Label-free rapid detection of urinary tumor cells based on surface-enhanced Raman scattering and deep learning

Abstract

Abstract In clinical practice, the primary objective of tumor detection is to acquire comprehensive diagnostic information while minimizing invasiveness and reducing patient discomfort. Urine cytology represents a non-invasive method frequently employed for urologic tumor detection. However, its sensitivity is limited. Enhancing the accurate identification of various urologic tumor cells and blood cells is crucial to improve the sensitivity of urine cytology. Surface-enhanced Raman spectroscopy (SERS), coupled with suitable machine learning algorithms, holds significant potential for rapid, sensitive, label-free, and non-destructive detection and identification of tumor cells. In this investigation, SERS spectra of urologic tumor cells and blood cells were acquired using an ordered substrate comprising Au-wrapped nanorod arrays. Notably, a remarkably high spectral resemblance was observed among the three distinct types of urologic tumor cells. Five machine learning algorithms were implemented for cell type differentiation and prediction. Among these, the classification network system integrating spatial attention mechanism with DenseNet exhibited the highest classification performance, yielding an accuracy rate of nearly 99%. Additionally, an attention heatmap was generated to highlight the wavenumber range that contributed the most in the SERS spectra, aiding in discriminating various cell species. This study demonstrates that SERS technology based on Au-wrapped nanorod arrays, in conjunction with deep learning algorithms, can promptly and accurately differentiate normal cells from tumor cells, thereby offering an effective approach to enhance the sensitivity of urine cytology tests.