Unsupervised Random Forest Based Analysis of Biomedical Raman Images

Raman imaging, a live cell imaging technique, is expected to provide a new means to diagnose the diseases, where the images contain both the spatial and spectral information of molecular vibrations of components of biosystems. Biomedical Raman images are the big data having rich molecular information. Useful data analysis is needed to analyze the complex Raman spectra to make the Raman imaging an efficient diagnostic tool. This study involves the application of unsupervised random decision forest (URF) and rate-distortion theory (RDT) based clustering, to analyze the Raman hyperspectral data in order to gain insights into the underlying spectral distributions and discrimination between spectra of various samples. The data are the Raman images of non-alcoholic fatty liver disease (NAFLD) of liver tissues of three different types of rats that were fed with three different diets. URF provides the proximity matrix (representing the similarity among the super pixel spectra (obtained by the extended simple linear iterative clustering (SLIC)), which is then transformed into Euclidean space, and taken as the input of the RDT clustering to extract the meaningful groups having the spectra with similar characteristics. Hierarchical clustering path plot is produced to show how the set of spectra hierarchically makes a group in full spectral dimension as a function of the number of clusters. We expect our proposed approach could be a nice tool for the quantitative analysis of biomedical Raman image data which could be effective to analyze various diseases.