Identifying histological concepts on basal cell carcinoma images using nuclei based sampling and multi-scale descriptors

David Romo-Bucheli, Ricardo Moncayo, Angel Cruz-Roa, Eduardo Romero

Abstract: Histopathological sample examination involves a sequential analysis of several fields of view (FoV) at different magnification levels. Experts integrate this information by implicitly fusing morphometric and spatial features, mainly related with cell appearance, spatial distribution and organization. By performing this analysis a pathologist recognizes several micro structures such as follicle, epidermis, carcinoma and eccrine glands in basal skin tissue samples. In this article we present a new approach to histopathology classification using a multi-scale nuclei descriptor, located at a set of detected nuclei and constructed as a multiresolution pyramid. The method was evaluated in a multiclass challenging problem, i.e, identifying epidermis, hair follicle, eccrine glands and nodular carcinoma in 240 histopathology images of basal cell carcinoma. The experimental results show an average Area Under the ROC Curve (AUC) of 0.93 in a 6-fold cross-validation for the set of four classes.