Deep Supervised and Self-Supervised Learning for Raman Spectroscopy signal analysis

Recent advances in Deep Learning have significantly increased the scope of application of powerful Neural Network models to produce highly informative representations of data of various types. Raman spectroscopy, specifically, has been recently shown to be a field in which such methods can bring substantial results. In this work, we discuss the building blocks of Machine Learning and Deep Learning. We train, validate, and test several Supervised architectures over Raman Spectra datasets of organic nature, comparing the efficacy of CNNs, Transformers, and Hybrid architectures. We propose, implement and test three, distinct novel uses of Deep Learning techniques in the analysis of Raman spectra: we employ Continuous Wavelet Transforms to input spectra to a 2D-CNN model to perform classification tasks, achieving competitive results compared to other techniques. We propose a Physics-based data generation procedure to estimate relative abundance in sugar mixtures, and we finally train a contrastive Siamese Network over thousands of different organic spectra, achieving promising zero-shot results in classification and performing zero-shot, highly detailed Raman Image segmentation over Human fibroblasts.