which is unnecessary information that degrades the accuracy of classification of skin diseases. Noise is mainly caused by skin conditions such as the patient's hair or dead skin cells. Such noise causes an error in extracting features of a skin lesion area, and causes a problem of deteriorating accuracy.

In order to compensate for this problem, the image preprocessor 140 removes noise based on the binary image acquired in the preprocessing process. Since noise information is characterized in that the size information is relatively small compared to skin diseases in the image, the image preprocessor 140 removes the noise using the size characteristic of the noise, and based on the image from which the noise is removed, the region of interest Can be extracted.

5 is an exemplary diagram of a size distribution diagram of objects extracted from a dermatoscope image. Typically, the size of the hair, skin keratin, etc. is smaller than the size of the skin lesion (Lesion component) (eg, melanoma), so dermatoscope images of small objects whose size difference with the skin lesion exceeds the set value. By removing it from, noise such as hair or dead skin cells can be removed. In the case of a skin lesion, the image contains the largest number of pixels, and relatively noise has a component having a size smaller than that of the lesion.

The image preprocessor 140 may compare the sizes of components (objects) based on the binary image acquired in the preprocessing process in order to remove noise such as hair, normal skin, and dead skin cells other than the skin lesion. That is, the image preprocessor 140 may detect objects (eg, hair, dead skin cells, etc.) having a size smaller than that of a skin lesion in the binary image (S34). The image preprocessor 140 removes noise including hair and dead skin cells based on the difference in size between the skin lesion and the object (S36).

6 is an exemplary diagram of a process of pre-processing a dermatoscope image according to an embodiment of the present invention. 6A is a binary image converted from a dermatoscope image, (b) a binary image with noise removed, (c) a result of extracting a skin lesion boundary, and (d) a result of extracting a region of interest. In (b) of FIG. 6, it can be seen that noise generated by hair or dead skin cells has been effectively removed.

When noise is removed from the binary image, the image preprocessor 140 extracts information on the skin lesion area using the image from which the noise is removed, and performs image reconstruction to minimize information on normal skin based on the extracted information. do. In an embodiment, the image preprocessor 140 may extract a region of interest including only skin lesions by converting a pixel value of normal skin to 0 in the dermatoscope image (S38).

In an embodiment, the image preprocessor 140 may extract a boundary line for a skin lesion portion of the image, and calculate a size of the skin lesion portion based on the extracted boundary line. The image preprocessor 140 removes information on normal skin from the lesion area based on the calculated size of the skin lesion area, and prevents features from being extracted from the area other than the skin lesion area. By converting the pixel value of the region (normal skin) to 0, the region of interest may be extracted as shown in (d) of FIG. 6. Accordingly, errors in features caused by normal skin and noise can be minimized.

Conventionally, a specialist diagnoses a disease by using the ABCD method or the like by checking a dermatoscopy image with the naked eye based on clinical experience, and such subjective diagnosis has a problem that may cause an error. In order to compensate for this problem, in an embodiment of the present invention, all pixel information in a dermatoscope image is utilized, two transformation matrices are generated based on the similarity of the pixel cluster, and features are extracted from the transformation matrices to classify skin diseases. do.

7 is a block diagram of a skin disease classification unit constituting an automatic skin disease classification apparatus according to an embodiment of the present invention. 2, 3, and 7, the skin disease classification unit 150 uses brightness values of pixels of a dermatoscope image, a histogram, and two transformation matrices (GLCM, GLRLM) generated from image information, From the region of interest, various features (pixel features, GLCM features, and GLRLM features) are extracted (S40 to S90), and the extracted features are learned by the learning unit 120 as a machine learning model (e.g. For example, a support vector machine model) may be applied to classify skin diseases such as melanoma (S100). In an embodiment, the skin disease classification unit 150 includes a histogram calculation unit 151, a pixel feature extraction unit 152, a GLCM conversion unit 153, a GLCM feature extraction unit 154, a GLRLM conversion unit 155, and GLRLM. A feature extraction unit 156 and a classifier 157 may be included.

The histogram calculator 151 calculates a histogram of brightness values of pixels in an ROI of an image photographing a skin lesion (S40). The pixel feature extractor 152 extracts pixel features by first statistically analyzing the histogram of brightness values of the pixels (S50). Pixel features using only pixel information are features related to pixel-specific information and a histogram using the frequency of the pixel information, and can be used as important features representing overall information of an image.

In an embodiment, the pixel feature extraction unit 152 includes a pixel including an average, standard deviation, skew, kurtosis, entropy, and root mean square of pixel values of an image. Features can be extracted. Equations (4) to (7) are equations of pixel features used for feature extraction for skin lesions.

