乳癌是目前台灣女性所有癌症死亡率排名的第四位，對國內婦女造成相當大的威脅。本研究之主要目的在開發一套乳癌診斷系統，藉由病理組織學原理及影像處理技術，對H&E染色法後之二維乳癌組織切片影像進行分析，自動偵測出有絲分裂細胞核與分析核大小多形性，評估乳導管結構程度。提供醫師定量的預後分析及組織惡性度分級，改善長久以來國內在乳癌診斷病理的準確率與效率，輔助醫師對於乳房腫瘤之診斷，以降低乳癌對於生命的危害。
研究過程中使用乳房組織切片之CCD數位元影像作分析，分為高倍(400X)與低倍(100X)顯微鏡所拍攝出來的圖像。在高倍成像，我們直接擷取乳房組織切片之CCD成像進行分析，透過實驗步驟包括：色彩轉換、分閥值演算法、區域成長、數學形態學演算法等，分割出感興趣細胞核部分，再計算細胞核之面積、緊密度、平均曲率差等特徵參數。在低倍成像，先透過亮度直方圖運用K-Means分群演算法、數學形態學、邊緣偵測、標記分水嶺等影像處理技術偵測感興趣之乳導管，量測乳腺管厚度、對稱性及內外輪廓的形狀相似性，藉此偵測正常或異常乳腺管區別所在，統計出良性或惡性導管比例。最後利用此病變細胞與導管型態有關特徵參數，結合機率類神經網路與支援向量機訓練選取出較好分類器，根據其分類標準評估乳癌發生的嚴重性，提供乳癌病患預後分析之指標。

The breast cancer is at the forth leading mortality cause of women carcinoma in Taiwan, and causes a big threat to the domestic women. The major goal of this research is to develop a diagnostic system for breast cancer bases on histology and cytology. It first analyzes the 2D H&E stained slide images of breast specimen to detect mitotic nuclei and nuclei pleomorphism, and extract feature parameters related to morphometry of mammary ducts. Moreover, the proposed system provides prognosis analysis and histological grade for clinical pathologists. The diagnostic accuracy and efficiency of breast carcinoma can thus be improved. In clinical medicine, the proposed system can assist doctor increase the diagnostic performance in detecting malignant tumors and reduce its threats to our lives.
In this study, high and low magnification histology-slide images of breast tissue specimens fetched from a CCD camera mounted on a microscope were selected carefully for our experimental samples. For high power images, image operations including color transform, thresholding, region growing, and mathematical morphology operations were performed on the slide images to separate the nuclei of interest. Next, feature parameters including area, compactness, and curvature difference were measured for following classification. For low power images, a series of image processing techniques such as K-Means clustering based on histogram, region growing, contour detection, and watershed operation were first designed to separate the mammary ducts of interest. Next, the thickness symmetry, and shape similarity between the inner contour and the outer contour of the detected mammary ducts were estimated in order to identify normal or abnormal duct, and evaluate the ratio of malignant carcinoma.
Finally, the parameters extracted from abnormal cell and the mammary duct were classified and trained based on probability neural network ( PNN ) and support vector machine (SVM) in order to choose a better classifier. The results of the classification can be used to evaluate the severity of the breast carcinoma. It can also provide an important indicator for prognosis analyses of the breast carcinoma.

