KR101297336B1 - Device For Classifying Mucinous Cystadenoma - Google Patents

Abstract

The present invention relates to a device for diagnosing mucous cystic tumors, wherein the diagnostic device according to the present invention identifies a luminal region from a tissue cell image input by a user, identifies an epithelial cell from the identified luminal region, and then identifies the identified epithelium. Features required for cystic tumors are extracted from the cells and mucinous cystic tumors are detected from tissue cell images using SVMs learned about the extracted features.

Description

Device for classifying mucinous cystadenoma

The present invention relates to a mucinous cystic tumor diagnosis apparatus for discriminating a mucinous cystic tumor by extracting features for diagnosing a mucinous cystic tumor by identifying a plurality of luminal regions and epithelial cells from an input tissue cell image.

Pancreatic cancer is the malignant tumor that causes cancer deaths in the West and fourth place in Korea, and sixth in Korea, accounting for 5% of cancer deaths. Although there are various techniques for the diagnosis of pancreatic cancer, it is difficult to diagnose the cancer early, so only 10 to 15% of patients are expected to undergo curative resection, and the recurrence rate is high even if the resection is performed. Five-year survival rates are known to be less than 20% [Park, Staging Work-up, r, How to Manage, Korean Pancreatobiliary Association Fall Conference, 2008]. Among these malignancies, pancreatic cystic tumors are very rare and show a frequency of 10-15% of pancreatic cystic lesions and less than 1% of pancreatic cancer [Warshaw AL, Rutledge PL. Cystic tumors mistaken for pancreatic pseudocysts. Ann Surg 1978, 205: 393-398. The cystic tumors of the pancreas are as follows (Sang Myung Woo, et al., Cystic Fluid Anaysis for the Differential Diagnosis of Pancreatic Cyst, The Korean Journal of Gastroenterology, 2002; 40: 394-401).

Serous cystic adenoma

Mucinous cystic tumor

Solid pseudo papillary tumor

Intraductal papillary mucinous tumor in the pancreatic duct

Acinar cell cystadeocarcinoma

Cystic islet cell tumor

Lymphangioma

Among these cystic tumors, mucinous cystic tumors and serous cystic tumors account for more than 75% of pancreatic cystic tumors. Serous cysts are benign tumors with no potential for malignancy, while mucinous cystic tumors are malignant when detected as premalignant tumors with potential for malignancy. In the case of mucinous cystic tumors, 70-80% of the body and tail are easy to be removed surgically when detected. The 5-year survival rate is also good at 60-70% [Taft DA, Freeny PC. Cystic neplasms of the pancreas. Am J Surg 1981, 142: 30-35; Yamaguchi K, Enjoji M. Cystic neoplasms of the pacreas. Gastroenterology 1987, 188: 679-684; Hodgkinson DJ, ReMine WH, Weiland LH. A clinicopathologic study of 21 cases of pancreatic cystadenocarcinoma. Ann Surg 1987, 188: 679-684. Therefore, it is clinically very important to discriminate mucosal cystic tumors with possible malignancy.

Pathologists perform the diagnosis by examining morphological features such as the histological structure of the biopsy and cell distribution obtained through electron microscopy. However, since such diagnosis depends on the subjective experience of the individual pathologist, there are various views on the same symptom [A. Andrion, C. Magnani, P.G. Betta, A. Donna, F. Mollo, M. Scelsi, P. Bernardi, M. Botta, B. Terracini, Malignant mesothelioma of the pleura: interobserver variability, J. Clin. Pathol. 48 (1995) 856-860; S.M. Ismail, A.B. Colclough, J.S. Dinnen, D. Eakins, D.M. Evans, E. Gradwell, J.P. O'Sullivan, J.M. Summerell, R.G. Newcombe, Observer variation in histopathological diagnosis and grading of cervical intraepithelial neoplasia, Br. Med. J. 298 (1989) 707-710. In order to overcome this subjective judgment problem and to perform objective pathological diagnosis, a computer-aided diagnostic system (CAD) based on quantitative numerical data has recently been developed.

