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

Area of Science:

• * Advanced analytical techniques for microbiology and biosensing.
• * Development of novel microfluidic devices for biological sample analysis.

Background:

• * Bacterial sensing faces challenges due to sample complexity and variability, requiring sophisticated analytical methods.
• * Surface-enhanced Raman spectroscopy (SERS) and microfluidics show potential for bacterial identification.
• * Previous research primarily focused on 1D SERS spectral classification, leaving 2D representations unexplored for on-chip applications.

Purpose of the Study:

• * To develop and evaluate a novel framework combining SERS-enabled microfluidics with 2D convolutional neural networks (2D-CNNs) for bacterial classification.
• * To explore the efficacy of various 1D-to-2D spectral transformations for enhanced bacterial identification.
• * To demonstrate the framework's adaptability and potential for low-data-volume diagnostics.

Main Methods:

• * Integration of SERS within microfluidic chips using direct laser writing for custom active areas and on-chip measurements.
• * Systematic evaluation of nine distinct 1D-to-2D spectral transformations for SERS data.
• * Application of optimized 2D-CNN models for bacterial classification using transformed spectral data.
• * Utilization of transfer learning to assess model adaptability to new datasets.

Main Results:

• * Spectrogram and continuous wavelet transform transformations achieved high test accuracies (99% and 97%) on controlled datasets.
• * The framework demonstrated 100% accuracy on an on-chip dataset using transfer learning, indicating excellent adaptability.
• * Other transformations, such as pairwise distance and autocorrelation, showed lower performance (<93%), highlighting the superiority of spectrogram and wavelet methods.
• * The developed framework offers enhanced sample control and parallelization capabilities.

Conclusions:

• * The combination of SERS microfluidics and 2D-CNNs provides a powerful and accurate approach for bacterial classification.
• * Spectrogram and continuous wavelet transform are effective methods for converting 1D SERS data into 2D representations for deep learning.
• * The framework's high accuracy and adaptability make it suitable for low-data-volume scenarios, such as diagnosing rare infections.
• * Further research can expand the bacterial database and refine the approach for real-world diagnostic applications.