KR101118211B1 - Diagnosis of submucosal tumor using endoscopic ultrasonic image analysis - Google Patents

Abstract

The present invention is to provide a method and apparatus for diagnosing diseases that can standardize images obtained by endoscopic ultrasonography, extract the characteristics of each disease appearing in the images, and apply them to objective diagnosis, thereby improving accuracy and reproducibility of prediction of malignancy. Standardizing the ultrasound region based on the contrast average value of the ultrasound region extracted from the endoscope ultrasound image through endoscopy, and extracting the tumor region from the normalized ultrasound region, and spots with high echoes among the extracted tumor regions ( separating a spot region from other regions, analyzing the tumor from which the spot region is extracted, calculating numerical data for analyzing symptoms or malignant information, and numerical data calculated through analysis of the tumor region. By applying the improved ART2-based RBF network algorithm based on Which comprises the step of classifying the diseases.

Description

Diagnosis of submucosal tumor using endoscopic ultrasonic image analysis}

The present invention relates to endoscopic ultrasonography, and more particularly, to a method for diagnosing a disease by ultrasound endoscopy image analysis of a submucosal tumor for standardizing an image acquired by endoscopic ultrasonography as a data for objective diagnosis.

In recent years, the use of ultrasound has been widely used. However, the ultrasonic waves sent from the surface of the body are strongly reflected by the gas in the digestive tract, so it is often difficult to observe deep parts. Therefore, an endoscope was developed for the purpose of observing organs around the endoscope by sending an ultrasonic vibrator to the end of the endoscope and sending an ultrasonic wave in the body cavity of the stomach.

In the past, the gastroenterology department mainly diagnosed using endoscopic ultrasound, but recently, it is used for various purposes such as tumor biopsy and drug administration. Submucosal tumors of the digestive tract are observed as spherical or hemispherical lesions that protrude into the lumen of the digestive tract by gastrointestinal endoscopy and are mostly detected by accident. Endoscopic ultrasomography (EUS) has a higher resolution than conventional imaging methods, so it is possible to directly observe the size, outline, development, internal appearance, and origin of the mosquito-derived layer of the submucosal tumor with higher resolution. It can be predicted to some extent, which is very useful for clinical diagnosis and treatment policy setting. In particular, there have been many attempts to predict whether malignant mesenchymal tumors are submucosal by endoscopy. In general, tumors with a diameter of 3 cm or more, uneven borders, and high-echo shadows or cyst spaces within tumors If observed or the surrounding lymph nodes are enlarged, the likelihood of a malignant submucosal tumor is high.

However, these studies are classified on the basis of general image photographs and subjective findings of doctors, so the reproducibility and accuracy are not high. Therefore, the subjects are diagnosed with the subjective findings of doctors. Have.

Accordingly, the present invention has been made to solve the above problems, by standardizing the image obtained by endoscopic ultrasonography, extracting the characteristics of each disease appearing in the image and applying it to the objective diagnosis, accuracy and reproducibility of the prediction of malignancy It is an object of the present invention to provide a method and apparatus for diagnosing a disease that can be enhanced.

In order to achieve the above object, a characteristic of the diagnosis method of the disease by the ultrasound endoscopy image analysis of the submucosal tumor according to the present invention is (A) the ultrasound region extracted from the endoscopy ultrasound image through the endoscopy ultrasonography Normalizing a region, (B) extracting a tumor region from the normalized ultrasound region, separating a spot region with high echo from the extracted tumor region from other regions, and (C) the spot Calculating the numerical data for analyzing symptoms or malignant information by analyzing the tumor from which the region has been extracted, and (D) applying an ART2-based RBF network algorithm developed based on the numerical data calculated through the analysis of the tumor region. To classify the disease.