W.N. Street and his colleagues proposed a method of distinguishing benign from malignant by quantifying the morphological features of breast cells [W.N. Street, W.H. Wolberg, O. L. Mangasarian, Nuclear feature extraction for breast tumor diagnosis, IS & T / SPIE 1993 International Symposium on Electronic Imaging: Science and Technology 1905 (1993) 861-870, San Jose, CA]. Shivang Naik uses low-level infomation (raw level information) and high leve-information (pathological structural information) to automatically identify the nucleus and gland and identify the morphological features and boronois Graph-based features, such as the minimum spanning tree, were used to diagnose prostate cancer and breast cancer [S. Naik, S. Doyle, S. Agner, A. Madabhushi, M. Feldman, J. Tomaszewski, "Automated gland and nuclei segmentation for grading of prostate and breast cancer histopathology," 5th IEEE international symposium on biomedical imaging, 2008, 284-284-. 287]. Makinaci et al. Extract surface characteristics using Gauss markov random field, correlation function and relative entropy techniques and use Support Vector Machine (SVM). Prostate cancer was diagnosed [Metehan Makinaci, "Support Vector Machine Approach for Classification of Cancerous Prostate Regions," In Proc. of world academy of science engineering and technology, 2005, 166-169]. Bilgin et al. Calculated quantitative metrics and diagnosed breast cancerous tissues using cell graphs based on graph theory [Cagatay Bilgin, Cigdem Demir, Chandandeep Nagi, Bulent Yener, "Cell-Graph Mining for Breast Tissue Modeling and Classification," IEEE EMBS, 2007, 5311-5314. In addition, various techniques for diagnosing cancerous tissue have been proposed.

In other words, research on a pathological support system for tissue cell imaging using a computer is currently insufficient. In addition, many of the previous studies focus on a few specific lesions, such as the breast or prostate, and most of the techniques focus on the morphological features of the nucleus. In particular, when diagnosing the area of the tube or sac (that is, the area including the lumen), which is the area of interest of many pathology diagnosis, automatic identification is not performed. Therefore, the user can identify the area by using a mouse or an electronic pen. There is a problem that must be addressed.

The present invention has been made to solve the above problems, the present invention is to provide a mucinous cystic tumor diagnosis device for automatically identifying a plurality of lumen areas from a given tissue cell image.

The present invention also provides a mucinous cystic tumor diagnosis apparatus for identifying epithelial cells covering areas including lumens such as sac and ducts using a lattice method.

The present invention also provides an apparatus for diagnosing mucous cystic tumors which extracts features such as epithelial cytoplasmic length and the presence of mucus from the identified epithelial cells.

Mucinous cystic tumor diagnosis related to an embodiment of the present invention for realizing the above object is an image input unit for providing a tissue cell image, a luminal region identification unit for identifying a plurality of lumen areas from the tissue cell image, the identification Using an epithelial cell identification unit for identifying epithelial cells from the luminal region, a feature extractor for extracting morphological features for cystic tumor diagnosis from the identified epithelial cells, and an SVM learned about the extracted features. And a diagnosis unit for detecting a mucus cystic tumor from the tissue cell image.

In accordance with at least one embodiment of the present invention configured as described above, the present invention may automatically identify a plurality of lumen areas from a given tissue cell image using a direction accumulation histogram.

In addition, the present invention can quickly identify the epithelial cells covering the lumen through the lattice color method, and can be widely applied to similar regions (such as tubes) including the lumen region.

In addition, even though the lesions are different, the morphological characteristics of the cells and the morphology of the nucleus are similar, the feature extraction techniques proposed in the present invention can be used to extract features of pancreatic tissue cell images as well as other tissue cell images. It can be applied in various forms.

In addition, the present invention quantifies the features of the mucinous cystic tumor, enabling the physician to use the quantified features as an aid for objective diagnosis and contribute to the objective and accurate diagnosis.

1 is a block diagram of a mucinous cystic tumor diagnosis apparatus according to an embodiment of the present invention.
2 is a diagram illustrating a process of diagnosing a mucinous cystic tumor associated with one embodiment of the present invention.
3 is a diagram illustrating a preprocessing process of a tissue cell image for identifying a lumen region according to the embodiment of FIG. 2.
4 is an exemplary diagram illustrating a preprocessed image and a pixel matrix thereof according to the embodiment of FIG. 3.
FIG. 5 shows the results of the direction accumulation histogram H (A) for FIG. 4. FIG.
6 is a diagram illustrating a process of extracting candidate seed points by a diagnostic apparatus according to an embodiment of the present invention.
Figure 7 illustrates a pretreatment process for identifying epithelial cells associated with one embodiment of the present invention.
FIG. 8 is a diagram illustrating a method for identifying epithelial cell nuclei from a set of cell nuclei extracted through the pretreatment process of FIG. 7.
9 is a diagram illustrating a method for measuring cytoplasmic length in accordance with an embodiment of the present invention.
10 is a view showing a measurement process of the presence of slime associated with one embodiment of the present invention.

