Research And Application Of Raman Spectroscopy For Blood Products Based On Deep Learning

The identification of multiple species blood is particularly important for entry-exit inspection and quarantine,forensic investigation and wildlife protection.The traditional methods often destroy blood samples and waste time.It is difficult to meet the requirements of rapid and non-destructive testing in identification.Raman Spectroscopy is a vibrational spectrum,which can obtain the information of molecular vibration and rotation so as to analyze the chemical composition of the material.Its characteristics of zero pollution,non-contact and rapid testing provide the possibility for rapid nondestructive identification of blood products.Currently,there are several methods of blood identification based on Raman spectroscopy,however,these methods often accompanied by the tedious preprocessing,which is easy to cause the model identification ability to decline due to improper pretreatment.At the same time,most of these methods use the linear discriminant model,ignoring nonlinear relationship between the spectrum and sample,and lead to the bad performance of the model,but the nonlinear discriminant model is less interpretable.Therefore,in order to overcome the shortcomings of the existing analytical methods,a new Raman spectrum analysis method of blood products is studied in this thesis,which is based on the excellent feature learning ability of deep learning.The main research contents include the following aspects:Firstly,this thesis studied the basic theories of Raman spectroscopy analysis methods,the basic principles of Raman spectroscopy analysis,common pretreatment methods and calibration models included.Moreover,discussed the shortcomings of Raman spectroscopy analysis methods,taken the strong characteristic learning ability of deep learning as the breakthrough point to overcome these shortcomings.Secondly,convolutional neural network(CNN)model was improved by optimized the network structure and objective function of deep learning.Raman-CNN was proposed,which integrated the denoising and baseline correction into the neural network.The adaptive data preprocessing of Raman spectra was implemented to overcome the shortcomings of denoising and baseline correction methods.In addition,the strong nonlinear mapping ability of the neural network improved the predictive performance of the discriminant model.The results shown that this method not only had excellent predictive performance,but also had better preprocessing effect than the traditional method.Thirdly,the idea of Squeeze-and-Excitation network(SENet)was explored to select the wavelength of spectrum in this thesis.Raman-SENet was proposed by simplified the network structure of SE module and increased the sparse constraint.It is a new method of sparse wavelength selection.It could not only self-learn the sparse weight to select the wavelength,but also can be used as a nonlinear calibration model with strongly interpretable.The results shown that Raman-SENet,as a wavelength selection method,could build a better model with fewer wavelength points than the traditional methods,and as a nonlinear correction model,it effectively overcomes the disadvantage of the traditional nonlinear method.What's more,the combination with Raman-CNN formed a powerful and complete deep learning analysis method of Raman spectroscopy,named Raman-DANet.It not only has strong prediction performance,but also has clear function and strong interpretability among the different layers of the network.It realizes the adaptive preprocessing process of de-noising,baseline correction and wavelength selection trinity.Finally,in view of the urgent need of import and export inspection and quarantine for the identification of blood species,a set of blood Raman spectrum identification software was designed and implemented for the micro Raman blood analyzer developed by the Microsystem Research Center of Chongqing University.34 blood samples from the Institute of Inspection and Quarantine were tested online.The test results are reliable and meet the requirements of entry-exit inspection and quarantine for the identification of blood species.