KR101527604B1 - A Method for Grading Pancreatic Ductal Adenocarcinomas Based On Morphological Features - Google Patents

Abstract

The present invention relates to a method for diagnosing pancreatic adenocarcinoma based on a morphologic feature of extracting morphological features from a given tissue cell image and diagnosing pancreatic adenocarcinoma based on the extracted features, ; A second step of constructing a diagnostic model using a map-based classifier; And a third step of diagnosing pancreatic adenocarcinoma of the tissue cell image using the diagnostic model, wherein the diagnostic model provides a configuration for determining the presence or degree of pancreatic adenocarcinoma with respect to input morphological characteristics.
According to the diagnosis method of pancreatic adenocarcinoma grade as described above, the morphological features are extracted and the diagnosis model is learned by the extracted features, so that the pancreatic adenocarcinoma can be accurately discriminated and each stage can be distinguished.

Description

TECHNICAL FIELD The present invention relates to a method for classifying pancreatic adenocarcinoma based on morphological characteristics,

The present invention also relates to a method for diagnosing pancreatic adenocarcinoma based on a morphological feature that extracts morphological features from a given tissue cell image and diagnoses pancreatic adenocarcinoma based on the extracted features.

Traditionally, pathologic diagnosis is manually reviewed for tissue slides given through an optical microscope and subjectively diagnosed by the experience and knowledge of the diagnosis. This traditional pathological diagnosis has several problems. First, since the screening process through the optical microscope proceeds at a high magnification, much time and effort are required [Non-Patent Document 1] and [Non-Patent Document 2]. The individual's ability to diagnose also has a decisive impact on the actual diagnosis. That is, since subjective diagnosis is performed instead of objectively diagnosed by quantitative feature analysis, various opinions can be shown to doctors about the same symptoms [Non-Patent Document 3] - [Non-Patent Document 5].

The emergence of digital pathology is changing the pathological diagnosis environment. Digital pathology digitized a glass slide, enabling objective diagnosis through convenient screening environment and quantitative feature analysis [Non-Patent Document 6] - [Non-Patent Document 8]. Nevertheless, pathologic diagnosis through screening of tissue images requires a lot of time and labor because it is performed passively. Therefore, there is still a problem due to subjective judgment.

To overcome these problems, there is an increasing number of studies to combine digital pathology with computer aided detection (CAD) technology. Many automated diagnosis-related studies cover not only the diagnosis of tumors but also the grading of the tumors found. Pathological grading is very valuable for identifying the progression of the disease and for proper prescribing to the patient [Non-Patent Document 9]. In fact, pathologists are known to do a lot of training in sorting the arms for the best prescription of the patient [Non-Patent Literature 7]. Many studies are now being conducted on prostate cancer and breast cancer.

First, most of the CAD studies related to prostate cancer diagnosis are based on the Gleason grading system developed by Gleason [Non-Patent Document 10]. Tabesh et al. [Non-Patent Document 11] proposed a machine learning-based automation system that diagnoses prostate cancer and distinguishes between low and high based on Gleason grading. They are used to distinguish color, texture, and morphological features from the historical (object) level and the global (image) level of a given pathology image for the diagnosis and staging of tumors (Gaussian, k-NN, SVM) using morphometric features extracted from them.

Naik et al. [Non-Patent Document 12] proposed a system for classifying intermediate Gleason grades. They use a Bayesian classifier that uses low-level information to identify the candidate gland region and use the Empirical domain information to identify the true falsehood identified as a gland (false positive) regions were excluded. Later, morphological features were extracted from the identified glands, and Gleason grade 3, Gleason grade 4, and benign were classified via SVM. Huwang and Lee [12] extracted features for Gleason grading using differential box-counting and entropy-based fractal dimension estimation, and Bayesina, k- NN and SVM classifiers were used to classify the prostate cancer images into four classes. In addition, many studies have been conducted to diagnose prostate tumor (post tumor) and to automate Gleason grading system [Non-Patent Document 14] - [Non-Patent Document 17].

Next, CAD studies related to breast cancer are as follows. Anderson et al. [Non-Patent Document 18] studied the problem of distinguishing between ductal hyperplasia (DH), which is benign in non-invasive breast lesions, and DCIS (ductal carcinoma in situ), which is malignant. They used the knowledge-guided machine vision in this study to automatically segment breast ducts and measure the duct pattern of proliferative lesions that are important for distinguishing between DH and DCIS We have proposed measures such as duct cribriformity and architectural complexity.

Bilgin et al. [Non-patent document 19] diagnosed breast cancer based on graph theoretical techniques. They segmented a given tissue cell image using a k-means algorithm and generated a cell-graph using the cell's positional coordinates. SVM models that classify a given tissue image as benign, invasive, or non-invasive are learned by quantitative metrics computed from the generated cell-graph. Basavanhally et al. [Non-patent document 20] proposed a system for identifying and grading the extent of lymphocytic infiltration known as a viable prognostic indicator of HER2 + breast cancer. They first identified lymphocytes using region grwowing and the Makov random field algorithm. We then extract the architectural features using the identified lymphocytes and use the SVM classifier learned by those features to determine the extent of LI as low, midium, high Respectively. There have been many studies for diagnosis and grading of breast cancer [Non-Patent Document 21] - [Non-Patent Document 24].

Many pathologic CAD studies are devoted to devising pathological features for each disease and methods for quantitatively measuring them. Because pathologic diagnosis has various diagnostic methods depending on the type of lesion and disease. In addition to the pathological diagnostic CAD for breast and prostate tumors, there are currently available colonic (non-patent document 25), bladder (non-patent document 26), neuroblastoma (non-patent document 27) 28] [Non-Patent Document 29], Follicular lymphoma [Non-Patent Document 30] [Non-Patent Document 31] Some studies are being conducted on tumors.

Pathology Diagnostic CAD research is still in its infancy and only a few are targeted. There are many methods for diagnosing various tumors. Therefore, more research on pathological CAD is needed.

