Raman spectroscopy has been promoted as a non-invasive diagnostic technique capable of distinguishing molecular changes in biological samples. The quantification of spectra is limited by physical, high frequency cosmic spikes, and broader fluorescence background artifacts.

Current preprocessing techniques eliminate artifacts but also distort Raman spectral bands in the process. The preprocessing techniques depend on many parameters that are difficult to optimize and rely on signal characteristics such as background shape and spectral and cosmic spike widths. This thesis presents an algorithm to remove cosmic spikes based on the principles of empirical mode decomposition and energy detection, and two algorithms for background fluorescence noise rejection by applying the principles of asymmetric penalized least squares minimization using non-quadratic cost functions. Performance analysis using simulated datasets show that the proposed algorithms outperform the state-of-the-art in both artifact suppression and Raman band preservation. This thesis also shows the importance of preprocessing Raman Spectra on classification.