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

In the following description, although a sac site identification is illustrated, the present invention is applicable not only to the sac site identification but also to the identification of similar other sites (eg, ducts) including the lumen. Here, the sac is composed of morphologically white luminal areas and epithelial cells covering them. This type of epithelial tissue is found in a similar form not only in the sac but also in the duct. Thus, the technique presented in the present invention can also be used to identify similar regions (eg, coffins), such as sacs.

According to the present invention, the identification of a mucinous cystic tumor proceeds through four steps as follows. The first step is to identify the lumen boundary to observe epithelial cells. The second step identifies epithelial cells of interest using the identified lumen boundary information. The third step extracts existing morphological features and newly defined morphological features from the identified epithelial cells. The fourth step is to diagnose using the learned SVM (Support Vector Machine) classifier and output the result.

1 is a block diagram of a mucinous cystic tumor diagnosis apparatus according to an embodiment of the present invention.

The mucous cystic tumor diagnosis apparatus includes an image input unit 110, a lumen region identification unit 120, an epithelial cell identification unit 130, a feature extraction unit 140, a diagnosis unit 150, a control unit 160, and a user input unit ( 170, the display unit 180, and the like.

The image input unit 110 obtains and outputs a tissue cell image. The image input unit 110 may access and output a tissue cell image previously stored in a memory (not shown).

The lumen region identification unit 120 receives a tissue cell image from the image input unit 110 and identifies a lumen region in the tissue cell image. The lumen region identification unit 120 selects candidate seed points of the lumen through a direction accumulation histogram algorithm, and automatically identifies the boundary region of the lumen using an area growth method based on the candidate seed points.

Epithelial cell identification unit 130 identifies the epithelial cells from the lumen area identified by the lumen area identification unit 120. Epithelial cell identification unit 130 identifies the epithelial cells around the identified lumen through an epithelial cell identification algorithm.

The feature extractor 140 extracts morphological features for identifying mucinous cystic tumors from the identified epithelial cells. The morphological features include features such as cytoplasmic length of the epithelial cells and the presence of mucus. The extracted morphological features are used to learn the SVM used for cystic tumor diagnosis.

The diagnosis unit 150 detects a mucinous cystic tumor from a given epithelial cell image using the SVM learned about the extracted features, and thus diagnoses the mucus cystic tumor. The SVM (classifier) obtains an optimal discriminant for classifying data through learning about the selected data when influential features are selected for the detection of a mucinous cystic tumor. The SVM classifies the mucinous cystic tumor from the given epithelial cell image using the obtained optimal discriminant.

The controller 160 controls each of the above components to control the overall operation of the diagnostic apparatus. Programs for operating the controller 160 and data input / output may be stored in a memory (not shown). The memory may store tissue cell images.

The user input unit 170 generates input data for controlling the operation of the diagnostic apparatus according to a user's manipulation. The user input unit 170 may be implemented as a keyboard, a mouse, an electronic pen, or the like.

The display unit 180 displays various data according to the operation of the diagnostic apparatus under the control of the controller 160. The display unit 180 may include a liquid crystal display (LCD), a thin film transistor-liquid crystal display (TFT LCD), an organic light-emitting diode (OLED), and a flexible display (flexible). and at least one of a 3D display.

2 is a diagram illustrating a process of diagnosing a mucinous cystic tumor associated with one embodiment of the present invention.

First, the controller 160 of the mucus cystic tumor diagnosis apparatus (hereinafter, referred to as a diagnosis apparatus) receives a tissue cell image from the image input unit 110 (S101). The tissue cell image may be an image captured by the image input unit 110 or an image stored in a memory (internal memory or external memory, etc.).