Preferably, the step (A) comprises converting the extracted ultrasound region to a gray level having a contrast value between 0 and 255, and (A2) applying a histogram smoothing to the ultrasound image converted to the gray level to contrast. Uniformly processing the distribution of values, and (A3) present all values that are less than or equal to a threshold value of the current pixel of all pixels constituting the edges existing in the ultrasound region to which the histogram smoothing is applied. Applying an edge-linking method for connecting and recording with the pixel; (A4) binarizing the threshold value of the ultrasonic region to which the edge link method is applied to 180; and (A5) the binarized ultrasound GrassFire knows to use the self-call on a region to in turn examine the neighboring neighboring pixels of the current pixel of interest until all adjacent elements are labeled (A6) extracting the lens area of the endoscope having a form having a ratio of 1: 1 horizontally and vertically among the labeled areas, and having more contrast 0 pixels than contrast 255 pixels. (A7) recording the center value of the extracted lens region, and then labeling the remaining regions except the lens region with the GrassFire algorithm, and (A8) removing the noise from the labeled image to remove the candidate region. (A9) extracting the candidate region nearest the distance to the final region by applying the Euclidean distance method to the center value of the recorded lens and the extracted candidate regions; A10) comparing the average intensity of the extracted final region with the average intensity of the ultrasonic region, characterized in that it comprises the step of normalizing the ultrasonic region do.

를 적용하여 거리(d)를 산출하는 것을 특징으로 한다.Preferably, the Euclidean distance method of (A9) is

It is characterized in that for calculating the distance (d) by applying.

Preferably, in the step (B), the spot region with high echo is extracted by applying the LVQ neural network algorithm proposed by Kohonen, in which the initial value of the center vector is randomly selected and the intensity value of the tumor region is applied as the input vector. To normalize to a value between 0 and 1.

를 이용하여 계산하는 단계와, 상기 수식에서의 입력 벡터와 승자 벡터가 같은 클러스터이면 수식

를 적용하고, 입력 벡터와 승자 벡터가 다른 클러스터이면 수식

를 적용하여 계산된 가중치를 갱신하는 단계와, 상기 갱신된 가중치를 기반으로 학습률

를 시간에 따라 감소시키면서 학습률을 조정하여 에코가 높은 스팟(spot) 영역을 추출하는 단계를 포함하는 것을 특징으로 한다.Preferably, the extraction of the high echo spot region comprises randomly selecting a center vector in the selected tumor region, initializing a learning rate and the number of repetitions, and calculating a winner node K of the selected center vector by the following equation.

And calculating if the input vector and the winner vector in the formula are the same cluster.

If the input vector and the winner vector are different clusters, the formula

Updating the calculated weight by applying a value, and learning rate based on the updated weight value

It is characterized in that it comprises the step of extracting a spot area with a high echo by adjusting the learning rate while decreasing over time.

Preferably, the method of analyzing the tumor in the step (C) is characterized in that it comprises any one of the distribution of high contrast, whether the calcium is deposited, analysis through the outer shape of the tumor and echo analysis.

Preferably, in the step (D), the improved ART2-based RBF network algorithm changes the economic variable by using the distance difference of each input pattern, and dynamically adjusts the learning rate by applying accuracy and inaccuracy to the fuzzy membership function. It is done.

As described above, the method for diagnosing a disease by the ultrasound endoscopy image analysis of the submucosal tumor according to the present invention has the effect of increasing the accuracy of prediction of malignancy by increasing the reproducibility of the subjective diagnosis classified based on the image photograph and the subjective findings of the doctor. There is.

1 is a flowchart illustrating a method for diagnosing a disease by ultrasound endoscopy image analysis of a submucosal tumor according to an embodiment of the present invention.
2 is a flowchart illustrating a method of standardizing the extracted ultrasound region in detail
[Figure 3] Image showing the extraction process of the water region
4 is a flowchart illustrating a method of extracting a spot region by applying an LVQ neural network algorithm
[Figure 5] Separation image of the spot (spot) using the LVQ algorithm
6 is a diagram showing the results of light and dark distribution analysis using a histogram
7 is a diagram illustrating a bit plane segmentation process
8 is an image showing the outer region of the tumor
9 is an image showing the outer appearance of the tumor
[Figure 10] Graph of the contrast of the echo
[Figure 11] TSS change rate according to the number of iterations
[Figure 12] ART2-based RBF network developed

Other objects, features and advantages of the present invention will become apparent from the following detailed description of embodiments taken in conjunction with the accompanying drawings.

Referring to the accompanying drawings, a preferred embodiment of a method for diagnosing a disease by ultrasound endoscopy image analysis of a submucosal tumor according to the present invention will be described. However, the present invention is not limited to the embodiments disclosed below, but can be implemented in various different forms, only the embodiments are to make the disclosure of the present invention complete and complete the scope of the invention to those skilled in the art. It is provided to inform you. Therefore, the embodiments described in the present specification and the configuration shown in the drawings are only the most preferred embodiments of the present invention and do not represent all of the technical idea of the present invention, and various equivalents may be substituted for them at the time of the present application. It should be understood that there may be variations and examples.

