使用細胞核特徵預測大腸癌的淋巴結轉移

Abstract

近年來台灣地區隨著生活品質的提升以及飲食結構的改變，大腸癌發生率逐年增加。大腸癌的危險因子包括與DNA修復相關的基因、多肉少蔬果的飲食習慣、運動、肥胖以及菸酒等都會增加罹患大腸癌之風險。大腸癌的形成是從正常的瘜肉逐漸長大並且演變成癌細胞。癌細胞持續生長可能會經由大腸周邊血管與淋巴結擴散轉移到身體其他部位。臨床上大腸癌分期的判斷標準仰賴切片取得組織以H&E染色後，病理科醫師用顯微鏡觀察組織型態，判斷淋巴結組織切片是否有癌細胞轉移之情況並做出診斷。 本研究使用的大腸周邊淋巴結組織切片為中國醫藥大學所提供。利用影像處理技術分割出組織切片中的細胞核，透過51個細胞核特徵擷取和特徵選取以及支持向量機等建立分類模型，預測淋巴結組織切片是否有癌細胞轉移，以及預測未觀察到癌細胞入侵淋巴結組織的癌化程度。預期提高判斷大腸癌分期的準確率以利醫師能針對不同癌化程度的大腸癌病患進行適當診斷與有效治療。 本研究使用一位大腸癌病患的四顆淋巴結組織切片影像作為訓練資料以及51個細胞核特徵值建立分類模型，目的為分類大腸周邊淋巴結組織切片影像是否有癌細胞入侵轉移。實驗結果為訓練資料中10折交叉驗證的AUC為0.9972，同病人預測的AUC為0.9997，跨病人預測的AUC為0.9008以及非大腸癌病患預測的準確率為96.55%。由測試結果可知，此分類模型具有良好的辨識率。並且利用此分類模型對8顆淋巴結組織切片影像進行預測取得決策值並用Heat map來表示淋巴結癌化程度。

Nowadays, the colorectal cancer has become an increased incidence of cancer in Taiwan since the promotion in qualities of life and changes in eating habits. The risk factors for colorectal cancer include DNA repair genes, low fruit and vegetable consumption, high-meat diet, exercise, obesity, smoking, alcohol and so on will increase risk of getting colorectal cancer. Colorectal cancer often comes from growing polyps of which will become cancer over time. The cancer has spread through the lymph nodes and blood to other parts of the body. The standard of diagnosis of the colorectal cancer is the lymph node biopsy with H&E stain that is able to stage the colorectal cancer and to offer the diagnosis of the cancer and suggestions for the prognosis. In this study, the lymph node biopsies in colorectal cancer are provided by China Medical University. This study use techniques of image segmentation, feature extraction, feature selection and SVM to build a high accuracy prediction model to predict lymph node metastasis or not in colorectal cancer and analysis the degree of cancer of lymph nodes without observing tumor cells. This study used one colorectal cancer’s 4 lymph node biopsies as training data and 51 nuclei-based features to build a classification model to predict lymph node metastasis or not in colorectal cancer. The result shows that the AUC from 10-fold validation of training data is 0.9972. The AUC of the prediction for the same patient is 0.9008 and the accuracy of the prediction for the patient with colorectal cancer is 96.55%. In short, the discrimination capacity of the classification model is good. Moreover, this study used this model to predict 8 whole lymph node biopsies to get the decision values and used heat map to show the degree of cancer.