【作者】 张岩；

【导师】 Peter Rolfe； 孙金玮；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2011， 博士

【摘要】 近红外光谱技术能够提供基于血红蛋白浓度变化的血液动力学信息,从而反映大脑皮质的血氧代谢状况,可用于脑功能活动的检测,被称为功能近红外光谱技术。与其它脑功能检测技术,如脑电图、脑磁图、正电子发射层析成像、以及功能磁共振成像等相比,近红外光谱技术具有使用方便、易实施、时间分辨率高、安全、便宜等优点,有非常广阔的应用前景。近红外脑功能检测会受到心动周期、呼吸、自发性低频振荡以及超低频振荡等人体生理活动的干扰,这种生理干扰不仅出现在头皮、颅骨和脑脊液等外层脑组织中,也出现在脑灰质和脑白质等深层脑组织中,严重影响脑功能信号的准确提取。因此,本文针对利用近红外光谱技术进行脑功能信号提取中存在的生理干扰问题,研究相应的解决方法,从而提高脑功能信号的检测精度,促进功能近红外光谱技术的开发与应用。本文的主要研究内容体现在如下几个方面:(1)在进行脑功能检测时,近红外光谱技术无法获得真实的脑功能信号,并且部分容积效应会造成测量信号远低于真实信号,从而难以定量分析脑功能信号提取方法的有效性。因此,本文基于五层脑部模型对脑功能活动进行模拟,利用MonteCarlo方法仿真近红外光在模型中的传输过程,通过补偿部分容积效应,开发了可用于对脑功能信号提取方法进行定量分析的仿真软件包。(2)单距测量方法具有探头结构简单和易于实现的优点,常用于血液动力学变化量的检测及基于阵列式光源检测器布局的脑功能成像研究。基于该方法的脑功能信号提取通常利用低通滤波技术抑制心动周期引起的生理干扰,但由于呼吸信号在频率上和脑功能信号有一定的重叠且是非平稳信号,使得呼吸干扰难以抑制。因此,本文提出基于经验模态分解的脑功能信号提取方法,该方法具有自适应时频特性,能够剔除心动周期和呼吸引起的生理干扰,从而提高脑功能信号的检测精度。利用自行开发的软件包,通过仿真验证了该方法在脑功能信号提取方面的有效性。(3)相比于心动周期和呼吸,低频振荡和超低频振荡等干扰信号与脑功能信号的频带严重重叠,采用常规的滤波方法不能消除这样的生理干扰。脑功能活动发生在深层大脑皮质,而生理干扰来自各种不同的脑组织。考虑探测深度与探测距离相关,提出基于多距测量方法的递归最小二乘脑功能信号提取方法。在该方法中,近端检测器用于获取参考信号,远端检测器用于获取期望信号,利用递推最小二乘自适应滤波技术进行处理从而完成脑功能信号的提取,并通过对比研究了递推最小二乘算法和最小均方算法的检测精度和收敛速率。此外,考虑基于多距测量方法的自适应滤波对近红外脑功能信号的提取精度与探头布局密切相关,进一步分析了多距测量方法中探头布局对测量结果的影响及探头布局对外层脑组织厚度的敏感性。仿真结果表明基于多距测量方法的递推最小二乘自适应滤波能够有效抑制生理干扰,对脑功能信号的检测精度和收敛速率均明显优于最小均方自适应滤波,并从统计意义上给出了不同探头布局和不同外层脑组织厚度时脑功能信号提取结果的均方误差。(4)近红外脑功能检测的生理干扰来源于人体不同的生理活动,当脑组织的非均匀性严重时,不同的生理活动在空间不同位置对生理干扰的影响也不尽相同。针对该问题,利用基于多距测量方法的自适应滤波能够实现对生理干扰的抑制,但并不能获得很好的检测精度。因此,针对多距测量方法,提出基于经验模态分解优化算法的脑功能信号提取方法。该方法首先对近端检测器测得的血液动力学变化进行经验模态分解,将分解的固有模态函数赋予不同的权系数以估计期望信号中的生理干扰,并通过递推最小二乘算法优化权系数。仿真结果表明优化算法能够依据生理干扰与固有模态函数的相关度自适应地调节权系数,在脑组织非均匀性严重时其脑功能信号检测精度优于最小二乘自适应滤波。(5)为了验证近红外光谱技术在脑功能信号检测中的有效性,本文基于连续波光谱技术和多距测量方法设计了近红外组织氧检测系统。通过离体模型实验和在体前臂阻断实验对近红外组织氧检测系统的工作性能进行了分析。针对音乐刺激诱发的颞叶区血液动力学变化,进行了基于听觉组块设计的脑功能实验研究。利用组织氧检测系统对实验过程中漫反射光强进行实时监测,并采用递推最小二乘自适应滤波对颞叶区的血液动力学变化进行了深入分析。听觉刺激的实验结果表明,颞叶区对音乐刺激敏感,经过对比原始的与提取的血液动力学变化,证明了基于多距测量方法的递推最小二乘自适应滤波在脑功能检测中可行性和有效性。更多还原

