基於超頻譜影像系統分類上皮組織與固有層實現黏膜組織癌症檢測

Abstract

惡性腫瘤(癌症)在過去數十年來早已成為開發中國家與已開發國家的主要死因，除了造成病人與家屬的經濟重擔外，相關的醫療照護也成為國民健康的主要議題。根據世界衛生組織(WHO)統計，在2008年，全球預估約有一千兩百萬的癌症新增病例，並造成了約七百六十萬人的死亡，因此癌症的篩檢與檢測已成為近代醫學研究的主要方向之一。 本研究利用本團隊研發的嵌入式繼光鏡顯微超頻譜影像系統進行黏膜組織的癌症檢測，利用癌化的黏膜組織，會由上皮組織產生細胞變異並入侵固有層組織的特性來實現癌症檢測。相較傳統用肉眼對彩色影像以經驗法則做主觀的診斷，超頻譜影像提供連續且寬頻的頻譜資訊與二維影像，在分析上有更多的資訊進行客觀的判斷。 首先，影像中的細胞核作為主要的判別目標，細胞核的光譜資訊抽取出來後以主成分分析(PCA)與費雪線性判別分析(Fisher's linear discriminant)進行上皮組織與固有層組織的鑑別訓練，最後根據訓練資料，將所有細胞核以單純貝氏分類器(NBC)、K最鄰近演算法(KNN)與支持向量機(SVM)進行分類。將所有細胞核分類完成後，為量化病理學上癌化細胞的變異，細胞核的熵 (Entropy)用以計算細胞核的亂度與生長異常、細胞的碎形維度(fractal dimension)用以計算上皮組織的破碎程度、最後再以型態學影像分析估算兩種組織的細胞混合程度，結合這三種方法可以準確的偵測出癌化的切片樣本。 三種篩檢方法結合後，最後的篩檢結果以靈敏度(sensitivity)與特異度(specificity)來衡量其偵測效果，我們以兩個方向取得篩檢閥值：一為最大準確率，另一為兼顧靈敏度與特異度。前者的最終篩檢效果為靈敏度97.06% 特異度88.24%，而後者的最終檢驗結果為靈敏度94.12%特異度度91.18%。

Malignant neoplasm (also known as cancer) has been the main cause of death for decades in both developing and developed countries. As such, cancer has become an economic burden and a huge national healthcare issue for all countries affected by this disease. Globally, 12.7 million cancer cases and 7.6 million cancer deaths were evaluated in 2008. Cancer detection and screening have been important issues for decades. In this study, a unique embedded relay lens hyperspectral imaging system was used for cancer detection in mucosa tissues, with oral mucosa tissues being obtained for the experiment. The analysis used both spectral profiles and spatial information to judge the experimental samples. All nuclei in the images were identified; the feature profiles of the hyperspectral training data were extracted by principal component analysis (PCA) and Fisher’s linear discriminant; and all nuclei were recognized by three classifiers. According to fundamental pathological changes in cancerous mucosa tissue, three methods were proposed to distinguish between healthy and cancerous tissue. The entropy of the nuclei was calculated for measuring the nuclei changes, and the fractal dimension was calculated as a measurement of completeness of the epithelial tissue. A combination or mixture of two classes of nuclei was also evaluated using morphological imaging processes. By combining the three methods, the defects of each method could be redressed by consulting the two other methods. There were two final results due to two sets of discrimination thresholds being chosen based on different methods. The sensitivity and specificity of the final results were 97.06% and 88.24% or 94.12% and 91.18%.