利用嵌入式繼光鏡顯微超頻譜影像系統進行口腔癌檢測

Abstract

癌症在國人十大死因的榜首居高不下，其中口腔癌是惡性腫瘤中最可能早期發現，並且藉由早期治療進而痊癒。相較於傳統檢測方法，以肉眼判斷是否有癌症細胞的侵蝕，高光譜影像提供了更多的資訊。我們以嵌入式繼光鏡高光譜成像系統(ERL-MHIS)對細胞切片樣本進行掃描，建立一個三維高光譜資訊矩陣，並且提出了型態與光譜兩類方法進行癌症的辨識。在癌症細胞的影像型態判斷方面，我們提出兩種方法，第一個方法為使用臨界值分離出細胞的基底層，並且使用碎形維度(fractal dimension)計算維度，因為癌症細胞的分裂失去限制，分維的維度值會較正常細胞的維度高。當口腔黏膜細胞發生癌病變時，基底層細胞會持續向內部的固有層持續侵蝕，造成固有層的型態產生變化，因此第二種方法為使用K最鄰近分類算法(KNN)對固有層的細胞核影像做分類，並且計算分類結果的準確率。光譜判斷方面，使用了光譜強度的比值、半波寬度(FWHM)、波形下面積與光譜波段範圍內的強度。將分析結果使用高斯分佈計算準確率。我們將準確率高的前三個方法做結合，並且計算新的準確率 98.45%.最後，藉由考慮樣本資料，我們提出雞尾酒方法，將判斷癌症之螢光光譜的準確率提升到87%。

Cancer has been the leading cause of death for years in Taiwan. Oral cancer has the greatest possibility for early detection and recovery after early treatment. Compared to the traditional method of using the naked eye to detect oral cancer, the method of using the hyperspectral image of tissue can offer more information. We used the embedded relay lens microscopic hyperspectral imaging system to scan the sections and save the hyperspectral image. In this study, we diagnosed oral cancer using two methods: morphology and spectrum. In diagnosis using morphology, we presented two techniques: calculation of the fractal dimension and classification of k-Nearest Neighbor (KNN). In diagnosis using the spectrum method, we presented six techniques: comparing intensity, ratio of intensity, wavelength of peak, area under spectral curve, maximum after spline and full width at half maximum. We calculated the sensitivity and specificity using Gaussian distribution. Combining the 3 methods of the highest specificity provides a specificity of 98.45%. Finally, in accordance with sample data, we presented a cocktail method to increase the specificity of spectral analysis with fluorescence excitation to 87%.