以邊緣偵測與區域成長為基礎之ThinPrep子宮頸抹片影像切割

Image Segmentation Base on Edge Detection and Region Growing for ThinPrep-Cervical Smear

陳俊杰(Chun-Chieh Chen)

指導教授 ： 林春宏

國立臺中科技大學/資訊學群/資訊科技與應用研究所/碩士(2009年)

https://doi.org/10.6826/NUTC.2009.00070

摘要

本論文主要是發展一套偵測(detection)與切割(segmentation)技術，精確地從新柏氏電腦超薄(ThinPrep)子宮頸抹片不同倍率之上皮細胞(epithelial cells)影像中，切割出細胞質(cytoplasm)與細胞核(cell nucleus)輪廓，以便分析出子宮頸細胞正常(normal)或異常(abnormal)現象。
子宮頸癌(cervical cancer)為婦女最常見的癌症之一，亦位居婦女十大死亡原因之首。採用子宮頸抹片檢查做為子宮頸癌之偵測已達數十年之久，該項技術同時也是預防子宮頸癌的最佳方法。ThinPrep是一項專門針對改善抹片採檢品質的一大革新發明，目前美國食品藥物衛生管理局(FDA)已通過ThinPrep為合格的抹片製作方式，且已經廣泛運用於婦產科採檢，其準確度高達93%，是子宮頸抹片細胞影像分析之優良題材。
該論文的主要目的，係希望在藉由電腦的輔助之下進行自動判讀，以減少人為疏忽、提高影像判讀的效能，及提升我國子宮頸癌細胞抹片的篩檢能力及檢查單位之篩檢品質。子宮頸細胞常依據細胞核大小、細胞核質輪廓長度、核質比及核染顏色的深淺等特徵，來進行抹片判讀。然而，上述特徵中除了核染顏色外，其餘特徵皆仰賴於精確的細胞輪廓切割技術。因此，如何從一子宮頸細胞抹片影像中，確地分割出背景、細胞質與細胞核影像區域，便顯得特別的重要。
本研究於影像處理技術中，首先採用mean filter，有效的去除影像中的雜訊。其次，利用Sobel邊緣偵測(Sobel edge detection)產生物件輪廓邊緣線之梯度，再經由假邊緣處理(false edge processing)，該方法係以分析真實邊緣梯度大小與真實邊緣梯度在其鄰近範圍內之特性，有效的保留真實邊緣梯度，去除雜訊梯度。在物件輪廓加強之梯度方面，主要是假設梯度呈 函數分佈，以做為加強物件輪廓梯度之依據。然而，於物件輪廓邊緣線之細線化方面，本文採用非最大值刪除(non-maximum suppression)技術將邊緣線以線形方式顯示，再配合遲滯性門檻(hysteresis thresholding)來決定物件的初始邊緣輪廓線。在邊緣萃取方面，主要係透過近似細胞輪廓修補之處理，以確保細胞核輪廓的完整性，藉由細胞核邊緣線段輪廓為簡單封閉曲線之定義，分析該曲線內之細胞核邊緣與細胞核內部的色階特徵，進而萃取出細胞核輪廓。最後，將細胞核輪廓線上的每一像元定義為初始種子點(initial seed)，並運用區域成長(region growing)演算法配合自適應性門檻與真實邊緣門檻值之設定，切割出細胞質區域。
於實驗部分，本文將分別以ThinPrep子宮頸抹片細胞之100倍、200倍、400倍率的表層細胞(superficial cell)、中層細胞(intermediate cell)、旁基底細胞(parabasal)及異常細胞等醫學影像，做一系列的細胞輪廓偵測與切割之實驗分析。

關鍵字

ThinPrep子宮頸抹片 ； 上皮細胞 ； 倍率 ； 偵測 ； 切割 ； 邊緣 ； 輪廓 ； 區域成長

並列摘要

In this thesis, a system is developed to detect contours of cytoplasm and cell nucleus and segment them from low-magnification ThinPrep image of epithelial cells. This system reads both normal and abnormal cells and determine whether it is necessary to observe epithelial cells on the smear at higher magnification.
Cervical cancer is one of the common cancers among women and tops the causes of death of women. Cervical smear has been the most effective means for screening and preventing cervical cancer in the last ten years. And ThinPrep, as a revolutionizing invention in improving the sampling quality of cervical smear, has been not only approved by FDA in the U.S., but also widely adopted in gynecological examination, as its accuracy of detection is as high as 93%. Therefore, ThinPrep-sampled cervical smear is most suitable for cell image analysis and research.
The objective of this thesis is to maximize the performance of cervical cell detection by applying computerized reading. Through computerized reading, human errors in detection can be avoided, and the quality of institutions conducting cervical smear examination in Taiwan can be improved. The smear reading technique is based upon the size of cell nucleus, the length of the contours of cytoplasm and cell nucleus, the ratio of the two, the shade of the chromatin color, and etc. Except the last feature, the rest of them all require accurate contour segmentation. Thus, accurate segmentation of cytoplasm and the cell nucleus from the background becomes very crucial. .
In order to effectively eliminate noises of image, we adopt the mean filter technique. Sobel edge detection is utilized to demonstrate the gradient of the contour edge. And we adopt the false edge processing to remove the noise gradient. The false edge processing is utilized to retain the gradient of real edge by analyzing the size of the gradient and the relationship between the neighboring pixels. The gradient value is calculated by function , and its curve line may be accentuated in order to serve as a basis for the reading. The contour edge line will be linearized by non-maximum suppression technique and its preliminary contour edge will be determined by Hysteresis Thresholding. In order to ensure the integrity of the contour of the nucleus, we adopt approximate contour of the cell repair technique. We have defined the contour of the nucleus for simple closed curve, so the contour of the nucleus is extracted by analyzing the feature of gray-level within each contour. Finally, we set the pixels which on the contour of the nucleus to initial seed and use region growing algorithm with adaptive threshold and the true edge of threshold to segment the region of the cytoplasm.
Finally, a series of detection and segmentation experiments on ThinPrep-smear cells, including superficial cells, intermediate cells, parabasal, abnormal cells, and cells magnified 100x, 200x, and 400x are conducted to demonstrate the performance of the proposed system.

並列關鍵字

ThinPrep-cervical smear ； epithelial cell ； low magnification ； detection ； segmentation ； edge ； contour ； region growing

參考文獻

[29] 黃俊魁，“子宮頸抹片細胞之電腦輔助診斷系統”，私立中原大學醫學工程所碩士論文，民國九十五年。

[32] 王仁宏，“子宮頸抹片細胞之參數分析”，私立中原大學醫學工程所碩士論文，民國九十五年。

[5] Koss LG. The Papanicolaou test for cervical cancer detection: a triumph and a tragedy. JAMA 1989;261:737-743.

[7] Gay JD, Donaldson LD, Goetliner JR. False negative results in cervical cytologic studies. Acta Cytologica 1985;29:1043-1046.

[9] Zahniser DJ, Hurley AA: Automated slide preparation system for the clinical laboratory. Cytometry 1996;26:60-64.