Dimensionality Reduction for Identification of Hepatic Tumor Samples Based on Terahertz Time-Domain Spectroscopy

Terahertz time-domain spectroscopy (THz-TDS) combining with chemometrics methods was proposed for the identification of hepatic tumors. Two linear compression methods, principle component analysis and locality preserving projections (LPPs), and a nonlinear method, Isomap, were used to reduce the dimensionality of the measured dataset. Comparing two-dimensional (2-D) data reduced by these three dimensionality reduction techniques, only 2-D Isomap plot could separate the distances between two classes for the THz timedomain data and LPP had capacity of distinguishing two types of samples building on frequency-domain data. The best classification accuracies from 2-D time-domain data were 99.81 ± 0.30% and 99.69 ± 0.61% given by Isomap probabilistic neural network (PNN) and Isomap support vector machine (SVM), respectively, while the best classification results of 2-D frequency-domain data were 100.00 ± 0.00%, 99.75 ± 0.32% provided by LPP-PNN, LPP-SVM. The results showed that Isomap and LPP are appropriate techniques to reflect the nonlinear manifold of the THz data. The THz technology either in time-domain or frequency-domain coupled with Isomap-PNN or LPP-PNN could offer a potential procedure to identify hepatic tumors.