The controller 160 controls the lumen region identification unit 120 to identify the lumen region in the tissue cell image (S120). The lumen region identification unit 120 performs preprocessing on the input tissue cell image (S121) and random seed points included in the lumen region from the preprocessed image using a direction accumulation histogram algorithm. Extracted as candidate seeds (candidate points) (S122). In addition, the lumen region identification unit 120 detects the boundary of the visceral region in the tissue cell image by applying a region growth method to the extracted candidate seed points (S123).

As described above, the diagnostic apparatus performs preprocessing on the input tissue cell image for applying the cumulative histogram algorithm, which will be described with reference to FIG. 3.

First, the lumen region identification unit 120 removes impurities in the lumen region by filtering the tissue cell image input through the intermediate filter. Thus, the lumen area is generally whiter.

In addition, the lumen area identification unit 120 converts the filtered tissue cell image into a gray image, and effects on various environmental factors (eg, lighting) on image acquisition through background correction. Decreases [K. Rodenacker, E. Bengtsson, A feature set for cytometry on digitized microscopic images, Anal. Cell. Pathol. 25 (2003) 1-36.

Thereafter, the lumen region identification unit 120 generates a binary image through a maximum entropy threshold (MET) process of the tissue cell image undergoing the background calibration. For the application of the direction accumulation histogram algorithm, the following preliminary information is assumed for the preprocessed image. First, the lumen area is white as the foreground and white is defined as 1. Second, the rest is black as the background and black is defined as zero.

4 is an exemplary diagram illustrating a preprocessed image and a pixel matrix thereof according to the embodiment of FIG. 3.

The given image is m × x matrix A, and the i th row of matrix A is denoted by a _i and the j th column is denoted by b _j .

delete

delete

The lumen area identification unit 120 obtains a direction accumulation histogram H ( A ) from the preprocessed binary image. This histogram is obtained by calculating the cumulative values in each direction by skipping 0 (background) for each of the four directions (right, left, up, down) and accumulating the foreground pixel values, and adding up the cumulative values. Lose. The operation to accumulate only the values of the foreground pixels without ignoring the pixel values of the background is defined as shown in Table 1.

Equations 2 and 3 are directional accumulation histograms H (A) defined using the binaryAcumulationBySkppingZero operation of Table 1.

는 주어진 조직세포 영상에 존재하는 다수의 내강 영역들의 후보 씨앗(seed)점들이 될 수 있다. 그러나 만약

를 H(A)로부터 직접 구하게 되면, 불필요하게 아주 작은 영역들에 대해서도 후보 씨앗(seed) 점들이 생성된다. 이러한 문제를 방지하고, 적당한 크기의 내강영역들에 대한 후보 씨앗(seed) 점들을 얻기 위해 임계치 값을 도입하여 H(A)의 값이 임계치보다 큰 경우에 대해서만 후보 씨앗점으로 선정(추출)한다. 이때 임계치가 적용된 방향누적히스토그램 HT(A)은 수4와 같이 정의된다.FIG. 5 is an exemplary diagram showing a table of direction accumulation histograms H (A) of FIG. 4. As shown in FIG. 5, the maximum point of H (A) is the seed point of the region growth method to be applied to identify the lumen from a given image. However, if there are many lumen areas to be identified from a given image, it is impossible to identify multiple lumen areas only by the maximum point of H (A). Therefore, in the present invention, the partial derivative of H (A) is calculated for x and y, and a vertical local maximum curve (LMC) and a horizontal local maximum curve are found for x and y, respectively. The intersections of these two regional maxima

May be candidate seed points of multiple luminal regions present in a given tissue cell image. But if

Is found directly from H (A), candidate seed points are generated for unnecessarily very small areas. To avoid this problem, the threshold value is introduced to obtain candidate seed points for appropriately sized lumen areas and selected (extracted) as a candidate seed point only when the value of H (A) is larger than the threshold value. . At this time, the direction cumulative histogram HT (A) to which the threshold is applied is defined as number 4.

는 주어진 조직세포 영상에 존재하는 다수의 내강 영역들의 후보 씨앗점들이 될 수 있다. 임계치가 적용된 H_T (A)를 이용한 내강 영역들의 후보 씨앗점들

는 수5와 같이 정의된다. 수5의

는 수직지역최대커브들에 있는 포인트들의 집합이고,

는 수평지역최대커브들에 있는 포인트들의 집합이다. Therefore, we compute the partial derivative of H _T ( A ) to which the threshold is applied and find the vertical local maximum curve (LMC) and horizontal local maximum curve for x and y, respectively. The intersections of these two regional maxima

May be candidate seed points for multiple luminal regions present in a given tissue cell image. Candidate Seed Points of Lumen Regions Using Threshold-Applied H _T ( A )

Is defined as Number 5

Is the set of points in the vertical maximum curves,

Is the set of points in the horizontal maximum curves.