中文摘要 i
Abstract iii
誌謝 v
目錄 vi
圖目錄 xi
表目錄 xvii
第一章、 導論 1
1.1 研究動機 1
1.2 研究背景及相關研究 4
1.2.1 國外相關研究現況 4
1.2.2 國內相關研究現況 5
1.2.3 參考文獻之評述 6
1.3 研究目的 7
1.4 論文架構 8
第二章、乳癌基礎病理及影像處理技術 9
2.1 乳癌基礎病理 9
2.1.1 乳癌的病理診斷 9
2.1.2 零期乳癌的病理變化 9
2.1.3 侵犯性乳癌的病理變化 11
2.1.4 乳癌的組織學分級及預後的關連性 12
2.2 影像處理技術探討 13
2.2.1 色彩空間 13
2.2.1.1 RGB彩色模型 14
2.2.1.2 HSV彩色模型 15
2.2.2 影像分割 16
2.2.2.1 Otsu分割演算法 17
2.2.2.2 K-Means 分割演算法 18
2.2.3 形態學後處理 19
2.2.4 Connected Component Labeling演算法 20
2.2.5 細線化 22
2.2.6 分水嶺演算法 24
2.2.7 支援向量機(Support Vector Machine) 27
2.3 機率神經網路（Probabilistic Neural Network ：PNN） 30
2.3.1 貝氏決策定理與Parzen估測理論 31
2.3.2 機率類神經網路設計 34
2.4 系統準確性評估 35
2.5 ROC曲線 36
第三章、乳癌細胞分割與惡性度分級 39
3.1 介紹 39
3.2 影像擷取與研究設備 41
3.2.1影像擷取 41
3.2.2研究設備 41
3.3 研究步驟－乳導管分割(100倍成像) 42
3.3.1 K-Means演算法之乳導管分割與標示 45
3.3.2 內外導管管腔分割與邊界取出 47
3.3.3 標記分水嶺進行導管分割 50
3.3.4 導管骨架偵測 51
3.3.5 導管特徵參數量測 52
3.3.5.1 外導管對稱性量測 52
3.3.5.2 導管厚度增生比例量測 56
3.3.5.3 外導管曲率變化比例量測 57
3.4 有絲分裂細胞核(400倍成像) 59
3.4.1 影像擷取 60
3.4.2 辨識流程探討 60
3.4.3 研究方法與步驟 62
3.4.3.1 Otsu分類分割法找出臨界值分割 62
3.4.3.2 區域成長法分割出病變細胞核 64
3.4.3.3 特徵參數量測面積與去除人工雜點區域 66
3.4.3.4 填滿細胞核內部空洞與初始輪廓偵測 67
3.4.3.5 特徵參數擷取(Feature Extraction) 68
3.4.3.6 量測病變細胞核之緊密度 68
3.4.3.7 細胞核所有輪廓點之曲率變化 69
3.5 核分裂程度(400倍成像) 71
3.5.1 影像強化 72
3.5.2 Otsu分閥值分割與後處理 73
3.5.3 形態學後處理 74
3.5.4 分離重疊細胞核 75
3.5.5 細胞核面積量測與分級 77
第四章、實驗結果與討論 78
4.1 組織惡性度分級系統介面 78
4.2 系統效能驗證 81
4.2.1 特徵選取與參數正確性評估 81
4.3 系統分類器比較 82
4.3.1 導管結構分析子系統 82
4.3.2 有絲分裂分析子系統 84
4.4 實驗分析結果及病理探討 86
4.4.1 導管結構分析 86
4.4.2 有絲分裂細胞核分析 88
4.4.3 細胞核大小分析 89
4.4.4 電腦預後分析之準確性評估 92
第五章、結論與未來展望 95
5.1 結論 95
5.2 未來展望 96
參考文獻 97

[1] Connolly JL, Fechner RE, Kempson RL, et al: Recommendations for the reporting of breast carcinoma. Association of directors of anatomic and surgical pathology. Human Pathol. 1996; 27: 221-4.

[2] Fitzgibbons PL, Conolly JL, Page DL: Updated protocol for the examination of specimens from patients with carcinomas of the breast. A basis for checklist. Arch Pathol Lab Med. 2000; 124: 1026-33.

[3] Latson L, Sebek B, Powell KA. “Automated cell nuclear segmentation in color images of hematoxylin and eosin-stained breast biopsy,” Anal. Quant. Cytol. Histol. 2003 Dec：25(6):321-31.

[4] E.J. Kaman, A.W.M. Smeulders, P.W. Verbeek, I.T. Young, and J.P.A. Baak. “Image Processing for Mitoses in Sections of Breast Cancer: A Feasibility Study,” Cytometry pp.1-6, 1984.

[5] Kate T.K., Belien J.A.M., Semulders A.W., and Baak J.P.A, “Method for counting mitosis by image processing in Feuglen stained breast cancer sections,” Cytometry Vol. 14, pp. 241-250, 1993.

[6] Yang, J. and J. Y. Yang, "Why can LDA be performed in PCA transformer space," Rapid and Brief Communication on Pattern Recognition, pp.563-566, 2003.

[7] Belien J.A.M., Baak J.P.A, Van diest P.J., and Van Ginkel A.H.M, “Counting mitosis by image processing in Feuglen stained breast cancer sections:the influence of resolution,” Cytometry, Vol. 28, pp. 135-140, 1997.

[8] H. Refai, T. K. Teague, “Automatic Count of Hepatocytes In Microscopic Images,” Image Processing, ICIP 2003. Proceedings. 2003 International Conference on Volume 2, 14-17 Sept. 2003 Page(s): II - 1101-4 vol.3.

[9] Elston CW, Ellis IO: Pathologic factors in breast cancer. The value of histological grade in breast cancer: Experience from a large study with long-term follow-up. Histopathology 1991; 19: 404-10.

[10] G. Paschos, “Perceptually uniform color spaces for color texture analysis: an empirical evaluation,” IEEE Transactions on Image Processing, Vol.10, No.6, June 2001, pp 932-937.

