An optimized algorithm for separating scattering and chemical absorption in biomedical infrared spectroscopy and imaging
MetadataShow full item record
- Master's theses (IMT) 
Infrarødspektroskopi av biologiske prøver har blitt utviklet til et lovende verktøy for ikke-destruktiv biokjemisk analyse gjennom de siste tiårene. Infrarøde absorbansspektre representerer molekylære fingeravtrykk. Enkeltceller og vev forårsaker imidlertid komplekse Mie-spredningsegenskaper i infrarøde absorbansspektre som forurenser de rene kjemiske signaturene. Flere prosesseringsteknikker har blitt foreslått for å håndtere spredning i infrarødspektroskopi. Mie-korreksjon [24, 5, 28, 26] basert på extended multiplicative signal correction (EMSC) [32, 18, 21, 35, 34] betraktes for tiden som det kraftigste verktøyet for å separere Mie-spredning og biokjemisk absorpsjon i infrarøde spektra av celler og vev.Over the past decades, infrared spectroscopy of biological samples has been developed to a promising tool for non-destructive biochemical analysis. Infrared absorbance spectra provide molecular ﬁngerprints. However, single cells and tissues cause complex Mie scattering features in infrared absorbance spectra contaminating the pure chemical signatures. Several preprocessing methods have been proposed to handle scattering in infrared spectroscopy. The Mie correction [24, 5, 28, 26] based on extended multiplicative signal correction (EMSC) [32, 18, 21, 35, 34] is currently considered as the most powerful tool for separating Mie scattering and biochemical absorption in infrared spectra of cells and tissues. Kohler et al.  developed an algorithm based on EMSC that could successfully predict Mie scattering features and remove them from infrared absorbance spectra. Bassan et al. developedtheMieEMSCmodelfurthertohandlethesocalleddispersioneffect. The model was implemented in an iterative algorithm, and a compiled program for Mie correction was published . This program is currently the mostly used pre-processing tool for infrared spectra of cells and tissues in the diagnosis of cancer by infrared imaging. However, the algorithm is observed to be strongly biased, since corrected spectra adapt features of the reference spectrum. During recent years, Konevskikh et al. improved the Mie EMSC model further, however a user-friendly program based on the improved algorithm is not yet available [28, 26]. The main aim of this thesis is to further develop the Mie correction algorithm, such that a user-friendly program for Mie correction can be published. This is achieved by proposing a number of improvements to the Mie correction algorithm related to stabilization and optimization. In addition, there is a need for establishing a simulated data set with known pure absorbance spectra and scatter features that mimic measured apparent absorbance spectra, in order to validate different features of the algorithm. The improvements of the Mie EMSC correction algorithm include a number of aspects. The algorithm presented in this thesis sets the number of principal components in the Mie EMSC model automatically by the program, based on a desired level of explained variance in the Mie extinction curves. A ﬂexible stop criterion, based on the convergence of a forward Mie EMSC model is implemented. Further, the initialization parameters are standardized by controlling the scaling of the reference spectrum. Additional stability is gained by weighting the reference spectrum and by setting negative parts of the reference spectrum to zero. A simple quality test for evaluating the correction based on the error of the forward model is implemented, which is used to optimize the initialization parameters. In order to validate the algorithm, a set of absorbance spectra mimicking measured apparent absorbance spectra was simulated. In the simulations, the underlying pure absorbance is known, and scattering features were based on measured spectra. The simulated spectra were used for validation, and to assess critical features of the algorithm. We demonstrate that the correction is not biased by the initial reference spectrum and that a more reliable amide I peak position is retrieved. Sensitivity towards the initialization parameters is further reviewed. It is further demonstrated that the estimated scatter parameters from the EMSC model are meaningful and can be used for clustering of samples with respect to morphological characteristics. The advantage of pre-processing for a subsequent multivariate analysis by chemometrics and machine learning is discussed and suggestions are made how the algorithm can be employed on big spectral data from FTIR imaging. As a result of the proposed improvements, a user-friendly code for correcting highly Mie scatter-distorted absorbance spectra is published at https://bitbucket.org/biospecnorway/mie-emsc-code.