[Non-Patent Document 1] M. N. Gurcan, L. E. Boucheron, A. Can, A. Madabhushi, N. M. Rajpoot, and B. Yener, "Histopathological Image Analysis: A Review," Biomedical Engineering, IEEE Reviews in, vol. 2, pp. 147 -171, 2009. [Non-Patent Document 2] A. E. Tutac, D. Racoceanu, T. Putti, W. Xiong, W.-K. Leow, and V. Cretu, "Knowledge-Guided Semantic Indexing of Breast Cancer Histopathology Images," in International Conference on BioMedical Engineering and Informatics, 2008. BMEI 2008, 2008, vol. 2, pp. 107 -112. [Non-Patent Document 3] Observer variation in histopathological diagnosis and grading of cervical intraepithelial neoplasia (JS) , "BMJ, vol. 298, no. 6675, pp. 707-710, Mar. 1989. [Non-Patent Document 4] A. Andrion, C. Magnani, PG Betta, A. Donna, F. Mollo, M. Scelsi, P. Bernardi, M. Botta, and B. Terracini, "Malignant mesothelioma of the pleura: interobserver variability, "J. Clin. Pathol., Vol. 48, no. 9, pp. 856-860, Sep. 1995. [Non-Patent Document 5] C. Demir and B. Yener, "Automated Cancer Diagnosis Based on Histopathological Systematic Images: A Systematic Survey." Rensselaer Polytechnic Institute, 2005. [Non-Patent Document 6] S. Al-Janabi, A. Huisman, and P. J. Van Diest, "Digital pathology: current status and future perspectives," Histopathology, vol. 61, no. 1, pp. 1-9, 2012. [Non-Patent Document 7] A. Madabhushi, "Digital pathology image analysis: opportunities and challenges," Imaging in Medicine, vol. 1, no. 1, pp. 7-10, Oct. 2009. [Non-Patent Document 8] RS Weinstein, AR Graham, LC Richter, GP Barker, EA Krupinski, AM Lopez, KA Erps, AK Bhattacharyya, Y. Yagi, and JR Gilbertson, "Overview of telepathology, virtual microscopy, : prospects for the future, "Human Pathology, vol. 40, no. 8, pp. 1057-1069, Aug. 2009. [Non-Patent Document 9] J. I. Epstein and G. J. Netto, Biopsy Interpretation of the Prostate, Fourth. Lippincott Williams & Wilkins, 2007. [Non-Patent Document 10] D. F. Gleason and G. T. Mellinger, "Prediction of prognosis for prostatic adenocarcinoma by combined histological grading and clinical staging," J. Urol., Vol. 111, no. 1, pp. 58-64, Jan. 1974. [Non-Patent Document 11] A. Tabesh, M. Teverovskiy, H.-Y. Pang, V. P. Kumar, D. Verbel, A. Kotsianti, and O. Saidi, "Multiparty Prostate Cancer Diagnosis and Gleason Grading of Histological Images," IEEE Transactions on Medical Imaging, vol. 26, no. 10, pp. 1366 -1378, Oct. 2007. [Non-Patent Document 12] S. Naik, S. Doyle, M. Feldman, J. Tomaszewski, and A. Madabhushi, Gland Segmentation and Computerized Gleason Grading of Prostate Histology by Integrating Low-, High-level and Domain Specific Information. MIAAB Workshop:, 2007. [Non-Patent Document 13] P.-W. Huang and C.-H. Lee, "Automatic Classification for Pathological Prostate Images Based on Fractal Analysis," IEEE Transactions on Medical Imaging, vol. 28, no. 7, pp. 1037-1050, Jul. 2009. [Non-patent Document 14] A. W. Wetzel, "Evaluation of prostate tumor grades by content-based image retrieval," 1999, vol. 3584, pp. 244-252. [Non-Patent Document 15] S. Doyle, M. Hwang, K. Shah, A. Madabhushi, M. Feldman, and J. Tomaszeweski, "AUTOMATED GRADING OF PROSTATE CANCER USING ARCHITECTURAL AND TEXTURAL IMAGE FEATURES, on Biomedical Imaging: From Nano to Macro, 2007. ISBI 2007, 2007, pp. 1284 -1287. [Non-Patent Document 16] S.-K. Tai, Y.-C. Wu, C.-Y. Li, Y.-J. Jan, and S.-C. Lin, "Computer-assisted Detection and Grading of Prostatic Cancer in Biopsy Image | www.libsearch.com - TEST," Mar-2010. [Online]. Available: http://www.libsearch.com/view/1046167. [Accessed: 19-Jul-2012]. [Non-Patent Document 17] Y. Peng, Y. Jiang, S.-T. Chuang, and X. J. Yang, "Computer-aided detection of prostate cancer on tissue sections," Appl. Immunohistochem. Mol. Morphol., Vol. 17, no. 5, pp. 442-450, Oct. 2009. [Non-patent Document 18] NH Anderson, PW Hamilton, PH Barnes, D. Thompson, R. Montironi, and JM Sloan, "COMPUTERIZED SCENE SEGMENTATION FOR THE DISCRIMINATION OF ARCHITECTURAL FEATURES IN DUCTAL PROLIFERATIVE LESIONS OF THE BREAST, of Pathology, vol. 181, no. 4, pp. 374-380, 1997. [Non-Patent Document 19] C. Bilgin, C. Demir, C. Nagi, and B. Yener, "Cell-Graph Mining for Breast Tissue Modeling and Classification," 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society , 2007. EMBS 2007, 2007, pp. 5311 -5314. [Non-patent document 20] AN Basavanhally, S. Ganesan, S. Agner, JP Monaco, MD Feldman, JE Tomaszewski, G. Bhanot, and A. Madabhushi, "Computerized image- based detection and grading of lymphocytic infiltration in HER2 + breast cancer histopathology, "IEEE Trans Biomed Eng, vol. 57, no. 3, pp. 642-653, Mar. 2010. [Non-patent Document 21] S. Doyle, S. Agner, A. Madabhushi, M. Feldman, and J. Tomaszewski, "Automated grading of breast cancer histopathology using spectral clustering with textural and architectural image features, on Biomedical Imaging: From Nano to Macro, 2008. ISBI 2008, 2008, pp. 496 -499. [Non-Patent Document 22] W. H. Wolberg, W. N. Street, D. M. Heisey, and O. L. Mangasarian, "Computer-derived nuclear features distinguish malignant from benign breast cytology," Hum. Pathol., Vol. 26, no. 7, pp. 792-796, Jul. 1995. [Non-Patent Document 23] R. Fernandez-Gonzalez, T. Deschamps, A. Idica, R. Malladi, and C. O. de Solorzano, "Automatic segmentation of histological structures in mammal gland tissue sections," Journal of Biomedical Optics, vol. 9, no. 3, pp. 444-453, May 2004. [Non-Patent Document 24] W. H. Wolberg, W. N. Street, and O. L. Mangasarian, Breast Cytology Diagnosis Via Digital Image Analysis. 1993. [Non-Patent Document 25] A. N. Esgiar, R. N. G. Naguib, B. S. Sharif, M. K. Bennett, and A. Murray, "Fractal analysis in the detection of colonic cancer images," IEEE Transactions on Information Technology in Biomedicine, vol. 6, no. 1, pp. 54 -58, Mar. 2002. [Non-Patent Document 26] HK Choi, T. Jarkrans, E. Bengtsson, J. Vasko, K. Wester, PU Malmstrom, and C. Busch, "Image analysis based on grading of bladder carcinoma. graph based methods and their reproducibility, "Anal Cell Pathol, vol. 15, no. 1, pp. 1-18, 1997. [Non-Patent Document 27] M. N. Gurcan, J. Kong, O. Sertel, B. B. Cambazoglu, J. Saltz, and U. Catalyurek, "Computerized pathological image analysis for neuroblastoma prognosis," AMIA Annu Symp Proc. 304-308, 2007. [Non-patent Document 28] O. Sertel, J. Kong, H. Shimada, UV Catalyurek, JH Saltz, and MN Gurcan, "Classification of Stromal Development," Pattern Recognition on Neuroblastoma on Whole- vol. 42, no. 6, pp. 1093-1103, Jun. 2009. [Non-Patent Document 29] BB Cambazoglu, O. Sertel, J. Kong, J. Saltz, MN Gurcan, and UV Catalyurek, "Efficient processing of pathological images using the grid: computer-aided prognosis of neuroblastoma," in Proceedings of the 5th IEEE workshop on Challenges of large applications in distributed environments, New York, NY, USA, 2007, pp. 35-41. [Non-Patent Document 30] K. Belkacem-Boussaid, M. Pennell, G. Lozanski, A. Shana'ah, and MN Gurcan, "Effect of pathologist agreement on evaluating a computer-aided assisted system: Recognizing centroblast cells in follicular lymphoma cases, "in < / RTI > 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2010, pp. 1411 -1414. [Non-Patent Document 31] O. Sertel, G. Lozanski, A. Shana'ah, and MN Gurcan, "Computer-aided detection of centroblasts for follicular lymphoma grading using adaptive likelihood-based cell segmentation," IEEE Trans Biomed Eng, Vol. . 57, no. 10, pp. 2613-2616, Oct. 2010. [Non-Patent Document 32] G. Klappeel, R. H. Huruban, D. S. Longnecker, G. Adler, S. E. Kern, and T. J. Partanen, "Word Health Organization Classification of Tumors in Ductal Adenocarcinoma of the Pancreas". Pathology and Genetics of Tumors of the Digestive System, 3rd ed., International Agency for Research on Cancer, 2000. [Non-Patent Document 33] RH Hruban, P. Boffetta, C. Iacobuzio-Donahue, Y. Kato, SE Kern, DS Klimstra, G. Kloppel, A. Maitra, GJA Offerhaus, and MB Pitman, "Ductal adenocarcinoma of the pancreas, "in WHO Classification of Tumors of the Digestive System, 4th ed., vol. 3, International Agency for Research on Cancer, 2010, pp. 217-224. [Non-Patent Document 34] Sixty hundred, LB Grochow, and RA Abrams, " Six hundreds, " fifty consecutive pancreaticoduodenectomies in the 1990s: pathology, complications, and outcomes., "Ann Surg, vol. 226, no. 3, pp. 248-260, Sep. 1997. [Non-Patent Document 35] R. Adams and L. Bischof, "Seeded region growing," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, no. 6, pp. 641-647, Jun. 1994. [Non-Patent Document 36] J. W. Song, J. H. Lee, S. Park, S. J. Chun, and J. H. Choi, "Automated Epithelial Tissue Detection Using Mucinous Cystadenoma Classification System," International Conference on Machine Vision, 2010. [Non-Patent Document 37] K. Rodenacker and E. Bengtsson, "A feature set for cytometry on digitized microscopic images," Anal Cell Pathol, vol. 25, no. 1, pp. 1-36, 2003. [Non-Patent Document 38] P. K. Sahoo, S. Soltani, A. K. C. Wong, and Y. C. Chen, "A survey of thresholding techniques," Comput. Vision Graph. Image Process., Vol. 41, no. 2, pp. 233-260, Feb. 1988. [Non-Patent Document 39] A. Z. Chitade and S. K. Katiyar, "Color based image segmentation using K-means clustering," International Journal of Engineering Science and Technology, vol. 2, no. 10, pp. 5319-5325, Oct. 2010. [Non-Patent Document 40] R. C. Gonz ㅱ lez and R. E. Woods, Digital Image Processing. Prentice Hall, 2008. [Non-Patent Document 41] C. Pan, C.-X. Zheng, and H.-J. Wang, "Robust color image segmentation based on mean shift and marker-controlled watershed algorithm," in 2003 International Conference on Machine Learning and Cybernetics, 2003, vol. 5, pp. 2752 - 2756 Vol.5. [Non-Patent Document 42] R. J. Hyndman and A. B. Koehler, "Another Look at Measures of Forecast Accuracy," International Journal of Forecasting, vol. 22, no. 4, pp. 679-688, Oct. 2006. [Non-Patent Document 43] F.-L. Chung, T.-C. Fu, V. Ng, and R. W. P. Luk, "An evolutionary approach to pattern-based time series segmentation," IEEE Transactions on Evolutionary Computation, vol. 8, no. 5, pp. 471 - 489, Oct. 2004. [Non-Patent Document 44] "Efficient Time Series Matching by Wavelets," in Proceedings of the 15th International Conference on Data Engineering, Washington, DC, USA, 1999, p. 126-. [Non-Patent Document 45] M. Kosmahl, U. Pauser, M. Anlauf, and G. Klul ppel, "Pancreatic ductal adenocarcinomas with cystic features: Pathol., Vol. 18, no. 9, pp. 1157-1164, Sep. 2005. [Non-Patent Document 46] M. Kallergi, "Evaluation Strategies for Medical-Image Analysis and Processing Methodologies," in Medical Image Analysis Methods, vol. 18, L. Costaridou, Ed. CRC Press, 2005. [Non-Patent Document 47] JK Udupa, VR LaBlanc, H. Schmidt, C. Imielinska, PK Saha, GJ Grevera, Y. Zhuge, LM Currie, P. Molholt, and Y. Jin, "Methodology for Evaluating Image- , pp. 266-277, May 2002. [Non-Patent Document 48] T. Allen, Particle Size Measurement: Volume 1: Powder sampling and particle size measurement. Springer, 1996. [Non-Patent Document 49] K. Rodenacker and E. Bengtsson, "A feature set for cytometry on digitized microscopic images," Analytical Cellular Pathology, vol. 25, no. 1, pp. 1-36, Jan. 2003. [Non-Patent Document 50] W. Rasband, "ImageJ," 2012-1997. [Online]. Available: http://imagej.nih.gov/ij/. [Accessed: 25-Aug-2012]. [Non-Patent Document 51] V. N. Vapnik, Statistical Learning Theory, 1st ed. Wiley-Interscience, 1998. [Non-Patent Document 52] C. J. C. Burges, "A Tutorial on Support Vector Machines for Pattern Recognition," Data Min. Knowl. Discov., Vol. 2, no. 2, pp. 121-167, Jun. 1998. [Non-patent Document 53] C.-C. Chang and C.-J. Lin, "LIBSVM: A library for support vector machines," ACM Trans. Intell. Syst. Technol., Vol. 2, no. 3, pp. 27: 1-27: 27, May 2011. [Non-Patent Document 54] D. L. Verbyla and J. A. Litvaitis, "Resampling methods for evaluating classification accuracy of wildlife habitat models," Environmental Management, vol. 13, no. 6, pp. 783-787, Nov. 1989. [Non-Patent Document 55] H. J. Ader, G. J. Mellenbergh, and D. J. Hand, Advising on research methods: a consultant's companion. Huizen, Netherlands: Johannes van Kessel Pub, 2008. [Non-Patent Document 56] J. A. Swets, "Measuring the accuracy of diagnostic systems," Science, vol. 240, no. 4857, pp. 1285-1293, Jun. 1988. [Non-Patent Document 57] M. Greiner, D. Pfeiffer, and R. D. Smith, "Principles and Practical Application of the Receiver-Operating Characteristic Analysis for Diagnostic Tests," Vet. Med., Vol. 45, no. 1-2, pp. 23-41, May 2000. [Non-Patent Document 58] L. J. Williams and H. Abdi, "Fisher's least significance difference (LSD) test," in Encyclopedia of Research Design, 1st ed., SAGE Publications, 491-494. [Non-Patent Document 59] R.L. Graham, "An efficient algorithm for determining the convex hull of a finite planar set", Info. Proc. Lett. 1, 132-133 (1972).

