Mirau全域式光學同調斷層掃描術結合近紅外光拉曼光譜用於皮膚細胞之影像與頻譜特性分析

Analysis of Image and Spectrum Properties on Skin Cells by Mirau-based Full-field Optical Coherence Tomography Combined with Near-infrared Raman Spectroscopy

游鈐(Cian You)

指導教授 ： 黃升龍

國立臺灣大學/電機資訊學院/光電工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900679

摘要

米繞式全域式光學同調斷層掃描術(Full-field optical coherence tomography; FF-OCT)具有高解析度、低震動雜訊與良好色散補償優勢，可用來觀察樣本形貌與強度分布資訊，但無法得知該樣本的化學組成；近紅外光拉曼光譜儀則具備量測樣本內化學分子組成且低螢光影響與穿透深度較深等優點，但無法獲得樣本的微結構影像。因此將兩者搭配應用，量測體外(in vitro)的五種皮膚細胞，包括角質細胞株、基底細胞癌細胞株、鱗狀細胞癌細胞株、黑色素細胞與黑色素癌細胞株，可同時獲得細胞的微結構、亮暗分布以及細胞組成化學分子資訊。而在臨床應用上，為提高判斷精確度與速度，利用機器學習的整體學習法進行演算與分類。
在in vitro實驗中，就FF-OCT三維影像上，發現癌細胞群的表面皆有明顯的突起，可能與其侵略性相關；在亮度上，角質系列的正常細胞表面較平滑且內部均質而較低。利用軟體半自動擷取出共283顆不同種細胞的三維資訊，含體積、致密度、表面粗糙度、內部平均亮度及內部亮度分布標準差，結果顯示後三項特徵，可以成為分辨癌細胞與否的良好指標，而平均亮度可用來區分出正常細胞中的黑色素細胞與角質細胞。對於癌細胞的分類仍可用致密度與體積的平均值做區分，但其標準差較大，無法成為良好指標，因此須搭配拉曼光譜儀。
在in vitro五種細胞株的拉曼頻譜中，顯現正常細胞的拉曼頻譜相較於癌細胞標準差較大，其可能是生物特性所致，正常細胞不像癌細胞單種繁衍，可能存在多種變異。而癌細胞的頻譜標準差相對較小，且能與文獻對照幾乎相符。在600-2000 cm-1拉曼頻譜中，可用6根peaks與4組bands (Peaks: 746, 780, 857, 1024, 1063, 1209 cm-1; Bands: 925-946, 990-1010, 1088-1130, 1281-1302 cm-1) / 7根peaks與3組bands (Peaks: 854, 898, 1064, 1158, 1191, 1233, 1452 cm-1; Bands: 923-946, 1007-1028, 1291-1336 cm-1)分別區分黑色素細胞癌對角質系列細胞癌 / 基底細胞癌對鱗狀細胞癌，因此在癌細胞分類上拉曼頻譜有較佳的表現。
藉由機器學習中的整體學習法進行分類，也能驗證以上所說。使用整體學習法中的決策樹分析法普遍有較好的分類效果，FF-OCT特徵在區分正常細胞與癌細胞可得準確度為85.9%，且在分辨正常細胞種類上可到達完全區分；而拉曼頻譜則是在分類三種癌細胞群上可達到完全區分。因此搭配FF-OCT與紅外光拉曼技術，並結合機器學習，可以很準確且快速的辨別五種正常與癌化皮膚細胞株，顯示此FF-OCT與紅外光拉曼光學技術搭配整體學習演算法可在皮膚科臨床研究與應用上成為重要的診斷工具與方向。