1 is a flowchart illustrating a method for diagnosing a disease by ultrasound endoscopy image analysis of a submucosal tumor according to an embodiment of the present invention.

As shown in FIG. 1, first, as an first step, an ultrasound region extracted from an endoscope ultrasound image through an endoscopy ultrasound is normalized based on an average value of contrast (S100).

2 is a flowchart illustrating a method of standardizing the extracted ultrasound region in detail.

As shown in FIG. 2, first, the extracted ultrasound region illustrated in FIG. 3A is converted into a gray level having a contrast value between 0 and 255 (S101). 3B illustrates an ultrasound image converted to gray level.

Subsequently, histogram smoothing is applied to the ultrasound image converted to gray level to uniformly distribute the contrast (S102). 3 (c) shows an image to which histogram smoothing is applied.

As shown in FIG. 3 (c), since there are countless edges in the ultrasonic region to which the histogram smoothing is applied, the following equation 1 is satisfied if all pixels constituting the edges are less than or equal to the threshold value of the current pixel minus the next pixel. Apply.

That is, as shown in FIG. 3 (d), an edge-linking method in which all values satisfying the criterion of Equation 1 in the 3 × 3 region adjacent to the current pixel is connected and recorded with the current pixel is applied. (S103). The applied ultrasound region is binarized by setting the threshold value to 180 through an experiment (S104). 3 (e) shows a threshold binarized image.

Next, the binarized ultrasound region is labeled by applying a GrassFire algorithm as shown in FIG. 3 (f) (S105). In this case, the GrassFire algorithm is a labeling method that sequentially inspects neighboring neighboring pixels of the current pixel of interest until all neighboring elements are labeled using self-calling.

Among the labeled areas, the lens area of the endoscope is extracted in a ratio of 1: 1 in width and length, and the number of pixels having zero contrast is greater than the number of pixels having contrast 255 (S106). 3 (g) shows an image obtained by extracting a lens region.

Subsequently, the center value of the extracted lens region is recorded as shown in FIG. 3 (h), and the remaining regions except for the lens region are labeled with the GrassFire algorithm (S107). The noise is removed from the image labeled with the GrassFire algorithm as shown in FIG. 3 (i) (S108), and the candidate region of water is extracted as shown in FIG. 3 (j) (S109).

Next, the center value of the recorded lens area and the center value of the extracted water candidate areas are calculated by applying the Euclidean distance method as shown in Equation 2, and then the nearest candidate area is extracted as the final water area ( S110). 3 (k) is an image showing a region of extracted final water.

Then, the contrast standard value of the extracted region of the final water and the contrast average value of the ultrasound region are compared to normalize the image (S111). The normalized image is shown in FIG. 3 (l), and standardizes all ultrasound regions in this manner.

Subsequently, as a second step, a tumor region is extracted from the standardized ultrasound region, and a spot region having a high echo among the extracted tumor regions is separated from other regions (S200). In this case, when the tumor area is automatically extracted from the standardized ultrasound area, the accuracy of diagnosis may be lowered due to the loss of surgical and morphological information. Therefore, a specialist may select the tumor area from the ultrasound area to analyze the accurate information. It is desirable to. However, this is not limited thereto as one embodiment.

In addition, the LVQ neural network algorithm proposed by Kohonen is applied to extract a spot region with high echo from the selected tumor region. For reference, the performance of LVQ algorithm depends on the method of selecting the initial value of the center vector. Since the center vector of the LVQ algorithm determines the class boundary and affects the convergence, selecting the initial value of the center vector is very important. In general, a method of selecting a center vector is a method of arbitrarily selecting among training data of an input pattern. In this specification, however, the initial value of the center vector is randomly selected and normalized to a value between 0 and 1 in order to apply the intensity value of the tumor region as an input vector.

4 is a flowchart illustrating a method of extracting a spot region by applying an LVQ neural network algorithm.

Referring to FIG. 4, first, a central vector is randomly selected from a selected tumor region, and a learning rate and a repetition number are initialized (S201).

 Next, the winner node K of the selected center vector is calculated as in Equation 3 below (S202).

In Equation 3, if the input vector and the winner vector are the same cluster, Equation 4 is applied. If the input vector and the winner vector are different clusters, the calculated weight is updated by applying Equation 5 (S203).

