Bacterial identification in SERS-integrated microfluidics using CNN-driven 2D classification of 1D spectra

Bacterial sensing involves complex and variable samples that require advanced handling and analytical methods. To address these challenges, machine learning-especially deep learning-and SERS-based microfluidics have shown great promise. While previous studies have majorly focused on 1D spectral classification, the use of 2D representations of SERS spectra has not yet been explored, particularly for on-chip bacterial identification. In this work, we introduce a novel framework that combines SERS-enabled microfluidics with optimized 2D convolutional neural networks (2D-CNNs) for bacterial classification. SERS integration inside microfluidic chips was achieved through direct laser writing, enabling custom active areas and efficient on-chip measurements. We systematically evaluated nine distinct 1D-to-2D spectral transformations, with spectrogram and continuous wavelet transform yielding test accuracies of 99 % and 97 %, respectively, on controlled datasets. Using transfer learning, we achieved 100 % accuracy on the on-chip dataset, demonstrating the model's adaptability to new data. In contrast, other transformations, like pairwise distance and autocorrelation, performed below 93 %, indicating their limited ability to capture subtle spectral features. This framework offers high sample control, parallelization, and the potential for expanding the bacteria database, making it ideal for low-data-volume situations such as rare infections. Further development and testing across strains, environments, and practical challenges can further improve our approach's reliability for real-world diagnostics.