It is an object of the present invention to provide a method for diagnosing pancreatic adenocarcinoma based on morphological features extracted from a given tissue cell image and diagnosing pancreatic adenocarcinoma based on extracted features . In other words, to identify PDACs (pancreatic ductal adenocarcinomas, pancreatic adenocarcinomas or pancreatic adenocarcinomas) and to distinguish their stage, new morphological features for diagnosis and grading of PDACs are extracted, And to provide a method of diagnosing the grade of line pressure.

Particularly, it is an object of the present invention to provide a method for diagnosing pancreatic adenocarcinoma by segmenting luminal and epithelial cells and identifying the pancreatic adenocarcinoma from the identified objects in a tube composed of luminal and epithelial cells, And to provide a method for diagnosing pancreatic adenocarcinoma grade based on morphological features extracting morphological features designed for the diagnosis of pancreatic adenocarcinoma.

In the diagnosis of pancreatic adenocarcinoma, the most common site is the lumen of the lumen and epithelium. Therefore, the present invention seg- ments each object (lumen, epithelial cells, non-epithelial cells) from a given tissue IV, extracts morphological features for diagnosis, and then uses the extracted features to learn the SVM model for diagnosis And to provide a method for diagnosing pancreatic adenocarcinoma which distinguishes diagnosis and stage of pancreatic adenocarcinoma.

In order to accomplish the above object, the present invention relates to a method for diagnosing pancreatic adenocarcinoma based on morphological characteristics, which extracts morphological characteristics from tissue cell images and diagnoses pancreatic adenocarcinoma. step; A second step of constructing a diagnostic model using a map-based classifier; And a third step of diagnosing pancreatic adenocarcinoma of the tissue cell image using the diagnostic model, wherein the diagnostic model determines the presence or degree of pancreatic adenocarcinoma with respect to the inputted morphological characteristics.

According to another aspect of the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on morphological characteristics, comprising the steps of: (a) segmenting a lumen area in a tissue cell image; (b) identifying nuclei in the tissue cell image; (c) dividing the identified nuclei into epithelial and non-epithelial cell nuclei; (d) extracting a luminal feature for the lumenal region; And (e) extracting epithelial cell characteristics from the epithelial cell nuclei.

Further, the present invention is characterized in that, in the diagnostic method of pancreatic adenocarcinoma grade based on morphological characteristics, the diagnostic model uses SVM (Support Vector Machine) classifier.

Further, the present invention provides a method for diagnosing pancreatic adenocarcinoma grade based on morphological characteristics, wherein in the second step, learning of the classifier is performed with learning data composed of pairs of classes of morphological features (hereinafter, referred to as characteristic vectors) Thereby constructing the diagnostic model.

Further, the present invention provides a method for diagnosing pancreatic adenocarcinoma based on morphological characteristics, wherein in the step (a), the lumenal region is identified and segmented by a seed region positive (SGR) method.

(A1) applying a median filtering and a background correlation algorithm to the tissue cell image, and applying a maximum entropy threshold to obtain a binary image of the pancreas adenocarcinoma based on a morphological feature, ; (a2) generating a direction cumulative map from the binary image; (a3) selecting local maximum points of the direction accumulation map as candidate seed points; And (a4) finding the boundary of the lumenal region using the candidate seed points as an initial parameter of the SRG method.

According to another aspect of the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic, wherein a candidate seed point is obtained by applying a threshold process to the direction cumulative map in the step (a3).

According to another aspect of the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic, comprising the steps of: (a) subjecting a tissue cell image to a threshold processing based on a k- .

According to another aspect of the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic, wherein in step (c), nuclei adjacent to the border of the lumenal region are divided into epithelial cell nuclei, And the remainder is divided into non-epithelial cell nuclei.

Further, the present invention is characterized in that in the method for diagnosing pancreatic adenocarcinoma grade based on morphological characteristics, the set N ^E of epithelial cell nuclei is obtained by the following [Expression 1].

[Equation 1]

Where B ^O is the boundary of the lumen region, centroid (-) is the function that returns the center point of the identified nucleus n, and distance (·, ·) is the function that returns the Euclidian distance between two given points.

According to another aspect of the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic, comprising the steps of: (d1) obtaining an ideal boundary from a boundary of the lumenal region (hereinafter referred to as an original boundary); (d2) obtaining a dyspeakable amplitude signature using a distance of points perpendicular to each other at the original boundary and the ideal boundary; And (d3) extracting the luminal feature from the dyspeakable amplitude signature.

According to another aspect of the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic, wherein in the step (d1), the ideal boundary is obtained by scaling the convex shell of the original boundary to a size of an inner region of the original boundary do.

According to the present invention, there is provided a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic, wherein in the step (d2), the emissivity amplitude is expressed as a distance between points on an original boundary vertically positioned at an ideal boundary point, Is expressed as a distance that differs in sign depending on whether the point on the boundary is the outer boundary or the inner boundary of the ideal boundary.

The present invention also provides a method for diagnosing pancreatic adenocarcinoma grade based on morphologic characteristics, wherein in the step (d3), the dysplastic amplitude signature is a dystrophic amplitude of a circumferential point (hereinafter, an ideal boundary point) with respect to a circumferential length of the ideal boundary And the emissivity amplitude is a distance having a sign between points on an original boundary located vertically at the ideal boundary point.

Further, the present invention is characterized in that in the method for diagnosing pancreatic adenocarcinoma grade based on morphological characteristics, the above-mentioned dyssine amplitude is obtained by the dyssine amplitude function A (t) of the following [Expression 2].

[Equation 2]

Where t is an index variable indicating the order of the points at the ideal boundary and sgn _t (q) is a function representing the sign of the distance between vertex q ∈B ⁰ in p _t ∈ B ^I , Returns -1 if q is at the outer boundary of the point p _t , or -1 if it is at the inner boundary.

The present invention also provides a method of diagnosing a pancreatic adenocarcinoma grade based on morphological features, wherein the luminal feature comprises a root mean square (RMSAA) of the root mean square for the dendritic amplitudes, a sum of the change amounts at the inflection point of the dendritic amplitude signature (TSAV) (AtypiaRatio), and the number of releasable regions (#AppliaRegions).

The present invention also provides a method of diagnosing pancreatic adenocarcinoma based on a morphological characteristic, wherein the epithelial cell characteristic is characterized by being at least one of a cytoplasm length and a standard deviation (CytoplasmLength _SD ) of a cytoplasmic length of epithelial cells .

As described above, according to the morphological feature-based diagnosis method of pancreatic adenocarcinoma grade according to the present invention, the extracted morphological characteristics can accurately discriminate pancreatic adenocarcinoma and distinguish each stage.