[Equations (4) to (7)]

는 관심 영역 안의 화소 정보(화소값),

는 관심 영역 안의 화소에 대한 히스토그램, N은 화소 개수, Ent는 히스토그램의 엔트로피, Kur는 화소값들의 커토시스, RMS는 화소값들의 제곱평균제곱근, STD는 화소값들의 표준편차이다.In equations (4) to (7),

Is the pixel information (pixel value) in the region of interest,

Is the histogram of the pixels in the region of interest, N is the number of pixels, Ent is the entropy of the histogram, Kur is the curtosis of the pixel values, RMS is the root mean square of the pixel values, and STD is the standard deviation of the pixel values.

The entropy of Equation (4) is a measure of the degree of disorder and indicates a characteristic of the frequency of histograms of pixels in an image. The curtosis of Equation (5) represents a measure of how much the distribution of pixel values in an image is distributed to a specific value by a probability distribution. The root mean square and standard deviation of Equations (6) and (7) can be used as a statistical measure of the magnitude of change of pixel values and a pixel feature related to the scatter of values in a matrix.

The GLCM converter 153 converts the pixel values in the region of interest of the image into a gray-level co-occurrence matrix (GLCM) representing the proximity between the pixel values in order to extract a feature related to the continuity of the pixels. Convert (S60). 8 is an exemplary diagram of a gray level co-occurrence matrix (GLCM) converted from an image according to an embodiment of the present invention.

The GLCM represents the frequency of pixel values of adjacent pixels, and is a matrix having a size of N×N (N is the number of total pixel values). Whenever an adjacent pixel group corresponding to an X-th (X is an integer of N or less) pixel value and a Y-th (Y is an integer of N or less) pixel value is found, the GLCM conversion unit 153 You can generate GLCM by increasing the value by 1. In the example of FIG. 8, since the continuous frequency of the pixels having the pixel value '1' is 1, the component of the first row and the first column of the GLCM is 1, and the pixel and the pixel value having the pixel value '2' (the second row of GLCM) Since the continuous frequency of the pixels having '1' (the first column of the GLCM) is 2, the component of the second row and one column of the GLCM is 2. By using the converted GLCM, it is possible to analyze the continuity between a pixel and an adjacent pixel in the image in the region of interest.

The GLCM feature extraction unit 154 extracts GLCM features based on the pixel values of the gray level co-occurrence matrix GLCM (S70). In an embodiment, the GLCM feature extraction unit 154 includes auto-correlation, contrast (contrast), correlation, and analogy of pixel values of the gray level co-occurrence matrix (GLCM). GLCM features including dissimilarity, energy, and homogeneity can be extracted. In an embodiment, the GLCM feature extractor 154 may extract GLCM features from the GLCM according to Equations (8) to (10) below.

[Equations (8) to (10)]

와

는 행렬의 위치,

는 GLCM의 화소값,

는 GLCM으로 변환되기 전의 영상의 화소값,

는

의 평균값,

는 GLCM에서 행에 대한 확률,

는

의 평균값,

는 GLCM에서 열에 대한 확률,

는

의 표준편차,

는

의 표준편차이다.In Equations (8) to (10), Cont is the contrast of GLCM pixel values, Corr is the correlation of GLCM, Homo is the homogeneity of GLCM,

Wow

Is the position of the matrix,

Is the pixel value of GLCM,

Is the pixel value of the image before being converted to GLCM,

Is

Mean value of,

Is the probability for a row in GLCM,

Is

Mean value of,

Is the probability for the column in GLCM,

Is

Standard deviation of,

Is

Is the standard deviation of

The contrast of Equation (8) is used as a feature of a measure of how much the contrast of pixels in the image is. The correlation of Equation (9) is used as a feature of a measure of how similar pixel values in an image are to each other. The homogeneity of Equation (10) is used as a feature of a measure of how similar pixel values are distributed among the pixels of an image.

The GLRLM conversion unit 155 converts the pixel values of the image into a gray level run length matrix (GLRLM) representing the continuous length of the pixel values (S80). 9 is an exemplary diagram of a gray level continuous length matrix (GLRLM) transformed from an image according to an embodiment of the present invention. The GLRLM is a matrix regarding how long pixels having a specific brightness value are continuously maintained at the same value (clusterability), and can be generated by accumulating the frequency of successive appearances of pixel values in an image in an ROI. Rows of GLRLM are pixel values, and columns are the number of consecutive pixel values. The GLRLM conversion unit 153 may generate the GLRLM by increasing the value of the (x, y) pixel by 1 whenever the x-th (x is an integer less than or equal to N) pixel values appear y times consecutively. In the example of FIG. 9, since the frequency in which the pixel value '1' (1 row of GLRLM) appears twice in succession (2 columns of GLRLM) is 1, the component value of column 1 and column 2 is 1, and the pixel value Since the frequency in which the pixel having '4' (4 rows of GLRLM) appears three times in succession (three columns of GLRLM) is 1, the component value of the 4 rows and 3 columns is 1.

