Classification of hyperspectral images for detection of hepatic carcinoma cells based on spectral-spatial features of nucleus

Abstract

A distinguishing characteristic of normal and cancer cells is the difference in their nuclear chromatin content and distribution. This difference can be revealed by the transmission spectra of nuclei stained with a pH-sensitive stain. Here, we used hematoxylin-eosin (HE) to stain hepatic carcinoma tissues and obtained spectral-spatial data from their nuclei using hyperspectral microscopy. The transmission spectra of the nuclei were then used to train a support vector machine (SVM) model for cell classification. Especially, we found that the chromatin distribution in cancer cells is more uniform, because of which the correlation coefficients for the spectra at different points in their nuclei are higher. Consequently, we exploited this feature to improve the SVM model. The sensitivity and specificity for the identification of cancer cells could be increased to 99% and 98%, respectively. We also designed an image-processing method for the extraction of information from cell nuclei to automate the identification process.