In particular, the SVM model, which is a representative statistical classification technique, was evaluated and the classification performance was evaluated. As a result, the taxonomic features extracted by the present invention showed a classification accuracy of 94.38% in discriminating between normal and pancreatic adenocarcinoma. The classification accuracy of 97.03% is shown for the Grade 1 and Grade 2 staging.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a configuration diagram of an entire system for implementing the present invention; Fig.
Figure 2 is a histological image of a pancreatic adenocarcinoma of the pancreas adenocarcinoma, which is a subject of the present invention. (A) Positive correlation, (b) well-differentiated, (c) moderately differentiated, (d) Fig.
FIGS. 3 and 4 are schematic diagrams and flowcharts illustrating a method for diagnosing a pancreatic adenocarcinoma grade based on a morphological feature according to an embodiment of the present invention. FIG.
FIG. 5 is a flowchart illustrating a morphological feature extraction method for classifying pancreatic adenocarcinoma according to an embodiment of the present invention. FIG.
6 is an exemplary image of each process for explaining the lumen area identification process according to the present invention.
Figure 7 is an exemplary image of a set of identified nuclei according to the present invention.
8 is an algorithm for obtaining an epithelium-nucleus set according to the present invention.
9 is a table showing main notations according to the present invention.
Figure 10 is an image of the original luminal boundary and the ideal luminal boundary according to the present invention.
11 is an image and a graph showing the process of measuring the releasing-amplitude according to the present invention.
Figure 12 is a diagrammatic-amplitude signature for the image of Figure 1 according to the present invention.
Figure 13 is a graph showing a portion of the release signature and PIP for Class 2 of Figure 12 in accordance with the present invention.
FIG. 14 shows (a) grade 2 tissue cell image and (a) imaging of a dysplastic area according to the present invention.
FIG. 15 is an image (blue line) of CytoplasmLength measurement results of the epithelial cells of the positive correlation and grade 1 according to the present invention.
16 is a table showing information of a digital slide according to the experiment of the present invention.
17 is a table showing the number of obtained experimental images according to the experiment of the present invention.
18 is a table for morphological features used in the classification experiment according to the experiment of the present invention.
Figure 19 is a table of symbols and dimensions for the feature sets used in the classification experiments according to the experiments of the present invention.
20 is a table for the number of test data sets and learning data for learning and evaluating the SVM classifier according to the experiment of the present invention;
FIG. 21 is a table showing evaluation results for classifying positive correlation and PDAC for each feature set according to the experiment of the present invention; FIG.
FIG. 22 is a graph showing the comparison between the positive correlation for each characteristic Jipp and the classification accuracy for PDAC according to the experiment of the present invention.
23 is a graph comparing the ROC curve and the AUC value for the classifier of Case 1 according to the experiment of the present invention, wherein (a) ROC curve and AUC of CNF; (b) ROC curves and AUC values for CLF, PLF, and ALF; (c) ROC curves and AUC values for CEF, PEF, and AEF; (d) ROC curves and AUC values for CDF, PDF, and ADF; (e) ROC curves and AUC values for CTF, PTF, and ATF.
24 is an evaluation result table for distinguishing between class 1 and class 2 for each feature set according to the experiment of the present invention.
25 is a graph showing a comparison of classification accuracy for Grade 1 and Grade 2 for each characteristic Jeep according to the experiment of the present invention.
FIG. 26 is a graph comparing the ROC curve and the AUC value for the classifier of case 2 according to the experiment of the present invention, wherein (a) ROC curve and AUC of CNF; (b) ROC curves and AUC values for CLF, PLF, and ALF; (c) ROC curves and AUC values for CEF, PEF, and AEF; (d) ROC curves and AUC values for CDF, PDF, and ADF; (e) ROC curves and AUC values for CTF, PTF, and ATF.
FIG. 27 is a statistical table for F-test results and non-epithelial cell characteristics according to the experiment of the present invention.
Fig. 28 is a statistical table for the results of the F-test and the luminal characteristics according to the experiment of the present invention.
29 is a statistical table for the results of the F-test and the epithelial cell characteristics according to the experiment of the present invention.
FIG. 30 is a table for LSD-test and F-test of the characteristics of non-epithelial cells according to the experiment of the present invention.
31 is a table for LSD-test and F-test of the lumenal characteristics according to the experiment of the present invention.
32 is a table for LSD-test and F-test of epithelial cell characteristics according to the experiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the drawings.

In the description of the present invention, the same parts are denoted by the same reference numerals, and repetitive description thereof will be omitted.

First, the overall system configuration for carrying out the present invention will be described with reference to Fig.

As shown in FIG. 1, the method for diagnosing pancreatic adenocarcinoma based on morphological characteristics according to the present invention can be implemented in the program system 30 on the computer terminal 20. At this time, the program system is referred to as a pancreatic adenocarcinoma grade assessment system (30).

That is, the respective functions of the pancreatic adenocarcinoma grade diagnosis system 30 are implemented in a computer program and installed in the computer terminal 20, so that the user can input commands or the like through the input device of the computer terminal 20, Diagnostic results, and the like through the output device of the computer terminal 13. [ On the other hand, data required in the pancreatic adenocarcinoma grade diagnosis system 30 is stored in a storage space such as a hard disk of the computer terminal 20 and used.

The data input to the pancreatic adenocarcinoma grade assessment system 30 is a tissue cell image 10, in particular an image or an image of the pancreas part.

In another embodiment, the pancreatic adenocarcinoma grade diagnostic system 30 may be implemented as a single dedicated IC chip configured as a microprogram and driven by a microprocessor, or as an electronic circuit such as an ASIC have. In other words, it may be configured in the form of software, an electronic circuit composed of an FPGA chip or a plurality of circuit elements. Other possible forms may also be practiced. However, for convenience of explanation, the pancreatic adenocarcinoma grade diagnosis system 30 implemented in the computer terminal 20 will be described below.

Next, before describing the morphological feature extraction method according to the present invention, pancreatic adenocarcinoma (PDAC, pancreatic duct adenocarcinoma), which is an object of feature extraction, will be described first with reference to FIG.

Pancreatic cancer is the next most common cause of death from gastrointestinal cancer and is the fifth most common cause of cancer death in the western world [Non-patent document 32]. Double pancreatic adenocarcinoma accounts for 85% to 90% of all pancreatic neoplasms. Eighty percent of them are seen in patients aged 60 to 80 years, and very rarely in people under 40 years of age. By gender, the incidence of women is slightly higher than that of men. Blacks have the highest incidence rate by race [Non-Patent Document 33]. Surgical treatment of pancreatic adenocarcinoma is known to be the best treatment method. However, because early diagnosis of pancreatic adenocarcinoma is difficult, about 5 ~ 22% of the patients are eligible for curative resection. Therefore, precise determination of the preoperative stage is a key factor in determining treatment strategy and prognosis [Non-Patent Document 34].

Pancreatic adenocarcinomas are classified into well-differentiated carcinomas, moderately differentiated carcinomas, and grade 3 (Poorly differentiated ductal adenocarcinomas) according to the structural features of the ducts and morphological features of the nucleus.

First, well-differentiated carcinomas have a duct-like structure and are joined by neoplastic glands. It is usually tubular or cribriform, and may have an irregular papillary appearance in the tube. The mitotic activity of the cells is less. Cells show columnar tendency and produce mucus. And some of the epithelial cells lose their polarity.

Grade 2 (Moderately differentiated carcinomas) is a mixed form of medium-sized duct-like and tubular structures and is present in the desmoplastic stroma. The shape of the tube is usually incomplete glands. Compared to grade 1, the size of the nucleus is larger, mitotic activity is more frequent and mucus production is decreased.

Poorly differentiated ductal carcinomas do not occur frequently. It shows a mixture of small irregular glands and solid tumour cell sheets and nets. In general, the tube shape is large and duct-like structures and intraductal tumor components are not well represented. In the cells, necrosis and haemorrhage occur and show active mitosis activity. As the carcinoma progresses, the glands and surrounding tissues are invaded by a cluster of small tumor cells.

Figure 2 shows tissue cell images for grade 1, grade 2, and grade 3 of pancreatic adenocarcinoma.

As described above, grade 3 does not occur well, and morphologically there is a clear difference between grade 1 and grade 2. Therefore, in the present invention, the main goal is to differentiate between the tissue of the positive correlation, the grade 1, and the grade 2.

Next, a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic according to an embodiment of the present invention will be described with reference to FIGS. 3 and 4. FIG. FIGS. 3 and 4 are flowcharts and flowcharts illustrating a method for diagnosing pancreatic adenocarcinoma based on a morphological characteristic according to the present invention.

3 and 4, the method for diagnosing pancreatic adenocarcinoma grade based on morphological characteristics according to the present invention comprises three steps as follows: (a) Image processing and feature extraction step (S100) of tissue cells, (b) a diagnosis model configuration and verification step (S200), (c) a diagnosis step using the learned model (S300).

In the first step, image processing and feature extraction, a given tissue cell image is segmented into three parts and the features of each part are extracted and stored in a feature database to extract the quantitative characteristics required for diagnosis. The second step is learning and verifying the diagnostic model using the feature database constructed in the first step as the model learning and verification step. The final step is to perform a diagnosis of tissue cell images using the generated diagnostic model.

First, a morphological feature extraction method for classifying pancreatic adenocarcinoma according to an embodiment of the present invention will be described with reference to FIG. The morphological feature extraction method is a concrete method for the image processing and feature extraction step (100) of the tissue cell.

As shown in FIG. 5, a morphological feature extraction method (or process) S100 for classifying pancreatic adenocarcinoma comprises segmenting the lumen area (step S10); Nuclear identifying step (S20); Epithelial cell nuclear differentiation step (S30); Luminal feature extraction step (S40); And an epithelial cell extraction step (S50).

This process identifies a given tissue cell image as three types of objects. The three object types are the lumen and epithelial nuclei constituting the tube, and the other one is the non-epithelial nucleus. FIG. 3B shows a process of segmenting an image into three object types. The method of identifying each object type will be described below.

First, the lumen segmentation step S10 will be described.

In the present invention, luminal area identification was performed using a "Seeded" region growing (SRG) method [Non-Patent Document 35]. The seeded region growing (SGR) method requires an initial starting point. In the previous study [Non-Patent Document 36], the identification of the lumen area using the SRG method was automated by looking for candidate seed points in the lumenal area.

Specifically, the SGR step S10 of the present invention comprises: (a1) obtaining a binary image S11; (a2) direction cumulative map generation step S12; (a3) a candidate seed point selection step (S13); And (a4) applying the SRG method to the candidate seed point to obtain the boundary of the lumenal region (S14).

First, a maximum entropy threshold [Non-Patent Document 38] is obtained after applying median filtering and background corrections (Non-Patent Document 37) algorithm to a given image so that SRG can be more easily applied. To generate a binary image A (S11).

Thereafter, a direction cumulative map H (A) is generated to obtain a seed point from the binary image A (S12). This map accumulates only white pixels with respect to each direction (left, right, up, and down) of the preprocessed binary image A and obtains the sum of the square root of the cumulative values to obtain a position corresponding to the center of each lumen It has a large value.

At this time, the local maximum points of H (A) are used as candidate seed points of the SRG algorithm (S13). However, if a candidate seed point is obtained directly from H (A), candidate seed points are generated even in an unnecessarily small area. To solve this problem, candidate seed points are obtained from H _T (A) processed with a small threshold value of H (A). The Otsus method was used as the thresholding method.

The Otsu method is a commonly used thresholding method among various thresholding methods. This method selects the point at which the variance between classes is maximized as the threshold value. That is, the Otsu method is a method of determining the threshold value to maximize the dispersion between two classes. If each pixel value of the image is greater than or equal to the selected threshold value, 1 is assigned. Use the Otsu method to remove unwanted small areas.

Now, the obtained candidate seed points are identified by using the initial parameters of the SRG algorithm (S14). At this time, the boundary of the identified lumenal region is represented by B ⁰ .

