Essential spectra to improve vibrational imaging of pharmaceutical samples

Raman and infrared vibrational imaging are invaluable techniques for obtaining both spectral and spatial information from complex biological and medical samples. However, the direct analysis of hyperspectral imaging datasets is often hindered by the physical and chemical complexity of raw samples, resulting in vast amounts of data and numerous uncontrolled sources of variance. To address these challenges, the selection of essential spectra – comprising the most linearly distinctive rows of a data matrix – and their targeted analysis offer an efficient method to significantly enhance spectral unmixing. An estimation of the set of essential spectra can be performed by convex hull analysis of the normalized scores resulting from a truncated singular value decomposition. Alternatively, Fourier coefficients at selected harmonic frequencies may be used. Both approaches are valid, even though the data point cloud is represented in two different intrinsic coordinate systems, which translates into two different approximations. In this paper, we picture the use and efficiency of data reduction by essential spectra selection using both alternatives for the analysis of pharmaceutical samples by FTIR and Raman hyperspectral microimaging. These examples provide very large datasets and challenging analytical situations including the identification of very minor compounds and the presence of strong scattering components whose effect can mask the spectral information of
other compounds. Our results clearly demonstrate the advantages of analyzing reduced datasets obtained by identifying essential spectra instead of full data in terms of speed, enabling a much faster determination of the drug composition. In addition, they show that the use of Fourier coefficients allows reaching better data reduction rates (down to 0.1% of the original number of measured spectra for the FTIR image investigated).