使用細胞拓墣特徵預測大腸癌的淋巴結轉移

Abstract

近年來由於國⼈⽣活和飲⾷習慣的改變，加上遺傳疾病和基因等因素，造成⼤腸癌的發⽣率快速攀升，⽽⼤腸癌在台灣⽬前已是發病數⽬最多的癌症。罹患⼤腸癌時，癌細胞可能經由⼤腸周圍的淋巴管引流進⼊淋巴結，進⽽擴散並轉移到⾝體其他器官或組織。⽬前臨床上，判斷⼤腸癌的分期仍需仰賴病理師以顯微鏡觀察組織切⽚的細胞型態，診斷其組織是否有癌細胞轉移之情況。現今已有許多影像分類的相關技術應⽤以及針對醫學影像的⾃動分類亦已被廣泛研究於癌症之診斷與預測中。本研究即利⽤影像處理技術，經由特徵擷取和特徵選取，並將這些細胞拓墣特徵組合以⽀持向量機進⾏訓練以及分類，再以此建⽴分類模型來預測⼤腸癌的淋巴結轉移。本研究使⽤的⼤腸周邊淋巴結WSI 是由中國醫藥⼤學所提供，並以⼤腸癌病患中病⼈代號1 的淋巴結組織切⽚影像作為訓練資料，再使⽤3 個細胞拓墣特徵的組合建⽴最佳分類模型以預測癌細胞⼊侵轉移程度。實驗結果分別為：訓練資料10 折交叉驗證的AUC 為0.9858、同⼀病⼈測試資料的AUC 為0.9919、跨⼤腸癌病患測試的AUC 為0.8979 以及測試⾮⼤腸癌病患的準確率為95.68%。綜合以上結果可知，以此分類模型測試組織切⽚影像的辨識率良好。最後使⽤libsvm tools 進⾏預測所獲得之決策值來畫 出heat map 圖藉以表⽰⼤腸癌周圍淋巴結組織之癌化程度。

Recently, the factors such as alternation in dietary, life style, gene and genetic diseases has made the incident rate of colorectal cancer raised rapidly and colorectal cancer has already become the most morbidity cancers in Taiwan at present. Cancer cells may transfer to other organs through the lymphatic nearby colon as suffering from colorectal cancer. At present, the way to diagnose the colorectal cancer grading in clinic still depends on observing cytology in biopsy by pathologist whether there is metastasis or not. Nowadays, there has existed many applications in techniques of image classification and the similar issues with automatic classification in histopathological images have been widely studied for detecting cancer. This study uses the techniques of image processing and combines the subsets of topological features of cells to conduct the training and classification steps in SVM. After that, the model is used to predict lymph node metastasis of colorectal cancer. The WSIs of lymph node nearby colon used in this study were provided by China Medical University. In this study uses the lymph nodes’ WSIs from patient 1 as training data. According to the result of feature selection, the optimal model used to predict metastasis is constructed by using 3 topological features. As the performance of predicting test data by using this optimal model, the AUC of 10-fold cross validation from training data is 0.9858; the AUC of predicting test data from same patient is 0.9919; the performance of AUC in cross subject is 0.8979 and the accuracy of predicting test data from patients without colorectal cancer is 95.68%. According to these results, this optimal model shows great performance in identifying histopathological images. Finally, plotting the heat map according to decision value from libsvm tools is used to express the lymph node metastasis of colorectal cancer.