Machine Learning Methods for Classifying Raman Scattering Spectra of the Skin

In this paper, the possibility of using machine learning methods to classify Raman skin spectra is considered. Registration of Raman skin spectra was carried out using a portable spectroscopic setup. More than 500 patients participated in the study. For each patient, the spectrum was recorded from the healthy skin area and from the area with the disease. In total, 1225 spectra were used in the work: 609 spectra of healthy skin and 616 spectra with various skin diseases (benign and malignant neoplasms). In addition to the Raman scattering spectra, information about the actual diagnoses for each patient was obtained. These diagnoses were made by doctors of the Samara Regional Clinical Oncology Dispensary on the basis of histological analysis. Classification models were developed for three cases: healthy skin versus skin with diseases, benign neoplasms versus malignant neoplasms, and malignant melanoma versus pigmented nevus. To reduce the dimensionality of the data, the method of multivariate curve resolution alternating least squares (MCR-ALS) was used. Thirty skin components and their relative concentrations in the test sample were obtained. These data were subsequently used as classification parameters. The following machine learning methods were used: logistic regression, random forest, k-nearest neighbors algorithm and gradient boosting. The effectiveness of the developed classifiers was compared. The classification accuracy varies from 73% to 90%.