Precise diagnosis of lung cancer enabled by improved FTIR-based machine learning

Abstract

Micro-morphologic lesions of lung cancers (LC) are usually related to the developed tumors, lack of sensitivity to early lung cancer and specificity to tissue lesions of different origins due to similarity detected by medical imaging techniques. Attenuated total Reflection-Fourier transform infrared spectroscopy (ATR-FTIR) using accessible blood samples has identified many biomolecular markers and characteristic spectral changes, which is related to early cancer diagnosis. In this study, for precise diagnosis of LCs, the in-situ serum-based ATR-FTIR spectroscopy experiments were carried out with 156 serum samples divided into 5 different types of LC groups and controls. The peak-to-peak area ratios between different functional groups contributed to the discovery of potential FTIR markers of LCs. Moreover, we also constructed a two-dimensional second-order derivative spectral (2D-SD-IR) feature dataset based on infrared molecular fingerprints (IMF) of LCs, including absorbance and wave number shifts of FTIR vibrational peaks for LC early diagnosis with low time cost-effectiveness and high classification accuracy. By comparing, the diagnosis models of partial least square discriminant analysis (PLS-DA) and back propagation (BP) neural network based on 2D-SD-IR are suitable to differentiate LCs and pathological stages accurately, and the high sensitivity and specificity of BP reached up to 98.7%. The proposed 2D-SD-IR machine learning protocol for LCs’ classification demonstrates the great advantage of aid clinical blood biochemical testing.