(A) and (b) shown in FIG. 6 are respectivelyxWowyon For H (AAre the partial derivatives, and the yellow areas (cells) represent the local maximum curve (LMC) in the vertical and horizontal directions. In FIG. 6C, the vertical and horizontal regional maximum curves LMC are indicated in yellow, and intersections (ie, candidate seed points) of the vertical and horizontal regional maximum curves LMC are indicated in red.

)을 이용하여 내강영역들에 대한 경계들을 추출할 수 있다. 표 2는 입력된 영상에 존재하는 내강영역의 경계를 식별하기 위한 알고리즘이다. Next, candidate seed points obtained using a cumulative histogram (

) To extract the boundaries of the lumens. Table 2 is an algorithm for identifying the boundary of the lumen area present in the input image.

는 입력된 영상에서 식별된 내강경계점 집합을 원소로 하는 집합으로 수 6과 같이 표현되며, 집합의 원소

는 i번째 내강영역의 경계점들의 집합을 의미한다.Results of the algorithm in Table 2

Is a set whose elements are the set of lumen boundary points identified in the input image, and are expressed as shown in Equation 6.

Denotes a set of boundary points of the i-th lumen region.

를 이용하여 상피세포들을 식별한다. 여기서, 상피세포는 내강을 피복하고 있는 세포로서 점액성 낭성종양을 식별하는데 중요한 역할을 한다. 상피세포 식별을 위해 상피세포 식별부(130)는 원본영상(조직세포 영상)에 대해 컬러문턱치처리(Color thresholding)를 수행하여 조직세포 영상에서 핵들을 식별한다(S131, S132). 상기 핵 식별 후, 상피세포 식별부(130)는 상기 식별된 핵들로부터 상피 세포의 핵들을 선별한다(S133). 상피세포를 식별하기 위한 전처리 과정은 도 7을 참조하여 설명한다. 먼저, 상피세포 식별부(130)는 주어진 조직세포 영상을 중간필터를 이용하여 불순물을 제거한다. 그리고, 상피세포 식별부(130)는 컬러문턱치처리(Color thresholding)를 통해 핵 이외의 불필요한 부분들을 제거하고, 핵의 홀을 메운다. 이어서, 상기 상피세포 식별부(130)는 분수선 기법[Chen Pan and Congxun Zheng, "Robust Color Image Segmentation based on Mean Shift and Marker-Controlled Watershed Algorithm," Conference on MLC, Vol.5, No.5, pp2752-2754, 2003 ]을 이용하여 핵들 분할하고 핵 집합 N을 생성한다.After extracting the boundary of the lumen area, the controller 160 controls the epithelial cell identification unit 130 to identify epithelial cells (S130). The epithelial cell identification unit 130 is a set of luminal boundary points obtained by the lumen area identification unit 120.

To identify epithelial cells. Here, epithelial cells are the cells that cover the lumen and play an important role in identifying mucinous cystic tumors. For epithelial cell identification, the epithelial cell identification unit 130 performs color thresholding on the original image (tissue cell image) to identify nuclei in the tissue cell image (S131, S132). After the nuclear identification, epithelial cell identification unit 130 selects the nuclei of epithelial cells from the identified nuclei (S133). The pretreatment process for identifying epithelial cells will be described with reference to FIG. 7. First, the epithelial cell identification unit 130 removes impurities from a given tissue cell image using an intermediate filter. The epithelial cell identification unit 130 removes unnecessary portions other than the nucleus through the color thresholding and fills the holes of the nucleus. Subsequently, the epithelial cell identification unit 130 uses a fractional technique [Chen Pan and Congxun Zheng, “Robust Color Image Segmentation based on Mean Shift and Marker-Controlled Watershed Algorithm,” Conference on MLC, Vol. 5, No. 5, pp2752. -2754, 2003] to split the nuclei and generate a nuclear set N.

를 피복하고 상피세포의 핵

을 식별한다.

