Variational approach to cancerous tissue identification from in vivo Raman spectra. (March 2019)

• Record Type: Journal Article
• Authors: Krbcova, Zuzana; Kukal, Jaromir; Mares, Jan; Habartova, Lucie; Setnicka, Vladimir
• Is Part Of: Biomedical signal processing and control. Volume 49(2019)
• Page Start: 520
• Page End: 527
• Publication Date: 2019-03
• Subjects: Variational classifier -- Cancerous tissue identification -- Raman spectra -- Hierarchical cross-validation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2018.12.026
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: New variational classifier of one-dimensional signal for two classes. Variational task is converted to application of Fast Fourier Transform. Fast and reliable classifier learning from Raman spectra of lung tissues. Directly applicable to any spectra of classified biomedical samples. Abstract: In biomedical applications, the human factor represents the major obstacle in the processing and classification of in vivo Raman spectra. Although several methods of statistical classification have been employed to address this issue, the outcomes thereof vary depending on the selection of spectral bands used for the evaluation. To avoid the use of such discrete values, we developed a variational classifier based on spectrum weighting, weight smoothing and regularization and the selection of a support vector set. Embedded regularization of the variational classifier smooths weight function and therefore original Raman spectra preprocessing don't need to include spectra smoothing. Due to a small data set, we proposed a hierarchical cross-validation technique performing better data generalization and overfitting avoiding. Accuracy of variational classifier obtained by hierarchical cross-validation for the data set, involving 30 Raman spectra of which 13 represent cancerous and 17 healthy tissues, was 93% with sensitivity of 92% and specificity of 94%.