本文選題：無創(chuàng)光譜檢測　切入點：舌診　出處：《天津大學》2014年博士論文

【摘要】：光譜檢測技術以其無創(chuàng)、便捷、高效等優(yōu)點已成為生物醫(yī)學領域的先進研究手段。隨著光譜分辨率的不斷提高以及波段范圍的拓展，生物組織的測量光譜蘊含著更為豐富的反映組織細胞生理、病理變化的微觀結(jié)構(gòu)及成分信息，使醫(yī)學活體無創(chuàng)光譜檢測技術具有很高的可行性。然而個體差異及各組分診斷光譜的大量交疊使得組織的光譜信息與待測目標間關系存在著很強的模糊性及復雜的非線性，以致于對這些高維、強相關的光譜數(shù)據(jù)分析及處理面臨著嚴峻挑戰(zhàn)。亟需借助并構(gòu)建合適的數(shù)據(jù)挖掘智能算法用以提取客觀表征測量目標的光學特性，以期能揭示組織生理、病理變化與光譜信息之間隱含的客觀規(guī)律。 舌是觀察體內(nèi)功能變化及疾病信息的重要窗口，是人體醫(yī)學無創(chuàng)檢測的最佳測量點之一�；谏嗖刻卣餍畔⑦M行臨床診斷是無創(chuàng)醫(yī)學檢測的重要命題之一。本文以光譜技術用于舌診客觀化為研究背景，將血清多種蛋白含量定量檢測作為研究載體，針對光譜數(shù)據(jù)與血清蛋白含量間的復雜模糊的非線性映射，致力于對高維醫(yī)學光譜數(shù)據(jù)智能分析方法的若干關鍵技術進行研究，以為推動此類醫(yī)學無創(chuàng)光譜檢測進一步的探索及發(fā)展提供思路、方法及技術支持。 1．鑒于舌與生理病理信息之間存在著復雜且模糊的映射關系，針對當前舌診客觀化信息采集的局限性及處理模式存在將混合信息體割裂提取以致重要內(nèi)涵丟失的缺陷，提出了采集攜帶組織微觀結(jié)構(gòu)變化的舌象高光譜信息來改善信息獲取方式，將舌體交疊混合的圖譜信息作為一個整體進行分析，結(jié)合多種線性與非線性數(shù)據(jù)挖掘算法以黑箱模式關聯(lián)生化、生理或病理信息，提取特異性光譜指標群的新模式；所提取的光譜指標群有望作為病因病機分析的客觀依據(jù)。該模式為后續(xù)相關研究提供了可借鑒的思路。 2．探討了一種基于舌近紅外光譜的人體血清白蛋白、球蛋白和總蛋白三種生化指標的無創(chuàng)檢測方法。并以此為載體分析了不同建模算法的擬合非線性映射的能力。運用不同的數(shù)據(jù)挖掘方法建立了蛋白成分的定量預測模型，通過實驗分析證明了基于舌的近紅外光譜進行血清蛋白含量檢測具有較高的可行性，將有望為臨床蛋白成分檢測提供一種便捷、無創(chuàng)的先進手段。同時也驗證了支持向量機可有效抵抗活體檢測定量分析中存在的非線性因素，提高模型的魯棒性，進一步地能夠作為最佳波長選擇的評判依據(jù)。 3．針對體內(nèi)各組分的特征譜峰存在嚴重混疊現(xiàn)象，從而導致診斷光譜信號微弱且存在不確定性，提出了一種對非線性映射敏感的波段組合提取方案。該方案以支持向量機交叉驗證預測精度作為各波長或波長組合的非線性映射辨識能力的評判標準，，分別設計了非線性區(qū)間選擇法和自適應遺傳尋優(yōu)算法，其中前者針對高維光譜數(shù)據(jù)進行粗選以鎖定特征波長所在區(qū)間，后者則在這些區(qū)間內(nèi)通過全局尋優(yōu)搜索策略精選最佳波段組合。將該方案用于舌的近紅外光譜數(shù)據(jù)的波長挑選，在降低了三種血清蛋白含量檢測模型復雜度的同時，有效的提高了非線性模型的預測能力，進一步克服了因譜峰混疊及個體差異等引起的非線性因素。 4．提高模型泛化能力及普適性需要對大動態(tài)范圍的大規(guī)模樣本集進行深度挖掘。基于這個前提，針對智能分析算法中最耗時的支持向量機交叉驗證的計算效率問題，選擇性價比高的GPU并行平臺，提出且開發(fā)了支持向量機交叉驗證的細粒度并行算法。通過對不同中、大尺度基準集的測試，該算法充分調(diào)度了并行資源，有效并發(fā)交叉驗證計算任務，在保證計算精度的前提下，顯著的提升了計算效率，特別對高維光譜等此類稠密數(shù)據(jù)性能提升更為明顯。這對將智能分析算法推廣到中大尺度光譜數(shù)據(jù)中進行深度挖掘提供了技術支持。此外，針對該方案對小樣本數(shù)據(jù)集加速不顯著問題，進一步擴展并提出了一種基于GPU支持向量機網(wǎng)格搜索并行策略，在對58例近紅外光譜數(shù)據(jù)的參數(shù)選擇測試中獲得了36.02倍的加速比。
[Abstract]:Spectrum detection technology for its non-invasive, convenient, high efficiency has become the advanced research tools in biomedical field. With the development and constantly improve the spectral resolution and spectral range, spectral measurement of biological tissue contains more abundant reflect tissue cell physiology, microstructure and composition of information science of disease, make the medicine in vivo a spectrum detection technology has high feasibility. However, a large number of overlapping individual differences and each component of the diagnostic spectral spectrum information and makes the organization to be tested the relationship between fuzziness exists strong nonlinear and complex, so for these high-dimensional spectral data analysis and processing, strong correlation is facing serious challenges need the help. And construct the appropriate data mining algorithm for intelligent optical characteristics extraction objective characterization of measurement targets, in order to reveal the physiological, pathological changes and spectrum The implicit objective law between information.
The tongue is an important window to observe the in vivo function changes and disease information, is one of the best measurement points for noninvasive measurement of human medicine. The tongue characteristic information in clinical diagnosis is one of the important proposition of a medical detection based on. The spectroscopy for tongue diagnosis as the research background, the content of serum protein quantitative detection as many the carrier, in view of the complex fuzzy nonlinear mapping of spectral data and serum protein content, commitment to research some key technologies of the intelligent analysis method for high dimensional medical spectral data, that promote the medicine of noninvasive spectroscopic detection of further exploration and development of ideas, methods and technical support.