는 다음의 수 8과 같이 정의된다.

Specific lumen from N

Epithelial cell nucleus

Identifies

Is defined as:

delete

를 구성하는 원소

들의 개수이고,

은 세포핵

의 중심점이다. 그러나 위와 같이 내강의 가장 가까운 상피세포들

들을 식별하기 위해서는

경계 포인트

와

핵중심점

사이의 거리를 구하고 오름차순으로 정렬한다. 이러한 과정은 매우 많은 비용을 수반한다. 예를 들어, 주어진 조직세포 영상의 경계포인트들의 개수가 m개이고, 식별된 핵의 수가 n 개인경우 거리 계산을 위해 O(mn)의 비용과 가까운 핵을 선별하기 위해 O(mnlog₂n)의 정렬비용이 든다. Where m is

Elements that make up

Number of

Silver cell nucleus

Is the center point. But as above, the closest epithelial cells in the lumen

To identify them

Boundary Point

Wow

Nuclear

Find the distance between them and sort in ascending order. This process is very expensive. For example, if the number of boundary points in a given tissue cell image is m and the number of identified nuclei is n , then alignment of O (mnlog ₂ n) to select nuclei close to the cost of O (mn) for distance calculation. It costs

가 속한 격자셀(cell)과 인접 격자셀들로 검색 범위를 줄일 수 있다. 이때,

에 인접한 격자셀은 모두 9개 이지만,

에 속하는 인접한 경계점들

,

,

을 포함하는 접선을 이용하여 6개의 격자셀들로 검색 범위를 줄일 수 있다. 왜냐하면 상피세포의 핵들은 내강의 바깥영역에 위치하고 있기 때문이다. 도 8은 격자색인에 대한 개념 설명을 위한 도이다. 격자색인을 이용하였을 때 상피세포의 핵들을 식별하는 비용은 다음과 같다. 먼저 격자색인비용은 O(n)이다. 격자색인을 이용한 상피세포핵 식별 시 수반되는 비용을 계산하기 위해서는 핵들이 주어진 조직세포 영상에 균등하게 분포한다고 가정한다. 즉, #grid 를 분할된 격자셀들의 개수이고, n을 식별된 핵의 개수라 했을 때, 한 격자셀에 포함되는 평균적인 핵의 개수는 n/#grid 이다. 이때, 격자색인을 이용한 거리 계산 비용은 O(mn/g)이 들며, 오름차순으로 정렬하는 비용은 O(mn/g log₂(n/g))이 든다. 이는 격자색인 방법을 이용하였을 때, 거리 계산 비용과 정렬비용에 있어서 각각

와

정도 더 효율적이다. 이는 격자셀의 크기가 작을수록 더 효율적이다. 아래 표 3은 상피 세포의 핵들을 식별하기 위한 절차(알고리즘)이다.In the present invention, in order to reduce such a cost, a given tissue cell image is divided into grids and the grids are indexed by grids. Grid index is the boundary point of the lumen

The search range may be reduced to the grid cells to which the cell belongs and the adjacent grid cells. At this time,

There are 9 grid cells adjacent to,

Adjacent boundary points belonging to

,

,

The search range may be reduced to six grid cells using a tangent line including. Because the nuclei of epithelial cells are located outside the lumen. 8 is a diagram for explaining a concept of the lattice index. The cost of identifying nuclei of epithelial cells using the lattice index is as follows. First, the lattice index cost is O (n). To calculate the costs involved in identifying epithelial cell nuclei using a lattice index, it is assumed that the nuclei are evenly distributed over a given tissue cell image. That is, when #grid is the number of grid cells divided and n is the number of identified nuclei, the average number of nuclei included in one grid cell is n / # grid. At this time, the distance calculation cost using the grid index is O (mn / g), the cost of sorting in ascending order is O (mn / g log ₂ (n / g)). When the grid color method is used, the distance calculation cost and the alignment cost respectively

Wow

Is more efficient. This is more efficient with smaller grid cells. Table 3 below is a procedure (algorithm) for identifying nuclei of epithelial cells.

The present invention automatically extracted nuclei of epithelial cells through the two processes described above.

