Optimal Feature Extraction for Discriminating Raman Spectra of Different Skin Samples using Statistical Methods and Genetic Algorithm

Introduction: Raman spectroscopy, that is a spectroscopic technique based on inelastic scattering of monochromatic light, can provide valuable information about molecular vibrations, so using this technique we can study molecular changes in a sample. Material and Methods: In this research, 153 Raman spectra obtained from normal and dried skin samples. Baseline and electrical noise were eliminated in the preprocessing stage with subsequent normalization of Raman spectra. Then, using statistical analysis and Genetic algorithm, optimal features for discrimination between these two classes have been searched.  In statistical analysis for choosing optimal features, T test, Bhattacharyya distance and entropy between two classes have been calculated. Seeing that T test can better discriminate these two classes so this method used for selecting the best features. Another time Genetic algorithm used for selecting optimal features, finally using these selected features and classifiers such as LDA, KNN, SVM and neural network, these two classes have been discriminated. Results: In comparison of classifiers results, under various strategies for selecting features and classifier, the best results obtained in combination of genetic algorithm in feature selection and SVM in classification. Finally using combination of genetic algorithm and SVM, we could discriminate normal and dried skin samples with accuracy of 90%, sensitivity of 89% and specificity of 91%. Discussion and Conclusion: According to obtained results, we can conclude that genetic algorithm demonstrates better performance than statistical analysis in selection of discriminating features of Raman spectra. In addition, results of this research illustrate the potential of Raman spectroscopy in study of different material effects on skin and skin diseases related to skin dehydration.