医学无创光谱检测中若干关键技术的研究

发布时间：2018-04-02 16:07

本文选题：无创光谱检测　切入点：舌诊　出处：《天津大学》2014年博士论文

【摘要】：光谱检测技术以其无创、便捷、高效等优点已成为生物医学领域的先进研究手段。随着光谱分辨率的不断提高以及波段范围的拓展，生物组织的测量光谱蕴含着更为丰富的反映组织细胞生理、病理变化的微观结构及成分信息，使医学活体无创光谱检测技术具有很高的可行性。然而个体差异及各组分诊断光谱的大量交叠使得组织的光谱信息与待测目标间关系存在着很强的模糊性及复杂的非线性，以致于对这些高维、强相关的光谱数据分析及处理面临着严峻挑战。亟需借助并构建合适的数据挖掘智能算法用以提取客观表征测量目标的光学特性，以期能揭示组织生理、病理变化与光谱信息之间隐含的客观规律。舌是观察体内功能变化及疾病信息的重要窗口，是人体医学无创检测的最佳测量点之一。基于舌部特征信息进行临床诊断是无创医学检测的重要命题之一。本文以光谱技术用于舌诊客观化为研究背景，将血清多种蛋白含量定量检测作为研究载体，针对光谱数据与血清蛋白含量间的复杂模糊的非线性映射，致力于对高维医学光谱数据智能分析方法的若干关键技术进行研究，以为推动此类医学无创光谱检测进一步的探索及发展提供思路、方法及技术支持。 1．鉴于舌与生理病理信息之间存在着复杂且模糊的映射关系，针对当前舌诊客观化信息采集的局限性及处理模式存在将混合信息体割裂提取以致重要内涵丢失的缺陷，提出了采集携带组织微观结构变化的舌象高光谱信息来改善信息获取方式，将舌体交叠混合的图谱信息作为一个整体进行分析，结合多种线性与非线性数据挖掘算法以黑箱模式关联生化、生理或病理信息，提取特异性光谱指标群的新模式；所提取的光谱指标群有望作为病因病机分析的客观依据。该模式为后续相关研究提供了可借鉴的思路。 2．探讨了一种基于舌近红外光谱的人体血清白蛋白、球蛋白和总蛋白三种生化指标的无创检测方法。并以此为载体分析了不同建模算法的拟合非线性映射的能力。运用不同的数据挖掘方法建立了蛋白成分的定量预测模型，通过实验分析证明了基于舌的近红外光谱进行血清蛋白含量检测具有较高的可行性，将有望为临床蛋白成分检测提供一种便捷、无创的先进手段。同时也验证了支持向量机可有效抵抗活体检测定量分析中存在的非线性因素，提高模型的鲁棒性，进一步地能够作为最佳波长选择的评判依据。 3．针对体内各组分的特征谱峰存在严重混叠现象，从而导致诊断光谱信号微弱且存在不确定性，提出了一种对非线性映射敏感的波段组合提取方案。该方案以支持向量机交叉验证预测精度作为各波长或波长组合的非线性映射辨识能力的评判标准，，分别设计了非线性区间选择法和自适应遗传寻优算法，其中前者针对高维光谱数据进行粗选以锁定特征波长所在区间，后者则在这些区间内通过全局寻优搜索策略精选最佳波段组合。将该方案用于舌的近红外光谱数据的波长挑选，在降低了三种血清蛋白含量检测模型复杂度的同时，有效的提高了非线性模型的预测能力，进一步克服了因谱峰混叠及个体差异等引起的非线性因素。 4．提高模型泛化能力及普适性需要对大动态范围的大规模样本集进行深度挖掘。基于这个前提，针对智能分析算法中最耗时的支持向量机交叉验证的计算效率问题，选择性价比高的GPU并行平台，提出且开发了支持向量机交叉验证的细粒度并行算法。通过对不同中、大尺度基准集的测试，该算法充分调度了并行资源，有效并发交叉验证计算任务，在保证计算精度的前提下，显著的提升了计算效率，特别对高维光谱等此类稠密数据性能提升更为明显。这对将智能分析算法推广到中大尺度光谱数据中进行深度挖掘提供了技术支持。此外，针对该方案对小样本数据集加速不显著问题，进一步扩展并提出了一种基于GPU支持向量机网格搜索并行策略，在对58例近红外光谱数据的参数选择测试中获得了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.

【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：O433;R445

【相似文献】