Study On Signal Denoising For Near-infrared Diffuse Optical Blood Flow Measurement With Nth-order Linear Algorithm

Diffuse correlation spectroscopy（DCS）is an emerging technology to extract the blood flow index（BFI）from light electric field temporal autocorrelation data.Due to the influence from environmental factors,all the measured optical data are contaminated by the noise.In order to obtain accurate and stable blood flow value,we developed the Nth-order linear（NL）algorithm to extract the blood flow values in both homogeneous and heterogeneous tissues with arbitrary geometry.Linear regression is a critical step to implement the NL algorithm,which also realizes the signal denoising for DCS data.In this thesis,four approaches are proposed to implement the DCS signal denosing in order to seek the best method for improving the accuracy of DCS blood flow extraction.First,the least-squared minimization（L₂ norm）is a widely-used method to perform data fitting and linear regression.However,the L₂ norm is severely affected by the data points with large derivations to the regression line,particularly in low signal intensity and low signal-to-noise ratio（SNR）.Hence,in order to overcome the limits of the L₂ norm method,we propose the least-absolute（L₁ norm）method which aims to reduce the weight of noise data.Compared with L₂ norm,the weight of noise data is substantially reduced through decreasing the norm order from 2 to 1.Secondly,in this thesis,for the first time,we explore the support vector regression（SVR）to denoise the optical signals for tissue blood flow measurements.By introducing theε-insensitive loss function,linear kernel function and Lagrange multiplier,the SVR regression model is established.The regression model introduces the idea of support vector and suppresses the noise data from the decision plane by adjusting the penalty parameters.Particularly,the severe noise data can be completely excluded from linear regression calculations.Thirdly,for the first time,the recursive neural network（RNN）regression model is utilized to extract the blood flow index by DCS signal denoising in this thesis.Through the forgetting gate structure of LSTM unit in the RNN network,the abnormal blood flow information（i.e.,noise points）are removed.In the process of constructing the regression model,the DCS temporal autocorrelation data with noise are taken as the input values,and the scatter points on the autocorrelation data without noise are used as the labels.The serialized output value is obtained through model training and,eventually,the regression model is established.In order to understand what are the appropriate regression methods for BFI estimation.We carried out the studies of computer simulations,liquid phantom and human in vivo experiments respectively.In the computer simulations,noise are added to tissue models of different curvatures and volumes（simulating human and small animal heads and limbs）,and combined with photon Monte Carlo（MC）simulation to generate DCS data;Accordingly,we designed and built the liquid phantom experiment device to obtain the DCS measurement data.In addition,we carried out the experiments of leg muscle blood flow and cerebral blood flow measurements,in order to obtain the DCS data from 10 healthy volunteers.The results show that both RNN and SVR approaches performed excellent in extracting blood flow values（i.e.,the highest accuracy and the smallest error value）and the best stability.These outcomes demonstrate that the new denoising methods proposed in this study have greatly improved the accuracy of blood flow measurements from DCS data,which is critical for translating the DCS technology to the clinic.