Neural Network Guided Global Least Squares for Quantitative Biochemical Component Analysis in Raman Spectroscopy.

Raman spectroscopy, as a label-free analytical tool, has been widely applied in many fields including pharmaceutical research, biological sample analysis, and food quality control due to its high chemical specificity and noninvasive nature. Least squares regression and neural networks are two frequently used methods for spectral analysis in Raman spectroscopy. To address the convergence issue of least-squares regression in the case of many parameters and the limitation of neural networks under the condition of small training data sets in the context of quantitative biochemical component analysis in Raman spectroscopy, this study proposes a hybrid algorithm in which the initial result of neural networks is used to guide the convergence of global modified least-squares (NN-GMLS). The method employs simulated data synthesized from modified experimental spectra for neural network training. It is validated on both a surface-enhanced Raman spectroscopy (SERS) data set measured from chemical phantoms and a spontaneous Raman spectroscopy data set measured from K562 leukemia cells. Results of comparison with standalone neural networks, modified least-squares (MLS), and NN-GMLS based transfer learning on the K562 leukemia cell data set demonstrate the superior accuracy of NN-GMLS in the quantitative biochemical component analysis of Raman spectra when reference spectra contain potential variations or inaccuracy. In addition, simulated data synthesized in NN-GMLS are used to train 1D ResNet-10 for the classification of spontaneous Raman spectra from live, apoptotic, and necrotic leukemia cells, which shows an overall accuracy of 87.5%. The NN-GMLS method is expected to play a significant role in scenarios with limited training data sets, such as microbial and cellular sample analysis.