의 점

에서의 접선을 T라 하고,

에서 이 접선을 T라 하고,

에서 이 접선과 직교하는 직선을 G라 할 때, 상피세포의 세포질 길이(cytoplasmlength)는

와

의

사이의 수직 거리로 정의된다. 세포질길이에 대한 수학적인 정의는 수 9와 같다. 도 9의 (a)는 본 발명에서 제안하는 상피세포의 세포질 길이를 측정하는 방법을 도시하며, (b)는 실제 조직세포 영상에 적용된 예를 도시한다.The controller 160 controls the feature extractor 140 to extract existing morphological features and new features specific to the mucinous cystic tumor from the extracted nuclei of the epithelial cell (S140). Existing morphological features include Area, Perimeter, Width, Height, Major axis, Minor axis, Angle, Circularity, Skewness, Feret Angle, MinFeret, Roundness, and Solidity. A detailed description of these features can be found in WN Street, WH Wolberg, OL Mangasarian, Nuclear feature extraction for breast tumor diagnosis, IS & T / SPIE 1993 International Symposium on Electronic Imaging: Science and Technology 1905 (1993) 861-870, San Jose , CA], Allen, T., 1997. Particle Size Measurement. Powder Sampling and Particle Size Measurement, Vol. 1, 5th Edition. Chapman & Hall, London, 525 pp. In the present specification, only new feature extraction methods specialized for mucinous cystic tumors will be described. The following features are measured to distinguish mucinous cystic tumors. First is the cytoplasmic length of the epithelial cells covering the lumen. The second is the presence of mucus, which measures the appearance of mucus, ring cells or goblet cells in epithelial cells. These two morphological features play an important role in the diagnosis of mucinous cystic tumors. First, the cytoplasmic length of epithelial cells can be obtained by measuring the vertical length from the boundary of the lumen to the epithelial cell nucleus. In other words,

Point of

The tangent to is called T,

This tangent is called T

When the straight line orthogonal to this tangent line is called G, the cytoplasm length of epithelial cells is

Wow

of

Is defined as the vertical distance between. The mathematical definition of cellular length is shown in Figure 9. Figure 9 (a) shows a method for measuring the cytoplasmic length of the epithelial cells proposed in the present invention, (b) shows an example applied to the actual tissue cell image.

와

의 위치정보를 이용하여 세포핵과 내강(lumen) 경계 사이에 있는 세포질을 관심영역으로 설정하고 나머지 부분들을 제거한다(도 10의 (b) 참고). 이후 문턱치처리을 적용하고, 방향누적히스토그램을 적용하여 다수개의 점액 부위들을 식별한다. 도 10의 (c)는 식별된 점액 부분을 보여주고 있다. The second feature proposed by the present invention is to confirm the presence of mucus in the cytoplasm. Mucus appears transparent in the cytoplasm or in the form of ring or goblet cells due to intracellular mucus fear [Kyungji Lee, et al., Fine Needle Aspiration Cytology of Mucinous Cystic Carcinoma of the Pancreas-A Case Report-, The Korean Journal of Cytopathology, 16 (2): 88-92, 2005]. The part indicated by a circle in Fig. 10 (a) is a cytoplasmic part containing mucus. To identify whether the cytoplasm is mucus

Wow

Using the location information of the cytoplasm between the nucleus and the lumen boundary to set the region of interest and remove the remaining parts (see Figure 10 (b)). After that, a threshold treatment is applied and a directional cumulative histogram is applied to identify a plurality of mucus sites. FIG. 10C shows the identified mucus portion.

The controller 160 controls the diagnosis unit 150 to detect a mucus cystic tumor from a given tissue cell image. The diagnosis unit 150 learns an SVM using the extracted morphological features, and diagnoses a mucus cystic tumor using the learned SVM.

Next, the experiment applying the mucinous cystic tumor diagnosis method proposed in the present invention will be described. In order to verify the influence of the features presented in the present invention, the experiments were carried out mucus on normal tissues and tissues containing mucinous cystic tumors in each case with and without new features. The classification performance of cystic tumors was evaluated. 100 pancreatic tissue cell samples used in the experiment were plated on a diagnostic slide by hematoxylin and eosin staining, and then 200X digital images were acquired using a slide scanner device, "ScanScope CS System". 40 of these are normal and 60 are tissue cell images of mucinous cystic tumors.

