Siamese networks in Raman spectroscopy: Towards a better performance against replicate variability.

Area of Science:

• Analytical Chemistry
• Biotechnology
• Machine Learning

Background:

• Raman spectroscopy combined with machine learning offers powerful biological and clinical insights.
• Model generalizability is a significant challenge due to variations between training and prediction datasets.
• Model transfer techniques can improve prediction accuracy without retraining models from scratch.

Purpose of the Study:

• To develop and evaluate a Siamese neural network (SNet) for enhanced generalizability in Raman spectral data analysis.
• To compare the performance of SNet against established methods like score movement (MS) and extensive multiplicative scattering correction (EMSC).

Main Methods:

• Development of a Siamese neural network (SNet) for spectral feature extraction and knowledge translation.
• Systematic performance verification using Raman spectral datasets from four bacterial species.
• Further testing of generalizability on Raman spectral data from mouse tissue samples.

Main Results:

• The Siamese neural network (SNet) demonstrated superior performance compared to MS and EMSC, particularly with large training datasets.
• SNet requires significantly less training data load than conventional networks.
• SNet does not necessitate test data information for model adjustment or adaptation, offering practical advantages.

Conclusions:

• Siamese neural networks present a robust and advantageous approach for improving the generalizability of machine learning models in Raman spectroscopy.
• The SNet method offers practical benefits by not requiring test data for model adaptation, making it highly suitable for real-world applications.