The GLRLM feature extraction unit 156 extracts GLRLM features based on the pixel values of the gray level continuous length matrix (S90). In an embodiment, the GLRLM feature extraction unit 156 is based on the pixel values of the gray level continuous length matrix (GLRLM), SRE (Short Run Emphasis), LRE (Long Run Emphasis), GLNU (Gray Level Non-Uniformity), RP GLRLM features including (Run Percentage), Run Length Non-Uniformity (RLNU), and High Gray Level Run Emphasis (HGLRE) can be extracted.

The classifier 157 classifies a skin disease by applying pixel features, GLCM features, and GLRLM features to a machine learning model (S100). In an embodiment, the machine learning model may include a Support Vector Machine (SVM) model. According to an embodiment of the present invention, by extracting objective skin disease features that are difficult to determine with the naked eye and using it for classification of skin diseases, skin diseases such as melanoma and nevus can be classified with high accuracy, and subjective judgment is excluded. And objectively classify skin diseases.

In order to evaluate the skin disease classification performance and effectiveness of the automatic skin disease classification method according to an embodiment of the present invention, learning and testing were conducted on 43 and 76 images of DerMIS and Dermquest certified melanoma data, and MATLAB R2017a ( MathWorks) environment, and the skin classification accuracy was compared with the conventional method using the same learning data and test data. For SVM learning, after preprocessing the training data and extracting the region of interest, pixel, GLCM, and GLRLM transformation were performed using the image of the region of interest, and feature values were extracted from the transformed matrices. 10 is an exemplary diagram of melanoma images (a) and nevus images (b) used for machine learning according to an embodiment of the present invention.

Table 1 shows the feature values (Pixel features, GLCM features, GLRLM features) for melanoma and nevus used in the training data. From Table 1, it can be seen that differences in pixel information that are not visible to the naked eye due to features of melanoma and nevus appear in feature values extracted from transformation matrices (GLCM, GLRLM). Skin disease classification was performed through an SVM classifier using random test images using these feature values.

11 is a diagram showing accuracy performance of skin disease classification in a method for automatically classifying skin diseases according to an embodiment of the present invention. In FIG. 11, Alpha is the classification accuracy when melanoma and nevus are classified using only the ABCD method, Beta is the classification accuracy when using the ABCD method and pixel-based features, and Theta is the ABCD method and pixel features and GLCM features. Classification accuracy in the case of using the ``, Delta'' is the classification accuracy in the case of applying all of the ABCD method, pixel features, GLCM features and GLRLM features.

According to an embodiment of the present invention, by applying subtle information (pixel features) of a pixel that is difficult to see with the naked eye, transformation matrices based on similarity of adjacent pixels, and GLCM/GLRLM features extracted from the transformation matrices. It can be seen that the classification accuracy is improved. The automatic classification method for skin diseases according to an embodiment of the present invention may be used for quantitative and objective analysis and diagnosis in determining skin diseases, and in particular, may be usefully used in an early detection system for skin diseases such as melanoma. In addition, the method for automatically classifying skin diseases according to an embodiment of the present invention can be used to effectively provide information on diseases before invasive diagnosis such as a biopsy is performed even for skin diseases other than melanoma.

The method for automatically classifying skin diseases according to an embodiment of the present invention may be implemented in a general-purpose digital computer that operates the program using, for example, a program that can be executed on a computer and operates the program using a computer-readable recording medium. . Computer-readable recording media include volatile memories such as SRAM (Static RAM), DRAM (Dynamic RAM), SDRAM (Synchronous DRAM), Read Only Memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Nonvolatile memory such as Electrically Erasable and Programmable ROM (EEPROM), flash memory device, phase-change RAM (PRAM), magnetic RAM (MRAM), resistive RAM (RRAM), ferroelectric RAM (FRAM), floppy disk, hard disk, or The optical reading medium may be, for example, a storage medium in the form of a CD-ROM or a DVD, but is not limited thereto.

It should be understood that the above embodiments have been presented to aid the understanding of the present invention, and do not limit the scope of the present invention, and various deformable embodiments are also within the scope of the present invention. The technical protection scope of the present invention should be determined by the technical spirit of the claims, and the technical protection scope of the present invention is not limited to the literal description of the claims per se, but the invention of the scope of which the technical value is substantially equal. It should be understood that it reaches to.

