LED激發之自體螢光影像感測系統透過多頻譜實現口腔癌檢測

Abstract

口腔癌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之一，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2012年，全球約有超過三十萬的口腔癌新增病例，並造成約十四萬五千例死亡。因此口癌症的篩檢與檢測已成為近代醫學研究的主要方向之一。 本研究開發一種手持式影像感測系統，搭配特定的LED激發光源以及濾光片拍攝口腔影像進而檢測口腔癌，其中手持式影像感測系統根據濾光片的數量以及LED激發光源的強度可分為四通道以及七通道兩種版本。手持式影像感測系統利用口腔癌化的黏膜組織，會對特定波長激發光源有不同於正常組織之螢光特性實現口腔癌症檢測。相較傳統用肉眼對彩色影像以經驗法則做主觀的診斷，此裝置提供多頻段之螢光影像，在分析上有更多的資訊進行客觀的判斷。 特定波長刺激光下的口腔黏膜組織會透過濾光片產生多頻段的螢光影像，影像將分成RGB三種灰階影像進行分析。藉由口腔黏膜組織的不正常增生所導致的型態變化可由影像中之熵值(Entropy)及碎形維度(Fractal Dimension)作為分析方法，而組織中的螢光變化即以影像中之強度(Intensity)以及直方圖(Histogram)作為分析方法。 每種分析數據將以整體當作訓練組及驗證組，以及分成兩組做交叉驗證。最後以七通道手持式影像感測系統下，以405nm光源搭配Bi505濾光片所拍攝得到的綠色(G)灰階影像在以影像強度方法做分析方法下能得到最好的平均準確率為83.48%

之後本研究將以兩種光源、各種濾光片、三種灰階影像以及四種分析方法以權重結合進而提升檢測效能，權重由整體訓練下的AUC所算出並以此權重將各種組合之分析方法結合，進而得到更好的準確率。最終以四通道手持式影像感測系統檢測下可得到整體驗證結果之靈敏度、特異度以及準確率分別為94.44%、85.18%以及89.81%，而七通道手持式影像感測系統檢測下則可得到靈敏度、特異度以及準確率分別為88.35%、88.35%以及88.35%。 最後為了找出最有效判別口腔癌的濾光片，將特定濾光片挑出並觀察其檢測結果。在四通道手持式影像感測系統下，在BP525濾光片下之方法結合能得出最好的結果，其靈敏度、特異度以及準確率分別為90.74%、68.52%以及79.63%，相較於沒有挑選濾光片下之方法結合之檢測準確率低了不少。七通道手持式影像感測系統則可在BN470以及Bi505兩種濾光片檢測下得出近九成的準確率，在BN470濾光片下的分法結合能得出靈敏度、特異度以及準確率分別為89.32%、86.41%以及87.86%，而在Bi505濾光片下則可得出最好的靈敏度、特異度以及準確率分別為89.32%、89.32%以及89.32%。

Oral cancer has been one of the main causes of death in both developing countries and developed countries over the past few decades. In addition to the economic burden of patients and their family, related medical care has been the main issue of national health. According to World Health Organization (WHO), more than 300000 new case of oral cancer has occur and caused 145000 death in 2012. Therefore, the detection of oral cancer has been the main topic of modern medical research. In this study, a hand-held imaging system with light-emitting diode (LED) excitation light and emission filter was used to screening oral mucosa tissues for oral cancer detection. The hand-held imaging system can be divided into two versions based on the number of the filter on the filter ring and the intensity of the light-emitting diode. The filter ring of version 1 and version2 imaging system contain four places and seven places for emission filter respectively.The difference between tumor tissue and normal tissue can be discriminate by the excitation light of LED with specific wavelength. Compare with the traditional method of naked-eye detection, the method of this study using multispectral fluorescent image provide a more objective detection method. Image will be excited by the light-emitting diode (LED) and pass through the selected spectral filter. Each grayscale intensity image of the fluorescent image’s RGB intensity is used for analysis. Both spatial and spectral methods are used in the analysis. Spatial analytical methods include entropy and fractal dimension, and spectral analytical methods include intensity and histogram. Data used in analysis will divide into two groups to implement the self-validation and cross-validation procedure. Finally, the average accuracy of self-validation and cross-validation is 83.48% using version 2 hand-held imaging system by 405nm excitation with Bi505 emission filter. In order to increase the efficiency of oral cancer detection, all kinds of excitation LED, emission filter and methods will be combined with the weight calculated by AUC. Finally, the sensitivity, specificity and accuracy of oral cancer detection are 94.44%, 85.18% and 89.81% respectively using version 1 imaging system. As for version2 imaging system, the sensitivity, specificity and accuracy of oral cancer detection are 88.35%, 88.35% and 88.35%. Finally, the specific emission filter will be picked up for oral cancer detection. Using version 1 imaging system with BP525 emission filter for oral cancer detection, the sensitivity, specificity and specificity are 90.74%, 68.52 and 79.63%.Compare with the full combination of methods, the oral cancer detection by BP525 emssion filter has much lower performance. Compare to version 1 imaging system, version 2 imaging system with BN470 and Bi505 emission filter have better performance. Version 1 imaging system with BN470 emission filter for oral cancer detection, the sensitivity, specificity and specificity are 89.32%, 86.41% and 87.86%, and the sensitivity, specificity and specificity can be better with Bi505 emission filter which are 89.32%, 89.32% and 89.32%.