子宮頸癌是台灣發生率與致死率高之疾病,各年齡層的女性都須防範。目前以子宮頸抹片檢查為防範子宮頸癌最普遍的方式。抹片之判讀在診斷中扮演非常重要的角色。因此本研究主要是透過影像處理技術,分析細胞特徵來協助醫師在抹片上異常細胞的偵測。 本研究主要是依據細胞染色特徵與細胞核參數來判斷異常細胞。細胞染色特徵由分析不同型態之細胞其細胞核與細胞質色彩在RGB比例變化來獲得。而細胞核參數部份,本研究利用影像量化與區域成長來判斷細胞核輪廓,藉以得到細胞核其色彩、面積與紋理等參數,並藉由ROC(Receiver Operating Characteristic)分析篩選各參數閥值來判斷細胞之型態。研究過程中,從40張抹片影像手動擷取細胞分析其染色與細胞核參數,進而得到自動圈選與細胞型態判斷之型態,並架構出自動處理整張抹片影像之方法。 研究結果顯示,細胞位置可由細胞核其色彩中(G/B)比值小於1來判斷,且異常細胞其B值比例有變大情況發生。而細胞型態辨識結果:正常與異常細胞核在RGB與面積參數之差異明顯,但抹片上染劑殘留與血球容易被誤判為異常細胞核,因此也利用細胞染色特徵來減低誤判。在10張正常與30張異常抹片辨識結果評估,其Accuracy為0.975、Sensitivity為1、Specificity為0.9,顯示本研究之處理方法對正常與異常抹片有良好之辨別率。而異常型態細胞之辨識部份,在異常細胞中融合型HSIL細胞可藉由細胞核聚集數量判斷,而在細胞核色彩與面積差異不大之HSIL與LSIL細胞,只能依據紋理參數作為判斷,因此LSIL影像有1/4被誤判為HSIL。在10張LSIL與20張HSIL抹片影像之辨識結果,其Accuracy為0.933、Sensitivity為1、 Specificity為0.8,測試結果以異常嚴重之抹片有比較大之辨識。 整體而言,本研究對於異常細胞核有不錯之偵測效率並能達到自動化處理。本研究之介面除了自動將抹片上細胞之型態分類外,也一併把細胞核圈選位置與其參數顯現出來,提供使用者參考與收集資料。而本研究採用影像量化與區域成長方法,降低處理步驟,對於大範圍與多細胞之影像將有不錯之圈選效率。
Cervical cancer is a disease of highly incidence and death rate in Taiwan, and the largest age group for the women must take precautions. At present, Pap smear screening is the most effective way to prevent cervical cancer. This study detects abnormal cells according to features of stained cells and parameters of each nucleus. Features of stained cells obtained via RGB proportions of the Nucleus and Cytoplasm of cells in each lesion. In parameters for each nucleus, we use image quantification and region growing to chose Nucleus contour, herewith we obtain color, area, and textures. Then, ROC analysis ( Receiver Operating Characteristic Analysis ) was used to obtain parameters threshold in order to choose cell type. In this study, we analyze color and nucleus parameters for cells chosen from 40 Pap smear images, then automatically detect and class cells of each lesion, at last accomplish an automatic process for Pap smear. Preliminary result shows the G/B ratio of nucleus color is smaller than 1, and abnormal cell has the higher blue percentage value. In discriminating result, it is obvious to differentiate between normal and abnormal nucleus according to their RGB and area parameters. However stain or bloods on the Pap smear were easily considered abnormal nucleus, so using other stain characteristics of cell to reduce error. In test for 10 normal and 30 abnormal images, the accuracy, sensitivity and specificity were 0.975,1, and 0.9, respectively. It indicated our study's process has a good identification in distinguishing normal and abnormal pap smears. In discriminating cell's lesion degree, HSIL syncytium detection can stand on nucleus's aggregate number, and we use texture parameters to choose HSIL and LSIL while nucleus has the similar color and area. Quarter of LSIL images were considered HSIL accordingly. In test for 10 LSIL and 20 HSIL images, the accuracy, sensitivity and specificity were 0.933,1, and 0.8, respectively. It indicated our study's process has a good identification in high grade lesion pap smears. In conclusion, this system can provide an automatic process serves and identify abnormal nucleus efficiently. Moreover, both positions and parameters of nucleus can provide users compare resources and compile data through the interface of system. Image quantification and region growing which can reduce process cost were used in this study, we expect this can process large image more efficiently.