Fig. 6 shows a process for identification of the lumen area described so far. 6 (a) is a binary image preprocessed to identify a lumen boundary, and FIG. 6 (b) shows H _T (A) and a candidate seed point (yellow point) for an image. At this time, the image was scaled from 0 to 255 for visualization. Fig. 6 (c) shows a three-dimensional graph for H _T (A), and Fig. 6 (d) shows the lumen boundary identified by the candidate seed (green line).

At this stage, the boundaries of the identified lumen areas are referred to as original lumen boundaries.

Next, the process of identifying epithelial and non-epithelial nuclei is described.

In this procedure, nuclei of a given tissue image (or tissue cell image) are identified (S20) and classified into epithelial nuclei and non-epithelial nuclei (S30). The process is as follows.

First, the nucleus identification step S20 will be described.

This step identifies all nuclei in a given tissue image (S20). First, a portion such as a cytoplasm, which is an unnecessary part for identifying a nucleus in a given tissue image, is removed by color thresholding based on k-means (Non-Patent Document 39).

Preferably, a hole-filling algorithm (non-patent document 40) is applied to holes existing in the nuclei in the threshold-processed image. The presence of holes in the nucleus does not occur well. Sometimes non-stained nuclei (where the nuclei are partially stained softly) can cause holes by thresholding. At this time, a hole filling algorithm is applied to fill these parts. This process can be omitted, but it is a process to prepare for the case where it may happen.

Finally, the Wathershed algorithm [Non-Patent Document 41] is applied to separate the nuclei. At this time, the separated nuclei are defined as the nucleus set N. The watermarker algorithm is one of the methods of image segmentation. The WaterShed algorithm is known to split objects more effectively than other image segmentation algorithms when objects are adjacent. Therefore, superposition of objects is the most commonly used method for separating nuclei from frequent cell images.

Figure 7 (a) shows the set N of identified nuclei. Figure 7 (b) shows the epithelial nucleus set N ^E (marked in red) selected in nucleus N by performing in the following steps.

Next, the differentiation step (S30) between epithelial and non-epithelial cell nuclei is described. In this step, the identified nuclei are separated into epithelial nucleus and non-epithelial nucleus clusters.

The epithelial cells surround the lumen. The epithelial cell nucleus can thus be selected by selecting nuclei adjacent to the identified luminal border B ^{< 0} > from the identified nucleus N. [ The epithelial cell nucleus set is denoted as N ^E and is defined as follows.

[Equation 1]

Where p is the point at which the lumenal boundary is made up of B ⁰ , and centroid (-) is a function that returns the center point of the identified nucleus n. And distance (·, ·) is a function that returns the Euclidean distance between two given points.

That is, the separation of the epithelial nuclei from the nucleus set N is accomplished by selecting the nuclei closest to each point p belonging to B ⁰ , and including the selected nuclei in the epithelial-nuclear set N ^E. Algorithm 1 of FIG. 8 shows this process.

Non-epithelial cell nuclei are obtained by excluding epithelial cell nuclei identified in the identified nucleus set as shown in Equation 2.

&Quot; (2) "

The table in Figure 9 summarizes the main notations used in the present invention.

Next, the process of extracting morphological features will be described. In the morphological feature extraction process, the luminal features are extracted for the segmented lumenal region (S40), and the epithelial cell features are extracted from the divided epithelial cell nuclei (S50).

The main site for the diagnosis of PDAC in the tissue image is the duct. As described above, PDACs are classified into Class 1, Class 2, and Class 3 depending on the degree of tube differentiation. The tube consists of the lumen and epithelial cells surrounding it. Thus, we describe methods for extracting morphological features for PDAC diagnosis from segmented luminal and epithelial cell nuclei.

First, the lumen feature extraction step S40 will be described.

In the PDAC, the papilla shows a papillary phase, and the epithelial cells lose their polarity and show an incomplete gland shape. As the armature progresses, the shape of the tube becomes complicated and the degree of releasing becomes more and more severe. In this step, the method of expressing the dysplasticity of the tube and the morphological feature extraction methods which quantitatively measure the dysplasia are explained.

The lumen feature extraction step S40 may include: obtaining an ideal boundary (S41); A step (S42) of obtaining a dysplasia amplitude signature; And extracting a luminal feature from the dyssomal amplitude signature (S43).

.

Describe the expression of Atypia of Duct.

In general, the lumen of a normal tube looks like a convex hull because it shows little diffusivity. On the other hand, PDAC ducts show more complicated tube dehiscence with complicated changes of the lumen boundary as the stage progresses. From this perspective, we estimate the ideal lumenal boundary of a given lumen and express the original lumenal lumen by comparison with the original lumen boundary.

First, an ideal lumen boundary is obtained (S41).

The ideal lumenal boundary is obtained by scaling the convex hull to the size of the original lumen by finding the convex hull (or convex hull boundary) from the original lumen boundary.

In this step, the boundary of the original lumen is expressed as B ^O and the ideal boundary of the lumen is estimated as B ^I. B ^O and B ^I are the sequence of points constituting each lumen boundary. The procedure for estimating the ideal lumenal boundary is as follows.

First, the convex shell of the original lumen boundary B ^O is obtained. The convex shell of B 2 ^{O thus} obtained is denoted by B ^C. A convex shell is defined as the smallest convex polygon containing given points. In other words, a point on the outermost point (a point on the smallest rectangle that encloses the points) is searched to find a point where all the points subtracted from it can be placed on the left or straight line. Again, finding the points satisfying the same condition for the remaining points on the basis of the found point in turn, we can obtain the group of the minimum points enclosing the given points, which constitutes the convex hull.

At this time, because the area bounded by B ^C is the area size bounded by B ^O , the ideal inner boundary B ^I is obtained as the boundary of the area bounded by B ^C , which is reduced to the size of the area bounded by B ^O .

The scaling factor s for reducing the region bounded by B ^C to the region bounded by B ^O is given by Equation 3. That is, the scaling factor for obtaining the ideal lumen boundary B ^I is obtained using the area of B ⁰ and B ^C.

 &Quot; (3) "

Here, R ^O means a region bounded by B ^O and R ^C means a region bounded by B ^C. Area (·) is a function that returns the size of the area. The lumen B ^I is the set of points constituting the boundary of the scaled region with respect to the center of R ^C.

Figure 10 shows the original luminal boundary B ^O (green) and the ideal luminal boundary B ^I (red).

Next, an Atypia-Amplitude Signature is obtained (S42)

In this step, we propose a 1-D signature that expresses the dissociation of the lumen using the original lumen boundary and the ideal lumen boundary. The proposed 1-D signature visualizes the dissociation in the original lumen by measuring the amplitude of the dissociation between the ideal lumen and the original lumen. At this time, the dysmorphic amplitude is measured by the dysmorphic-amplitude function A (t) as the vertical distance having the sign between the ideal lumen and the original lumen.

Equation (4) shows the definition of the damping amplitude function.

&Quot; (4) "

Where t is an index variable indicating the order of the points at the ideal lumen boundary, and sgn _t (q) is a function representing the sign of the distance between vertex q ∈B ⁰ in p _t ∈ B ^I , +1 if it is located at the outer boundary of the point p _t , or -1 if it is at the inner boundary.

That is, the emissivity amplitude is a distance between points on an original boundary located vertically at an ideal boundary point, and is represented by a distance having a sign.

Fig. 11 (a) shows the process of measuring the dyssymmetry-amplitude at p _t ∈ B ^I. In this example, q is a point at the outer boundary of p _t , so A (t) is measured as a positive value.

In the present invention, the 1-D signature using the releasing-amplitude function A (t) is called a releasing amplitude signature. The Atypia-Amplitude Signature is the distance from the origin (p ₀ ) to p _t of the ideal lumen boundary to the x-coordinate of the circumference L (t) and the emissivity amplitude A (t) obtained from the point p _t of B ^I as y (T (t), A (t)) of coordinates (Fig. 12 (b)). Here, L (t) representing the circumferential length from the origin (p ₀ ) to the point p _t is defined as the sum of the distances between the points as shown in Equation (5).

The table of Fig. 12 shows the dummy amplitude signatures for Figs. 1 (a), (b) and (c).

Next, the dissimilarity measurement features of the tube to the lumenal region are extracted (S43).

In this step, we describe the characteristics of quantitatively measuring the lumenal dysplasia using the ideal lumenal boundary and the dysplasia-amplitude signature obtained in the previous step. The proposed features are Root Mean Square Atypia Amplitude (RMSAA), Total Sum of Atypia Volatility (TSAV), AtypiaRatio, and #AtypiaRegions.

First, the Root Mean Square Atypia Amplitude (RMSAA) will be described.

The RMSAA is measured from the dyssomnal amplitude signature obtained in the previous step. This is the square root of the average squared atypia amplitudes for the discrete amplitude A (t). A (t) is the vertical distance between the original ideal lumen boundary B ^{I and} the original lumen boundary B ^O , which represents how far the original lumen B ^O deviates from the estimated normal lumen B ^I.

Thus, the measured A (t) can be interpreted as a residual or error indicating the difference between the observation and the estimated function value in the regression analysis. Similarly, it corresponds to an MSE (Root Mean Square Error) [Non-Patent Document 42] which measures an average error of the MSAA regression equation. RMSAA is defined by Equation 6.

&Quot; (6) "

Where m is the number of points in B ^I.

TSAV (Total Sum of Atypia Volatility) will be described.

Pancreatic adenocarcinoma becomes more complex as the lumen becomes more complicated as the stage progresses. TSAV measures the border changes of the lumen complicating the progression of the stage. The TSAV measurement is calculated by finding the major inflection point of the dysplasia-amplitude signature and summing the variations at those points.

In the present invention, a PIP algorithm [Non-Patent Document 3] and [Non-Patent Document 4] were used to find the main inflection point of the dyssomance-amplitude signature. The PIP algorithm is a method of finding important points (PIPs) representing key trends from time series data.

In the present invention, the existing PIP algorithm for detecting a fixed number of PIPs is modified to find all inflexion points present in the dysplasia-amplitude signature.

Figure 13 shows a portion of the dormancy signature for Grade 2 in Figure 12 and the PIPs identified by the modified PIP algorithm.