關鍵字

光學同調斷層掃描 ； 細胞三維特徵擷取 ； 影像分析 ； 拉曼光譜學 ； 皮膚癌 ； 機器學習 ； 整體學習法

並列摘要

Mirau-based full-field optical coherence tomography (FF-OCT) has many advantages such as high resolution, low vibration noise, and good dispersion compensation. Thus it can observe the morphology and intensity distribution of the sample but not its chemical component. Near-infrared (NIR) Raman spectroscopy can recognize the chemical institution of the sample with the advantages of low fluorescent effect and deep penetration, but not its microstructure. Therefore, measuring the five skin in vitro cell lines, including Keratinocyte cell lines, Basal cell carcinoma cell lines, Squamous cell carcinoma cell lines, Melanocyte cell and Melanoma cell lines, through combining both two techniques can acquire the microstructure, intensity distribution and chemical molecular information inside cells. Then in order to increase the accuracy and speed of the judgment in the clinical application, we use the ensemble learning algorithm of machine learning to calculate and classify them.
In the in vitro experiments, the FF-OCT three-dimensional (3D) images of the five skin cell lines show the obvious protrusion on the surface of cancerous cells, that may relate to the aggressiveness of cell. And due to the smoother surface and more homogeneous internal space of the normal keratinocyte cell lines, they have lower average intensity. We utilized software to semi-automatically capture totally 283 different cells’ 3D information, including volume, compactness, surface roughness, internal average intensity, and internal intensity standard deviation. The results imply the last three features can be great parameters to distinguish between cancerous and normal cells, and the internal average intensity can be used to classify the normal melanocyte cell and keratinocyte cell. The features included compactness and volume can also distinguish different cancerous cells, but the standard deviation of these two features is too large, thus they can not be excellent indicators for classification. As a result, it needs combining the Raman spectroscopy.
On the Raman spectrum of different in vitro skin cell lines, they indicate the standard deviation of the Raman spectrum from normal cells is larger than ones from cancerous cells, that may due to the biological characteristics of normal cells, that is to say, normal cells have larger variation because the cancerous cells reproduce them within one species. On the other hand, the standard deviation of Raman spectrum form cancerous cells is smaller, and also these Raman spectrums are seemingly corresponding with the literature. We can successfully classify the melanoma cells with keratinocyte-based cancerous cells / the basal cell carcinoma cells with the squamous cell carcinoma cells by six peaks and four bands (Peaks: 746, 780, 857, 1024, 1063, 1209 cm-1; Bands: 925-946, 990-1010, 1088-1130, 1281-1302 cm-1) / seven peaks and three bands (Peaks: 854, 898, 1064, 1158, 1191, 1233, 1452 cm-1; Bands: 923-946, 1007-1028, 1291-1336 cm-1) in the Raman spectrum of 600-2000 cm-1. Consequently, the classification of Raman spectrums on cancerous cells has better performance.
Eventually, we employ the ensemble learning form machine learning to classify them, which also verify the above results. Decision tree method in ensemble learning has better-classified results generally. The accuracy can reach 85.9% on distinguishing normal and cancerous cell by FF-OCT features. Also, these features can distinguish every species of normal skin cell. Then Raman spectrums can completely classify three kinds of cancerous skin cell. Therefore, these techniques can distinguish these five normal and cancerous skin cell-lines very accurately and fast. And it shows the method of integrating FF-OCT, NIR Raman spectroscopy and ensemble learning can be an important diagnostic tool and direction for clinical research and application.

並列關鍵字

Optical coherence tomography ； 3D features extraction from cells ； image analysis ； Raman spectroscopy ； skin cancer ； machine learning ； ensemble learning

參考文獻

[1] N. Kourkoumelis, I. Balatsoukas, V. Moulia et al., “Advances in the in Vivo Raman Spectroscopy of Malignant Skin Tumors Using Portable Instrumentation,” International Journal of Molecular Science, vol. 16, no. 7, pp. 14554-14570, 2015.

[2] H. W. Rogers, M. A. Weinstock, A. R. Harris et al., “Incidence Estimate of Nonmelanoma Skin Cancer in the United States, 2006,” American Medical Association, vol. 146, no. 3, pp. 283-287, Mar. 2010.

[3] S. E. Hoey, C. E. Devereux, L. Murray et al., “Skin cancer trends in Northern Ireland and consequences for provision of dermatology services,” British Journal of Dermatology, vol. 156, pp. 1301-1307, 2007.

[4] M. R. Donaldson and B. M. Coldiron, “No End in Sight: The Skin Cancer Epidemic Continues,” Seminars in Cutaneous Medicine and Surgery, vol. 30, pp. 3-5, 2011.

[5] J. Sng, D. Koh, W. C. Siong et al., “Skin cancer trends among Asians living in Singapore from 1968 to 2006,” Journal of the American Academy of Dermatology, vol.61, no. 3, pp. 426-432, 2009.