Multicomponent Raman spectral regression using complete and incomplete models and convolutional neural networks
Derrick Ampadu Boateng, Chuanzhen Hu, Yichuan Dai, Kaiqin Chu, Jun Du, Zachary J. Smith
Abstract
an underdetermined model). An advantage of the model is that it combines background correction and regression into a single step, and does not require user-selected parameters. We compare our results with traditional least squares methods, including the popular asymmetric least squares (AsLS) approach. Our results demonstrate that the proposed CNN model boasts less sensitivity to parameter selection, and with a rapid processing speed, with performance equal to or better than comparison methods. The performance is validated on synthetic spectral mixtures, as well as experimentally measured single-vesicle liposome data.
Topics & Concepts
Hyperspectral imagingComputer scienceConvolutional neural networkPreprocessorPartial least squares regressionSensitivity (control systems)Least-squares function approximationBottleneckArtificial intelligencePattern recognition (psychology)RegressionArtificial neural networkAlgorithmMachine learningMathematicsStatisticsEngineeringEstimatorEmbedded systemElectronic engineeringSpectroscopy Techniques in Biomedical and Chemical ResearchSpectroscopy and Chemometric AnalysesPhotoacoustic and Ultrasonic Imaging