Tailored Deep Learning–Assisted In Situ SERS: Overcoming Surface Irregularities-Induced Large Signal Variation on Biological Tissues

Accurate in situ quantification on biological tissues with uneven surfaces using surface-enhanced Raman spectroscopy (SERS) remains a persistent challenge due to severe signal variability arising from surface irregularities and the coffee-ring effect. Herein, we present a tailored deep learning-assisted in situ SERS strategy that integrates minimal sample preparation, a low-cost SERS substrate, and a tailored one-dimensional convolutional neural network (1D-CNN) for highly reproducible SERS quantification on uneven biological surfaces. Detection of thiabendazole on apple skin served as a representative model. We highlight the critical role of sample preparation and SERS substrate selection in minimizing spectral variation. To address the remaining substantial variability in intensity after preprocessing, a tailored 1D-CNN with decreasing kernel sizes (57–37–11–3) was compared with single-peak intensity calibration (SPIC), partial least squares regression, random forest, and fixed-kernel 1D-CNNs (3 and 5). The tailored 1D-CNN consistently outperformed all other models, improving the R 2 for AuNPs-enhanced thiabendazole quantification from 0.332 (SPIC) to 0.935 while maintaining a short training time of 242 s. This work establishes a deep learning–enabled framework to mitigate surface-induced signal variability in in situ SERS and improve quantitative robustness on uneven biological surfaces, providing a promising strategy for rapid surface analysis of heterogeneous biological samples, including but not limited to residue screening.