Integrating spatial and chemical information enhances differentiation of non-alcoholic steatohepatitis states in Raman imaging.

Machine learning studies for Raman imaging have addressed the differentiability of normal and diseased states in biomedical applications by grouping a set of Raman spectra in terms of spectral similarity over the sample. However, Raman imaging provides both chemical information relevant to the underlying chemical species and their spatial distribution across the biological samples. Utilizing both the chemical and spatial information may further discriminate the sample states more than just using the spectral similarity free from the spatial structure. Here, we develop a Raman image analysis method integrating spatial and chemical information. The crux of our method is to introduce a measure to quantify spatial heterogeneity among Raman spectra at each pixel, and to classify Raman images using information theory, based not directly on the Raman spectra themselves at individual pixels but on the spatial heterogeneity in the spectral space over the surroundings. In this paper, we applied the method to a set of liver tissues dissected from non-alcoholic fatty liver disease (NAFLD) rat model, each of which is pathologically classified into normal, nonalcoholic fatty liver (NAFL) and non-alcoholic steatohepatitis (NASH), respectively. We show how the pathologically-identified liver states can be further classified using chemical information, and both chemical and spatial information. All NASH tissues that are belonging to a same cluster in spectral similarity are found to be further divided into substates that correlate the progression of the NAFLD disease, and subtle contamination of bloods that often prevents from appropriate pathological judgments. The potential of a use of both chemical and spatial information in Raman imaging is expected to enhance the differentiability of disease states of biological samples.