Differential Diagnosis of Lung Cancer Using Artificial Intelligence-Based Extracellular Vesicle Analysis: A Pilot Study

Abstract

In cancer diagnosis, liquid biopsy is a minimally invasive method, yet its diagnostic performance is constrained by insufficient sensitivity and restricted blood biomarkers. Therefore, the use of additional invasive methods is unavoidable to identify cancer-specific information, such as subtypes, stages, and mutation status. Here, we propose a deep-learning-based method for differential diagnosis, in which group-specific feature importance is assigned to Raman spectra of extracellular vesicles (EVs) and the spectra are decoded. The feature importance was extracted from the training dataset using explainable AI (XAI) to emphasize key spectral patterns. These features were then applied to the original spectra as weights to develop an AI-based diagnostic algorithm. As a result, the algorithm successfully identified the presence of lung cancer, achieving an area under the curve (AUC) value of 0.98. Moreover, it demonstrated the capability of precision cancer diagnosis with an average AUC of 0.96 across the detailed cancer information. These preliminary findings suggest the potential of liquid-biopsy-based differential diagnosis and its possible contribution to the establishment of rapid and tailored treatment strategies.