以GPU加速蒙地卡羅演算法並分析漫反射和螢光光譜

Using GPU to accelerate Monte Carlo simulations and analyze diffuse reflectance and fluorescence spectra

許芳瑋(Fang-Wei Hsu)

指導教授 ： 宋孔彬

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2014年)

https://doi.org/10.6342/NTU.2014.02978

摘要

癌前病變會造成組織型態結構的改變，進而影響組織的光學參數以及量測到的漫反射光譜和螢光光譜。本研究採用蒙地卡羅演算法模擬光子在組織內傳遞的情形，並使用反向擬合工具擷取漫反射和螢光光譜的組織光學參數。反向擬合工具是由GPU版本的蒙地卡羅程式和MATLAB擬合介面所組成。
為了驗證漫反射光譜的測量與反向擬合工具，我們製作多個雙層仿體，並用不同的光纖架構量測光譜。進行反向擬合的過程中，蒙地卡羅程式會分別套用兩種光纖架構和不同的相位函數。相較於垂直光纖架構，斜角光纖架構對相位函數的改變較為靈敏，且對組織上皮層的參數有較準確的擷取結果，例如上層厚度的誤差約9.83%、上層散射係數的方均根百分誤差約12.17%。
螢光光譜的分析部份，我們首先參考文獻制定正常組織和病變組織的光學參數。並模擬不同的光纖架構下，偵測到的上層螢光比率是否提升。根據模擬結果，我們發現病變組織的螢光強度較低，且偵測到的螢光大多源自於組織上層。除此之外，斜角光纖架構比垂直光纖架構更能有效的偵測組織上層螢光，但一直增加光纖傾斜的角度並不一定會提升上層螢光被偵測的比率。至於螢光光譜的反向擬合，我們首先將理論光譜加上3%的雜訊作為輸入值，接著反向萃取組織上下層的螢光參數，包含螢光物質的量子效率和在激發波長下的吸收係數。螢光效率，也就是量子效率和吸收係數的乘積，是我們觀察的重點，因為螢光效率的大小反映了該螢光物質的放光能力，可當作不同螢光物質間相互比較的依據。雖然反向擬合所萃取的螢光參數和理論值差異頗大，但正常組織上層螢光效率的平均誤差為20.75%，下層為15.5%；病變組織上層螢光效率的平均誤差為17.75%，下層為12.5%，誤差都落在可接受的範圍。此結果代表本論文描述的螢光反向擬合流程可確實幫助我們萃取螢光效率的值，對日後實驗量測螢光光譜有很大的幫助。

關鍵字

漫反射光譜 ； 螢光光譜 ； 蒙地卡羅 ； 仿體 ； GPU

並列摘要

The development and progression of neoplasia will change the morphological structures of tissue, thus in turn affect the optical properties of tissue, diffuse reflectance spectra and fluorescence spectra. This research employs Monte Carlo algorithm to simulate how photons propagate in tissue and uses a curve-fitting tool to inversely extract the optical properties of diffuse reflectance spectra and fluorescence spectra. The inverse curve-fitting tool comprises GPU-version Monte Carlo program and MATALB interface for fitting.
In order to analyze diffuse reflectance spectra, we fabricate several two-layered tissue phantoms and use different fiber configurations to measure their spectra. In the process of curve-fitting, the Monte Carlo program would take different fiber configurations and phase functions into account. Compared with perpendicular fiber configuration, oblique fiber configuration is more sensitive to change of phase function and more capable of extracting optical properties of epithelium. For example, the error for thickness of upper layer is about 9.83% and the RMS percentage error for scattering coefficient of upper layer is about 12.17%.
As for fluorescence spectra, we first refer some literature to set the optical properties of normal tissue and pathological tissue. And then we simulate spectra under different fiber conditions to see whether the fraction of detected fluorescence from upper layer would increase or not. According to the simulation results, we find out the fluorescence intensity measured from dysplasia tissue is lower than the one from normal tissue and most of the detected fluorescence from dysplasia tissue comes from upper layer. In addition, oblique fiber configuration can more effectively detect fluorescence from upper layer than perpendicular fiber configuration. However, increasing the tilted angle of fibers doesn’t necessarily mean the increase in fraction of detected fluorescence from upper layer. In order to do the curve-fitting for fluorescence spectra, we add 3% of noise to the theoretical spectra and use the noise-added spectra as the input. Then we inversely extract the fluorescence parameters, such as the quantum yield and the absorption coefficient at excitation wavelength of certain fluorophore. Fluorescence efficiency, the product of quantum yield and absorption coefficient, is what we pay most attention in this research. The reason is because a fluorophore’s fluorescence efficiency represents the fluorophore’s capacity for emitting fluorescence and it also serves as a standard for comparing different fluorophores. Though the extracted fluorescence parameters vary significantly from theoretical values, the analysis of fluorescence efficiency shows a much more acceptable result. The average error of normal tissue’s fluorescence efficiency is about 20.75% for upper layer and 15.5% for bottom layer. The average error of dysplasia tissue’s fluorescence efficiency is about 17.75% for upper layer and 12.5% for bottom layer. This result demonstrates our proposed fitting procedure for fluorescence spectra can really help us extract the real value of fluorescence efficiency, which helps a lot for measuring fluorescence spectra by experiment in the future.

並列關鍵字

diffuse reflectance spectrum ； fluorescence spectrum ； Monte Carlo algorithm ； phantom ； GPU

參考文獻

[13] 陳錫勳, “利用斜角光纖定量雙層組織的光學參數,” (2012)

[27] 施光偉, “移動式漫反射光譜系統建立與人體口腔黏膜參數萃取,” (2013)

[2] V. Backman, M.B. Wallace, L.T. Perelman, “Detection of preinvasive cancer cells,” Nature, vol. 406, pp. 35-36 (2000)

[3] D. De Veld, M. Witjes, H. Sterenborg, J. Roodenburg, "The status of in vivo autofluorescence spectroscopy and imaging for oral oncology," Oral Oncology, vol. 41, pp. 117-131 (2005)

[4] J. Peters, T. Watson, “Endoscopic Mucosal Resection of Barrett’s Esophagus and Early Esophageal Cancer,” Journal of Gastrointestinal Surgery, vol. 15, no. 8, pp. 1299-1302 (2011)