In Fig. 13, p _{i, which} is AV _i in the PIP, is measured at an angle (? _I ) between a _i and b _i .

TSAV measures the AV at each point of the identified PIPs and is calculated as the sum of the values. In this case, AV _i in the i-th PIP _i is defined as the two vectors, a _i = _i PIP -PIP _{i -1} b _i = the angle (θ _i) between the between the PIP -PIP _{i +1 i.} Figure 13 shows how AV _i is measured in PIP _i .

&Quot; (7) "

Where m is the number of PIPs identified from the dormant amplitude signature.

The releasability ratio and the number of releasable regions will be described.

Tubular adenocarcinoma (pancreatic adenocarcinoma, pancreatic duct adenocarcinoma) has a complicated tubular shape and a distinct papillary phase as the stage progresses. The actual lumenal area R ^O in which the armature is advanced deviates or invades the ideal tube area R ^I. Therefore, in this step, set operation is applied to the original tube region and the ideal tube region to obtain the dissociation regions, and the number of dissociation ratios and the number of dissociation in the original tube are counted.

To obtain these two features, we obtain the set of homogeneous regions AR = {ar ₁ , ar ₂ , ..., ar _m }. AR is defined as the set of regions that divide (R ^I ∪ R ^O ) - (R ^I ∩ R ^O ) by B ^I. Figure 14 shows the emissive regions identified in Class 2.

In Fig. 14, (a) is a tissue cell image of Grade 2, and (b) is a releasability region generated by regions R ⁰ and R ^I of (a).

Now we obtain the ratio (AtypiaRatio) and the number of heterogeneous regions (#AtypiaRegions) that the heterogeneous region occupies in the actual tube from the acquired heterogeneous regions.

&Quot; (8) "

&Quot; (9) "

Here Area () is a function that returns the size of a given area. m is the number of elements in the AR set. AtypiaRatio indicates how much the tube is distorted overall. #AtypiaRegions expresses the degree of papillary phase by counting the number of dissociative regions in which the actual tube region deviates or invades the ideal tube region. At this time, areas that are too small to express the nipple image are excluded from counting by applying thresholding.

Preferably, the threshold value (or threshold value) is set to 300 μm ² .

Next, an epithelial cell feature is extracted from epithelial cell nuclei (S50).

The second component of the tube is epithelial cells. In many cases, the epithelial cells in mucinous adenocarcinoma produce mucin, which exhibits columnar tendency and polarity of nucleus is lost [Non-Patent Document 33] [Non-Patent Document 45]. We describe how to extract these features using the epithelial cells identified in this step.

First, the characteristics of the epithelial cytoplasm length will be described.

In general, ducts of normal tissues are surrounded by cuboidal epithelial cells. On the other hand, the ducts of PDAC are surrounded by columnar epithelium due to abundant cytoplasm. The nucleus of the columnar cell has an egg shape. The cytoplasmic length of columnar epithelial cells with abundant cytoplasm is longer than that of cuboidal epithelial cells.

Thus, the measurement of the cytoplasmic length of epithelial cells expresses whether epithelial cells are columnar cells or not. In the prior art [Non-Patent Document 36], Cytoplasm Length is used to measure the cytoplasmic length of epithelial cells. The cytoplasmic length feature is the vertical distance between the original luminal border B ^O and the epithelial nucleus n. Equation 16 expresses the cytoplasm length characteristic CytoplasmLength.

&Quot; (10) "

Where m is the number of elements in the set N ^E and the point q of B ^O is the vertex of the centroid (n).

FIG. 15 shows the result of measurement of the CytoplasmLength of normal and Grade 1. In Fig. 15, the red dot is the epithelial nucleus and the green line is the identified luminal border. And the wave line between the identified luminal border and epithelial nuclei is the measured CytoplasmLength.

Next, the standard deviation of the cytoplasm length characteristic (CytoplasmLength) will be described.

This feature measures the loss of nuclear polarity, one of the characteristics of pancreatic adenocarcinoma. The epithelial cells that exhibit this feature are irregularly arranged along the lumen, and thus the deviation of the cytoplasmic length is large. On the other hand, the epithelial cells are regularly arranged along the lumen because the polarity loss does not occur in the normal case, and the deviation of the cytoplasmic length is very small.

Thus, the loss of nuclear polarity is measured by determining the standard deviation of the cytoplasmic length of epithelial cells.

&Quot; (11) "

Next, the diagnosis model configuration and verification step (S200) will be described.

The diagnostic model is constructed using a classifier based on map learning. Preferably, a SVM (Support Vector Machine) classifier is used. Diagnostic models are constructed by using learning data that pairs each feature with the diagnostic result (or grade) of pancreatic adenocarcinoma. And verified through cross validation.

The diagnostic model is a 'classifier' that distinguishes between pancreatic adenocarcinoma and stage. In other words, diagnosis of pancreatic adenocarcinoma is similar to classification of pancreatic adenocarcinoma. As a way of dealing with classification problems using computers, we use methods based on map learning. Preferably, SVM (Support Vector Machine), which is a statistical based map learning method, is used.

The map learning method learns a diagnostic model or classification model (classifier) using learning data. Here, the learning data is a set composed of elements of a pair of a feature vector ( x ) and a class label (y). Let D denote the learning data as follows.

D = {( x ₁ , y ₁ ), ( x ₂ , y ₂ ), ..., ( x _n , y _n )}.

The feature vector ( x ) is a vectorized feature of the previously obtained pancreatic adenocarcinoma.

The class label y indicates the actual class label of the feature vector x , which is a classification of the diagnostic result (or the grade of the diagnostic result) of pancreatic adenocarcinoma. For example, in the case of normal and pancreatic adenocarcinoma classification, y may have a value of 'normal' or 'pancreatic adenocarcinoma', and in the case of pancreatic adenocarcinoma staging, y has a value of 'Grade 1' and 'Grade 2' .

At this time, the learning of the diagnostic model is heavily influenced by the learning parameters (C ^* , γ ^* ) determined at SVM learning (where C is the model parameter and γ is the kernel parameter). Of course, learning data is most important, but these parameters play an important role in learning learning data. Specific examples refer to those described in the experiment of the present invention with reference to Fig.

Verification of the learned diagnostic model is performed through k-fold cross validation. k-fold crossover validation divides training data D into k data sets and k-1 data sets to learn the SVM model. The remaining one data set is used to verify the performance of the diagnostic model. This process is performed k times. That is, k diagnostic models are generated using the divided data sets, and the model is verified as an average of the performance evaluation of k diagnostic models. Learning for k-th cross-validation uses learning parameters (C ^* , γ ^* ) obtained in SVM learning.

Next, the diagnosis step (S300) using the learned model will be described.

It is a process of diagnosing a given tissue image using a learned diagnostic model (or a diagnostic classifier). That is, the feature vector is input to the diagnostic model (or the diagnostic classifier), and the result (or the grade of the diagnostic result) output by the diagnostic model is the diagnosis result. That is, the diagnostic model is like a function that classifies classes.

Next, the effects of the present invention will be described in detail through experiments.

For the experiments of the present invention, 21 normal tissue slides and 26 PDAC tissue slides were provided from Yeungnam University Pathology Department. The tissue slides provided were stained with Hematoxylin & Eosin method. And these tissue slides were converted into 20X magnification digital slides using ScanScopeCS system. The size of each digital slide image is variable depending on the acquired organization. Figure 16 shows information on a digital slide.

In order to evaluate the features of the present invention, we have created an experimental image to include one tube from the obtained digital slide. Experimental images were stored in 24bit tiff format and the size of each image was variable according to the size of the tube. An important issue in diagnosis is inter- and intra-observer variability. These problems can lead to inaccurate, inconsistent, and biased diagnostic results [Non-Patent Document 46]. These problems also appear in the experts' ground truth data for learning diagnostic systems and evaluating performance. To reduce the variability problem of ground thruth, a well-known method involves the participation of several experts to construct ground thruth [Non-Patent Document 47]. In the present invention, three pathologists of the pathology department of Yeungnam University received the help of a ground pathologist for evaluation of ground throuth. Each tube image generated from the digital slide was carefully labeled by each of the three pathologists.

Figure 17 shows the number of experimental images labeled by experts for ground thruth evaluation.

A given tissue cell image was segmented into three parts (lumen, epithelium, and non-epithelial nuclei) and new features specific to tubal carcinoma were extracted at each site along with traditional morphological features. Figure 18 is a characteristic used in the experiments for diagnosis of pancreatic adenocarcinoma. The features extracted from each segment are indicated by "*". The shaded rows in FIG. 18 are characteristics specific to the tuberculosis proposed in the present invention, and the remaining rows are the characteristics of the existing morphological features [Non-Patent Document 5], [Non-Patent Document 24], [Non-Patent Document 48] [Non-Patent Document 49]. In the case of epithelial nuclei and non - epithelial nuclei, since there are several objects in one tissue cell image, feature values of each image are expressed by obtaining and averaging the feature values for each object. Experimental environment for feature extraction of experimental images was conducted on AMD Athlon II 3GHz CPU, 2G RAM and Windows7-64bit. Existing feature extraction methods and proposed feature extraction methods are implemented through Image J [Non-Patent Document 50], a public image analysis package based on JAVA language

In order to demonstrate the superiority of the morphological features presented in the present invention for PDAC diagnosis, the classification performance of the classifiers learned by the proposed features and the existing features is compared. In the present invention, learning of the classifier employs SVM [Non-Patent Document 51], which is well known as a statistical learning technique. SVM is generally known to have good performance because it minimizes VC dimension and empirical risk [Non-Patent Document 52]. The experiment was carried out to measure the classification results for the following two cases. The first is the classification of normal and tubal adenocarcinoma. The second is the stage classification of tubal adenocarcinoma grade 1 and grade 2. In order to measure the contribution of the features of the present invention to the improvement of the classification ability, the classifier was learned by combining the features of the present invention and the features of the present invention, and the characteristics of the present invention were evaluated through comparison of the classification results of the classifiers. Figure 19 shows the symbols and dimensions for the feature sets used in the classification experiments.