In order to measure how much the proposed features contribute to the improvement of the classification ability, we studied the classification model SVM with the existing classification model and the proposed features. In this experiment, because the amount of sample was not enough, the classification performance was evaluated by bootstrap resembling technique. Sampling was performed independently 10 times, and the final performance was measured as the average of the results. The classification accuracy was improved by 60% when using the existing morphological features and by 85% when the proposed features were added. The 2 × 2 confusion matrix for evaluating diagnostic capability as a diagnostic system is as follows. Table 5 shows the existing features [W.N. Street, W.H. Wolberg, O. L. Mangasarian, Nuclear feature extraction for breast tumor diagnosis, IS & T / SPIE 1993 International Symposium on Electronic Imaging: Science and Technology 1905 (1993) 861-870, San Jose, CA], Allen, T., 1997. Particle Size Measurement. Powder Sampling and Particle Size Measurement, Vol. 1, 5th Edition. Chapman & Hall, London, 525 pp] is the evaluation results when using only, Table 6 is the evaluation results when the features presented in the present invention are added.

As can be seen from the results, by adding the features presented in this paper, the performance was improved in both sensitivity and specificity as well as improvement of classification ability. Sensitivity is defined as the ratio of positive to positive among patients with disease.

Specificity is defined as shown in Figure 11 as the ratio judged to be negative among patients without disease.

As described above, although described with reference to a preferred embodiment of the present invention, those skilled in the art various modifications of the present invention without departing from the spirit and scope of the invention described in the claims below Or it may be modified.

110: image input unit 120: lumen area identification unit
130: epithelial cell identification unit 140: feature extraction unit
150: diagnostic unit 160: control unit
170: user input unit 180: display unit

Claims (11)

An image input unit for providing a tissue cell image,
A lumen region identification unit for identifying a plurality of lumen regions from the tissue cell image;
An epithelial cell identification unit for identifying epithelial cells from the identified lumen area;
A feature extraction unit for extracting morphological features for cystic tumor diagnosis from the identified epithelial cells;
Mucus cystic tumor diagnosis device, characterized in that it comprises a diagnostic unit for detecting a mucus cystic tumor from the tissue cell image using the SVM learned about the extracted features. The method of claim 1, wherein the lumen area identification unit,
The mucus cystic is characterized by preprocessing the tissue cell image, extracting candidate seed points for identifying the lumen region from the preprocessed image, and detecting the boundary of the lumen region using the extracted candidate seed points. Tumor diagnostic device. The method of claim 2, wherein the pretreatment is
The tissue cell image is filtered through an average filter to remove impurities in the lumen region, convert the filtered image into a gray image, and reduce the environmental factors for the image through background correction. A mucus cystic tumor diagnosis device, characterized in that for generating a binary image by applying the maximum entropy threshold to the background corrected image. The method of claim 2, wherein the extraction of the candidate seed points,
Apparatus for detecting a mucinous cystic tumor, characterized in that for extracting the candidate seed points of a plurality of lumen areas by applying a direction cumulative histogram to the pre-processed image. The method of claim 2, wherein the boundary detection of the lumen area,
Apparatus for diagnosing a mucinous cystic tumor characterized in that the boundary of the lumen area is extracted by applying a region growth method to the candidate seed points. The method of claim 1, wherein the epithelial cell identification unit,
And detecting nuclei by preprocessing the tissue cell image, and extracting nuclei of epithelial cells using the detected nuclei and the identified lumen regions. The method of claim 6, wherein the nuclear detection,
The tissue cell image is detected by nucleus through color threshold processing, and mucus cystic tumor diagnosis device, characterized in that for generating a nucleus by dividing the detected nucleus. The method of claim 7, wherein the nuclear extraction of epithelial cells,
Mucus cystic tumor diagnosis device, characterized in that to identify the nucleus of the epithelial cell using the generated nucleus set and the boundary of the lumen area. The method of claim 1, wherein the feature extraction unit,
An apparatus for diagnosing a mucinous cystic tumor characterized in that it extracts features such as the presence of mucus, which measures the cytoplasmic length of epithelial cells covering the lumen and whether mucus, ring cells or goblet cells are present in the epithelial cells. The method of claim 1, wherein the diagnostic unit,
A mucinous cystic tumor diagnosis device, characterized in that for detecting a mucinous cystic tumor from the tissue cell image using the SVM classification model trained on the extracted features. delete

Images