Label-free Diagnosis of Lung Cancer by Fourier Transform Infrared Microspectroscopy Coupled with Domain Adversarial Learning

Abstract

Lung cancer is one of the most prevalent malignancies, characterized by high morbidity and mortality rates. Current diagnostic approaches primarily rely on CT imaging and histopathological evaluations, which are time-consuming, heavily dependent on pathologists' expertise, and prone to misdiagnosis. Fourier transform infrared (FTIR) microspectroscopy is a promising label-free technique that can offer insights into morphological and molecular pathologic alterations from biological tissues. Here, we present a novel FTIR microspectroscopic method enhanced by a deep learning model for differentiating lung cancer tissues, which serves as a crucial adjunct to clinical diagnosis. We propose an Infrared Spectral Domain Adversarial Neural Network (IRS-DANN), which employs a domain adversarial learning mechanism to mitigate the impact of inter-patient variability, thereby enabling the accurate discrimination of lung cancer tissues. This method demonstrates superior classification performance on a real clinical FTIR dataset, even with limited training samples. Additionally, we visualize and elucidate the FTIR fingerprint peaks, which are linked to the corresponding biological components and crucial for lung cancer differentiation. These findings highlight the great potential of incorporating FTIR microspectroscopy with the deep learning model as a valuable tool for the diagnosis and pathological studies of lung cancer.