1. in the light of fuzzy and complicated mappings between the tongue and the physiological and pathological information, according to the limitations of the current treatment model and tongue objective information collection are mixed information that separates defects from the important connotation of lost, proposed acquisition organizations carry the micro structure change of tongue hyperspectral information to improve information retrieval method the tongue, the overlapping of mixed information are analyzed as a whole. The combination of linear and nonlinear data mining algorithm in the model of black box related biochemical, physiological or pathological information extraction, new model specific spectral index group; as the objective basis for the etiology and pathogenesis analysis of the extracted spectral index group is expected to provide. The reference model for future research.
The 2. discusses a tongue near infrared spectra of human serum albumin based on non-invasive detection methods of globulin and total protein of three kinds of biochemical indexes. And as the carrier of the ability of fitting nonlinear mapping different modeling algorithm. Using different data mining method to establish the quantitative prediction model of protein components, through experiments near infrared spectroscopy analysis proves that the tongue based on detection of serum protein content has the high feasibility, is expected to detect clinical protein components provide a convenient means, no advanced invasive. Also verified the support vector machine can effectively resist the nonlinear factors existing in vivo detection in the quantitative analysis, to further improve the robustness of the model. Can as the optimum wavelength selection criteria.
According to the characteristics of each component in 3. peaks exist serious aliasing, which leads to the diagnostic spectral signal is weak and there is uncertainty, we propose an extraction for nonlinear mapping sensitive band combination scheme based on support vector machine prediction of cross validation accuracy as identification of nonlinear mapping ability of each wavelength or wavelength combination evaluation the standard design of nonlinear interval selection method and adaptive genetic algorithm, in which the former for high dimensional spectral data for roughing to lock the characteristic wavelength where the interval, the latter through global optimization search strategy to select the optimal band combination in the range. The scheme for near infrared spectral data of tongue wavelength selection in the lower three kinds of serum protein content of model complexity and improve the prediction ability of the nonlinear model, further to overcome Nonlinear factors caused by the mixing of spectral peaks and individual differences.
4. improve the generalization ability and the universality of the large-scale sample of large dynamic range in the depth of excavation. Based on this premise, the computation efficiency of the intelligent analysis support vector machine algorithm is the most time-consuming in cross validation, choose cost-effective GPU parallel platform, fine-grained parallel algorithm proposed and developed a support vector machine cross validation. Based on different, large scale benchmark test set of the algorithm fully parallel resource scheduling, effective concurrent cross validation task in ensuring the accuracy, significantly improve the computational efficiency, especially to enhance the high dimensional spectral properties such dense data is more obvious. The the intelligent analysis algorithm is extended to large scale spectral data mining provides technical support. In addition, the scheme set is not significant to accelerate the small sample data, further A parallel search strategy based on GPU support vector machine for grid search is extended and proposed, and 36.02 times speedup is achieved in the parameter selection test of 58 NIR data.

【學位授予單位】：天津大學
【學位級別】：博士
【學位授予年份】：2014
【分類號】：O433;R445

【相似文獻】

相關會議論文 前1條

1 李星星;黃在銀;范高超;;光催化原位過程的熱力學、動力學及其機理的協(xié)同表征[A];中國化學會第29屆學術年會摘要集——第21分會：光化學[C];2014年

相關博士學位論文 前2條

1 李家星;醫(yī)學無創(chuàng)光譜檢測中若干關鍵技術的研究[D];天津大學;2014年

2 劉麗娜;結(jié)腸癌內(nèi)源性生物標記物的光譜檢測與分析[D];福建師范大學;2014年

相關碩士學位論文 前4條

1 嚴峗;多光譜污水有機污染物綜合指標在線檢測裝置的設計與開發(fā)[D];浙江大學;2011年

2 劉林美;煙氣中重金屬元素的激光擊穿光譜檢測研究[D];中南民族大學;2012年

3 張親富;α-Fe_2O_3光催化降解有機物的光譜分析與研究[D];華中科技大學;2012年

4 汪露;牛奶中違禁添加劑快速檢測方法的研究[D];浙江師范大學;2014年

本文編號：1701118