Experimental data were generated according to each feature set in Fig. 19 for two experiments (Normal vs. PDAC and Grade 1 vs. Grade 2). In the first PDAC diagnostic experiment, 13 data sets were generated as follows: D (CNF), D (CLF), D (PLF), D (ALF), D (CEF) (AFT), D (CDF), D (PDF), D (ADF), D (CTF), D (PTF), and D (ATF). Where D (·) is a data set consisting of a specific feature set given as a factor. D (·) is written as D (·) = {(x (·) ₁ , y ₁ ), ..., (x (·) _m , y _m )}. Where x (·) _i is an i-th feature vector with feature set given as a factor. The class label y _i is the class label of x (·) _i and has a value of -1 which means Normal class or a value of 1 meaning PDAC class. Similarly, fifteen data sets were generated in the second PDAC arm sorting experiment. If the class y _i is -1, it means class 1, and if 1, it means class 2.

In order to evaluate the classification performance of SVM according to feature set, training set and test set were constructed from data set generated by feature set. The ratio of training set to test set for learning and evaluation was set at 60:40. In the first experiment (normal vs. PDAC), the PDAC experimental data was constructed by sampling 80 data from 160 PDAC data (data of class 1 and data of class 2). Since the number of normal samples is 80, we have limited the number of PDAC samples to 80 for fair evaluation of the classifier.

Figure 20 shows the number of training data and test data used in Experiment 1 (Classification between the normal and PDAC) and Experiment 2 (Classification between the normal and PDAC).

In the present invention, a soft-margin method and an RBF kernel (Non-Patent Document 51) are used for the SVM classifier. Therefore, the model parameter C and the kernel parameter γ are required. The optimal classification parameter (C ^* , γ ^* pair) is a pair of (C, γ) ∈ {10 ^-1 , 10 ^-0.5 , ..., 10 ⁴ ) that maximizes the accuracy of the training set using 10 cross validation methods. } {2 ^-5 , 2 ^-4.5 , ..., 2 ⁰ } were selected through grid search [Non-Patent Document 53]. The number of experimental images used in the present invention is very low. Thus, the accuracy of the generated classifier can be biased [Non-Patent Document 54]. Statistically, to solve this problem, a bootstrap technique (Non-Patent Document 55) is used. The bootstrap technique is used for unbiased evaluation of the classifier for each feature set. First, we created 10 training sets and 10 test sets from the data set D (·) corresponding to a given feature set (Table 6) for bootstrap evaluation. Test a set of 10 training sets are D (·) = {(L (·) 1, T (·) 1), ..., (L (·) 10, T (·) 10)}, ∀L (· ) _i , T (·) _i ⊂ D (·) ( ₁ ≤ _i ≤ 10). Where L ( _i ) and T ( _i ) denote training data and testing data, respectively. Thus, the classification performance for each feature set is measured as the average of 10 training data sets and the sorter evaluation results individually optimized for the test data set.

The evaluation items used for the performance evaluation of the classifier in the experiment are rue Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN), Sensitivity (SN) , Positive Predictive Value (PPV), Negative Predictive Value (NPV), and Accuracy (ACC). The description of TP, TN, FP, FN is described in the experiment of each case. The remaining performance evaluation items are defined as follows.

&Quot; (12) "

&Quot; (13) "

&Quot; (14) "

&Quot; (15) "

&Quot; (16) "

Next, the experimental results will be described.

First, case 1 (positive correlation vs. PDAC) will be described.

21 and 22 show the bootstrap evaluation of the classifier learned by each feature set for distinguishing Normal and PDAC. The numerical values in parentheses in FIG. 21 represent standard deviations for 10 bootstrap evaluation results. First, let's look at a comparison of classifiers learned by feature sets extracted from cabinet objects. In this comparison, the accuracy of the classifier learned by the PLF was 91.56. It was about 18% more accurate than the classifier learned in CLF (73.44%). The accuracy of the classifier learned by the ALF feature set composed of CLF and PLF is 87.35, which is lower than that of PLF. These results show that the proposed PLF feature set is better for diagnosing PDAC. It also shows that there is no additional performance improvement due to the combination of PLF and CLF. This shows that simple morphological features such as Area and Perimeter are not suitable for expressing complex features of PDAC.

Next, let us look at the classification results of feature sets extracted from epithelial nuclei. First, CEF, which is a set of existing morphological features, showed classification performance similar to CLF of lumen with an accuracy of 72.66. PEF and AEF showed the same classification accuracy of 87.50%. However, the classifier with AEF has a standard deviation of classification accuracy of 1.47, which is more stable than the PEF with a standard deviation of 3.21. Considering that the AEF is a combination of the proposed PEF and the existing CEF, the accuracy of the AEF can be attributed to the proposed PEF.

It is also noteworthy that the characteristic dimension of the PEF is two-dimensional. These results show that PEF is very suitable for identifying PDACs. It is also very cost effective. In the PDAC diagnosis, the tube is an important area of interest in diagnosis. The tube consists of lumen and epithelial cells. Thus, the combination of features extracted from lumen and epithelial cells will be meaningful. To this end, we have prepared three feature sets. CDF (CLF + CEF), PDF (PLF + PEF), and ADF (CDF + PDF). And we used them to evaluate the classification performance. In this experiment, the accuracy of the classifier was 94.38% with PDF. This is about 3-7% higher accuracy than PLF and AEF (or PEF), which showed the best classification accuracy in each object. This demonstrates that the combination of luminal and epithelial cell features are helpful in the diagnosis of PDAC.

Next, the experimental results of the CTF, PTF, and ATF using the combination of the features extracted from the three objects can be seen to depend on the experimental results performed on the vessel object. Therefore, this shows that there is no performance improvement through combining features of all objects. In the case of the PTF, by adding the CNF feature set to the PDF, the accuracy is lowered by about 2% compared to the classification accuracy of the PDF.

Next, ROC analysis was performed on the classifiers learned by each feature set. ROC analysis is widely used in the medical field as a criterion for comparing and indicating the accuracy of diagnostic tests. The ROC analysis is analyzed through ROC curves plotted with TP rate (Sensitivity) and FP rate (1-Specificity). In this analysis, the area under the ROC curve (AUC) of the ROC curve represents the accuracy of the diagnosis. Swets are classified as non-informative (AUC = 0.5), less accurate (0.5 <AUC≤0.7), moderately accurate (0.7 <AUC≤0.9), very accurate , and highly accurate (0.9 <AUC <1) and perfect tests (AUC = 1) [Non-Patent Document 57] and [Non-Patent Document 58]. In other words, it can be interpreted that the closer the ROC curve is to the left edge, the more accurate diagnosis is obtained. 23 shows the averaged ROC graph and AUC values of the 10 classifiers learned by the feature sets for each object type.

In the ROC analysis, the AUC value of the classifier learned by PDF was the highest at 0.96. This shows that the combination of lumen and epithelial nuclear features is significant, as in previous classifier accuracy analyzes. Next, the AUC value of the classifier learned by the PEF was 0.94, which was higher than that of the PLF with an AUC value of 0.93 than that of the lumen. Except for the CNF feature, the classifiers learned by the feature sets including the proposed PDF or PEF can be interpreted to perform highly accurate diagnosis with AUC value of 0.9 or higher.

Overall, experiments involving the feature set of the present invention showed better performance than classifiers with existing feature sets. As mentioned earlier, we have shown that tubes are an important area for PDAC diagnosis. The PDF consisting of PLF and PEF led to improved performance of the classifier. PLF and PEF extracted from the proposed lumen and epithelial cells showed higher performance than the existing feature sets CLF, CEF and CNF.

Next, Case 2 (Grade 1 vs. Grade 2) will be described.

In this step, the distinction between armor class 1 and class 2 of the PDAC is performed according to the feature set. Similar to the experiment in Case 1, we generated 10 training sets and test sets for class 1 and class 2 distinction and evaluated classification accuracy. 24 and 25 show the performance evaluation results of the classifier. The classification results of each feature set to distinguish between stages 1 and 2 showed lower classification accuracy than those of case 1 (Normal vs PDAC).

First, PLF (77.03%) showed about 19% higher accuracy than CLF (57.97%) in the experiment on the lumen object. In particular, the PLF specificity (79.69) shows a performance gain of about 73% over CLF (45.94). In an experiment with ALF, a combination of CLF and PLF, the classification accuracy was 68.44%. This shows that the performance of the classifier learned by the PLF is lower than that of the classifier.

Next, experiments with feature sets extracted from epithelial-nuclei show overall lower performance than that of case 1 (normal vs pdac). In Grad1 and grade 2 of PDAC, epithelial cells show a columnar shape and polarity loss characteristic of nucleus. It is therefore difficult for these features to distinguish between class 1 and class 2. Nevertheless, the classification accuracy of PEF (70.78%) was 14% higher than that of CEF (56.41%). Experiments using AEF were lower than those of PEF. This shows no performance improvement through CEF and PEF.

Experiments using the tube feature sets are identical to those of the feature sets extracted from the lumen object. Unlike Case 1 (Normal vs. PDAC) experiments using PDF, there was no improvement in performance through PLF and PEF, and the same experimental results as PLF were obtained. This shows that the performance improvement of the classifier via PDF is entirely dependent on PLF, not PEF. Experiments on the combination of feature sets (CTF, PTF, ATF) extracted from three objects showed the same results as those of the feature set extracted from the tube.

26 shows the averaged ROC graph and the AUC value of the staging classifiers according to the feature sets for each object type. Compared to Case 1, the performance difference was large. For the classifier using existing morphological feature sets such as CNF, CLF, and CEF, the AUC values were 0.61, 0.45, and 061, respectively, indicating less accurate diagnostic results. On the other hand, the classifiers using the feature sets of the present invention, such as PLF and PEF, show a moderately accurate test result of 0.79 and 0.7, respectively.

The performance of the classifier with PLF showed the best performance in this stage classification experiment. Unlike Case 1, which distinguishes between normal and PDAC, performance improvement through PEF and PLF combination is not shown. What is unique in this experiment is that the experimental results of the feature sets of the Lumen object are the same as the experimental results of the duct and Tissue object feature sets. The duct object feature set and the tissue feature set consist of features extracted from a lumen object and features extracted from other objects. This means that the features extracted from the lumen object directly affect the performance of the classifier and the lumen part contains the most information necessary for the diagnosis.

