- 學位論文
電腦輔助子宮頸抹片異常細胞辨識之初期研究
Computer-aided Recognition for Abnormal Cells in Pap Smear
摘要
子宮頸癌為台灣地區癌症的致命疾病之一，其發生率高居五大婦癌的榜首。傳統上以子宮頸抹片檢查為預防子宮頸癌最好的方式，一般配合子宮頸陰道鏡影像檢查，提供醫師在細胞型態、癌化程度與是否需要進一步施行組織切片檢查的主要依據。本研究主要為『子宮頸抹片異常細胞辨識之初期研究』，希望透過影像處理技術的應用，分析與子宮頸癌發生之可能相關細胞型態及特徵參數，藉以提供臨床醫師在子宮頸癌前期之診斷輔助。 本研究直接擷取子宮頸抹片影像作分析，將影像分為彩色影像與灰階影像兩部分進行。在彩色影像部分，透過RGB與HIS模型運算所得之參數提供不同損傷程度之細胞其表現差異。而在灰階影像部分，針對影像運用Histogram Equalization提高影像之對比度，利用灰度伴隨矩陣分析影像之紋理表現，最後由主動輪廓模型圈選出欲分析之細胞，計算其細胞核大小、核質比及細胞輪廓長度等參數。研究過程中，利用假體影像及標準圖譜共30張影像來測試系統之參數準確性與可行性，並針對臨床子宮頸癌病例共149張影像進行實際診斷鑑別及分析探討。此外，也藉由臨床醫師之協助比較系統與傳統人工鑑別細胞正常或異常的差異性。 初步結果顯示，系統不論在薄層抹片或傳統抹片影像上初步辨識細胞正常或是異常時，其Accuracy、Sensitivity與Specificity值均為1，表示系統在辨識正常細胞與異常細胞上具有良好的鑑別力。而在異常細胞鑑別型態部分，薄層抹片其Accuracy、Sensitivity與Specificity值分別為0.9、1與0.8；而傳統抹片其Accuracy、Sensitivity與Specificity值分別為0.841、0.905與0.783。導致Accuracy與Specificity不高的原因是在薄層抹片分析中有一筆FP值，因其細胞核質比較大，因此系統將該細胞判斷為HSIL；而傳統抹片中有2張HSIL的影像，其在特徵表現上與LSIL近似，因此系統鑑別錯誤；而將系統鑑識結果與人工比對發現，系統可以有效判斷肉眼較無法辨識之異常細胞。 整體而言，系統之完成能夠提供子宮頸癌細胞之形態分析與初步診斷，並能實際輔助臨床上之子宮頸癌細胞鑑別診斷，資料庫之建立亦可提供子宮頸癌治療前後之評估。未來可增加病例數與其他特徵參數之收集，使貝氏網路之訓練更加具有意義以提高系統之正確率，亦可加入異常細胞自動偵測來提高系統之效率。
並列摘要
Cervical cancer is the one of the deadly diseases of cancers in Taiwan, and its occurrence rate is the top of five women cancers. Traditionally, Pap smear is the best treatment for preventing cervical cancer. Originally cooperation with cervical colposcopy to provide doctors cell types, degree of cancer or whether to apply tissue section. This study is focused on the primarily study of Pap smear abnormal cells recognition. We hope to analyze cell types and characteristics parameters relation to cervical cancer via the allocations in image processing to provide diagnostic assistant for clinicians in cervical pre-cancer. Pap smear images were caught to analyze in two different models. In color images, the parameters obtained via RGB and HIS model calculating provided the difference performance between different degree lesion cells. In grey images, histogram equalization was applied to images to enhance the contrast of images and co-occurrence matrix was used to analyze the textures of images. Finally by applied the active contour model to circle the interested cell, and then its cell nucleus’s size, N/C Ratio and cytoplasm’s path were calculated. To totally 30 phantom images and standard plates were used to train this system and 149 clinical cases were used to test the accuracy and feasibility of the system. Furthermore, through the help of clinicians, the comparison between the system and traditional method for the differences in distinguishing the normal and abnormal cells also done. Preliminary results showed the accuracy, sensitivity and specificity were 1 in distinguishing normal and abnormal cells for ThinPrep and Pap smear images. It indicated the system has good identification in distinguishing normal and abnormal cells. In discriminating cells’ type, the accuracy, sensitivity and specificity of ThinPrep were 0.9, 1 and 0.8 individually which due to a FN case caused by its high N/C Ratio, and the accuracy, sensitivity and specificity of Pap smear were 0.841, 0.905 and 0.783 individually which due to two HSIL cases’ features are similar to LSIL. To compare our system with artificial detection can observe our system can diagnose abnormal cells identified by naked eyes. In conclusion, the accomplishment of system can provide the cervical cancer cells’ type analysis and preliminary diagnosis, and practically assist the clinical cervical cancer cells’ discrimination diagnosis. The developed database also can provide the estimation of the treatment before and after. In the future, to increase cases and other characteristic parameters collection can make the Bayesian’s training more meaningful to improve the accuracy of the system, and also can add automatically detect for abnormal cells to raise the system’s efficiency.