Single-layer multiple-kernel-based convolutional neural network for biological Raman spectral analysis

In this study, we propose a single-layer multiple-kernel-based convolutional neural network (SLMK-CNN) as an analysis tool for biological Raman spectra. We investigated the characteristics of SLMK-CNN and then analyzed and classified the biological Raman spectra by optimizing the structure of SLMK-CNN. We have found that the kernel size used in SLMMK-CNN plays an important role in changing the characteristics of Raman spectra such as intensity and peak position. As a result, the kernel size affects the classification performance and histological interpretation of biological Raman spectra. We also evaluated the classification performance of SLMK-CNN using Raman spectra obtained from the porcine skin samples irradiated by an ultraviolet (UV) source for different time. For three sample groups according to UV irradiation time (0, 10, and 24 hr), SLMK-CNN showed the classification accuracy of 96.4% and 92.5% for the preprocessed and raw Raman spectra, respectively. This is superior to other classification methods such as single-layer single-kernel-based CNN and principal component-linear discriminant analysis in this study.