Machine learning assisted spectral analysis for non-invasive skin lesion classification
Ekaterina A. Beshanova1, Irina A. Matveeva1; 1Samara National Research University, Samara, Russia
Abstract
This research develops an end-to-end workflow that combines preprocessing, multivariate curve resolution (MCR), and machine learning to classify skin lesions from Raman spectra. Raw spectra are denoised (Savitzky–Golay filtering, wavelet and airPLS baseline correction) to improve signal-to-noise before feature extraction. MCR decomposition and PCA/PLS dimensionality reduction are used to separate overlapping biochemical components and to generate interpretable spectral factors that represent chromophores and molecular vibrations relevant to skin pathology. Classical classifiers (k-NN, SVM, PLS-DA) are compared with neural architectures (MLP, 1D-CNN, LSTM) using balanced training, augmentation, and evaluation by Accuracy, ROC AUC, Precision, Recall and F1-score. The work includes analysis of discriminative Raman bands, quantitative evaluation of preprocessing choices, and recommendations for deploying the pipeline in diagnostic studies. Results demonstrate that integrating chemometric decomposition with modern classifiers yields robust, interpretable models suitable for supporting non-invasive skin lesion diagnostics from Raman measurements.
Speaker
Ekaterina A. Beshanova
Samara National Research University
Russia
Discussion
Ask question
