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

The power of Raman spectroscopy is largely enhanced by machine learning and chemometrics, which extract and translate the spectral features into high-level biological or clinical knowledge by constructing classical or deep learning models. The generalizability of such models, however, is often degraded due to the large variations between the training data and the data to be predicted. Model transfer showed great potential in this regard, which improved the prediction on the test data without re-building a new model from scratch. We developed a method based on Siamese neural network (SNet) and compared it with two basis models as well as two model transfer methods score movement (MS) and extensive multiplicative scattering correction (EMSC). The performance was systematically verified with a Raman spectral dataset measured from four bacterial species, each consisting of nine biological replicates. Its generalizability was further tested on a second Raman dataset from mice tissue samples. Siamese network was demonstrated to outperform the MS and EMSC, especially given large training datasets. The load on training data, however, is substantially lower than conventional networks and can be slightly reduced when variability between training and test data is properly incorporated into the loss function. Unlike MS and EMSC, more importantly, Siamese network does not require information of test data for model adjustment or data space adaptation, which makes it more advantageous in practice.