를 시간에 따라 감소시키면서 학습률을 조정하여 에코가 높은 스팟(spot) 영역을 추출한다(S204). 그리고 종료 조건에 만족하면 학습을 종료하고, 그렇지 않으면 상기 수학식 3을 통한 계산과정을 반복 수행한다(S205).Learning rate based on the updated weight

By adjusting the learning rate while decreasing over time to extract a spot area with a high echo (S204). If the end condition is satisfied, the learning is terminated. Otherwise, the calculation process through Equation 3 is repeated (S205).

FIG. 5 (a) shows a selected tumor image, FIG. 5 (b) is an image showing a result of VLO application, and FIG. 5 (c) is an image showing a spot region finally extracted.

Subsequently, as a third step, numerical data for analyzing symptoms or malignant information is calculated by analyzing the tumor from which the spot region is extracted (S300). At this time, the method for analyzing the tumor includes the distribution of high contrast, the presence of calcium, the analysis through the outer shape of the tumor and echo analysis.

In the first method, the intensity distribution analysis, the histogram is used to analyze the distribution of the intensity increase. As shown in FIG. 6, the histogram enables the physician to analyze the intensity distribution of the tumor area.

The second method, calcium deposition area extraction analysis, can be used to determine the extent of symptoms or malignancy through the deposition of calcium in selected tumor areas. The deposited calcium has a brighter contrast than other tissues and the tissues are clustered. Therefore, the bit plane segmentation method, which provides important information for analyzing the importance of a particular bit, is applied when extracting calcium deposition regions. Histogram stretching is applied to selected tumor areas to make the contrast more pronounced before applying the beat plane segmentation. In this case, the simple histogram stretching is effective when the pixel contrast is concentrated only in a specific brightness region, and when the low value is 0 and the high value is 255, the effect of stretching is reduced. Histogram stretching using end-in search is applied to the tumor area. That is, the bit plane segmentation method is applied to the stretched tumor area.

The bit-plane partitioning method is a method of binarization using a feature in which a bit corresponding to the most significant bit holds the outline information of the image best in a gray scale including 8 bits. As a result of extracting the calcium deposition region by applying the bit plane segmentation method to the selected tumor region as shown in Figure 7 (a) to Figure 7 (h) is shown in Figure 7 (h).

The third method, tumor morphology analysis, can distinguish between benign and malignant tumors according to the morphology of the selected tumor area. Benign tumors, ie, non-cancers, are often round or oval because the tumors grow uniformly. On the other hand, malignant tumors, i.e. tumors having cancerous characteristics, do not grow uniformly, and thus have irregularities like an amoeba. Therefore, in the present specification, the slope information of the outline is used to distinguish the features of the outline form of the tumor. When a specialist selects a tumor area, the boundary is considered the outer area. 8 (b) shows the result of displaying the outer region in the selected tumor region.

At this time, the outer area is applied to the following equation (6) to calculate the slope of the outline.

과

은 기울기를 계산할 위치의 좌표 값이며,

과

는

과

에서 외곽선을 따라 10픽셀 이동한 위치의 좌표 값이다.At this time, in Equation 6

and

Is the coordinate value of the location to calculate the slope of,

and

Is

and

The coordinate value of the position moved 10 pixels along the outline in.

As a result of examining the change of the calculated slope value, it can be seen that the value of the slope suddenly changes in the region in which the amoeba shape appears as shown in FIG. 9.

The fourth method, echo analysis, shows that the characteristics of the intensity of echo have a great influence on the tumor malignancy. Therefore, detailed echo analysis is needed. Outputs distances and gray levels for two randomly selected points for detailed echo analysis.

Therefore, the distance between two selected points is calculated using the Euclidean distance method as shown in Equation 7, and the slope of the selected two points is calculated using the same method as in Equation 8.

Using the distance and the slope calculated as described above, the intensity graph is output as shown in FIG. 10.

As a fourth and final step, the disease is classified by applying the improved ART2-based RBF network algorithm based on the numerical data calculated through the analysis of the tumor region (400). At this time, the developed ART2-based RBF network algorithm is an algorithm that dynamically adjusts the learning rate by changing the economic variable by using the distance difference of each input pattern, and applying accuracy and inaccuracy to the fuzzy membership function.