[11] N. Otsu, “A Threshold Selection Method from Gray-Level Histogram,” IEEE Trans. System Man Cybernetics, SMC-9(1): no.1, 1979, pp 62-66.

[12] Thomas Dietterich, “Introduction to Machine Learning,” Ethem Alpaydin, 2004.

[13] Wu, H.-S. Barba, J. Gil, J., “Region growing segmentation of textured cell images,” Journal, Vol.32, No.6, June 1996, pp 1084-1085.

[14] J.M. Park, C.G. Looney and H.C. Chen, “Fast Connected Component Labeling algorithm Using A Divide and Conquer Technique,” Conference on Computers and Their Applications, 2000, pp 373-376.

[15] Milan Slnka, Vaclac Hlavac, Roger Boyle, “Image Processing Analysis and Machine Vision,” 1997, pp 177-181.

[16] R.C. Gonzalez and R.E. Woods, “Digital image processing, 2ed,” Massachusetts: Addison Wesley, 2002.

[17] Zhung and Suen(T.Y. Zhang and C.Y. Suen, “A fast parallel algorithm for thinning digital pattern,” Communications of the ACM, Vol. 27, No. 3, pp. 236-239,1984.

[18] K. Haris, S.N. Efstratiadis, and N. Maglaveras (1998) “Watershed-based image segmentation with fast region merging,” in Proceedings of the IEEE International Conference on Image Processing, vol. 3, pp. 338-342.

[19] L. Vincent and P. Soille (June 1991) “Watersheds in digital spaces: an efficient algorithm based on immersion simulations,” IEEE Trans. On Pattern Anal. Machine Intell., vol. 13, no. 6, pp. 583–598.

[20] K. Haris and S.N. Efstratiadis (Dec. 1998) “Hybrid image segmentation using watersheds and fast region merging” IEEE Trans. on Image Processing, Vol. 7, No. 12, pp. 1684-1699.

[21] Chen Pan, Cong-Xun Zheng, and Hao-Jun Wang, “Robust color image segmentation based on mean shift and marker-controlled watershed algorithm,” Second International Conference on Machine Learning.

[22] S. Taylor, N.C.a.J., An Introduction ro Support Vector Machines and Other Kernel-based Learning Methods. 2000: Cambridge University Press.

[23] D.F. Specht, “Probabilistic Neural Network for Classification, Mapping, or Associative Memory,” IEEE Int Conference. Neural Network, San Diego, CA, Vol.1, July 1988, pp 525-532.

[24] Ethem Alpaydin, Introduction to Machine Learning, October 2004, pp 39-58.

[25] Parzen, E. , “ On Estimation of a Probability Density Function and Mode,” Annuals of Mathematical Statistics,1962, pp 1065-1076.

[26] Specht,D. F., “Probabilistic Neural Networks (original contribution) ”,Neural Networks .vol.3,no.1 Jan 1990,pp. 109-118.

[27] The Eurographics Association and Blackwell Publishing, Computer Graphics Forum, Vol.23 Issue 3, pp 271-279.

[28] A. Rosenfeld and E. Johnston, “An improved method of Angle Detection on Digital Curves”, IEEE Transaction on Computers, Vol. C-22, 1973, pp 875-878.

[29] J.M. Thijssen, B.J. Oosterveld, and R.F. Wagner, “Gray level transforms and lesion detectability in echographic images,” Ultrason. Imaging, vol.10, no.3, pp. 171-195, July, 1995.

[30] 朱娟秀，”乳癌病理診斷”，台灣醫學第五卷第五期，P.552-556，2001。

[31] 馮齡儀， ”電腦輔助子宮頸抹片異常細胞辨識之初期研究”，中原大學醫學工程研究所碩士論文，2004。

[32] 楊勝智, ”乳房醫學影像之腫瘤電腦輔助診斷系統”，國立成功大學電機工程學系博士論文，2006。

[33] 張祐育,”子宮內膜異位症之組織影像特徵參數分析”，南台科技大學電機工程研究所碩士論文，2004。

[34] 林俊志,”以細胞學及影像處理技術分析乳房腫瘤”，國立嘉義大學資訊工程所碩士論文，2007。

[35] 許迪程,”乳癌免疫組織化學染色細胞分割與判讀”，國立嘉義大學資訊工程所碩士論文，2007。

[36] 林儀信,”以組織學為基礎之乳癌細胞分割及辨識”， 國立嘉義大學資訊工程所碩士論文，2006。

[37] 乳癌治療相關因子，財團法人天主教耕莘醫院檢驗科，李麗秀、侯惠真。

[38] 鐘國亮，”影像處理與電腦視覺”，東華書局，2002。