【Abstract】 Near infrared spectroscopy (NIRS) allows the non-invasive measurement of haemody-namic variables and is particularly suited to the detection of changes in concentrationsof oxy- and deoxy-haemoglobin in the brain, thereby providing insights into metabol-ic events in the cerebral cortex. Consequently, NIRS has been developed to measurebrain activity, leading to what has become the well-recognised method of functional near-infrared spectroscopy (fNIRS). fNIRS may be compared with other techniques, such aselectroencephalography (EEG), magnetoencephalography (MEG), positron emission to-mography (PET), and functional magnetic resonance imaging (fMRI). It does appear tohave several advantages over these other methods, such as portability, fewer physicalrestrictions and greater practicality, good temporal resolution, safety, and inexpensiveinstrumentation, and thus has a very broad application prospect. However, there are prob-lems in using fNIRS due to the presence of physiological interference.The physiological interference when using fNIRS arises mainly from perturbationscaused by cardiac events, breathing, low frequency oscillations (LFOs), and very lowfrequency oscillations (VLFOs). All of these interference sources are located both in thevasculature of the superficial layer of the brain and deeper inside the brain. This has meantthat without appropriate interference reduction the full potential of fNIRS has not yet beenrealised. Therefore, this thesis studies several practical methodologies to overcome thephysiological interference problem in fNIRS, aiming to improve the detection accuracyof brain activity measurement and promote the further development and utilization of themethod.The main contents of this dissertation are as follows:(1) A justification for the use of Monte Carlo simulation is given. A truly rigorousevaluation of fNIRS in vivo requires an uncontaminated evoked brain activity responsesignal as a standard, which, unfortunately, is unavailable. In addition, the partial volumeeffect (PVE) cannot be precisely compensated for in vivo and the quantitative comparisonof the recovery response and the true response of brain activity is therefore difficult. Thus,the Monte Carlo method, based on a five-layered adult head model, was implemented tosimulate the brain activity process for optical measurement. By compensating for the PVE, a simulation software package was developed to serve as the evaluation tool fordifferent methods for the extraction of brain activity measurements.(2) The use of Empirical Mode Decomposition (EMD) for brain signal extraction isdescribed. The single-distance NIRS probe configuration is often used to measure thehaemodynamic changes, both for monitoring and for imaging based on grids of sourcesand detectors, because it has the advantages of the simplicity of the optical probe andgreater practicality. Low pass filtering techniques have been used in attempts to suppressphysiological interference and these have been moderately successful for removal of theinterference caused by cardiac oscillations. However, low pass filtering may not be appro-priate for other specific physiological noise, such as that produced by breathing since suchnoise is difficult to be distinguished from the genuine haemodynamic response to brainactivity by frequency characteristics alone and thus it is not possible to design the low passfilter with a fixed cut-off frequency. Therefore, a methodology based on EMD is proposedto extract the signal of brain activity for single-distance measurement. The accuracy ofthe brain activity measurement is improved by utilizing EMD because it can be used toremove interference arising from the cardiac events and breathing. The effectiveness ofthis methodology has been proved by means of the software package.(3) The application of Recursive Least Squares (RLS) adaptive filtering is described.Compared with cardiac and respiratory interference, the suppression of LFOs and VL-FOs is relatively difficult with ordinary filtering techniques because these frequencies andthose of the functional activity may severely overlay each other. In fNIRS measurementvery useful information may be in the deep tissue (gray matter) and light inevitably inter-acts with blood in layers other than gray matter. Considering that the penetration depth ofNIRS is related to the source-detector separation, a methodology of combining a multi-distance probe and recursive least square (RLS) adaptive filtering is proposed. We madethe measurement acquired from NIRS short source-detector separation as the referencesignal and the measurement acquired from NIRS long source-detector separation as thedesired signal. The least mean square (LMS) and RLS algorithms are implemented tocompare the accuracy and the convergence rate. We derived measurements by adoptingdifferent interoptode distances, which is relevant to the process of optimizing the NIRSprobe configuration. The in?uence of superficial layer thickness on the performance of theRLS algorithm was also investigated. The simulation results demonstrated that the RLS algorithm has a faster convergence and smaller mean squared error (MSE) than the LMSalgorithm. The MSE for different probe configuration and superficial layer thickness arealso calculated based on statistical methods.(4) The combination of EMD and RLS was explored. Physiological interference can beinduced by different physiological phenomena and thus it contains multiple components.When the brain exhibits some haemodynamic heterogeneity, the different interferencecomponents may produce dissimilarities between the superficial layers and the cortex, orin different locations. In our study presented here we adopt the multidistance measure-ment method and a theoretical analysis of global interference reduction based on EMDand the least squares criterion. The short-distance fNIRS measurement is treated as com-prising of superficial haemodynamic changes induced by physiological ?uctuations andthe long-distance fNIRS measurement is the functional haemodynamic response contam-inated by global interference. By decomposing superficial haemodynamic ?uctuationswith the EMD algorithm, we separated the interference into different intrinsic mode func-tions (IMFs) possessing distinct frequency characteristics. The recursive least squaresmethod was then used to adjust the corresponding weighting coefficients to estimate glob-al interference with the obtained IMFs. The experimental results demonstrate that optimalalgorithms have higher precision that RLS adaptive filter when the brain tissue presentssome degree of heterogeneity.(5) In vivo measurements with multi-distance NIRS were investigated. To further studythe brain activity with fNIRS and evaluate the effectiveness of the proposed method, aNIRS system was developed based on a multi-distance measurement configuration andcontinuous wave spectroscopy. The performance of the system was verified by the in vit-ro model experiment and in vivo forearm occlusion experiments. Subsequently a block-design experiment was conducted on auditory stimuli and the evoked response of thecortex in the temporal region was continuously monitored and further analyzed by theRLS algorithm. The experimental results show that the temporal area is sensitive to mu-sic stimuli. The comparison of the original results and the RLS results demonstrate thefeasibility and effectiveness of RLS adaptive filtering for fNIRS.更多还原