100: automatic classification device for skin diseases
110: control unit
120: Learning Department
130: image acquisition unit
140: image preprocessor
150: skin disease classification unit
151: histogram calculation unit
152: pixel feature extraction unit
153: GLCM conversion unit
154: GLCM extraction unit
155: GLRLM conversion unit
156: GLRLM feature extraction unit
157: classifier
160: storage unit

Claims (15)

Obtaining a dermatoscope image by photographing a skin lesion using a dermoscopy;
Converting the dermatoscope image to a binary image, and detecting an object having a size smaller than that of the skin lesion in the binary image;
Removing noise including hair and dead skin cells based on the difference in size between the skin lesion and the object;
Extracting a region of interest including only skin lesions by converting a pixel value of normal skin to 0 in the dermatoscope image;
Calculating a histogram of brightness values of pixels in the region of interest, and statistically analyzing the histogram to extract pixel features;
Converting the pixel values of the image into a gray-level co-occurrence matrix (GLCM) representing proximity between the pixel values;
Extracting GLCM features based on pixel values of the gray level co-occurrence matrix;
Converting the pixel values of the image into a gray level run length matrix (GLRLM) representing a continuous length of the pixel values;
Extracting a GLRLM feature based on the pixel values of the gray level continuous length matrix; And
And classifying a skin disease by applying the pixel feature, the GLCM feature, and the GLRLM feature to a machine learning model. The method of claim 1,
The pixel feature is an automatic classification method for skin diseases including a mean, standard deviation, skew, curtosis, entropy, and root mean square of pixel values of the image. The method of claim 1,
The GLCM features are auto-correlation, contrast, correlation, dissimilarity, energy, and homogeneity of pixel values of the gray level co-occurrence matrix. Automatic classification method for skin diseases comprising a. The method of claim 1,
The GLRLM feature is SRE (Short Run Emphasis), LRE (Long Run Emphasis), GLNU (Gray Level Non-Uniformity), RP (Run Percentage), RLNU (Run Percentage) extracted based on the pixel values of the gray level matrix. Automatic classification method for skin diseases including Length Non-Uniformity) and High Gray Level Run Emphasis (HGLRE). The method of claim 1,
The machine learning model is a method for automatically classifying skin diseases including a support vector machine (SVM) model. delete The method of claim 1,
Generating the machine learning model by learning the pixel feature, the GLCM feature, and the GLRLM feature based on training image data including melanoma images and nevus images,
Classifying the skin disease comprises the step of classifying the skin lesion into one of melanoma and nevus based on the learned machine learning model. A computer-readable recording medium in which a program for executing the automatic classification method for skin diseases according to any one of claims 1 to 5 is recorded. In the skin disease automatic classification apparatus comprising at least one processor,
The at least one processor,
Converting a dermatoscopy image taken of a skin lesion by dermoscopy into a binary image;
Detecting an object having a size smaller than that of the skin lesion in the binary image;
Removing noise including hair and dead skin cells based on the difference in size between the skin lesion and the object;
Converting a pixel value of normal skin to 0 in the dermatoscope image to extract a region of interest including only skin lesions;
Calculating a histogram of brightness values of pixels in the region of interest and statistically analyzing the histogram to extract pixel features;
Converting the pixel values of the image into a gray-level co-occurrence matrix (GLCM) representing the proximity between the pixel values;
Extracting GLCM features based on pixel values of the gray level co-occurrence matrix;
Converting pixel values of the image into a gray level run length matrix (GLRLM) representing a continuous length of the pixel values;
Extracting GLRLM features based on the pixel values of the gray level continuous length matrix; And
An automatic skin disease classification apparatus configured to classify skin diseases by applying the pixel characteristics, the GLCM characteristics, and the GLRLM characteristics to a machine learning model. The method of claim 9,
The pixel feature is an automatic skin disease classification apparatus including a mean, standard deviation, skew, curtosis, entropy, and root mean square of the pixel values of the image. The method of claim 9,
The GLCM features are auto-correlation, contrast, correlation, dissimilarity, energy, and homogeneity of pixel values of the gray level co-occurrence matrix. Automatic classification device for skin diseases comprising a. The method according to any one of claims 9 to 11,
The GLRLM feature is SRE (Short Run Emphasis), LRE (Long Run Emphasis), GLNU (Gray Level Non-Uniformity), RP (Run Percentage), RLNU (Run Percentage) extracted based on the pixel values of the gray level matrix. Automatic classification device for skin diseases including Length Non-Uniformity) and High Gray Level Run Emphasis (HGLRE). The method according to any one of claims 9 to 11,
The machine learning model is an automatic skin disease classification apparatus including a support vector machine (SVM) model. delete The method according to any one of claims 9 to 11,
The at least one processor,
Generating the machine learning model by learning the pixel feature, the GLCM feature, and the GLRLM feature using training image data including melanoma images and nevus images; And
An automatic skin disease classification apparatus configured to classify the skin lesion into either melanoma or nevus based on the machine learning model learned using the training image data.

Images