Overall, stage 1 and level 2 armament trials have shown poor flight performance in tests distinguishing normal and PDAC from both the features of the present invention and the existing features. This is because the characteristics that appear in the PDAC appear in Class 1 and Class 2 arms. In addition, the features of the present invention are designed to classify pancreatic adenocarcinoma and normal tissue preferentially, thus limiting the diagnosis of pancreatic adenocarcinoma including similar morphological characteristics. However, the features of the present invention show higher classification accuracy than the existing features.

Next, the fitness test of the extracted features will be described.

In other words, a statistical analysis of the features is performed to show how the features extracted from the segmented details (non-epithelial cells, epithelial cells, lumen) are suitable for diagnosing pancreatic adenocarcinoma. If the extracted features are suitable for diagnosis, it is assumed that the feature values will be differentiated among the groups. Analysis of variance (ANOVA) was performed for each feature to statistically show whether the extracted features differed among the three groups. The hypotheses to test for differences in extracted features are as follows.

&Quot; (17) &quot;

Where μ _N , μ _G1 , and μ _G2 are the mean of the mean for Normal, Grade 1, and Grade 2, respectively. The significance test of ANOVA is tested by F-statistic. Figures 27, 28, and 29 are F-test results of the features obtained from the segmented cells (non-epithelial cells, epithelial cells, lumen).

The F-test results of the features for the three object types statistically show that most features differ between groups. It is shown that all the features differ between the groups (normal, grade 1, and grade 2) except for the F-test case Roundness and Solidity feature for the features extracted from the endurance object, The F-test results of the extracted features show that all extracted features differ among the groups. In non-epithelial nuclei, all features except MinorAxis and skewness showed differences between groups.

개 쌍들간에 대해 차이가 있는지를 검정한다. 각 쌍들을 검정하기 위한 영가설은 다음과 같다.We conducted a multiple range test to find out which differences exist in the comparison of the groups for the above null hypothesis-rejected features (ie, there is a difference between the three groups). In the present invention, was used Fisher's LSD (LSD (Least Significant Difference) test [ Non-Patent Document 58] The most commonly of the number of multiple range test method used. LSD is the average of the two populations to be compared, μ _i and μ _j In the present invention, there are three groups, normal, grade 1, and grade 2, so that the total

Check for differences between pairs of dogs. The null hypothesis for testing each pair is as follows.

&Quot; (18) &quot;

The LSD-test was performed using Eq. 17 is rejected. The LSD-test results are shown in the tables of FIGS. 30 to 32.

The shaded rows in FIGS. 30-32 show the Eq. These are the features that rejected all three spiritual theories of 30.

The features marked '-' in the table indicate that the LSD-test was not performed because the F-test did not reject the Eq. 29 hypothesis. First, twelve features except Circularity and AspectRati in the LSD-test for features extracted from the lumen object are Eq. I rejected all the spiritual theories of 18. Eight of these are existing features, and four are features of the present invention. Despite the fact that many of the existing features show differences between groups (Normal, Grade 1, Grade 2), the results of experiments using the proposed PLF showed good performance for CLF in Case 1 and Case 2.

In the epithelial nucleus object, five features, including the proposed CytoplasmLength and CytoplasmLength _SD features, I rejected the null hypothesis of 18. In the case of non-epithelial nuclei, only Perimeter, Width, Major Axis, and Fereter's Diameter rejected all null hypotheses (Eq. 30).

In the LSD-test, the null hypothesis H ₀₃ of Eq. 18 (to test for a significant difference between grade 1 and grade 2) was not rejected in many of the features extracted from epithelial and non-epithelial nuclei. In the F-test, the null hypothesis (Eq. 29) was rejected. Among the features rejected, H ₀₃ of Eq. The feature that failed to dismiss 30 is only two features of Circularity and AspectRatio. However, in the case of the epithelial nucleus type, in which all the features rejected the null hypothesis (Eq. 29) in the F-test, only 5 of the 14 features were identified in the LSD-test as H ₀₃ of Eq. 30 was rejected. In the non-epithelial nuclear type, only four features were identified in the LSD-test, H ₀₃ of Eq. 30 rejected.

This LSD-test result shows that lumen is the most important object in PDAC diagnosis and staging. In addition, these results are statistically explained because the performance of Case 2 for staging is lower than that of Case 1 for normal and PDAC diagnosis.

The present invention has described methods for automatically diagnosing tuberous carcinoma. Diagnosis of ductal carcinoma is mainly performed by observing the degree of differentiation of the duct and epithelial cells surrounding the duct. The present invention segmented a given tissue cell image into luminal, epithelial, and non-epithelial cells. From the segmented duct and epithelial cell line, we have proposed new morphological features that are specialized to diagnose tuberous carcinoma with traditional features. In the case of tubular adenocarcinoma, the tube is more complex in shape than the tube of normal tissue and the bending change of the border is very severe. From this viewpoint, the present invention has proposed features for measuring pipe volatility and pipe release ratio. In order to intuitively express the volatility of the pipe, the pipe is represented by a 1-D signature using the damping amplitude. The RMSAA feature for measuring the variation width of the emissivity amplitude from this emissive amplitude signature and the PIP - based TSAV feature for measuring the variation at the inflection point are proposed. Using the ideal lumen and original lumen areas, AtypiaRatio for measuring the degree of distortion of the tube and #AtypiaRegions for measuring the papillary level of the tube were proposed. In addition, CytopasmLength and CytoplasmLength _SD features were used to measure the cytoplasmic length and polarity from the segmented epithelial cells. Statistical analysis of the classification results and features using the SVM shows that the features of the present invention are suitable for diagnosis, grade 1 and grade classification.

Although the present invention has been described in detail with reference to the above embodiments, it is needless to say that the present invention is not limited to the above-described embodiments, and various modifications may be made without departing from the spirit of the present invention.

10: tissue cell image 20: computer terminal
30: Program system

Claims (14)

A method for classifying pancreatic adenocarcinoma based on morphologic features that extract morphologic features from tissue cell images and classify the presence or grade of pancreatic adenocarcinoma,
A first step of extracting the morphological feature vector from the tissue cell image;
A second step of constructing a SVM (Support Vector Machine) classifier for classifying the presence or absence of pancreatic adenocarcinoma upon receipt of the morphological feature vector; And
And a third step of classifying whether the tissue cell image is pancreatic adenocarcinoma or grade using the classifier,
In the first step,
(a) identifying and segmenting a luminal area in a tissue cell image by a seed area positive (SGR) method;
(b) subjecting the tissue cell image to threshold processing based on k-means and then applying a Wathershed algorithm to identify nuclei;
(c) dividing the identified nuclei into epithelial cell nuclei and non-epithelial cell nuclei, wherein the nuclei adjacent to the border of the lumenal region are divided into epithelial cell nuclei, and the remaining epithelium nuclei except for the epithelial cell nucleus are divided into non-epithelial cell nuclei A step of discriminating;
(d) obtaining an ideal boundary using a block skin of the original boundary of the lumenal region, determining a distinction amplitude from a distance between the original boundary and the ideal boundary, and extracting a luminal characteristic using the dissimilarity amplitude; And
(e) extracting epithelial cell characteristics from the cytoplasmic length by obtaining a cytoplasmic length, which is a distance between the boundary of the lumenal region and the epithelial cell nucleus; And
(f) generating a morphological feature vector comprising the lumenal feature and the epithelial cell feature.
delete delete The method according to claim 1,
Wherein in the second step, the classifier is learned with learning data composed of a pair of classes of the morphological feature vector and classification result.
delete The method of claim 1, wherein the step (a)
(a1) acquiring a binary image by applying a median filtering and a background correlation algorithm to the tissue cell image and applying a maximum entropy threshold;
(a2) generating a direction cumulative map from the binary image;
(a3) selecting local maximum points of the direction accumulation map as candidate seed points; And
(a4) determining the boundary of the lumenal region using the candidate seed points as an initial parameter of the SRG method.
delete delete The method of claim 1, wherein the step (d)
(d1) obtaining an ideal boundary from the original boundary;
(d2) obtaining a dyspeakable amplitude signature using a distance of points perpendicular to each other at the original boundary and the ideal boundary; And
(d3) extracting the luminal feature from the dyspeptic amplitude signature. &lt; Desc / Clms Page number 19 &gt;
10. The method of claim 9,
Wherein the ideal boundary is obtained by scaling the convex shell of the original boundary to a size of an inner region of the original boundary in the step (d1).
10. The method of claim 9,
Wherein in the step (d3), the dyspeakable amplitude signature is expressed by a dummy amplitude of a circumferential point (hereinafter, an ideal dummy point) with respect to the circumferential length of the ideal boundary, and the dummy amplitude indicates an original boundary Wherein the distance between the points on the pancreatic adenocarcinoma is a distance having a sign between points on the pancreas.
12. The method of claim 11,
Wherein the dissociation amplitude is determined by a dyssymic amplitude function A (t) of the following formula (4).
&Quot; (4) &quot;

Where t is an index variable indicating the order of the points at the ideal boundary and sgn _t (q) is a function representing the sign of the distance between vertex q ∈B ⁰ in p _t ∈ B ^I , Returns -1 if q is at the outer boundary of the point p _t , or -1 if it is at the inner boundary.
12. The method of claim 11,
The luminal feature may be any one or more of the square root of the root mean square (RMSAA) for the emissive amplitudes, the sum of the amounts of change at the inflection points of the emissive amplitude signature (TSAV), the emissivity ratio (AtypiaRatio), and the number of emissive regions (#AtypiaRegions) Characterized by morphological features based on classification of pancreatic adenocarcinoma grade.
12. The method of claim 11,
Wherein the epithelial cell characteristic is at least one of an epithelial cytoplasm length (CytoplasmLength) and a standard deviation (CytoplasmLength _SD ) with respect to a cytoplasmic length of epithelial cells.

Images