In this way, the intensity of the tumor area was analyzed from the general images of the submucosal tumors taken during the endoscopy, and 15 ultrasound endoscopy images provided by the gastroenterologist were used to classify the diseases. For reference, five studies were conducted on three diseases of GCP, lipoma, and GIST. The boundary parameter of the improved ART2-based RBF network algorithm applied to classify disease was set to 0.2 and the initial learning rate was set to 1.0. The learning was terminated when the total error squared sum (TSS) became 0.05 or less.

Next, FIG. 11 is a learning result of the developed ART2-based RBF network algorithm, and FIG. 12 shows an ART2-based RBF network algorithm developed for disease classification.

By applying the ART2 based RBF network algorithm developed as shown in FIG. 11, it was confirmed that early saturation does not occur in the process of learning and classifying three types of disease data of GCP, lipoma, and GIST. The ART2-based RBF network algorithm is a well-known technique and detailed description thereof will be omitted.

Although the technical spirit of the present invention described above has been described in detail in a preferred embodiment, it should be noted that the above-described embodiment is for the purpose of description and not of limitation. In addition, those skilled in the art will understand that various embodiments are possible within the scope of the technical idea of the present invention. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

Claims (7)

(A) standardizing the ultrasound region extracted from the endoscope ultrasound image by endoscope ultrasound based on the contrast mean value;
(B) extracting a tumor region from the standardized ultrasound region, and separating the extracted spot region from other regions by applying the LVQ neural network algorithm proposed by Kohonen to the extracted tumor region; ) Separating the region from other regions,
(C) calculating the numerical data for analyzing the symptoms or malignant information by analyzing the tumor from which the spot region is extracted;
(D) classifying the disease by applying the ART2-based RBF network algorithm developed based on the numerical data calculated through the analysis of the tumor region. The method of claim 1, wherein step (A)
(A1) converting the extracted ultrasound region to a gray level having a contrast value between 0 and 255,
(A2) applying a histogram smoothing to the ultrasound image converted to the gray level to uniformly process the distribution of contrast values;
(A3) an edge link connecting and recording all values of the current pixel of all pixels constituting the edges existing in the ultrasound region to which the histogram smoothing is applied, with the current pixel having an absolute value less than or equal to a threshold value equal to or less than a threshold. applying a linking method,
(A4) binarizing the threshold value of the ultrasound region to which the edge link method is applied to 180;
(A5) labeling the binarized ultrasound region by applying a GrassFire algorithm that sequentially inspects neighboring neighboring pixels of the current pixel of interest until all adjacent elements are labeled using self-calling;
(A6) extracting the lens area of the endoscope having a form in a ratio of 1: 1 in width and length among the labeled areas, and having more contrast 0 pixels than contrast 255 pixels;
(A7) recording the center value of the extracted lens region, and then labeling the remaining regions except the extracted lens region with the GrassFire algorithm;
(A8) extracting candidate regions by removing noise from the labeled regions;
(A9) applying the Euclidean distance method to the center of the recorded lens and the center of the extracted candidate regions to extract the candidate region closest to the final region;
(A10) comparing the mean intensity of the extracted final region and the mean intensity of the ultrasonic region to normalize the ultrasound region. The method of claim 2,
The Euclidean distance method of (A9) is

The method of diagnosing a disease, characterized in that to calculate the distance (d) by applying. The method of claim 1,
In the LVQ neural network algorithm proposed by Kohonen in step (B), the initial value of the center vector is randomly selected and normalized to a value between 0 and 1 to apply the intensity value of the tumor region as an input vector. How to diagnose the disease. The method of claim 4, wherein the spot region of step (B) is
Randomly selecting a center vector in the selected tumor area, initializing a learning rate and the number of repetitions;
The winner node K of the selected center vector is

Calculating using;
If the input vector and the winner vector in the formula are the same cluster, the formula

If the input vector and the winner vector are different clusters, the formula

Updating the calculated weight by applying a;
Learning rate based on the updated weight Adjusting the learning rate while decreasing over time. The method of claim 1,
The method of analyzing the tumor in the step (C) is a method for diagnosing a disease, characterized in that any one of the distribution of high contrast, calcium deposition, analysis of the tumor through the outer shape and echo analysis. The method of claim 1,
The improved ART2-based RBF network algorithm in step (D) is characterized by changing the economic variable by using the distance difference of each input pattern, and dynamically adjusting the learning rate by applying accuracy and inaccuracy to the fuzzy membership function. How to diagnose the disease.

Images